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
|
|---|---|---|---|---|---|---|---|---|---|
1212.1604 | 1 | 1212 | 2012-12-07T13:17:37 | On some inequalities for s-logarithmically convex functions in the second sense via fractional integrals | [
"math.FA",
"math.CA"
] | In this paper, we establish some new Hadamard type inequalities for s-logarithmically convex functions in the second sense via fractional integrals by using Lemma 1 which has been proved by Sarikaya et al. in the paper [3]. | math.FA | math |
ON SOME INEQUALITIES FOR s-LOGARITHMICALLY
CONVEX FUNCTIONS IN THE SECOND SENSE VIA
FRACTIONAL INTEGRALS
HAVVA KAVURMACI AND MEVL UT TUNC¸
Abstract. In this paper, we establish some new Hadamard type inequalities
for s-logarithmically convex functions in the second sense via fractional inte-
grals by using Lemma 1 which has been proved by Sarıkaya et al. in the paper
[3].
1. Introduction
The following result is well known in the literature as Hadamard's inequality [1].
Theorem 1. Let f : I ⊂ R → R be a convex function on the interval I of real
numbers and a, b ∈ I with a < b. Then
(1.1)
f(cid:18) a + b
2 (cid:19) ≤
1
b − aZ b
a
f (x) dx ≤
f (a) + f (b)
2
The following definitions is well known in the literature:
Definition 1. A function f : I → R, ∅ 6= I ⊆ R, where I is a convex set, is said
to be convex on I if inequality
f (tx + (1 − t) y) ≤ tf (x) + (1 − t) f (y)
holds for all x, y ∈ I and t ∈ [0, 1].
In [2], Akdemir and Tun¸c were introduced the class of s-logarithmically convex
functions in the first and second sense as the following:
Definition 2. A function f : I ⊂ R0 → R+ is said to be s-logarithmically convex
in the first sense if
(1.2)
for some s ∈ (0, 1], where x, y ∈ I and αs + βs = 1.
f (αx + βy) ≤ [f (x)]αs
[f (y)]βs
Definition 3. A function f : I ⊂ R0 → R+ is said to be s-logarithmically convex
in the second sense if
(1.3)
f (tx + (1 − t) y) ≤ [f (x)]ts
for some s ∈ (0, 1], where x, y ∈ I and t ∈ [0, 1].
[f (y)](1−t)s
2000 Mathematics Subject Classification. 26D10, 26A15, 26A16, 26A51.
Key words and phrases. Hadamard's inequality, s-geometrically convex functions .
1
2
HAVVA KAVURMACI AND MEVL UT TUNC¸
Clearly, when taking s = 1 in Definition 2 or Definition 3, then f becomes the
standard logarithmically convex function on I.
Definition 4. Let f ∈ L1[a, b]. The Riemann-Liouville integrals J α
order α > 0 with a ≥ 0 are defined by
a+ f and J α
b−f of
and
J α
a+ f (x) =
1
Γ(α)
J α
b−f (x) =
1
Γ(α)
∞
x
Za
b
Zx
(x − t)α−1 f (t)dt, x > a
(t − x)α−1 f (t)dt, x < b
respectively where Γ(α) =
e−uuα−1du. Here is J 0
a+ f (x) = J 0
b−f (x) = f (x).
In the case of α = 1, the fractional integral reduces to the classical integral. For
some recent results connected with fractional integral inequalities see [3]- [11].
In [3], Sarıkaya et. al. proved the following results for fractional integrals.
Lemma 1. Let f : [a, b] → R be a differentiable mapping on (a, b) with a < b. If
f ′ ∈ L [a, b] , then the following equality for fractional integrals holds:
R0
f (a) + f (b)
−
Γ (α + 1)
2 (b − a)α [J α
b− f (a) + J α
a+ f (b)]
[(1 − t)α − tα] f ′ (ta + (1 − t) b) dt.
=
b − a
2
2
Z 1
0
Theorem 2. Let f : [a, b] → R be a differentiable mapping on (a, b) with a < b. If
f ′ is convex on [a, b] , then the following inequality for fractional integrals holds:
−
f (a) + f (b)
2
b − a
(cid:12)(cid:12)(cid:12)(cid:12)
2 (α + 1)(cid:18)1 −
b− f (a) + J α
Γ (α + 1)
2 (b − a)α [J α
2α(cid:19) [f ′ (a) + f ′ (b)] .
1
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
=
In the present paper, we will establish several Hermite-Hadamard type inequali-
ties for the class of functions whose derivatives in absolute value are s-logarithmically
convex functions in the first and second sense via Riemann-Liouville fractional in-
tegral.
2. Hadamard type inequalities for s-logarithmically convex
functions
Theorem 3. Let I ⊃ [0, ∞) be an open interval and f : I → (0, ∞) is differentiable.
If f ′ ∈ L [a, b] and f ′ is s-logarithmically convex functions in the second sense on
ON SOME INEQUALITIES
3
[a, b] for some fixed s ∈ (0, 1] and µ, η > 0 with µ + η = 1, then the following
inequality for fractional integrals with α > 0 holds:
f (a) + f (b)
2
−
Γ (α + 1)
2 (b − a)α [J α
b− f (a) + J α
(2.1)
≤
(cid:12)(cid:12)(cid:12)(cid:12)
b − a
2 (Z 1/2
0
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
1
µ dt +Z 1
1/2
µ [(1 − t)α − tα]
µ [tα − (1 − t)α]
1
µ dt
+ η × f ′ (b)
s
η ψ(cid:18) s
η
,
s
η(cid:19)(cid:27)
where
(2.2) Ψ (ψ) =(cid:26) 1,
ψ−1
ln ψ ,
ψ = 1,
0 < ψ < 1
and ψ (u, v) = f ′ (a)u f ′ (b)−v , u, v > 0.
Proof. By Lemma 1 and since f ′ is s-logarithmically convex functions in the second
sense on [a, b] , we have
≤
≤
≤
(2.3)
(cid:12)(cid:12)(cid:12)(cid:12)
0
0
2
2
b − a
b − a
Z 1
Z 1
2 (Z 1/2
+Z 1
b − a
1/2
0
f (a) + f (b)
2
−
Γ (α + 1)
2 (b − a)α [J α
b− f (a) + J α
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
(1 − t)α − tα f ′ (ta + (1 − t) b) dt
(1 − t)α − tα f ′ (a)ts
f ′ (b)(1−t)s
dt
[(1 − t)α − tα] f ′ (a)ts
f ′ (b)(1−t)s
dt
[tα − (1 − t)α] f ′ (a)ts
f ′ (b)(1−t)s
dt) ,
for all t ∈ [0, 1] . Using the well known inequality mn ≤ µm
side of (2.3), we have
1
µ + ηn
1
η , on the right
≤
=
0
2
b − a
f (a) + f (b)
(cid:12)(cid:12)(cid:12)(cid:12)
2 (Z 1/2
+Z 1
2 (Z 1/2
+ ηZ 1
b − a
1/2
0
0
−
Γ (α + 1)
2 (b − a)α [J α
b−f (a) + J α
µ [(1 − t)α − tα]
1
µ dt +Z 1/2
0
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
η f ′ (a)
ts
η f ′ (b)
(1−t)s
η
dt
µ [tα − (1 − t)α]
1
µ dt +Z 1
1/2
η f ′ (a)
ts
η f ′ (b)
(1−t)s
η
dt)
µ [(1 − t)α − tα]
f ′ (a)
ts
η f ′ (b)
(1−t)s
η
µ [tα − (1 − t)α]
1
µ dt
1/2
1
µ dt +Z 1
dt(cid:27)
If 0 < λ ≤ 1, 0 < u, v ≤ 1, then
(2.4)
λuv
≤ λuv.
4
HAVVA KAVURMACI AND MEVL UT TUNC¸
When ψ (u, v) ≤ 1, by (2.4), we get that
(2.5)
f ′ (a)
ts
η f ′ (b)
(1−t)s
η
dt ≤Z 1
0
f ′ (a)
st
η f ′ (b)
s(1−t)
η
dt = f ′ (b)
s
η ψ(cid:18) s
η
,
s
η(cid:19) .
(cid:3)
From (2.3) to (2.5), (2.1) holds.
Remark 1. If we take α = 1, in Theorem 3, then the inequality (2.1) become the
inequality
f (a) + f (b)
2
−
1
b − aZ b
a
≤
b − a
2
(cid:20) µ2
µ + 1
+ η × f ′ (b)
s
η ψ(cid:18) s
η
,
s
η(cid:19)(cid:21) .
The corresponding version for powers of the absolute value of the first derivative
f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
is incorporated in the following result:
Z 1
0
(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)
Theorem 4. Let I ⊃ [0, ∞) be an open interval and f : I → (0, ∞) is differentiable.
If f ′ ∈ L [a, b] and f ′ is s-logarithmically convex functions in the second sense on
[a, b] for some fixed s ∈ (0, 1] and µ, η > 0 with µ + η = 1 and p, q > 1, then the
following inequality for fractional integrals with α > 0 holds:
(2.6)
≤
f (a) + f (b)
2
b − a
2 (αp + 1)
1
p
−
Γ (α + 1)
2 (b − a)α [J α
f ′ (b)s (ψ (sq, sq))
1
q
b− f (a) + J α
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
where 1/p + 1/q = 1, and ψ (u, v) is defined as in (2.2).
Proof. By Lemma 1 and since f ′ is s-logarithmically convex functions in the second
sense on [a, b] , we have
(2.7)
≤
b − a
2
Z 1
0
f (a) + f (b)
2
−
Γ (α + 1)
2 (b − a)α [J α
b− f (a) + J α
(1 − t)α − tα f ′ (a)ts
f ′ (b)(1−t)s
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
dt
for all t ∈ [0, 1] . Using the well known Holder inequality, on the right side of (2.7)
and making the change of variable we have
(2.8)
≤
(cid:12)(cid:12)(cid:12)(cid:12)
f (a) + f (b)
2
b − a
2 (cid:18)Z 1
0
−
Γ (α + 1)
2 (b − a)α [J α
(1 − t)α − tαp dt(cid:19)
1
p (cid:18)Z 1
0
b− f (a) + J α
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
f ′ (a)qts
It is know that for α, t1, t2 ∈ [0, 1] ,
f ′ (b)q(1−t)s
1
q
.
dt(cid:19)
tα
1 − tα
2 ≤ t1 − t2α ,
therefore
(2.9)
Z 1
0
(1 − t)α − tαp dt ≤Z 1
0
1 − 2tαp dt =
1
αp + 1
.
ON SOME INEQUALITIES
5
Since f ′ is s-logarithmically convex functions on [a, b] and ψ (u, v) ≤ 1, we obtain
(2.10)
Z 1
0
f ′ (a)qts
f ′ (b)q(1−t)s
dt ≤ f ′ (b)sq ψ (sq, sq)
From (2.8) to (2.10), (2.6) holds.
(cid:3)
A different approach leads to the following result.
Theorem 5. Let I ⊃ [0, ∞) be an open interval and f : I → (0, ∞) is differentiable.
If f ′ ∈ L [a, b] and f ′q is s-logarithmically convex functions in the second sense
on [a, b] for some fixed s ∈ (0, 1] and µ, η > 0 with µ + η = 1 and q ≥ 1, then the
following inequality for fractional integrals with α > 0 holds:
(2.11)
(cid:12)(cid:12)(cid:12)(cid:12)
≤
f (a) + f (b)
2
b − a
q−(1−α)(q−1)
q
2
b−f (a) + J α
−
Γ (α + 1)
2 (b − a)α [J α
α + 1 (cid:19)1− 1
(cid:18) 2α − 1
q (cid:18) µ2
α + µ
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
sq
+ η f ′ (b)
η ψ(cid:18) sq
η
1
q
,
sq
η (cid:19)(cid:19)
where ψ (u, v) is defined as in (2.2).
Proof. By Lemma 1 and using the well known power mean inequality, we have
(cid:12)(cid:12)(cid:12)(cid:12)
≤
≤
b − a
0
2
Z 1
2 (cid:18)Z 1
0
b − a
f (a) + f (b)
2
−
Γ (α + 1)
2 (b − a)α [J α
b− f (a) + J α
(1 − t)α − tα f ′ (ta + (1 − t) b) dt
(1 − t)α − tα dt(cid:19)
1− 1
q (cid:18)Z 1
0
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
(1 − t)α − tα f ′ (ta + (1 − t) b)q dt(cid:19)
1
q
It is easily check that
Z 1
0
(1 − t)α − tα dt =
2
α + 1(cid:18)1 −
1
2α(cid:19) .
Since f ′q is s-logarithmically convex and using the well known inequality mn ≤
µm
η , we obtain
µ + ηn
1
1
Z 1
0
(1 − t)α − tα f ′ (ta + (1 − t) b)q dt ≤ Z 1
≤ Z 1
≤ µZ 1
0
0
0
It is easily check that
(1 − t)α − tα f ′ (a)qts
f ′ (b)q(1−t)s
dt
1 − 2tα f ′ (a)qts
f ′ (b)q(1−t)s
dt
1 − 2t
α
µ dt + ηZ 1
0
f ′ (a)
qts
η f ′ (b)
q(1−t)s
η
dt.
µZ 1
0
1 − 2t
α
µ dt = µ
1
α
µ + 1
=
µ2
α + µ
.
6
HAVVA KAVURMACI AND MEVL UT TUNC¸
Afterwards, when ψ (u, v) ≤ 1, by (2.4), we get that
(2.12)
f ′ (a)
qts
η f ′ (b)
q(1−t)s
Z 1
0
Therefore
f (a) + f (b)
2
−
(cid:12)(cid:12)(cid:12)(cid:12)
≤
≤
b − a
2 (cid:18) 2
2 (cid:18) 2
α + 1(cid:18)1 −
α + 1(cid:18)1 −
b − a
which comletes the proof.
η
0
dt ≤Z 1
Γ (α + 1)
2 (b − a)α [J α
2α(cid:19)(cid:19)1− 1
2α(cid:19)(cid:19)1− 1
1
1
f ′ (a)
sqt
η f ′ (b)
sq(1−t)
η
dt = f ′ (b)
sq
η ψ(cid:18) sq
η
,
sq
η (cid:19) .
b− f (a) + J α
1 − 2t
α
a+ f (b)](cid:12)(cid:12)(cid:12)(cid:12)
µ dt + ηZ 1
η ψ(cid:18) sq
η
sq
0
+ η f ′ (b)
0
q (cid:18)µZ 1
q (cid:18) µ2
α + µ
1
q
dt(cid:19)
f ′ (a)
qts
η f ′ (b)
q(1−t)s
η
1
q
,
sq
η (cid:19)(cid:19)
(cid:3)
References
[1] J. Hadamard: ´Etude sur les propri´et´es des fonctions enti`eres et en particulier d'une fonction
consider´ee par Riemann, J. Math Pures Appl., 58, (1893) 171 -- 215.
[2] A. O. Akdemir and M. Tun¸c: On some integral inequalities for s-logarithmically convex
functions, submitted.
[3] Sarıkaya, M. Z., Set, E., Yaldiz, H. and Ba¸sak, N.: Hermite-Hadamard's inequalities for frac-
tional integrals and related fractional inequalities, Mathematical and Computer Modelling,
In Press.
[4] Belarbi, S. and Dahmani, Z.: On some new fractional integral inequalities, J. Ineq. Pure and
Appl. Math., 10(3), Art. 86 (2009).
[5] Dahmani, Z.: New inequalities in fractional integrals, International Journal of Nonlinear
Science, 9(4), 493-497 (2010).
[6] Dahmani, Z.: On Minkowski and Hermite-Hadamard integral inequalities via fractional inte-
gration, Ann. Funct. Anal. 1(1), 51-58 (2010).
[7] Dahmani, Z., Tabharit, L. and Taf, S.: Some fractional integral inequalities, Nonl. Sci. Lett.
A., 1(2), 155-160 (2010).
[8] Dahmani, Z., Tabharit, L. and Taf, S.: New generalizations of Gruss inequality using
Riemann-Liouville fractional integrals, Bull. Math. Anal. Appl., 2(3), 93-99 (2010).
[9] Ozdemir, M.E., Kavurmacı, H. and Avcı, M.: New inequalities of Ostrowski type for map-
pings whose derivatives are (α, m)-convex via fractional integrals, RGMIA Research Report
Collection, 15, Article 10, 8 pp (2012).
[10] Ozdemir, M.E., Kavurmacı, H. and Yıldız, C¸ .: Fractional integral inequalities via s-convex
functions, arXiv:1201.4915v1 [math.CA] 24 Jan 2012.
[11] Ozdemir, M.E., Dragomir, S. S. and Yıldız, C¸ .: The Hadamard inequality for convex function
via fractional integrals, RGMIA Volume 15, art.14, 2012.
Agrı Ibrahim C¸ ec¸en University, Faculty of Science and Arts, Department of Math-
ematics, Agrı, Turkey.
E-mail address: [email protected]
Kilis 7 Aralık University, Faculty of Science and Arts, Department of Mathematics,
Kilis, 79000, Turkey.
E-mail address: [email protected]
|
1808.06225 | 4 | 1808 | 2019-11-14T13:20:46 | Inversion problem in measure and Fourier-Stieltjes algebras | [
"math.FA"
] | In this paper we study the inversion problem in measure and Fourier-Stieltjes algebras from qualitative and quantitative point of view extending the results obtained by N. Nikolski. | math.FA | math |
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES
ALGEBRAS
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
Abstract. In this paper we study the inversion problem in measure and Fourier -- Stieltjes
algebras from qualitative and quantitative point of view extending the results obtained
by N. Nikolski in [N].
1. Introduction
We are going to collect first some basic facts from Banach algebra theory and harmonic
analysis in order to fix the notation (our main reference for Banach algebra theory is
[Ż], for harmonic analysis check [R]). For a commutative unital Banach algebra A, the
Banach algebra of all complex-valued Borel regular measures equipped with the convolution
Gelfand space of A (the set of all multiplicative-linear functionals endowed with weak∗-
topology) will be denoted by △(A) and the Gelfand transform of an element x ∈ A is
a surjection bx : △(A) → σ(x) defined by the formula: bx(ϕ) = ϕ(x) for ϕ ∈ △(A),
where σ(x) := {λ ∈ C : µ − λδ0 is not invertible} is the spectrum of an element x. Let
G be a locally compact Abelian group with its unitary dual bG and let M(G) denote the
product and the total variation norm. It is also a ∗-algebra with involution µ → eµ defined
for any Borel set E ⊂ G by eµ(E) := µ(−E). A measure µ ∈ M(G) is called hermitian, if
µ = eµ or, equivalently, its Fourier-Stieltjes transform is real-valued. The Fourier-Stieltjes
transform will be treated as a restriction of the Gelfand transform to bG. Note that we
have a direct sum decomposition M(G) = Mc(G) ⊕ Md(G) where Mc(G) is the ideal
Date: November 15, 2019.
2010 Mathematics Subject Classification. Primary 43A05; Secondary 43A30.
Key words and phrases. Inverse, Measure Algebras, Fourier -- Stieltjes algebras.
PO was supported by foundations managed by The Royal Swedish Academy of Sciences. MW was
partially supported by National Science Centre (NCN) grant 2016/21/N/ST1/02499, long term structural
funding -- Methusalem grant of the Flemish Government -- and by European Research Council Consolidator
Grant 614195 RIGIDITY..
1
2
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
of continuous (non-atomic) measures and Md(G) is the subalgebra of discrete (atomic)
measures. For µ ∈ M(G) we will write µ = µc + µd with µc ∈ Mc(G) and µd ∈ Md(G).
We recall the problem investigated by N. Nikolski in [N], which will be our main point
of interest.
Problem 1. Let µ ∈ M(G) satisfy kµk ≤ 1 and suppose that inf γ∈bG bµ(γ) = δ > 0. What
is the minimal value of δ0 > 0 such that for any δ > δ0 the measure µ is automatically
invertible? What can be said about the norm of the inverse?
It is clear that δ0 < 1
2 ν where ν
is a probability measure with non-negative Fourier-Stieltjes transform satisfying σ(ν) = D
2 is not enough as the example of a measure µ := 1
2 δ0 + 1
shows (here δ0 is the Dirac delta at the point 0). On the other hand, it was proved in [N] that
any δ > 1√2 does the job and moreover the norm of the inverse is bounded by (2δ2 − 1)−1.
For the readers convenience we will reprove this result now with a slightly simpler (but
less general) approach. Let µ ∈ M(G) satisfy kµk ≤ 1 and inf γ∈bG bµ(γ) = δ > 1√2. Then,
by the generalization of Wiener lemma (see 5.6.9 in [R]):
(µ ∗ eµ)({0}) = Xx∈Gµ({x})2 = lim
α ZΓcfα(γ)bµ(γ)2dγ ≥ δ2 >
1
2
,
where {fα} is a system of continuous positive-definite functions with compact supports
subordinated to a neighborhood base {Vα} of 0 with fα(0) = 1. The second ingredient is
a version of Lemma 1.4.3 from [N] adapted to our situation.
Lemma 1.1. Let µ ∈ M(G) and let µ = λδ0 + ν where ν({0}) = 0 and 1
kµk ≤ 1. Then µ is invertible and kµ−1k ≤ 1
By the lemma, k(µ ∗ eµ)−1k ≤ 1
observe that µ−1 = (µ ∗ eµ)−1 ∗ eµ.
2δ−1 .
2δ2−1 and in order to finish the argument one has only to
However, the minimal value of δ0 > 0 seems to be unknown and the question on improv-
2 < δ ≤ λ ≤
ing the aforementioned bounds is stated in [N].
The first aim of this paper is to give a proof of the following fact (Theorem 2.4):
if
2 then µ is invertible, which is the final
µ ∈ M(G) satisfies kµk ≤ 1 and inf γ∈bG bµ(γ) > 1
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
3
solution of the problem (proving δ0 = 1
sharp.
2) -- in view of the previous discussion the result is
In the next part of this article we attack the problem of estimating the norm of the
inverse. In addition to the upper bounds mentioned above, the paper [N] contains a result
on the lower bound. We will not reproduce the whole discussion here but what is crucial
for us is that one cannot hope for a bound better than (2δ − 1)−1 and the norm-controlled
inversion cannot hold for any δ ≤ 1
2 for any infinite locally compact Abelian group. The
most important result is Theorem 2.8 which states that under additional group-theoretic
assumption the norm-controlled inversion holds true for δ > −1+√33
≃ 0, 593 - this improves
the result of N. Nikolski (δ > 1√2 ≃ 0, 707). Later, we show in Theorem 2.14 that for
measures with discrete parts supported on independent subsets of locally compact Abelian
8
groups consisting of elements of infinite order we obtain the optimal bound. The section is
concluded with a discussion of another special case -- real measures on groups of exponent
two.
The last, independent part of the paper is devoted to the study of analogous problems
for Fourier-Stieltjes algebras. Particularly, we prove in Theorem 3.2 that the qualitative
2. We also prove (Theorem 3.7) the
inversion problem has a positive solution for any δ > 1
norm-controlled inversion for δ > 1√2 in this context.
2. Measure algebras
2.1. Qualitative result.
is invertible.
Lemma 2.1. Let µ ∈ M(G) satisfy kµk ≤ 1 and suppose that inf γ∈Γ cµd(γ) > 1
Proof. Recalling that σ(µd) = σMd(G)(µd) and △(Md(G)) = △(l1(G)) = bbG (the Bohr
compactfication of bG) we get
2. Then µ
(2.1)
∀ϕ∈△(M (G)) cµd(ϕ) >
1
2
.
4
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
Of course, 1
2 < r(µd) ≤ kµdk so
(2.2)
kµck = kµk − kµdk <
1
2
.
Suppose now, towards the contradiction, that µ is not invertible. Then 0 ∈ σ(µ) and there
exists ϕ0 ∈ △(M(G)) for which ϕ0(µ) = 0. This gives ϕ0(µd) = −ϕ0(µc) implying that
ϕ0(µd) = ϕ0(µc). But by (2.1) we have ϕ0(µd) > 1
2 and by (2.2) we get ϕ0(µc) ≤
r(µc) ≤ kµck < 1
2 which is a contradiction.
(cid:3)
Here we need to cite the theorem of I. Glicksberg and I. Wik (see (2) in [GW]).
A straightforward corollary of this theorem is the following observation.
Theorem 2.2. Let µ ∈ M(G). Then cµd(bG) ⊂ bµ(bG).
Corollary 2.3. Let µ ∈ M(G). Then inf γ∈bG cµd(γ) ≥ inf γ∈bG bµ(γ).
Theorem 2.4. Let µ ∈ M(G) satisfy kµk ≤ 1 and inf γ∈bG bµ(γ) > 1
Proof. By Corollary 2.3 we have inf γ∈bG cµd(γ) ≥ inf γ∈bG bµ(γ) > 1
We are prepared now to ultimately solve the inversion problem.
just need to apply Lemma 2.1.
2. Then µ is invertible.
2. To finish the proof we
(cid:3)
2.2. Quantitative results.
2.2.1. The general case. In this section we will attack the problem of improving the con-
stant δ0 giving the following assertion: if µ ∈ M(G) satisfy kµk ≤ 1 and inf γ∈bG bµ(γ) ≥
δ > δ0 then kµ−1k can be bounded as a function of δ (norm-controlled inversion). As it was
explained in the introduction, if G is non-discrete, then δ0 ≥ 1
2. In addition, a non-trivial
argument from [N] shows that the same statement holds true for arbitrary infinite discrete
groups. Let us also recall that δ0 ≤ 1√2
We start with an elementary lemma.
.
Lemma 2.5. Let (xn)∞n=1 be a non-increasing sequence of non-negative numbers satisfying
(2.3)
∞Xn=1
xn ≤ 1,
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
5
(2.4)
x2
n ≥ δ2, where 1 ≥ δ >
1
2
.
∞Xn=1
Then the following inequalities hold true
(2.5)
(2.6)
x1 ≥ δ2,
x1 + x2 ≥ δ.
Proof. In order to prove (2.5) we use consecutively (2.4), monotonicity of the sequence
(xn)∞n=1 and then (2.3)
δ2 ≤
x2
n ≤ x1
xn ≤ x1.
∞Xn=1
∞Xn=1
For the second inequality we first observe that by (2.4) and the monotonicity:
δ2 ≤
∞Xn=1
x2
n = x2
1 +
Now, we apply (2.3) which gives
∞Xn=2
x2
n ≤ x2
1 + x2
∞Xn=2
xn.
(2.7)
δ2 ≤ x2
1 + x2(1 − x1).
If x1 ≥ δ then inequality (2.6) is obvious, so we are allowed to assume that x1 < δ and
then (2.7) can be rewritten as
(2.8)
Thus, it is enough to justify
(2.9)
This is equivalent to
x2 ≥
1
δ2 − x2
1 − x1
.
x1 +
1
δ2 − x2
1 − x1 ≥ δ for x1 ∈ [δ2, δ].
2x2
1 − (1 + δ)x1 + δ − δ2 ≤ 0 for x1 ∈ [δ2, δ].
This is a quadratic inequality with the discriminant (3δ−1)2 > 1
lead to x1 ∈ [ 1−δ
the last inequality simply follows from the assumption δ > 1
2 , δ] and therefore the assertion will be proved as long as 1−δ
4. Elementary calculations
2 ≤ δ2. However,
2 establishing (2.6) and finishing
(cid:3)
the proof of the lemma.
This lemma can be applied to obtain the following proposition.
6
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
µd =
2 for every γ ∈ bG. If
anδτn with a1 ≥ a2 ≥ . . . and for some τn ∈ G,
Proposition 2.6. Let G be a locally compact Abelian group and let µ ∈ M(G) satisfy
kµk ≤ 1 and bµ(γ) ≥ δ > 1
∞Xn=1
then a1 ≥ δ2 and a1 + a2 ≥ δ.
Proof. Clearly, kµdk ≤ 1 which is equivalent to the assumption (2.3) from Lemma 2.5.
We can treat µd as an element of l2(Gd) (Gd is the original group G equipped with the
It is well-known that (Gd) is canonically
discrete topology). Then cµd ∈ L2(cid:16)(Gd)(cid:17).
isomorphic to bbG -- the Bohr compactification of bG. By Theorem 2.2 we have cµd(γ) ≥ δ
for every γ ∈ bG and since bG is dense in bbG we obtain by Parseval's identity:
∞Xn=1an2 =ZbbG cµd2(x)dx ≥ δ2 (here dx is the normalized Haar measure on bbG),
which proves (2.4) from Lemma 2.5 and enables us to use the assertion of the lemma to
finish the argument.
(cid:3)
In order to proceed further, we need yet another technical lemma.
Lemma 2.7. Let G be a locally compact Abelian group and for x ∈ G define
Sx = {γ(x) : γ ∈ bG} ⊂ {z ∈ C : z = 1}.
If x is of order n < ∞ then Sx = Zn and if x is of infinite order then Sx is a dense subgroup
of {z ∈ C : z = 1}.
Proof. In case of element x ∈ G of finite order n we simply observe that the closed subgroup
generated by x is isomorphic to Zn and as Znb = Zn the result follows by the well-known
fact that characters on the closed subgroup can be extended to the whole group.
If x is of infinite order we formally verify first that Sx is a subgroup of the circle group and
since every infinite subgroup of the circle group is dense it is enough to show #Sx = ∞.
Let H be a closed subgroup of G generated by x. By definition, H is a monothetic group1
and by the results from Section 2.3 in [R] we obtain that either H = Z or H is a compact
1A locally compact Abelian group is called monothetic if it contains a dense homomorphic image of Z.
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
7
group whose dual is a subgroup of Td (the circle group with the discrete topology). The
first case is elementary and for the second one we continue as follows: since H is infinite, cH
is also infinite. Now, if cH contains an element of infinite order then we are done, as every
character in cH is uniquely defined by its action on x. Otherwise, cH contains elements of
arbitrarily high order and again since x topologically generates H we obtain that for a
fixed γ0 ∈ cH the set {γk
0 (x) : k ∈ N} ⊂ Sx has the same number of elements as the order
of γ0, which finishes the proof.
(cid:3)
We are ready now to prove one of the two main theorems of this section. Note that
the group-theoretic assumption from the next theorem is automatically satisfied for some
classical groups such as R and Z.
Theorem 2.8. Let µ ∈ M(G) satisfy kµk ≤ 1 and bµ(γ) > δ > 1
If the order of element τ2 − τ1 is infinite then
anδτn, a1 ≥ a2 ≥ . . . .
∞Xn=1
µd =
2 for every γ ∈ bG. Let
1 − δ + √17δ2 + 6δ − 7
a1 ≥
4
1
3
2
≥
1
2
and
δ −
2
3
,
for δ >
kµ−1k ≤
3δ − 2
−(1 + δ) + √17δ2 + 6δ − 7
2
kµ−1k ≤
for δ > −1 + √33
8
≃ 0, 593.
Proof. Since both multiplication by a constant of modulus one and convolution with a
Dirac delta do not violate the assumptions of the theorem and do not change the norm of
the inverse, we are allowed to work with a measure ν := cµ ∗ δ−τ1 where ca1 = a1 instead
of the original measure µ. Then
νd = a1δ0 + ca2δτ2−τ1 + ρ, where ρ := c
anδτn−τ1.
∞Xn=3
As the order of τ2 − τ1 is infinite, basing on Lemma 2.7, we are allowed to pick a sequence
(γn)∞n=1 ⊂ bG such that
(2.10)
bδτ2−τ1(γn) · ca2 = γn(τ2 − τ1) · ca2 −−−→n→∞ −a2.
8
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
By our assumptions and Theorem 2.2 we have
(2.11)
a1 + ca2bδτ2−τ1(γn) + bρ(γn) ≥ δ.
Using elementary inequalities, kρk = kµdk − a1 − a2 and passing with the parameter n
to infinity we obtain, by (2.10),
(2.12)
kµdk − 2a2 ≥ δ, which is equivalent to a2 ≤ kµdk − δ
2
≤
1 − δ
2
.
By Proposition 2.6 we get a1+a2 ≥ δ. Combining this with (2.12) we obtain a1 ≥ 3
and by Lemma 1.1 we get kµ−1k ≤ 1
In order to get a more refined bound we recall that by (2.8) (from the proof of Lemma 2.5)
3δ−2 for δ > 2
3.
2 δ− 1
2
and (2.12):
(2.13)
or equivalently a12 −
It is a quadratic inequality with discriminant D := 17
δ2 − a12
1 − a1
1 − δ
2 ≥
1 − δ
2 a1 +
2 δ − 7
4 δ2 + 3
be positive as δ > 1
2. With the aid of Viète formulas (note that 1−δ
1 − δ
2 − δ2 ≥ 0.
4, which is easily shown to
2 − δ2 < 0), we finally
get
a1 ≥
The elementary verification of the inequality
2 iff δ > −1+√33
1
8
finishes the argument by Lemma 1.1.
2 + √D
1−δ
2
2 +√D
1−δ
.
2
≥ 3
2 δ− 1
2 and the condition
1−δ
2 +√D
2
>
(cid:3)
It is worth to state the full formulation for the most classical case of G = Z.
Theorem 2.9. Let f ∈ A(T) satisfy kfk ≤ 1 and f(t) ≥ δ > 1
2 for every t ∈ T. Then
1
for δ >
2
3
,
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤
3δ − 2
−(1 + δ) + √17δ2 + 6δ − 7
2
for δ > −1 + √33
8
≃ 0, 593.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤
At last, we deal with the case of elements of finite order.
Theorem 2.10. Let µ ∈ M(G) satisfy kµk ≤ 1 and bµ(γ) > δ for every γ ∈ bG. Let
µd =
anδτn, a1 ≥ a2 ≥ . . . .
∞Xn=1
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
9
If the order of element τ2 − τ1 is equal to n ∈ N then
a1 ≥ δ −
kµ−1k ≤
2f(δ) − 1
1 − δ
2(1 − sin π
2n)
1
for δ >
:= f(δ) and
2n
2 − sin π
3 − sin π
2n
.
Proof. We perform the same reductions as in the proof of Theorem 2.8 and arrive to the
following point (check (2.9)):
a1 + ca2γ(τ2 − τ1) + bρ(γ) ≥ δ for every γ ∈ bG.
the set of all vertices of a regular n-gon inscribed in the unit circle with one vertex at the
As the order of τ2 − τ1 is equal to n we know, by Lemma 2.7, that {γ(τ2 − τ1) : γ ∈ bG} is
point 1. Using elementary geometry2 we find γ0 ∈ bG such that
csgn(a2)γ0(τ2 − τ1) + 1 ≤ 2 sin
π
2n
.
This gives
a1 + ca2γ0(τ2 − τ1) − a1 − a2 ≤ a1 + ca2γ(τ2 − τ1) − (a1 − a2) ≤ 2a2 sin
1−δ
Repeating the estimates from the proof of Theorem 2.8 we easily get a2 ≤
a1 ≥ δ − 1−δ
2(1−sin π
2n ) and finish the argument using Lemma 1.1.
2(1−sin π
π
2n
2n ),
(cid:3)
Remark 2.11. It is possible to obtain a more refined bound in Theorem 2.10 applying the
same method as in the proof of Theorem 2.8.
2.2.2. Independent case. We pass now to the analysis of a restricted class of measures
(in fact, the discrete part is the most important) in order to get the estimates for the
inverse. The first (trivial) lemma will help us to deal with measures with the discrete part
consisting of a finite sum of Dirac deltas at points constituting an independent set.
Lemma 2.12. Let x1, x2, . . . , xN , N ∈ N be a finite sequence of non-negative numbers
satisfying x1 ≥ x2 ≥ . . . ≥ xN and PN
2. Assume also that the
n=1 xn ≤ 1 and let δ > 1
2We estimate the distance of any point of the unit circle from the nearest vertex of the regular n-gon
inscribed in it by the distance of a midpoint of an arc joining two neighboring vertices.
Then
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(−1)n−1xn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ δ and (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
NXn=1
NXn=2
x1 ≥ δ +
NXn=2
xn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ δ
x1 −
xn.
10
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
following inequalities hold true:
Proof. As the sequence (xn)N
x1 ≥ PN
n=1(−1)n−1xn ≥ δ. Since PN
n=1 is non-increasing we have PN
n=1(−1)n−1xn ≥ 0. Of course,
n=1 xN ≤ 1 the second inequality proves the assertion.
(cid:3)
We also need a generalization of the classical Kronecker's approximation theorem (see
5.1.3 in [R]). By definition, a subset E ⊂ G is independent if for every choice of distinct
points x1, . . . , xk of E and integers n1, . . . , nk either
or
n1x1 = n2x2 = . . . = nkxk = 0
n1x1 + n2x2 + . . . + nkxk 6= 0.
Theorem 2.13. Suppose E is a finite independent subset of a locally compact Abelian
group G consisting of elements of infinite order and let f be a function on E satisfying
f(x) = 1 for every x ∈ E. Then, for any fixed ε > 0, there exists γ ∈ bG such that
γ(x) − f(x) < ε for every x ∈ E.
Now, we are ready to prove the theorem on the norm of the inverse of a measure with a
discrete part supported on an independent set of points of infinite order.
Theorem 2.14. Let µ ∈ M(G) satisfy kµk ≤ 1, inf γ∈bG bµ(γ) = δ > 1
of µd be independent and consist of elements of infinite order. Then kµ−1k ≤ 1
Proof. First of all, by Corollary 2.3 we have inf γ∈bG cµd(γ) ≥ δ. We will deal first with a
simpler case, so let us assume that the support of µd is finite, i.e.
2 and let the support
2δ−1 .
µd =
akδyk for some N ∈ N, complex numbers ak and yk ∈ G, k = 1, 2, . . . , N .
NXk=1
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
We can, of course, arrange the numbers (ak)N
k=1 so that a1 ≥ a2 ≥ . . . ≥ aN.
addition, as convolution of a measure with a Dirac delta and multiplication by a constant
of modulus one do not change the norm of a measure, the norm of the inverse and do not
violate the estimate bµ ≥ δ, we can assume, without losing the generality, the following
form of µd:
11
In
µd = a1δ0 +
akδyk where a1 > 0.
NXk=2
a1 +
a1 +
(2.14)
for k ∈ {2, . . . , N}.
2, any sequence of signs (ξn)N
n=2 and let (αk)N
γ0(yk) − αk ≤
Now, we perform elementary estimates
Let us now fix ε > 0 with δ − ε > 1
k=2 be a
sequence of complex numbers of modulus one such that αkak = ξkak for k = 2, . . . , N.
By Theorem 2.13 there is γ0 ∈ bG such that
ε
PN
k=2 ak
cµd(γ0) −(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
NXk=2akξk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
akγ0(yk)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) −(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
NXk=2akξk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
NXk=2
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
NXk=2akξk!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
a1 +
akγ0(yk)! − a1 +
NXk=2
NXk=2akγ0(yk) − akξk =
NXk=2akγ0(yk) − αk ≤ ε by (2.14).
NXk=2akγ0(yk) − akαk =
NXk=2akξk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ δ − ε.
Recalling that cµd(γ0) ≥ δ we get
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
This estimate holds true for every choice of signs ξk and every sufficiently small ε > 0 so
a1 +
a1 +
=
we obtained, in fact, the following conclusion:
a1 +
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
NXk=2akξk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ δ >
1
2
for every choice of signs ξk
Now, we are in position to use Lemma 2.12 which gives
The assertion of the theorem follows now from Lemma 1.1.
a1 ≥ δ +
N−1Xk=2 ak ≥ δ >
1
2
.
Let us move on now to the general case, so let µd be given by the formula (we use the same
12
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
reduction as before):
µd = a1δ0 +
akδyk where
∞Xk=2
a1 > 0, a1 ≥ a2 ≥ a3 ≥ ... and for every N ∈ N the set {yk}N
of elements of infinite order and Pn an < ∞. Put
NXk=2
νN = a1δ0 +
akδτk .
k=2 is independent, consists
2. Then there exists N ∈ N, N := N(ε) such that
Fix ε > 0 such that δ − ε > 1
inf γ∈bG cνN (γ) ≥ δ − ε. By the initial part of the proof we get
A moment of consideration (we can assume that N(ε) tends to ∞ as ε → 0) leads to
a1 ≥ δ − ε +
NXn=2an.
∞Xn=2an + δ ≥ δ
which finishes the proof.
a1 ≥
(cid:3)
The proof of the last theorem gives us a bit more.
Theorem 2.15. Let µ ∈ M(G) be a measure satisfying the assumptions of Theorem 2.14
and let a > 0 be the mass of the greatest atom in µd. Then
kµ−1k ≤
1
2(kµdk + δ − a) − 1
Remark 2.16. The assumptions of Theorem 2.14 might seem to be quite restrictive but for
the classical groups of harmonic analysis: Tn and Rn the situation covered by the theorem
is a generic one.
2.2.3. Groups of exponent two. In this section we will deal with a special case of real
measures on groups of exponent two (i.e. groups such that 2x = 0 for every x), which has
connection to the so-called dyadic analysis. As it will be shown in Theorem 2.18 we obtain
the norm-controlled inversion for every δ > 1
2.
We start with finite groups Zn
2 : for x, y ∈ Zn
ordering on Zn
2 . For convenience, we introduce the following (non-standard)
2 = {0, 1}n we define x < y if the last digit of x non-equal to
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
13
the corresponding digit of y is equal to 0. This order is sometimes called the colexicographic
order; the only difference is that we read the strings of digits from right to left. For example,
for n = 3 we have
000 < 100 < 010 < 110 < 001 < 101 < 011 < 111.
The most important property of this order is that the second half of strings end with 0,
and the second half ends with 1. What is more, if we have a string ending with 0, it is
easy to find the corresponding string that ends with 1 -- you need to shift by 2n−1.
the set of all homomorphisms f : Zn
We will also need a precise description of the dual group of Zn
2 which is by definition
2 → {−1, 1}. It is clear that every such f is uniquely
determined by its values on the standard system of generators of Zn
2 : 1000...0, 01000...0, ...,
000...1 and for every choice of signs from {−1, 1}n we get a homomorphism. Thus we can
write that every homomorphism f ∈ cZn
2 is bijectively identified with an element y ∈ Zn
2
by the formula fy(x) = (−1)<x,y>n where < x, y >n is a formal scalar product on Zn
2 (for
example: < 110, 011 >3= 1 · 0 + 1 · 1 + 0 · 1 = 1). Thus, the Fourier-Stieltjes transform of
δx for x ∈ Zn
2 is of the form:
We will use the following two symbolic operations on Zn
2 :
cδx(y) = (−1)<x,y>n for y ∈ Zn
2 .
(1) x′ for x ∈ Zn
(2) x0, x1 for x ∈ Zn−1
2
(respectively) after the last digit of x.
2 is the element of Zn−1
2
obtained from x by erasing the last digit.
is the element of Zn
2 obtained from x by inserting 0 or 1
Lemma 2.17. Let n be a positive integer and let µ ∈ M(Zn
kµk ≤ 1 and bµ(γ) ≥ δ > 1
support of µ is greater or equal then δ.
2 for every γ ∈ Zn
2 ) be a real measure satisfying
2 . Then the mass of the greatest atom in the
Proof. Let
µ =
2n
Xk=1
axk δxk,
where the elements xk ∈ Zn
2 are ordered as explained before.
14
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
Then, using the introduced notation, our assumption can be rewritten as:
∀l∈{1,...,2n}(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2n
(−1)<xk,xl>naxk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ δ.
Xk=1
The proof will proceed by induction on n.
The case of n = 1 is elementary as the inequalities ax1 + ax2,ax1 − ax2 ≥ δ for real
numbers ax1, ax2 clearly imply that the modulus of one of these numbers is greater than δ.
Suppose that we have already proven the assertion for s ≤ n − 1. By the definition of
our order, the last digit of xl is 0 for l ∈ {1, . . . , 2n−1} which gives for l ∈ {1, . . . , 2n−1}:
< x′l, x′k >n−1 for k ∈ {1, . . . , 2n−1}
< x′l, x′k−2n−1 >n−1 for k ∈ {2n−1 + 1, 2n}
Inserting this to (2.15), we obtain
< xl, xk >n=< x′l0, x′k0 >n=(
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2n−1Xk=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∀l∈{1,...,2n−1}
l∈{2n−1+1,...,2n}
2n−1Xk=1
∀
Using similar arguments for elements ending with 1, we get
(axk + axk+2n−1 )(−1)<x′
k,x′
.
l>n−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
l>n−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(axk − axk+2n−1 )(−1)<x′
k,x′
.
(2.15)
(2.16)
(2.17)
We perform a substitution bxk := axk + axk+2n−1 , ck := axk − axk+2n−1 for k ∈ {1, . . . , 2n−1}.
It is straightforward to verify that the original assertion (for n) reduces to two analogous
problems of reduced order (n − 1). Hence we are allowed to use the inductive hypothesis,
which yields bxk1 ≥ δ and cxk2 ≥ δ for some k1, k2 ∈ {1, . . . , 2n−1}. If k1 6= k2, we get
a contradiction with the norm assumption: 1 ≥ bxk1 + cxk2 ≥ 2δ > 1; indeed, note that
bxk1 + ck2 6 axk1 + axk1+2n−1 + axk2 + axk2+2n−1. Thus k1 = k2 and it is enough to
use the argument for n = 1 to finish the proof.
(cid:3)
Theorem 2.18. Let G be a locally compact Abelian group satisfying 2x = 0 for every
x ∈ G. Let µ ∈ M(G) be a real measure such that kµk ≤ 1, bµ(γ) ≥ δ > 0 for every
γ ∈ bG. Then the mass of the greatest atom in the support of µd is greater than or equal to
δ. Moreover, µ is invertible and kµ−1k ≤ 1
2δ−1 .
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
15
Proof. By Theorem 2.2 we have cµd(γ) ≥ δ for every γ ∈ bG and obviously kµdk ≤ 1. Let
µd have the following form:
(2.18)
µd =
∞Xk=1
Fix ε > 0 such that δ − ε > 1
2 and find Nε ∈ N for which
akδxk for some ak ∈ R and xk ∈ G.
∞Xk=Nε+1ak < ε.
Define µε as a sum as in (2.18) from k = 1 to Nε. Then clearly kµεk ≤ 1 and cµε(γ) ≥
2 for every γ ∈ bG. Now it is enough to observe that µε may be treated as an
δ − ε > 1
element of M(ZNε
2 ) and apply Lemma 2.17. To finish the proof we pass with ε to zero and
use Lemma 1.1.
(cid:3)
3. Fourier -- Stieltjes algebras
We switch now to Fourier -- Stieltjes algebras (the standard reference for this part is [Ey],
see also a recent monograph [KL]). Let G be a locally compact group and let B(G) be
the Fourier -- Stieltjes algebra of the group G i.e. the linear span of continuous positive-
definite functions on G equipped with the norm given by the duality B(G) = (C∗(G))∗,
where C∗(G) is the full group C∗-algebra of G. It is well-known that B(G) with pointwise
product is a commutative semisimple unital Banach algebra with G embedded in △(B(G))
via evaluation functionals. A continuous bounded function on G is called (weakly) almost
periodic if the set of all its translates is precompact in the (weak) uniform topology, the set
of all such functions will be denoted by AP (G) (W AP (G)). A classical result of Eberlein
(see [E] for the original work and [GL] for a different proof) states that B(G) ⊂ W AP (G).
Moreover, there exists an invariant mean M ∈ (W AP (G))∗ (the existence of the mean and
related matters are discussed in detail in the book [B]). Restricting this mean to B(G) we
obtain a decomposition B(G) = Bc(G) ⊕ (B(G) ∩ AP (G)) where Bc(G) = {f ∈ B(G) :
M(f) = 0} is a closed ideal in B(G) and B(G)∩AP (G) is a closed subalgebra. Here Bc(G)
is the analogue of the ideal of continuous measures and B(G)∩AP (G) is the analogue of the
16
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
subalgebra of discrete measures for commutative groups. The investigations on Fourier --
Stieltjes algebras are natural generalizations of those concerning measure algebras, as in
case of an Abelian group G we have a canonical isomorphism of B(G) with M(bG).
As we work already 'on the Fourier transform side', the variant of Problem 1 can be
rephrased in our context as follows.
Problem 2. Let G be a locally compact group and let f ∈ B(G) satisfy kfk ≤ 1 and
inf x∈G f(x) = δ > 0. What is the minimal value of δ0 > 0 such that for any δ > δ0 the
element f is automatically invertible? What can be said about the norm of the inverse?
3.1. Qualitative result. The inspection of the proof of Theorem 2.4 shows that in order
to obtain an analogue for Fourier -- Stieltjes algebras (and answer the first question from
Problem 2) we need only to show the version of Theorem 2.2 for Fourier-Stieltjes algebras
(indeed, we have only used the basic facts from Gelfand theory, the orthogonal decompo-
sition of a measure into continuous and discrete part and the density of bG in △(Md(G)) --
all of these facts have standard counterparts for Fourier -- Stieltjes algebras). However, the
original proof cannot be repeated verbatim and we use a different approach in the second
part of the argument.
Theorem 3.1. Let G be a locally compact group and let f ∈ B(G) have the decomposition
f = g + h with g ∈ (AP (G) ∩ B(G)) and h ∈ Bc(G). Then g(G) ⊂ f(G).
Proof. Without loss of generality, it is enough to show g(e) ∈ f(G) (the general case being
obtained via considering shifted f). First we prove the existence of distinct group elements
(xn)∞n=1 satisfying
(3.1)
h(x1) < ε and h(xnx−1
j ) < ε for j < n and n > 1.
Since h ∈ Bc(G) and the mean M satisfies M(f) ≥ 0 if f ≥ 0 we clearly have inf x∈G h(x) =
0 so there is no problem with the choice of x1. Suppose that we have already picked
x1, . . . , xn−1 and consider an auxiliary function u ∈ Bc(G) defined as follows:
u(x) =
n−1Xj=1 h(xx−1
j )2.
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
17
As u ∈ Bc(G) w also have inf x∈G u(x) = 0 and we are able to choose xn different from
x1, . . . , xn−1 or h(e) < ε which immediately finishes the proof.
Let us consider now the following set of functions in B(G) ∩ AP (G).
X = {gxn : n ∈ N} where gx(y) = g(xy) for x, y ∈ G.
Since g ∈ AP (G) this set is relatively compact in the uniform topology. Hence there exists
a subsequence (gxnk )k∈N which is a Cauchy sequence. It follows that for k > N we have
kgxnk+1 − gxnkk∞ < ε.
In particular,
But now,
gxnk+1
nk ) − gxnk (x−1
(x−1
nk ) = g(xnk+1x−1
nk ) − g(e) < ε.
f(xnk+1x−1
nk ) − g(e) = h(xnk+1x−1
nk ) + g(xnk+1x−1
nk ) − g(e) ≤ 2ε.
(cid:3)
For completeness, let us state a variant of Theorem 2.4 for Fourier-Stieltjes algebras.
2. Then f is invertible.
Theorem 3.2. Let G be a locally compact group and let f ∈ B(G) satisfy kfk ≤ 1 and
inf x∈G f(x) > 1
3.2. Quantitative results. The aim of this short section is to show that Nikolski's esti-
mate for δ > 1√2 can be achieved for Fourier-Stieltjes algebras as well. In order to do that,
we need to recall some facts about absolute continuity and L-spaces from [OW].
Definiton 3.3 (cf.
[Ex, Definition 2.] and [OW, Definition 4.7]). Let G be a locally
compact group and let f, g ∈ B(G). Let zs(f) and zs(g) be the central supports of f and
g, respectively, where f and g are viewed as functionals on C∗(G). We say that f is
absolutely continuous with respect to g (denoted by f ≪ g) if zs(f) ≤ zs(g). We call f
and g mutual ly singular if zs(f)zs(g) = 0.
Proposition 3.4 (cf. [OW, Proposition 4.18] and [OW, Corollary 4.19]). Let X ⊂ B(G)
be a closed subspace that is left and right invariant by the action of G. Then there exists
18
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
a central projection z ∈ W ∗(G) := (C∗(G))∗∗ such that X = z · B(G). Consequently, we
have a splitting B(G) = X ⊕ X⊥, where X⊥ := (1 − z) · B(G), which is orthogonal, i.e.
all elements from X⊥ are mutually singular with elements from X.
We are now able to show the following fact.
Proposition 3.5. Let f ∈ B(G). Then there exists an orthogonal decomposition f =
c1 + f0, where f0 and 1 are mutually singular. Moreover, c = m(f), where m is the unique
invariant mean on B(G).
Proof. Since the subspace consisting of constant functions is G-invariant, we get an orthog-
onal decomposition by the previous proposition. Denote the constant appearing therein
by ϕ(f). By uniqueness of the decomposition, it is easy to see that ϕ : B(G) → C is a
linear functional that satisfies ϕ(1) = 1. Since the decomposition is orthogonal, we get
kfk = ϕ(f) +kf0k, so ϕ is a continuous functional. Note also that ϕ is G-invariant (from
both left and right).
Indeed, for any g ∈ G the equality f = ϕ(f)1 + f0 implies that
g · f = ϕ(f)1 + g · f0, but also g · f = ϕ(g · f)1 + (g · f)0, hence ϕ(g · f) = ϕ(f) by
uniqueness; in the same way we obtain invariance from the right. This is almost exactly
the definition of an invariant mean, but we do not know if ϕ is a positive functional. We
will not show this directly; in the next lemma we will prove that there exists a unique
G-invariant functional that attains the value 1 at 1, hence it has to coincide with the
invariant mean.
(cid:3)
Lemma 3.6 (see also [LL, Proposition 4.10]). Let ϕ : B(G) → C be a linear functional
such that ϕ(1) = 1 and ϕ(f) = ϕ(g · f) = ϕ(f · g) for any g ∈ G. Then it is equal to the
unique invariant mean on B(G).
Proof. It suffices to show that there is at most one functional ϕ satisfying the assumptions
of the lemma. Recall that in order to show that two such functional are equal, it suffices to
show that their kernels coincide; this will force them to be proportional and the condition
ϕ(1) = 1 will finish the proof. Therefore we have to check that the kernel of ϕ is uniquely
specified by the conditions spelled out in the lemma. Note that kerϕ is a G-invariant
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
19
subspace, so it is of the form z · B(G) for some central projection in W ∗(G). Let ztriv be
the central projection such that ztrivB(G) = C1. We will show that z = 1 − ztriv. Since
ϕ(1) = 1, it cannot be the case that ztriv ≤ z. As ztriv is a minimal projection, it follows
that z ≤ 1 − ztriv. If the inequality was strict, then the codimension of the kernel would
be greater than 1, which cannot happen for a linear funcitonal, hence z = 1 − ztriv.
(cid:3)
With these tools we are able to conclude.
2δ2−1.
f k 6 1
Proposition 3.7. Let f ∈ B(G) satisfy kfk 6 1 and f > δ, where δ > 1√2. Then f is
invertible and k 1
Proof. As f > δ, f2 > δ2 > 1
2. By Proposition 3.5 we get f2 = m(f2)1 + f0.
Since the mean m(f) is characterised as being the unique constant that can be uniformly
approximated with convex combinations of translates of f, we get m(f2) ≥ δ2 > 1
2. As
kf2k 6 1, we get that kf0k ≤ 1 − δ2 < 1
2. This means that we can write the inverse of
f2 using the Neumann series (just like in Lemma 1.1) and quickly obtain the estimate for
the norm of the inverse by
(cid:3)
1
2δ2−1; it just remains to note that 1
f = f
f2 .
The proofs contained in this note clearly suggest that the most important information
concerning invertibility is contained in the discrete (or almost periodic) part. In the non-
commutative setting it is possible that the almost periodic part is very simple. Namely,
there is a class of groups, called minimally almost periodic, for which constants are the
only almost periodic functions; an example of such a group is SL2(R). For them we get a
better result.
2. Then f is invertible and k 1
Proposition 3.8. Let G be a minimally almost periodic group and let f ∈ B(G) satisfy
kfk 6 1 and f ≥ δ > 1
Proof. Let f = fap + fc be a decomposition of f into almost periodic and continuous parts,
and recall that m(fc) = 0. It follows that m(f) = m(fap). As the only almost periodic
functions are the constants, we get fap = m(f)1, hence m(f) = m(f) ≥ δ. We have an
2, so kfck ≤ 1 − δ < 1
orthogonal decomposition f = m(f)1 + fc, where m(f) ≥ δ > 1
f k ≤ 1
2δ−1 .
2
20
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
and the usual procedure involving the Neumann series proves invertibility and provides the
estimate.
(cid:3)
There are two more simple cases in which we get the same estimate for the inverse; this
time there are very strong assumptions about the function.
Proposition 3.9. Let G be a locally compact group and let f ∈ B(G) satisfy kfk ≤ 1 and
f > δ > 1
2δ−1 if one of the following conditions is
satisfied:
2. Then f is invertible and k 1
f k 6 1
(i) f is real and has constant sign;
(ii) the absolute value of f is constant.
Proof.
(i) We may assume that f is positive. Recall the decomposition from Proposition
3.5: f = m(f)1 + f0. Since f > δ, we get m(f) > δ, so f = m(f)(cid:16)1 + f0
inverted using the Neumann series, which also gives the estimate.
c , which gives k 1
(ii) In this case we have f f = c > δ2, hence 1
m(f )(cid:17) can be
2δ−1. Note
f k 6 1
δ2 6 1
that here the assumption δ > 1
f = f
2 is not necessary.
4. Concluding remarks
(1) The critical constant δ0 = 1
2 is the same as for measure algebras on 'analytic'
semigroups as discussed in part 2.2. of [N].
(cid:3)
(2) The problem of obtaining an estimate of the inverse for measures µ ∈ M(G) sat-
2 , 1√2(cid:17) seems to be very difficult even for
isfying kµk ≤ 1 and inf γ∈bG bµ(γ) ∈ (cid:16) 1
G = T and the considered measures being finite sums of Dirac deltas supported at
rational points of the circle. Then the estimation of the inverse is translated to an
elementary but surprisingly hard problem on bounding the entries of the inverses
of circulant matrices in terms of their eigenvalues.
It should be also noted that
it cannot be solved by simply proving that the mass of a single atom is greater
2 δ1 ∈ M(Z2) shows (such examples can be also
than δ as the example µ = 1
2 δ0 + i
constructed for non-torsion groups such as Z).
INVERSION PROBLEM IN MEASURE AND FOURIER -- STIELTJES ALGEBRAS
21
(3) Theorem 2.4 is sharp for any non-discrete locally compact Abelian group as it was
discussed before but Theorem 3.2 is not necessarily sharp even for non-compact
groups as there are examples of non-compact groups for which the Wiener-Pitt
phenomenon does not occur (check [C]). However, this Theorem is sharp for discrete
groups containing an infinite Abelian subgroup (consult [OW])
5. Acknowledgements
The authors would like to thank Maria Roginskaya and Michał Wojciechowski for many
valuable discussions. They would also like to thank Nikolai Nikolski for careful reading of
the manuscript and useful remarks. Parts of this work were completed during a visit of
Mateusz Wasilewski to Chalmers University of Technology; he thanks the university and
the host for a stimulating research environment.
References
[B]
[C]
[E]
[Ey]
[Ex]
[GW]
[GL]
[KL]
[LL]
R. B. Burckel: Weakly almost periodic functions on semigroups, Gordon and
Breach, New York, 1970
M. Cowling: The Fourier-Stieltjes algebra of a semisimple group, Colloq. Math.,
vol. 41, no. 1, pp. 89 -- 94, 1979.
W. F. Eberlein: Abstract ergodic theorems and weak almost periodic functions,
Transactions of the Amer. Math. Soc., vol. 67, pp. 217 -- 240, 1949.
P. Eymard: L'algebre de Fourier d'un groupe localement compact, Bull. Soc. Math.
France, vol. 92, pp. 181 -- 236, 1964.
R. Exel: The F. and M. Riesz theorem for C∗-algebras, J. Operator Theory, vol.
23, no.2, pp. 351 -- 368, 1990.
I. Glicksberg, I. Wik: The range of Fourier-Stieltjes transforms of parts of a mea-
sures, Lecture Notes in Mathematics, vol. 266, pp 73 -- 77, 1971.
Colin C. Graham, Anthony T. M. Lau: Relative weak compactness of orbits in
Banach spaces associated with locally compact groups, Trans. Amer. Math. Soc.,
vol.359, no. 3, pp. 1129 -- 1160, 2007.
E. Kaniuth, Anthony T. M. Lau: Fourier and Fourier-Stieltjes algebras on Locally
Compact Abelian Groups, AMS 2018.
Anthony T. M. Lau, J. Ludwig: Fourier-Stieltjes algebra of a topological group,
Adv. Math., vol. 229, no. 3, p. 2000 -- 2023, 2012.
22
[N]
PRZEMYSŁAW OHRYSKO AND MATEUSZ WASILEWSKI
N. Nikolski: In search of the invisible spectrum, Ann. Inst. Fourier (Grenoble), vol.
49, no. 6, pp. 1925 -- 1998, 1999.
[OW] P. Ohrysko, M. Wasilewski: Spectral theory of Fourier-Stieltjes algebras, J. Math.
Anal. Appl., vol. 473, no. 1, pp. 174 -- 200, 2019.
W. Rudin: Fourier analysis on groups, Wiley Classics Library, 1990.
W. Żelazko: Banach algebras, Elsevier Science Ltd and PWN, February 1973.
[R]
[Ż]
Chalmers University of Technology and the University of Gothenburg
E-mail address: [email protected]
KU Leuven
E-mail address: [email protected]
|
1302.0652 | 1 | 1302 | 2013-02-04T11:27:50 | Coupling and relaxed commutant lifting | [
"math.FA"
] | A Redheffer type description of the set of all contractive solutions to the relaxed commutant lifting problem is given. The description involves a set of Schur class functions which is obtained by combining the method of isometric coupling with results on isometric realizations. For a number of special cases, including the case of the classical commutant lifting theorem, the description yields a proper parameterization of the set of all contractive solutions, but examples show that, in general, the Schur class function determining the contractive lifting does not have to be unique. Also some sufficient conditions are given guaranteeing that the corresponding relaxed commutant lifting problem has only one solution. | math.FA | math |
COUPLING AND RELAXED COMMUTANT LIFTING
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
Abstract. A Redheffer type description of the set of all contractive solutions
to the relaxed commutant lifting problem is given. The description involves
a set of Schur class functions which is obtained by combining the method
of isometric coupling with results on isometric realizations. For a number of
special cases, including the case of the classical commutant lifting theorem,
the description yields a proper parameterization of the set of all contractive
solutions, but examples show that, in general, the Schur class function deter-
mining the contractive lifting does not have to be unique. Also some sufficient
conditions are given guaranteeing that the corresponding relaxed commutant
lifting problem has only one solution.
0. Introduction
This paper is devoted to the relaxed commutant lifting theorem in [14]. This
theorem is a generalization of the classical commutant lifting theorem [19], and it
includes as special cases the Treil-Volberg lifting theorem [20], and its weighted
version due to Biswas, Foias and Frazho [12].
To state the relaxed commutant lifting theorem, let us first recall the general
setup. The starting point is a lifting data set {A, T ′, U ′, R, Q} consisting of five
Hilbert space operators. The operator A is a contraction mapping H into H′, the
operator U ′ on K′ is a minimal isometric lifting of T ′ on H′, and R and Q are
operators from H0 to H, satisfying the following constraints
T ′AR = AQ and R∗R ≤ Q∗Q.
Given this data set the relaxed commutant lifting theorem in [14] states that there
exists a contraction B from H to K′ such that
(0.1)
ΠH′ B = A and U ′BR = BQ.
Here ΠH′ is the orthogonal projection from K′ onto H′. In fact, [14] provides an
explicit construction for a contraction B satisfying (0.1). In the sequel we say that
B is a contractive interpolant for {A, T ′, U ′, R, Q} if B is a contraction from H into
K satisfying (0.1).
In this paper we present a Redheffer type formula to describe the set of all
contractive interpolants for {A, T ′, U ′, R, Q}. In order to state our main results we
need some auxiliary operators. To this end, let D◦ be the positive square root of
1991 Mathematics Subject Classification. Primary 47A20, 47A57; Secondary
47A48.
Key words and phrases. commutant
lifting,
isometric coupling,
isometric realization,
parameterization.
1
2
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
Q∗Q − R∗R, and set
(0.2)
F = DAQH0
and F ′ =
D◦
DT ′ AR
DAR
H0.
Notice that F is a subspace of DA and F ′ is a subspace of D◦ ⊕ DT ′ ⊕ DA. Here
we follow the convention that for a contraction C, the symbol DC denotes the
positive square root of I − C∗C and DC stands for the closure of the range of DC.
Furthermore, D◦ = D◦H0. Since T ′AR = AQ, we know from formula (4.11) in [14]
that there exists a unique unitary operator ω mapping F onto F ′ such that
(0.3)
ω(DAQh) =
D◦
DT ′ AR
DAR
h,
h ∈ H0.
We also need the projections ΠT ′ and ΠA defined by
(0.4)
ΠT ′ = (cid:2) 0
I
0 (cid:3) :
D◦
DT ′
DA
→ DT ′ , ΠA = (cid:2) 0
0 I (cid:3) :
D◦
DT ′
DA
→ DA.
Notice that the previous definitions only relied upon the operators A, T ′, R and
Q. The minimal isometric lifting U ′ did not play a role. Recall that all minimal
isometric liftings of the same contraction are isomorphic. So without loss of gen-
erality, in our main theorem, we can assume that U ′ = V is the Sz.-Nagy-Schaffer
minimal isometric lifting of T ′ which acts on H′ ⊕ H 2(DT ′ ). The definitions of
a minimal isometric lifting and the Sz.-Nagy-Schaffer lifting are presented in the
next section. Finally, given Hilbert spaces U and Y, we write S(U , Y) for the set
of all operator-valued functions which are analytic on the open unit disk D and
whose values are contractions from U to Y. We refer to S(U , Y) as the Schur class
associated with U and Y. We are now ready to state our first main result.
Theorem 0.1. Let {A, T ′, V, R, Q} be a lifting data set, where V on H′ ⊕ H 2(DT ′ )
is the Sz.-Nagy-Schaffer minimal isometric lifting of T ′. Then all contractive in-
terpolants for this date set are given by
(0.5)
Bh = (cid:20)
ΠT ′ F (λ)(IDA − λΠAF (λ))−1DAh (cid:21) ,
Ah
h ∈ H,
where F is any function from the Schur class S(DA, D◦ ⊕ DT ′ ⊕ DA) satisfying
F (0)F = ω.
In general, formula (0.5) does not establish a one to one correspondence between
B and the parameter F . It can happen that different F 's yield the same B. For
instance, assume H0, H and H′ to be equal to C, let A, R and Q be the zero operator
on C, and take for T ′ the identity operator on C. Since T ′ is an isometry, the Sz.-
Nagy-Schaffer minimal isometric lifting V of T ′ is equal to T ′. The latter implies
that there is only one contractive interpolant B for the data set {A, T ′, V, R, Q},
namely B = A. The fact that R and Q are the zero operators on C implies that
F = {0} and F ′ = {0}. It follows that for this data set {A, T ′, V, R, Q} the only
contractive interpolant B is given by formula (0.5) where for F we can take any
function in the Schur class S(C, C). The previous example can be seen as a special
case of our second main theorem.
COUPLING AND RELAXED COMMUTANT LIFTING
3
Theorem 0.2. Let B be a contractive interpolant for the data set {A, T ′, V, R, Q}
where V is the Sz.-Nagy-Schaffer minimal isometric lifting of T ′. Then there is a
one to one mapping from the set of all F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω
such that B is given by (0.5) onto the set S(GB , G′
B being given
by
B), with GB and G′
(0.6)
GB = DB ⊖ DBQH0
and G′
B = (D◦ ⊕ DB) ⊖(cid:20) D◦
DBR (cid:21) H0.
Our proof of the above theorem also provides a procedure to obtain a mapping
of the type referred to in the theorem.
It is interesting to specify Theorems 0.1 and 0.2 for the case when in the lifting
data set {A, T ′, V, R, Q} the operators A, T ′, R and Q are zero operators.
In
this case the intertwining condition V BR = BQ, where V is the Sz.-Nagy-Schaffer
minimal isometric lifting of T ′ = 0, is trivially fulfilled, and hence B is a contractive
interpolant if and only if
0
Bh = (cid:20)
where Θ is any function in H 2
ball(L(H, H′)). The latter means that Θ is a L(H, H′)-
valued analytic function on D such that for each h ∈ H the function Θ(·)h belongs
to the Hardy space H 2(H′), and kΘ(·)hkH 2(Y) ≤ khk. It follows that Theorems 0.1
and 0.2 have the following corollaries.
Θ(·)h (cid:21) ,
h ∈ H,
Corollary 0.3. Let F be any function in the Schur class S(H, H′ ⊕ H), and let Π
and Π′ be the orthogonal projections of H′ ⊕ H on H and H′, respectively. Then
the function Θ defined by
(0.7)
Θ(λ) = Π′F (λ)(IH − λΠF (λ))−1
belongs to H 2
way
ball(L(H, H′)), and any function in H 2
ball(L(H, H′)) is obtained in this
Corollary 0.4. Let Θ ∈ H 2
ball(L(H, H′)). Then there is a one to one mapping
from the set of all F in S(H, H′ ⊕ H) such that (0.7) holds onto the set S(DΓ, DΓ),
where Γ is the contraction from H into H 2(H′) defined by
(Γh)(λ) = Θ(λ)h,
h ∈ H, λ ∈ D.
When H = H′ = C, and hence Θ is a scalar function, Corollary 0.3 can be found
in [18], page 490, provided Θ is of unit H 2 norm. For p × q matrix functions Θ,
when H = Cq and H′ = Cp, Corollary 0.3 is Theorem 2.2 in [3]. For the general
operator valued case Corollary 0.3 seems to be new. Corollary 0.4 seems to be new
even in the scalar case. Notice that in the scalar case the space DΓ in Corollary 0.4
consists of the zero element only if Θ is of unit H 2 norm, and DΓ = C otherwise.
Another case of special interest is the classical commutant lifting problem. As
we know from [14] the commutant lifting theorem can be obtained by applying
the relaxed commutant lifting theorem to the data set {A, T ′, U ′, IH, Q} where
H0 = H, the operator R is the identity operator on H and Q is an isometry; see
[14]. In this case, the space G′
B in Theorem 0.2 consists of the zero element for
any choice of the contractive interpolant B.
In other words, for the case of the
classical commutant lifting formula (0.5) provides a proper parameterization, that
is, for every contractive interpolant B for {A, T ′, V, R, Q} there exists a unique F
4
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω such that B is given by (0.5). Finally,
it is noted that this formula also yields the Redheffer type parameterization for the
commutant lifting theorem presented in Section XIV of [13].
If in Theorem 0.1 we take F (λ) ≡ ωΠF , where ΠF is the orthogonal projection
of DA onto F , then the contractive interpolant B in (0.5) is precisely the central
solution presented in [14].
From Theorem 0.1 we see that F = DA implies that there is a unique contrac-
tive interpolant (which is known from Theorem 3.1 in [14]). Other conditions of
uniqueness will be given in the final section of the paper.
We shall prove Theorems 0.1 and 0.2 by combining the method of isometric
coupling with some aspects of isometric realization theory. The theory of isometric
couplings originates from [1], [2], and was used to study the commutant lifting
problem for the first time in [5] -- [9]; see also, Section VII.7 in [13].
The paper consist of six sections not counting this introduction. The first two
sections have a preliminary character, and review the notions of an isometric lifting
(Section 1), and an isometric realization (Section 2). In the third section we develop
the notion of an isometric coupling of a pair of contractions which provides the main
tool in this paper. In Section 4 we prove Theorem 0.1 for the case when R∗R = Q∗Q,
and in Section 5 we prove Theorem 0.1 in its full generality. In the final section
we prove Theorem 0.2, and we present a few sufficient conditions for the case when
(0.5) provides a proper parameterization, and also conditions for uniqueness of the
solution.
We conclude this introduction with a few words about notation and terminology.
Throughout capital calligraphic letters denote Hilbert spaces. The Hilbert space
direct sum of U and Y is denoted by
U ⊕ Y or by
(cid:20) U
Y (cid:21) .
The set of all bounded linear operators from H to H′ is denoted by L(H, H′). The
identity operator on the space H is denoted by IH or just by I, when the underlying
space is clear from the context. By definition, a subspace is a closed linear manifold.
If M is a subspace of H, then H ⊖ M stands for the orthogonal complement of
M in H. Given a subspace M of H, the symbol ΠM will denote the orthogonal
projection of H onto M viewed as an operator from H to M, and PM will denote
the orthogonal projection of H onto M viewed as an operator on H. Note that Π∗
M
is the canonical embedding from M into H, and hence PM = Π∗
MΠM. Instead of
Π∗
M we shall sometimes write EM, where the capital E refers to embedding. A
subspace M of H is said to be cyclic for an operator T on H whenever
H =
∞
_n=0
T nM = span {T nM n = 0, 1, 2, . . .}.
Finally, by definition, a L(H, H′)-valued Schur class function is a function in
S(H, H′), i.e., an operator-valued function which is analytic on the open unit disk
D and whose values are contractions from H to H′.
COUPLING AND RELAXED COMMUTANT LIFTING
5
1. Isometric liftings
In this section we review some facts concerning isometric liftings that are used
throughout this paper. For a more complete account we refer to the book [16] (see
also Chapter VI in [13], and Section 11.3 in [14]).
Let T ′ on H′ be a contraction. Recall that an operator U on K is a isometric
lifting of T ′ if H′ is a subspace of K and U is an isometry satisfying ΠH′ U = T ′ΠH′.
Isometric liftings exist. In fact, the Sz.-Nagy-Schaffer isometric lifting V of T ′ is
given by
(1.1)
V = (cid:20)
T ′
EDT ′ S (cid:21) on (cid:20)
0
H′
H 2(DT ′ ) (cid:21) .
Here S is the unilateral shift on the Hardy space H 2(DT ′ ) and E is the canonical
embedding of DT ′ onto the space of constant functions in H 2(DT ′ ). To see that V
in (1.1) is an isometric lifting of T ′ note that any operator U on K = H′ ⊕ M is an
isometric lifting of T ′ if and only if U admits an operator matrix representation of
the form
(1.2) U = (cid:20)
is an isometry.
T ′
Y1DT ′ Y2 (cid:21) on (cid:20) H′
0
M (cid:21) where Y = (cid:2) Y1 Y2 (cid:3) : (cid:20) DT ′
M (cid:21) → M
An isometric lifting U of T ′ is called minimal when H′ is cyclic for U . The
Sz.-Nagy-Schaffer isometric lifting of T ′ is minimal.
If the isometric lifting U is
given by (1.2), then the lifting is minimal if and only if the space Y1DT ′ is cyclic
for Y2.
Two isometric liftings U1 on K1 and U2 on K2 of T ′ are said to be isomorphic if
there exists a unitary operator Φ from K1 onto K2 such that
ΦU1 = U2Φ and Φh = h for all h ∈ H′.
Minimality of an isometric lifting is preserved under an isomorphism, and two
minimal isometric liftings of T ′ are isomorphic.
Finally, when U on K is a isometric lifting of T ′, then the subspace K′, given by
K′ =
∞
_n=0
U nH′,
is reducing for U , that is, both K′ and its orthogonal complement K = K ⊖ K′ are
invariant under U . Furthermore, in that case the operator U ′ = ΠK′ U K′ on K′
is a minimal isometric lifting of T ′, and the operator U admits a operator matrix
decomposition of the form
(1.3)
U = (cid:20) U ′
0
0
U (cid:21) on (cid:20) K′
K (cid:21) ,
where U is an isometry on K. We shall call U ′ in (1.3) the minimal isometric lifting
of T ′ associated with U .
The following proposition summarizes the results referred to above in a form
that will be convenient for this paper. For details we refer to Section 11.3 in [14].
Theorem 1.1. Let T ′ be a contraction on H′, let V on H′ ⊕ H 2(DT ′ ) be the Sz.-
Nagy-Schaffer (minimal) isometric lifting of T ′, and let U on H⊕M be an arbitrary
isometric lifting of T ′ given by (1.2). Then there exists a unique isometry Φ from
6
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
H′ ⊕ H 2(DT ′ ) into H′ ⊕ M such that U Φ = ΦV and ΦH′ = IH′ . In fact, Φ is
given by
Φ = (cid:20) IH′
0
0
Λ (cid:21) : (cid:20)
where Λ is defined by
H′
H 2(DT ′ ) (cid:21) → (cid:20) H′
M (cid:21) ,
Λh =
∞
Xn=0
Y n
2 Y1hn,
h(λ) =
∞
Xn=0
λnhn ∈ H 2(DT ′ ),
2 )−1m for each
with Y1 and Y2 as in (1.2). Moreover, (Λ∗m)(λ) = Y ∗
m ∈ M. Finally, Φ is unitary if and only if U is a minimal isometric lifting of T ′,
and in that case the isometric liftings V and U of T ′ are isomorphic.
1 (I − λY ∗
The isometry Φ introduced in the above theorem will be referred to as the unique
isometry associated with T ′ that intertwines V with U . Since V is uniquely deter-
mined by T ′, we shall denote this isometry simply by ΦU, T ′ . When U on K is
an isometric lifting of T ′ and U ′ on K′ is the minimal isometric lifting of T ′ as-
sociated with U , then the operator ΠK′ΦU, T ′ is the unique isometry associated
with T ′ that intertwines V with U ′, that is, ΦU ′, T ′ = ΠK′ΦU, T ′ or, equivalently,
Π∗
K′ΦU ′, T ′ = ΦU, T ′ .
2. Isometric realizations
In this section we review some of the classical results on controllable isometric
realizations, and we prove a few additional results that will be useful in the later
sections.
We say that {Z, B, C, D; X , U , Y} (or simply {Z, B, C, D}) is a realization of a
L(U , Y)-valued function G if
(2.1)
G(λ) = D + λC(IX − λZ)−1B
for all λ in some open neighborhood of the origin in the complex plane. Here Z
is an operator on X and B is an operator from U into X while C is an operator
mapping X into Y and D is an operator from U into Y (where X , U and Y are
all Hilbert spaces). In this case, we refer to the function defined by the right hand
side of (2.1) as the associated transfer function. A realization {Z, B, C, D} is called
isometric if the operator
(2.2)
M = (cid:20) D C
B Z (cid:21) : (cid:20) U
X (cid:21) → (cid:20) Y
X (cid:21)
is an isometry. The 2 × 2 operator matrix in (2.2) is called the system matrix
associated with the realization {Z, B, C, D}. The transfer function of an isometric
realization belongs to the Schur class S(U , Y), that is, if {Z, B, C, D} is an isometric
realization, then the function G defined by (2.1) is a contractive analytic L(U , Y)-
valued function on D. Conversely, if G ∈ S(U , Y), then there is an isometric
realization {Z, B, C, D} such that (2.1) holds for all λ in D.
The transfer function of a realization can also be expressed in terms of the system
matrix M . In fact, if {Z, B, C, D} is a realization and M is the associated system
matrix, then in a neighborhood of the origin the transfer function G is also given
by
(2.3)
G(λ) = ΠY M (IU ⊕X − λJX M )−1Π∗
U ,
COUPLING AND RELAXED COMMUTANT LIFTING
7
where JX is the partial isometry from Y ⊕ X to U ⊕ X given by
JX = (cid:20) 0
0
0
IX (cid:21) : (cid:20) Y
X (cid:21) → (cid:20) U
X (cid:21) .
Indeed, for λ sufficiently close to zero we have
G(λ) = D + λC(IX − λZ)−1B = (cid:2) D C (cid:3)(cid:20)
IU
λ(IX − λZ)−1B (cid:21)
0
IU
= (cid:2) D C (cid:3)(cid:20)
= (cid:2) D C (cid:3)(cid:20) IU
= (cid:2) D C (cid:3)(cid:18)IU ⊕X − λ(cid:20) 0
λ(IX − λZ)−1B (IX − λZ)−1 (cid:21)(cid:20) IU
0 (cid:21)
−λB IX − λZ (cid:21)−1(cid:20) IU
0 (cid:21)
B Z (cid:21)(cid:19)−1(cid:20) IU
0 IX (cid:21)(cid:20) D C
0 (cid:21)
0
0
= ΠY M (IU ⊕X − λJX M )−1Π∗
U .
Since the right side of (2.1) is a Schur class function if M in (2.2) is an isometry,
the same holds true for the right hand side of (2.3). Notice that the function G
defined by (2.3) can also be written in the form:
If for a realization {Z, B, C, D; X , U , Y} the space W∞
G(λ) = ΠY (IY⊕X − λM JX )−1M Π∗
U .
n=0 Z nBU is equal to X ,
then the realization or the pair {Z, B} is called controllable.
In other words, a
realization is controllable if and only if the space BU is cyclic for Z. In terms of the
system matrix M in (2.2) the realization {Z, B, C, D} is controllable if and only if
(2.4)
X = ΠX
∞
_n=0
(JX M )n(cid:20) U
{0} (cid:21) .
The above condition (2.4) is also equivalent to the requirement that {JX M, Π∗
U } is
a controllable pair. In the particular case when U = Y in (2.2), condition (2.4) can
be written in an even simpler form. This is the contents of the next lemma.
Lemma 2.1. Let M be as in (2.2), and assume U = Y. Then {Z, B} is controllable
if and only if U ⊕ {0} is cyclic for M , that is,
(2.5)
∞
_n=0
M n(cid:20) U
{0} (cid:21) = (cid:20) U
X (cid:21) .
Proof. Let EU be the canonical embedding of U into U ⊕ X , and define M0 to be
the operator
M0 = (cid:20) 0
B Z (cid:21) : (cid:20) U
X (cid:21) → (cid:20) U
X (cid:21) .
0
Then M0 = M − EU (cid:2) C D (cid:3). This feedback relation implies that the pair
{M0, EU } is controllable if and only if the pair {M, EU } is controllable. Thus
(2.5) holds if and only if {M0, EU } is controllable. Now notice that for all integers
n ≥ 1, we have
M n
0 EU = (cid:20)
0
Z n−1B Z n (cid:21)(cid:20) IU
0 (cid:21) = (cid:20)
0
0
Z n−1B (cid:21) : U → (cid:20) U
X (cid:21) .
8
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
It follows that
∞
_n=0
M n
0 EU U = (cid:20) U
{0} (cid:21) ⊕
∞
_n=1
M n
0 EU = U ⊕
∞
_n=1
Z n−1BU .
We conclude that (2.5) holds if and only if the pair {Z, B} is controllable.
(cid:3)
A realization {Z, B, C, D} or the pair {C, Z} is called observable if CZ nx =
0 for all integers n ≥ 0 implies that the vector x is equal to zero. Since the
orthogonal complement of Ker CZ n is equal to the closure of Im (Z ∗)nC∗, we see
that observability of the realization {Z, B, C, D} is equivalent to the controllability
of the dual realization {Z ∗, C∗, B∗, D∗}.
Two realizations {Z1 on X1, B1, C1, D1} and {Z2 on X2, B2, C2, D2} are said to
be unitarily equivalent if D1 = D2 and there exists a unitary operator W mapping
X1 onto X2 such that
W Z1 = Z2W, W B1 = B2
and C2W = C1.
Unitary equivalence does not change the transfer function. More precisely, when
two realizations are unitary equivalent, then their transfer functions coincide in a
neighborhood of zero. For isometric controllable realizations the converse is also
true. In fact we have the following theorem.
Theorem 2.2. Let G be a L(U , Y)-valued function. Then G ∈ S(U , Y) if and only
if G admits an isometric realization. In this case, G admits a controllable isomet-
ric realization and all controllable isometric realizations of G are unitarily equiv-
alent. In particular, formula (2.1) provides a one to one correspondence between
the L(U , Y)-contractive analytic functions on D and (up to unitary equivalence) the
controllable isometric realizations of L(U , Y)-valued functions.
The above result appears in a somewhat different form in [19] as a theorem
representing a Schur class function as a characteristic operator function. A full
proof, with isometric systems replaced by their dual ones, can be found in [4] which
also gives additional references. In Section 1.3 of [15] the theorem is proved using
the Naimark dilation theory.
We conclude this section with a proposition that will be useful in the later
sections. The starting point is an isometry Y of the type appearing in (1.2). More
precisely,
(2.6)
Y = (cid:2) Y1 Y2 (cid:3) : (cid:20) D′
M (cid:21) → M.
Proposition 2.3. Let Y in (2.6) be an isometry. Assume M = D ⊕ X , and let
ΠD and ΠX be the orthogonal projections of M onto D and X , respectively. Put
(2.7)
where ΠD ′⊕D is the orthogonal projection of D′ ⊕ M onto D′ ⊕ D, and
F (λ) = ΠD ′⊕DY ∗(IM − λJ ′
X Y ∗)−1Π∗
D,
λ ∈ D,
X : D′ ⊕ M → M,
J ′
X (d′ ⊕ m) = ΠX m.
J ′
Then F belongs to the Schur class S(D, D′ ⊕ D) and
(2.8)
1 (IM − λY ∗
Y ∗
2 )−1Π∗
where Π and Π′ are the orthogonal projections of D′⊕D onto D and D′, respectively.
D = Π′F (λ)(cid:16)ID − λΠF (λ)(cid:17)−1
,
λ ∈ D,
COUPLING AND RELAXED COMMUTANT LIFTING
9
It will be convenient first to prove a lemma. Let Γ be a contraction from M into
E1 ⊕ M. Partition Γ as a 2 × 1 operator matrix, as follows
(2.9)
Γ = (cid:20) Γ1
Γ2 (cid:21) : M → (cid:20) E1
M (cid:21) .
Furthermore, let E2 be a subspace of M, and consider the function
(2.10)
Ξ(λ) = Γ1(IM − λΓ2)−1Π∗
E2 ,
λ ∈ D,
Here ΠE2 is the orthogonal projection of M onto E2. Since Γ is a contraction, the
same holds true for Γ2, and hence I − λΓ2 is invertible for each λ ∈ D. Thus Ξ is
well-defined on D. Next, let X be the orthogonal complement of E2 in M, and thus
M = E2 ⊕ X . Then Γ also admits a 3 × 2 operator matrix representation, namely
(2.11)
Put
(2.12)
Γ =
D1 C1
D2 C2
B Z
: (cid:20) E2
X (cid:21) →
E1
E2
X
.
F (λ) = (cid:20) D1
D2 (cid:21) + λ(cid:20) C1
C2 (cid:21) (IX − λZ)−1B,
λ ∈ D.
Again, since Γ is a contraction, the operator Z is a contraction, and hence F is
well-defined on D.
Lemma 2.4. Let Ξ and F be the functions defined by (2.10) and (2.12), respectively.
Then F belongs to the Schur class S(E2, E1 ⊕ E2) and
(2.13)
Ξ(λ) = Π1F (λ)(I − Π2F (λ))−1 λ ∈ D,
where Π1 and Π2 are the orthogonal projections of E1 ⊕ E2 onto E1 and E2, respec-
tively.
Proof. The function F is the transfer function of the system
(cid:8)Z, B,(cid:20) C1
C2 (cid:21) ,(cid:20) D1
D2 (cid:21) ; X , E2, E1 ⊕ E2(cid:9).
By (2.11) the system matrix corresponding to this system is equal to Γ, and hence
it is a contraction. This implies that F belongs to the Schur class S(E2, E1 ⊕ E2);
cf., Theorem 4.1 in [10] where this is proved for time-variant systems.
To prove (2.13) fix λ ∈ D. Using the partitioning of Γ in (2.11) we see that for
each e ∈ E2 we have
Ξ(λ)e = (cid:2) D1 C1 (cid:3)(cid:18)I − λ(cid:20) D2 C2
B Z (cid:21)(cid:19)−1(cid:20) e
0 (cid:21) .
To find Ξ(λ)e we have to compute the first column of the inverse of the 2 × 2
operator matrix
(2.14)
(cid:20) I − λD2 −λC2
I − λZ (cid:21) .
−λB
Since I − λZ is invertible, the Schur complement ∆(λ) of I − λZ in (2.14) is well-
defined and is given by
∆(λ) := I − λD2 − λ2C2(I − λZ)−1B = I − λΠ2F (λ).
10
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
It follows (cf., Remark 1.2 in [11]) that
(cid:20) I − λD2 −λC2
I − λZ (cid:21)−1
−λB
= (cid:20)
(I − λΠ2F (λ))−1
λ(I − λZ)−1B(I − λΠ2F (λ))−1
∗
∗ (cid:21) .
Thus
Ξ(λ)e = (cid:2) D1 C1 (cid:3)(cid:20)
(I − λΠ2F (λ))−1e
λ(I − λZ)−1B(I − λΠ2F (λ))−1e (cid:21)
= (D1 − λC1(I − λZ)−1B)(I − λΠ2F (λ))−1e
= Π1F (λ)(I − λΠ2F (λ))−1e.
Since e is an arbitrary element of E2, this proves (2.13).
Proof of Proposition 2.3. Since Y is assumed to be an isometry, Y ∗ is a con-
traction. Now apply Lemma 2.4 with D′ in place of E1, with Y ∗ in place of the
contraction Γ in (2.9), and with D in place of E2. With these choices the function
Ξ in (2.10) coincides with the function defined by the left hand side of (2.8). Thus
in order to finish the proof it remains to show that with Γ = Y ∗, E1 = D′, and
E2 = D the function F in (2.12) is also given by (2.7). But this follows by applying
to F in place of G that the function G in (2.1) is also given by (2.3). Indeed, since
F is the transfer function of the system
(cid:3)
(cid:8)Z, B,(cid:20) C1
C2 (cid:21) ,(cid:20) D1
D2 (cid:21) ; X , D, D′ ⊕ D(cid:9),
and the system matrix of this system is equal to Y ∗, the equivalence between (2.1)
and (2.3) yields in a straightforward way that F in (2.12) is also given by (2.7). (cid:3)
3. Isometric couplings
Throughout this section {T ′, A} is a pair of contractions, T ′ on a Hilbert space
H′ and A from a Hilbert space H to H′.
An isometric coupling of {T ′, A} is a pair {U on K, τ} of operators such that U
is an isometric lifting of T ′, acting on K (and thus H′ ⊂ K), and τ is an isometry
from H to K with ΠH′ τ = A. If the space K is of no interest, then we will just
write {U, τ}. An isometric coupling {U on K, τ} of {T ′, A} is called minimal if, in
addition, the space H′ ∨ τ H is cyclic for U , that is,
K =
∞
_n=0
U n(H′ ∨ τ H).
There exist minimal isometric couplings of {T ′, A}. To see this, let U be the
operator on H′ ⊕ H 2(DT ′ ) ⊕ H 2(DA) given by the following operator matrix rep-
resentation
T ′
0
EDT ′ SDT ′
0
0
SDA
0
0
.
U =
Here EDT ′ is the canonical embedding of DT ′ onto the space of constant functions
of H 2(DT ′ ), and SDT ′ and SDA are the unilateral shifts on H 2(DT ′ ) and H 2(DA),
respectively. Notice that the operator defined by the 2 × 2 operator matrix in the
left upper corner of the matrix for U is the Sz.-Nagy-Schaffer minimal isometric
COUPLING AND RELAXED COMMUTANT LIFTING
11
lifting of T ′. Since SDA is an isometry, we conclude that U is also an isometric
lifting of T ′. Now let τ be the isometry defined by
τ =
A
0
EDA DA
: H →
H′
H 2(DT ′ )
H 2(DA)
where EDA is the canonical embedding of DA onto the space of constant functions
of H 2(DA). Then {U, τ} is a minimal isometric coupling of {T ′, A}.
Two isometric couplings {U1 on K1, τ1} and {U2 on K2, τ2} of {T ′, A} are said
to be isomorphic if there exists a unitary operator Ψ from K1 to K2 such that
ΨU1 = U2Ψ, Ψτ1 = τ2
and Ψh = h for all h ∈ H′.
In this case
(3.1)
ΨU1τ1 = U2τ2.
Minimality is preserved under isomorphic equivalence. Indeed, when the pairs
{U1 on K1, τ1} and {U2 on K2, τ2} are isomorphic isometric couplings of {T ′, A},
and Ψ from K1 to K2 is an isomorphism between the two isometric couplings, then
∞
_n=0
U n
2 (H′ ∨ τ2H) =
∞
∞
(ΨU1Ψ∗)n(H′ ∨ Ψτ1H) =
_n=0
1 ((Ψ∗H′) ∨ Ψ∗Ψτ1H) = Ψ
U n
_n=0
_n=0
∞
= Ψ
ΨU n
1 Ψ∗(H′ ∨ Ψτ1H)
∞
_n=0
U n
1 (H′ ∨ τ1H).
We say that an isometric coupling {U on K, τ} of {T ′, A} is special
if K is a
Hilbert direct sum of the space H′, the space DA and some Hilbert space X , that
is, K = H′ ⊕ DA ⊕ X , and the action of τ is given by τ h = Ah ⊕ DAh ⊕ 0, where 0 is
the zero vector in X . In other words, an isometric coupling {U on K, τ} of {T ′, A}
is special if, in addition, DA is a subspace of M, where M = K ⊖ H′, and τ admits
a matrix representation of the form
τ = (cid:20)
A
DA DA (cid:21) : H → (cid:20) H′
M (cid:21) .
Π∗
The importance of special isometric couplings follows from Theorem 3.4 below.
To prove this theorem we need a few auxiliary propositions. The first also settles
the question of existence of special isometric couplings.
Proposition 3.1. Every isometric coupling is isomorphic to a special isometric
coupling.
Proof. Let {U on K, τ} be an isometric coupling of {T ′, A}, and put M = K ⊖ H′.
Since τ is an isometry and ΠH′ τ = A, the operator τ admits a matrix representation
of the form:
(3.2)
τ = (cid:20) A
ΓDA (cid:21) : H → (cid:20) H′
M (cid:21) where Γ : DA → M is an isometry;
see Section IV.1 of [13] or Section XXVII.5 of [17]. Now let D = Im Γ, and put
X = M ⊖ D. Then D is closed, and we can view Γ as a unitary operator from DA
12
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
onto D. Define the unitary operator σ by
σ =
IH′
0
0
0
Γ
0
0
0
IX
:
H′
DA
X
→
H′
D
X
.
Also define U0 = σ∗U σ and τ0 = σ∗τ . Then {U0, τ0} is a special isometric coupling
of {T ′, A} which is isomorphic to {U, τ}.
(cid:3)
Since minimality of isometric couplings is preserved under isomorphisms, and
isometric couplings do exist (see the third paragraph of this section), the above
proposition shows that any {T ′, A} admits a special minimal isometric coupling.
Recall that an isometric lifting U of T ′ can always be represented (see (1.2)) in
the following form:
(3.3) U = (cid:20)
T ′
Y1DT ′ Y2 (cid:21) on (cid:20) H′
0
M (cid:21) where Y = (cid:2) Y1 Y2 (cid:3) : (cid:20) DT ′
M (cid:21) → M
is an isometry. According to (1.3) this U also admits a matrix representation of
the form:
U = (cid:20) U ′
0
0
U (cid:21) on (cid:20) K′
K (cid:21) where K′ =
∞
_n=0
U nH′,
K = K ⊖ K′.
Here U ′ on K′ is the minimal isometric lifting of T ′ associated with U (see Section 1),
and U is an isometry on K. We can now state the next proposition.
Proposition 3.2. Let {U, τ } be an isometric coupling of {T ′, A}, where U is de-
termined by (3.3) and τ by (3.2). Set D = Im Γ, where Γ is given by (3.2), and for
Y in (3.3) consider the following operator matrix representation:
(3.4)
Y = (cid:20) D C
B Z (cid:21) : (cid:20) DT ′ ⊕ D
X
(cid:21) → (cid:20) D
X (cid:21) where X = M ⊖ D.
Then {U, τ } is a minimal isometric coupling of {T ′, A} if and only if the pair {Z, B}
is controllable.
Proof. Since τ is given by (3.2), the space H′ ⊕ τ H is equal to H′ ⊕ D. Thus we
have to show that H′ ⊕D is cyclic for U if and only if the pair {Z, B} is controllable.
To do this we associate with U two auxiliary operators, namely
U = (cid:20)
0
Y1DT ′ Y2 (cid:21) on (cid:20) H′
M (cid:21) , M = (cid:20) 0
Y1 Y2 (cid:21) on (cid:20) DT ′
M (cid:21) .
0
0
Notice that the range of U − U belongs to H′. Since H′ ⊂ H′ ⊕ D, this implies that
H′ ⊕ D is cyclic for U if and only if H′ ⊕ D is cyclic for U . By induction one proves
that for n = 1, 2, 3, . . . we have
U n = (cid:20)
0
0
Y n−1
Y1DT ′ Y n
2
2 (cid:21) on (cid:20) H′
M (cid:21) , M = (cid:20)
0
Y n−1
2
0
Y1 Y n
2 (cid:21) on (cid:20) DT ′
M (cid:21) .
On the other hand
(cid:2) Y n−1
2
Y1DT ′ Y n
2
(cid:3)(cid:20) H′
D (cid:21) = (cid:2) Y n−1
2
Y1 Y n
2
(cid:3)(cid:20) DT ′
D (cid:21), n = 1, 2, 3, . . . ,
and hence H′ ⊕ D is cyclic for U if and only if DT ′ ⊕ D is cyclic for M .
COUPLING AND RELAXED COMMUTANT LIFTING
13
It remains to prove that DT ′ ⊕ D is cyclic for M if and only if the pair {Z, B}
is controllable. Using Y = (cid:2) Y1 Y2 (cid:3) and (3.4), we see that
M = (cid:20) D C
B Z (cid:21) ,
DD and C = Π∗
where D = Π∗
DC with ΠD equal to the orthogonal projection of
DT ′ ⊕ D onto D. By employing Lemma 2.1 with U = Y = DT ′ ⊕ D, we see that
DT ′ ⊕D is cyclic for M if and only if the pair {Z, B} is controllable, which completes
the proof.
(cid:3)
Proposition 3.3. Let {U1 on K1, τ1} and {U2 on K2, τ2} be special isometric cou-
plings of {T ′, A}. For j = 1, 2 set Xj = Kj ⊖ (H′ ⊕ DA), and let Y (j) be the
isometry from DT ′ ⊕ DA ⊕ Xj into DA ⊕ Xj corresponding to Uj via (3.3). Consider
the following operator matrix representation
Y (j) = (cid:20) Dj Cj
Bj Zj (cid:21) : (cid:20) DT ′ ⊕ DA
Xj
Xj (cid:21)
(cid:21) → (cid:20) DA
for j = 1, 2.
Then {U1, τ1} and {U2, τ2} are isomorphic if and only if {Z1, B1, C1, D1} and
{Z2, B2, C2, D2} are unitarily equivalent realizations.
Proof. Assume that {Z1, B1, C1, D1} and {Z2, B2, C2, D2} are unitarily equivalent,
that is, D1 = D2 and there exists a unitary operator W from X1 onto X2 such that
(3.5)
W Z1 = Z2W, W B1 = B2
and C1 = C2W.
Now let Φ be the unitary operator from K1 onto K2 defined by
Then Φh = h for all h in H′. Because {U1, τ1} and {U2, τ2} are special, we see that
Φ =
IH′
0
0
τj =
A
DA
0
0
IDA
0
0
0 W
:
: H →
H′
DA
Xj
H′
DA
X1
→
H′
DA
X2
.
for j = 1, 2.
(3.6)
(3.7)
(3.8)
Hence Φτ1 = τ2. Using the appropriate operator matrix decomposition we arrive
at
IH′
0
0
ΦU1 =
=
D1Π∗
W B1Π∗
T ′
0
DT ′ DT ′ D1Π∗
DT ′ DT ′ B1Π∗
0
DA C1
DA Z1
0
IDA
D1Π∗
B1Π∗
0
0
0 W
T ′
0
0
D1Π∗
DT ′ DT ′
C1
DA
DT ′ DT ′ W B1Π∗
DA W Z1
.
A similar calculation shows that
U2Φ =
=
(3.9)
T ′
0
DT ′ DT ′ D2Π∗
DT ′ DT ′ B2Π∗
T ′
0
DT ′ DT ′ D2Π∗
DT ′ DT ′ B2Π∗
0
DA C2
DA Z2
0
DA C2W
DA Z2W
D2Π∗
B2Π∗
D2Π∗
B2Π∗
0
IDA
0
0
0 W
IH′
0
0
.
14
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
Because D1 = D2 and (3.5) holds, we see that ΦU1 = U2Φ. In other words, {U1, τ1}
and {U2, τ2} are isomorphic.
Conversely assume that {U1, τ1} and {U2, τ2} are isomorphic. Then there exists a
unitary operator Φ from K1 onto K2 such that Φh = h for all h in H′ and Φτ1 = τ2
and ΦU1 = U2Φ. Because τ1 and τ2 admit matrix representations of the form
presented in (3.7) and Im DA is dense in DA, we see that Φh = h for all h ∈ DA. So
Φ admits a matrix representation as in (3.6) where W is a unitary operator from
X1 onto X2. By combining ΦU1 = U2Φ with the matrix representations for ΦU1 in
(3.8) and U2Φ in (3.9), we see that
D1 = D2, W Z1 = Z2W, W B1 = B2
and C1 = C2W.
Hence {Z1, B1, C1, D1} and {Z2, B2, C2, D2} are unitarily equivalent realizations.
(cid:3)
Theorem 3.4. Let {T ′, A} be a pair of contractions, T ′ acting on H′ and A from
H into H′. Then there is a one to one map from the set of minimal isometric
couplings of {T ′, A}, with isomorphic ones being identified, onto the Schur class
S(DA, DT ′ ⊕DA). This map is defined as follows. Let {U, τ } be a minimal isometric
coupling of {T ′, A}, which may be assumed to be special, by Proposition 3.1. Define
(3.10)
F{U,τ }(λ) = ΠDT ′ ⊕DA Y ∗(IM − λJ ′
X Y ∗)−1Π∗
DA ,
where Y is the isometry uniquely determined by U via (3.3), X = M ⊖ DA, and
J ′
X is the partial isometry from DT ′ ⊕ DA ⊕ X to DA ⊕ X given by
J ′
X = (cid:20) 0
0
0
IX (cid:21) : (cid:20) DT ′ ⊕ DA
X
(cid:21) → (cid:20) DA
X (cid:21) .
Then {U, τ } 7→ F{U,τ } is the desired map.
Proof. We know from Proposition 3.1 that every isometric coupling is isomorphic
to a special one. So without loss of generality we can assume the isometric couplings
to be special.
From Proposition 3.2 and Section 1 it is clear that there is a one to one cor-
respondence between the special minimal isometric couplings of {T ′, A} and the
isometries Y mapping the space DT ′ ⊕ DA ⊕ X into DA ⊕ X , where X is some
Hilbert space and the pair {ΠX Y Π∗
DT ′ ⊕DA} is controllable. In fact, this
one to one correspondence is provided by (3.3). Furthermore, formula (3.4) estab-
lishes a one to one correspondence between the isometries Y mapping the space
DT ′ ⊕ DA ⊕ X into DA ⊕ X and the isometric realizations
X , ΠX Y Π∗
{Z, B, C, D; X , DT ′ ⊕ DA, DA},
X and B = ΠX Y Π∗
and in this one to one correspondence Z = ΠX Y Π∗
DT ′ ⊕DA.
From Theorem 2.2 we know that there is a one to one correspondence between the
controllable isometric realizations, with the unitarily equivalent ones being iden-
tified, and the S(DT ′ ⊕ DA, DA) Schur class functions. Next, note that the map
G 7→ F , where F (λ) = G(λ)∗, is a one to one map from S(DT ′ ⊕ DA, DA) onto
S(DA, DT ′ ⊕ DA). Following up all these one to one correspondences and using the
results of Section 2 we see that the map from a special minimal isometric coupling
{U, τ } to F is given by F = F{U,τ }. To complete the proof, it remains to apply
Proposition 3.3.
(cid:3)
We conclude this section with a lemma that will be useful in the next section.
COUPLING AND RELAXED COMMUTANT LIFTING
15
Lemma 3.5. Let {U1 on K1, τ1} and {U2 on K2, τ2} be isomorphic isometric cou-
plings of {T ′, A}, and let V on H′ ⊕ H 2(DT ′ ) be the Sz.-Nagy-Schaffer minimal
isometric lifting of T ′. For j = 1, 2 let Φj be the unique isometry associated with
T ′ intertwining V and Uj. Then
Φ∗
1τ1 = Φ∗
2τ2.
Proof. Let Ψ from K1 to K2 be an isomorphism from {U1, τ1} to {U2, τ2}. Define
Θ from H′ ⊕ H 2(DT ′ ) into K1 by setting Θ = Ψ∗Φ2. Then Θ is an isometry,
Θh = Ψ∗Φ2h = Ψ∗h = h for all h ∈ H′, and
ΘV = Ψ∗Φ2V = Ψ∗U2Φ2 = U1Ψ∗Φ2 = U1Θ.
So, by Theorem 1.1 (see also the last paragraph of Section 1), the operator Θ is
the unique isometry associated with T ′ intertwining V and U1, that is, Θ = Φ1. It
follows that Φ1 = Ψ∗Φ2, and hence Φ∗
2τ2, which completes the
proof.
(cid:3)
2Ψτ1 = Φ∗
1τ1 = Φ∗
4. Main theorem for the case when R∗R = Q∗Q
In this section {A, T ′, V, R, Q} is a lifting data set, with V on H′ ⊕H 2(DT ′ ) being
the Sz.-Nagy-Schaffer minimal isometric lifting of T ′. In particular, T ′AR = AQ
and R∗R ≤ Q∗Q. Recall that B from H into H′ ⊕ H 2(DT ′ ) is a contractive
interpolant for {A, T ′, V, R, Q} if B is a contraction satisfying ΠH′ B = A and
V BR = BQ.
Our aim is to prove Theorem 0.1 assuming that R∗R = Q∗Q. First let us
reformulate Theorem 0.1 for this case. For this purpose note that for R∗R = Q∗Q
the spaces F and F ′ defined by (0.2) are given by
F = DAQH0
and F ′ = (cid:20) DT ′ AR
DAR (cid:21) H0.
Observe that F ⊂ DA and F ′ ⊂ DT ′ ⊕ DA. Furthermore, the unitary operator ω
mapping F onto F ′ in (0.3) is now determined by
(4.1)
ω(DAQh) = (cid:20) DT ′ AR
DAR (cid:21) h,
h ∈ H0.
The following is the main result of this section.
Theorem 4.1. Let {A, T ′, V, R, Q} be a lifting data set, where V on H′⊕H 2(DT ′ ) is
the Sz.-Nagy-Schaffer minimal isometric lifting of T ′, and assume that R∗R = Q∗Q.
Then all contractive interpolants B for {A, T ′, V, R, Q} are given by
(4.2)
Bh = (cid:20)
ΠT ′ F (λ)(I − λΠAF (λ))−1DAh (cid:21) ,
Ah
h ∈ H,
where F is any function in S(DA, DT ′ ⊕ DA) satisfying F (0)F = ω. Here ω is the
unitary operator defined in (4.1) while ΠT ′ and ΠA are the projections given by
ΠT ′ = (cid:2) I
0 (cid:3) : (cid:20) DT ′
DA (cid:21) → DT ′
and ΠA = (cid:2) 0 I (cid:3) : (cid:20) DT ′
DA (cid:21) → DA.
The proof of the above theorem will be based on a further refinement (which we
present in two propositions) of the theory of isometric couplings presented in the
previous section. In fact, to obtain contractive interpolants for {A, T ′, V, R, Q} we
16
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
shall need isometric couplings {U, τ } of {T ′, A} satisfying the additional intertwin-
ing relation U τ R = τ Q. This is the contents of the first proposition (Proposition
4.2 below). The existence of such couplings is guaranteed by the second proposi-
tion (Proposition 4.3 below), which is based on Theorem 3.4. In the sequel, for
simplicity, we shall write V for the space H′ ⊕ H 2(DT ′ ).
Proposition 4.2. Let {A, T ′, V, R, Q} be a lifting data set, with V on V being the
Sz.-Nagy-Schaffer minimal isometric lifting of T ′, and assume that R∗R = Q∗Q.
Let {U on K, τ } be an isometric coupling of {T ′, A} satisfying U τ R = τ Q, and let
Φ be the unique isometry from V into K associated with T ′ intertwining V with U .
Then
(4.3)
B = Φ∗τ
is a contractive interpolant for {A, T ′, V, R, Q}, and all contractive interpolants for
this data set are obtained in this way. More precisely, if B a contractive interpolant
for {A, T ′, V, R, Q}, then there exists a minimal special isometric coupling {U, τ }
of {T ′, A} such that B = Φ∗τ and U τ R = τ Q.
Proof. First let us show that B defined by (4.3) is a contractive interpolant for the
data set {A, T ′, V, R, Q}. Obviously, B is a contraction. Put K′ = Im Φ. Recall
(see Section 1) that ΦΦ∗ is the orthogonal projection of K onto K′. From Theorem
1.1 we know that K′ = Wn≥0 U nH′ is a reducing subspace for U . It follows that U
commutes with ΦΦ∗. Since Φ∗Φ is the identity operator on V = H′ ⊕ H 2(DT ′ ), we
obtain
V BR = Φ∗U Φ(Φ∗τ )R = Φ∗(U ΦΦ∗)τ R = Φ∗ΦΦ∗U τ R = Φ∗τ Q = BQ.
Thus B is a contractive interpolant for {A, T ′, V, R, Q}.
To prove the reverse implication, assume that B is a contractive interpolant. We
have to construct a minimal special isometric coupling {U, τ } of {T ′, A} satisfying
U τ R = τ Q such that B is given by (4.3). Since B is a contraction, we may consider
the subspaces
F = DBRH0
and
F ′ = DBQH0.
Using V BR = BQ with R∗R = Q∗Q, and the fact that V is an isometry, we see
that for each h ∈ H0 we have
kDBQhk2 = kQhk2 − kBQhk2 = kRhk2 − kV BRhk2
= kRhk2 − kBRhk2 = kDBRhk2.
Hence there exists a unique unitary operator ω from F onto F ′ such that ωDBR =
DBQ. Next, define the subspaces
G = DB ⊖ F and
G′ = DB ⊖ F ′.
Notice that DB = F ⊕ G and DB = F ′ ⊕ G′. Thus ω defines a partial isometry Ω
on DB as follows:
Ω = (cid:20) ω 0
0
0 (cid:21) : (cid:20) F
G (cid:21) → (cid:20) F ′
G′ (cid:21) .
Observe that DΩ coincides with G. Define VΩ to be the Sz.-Nagy-Schaffer minimal
isometric lifting of Ω on VΩ = DB ⊕ H 2( G). Thus VΩ has the following operator
COUPLING AND RELAXED COMMUTANT LIFTING
17
matrix representation
VΩ =
0
0
0
ω
0
0
0 E G S G
:
F
G
H 2( G)
→
F ′
G′
H 2( G)
.
Here E G is the canonical embedding of G onto the space of constant functions in
H 2( G), and S G is the unilateral shift on the Hardy space H 2( G). Since VΩ is a
minimal isometric lifting of Ω, we have VΩ = DB ⊕ H 2( G) = W∞
Ω DB. Now,
n=0 V n
put
(4.4)
(4.5)
UΩ = (cid:20) V
τΩ = (cid:20)
VΩ (cid:21) ,
0 VΩ (cid:21) on (cid:20) V
DB DB (cid:21) : H → (cid:20) V
Π∗
VΩ (cid:21) .
0
B
Since V on V = H′ ⊕ H 2(DT ′ ) is an isometric lifting of T ′, the operator UΩ is an
isometric lifting of T ′, and A = ΠH′ τΩ. It follows that {UΩ, τΩ} is an isometric
coupling of {T ′, A}. Notice that V ∨ τΩH = V ⊕ DB. Because H′ is cyclic for V and
DB is cyclic for VΩ, the reducing decomposition of UΩ in (4.4) shows that H′ ∨ τΩH
is cyclic for UΩ. In other words, the isometric coupling {UΩ, τΩ} is minimal. Since
V BR = BQ, the construction of UΩ and τΩ implies that
(4.6)
UΩτΩR = τΩQ and B = ΠV τΩ.
Indeed, for h ∈ H0 we have
UΩτΩRh = (cid:20) V
0 VΩ (cid:21)(cid:20) BRh
DBRh (cid:21) = (cid:20) V BRh
VΩDBRh (cid:21) .
0
However, DBRH0 ⊂ F , and hence VΩDBRh = ωDBRh = DBQh, which follows
from the definition of ω. Since, by assumption, V BRh = BQh, we see that
UΩτΩRh = (cid:20) BQh
DBQh (cid:21) = τΩQh,
which proves the first identity in (4.6). The second is clear from the definition of
τΩ.
From the construction of UΩ it follows that the unique isometry ΦΩ associated
with T ′ that intertwines V with UΩ is equal to Π∗
V , where ΠV is the orthogonal
projection of V ⊕ VΩ onto V. This together with the second identity in (4.6) yields
(4.7)
B = ΠV τΩ = Φ∗
ΩτΩ.
By Proposition 3.1 and the fact that minimality of isometric couplings is pre-
served under isomorphisms, there exists a minimal special isometric coupling {U, τ }
of {T ′, A} which is isomorphic to {UΩ, τΩ}. Using Lemma 3.5 and formula (4.7) we
obtain B = Φ∗τ , where Φ is the unique isometry associated with T ′ that intertwines
V with U .
It remains to prove that U τ R = τ Q. Let Ψ be the isomorphism that trans-
forms {UΩ, τΩ} into {U, τ }. In particular, ΨτΩ = τ . Moreover formula (3.1) yields
ΨUΩτΩ = U τ . Since UΩτΩR = τΩQ, it follows that U τ R = ΨUΩτΩR = ΨτΩQ =
τ Q.
(cid:3)
18
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
Proposition 4.3. Let {A, T ′, V, R, Q} be a lifting data set. Assume that R∗R =
Q∗Q, and let ω be the unitary operator defined in (4.1). Consider a minimal special
isometric coupling {U, τ } of {T ′, A}, and let F{U,τ } be the function in the Schur class
S(DA, DT ′ ⊕ DA) defined by (3.10). Then U τ R = τ Q if and only F{U,τ }(0)F = ω.
In particular, there exists a special isometric coupling {U, τ } of {T ′, A} satisfying
U τ R = τ Q.
It will be convenient first to prove the following lemma.
Lemma 4.4. Let {A, T ′, U ′, R, Q} be a lifting data set satisfying R∗R = Q∗Q. Let
{U, τ } be a special isometric coupling of {T ′, A}, and consider its operator matrix
representation of the form
(4.8) U = (cid:20)
T ′
Y1DT ′ Y2 (cid:21) on (cid:20) H′
0
M (cid:21) where Y = (cid:2) Y1 Y2 (cid:3) : (cid:20) DT ′
M (cid:21) → M
is an isometry. Then U τ R = τ Q if and only if Y F ′ = ω∗.
Proof. Since the coupling is special, the space DA is a subspace of M and
τ = (cid:20)
A
DA DA (cid:21) : H → (cid:20) H′
M (cid:21) .
Π∗
It follows that for h in H0, we have
T ′ARh
Y1DT ′ ARh + Y2DARh (cid:21)
U τ Rh = (cid:20)
=
Thus
T ′
0
AQh
DARh (cid:21) = (cid:20)
Y1DT ′ Y2 (cid:21)(cid:20) ARh
DARh (cid:21)
.
DAR (cid:21) h = DAQh,
Y (cid:20) DT ′ ARh
U τ R = τ Q ⇐⇒ Y (cid:20) DT ′ AR
h ∈ H0.
Since Y is an isometry, we see that U τ R = τ Q if and only if Y F ′ is a unitary
operator from F ′ onto F with the same action as ω∗. Because of the uniqueness of
ω, this proves the lemma.
(cid:3)
Proof of Proposition 4.3. Let Y be the isometry determined by the operator
matrix representation for U in (4.8), and set F = F{U,τ }. From Lemma 4.4 we
know that U τ R = τ Q if and only if Y F ′ = ω∗. Thus we have to show that
F (0)F = ω ⇐⇒ Y F ′ = ω∗.
(4.9)
By consulting (3.10) we see that F (0) = ΠDT ′ ⊕DA Y ∗Π∗
DA. The fact that F ′ ⊂
DT ′ ⊕ DA allows us to view ω as an isometry from F into DT ′ ⊕ DA. Since F ⊂ DA,
it follows that the first condition in (4.9) is equivalent to
Y ∗F = (cid:20) ω
γ (cid:21) : F → (cid:20) DT ′ ⊕ DA
X
(cid:21) ,
where γ is some operator from F into X . However, ω is an isometry and Y ∗F is
a contraction. This implies that γ = 0. We conclude that the first condition in
(4.9) is equivalent to Y ∗F = ω. By taking adjoints, and using that ω is a unitary
operator from F onto F ′, we see that the same holds true for the second condition
in (4.9).
COUPLING AND RELAXED COMMUTANT LIFTING
19
Finally, for each λ ∈ D define the operator F (λ) from DA into DT ′ ⊕DA by setting
F (λ)d = ωΠF d for d ∈ DA. Here ΠF is the orthogonal projection of DA onto F .
Then F belongs to the Schur class S(DA, DT ′ ⊕ DA). Hence, by Theorem 3.4, there
exists a minimal special isometric coupling {U, τ } of {T ′, A} such F = F{U,τ }. Since
F (0)F = ω, we conclude that U τ R = τ Q.
(cid:3)
Proof of Theorem 4.1. We split the proof into two parts.
Part 1. Assume that B is a contractive interpolant for the data set {A, T ′, V, R, Q}.
Since R∗R = Q∗Q, we know from Proposition 4.2 that there exists a (minimal)
special isometric coupling {U on K, τ } of {T ′, A} such that U τ R = τ Q and B =
Φ∗τ , where Φ is the unique isometry from V = H′ ⊕ H 2(DT ′ ) into K associated
with T ′ intertwining V with U . Now write
U = (cid:20)
T ′
Y1DT ′ Y2 (cid:21) on (cid:20) H′
M (cid:21) ,
0
where M = K ⊖ H′. Since {U, τ } is special, we have DA ⊂ M, and
τ = (cid:20)
A
DA DA (cid:21) : H → (cid:20) H′
M (cid:21) .
Π∗
The identity B = Φ∗τ and the formula for Φ in Theorem 1.1 show that B =
A ⊕ Λ∗Π∗
DA DA, where
(Λ∗m)(λ) = Y ∗
1 (I − λY ∗
2 )−1, m ∈ M and λ ∈ D.
It follows that
Bh = (cid:20)
Y ∗
1 (I − λY ∗
Ah
2 )−1Π∗
DA DAh (cid:21) ,
h ∈ H.
To obtain the expression for B given in (4.2) we apply Proposition 2.3 with D = DA
and D′ = DT ′ . It follows that (4.2) holds with F ∈ S(DA, DT ′ ⊕ DA) given by
F (λ) = ΠDT ′ ⊕DAY ∗(IM − λJ ′
X Y ∗)−1,
λ ∈ D.
Here X = M ⊖ DA, the operator ΠDT ′ ⊕DA is the orthogonal projection of DT ′ ⊕ M
onto DT ′ ⊕ DA, and
J ′
X = (cid:2) 0 PX (cid:3) : (cid:20) DT ′
M (cid:21) → M, Y = (cid:2) Y1 Y2 (cid:3) : (cid:20) DT ′
M (cid:21) → M.
In other words, using the terminology introduced in Theorem 3.4, we have F =
F{U,τ }. Since U τ R = τ Q, Proposition 4.3 shows that F (0)F = ω, which completes
the first part of the proof.
Part 2. Let F be any function in S(DA, DT ′ ⊕ DA) satisfying F (0)F = ω. We
have to show that B defined by (4.2) is a contractive interpolant for the given data
set. According to Theorem 3.4 there is a minimal special isometric coupling {U, τ }
of {T ′, A} such that F = F{U,τ }, where F{U,τ } is defined by (3.10). The fact that
F (0)F = ω yields U τ R = τ Q, by Proposition 4.3.
Since B is given by (4.2), we can use Proposition 2.3 (with D = DA and
D′ = DT ′ ) and Theorem 1.1 to show that B = Φ∗τ , where Φ is the unique isom-
etry associated with T ′ intertwining V (the Sz.-Nagy-Schaffer minimal isometric
lifting of T ′) with U . This allows us to apply Proposition 4.2 to show that B is a
contractive interpolant.
(cid:3)
20
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
5. Proof of the first main theorem
In this section we shall prove Theorem 0.1. The proof will be based on the
analogous result for the case when R∗R = Q∗Q, which was proved in the preceding
section, and on Proposition 5.1 below, which allows us to reduce the general case
to the case when R∗R = Q∗Q.
Throughout this section {A, T ′, V, R, Q} is a lifting data set with V being the
Sz.-Nagy-Schaffer minimal isometric lifting of T ′. As before, put D◦ = D◦H0, where
D◦ is the positive square root of Q∗Q − R∗R. Introduce the following operators:
D◦ (cid:21) ,
T ′
0
0
0
A◦ = (cid:20) A
R◦ = (cid:20) R
ID◦ (cid:21) : (cid:20) H
D◦ (cid:21) : H0 → (cid:20) H
D◦ (cid:21) → (cid:20) H′
D◦ (cid:21) , Q◦ = (cid:20) Q
D◦ (cid:21) ,
0
◦ = (cid:20) T ′
0 (cid:21) on (cid:20) H′
D◦ (cid:21) ,
0 (cid:21) : H0 → (cid:20) H
H 2(DT ′ )
H 2(D◦)
H′
D◦
.
on
V =
T ′
0
EDT ′ DT ′
0
0
0
0
ED◦
0
0
SDT ′
0
0
0
0
SD◦
◦, and that the quintet
◦, V , R◦, Q◦}
Here EDT ′ and ED◦ are the canonical embeddings of DT ′ and D◦ onto the spaces of
constant functions of H 2(DT ′ ) and H 2(D◦), respectively, and SDT ′ and SDA are the
forward shifts on H 2(DT ′ ) and H 2(DA), respectively. Identifying H 2(DT ′ ⊕D◦) with
H 2(DT ′ ) ⊕ H 2(D◦) it is straightforward to check that V is the Sz.-Nagy-Schaffer
minimal isometric lifting of T ′
(5.1)
is a lifting data set satisfying R∗
Proposition 5.1. If B from H⊕D◦ to H′ ⊕D◦⊕H 2(DT ′ )⊕H 2(D◦) is a contractive
interpolant for the data set (5.1), then the operator B from H to H′ ⊕ H 2(DT ′ ),
defined by
{A◦, T ′
◦R◦ = Q∗
◦Q◦.
(5.2)
B = ΠH′⊕H 2(DT ′ )
BΠ∗
H,
is a contractive interpolant for the data set {A, T, V, R, Q}, and all contractive
interpolants for {A, T, V, R, Q} are obtained in this way.
Proof. Let B be a contractive interpolant for the data set (5.1). Then B is of the
following form
(5.3)
where
B =
A
0
H′
D◦
Γ1DA
Γ2DA
: (cid:20) H
0
ID◦
0
0
D◦ (cid:21) →
Γ2 (cid:21) : DA → (cid:20) H 2(DT ′ )
(cid:20) Γ1
H 2(D◦) (cid:21) is a contraction.
H 2(DT ′ )
H 2(D◦)
,
Moreover, V BR◦ = BQ◦. Now, using this B, let B be the operator defined by
(5.2). In other words
B = (cid:20) A
Γ1DA (cid:21) : H → (cid:20)
H′
H 2(DT ′ ) (cid:21) .
COUPLING AND RELAXED COMMUTANT LIFTING
21
By virtue of V BR◦ = BQ◦ it follows that
(cid:20)
T ′
EDT ′ DT ′ SDT ′ (cid:21)(cid:20) AR
Γ1DAR (cid:21) = (cid:20) AQ
Γ1DAQ (cid:21) .
0
Thus B is a contraction, A = ΠH′ B and V BR = BQ, that is, B is a contractive
interpolant for the data set {A, T ′, V, R, Q}.
Next, let B from H to V = H′ ⊕ H 2(DT ′ ) be an arbitrary contractive interpolant
for the data set {A, T ′, V, R, Q}. We have to show that B is given by (5.2), where
B is some contractive interpolant for the data set (5.1). In fact, from (5.3) we see
that it suffices to find a contraction Γ from DB into H 2(D◦) such that the operator
B, given by
(5.4)
B =
B
0
ΓDB
0
ID◦
0
: (cid:20) H
D◦ (cid:21) →
V
D◦
H 2(D◦)
,
satisfies the intertwining relation W BR◦ = BQ◦, where W is the operator which
one obtains by interchanging the second and the third column and the second and
third row in the operator matrix for V . Put
D◦ (cid:21) → (cid:20) V
(5.5) B◦ = (cid:20) B
0 (cid:21) on (cid:20) V
ID◦ (cid:21) : (cid:20) H
V◦ = (cid:20) V
D◦ (cid:21) .
D◦ (cid:21) ,
0
0
0
0
Since V BR = BQ, we have V◦B◦R◦ = B◦Q◦. Now, notice that B◦ = B ⊕ ID◦
is a contraction. Furthermore, V◦ is a partial isometry, and the Sz.-Nagy-Schaffer
minimal isometric lifting of V◦ is equal to W . Thus
(cid:8)B◦, V◦, W , R◦, Q◦(cid:9)
is a lifting data set. Since R∗
set (5.6) has a contractive interpolant B. By identifying the spaces
◦Q◦, we know from Theorem 4.1 that the data
◦R◦ = Q∗
(5.6)
H′ ⊕ D◦ ⊕ H 2(DT ′ ) ⊕ H 2(D◦)
and H′ ⊕ H 2(DT ′ ) ⊕ D◦ ⊕ H 2(D◦).
one sees that this operator B is also a contractive interpolant for the data set (5.1),
and from (5.4) it follows that with this choice of B the identity (5.2) holds.
(cid:3)
Proof of Theorem 0.1. We split the proof into two parts.
Part 1. Let B be a contractive interpolant for the data set {A, T ′, V, R, Q}. Then
B is of the form (5.2) for some contractive interpolant B for {A◦, T ′
◦, V , R◦, Q◦}.
◦Q◦, we can use Theorem 4.1 to find a formula for B. To write
Since R∗
this formula, we need the subspaces
◦R◦ = Q∗
F◦ = DA◦ Q◦H0
and F ′
◦ = (cid:20) DT ′
DA◦ R◦
◦ A◦R◦
and the unitary operator ω◦ from F◦ onto F ′
ω◦DA◦Q◦ = (cid:20) DT ′
◦ given by
◦ A◦R◦
DA◦ R◦
(cid:21) .
In this setting,
(5.7)
DA◦ = (cid:20) DA 0
0 (cid:21)
0
and DT ′
◦ = (cid:20) DT ′
0
(cid:21) H0,
0
ID◦ (cid:21) .
22
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
A straightforward computation shows that
DA◦ Q◦ = (cid:20) DAQ
0
(cid:21) : H0 → (cid:20) H
D◦ (cid:21)
and
◦A◦R◦
(cid:20) DT ′
DA◦ R◦
(cid:21) =
DT ′ AR
D◦
DAR
0
: H0 →
H′
D◦
H
D◦
.
By interchanging in the last column the first two coordinate spaces and identifying
the vector x ⊕ 0 with the vector x, we see that
(5.8)
F◦ = F ,
F ′
◦ = F ′
and ω◦ = ω,
where the subspaces F and F ′ and the unitary operator ω are defined in Section
0. Let us now apply Theorem 4.1 to B. It follows that
(5.9)
Bx = (cid:20)
A◦x
DA◦ x (cid:21) ,
x ∈ (cid:20) H
D◦ (cid:21) ,
ΠT ′
◦ F◦(λ)(cid:0)IDA◦ − λΠA◦ F◦(λ)(cid:1)−1
where F◦ ∈ S(DA◦ , DT ′
◦ ⊕ DA◦ ) satisfies F◦(0)F◦ = ω◦. Here
ΠT ′
◦ = (cid:2) I
0 (cid:3) : (cid:20) DT ′
DA◦ (cid:21) → DT ′
◦
◦
and ΠA◦ = (cid:2) 0 I (cid:3) : (cid:20) DT ′
DA◦ (cid:21) → DA◦ .
◦
From (5.7) we see that we can identify in a canonical way DA◦ with DA, and DT ′
with D◦ ⊕ DT ′ . This together with (5.8) shows that we can view F◦ as a function
F from the Schur class S(DA, D◦ ⊕ DT ′ ⊕ DA) satisfying F (0)F = ω and
(5.10)
◦
B(cid:20) h
d0 (cid:21) =
Ah
0
ΠT ′ F (λ)(I − λΠAF (λ))−1DAh
ΠD◦ F (λ)(I − λΠAF (λ))−1DAh
0
d0
0
0
,
Here ΠT ′ and ΠA are the projections given by (0.4) and
d0 (cid:21) ∈ (cid:20) H
(cid:20) h
D◦ (cid:21) .
ΠD◦ = (cid:2) I
0
0 (cid:3) :
D◦
DT ′
DA
→ D◦.
Since B is obtained from B via (5.2), we conclude that B has the desired form
(0.5).
Part 2. The reverse implication is proved in a similar way. Indeed, assume that
B is given by (0.5), where F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) satisfies F (0)F = ω. Using
the identifications made in the first part of the proof, we can view F as a function
◦ ⊕ DA◦) satisfying F◦(0)F◦ = ω◦. But then we can use Theorem
F◦ ∈ S(DA◦ , DT ′
4.1 to show that B defined by (5.9) is a contractive interpolant for the data set
◦, V , R◦, Q◦}. Since B is also given by (5.10), we conclude that B and B are
{A◦, T ′
related as in (5.2). Thus Proposition 5.1 implies that B is a contractive interpolant
for {A, T ′, V, R, Q}.
(cid:3)
COUPLING AND RELAXED COMMUTANT LIFTING
23
6. Parameterization and uniqueness of solutions
In this section we prove the second main theorem (Theorem 0.2). As a conse-
quence of this theorem we obtain conditions on the lifting data set {A, T ′, V, R, Q}
guaranteeing that the parameterization in Theorem 0.1 is proper, that is, con-
ditions on {A, T ′, V, R, Q} implying that for every contractive interpolant B for
{A, T ′, V, R, Q} there exists a unique F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω
such that B = BF . Here BF is the contractive interpolant for {A, T ′, V, R, Q} pro-
duced by the Schur class function F from S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω
as in Theorem 0.1, that is,
(6.1)
BF h = (cid:20)
ΠT ′ F (λ)(IDA − λΠAF (λ))−1DAh (cid:21) ,
Ah
h ∈ H.
We shall also present conditions on {A, T ′, V, R, Q} implying the existence of a
unique interpolant for {A, T ′, V, R, Q}.
To shorten the notation in this section we define
V = H′ ⊕ H 2(DT ′ ) and
(6.2)
Also, for a given contractive interpolant B for {A, T ′, V, R, Q} we define the spaces
FB and F ′
V = H′ ⊕ D◦ ⊕ H 2(DT ′ ) ⊕ H 2(D◦).
B by
(6.3)
FB = DBQH0
and F ′
B = (cid:20) D◦
DBR (cid:21) H0.
Notice that GB and G′
B in (0.6) are then given by
(6.4)
GB = DB ⊖ FB and G′
B = (D◦ ⊕ DB) ⊖ F ′
B.
With the above notation and definitions we can reformulate Theorem 0.2 as follows.
Theorem 6.1. Let {A, T ′, V, R, Q} be a lifting data set with V the Sz.-Nagy-
Schaffer minimal isometric lifting of T ′, and let B be a contractive interpolant
for the data set {A, T ′, V, R, Q}. Then there exists a one to one mapping from the
set of all F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω such that B = BF onto the
Schur class S(GB, G′
B), with GB and G′
B as in (6.4).
For the proof of Theorem 6.1 it will be convenient to first prove two lemma's. Let
◦, V , R◦, Q◦} be as defined in Section 5. Given a contractive interpolant B for
{A◦, T ′
the data set {A, T ′, V, R, Q}, we let B◦ and V◦ be the operators defined by (5.5).
Furthermore, as in the previous section, we define W to be the operator which
one obtains by interchanging the second and the third column and the second and
third row in the operator matrix for V . Recall that both {A◦, T ′
◦, V , R◦, Q◦} and
{B◦, V◦, W , R◦, Q◦} are lifting data sets. From the construction of W from V we
see that both W and V are minimal isometric liftings of both T ′
◦ and V◦.
Lemma 6.2. Let B be a contractive interpolant for {A, T ′, V, R, Q} and let the
pair {U on K, τ} be an isometric coupling of {V◦, B◦}. Then
◦, A◦};
(i) the pair {U, τ} is an isometric coupling of {T ′
(ii) the pair {U, τ} is minimal as an isometric coupling of {V◦, B◦} if and only
if {U, τ} is minimal as an isometric coupling of {T ′
◦, A◦};
(iii) the operator U is an isometric lifting of both T ′ and T ′
◦; moreover, ΦU,T ′ is
the canonical embedding of V into K, and ΦU,T ′
◦ v = v for all v ∈ V;
(iv) the contractive interpolant B = ΠV Φ∗
U,T ′
◦
τ Π∗
H.
24
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
Furthermore, two isometric couplings {U1, τ1} and {U2, τ2} of {V◦, B◦} are iso-
morphic as isometric couplings of {V◦, B◦} if and only if they are isomorphic as
isometric couplings of {T ′
◦, A◦}.
Proof. First remark that K can be decomposed as V ⊕ D◦ ⊕ M for some Hilbert
space M. Relative to this direct sum decomposition the operators U and τ admit
operator matrix representations of the form
(6.5)
U =
V
0
0
0
0
∗
0
0
∗
on
V
D◦
M
and τ =
B
0
∗
0
ID◦
0
: (cid:20) H
D◦ (cid:21) →
V
D◦
M
,
where ∗ represents operators which are not specified any further.
(i) Because V is an isometric lifting of T ′ and U is an isometric lifting of V , as we
can see from (6.5), we obtain that U is an isometric lifting of T ′. From (6.5) we
can immediately see that U also is an isometric lifting of the zero operator on D◦.
Hence U is an isometric lifting of T ′
◦. Since τ is as in (6.5) and ΠH′ B = A, we see
that ΠH′⊕D◦ τ = A◦. So {U, τ} is an isometric coupling of {T ′
(ii) Assume that {U, τ} is minimal as an isometric coupling of {V◦, B◦}, and thus
that the space (V ⊕ D◦) ∨ τ (H ⊕ D◦) is cyclic for U .
◦, A◦}.
Notice that in general we have for every operator W on a Hilbert space L with
U and Y subspaces L that
∞
_n=0
∞
∞
W n(U ∨ Y) = (
_n=0
W nU) ∨ (
_n=0
W nY).
Applying this with U in (6.5), the fact that V on V is a minimal isometric lifting
of T ′ and the fact that (V ⊕ D◦) ∨ τ (H ⊕ D◦) is cyclic for U yield
∞
K =
U n((V ⊕ D◦) ∨ τ (H ⊕ D◦))
∞
U nV) ∨ (
_n=0
∞
U n(D◦ ∨ τ (H ⊕ D◦)))
= V ∨ (
U n(D◦ ∨ τ (H ⊕ D◦)))
_n=0
V V nH′) ∨ (
Π∗
∞
= (
U nH′) ∨ (
U n(D◦ ∨ τ (H ⊕ D◦)))
_n=0
U n(D◦ ∨ τ (H ⊕ D◦)))
∞
_n=0
= (
= (
_n=0
_n=0
∞
∞
∞
_n=0
_n=0
_n=0
∞
=
U n((H′ ⊕ D◦) ∨ τ (H ⊕ D◦)).
Hence (H′ ⊕ D◦) ∨ τ (H ⊕ D◦) is cyclic for U , and thus {U, τ} is minimal as an
isometric coupling of {T ′
◦, A◦}.
Conversely, assume that {U, τ} is minimal as an isometric coupling of
{T ′
◦, A◦}. In other words, (H′ ⊕ D◦) ∨ τ (H ⊕ D◦) is cyclic for U . Note that H′ is a
subspace of V and hence (H′ ⊕D◦)∨τ (H⊕D◦) is a subspace of (V ⊕D◦)∨τ (H⊕D◦).
COUPLING AND RELAXED COMMUTANT LIFTING
25
This implies that (V ⊕D◦)∨τ (H⊕D◦) is cyclic for U as well. Thus {U, τ} is minimal
as an isometric coupling of {V◦, B◦}.
(iii) We already showed, in (i), that U is an isometric lifting of both T ′ and T ′
◦.
From (6.5) we see that V is the minimal isometric lifting of T ′ associated with U
and thus, using the remark in the final paragraph of Section 1, we obtain that
ΦU,T ′ = Π∗
V IV = Π∗
V .
◦ is the unique isometry associated with T ′
V ΦV,T ′ = Π∗
Because ΦU,T ′
U , we see that the isometry ΦU,T ′
◦ Π∗
V satisfies ΦU,T ′
◦ Π∗
U ΦU,T ′
◦ Π∗
V = ΦU,T ′
◦
V Π∗
V = ΦU,T ′
◦ that intertwines V with
V h = h for all h ∈ H′ and
◦ Π∗
V V.
Hence ΦU,T ′
U . Thus for all v ∈ V we have
◦ Π∗
V is the unique isometry associated with T ′ that intertwines V and
ΦU,T ′
◦ v = ΦU,T ′
◦ Π∗
V v = ΦU,T ′ v = v.
(iv) In the proof of (iii) we saw that ΦU,T ′
obtain that
◦ Π∗
V = ΦU,T ′ = Π∗
V . Hence from (6.5) we
B = ΠV τ Π∗
H = ΠV Φ∗
U,T ′
◦
τ Π∗
H.
It remains to prove the final statement of the lemma. For this purpose, let
{U1 on K1, τ1} and {U2 on K2, τ2} be isometric couplings of {V◦, B◦}.
First assume that {U1, τ1} and {U2, τ2} are isomorphic as isometric couplings of
{V◦, B◦}. We can immediately see from the definition of an isomorphism and the
fact that H′ ⊕ D◦ is a subspace of V ⊕ D◦, that every isomorphism from {U1, τ1}
to {U2, τ2} as isometric couplings of {V◦, B◦} also is an isomorphism from {U1, τ1}
to {U2, τ2} as isometric couplings of {T ′
◦, A◦}. Hence {U1, τ1} and {U2, τ2} are
isomorphic as isometric couplings of {T ′
◦, A◦}.
Conversely, assume that {U1, τ1} and {U2, τ2} are isomorphic as isometric cou-
◦, A◦} and that Ψ is an isomorphism from {U1, τ1} to {U2, τ2}. Then
V h = h for each h ∈ H′. Since V is the
plings of {T ′
ΨΠ∗
minimal isometric lifting of T ′ associated with U1, we obtain
V is an isometry from V to K2 with ΨΠ∗
U2ΨΠ∗
V = ΨU1Π∗
V = ΨΠ∗
V V.
Thus with (iii) we see that
ΨΠ∗
V = ΦU2,T ′ = Π∗
V .
Since Ψ is an isomorphism between isometric couplings of {T ′
◦, A◦}, the operator
Ψ is the identity on H′ ⊕ D◦. In particular, Ψd = d for each d ∈ D◦. Hence Ψ also
is an isomorphism from {U1, τ1} to {U2, τ2} as isometric couplings of {V◦, B◦}. (cid:3)
Lemma 6.3. Let B be a contractive interpolant for the data set {A, T ′, V, R, Q},
and let {U on K, τ} be an isometric coupling of {T ′
◦, A◦} such that
B = ΠV Φ∗
U,T ′
◦
τ Π∗
H.
Then there exists an isometric coupling { U , τ } of {T ′
such that { U , τ } also is an isometric coupling of {V◦, B◦}.
◦, A◦}, isomorphic to {U, τ},
Proof. From the remark in the last paragraph of Section 1 we can conclude that
ΦU,T ′
◦ associated
with U . Since U ′ is minimal and V is the Sz.-Nagy-Schaffer minimal isometric lifting
◦ , where U ′ on K′ is the minimal isometric lifting of T ′
K′ ΦU ′,T ′
◦ = Π∗
26
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
◦, we see that the unique isometry ΦU ′,T ′
of T ′
V with U ′ is unitary. Define Ψ from V ⊕ W to K, where W = K ⊖ K′, by
◦ associated with T ′
◦ that intertwines
Ψ = (cid:20) ΦU ′,T ′
0
◦
0
IW (cid:21) : (cid:20) V
W (cid:21) → (cid:20) K′
W (cid:21) .
Then Ψ is a unitary operator with Ψx = x for all x ∈ H′ ⊕ D◦. So the operators
U = Ψ∗U Ψ and τ = Ψ∗τ form an isometric coupling { U , τ } of {T ′
◦, A◦} that is
isomorphic to {U, τ}.
Since ΦU ′,T ′
◦ intertwines V and U ′, we obtain that V is the minimal isometric
◦ associated with U . Recall that V also is an isometric lifting of V◦.
lifting of T ′
Hence U is an isometric lifting of V◦.
Φ U ,T ′
◦
Because V is the minimal isometric lifting of T ′
◦ associated with U , we have
. Since { U , τ } and {U, τ} are isomorphic, Lemma 3.5 implies that
= Π∗
V
B = ΠV Φ∗
H = ΠV τ Π∗
H.
H = ΠV Π V τ Π∗
H = ΠV Φ∗
τ Π∗
τ Π∗
U,T ′
U ,T ′
◦
◦
Note that ΠH′⊕D◦ τ = A◦, and thus, since both τ and A◦D◦ are isometries, we get
that ΠV⊕D◦ τ = B◦. Hence { U , τ } is an isometric coupling of {V◦, B◦}.
(cid:3)
Proof of Theorem 6.1. Let SB be the set defined by
(6.6)
We have to show that there exists a one to one mapping from SB onto S(GB , G′
SB = {F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) F (0)F = ω and B = BF }.
B).
◦, A◦} and Proposition 4.3 to the lifting
◦, V , R◦, Q◦}, and using the identities in (5.8), we obtain that the
By applying Theorem 3.4 to the pair {T ′
data set {A◦, T ′
mapping
(6.7)
{U, τ} 7→ F{U,τ}
given by Theorem 3.4 is a one to one mapping from the set of (equivalence classes
of) minimal isometric couplings {U, τ} of {T ′
◦, A◦} satisfying U τ R◦ = τ Q◦ onto the
set of all functions F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) satisfying F (0)F = ω. Moreover,
◦, V , R◦, Q◦}, we obtain
from Proposition 4.2 and Proposition 5.1, applied to {A◦, T ′
that the mapping (6.7) maps the set of (equivalence classes of) minimal isometric
couplings {U, τ} of {T ′
H onto
the set SB defined by (6.6).
◦, A◦} satisfying U τ R◦ = τ Q◦ and B = ΠV Φ∗
τ Π∗
U,T ′
◦
Then, using Lemma 6.2 and Lemma 6.3, we obtain that there exists a one to
one mapping from SB onto the set of (equivalence classes of) minimal isometric
couplings {U, τ} of {V◦, B◦} satisfying U τ R◦ = τ Q◦.
Note that, because B is a contractive interpolant for {A, T ′, V, R, Q} and thus
V BR = BQ, we have that {B, V, V, R, Q} is a lifting data set, and that the lifting
data set {B◦, V◦, W , R◦, Q◦} is constructed from {B, V, V, R, Q} in the same way
◦, V , R◦, Q◦} from {A, T ′, V, R, Q} in Section 5. Moreover,
as we constructed {A◦, T ′
since V is an isometry and thus DV = {0}, we get that FB and F ′
B in (6.3)
correspond to {B, V, V, R, Q} as F and F ′ correspond to {A, T ′, V, R, Q}. Hence
there exists a unique unitary operator ωB from FB to F ′
B defined by
ωBDBQ = (cid:20) D◦
DBR (cid:21) .
By again applying Theorem 3.4, Proposition 4.3 and the identities in (5.8), but
now to the pair {V◦, B◦} and the lifting data set {B◦, V◦, W , R◦, Q◦}, we obtain that
COUPLING AND RELAXED COMMUTANT LIFTING
27
there exists a one to one mapping from the set of (equivalence classes of) minimal
isometric couplings {U, τ} of {B◦, V◦, W , R◦, Q◦} satisfying U τ R◦ = τ Q◦ onto the
set of all functions H ∈ S(DB, D◦⊕DB) satisfying H(0)FB = ωB. Thus there exists
a one to one mapping from SB onto the set of all functions H ∈ S(DB , D◦ ⊕ DB)
satisfying H(0)FB = ωB.
For each H ∈ S(DB, D◦ ⊕ DB) we have that H(0)FB = ωB if and only if there
exists a (unique) G in S(GB , G′
B), with GB and G′
B as in (6.4), such that
(6.8)
H(λ) = (cid:20) ωB
0 G(λ) (cid:21) : (cid:20) FB
GB (cid:21) → (cid:20) F ′
B
G′
0
B (cid:21) ,
λ ∈ D.
Hence there exists a one to one mapping from the set SB onto S(GB , G′
B).
(cid:3)
In fact, in the proof of Theorem 6.1 we do not only show that there exists a one
to one mapping from the set of F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω such
that B = BF onto S(GB , G′
B), but we actually indicate how such a mapping can
be constructed. To be more specific, the construction in the reverse way goes as
follows.
Assume that G is a Schur class function from S(GB, G′
B), with B some contractive
interpolant for {A, T ′, V, R, Q}. Define H ∈ S(DB, D◦ ⊕ DB) by (6.8). Then H
satisfies H(0)FB = ωB, and thus from Section 2 we obtain that there exists an
isometry M from D◦ ⊕ DB ⊕ Y to DB ⊕ Y, for some Hilbert space Y, such that
H(λ) = ΠD◦⊕DB M ∗(IDB ⊕Y − λJ ∗
Y M ∗)−1Π∗
DB ,
λ ∈ D,
where M satisfies the controllability type condition
(6.9)
Y = ΠY
∞
_n=0
(JY M )n(cid:20) D◦ ⊕ DB
{0}
(cid:21)
and M F ′
B = ω∗
B.
Here JY is the partial isometry given by
JY = (cid:20) 0
0 IY (cid:21) : (cid:20) DB
Y (cid:21) → (cid:20) D◦ ⊕ DB
Y
0
(cid:21) .
Notice that because V◦ and B◦ are as in (5.5) we obtain that DV◦ = ΠD◦ , DV◦ = D◦,
DB◦ = DB and DB◦ = DBΠH. Thus we can define
U = (cid:20)
V◦
M D◦ΠD◦ M (DB ⊕ Y) (cid:21) on (cid:20) V ⊕ D◦
DB ⊕ Y (cid:21)
0
and
τ = (cid:20)
B◦
DB DBΠH (cid:21) : H ⊕ D◦ → (cid:20) V ⊕ D◦
DB ⊕ Y (cid:21) .
Π∗
Then { U , τ } is a special isometric coupling of {V◦, B◦}. Because M satisfies (6.9),
the coupling { U , τ } is minimal and U τ R◦ = τ Q◦. Hence by Lemma 6.2 we obtain
that { U , τ } also is a minimal isometric coupling of {T ′
τ Π∗
H.
According to Proposition 3.1, the coupling { U , τ } is isomorphic to a special isomet-
ric coupling {U on K, τ} of {T ′
◦, A◦}. This isometric coupling {U, τ} is minimal,
satisfies U τ R◦ = τ Q◦ and, by Lemma 3.5, we have that B = ΠV Φ∗
H. The
isometry U defines an isometry Y from DT ′ ⊕ D◦ ⊕ DA ⊕ X to DA ⊕ X , with
◦ on H′ ⊕ D◦ instead of T ′ on H′. Then
X = K ⊖ (H′ ⊕ D◦ ⊕ DA), by (1.2), with T ′
◦, A◦} with B = ΠV Φ U ,T ′
Π∗
U,T ′
◦
◦
28
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
the functions F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω satisfying B = BF
corresponding to the function G ∈ S(GB, G′
B) is given by
F (λ) = ΠDT ′ ⊕D◦⊕DA Y ∗(I − λJ ∗
X Y ∗)−1ΠDA , λ ∈ D,
with JX being the partial isometry given by
JX = (cid:20) 0
0
0
IX (cid:21) : (cid:20) DA
X (cid:21) → (cid:20) DT ′ ⊕ D◦ ⊕ DA
X
(cid:21) .
From Theorem 6.1 we immediately obtain the next corollary.
Corollary 6.4. Let B be a contractive interpolant for {A, T ′, V, R, Q}. Then there
is a unique F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω such that B = BF if and
only if FB = DB or F ′
B = D◦ ⊕ DB.
The next lemma gives some sufficient conditions on {A, T ′, V, R, Q} under which
the parameterization in Theorem 0.1 is proper. To this end, define the subspace
F ′
A of D◦ ⊕ DA by
(6.10)
F ′
A = (cid:20) D◦
DAR (cid:21) H0.
Lemma 6.5. Let {A, T ′, V, R, Q} be a lifting data set. Then
(i) F = DA implies that there exists a unique contractive interpolant B and
that FB = DB;
A = D◦ ⊕ DA implies that F ′
B = D◦ ⊕ DB for every contractive interpolant
(ii) F ′
B.
If either F = DA or F ′
A = D◦ ⊕ DA holds, then the mapping F → BF given by (6.1)
is one to one from the set of all F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) satisfying F (0)F = ω
onto the set of all contractive interpolants.
Proof. Let B be a contractive interpolant for {A, T ′, V, R, Q}. Then ΠH′ B = A,
hence there exists a contraction Γ from DA to H 2(DT ′ ) such that
B = (cid:20) A
ΓDA (cid:21) : H → (cid:20)
H′
H 2(DT ′ ) (cid:21) .
From this we obtain that for all h ∈ H
kDBhk2 = khk2 − kBhk2 = khk2 − kAhk2 − kΓDAhk2
= kDAhk2 − kΓDAk2 = kDΓDAhk2.
Note that DΓ ⊂ DA and thus DΓDAH = DΓ. Hence there exists a unitary operator
γ from DΓ onto DB such that DB = γDΓDA.
(i) Assume that F = DA. Then there is only one F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with
F (0)F = ω. Hence, by Theorem 0.1, there can be only one contractive interpolant
B for {A, T ′, V, R, Q}, and for this contractive interpolant B there can be only one
F in S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω such that B = BF . Moreover, we
have
FB = DBQH0 = γDΓDAQH0 = γDΓDAQH0 = γDΓF = γDΓDA = DB.
COUPLING AND RELAXED COMMUTANT LIFTING
29
(ii) Assume that F ′
A = D◦ ⊕ DA. Then we have that
F ′
0
0
DAR (cid:21) H0
B = (cid:20) D◦
= (cid:20) ID◦
= (cid:20) ID◦
DBR (cid:21) H0 = (cid:20) ID◦
γDΓ (cid:21)(cid:20) D◦
γDΓ (cid:21)(cid:20) D◦
γDΓ (cid:21)(cid:20) D◦
DAR (cid:21) H0 = (cid:20) ID◦
DA (cid:21) = (cid:20)
0
0
0
0
0
γDΓ (cid:21) F ′
DB (cid:21) .
γDΓDA (cid:21) = (cid:20) D◦
D◦
A
0
The final statement of the lemma follows immediately from (i), (ii) and Corol-
(cid:3)
lary 6.4.
For the classical commutant lifting theorem, that is, when H = H0, R = IH and
Q is an isometry on H, we have already seen in Section 0 that the parameterization
in Theorem 0.1 is proper. This result also follows from Lemma 6.5 (ii). Indeed, if
H = H0, R = IH and Q is an isometry on H, then
F ′
A = {0} ⊕ DAIHH = {0} ⊕ DA = D◦ ⊕ DA.
Finally we derive some sufficient conditions on {A, T ′, V, R, Q} guaranteeing that
there is only one contractive interpolant. From Lemma 6.5 we already know that
the condition F = DA is such a condition. In the same way we can see that the
condition F ′ = DT ′ ⊕ D◦ ⊕ DA is sufficient.
For the classical commutant lifting theorem the combination of these two con-
ditions is also a necessary condition. That is, if H = H0, R = IH and Q is an
isometry on H, then there is only one contractive interpolant if and only if F = DA
or F ′ = DT ′ ⊕ D◦ ⊕ DA. We can see this as follows.
If the parametrization in
Theorem 0.1 for the lifting data set {A, T ′, V, R, Q} is proper, then there is only
one contractive interpolant for {A, T ′, V, R, Q} if and only if there is only one F in
S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω. The latter is equivalent to the condition
'F = DA or F ′ = DT ′ ⊕ D◦ ⊕ DA'.
Notice that when T ′ is an isometry, then the Sz.-Nagy-Schaffer minimal isometric
lifting of T ′ is T ′ itself. So in that case there also is only one contractive interpolant
B for {A, T ′, V, R, Q}, namely B = A. In the next lemma we summarize the above,
and improve the condition F ′ = DT ′ ⊕ D◦ ⊕ DA a bit further.
Proposition 6.6. Assume that for {A, T ′, V, R, Q} either T ′ is an isometry, F =
DA or DT ′ ⊕ DA ⊂ F ′. Then there exists a unique contractive interpolant for
{A, T ′, V, R, Q}.
Proof. We have already seen above that the requirement T ′ is an isometry and
the equality F = DA are both sufficient conditions. So assume that we have
DT ′ ⊕ DA ⊂ F ′. Define for all F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) the Schur class functions
F◦ = ΠD◦ F , FT ′ = ΠDT ′ F and FA = ΠDA F . Hence for all λ ∈ D
F (λ) =
F◦(λ)
FT ′ (λ)
FA(λ)
: DA →
D◦
DT ′
DA
.
Then we have
(6.11) ΠT ′ F (λ)(IDA − λΠAF (λ))−1DA = FT ′ (λ)(IDA − λFA(λ))−1DA,
λ ∈ D.
30
A.E. FRAZHO, S. TER HORST, AND M.A. KAASHOEK
All F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω admit a matrix representation of
the form
F (λ) = (cid:20) ω
0 G(λ) (cid:21) : (cid:20) F
0
G (cid:21) → (cid:20) F ′
G′ (cid:21) ,
λ ∈ D,
for some G ∈ S(G, G′) where G = DA ⊖ F and G′ = (D◦ ⊕ DT ′ ⊕ DA) ⊖ F ′. Hence,
because DT ′ ⊕ DA ⊂ F ′ all F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω have
identical FT ′ and FA and thus from (6.11) we see that BF is the same operator for
all F ∈ S(DA, D◦ ⊕ DT ′ ⊕ DA) with F (0)F = ω. Hence by Theorem 0.1 there is
only one contractive interpolant.
(cid:3)
Acknowledgement. We thank Ciprian Foias for useful discussions on an earlier
version of this paper.
Added in proof. At the IWOTA 2004 conference in Newcastle, when a preliminary
version of this paper had been completed, the authors learned that W.S. Li and
D. Timotin had a preprint ready in which the coupling method was also used to
study the relaxed commutant lifting problem and the set of its solutions. Although
the same method was used in the same area the two papers turned out to be quite
complementary in style and results. We are happy that the editors of Integral
Equations and Operator Theory agreed to publish the final versions of both papers,
one directly after the other in this volume.
References
[1] V.M. Adamjan and D.Z. Arov, Scattering operators and contraction semigroups in Hilbert
space, Doklady 165 (1965), 1377 -- 1380.
[2] V.M. Adamjan and D.Z. Arov, On the unitary couplings of isometric operators, Mat. Issled.
Kisinev 1 (1966), 3- 66 (Russian).
[3] D. Alpay, V. Bolotnikov, Y. Peretz, On the tangential interpolation problem for H2 functions.
Trans. Amer. Math. Soc. 347 (1995), 675 -- 686.
[4] T. Ando, De Branges spaces and analytic functions, Lecture notes of the division of Applied
Mathematics Research Institute of Applied Electricity, Hokkaido University, Sapporo, Japan,
1990.
[5] R. Arocena, Generalized Toeplitz kernels and dilations of intertwining operators, Integral
Equations and Operator Theory, 6 (1983), 759 -- 778.
[6] R. Arocena, On the parameterization of Adamjan, Arov and Krein, Publ. Math. Orsay 83
(1983), 7 -- 23.
[7] R. Arocena, On generalized Toeplitz kernels and their relation with a paper of Adamjan,
Arov and Krein, in: Functional Analysis Homomorphy and Approximation Theory Math.
Studies 86, North-Holland Amsterdam, 1984, pp. 1 -- 22.
[8] R. Arocena, A theorem of Naimark, linear systems and scattering operators, J. Funct. Anal.
69 (1986), 281 -- 288.
[9] R. Arocena, Unitary extensions of isometries and contractive intertwining dilations, in: The
Gohberg Anniversary Collection II, OT 41, Birkauser Verlag Basel, 1989, pp. 13 -- 23.
[10] D.Z. Arov, M.A. Kaashoek and D.R. Pik, Optimal time-variant systems and factorization
of operators, I: minimal and optimal systems, Integral Equations and Operator Theory, 31
(1998), 389 -- 420.
[11] H. Bart, I. Gohberg and M.A. Kaashoek, Minimal factorization of matrix and operator
functions, OT 1, Birkhauser Verlag, Basel, 1979.
[12] A. Biswas, C. Foias and A. E. Frazho, Weighted Commutant Lifting, Acta Sci. Math.
(Szeged), 65 (1999), 657-686.
[13] C. Foias and A. E. Frazho, The Commutant Lifting Approach to Interpolation Problems, OT
44, Birkhauser Verlag, Basel, 1990.
[14] C. Foias, A.E. Frazho, and M.A. Kaashoek, Relaxation of metric constrained interpolation
and a new lifting theorem, Integral Equations and Operator Theory, 42 (2002), 253 -- 310.
COUPLING AND RELAXED COMMUTANT LIFTING
31
[15] A.E. Frazho and M.A. Kaashoek, A Naimark dilation perspective of Nevanlinna-Pick inter-
polation, Integral Equations and Operator theory, 42 (2002), 253 -- 310.
[16] B. Sz.-Nagy and C. Foias, Harmonic Analysis of Operators on Hilbert Space, North Holland
Publishing Co., Amsterdam-Budapest, 1970.
[17] I. Gohberg, S. Goldberg and M. Kaashoek, Classes of Linear Operators Vol.II, OT 63,
Birkhauser Verlag, Basel, 1993.
[18] D. Sarason, Exposed points in H 1. I, in: The Gohberg anniversary collection, Vol. II, OT
41, Birkhauser Verlag, Basel, 1989, pp. 485 -- 496.
[19] B. Sz.-Nagy and C. Foias, Dilation des commutants d'op´erateurs, C. R. Acad. Sci. Paris,
s´erie A, 266 (1968), 493-495.
[20] S. Treil and A. Volberg, A fixed point approach to Nehari's problem and its applications, in:
The Harold Widom Anniversary Volume, OT 71, Birkhauser Verlag, Basel, 1994, pp.165-186.
Department of Aeronautics and Astronautics, Purdue University, West Lafayette,
IN 47907, USA
E-mail address: [email protected]
Afdeling Wiskunde,, Faculteit der Exacte Wetenschappen, Vrije Universiteit, De
Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands
E-mail address: [email protected]
Afdeling Wiskunde, Faculteit der Exacte Wetenschappen, Vrije Universiteit, De
Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands
E-mail address: [email protected]
|
1206.1244 | 1 | 1206 | 2012-06-06T14:33:37 | New examples of K-monotone weighted Banach couples | [
"math.FA"
] | Some new examples of K-monotone couples of the type (X, X(w)), where X is a symmetric space on [0, 1] and w is a weight on [0, 1], are presented. Based on the property of the w-decomposability of a symmetric space we show that, if a weight w changes sufficiently fast, all symmetric spaces X with non-trivial Boyd indices such that the Banach couple (X, X(w)) is K-monotone belong to the class of ultrasymmetric Orlicz spaces. If, in addition, the fundamental function of X is t^{1/p} for some p \in [1, \infty], then X = L_p. At the same time a Banach couple (X, X(w)) may be K-monotone for some non-trivial w in the case when X is not ultrasymmetric. In each of the cases where X is a Lorentz, Marcinkiewicz or Orlicz space we have found conditions which guarantee that (X, X(w)) is K-monotone. | math.FA | math | New examples of K-monotone
weighted Banach couples
Sergey V. Astashkin∗, Lech Maligranda†
and Konstantin E. Tikhomirov∗
Abstract
Some new examples of K-monotone couples of the type (X, X(w)), where X is a symmetric
space on [0, 1] and w is a weight on [0, 1], are presented. Based on the property of the w-
decomposability of a symmetric space we show that, if a weight w changes sufficiently fast, all
symmetric spaces X with non-trivial Boyd indices such that the Banach couple (X, X(w))
is K-monotone belong to the class of ultrasymmetric Orlicz spaces.
If, in addition, the
fundamental function of X is t1/p for some p ∈ [1,∞], then X = Lp. At the same time a
Banach couple (X, X(w)) may be K-monotone for some non-trivial w in the case when X
is not ultrasymmetric. In each of the cases where X is a Lorentz, Marcinkiewicz or Orlicz
space we have found conditions which guarantee that (X, X(w)) is K-monotone.
2
1
0
2
n
u
J
6
]
.
A
F
h
t
a
m
[
1
v
4
4
2
1
.
6
0
2
1
:
v
i
X
r
a
1
Introduction
One of the fundamental problems in interpolation theory is to find a description of all
interpolation spaces between two fixed Banach spaces X0 and X1, which form a Banach
couple ¯X = (X0, X1), i.e., the description of all intermediate Banach spaces X with
respect to ¯X such that every linear operator T : ¯X → ¯X maps X into X boundedly.
between fixed Banach spaces X0 and X1, which are defined as follows:
X0 + X1, and the inequality
An important role in the interpolation theory is played by the K-monotone spaces
if x ∈ X, y ∈
K(t, y; X0, X1) ≤ K(t, x; X0, X1) holds for all t > 0,
then y ∈ X and kykX ≤ CkxkX for some constant C ≥ 1 independent of x and y. Here
K(t, x; X0, X1) = inf{kx0kX0 + tkx1kX1 : x = x0 + x1, x0 ∈ X0, x1 ∈ X1}
is the classical K-functional of Peetre.
∗Research partially supported by RFBR grant 10-01-00077, G. S. Magnusons found of the Royal
Swedish Academy of Sciences-project number FOAMagn09-028 and Lulea University of Technology.
†Research partially supported by the Swedish Research Council (VR) grant 621-2008-5058.
2010 Mathematics Subject Classification: 46E30, 46B20, 46B42
Key words and phrases: K-monotone couples, w-decomposable Banach lattices, symmetric spaces,
ultrasymmetric spaces, weighted symmetric spaces, Lorentz spaces, Marcinkiewicz spaces, Orlicz spaces,
regularly varying functions
1
A couple ¯X = (X0, X1) is called K-monotone (or Calder´on-Mityagin couple) if all
interpolation spaces between X0 and X1 are K-monotone.
By a theorem due to Brudnyı and Krugljak [9, Theorem 4.4.5] all interpolation spaces
with respect to a K-monotone Banach couple (X0, X1) can be represented in the form
X = (X0, X1)K
Φ , where Φ is a Banach lattice of measurable functions on (0,∞) and
kxk(X0,X1)K
Φ
= kK(·, x; X0, X1)kΦ.
Moreover, even if (X0, X1) is not K-monotone, every interpolation space X with respect
to (X0, X1) which happens to be a K-monotone space satisfies X = (X0, X1)K
Φ for some
suitable Φ, and of course this is only up to equivalence of norms ( Brudnyı and Krugljak
[9, Theorem 3.3.20]). Therefore, the problem of finding new examples of K-monotone
couples or K-monotone spaces becomes very important.
Calder´on [10] and independently Mitjagin [26] proved that the couple (L1, L∞) is K-
monotone. Several years later Sedaev and Semenov [33] proved that a weighted couple
(L1(w0), L1(w1)) is K-monotone (cf. also Cwikel-Kozlov [13] for another proof) and then
Sedaev [32] generalized this result to the couples of the form (Lp(w0), Lp(w1)) (1 ≤ p ≤
∞). Finally, Sparr [35], [36] showed that (Lp(w0), Lq(w1)) is a K-monotone couple for
0 < p, q ≤ ∞. There are other proofs of Sparr's result, for example, in papers of Dmitriev
[17], Cwikel [11] and of Arazy-Cwikel [2].
In [15], Cwikel and Nilsson considered the problem of K-monotonicity from a some-
what different point of view. Namely, they studied the problem when a weighted Banach
couple (X(w0), Y (w1)), with X, Y being separable Banach lattices with the Fatou prop-
erty on a measure space (Ω, Σ, µ), is K-monotone for all weights w0, w1 on Ω. They proved
that this can happen if and only if X = Lp(v0) and Y = Lq(v1) for some weights v0, v1
and some numbers 1 ≤ p, q < ∞. In their proof the concept of a decomposable Banach
lattice on a measure space is essentially used. A Banach lattice X is called decomposable
if for any convergent series P∞n=1 fn in X with pairwise disjoint fn (n = 1, 2, . . . ) and any
(formal) series P∞n=1 gn, gn ∈ X, kgnkX ≤ kfnkX (n = 1, 2, . . . ), such that all gn are pair-
wise disjoint, we have P∞n=1 gn ∈ X and kP∞n=1 gnkX ≤ C kP∞n=1 fnkX with a constant
C independent of fn, gn. This notion or some variants of it were introduced earlier by
Cwikel [12] and Cwikel-Nilsson [14].
Note that the problem of K-monotonicity of weighted couples (X(w0), Y (w1)) can be
reduced to considering couples of the form (X, Y (w)). Therefore, in what follows, we will
examine couples with one weight only. We will say that a weight w is non-trivial if either
w or 1/w is unbounded.
In [37], the concept of w-decomposability of a Banach lattice, which generalizes in a
sense the previous one due to Cwikel, was introduced. A theorem proved in [37] states
that, whenever X is a Banach lattice with the Fatou property, the couple (X, X(w))
is K-monotone if and only if X is w-decomposable (see Theorem 3.1 below in Section
3). Earlier Kalton [18] showed that in the case of symmetric sequence spaces with the
Fatou property the K-monotonicity of a couple (X, Y (w)) for some non-trivial weight
w implies that X = lp and Y = lq for some 1 ≤ p, q ≤ ∞ (note, however, that there
exist examples of shift-invariant sequence spaces X with the Fatou property, such that
(X, X(2−k)) is K-monotone but X is not isomorphic to lp for any 1 ≤ p ≤ ∞ [5], [6]).
2
Tikhomirov's theorem from [37] allows us to examine whether the result of Kalton extends
to symmetric function spaces. We will see that this is not the case and the situation here
will be essentially different.
The paper is organized as follows. After the introduction, in Section 2, some necessary
definitions and notations are collected. In the first part, we recall necessary information
about symmetric spaces on [0, 1] and then, in the second part, regularly varying convex
Orlicz functions on [0,∞) and regularly varying quasi-concave functions on [0, 1] are
discussed.
In Section 3 we consider the notion of a w-decomposable Banach lattice, which plays
a central role in these investigations. Using the Krivine theorem we show that it can be
essentially simplified in the case of symmetric function spaces. Namely, we prove condition
(9) which means that for any w-decomposable symmetric space X there exists p ∈ [1,∞]
(depending on X) such that X has, roughly speaking, both "restricted lower and upper
p-estimates". In particular, its fundamental function ϕ satisfies condition (13) for some
p, which means that the function ϕp is "almost additive" near zero.
Section 4 contains results on the w-decomposability of Lorentz and Marcinkiewicz
spaces on [0, 1]. If ϕ is a concave increasing function on [0, 1] with γϕ > 0 and 1 ≤ p < ∞,
then the couple (X, X(w)) with X = Λp,ϕ([0, 1]) and a given non-trivial weight w is K-
monotone if and only if condition (10) holds. This couple is K-monotone for some weight
w if and only if ϕ is equivalent to a regularly varying function at 0 of order p. Moreover,
for any weight w on [0, 1] we can construct a concave function ϕ on [0, 1] such that the
couple (X, X(w)) with X = Λ1,ϕ([0, 1]) is K-monotone and Λ1,ϕ([0, 1]) 6= L1[0, 1].
We obtain analogous results for Marcinkiewicz spaces, as a consequence of a new
duality theorem which is of independent interest.
It states that under suitable mild
conditions on a Banach lattice X, the weighted couple (X, X(w)) is K-monotone if and
only if the couple (X′, X′(w)) is K-monotone, where X′ means the Kothe dual to X.
Section 5 deals with conditions of w-decomposability of Orlicz spaces LF [0, 1]. It is
shown, in Theorem 6, that if an Orlicz function F satisfies the ∆2-condition for large
arguments, then LF [0, 1] is w-decomposable if and only if it satisfies some restricted p-
upper and p-lower estimates (see condition (32)). Moreover, it is proved, in Theorem 7,
that if an Orlicz function F is equivalent to an Orlicz function which is regularly varying
at ∞ of order p ∈ [1,∞), then the Orlicz space LF = LF [0, 1] is w-decomposable for some
weight w on [0, 1] and therefore the couple (LF , LF (w)) is K-monotone.
Finally, in Section 6, we prove that if a symmetric space X on [0, 1] with non-trivial
Boyd indices is w-decomposable with respect to a weight changing sufficiently fast, then X
is an ultrasymmetric Orlicz space. The result implies that, for such a weight w, every K-
monotone couple (X, X(w)) with X having the Fatou property must be an ultrasymmetric
Orlicz space. Moreover, if its fundamental function is of the form ϕX(t) = t1/p for some
1 ≤ p ≤ ∞, then X = Lp.
2 Preliminaries
Let us collect necessary information and results, in two parts, on symmetric (rear-
rangement invariant) spaces and regularly varying functions.
3
2a. Symmetric spaces. Let (Ω, Σ, µ) be a complete σ-finite measure space and
L0 = L0(Ω) be the space of all classes of µ-measurable real-valued functions defined on Ω.
A Banach space X = (X,k · kX ) is said to be a Banach lattice on Ω if X is a linear subspace
of L0(Ω) and satisfies the so-called ideal property, which means that if y ∈ X, x ∈ L0 and
x(t) ≤ y(t) for µ-almost all t ∈ Ω, then x ∈ X and kxkX ≤ kykX. We also assume that
the support of the space X is Ω (supp X = Ω), that is, there is an element x0 ∈ X such
that x0(t) > 0 µ-a.e. on Ω.
We will say that X has the Fatou property if 0 ≤ xn ↑ x ∈ L0 with xn ∈ X and
supn∈N kxnkX < ∞ imply that x ∈ X and kxnkX ↑ kxkX .
A Banach lattice X is said to be p-convex (1 ≤ p < ∞), respectively q-concave
(1 ≤ q < ∞), if there is a constant C > 0 such that
Xk=1
xkp)1/pkX ≤ C(
Xk=1
n
n
k(
respectively,
n
n
kxkkp
X )1/p,
(
Xk=1
kxkkq
X )1/q ≤ Ck(
xkq)1/qkX,
Xk=1
for any choice of vectors x1, x2, . . . , xn in X and any n ∈ N. If in the above definitions vec-
tors x1, x2, . . . , xn ∈ X are pairwise disjoint, then X is said to satisfy an upper p-estimate
and lower q-estimate, respectively. Of course, p-convexity implies upper p-estimate and
q-concavity implies lower q-estimate of a Banach lattice X. More properties can be found
in the book [22].
Let w be a weight on (Ω, Σ, µ), i.e., positive finite a.e.
function, and let X be a
Banach lattice on (Ω, Σ, µ). Then the weighted space X(w) on (Ω, Σ, µ) is defined by
X(w) = {x ∈ Ω : xw ∈ X} with the norm kxkX(w) = kxwkX.
In what follows, we
will always suppose that the weight w is non-trivial, that is, w or 1/w is an unbounded
function on (Ω, Σ, µ).
C
For two Banach spaces E and F the symbol E
֒→ F means that the embedding E ⊂ F
is continuous with the norm which is not bigger than C, i.e., kxkF ≤ CkxkE for all x ∈ E.
By a symmetric space (symmetric Banach function space), we mean a Banach lattice
X = (X,k · kX) on I = [0, 1] with the Lebesgue measure m satisfying the following
for any two equimeasurable functions x, y ∈ L0(I) (that is, they
additional property:
have the same distribution functions dx(λ) = dy(λ), where dx(λ) = m({t ∈ I : x(t) >
λ}), λ ≥ 0) the condition x ∈ X implies that y ∈ X and kxkX = kykX. In particular,
kxkX = kx∗kX, where x∗(t) = inf{λ > 0 : dx(λ) ≤ t}, t ≥ 0.
Recall that a non-negative function ϕ : [0, 1] → [0,∞) is called quasi-concave if it
is non-decreasing on [0, 1] with ϕ(0) = 0 and if ϕ(t)
is non-increasing on (0, 1]. The
t
fundamental function ϕX of a symmetric space X on I is defined by the formula ϕX(t) =
kχ[0, t]kX, t ∈ I.
It is well known that every fundamental function is quasi-concave on
I. Taking ϕX(t) := inf s∈(0,1)(1 + t
s )ϕX(s) we obtain a concave function ϕX satisfying
ϕX (t) ≤ ϕX(t) ≤ 2ϕX(t) for all t ∈ I. For any quasi-concave function ϕ on I the
4
Marcinkiewicz space Mϕ is defined by the norm
kxkMϕ = sup
t∈I,t>0
ϕ(t)x∗∗(t), x∗∗(t) =
1
t Z t
0
x∗(s)ds.
֒→
This is a symmetric space on I with the fundamental function ϕMϕ(t) = ϕ(t) and X
MϕX . The fundamental function of a symmetric space X = (X,k · kX) is not necessarily
concave but we can introduce an equivalent norm on X in such a way that the fundamental
function will be concave (take kxk1
For any symmetric function space X with a concave fundamental function ϕ = ϕX
there is also the smallest symmetric space with the same fundamental function. This
space is the Lorentz space given by the norm
X = max(kxkX ,kxkM ϕX
), x ∈ X).
1
kxkΛϕ =ZI
x∗(t)dϕ(t) := ϕ(0+)kxkL∞(I) +ZI
x∗(t)ϕ′(t)dt.
1
1
֒→ X
֒→ MϕX . A non-trivial symmetric function space X
C1֒→
C2֒→ L1(I), where C1 = ϕX (1), C2 = 1/ϕX(1) (see [7], Corollary 6.7 on page 78 or
We have then embeddings ΛϕX
on I = [0, 1] is an intermediate space between the spaces L1(I) and L∞(I) and L∞(I)
X
Theorem 4.1 on page 91 of [21] for a similar result when the underlying measure space is
(0,∞).)
The lower and upper Boyd indices αX resp. βX and the dilation indices γX resp. δX
of a symmetric space X on I = [0, 1] with the fundamental function ϕX = ϕ are defined
as follows:
αX := lim
t→0+
lnkσtkX→X
ln t
, βX := lim
t→∞
lnkσtkX→X
ln t
, σtx(s) = x(s/t)χI(s/t)
and
γX := γϕ = lim
t→0+
ln ¯ϕ(t)
ln t
, δX := δϕ = lim
t→∞
ln ¯ϕ(t)
ln t
, ¯ϕ(t) = sup
s,st∈I
ϕ(st)
ϕ(s)
.
We have the relations 0 ≤ αX ≤ γX ≤ δX ≤ βX ≤ 1 (see [21], pp. 101-102 and [24], p.
28).
A function F : [0,∞) → [0,∞) is called an Orlicz function if it is convex and increasing
with F (0) = 0. For a given Orlicz function F the Orlicz space LF = LF (I) on I = [0, 1]
is defined as
LF (I) = {x ∈ L0(I) : IF (cx) < ∞ for some c = c(x) > 0},
where IF (x) := RI F (x(t))dt. The Orlicz space LF is a symmetric space on I with the
so-called Luxemburg-Nakano norm defined by
kxkLF = inf {λ > 0 : IF (x/λ) ≤ 1} .
An Orlicz function F satisfies the ∆2-condition for large u if there exist constants
C ≥ 1, u0 ≥ 0 such that F (2u) ≤ CF (u) for all u ≥ u0.
5
The following notation will be used throughout the text: f
≈ g means that the
functions f and g are equivalent with the constant C > 0, that is, C−1f (t) ≤ g(t) ≤ Cf (t)
for all points t of the whole set on which these functions are defined, or at all points of
some explicitly designated subset of that set. In the case when the constant of equivalence
is not important we will write just f ≈ g. By [r] we will denote the integer part of a real
number r.
More information about Banach lattices and symmetric spaces can be found, for ex-
C
ample, in [7], [21] and [22]; about Orlicz spaces one can read e.g. in [20] and [25].
2b. Regularly varying convex and concave functions. An Orlicz function F on
[0,∞) is called regularly varying at ∞ of order p (1 ≤ p < ∞) if
F (tu)
F (t)
lim
t→∞
= up for all u > 0.
(1)
The following result is due to Kalton [18, Lemma 6.1].
Lemma 2.1. Let F be an Orlicz function. The following three conditions are equivalent:
(a) F is equivalent to a regularly varying Orlicz function at ∞ of order p ∈ [1,∞).
(b) There exists a constant C > 0 such that for any u ∈ (0, 1] we can find t0 = t0(u)
with
F (tu)
F (t)
C
≈ up
for all
t ≥ t0.
Although we do not need it here, there is an analogous definition to the one above for
Orlicz functions which are regularly varying of order p at 0 instead of at ∞ (see e.g. [18]).
However, we do need to consider quasi-concave functions which are regularly varying of
order p at 0. Before recalling the definition of these we should point out that it is not quite
analogous to the definitions for regularly varying Orlicz functions, because the power p
which appeared in (1) and in the corresponding definition in [18] will be replaced in (2)
by the power 1/p.
A function ϕ : [0, 1] → [0,∞) which is quasi-concave and satisfies ϕ(0) = 0 is said to
be regularly varying at zero of order p (1 ≤ p ≤ ∞) if
ϕ(tu)
ϕ(t)
lim
t→0+
= u1/p for all u > 0.
(2)
Abakumov and Mekler [1, Theorem 5] proved that a quasi-concave function ϕ is equiv-
alent to a quasi-concave regularly varying function at zero of order p ∈ [1,∞] if and only
if
ϕ(tu)
ϕ(t) ≈ u1/p for all u > 0.
lim sup
t→0+
The following lemma is an immediate consequence of this result (see also the proof of
Theorem 5 in [1]).
6
Lemma 2.2. A quasi-concave function ϕ on [0, 1] is equivalent to a quasi-concave function
which is regularly varying at zero of order p ∈ [1,∞] if and only if for some C > 0 and
any N ∈ N there exists τ (N) ∈ (0, 1] such that for all 0 < t ≤ τ (N), 0 < tN ≤ 1 we have
ϕ(Nt)
ϕ(t)
C
≈ N 1/p.
(3)
Recall that the fundamental function of an Orlicz space LF on [0, 1] with the Luxemburg-
F −1(1/t) for 0 < t ≤ 1 and ϕLF (0) = 0, where F −1 is the inverse
Nakano norm is ϕLF (t) =
of F (see formula (9.23) in [20] on page 79 of the English version or Corollary 5 in [25]
on page 58). The function ϕLF is quasi-concave but not necessarily concave on [0, 1] (see
[20] or [25]).
1
The notions of regularly varying Orlicz and quasi-concave functions are closely inter-
related. Using Lemmas 2.1 and 2.2 and routine arguments we establish the following
quantitative result showing a connection between an regularly varying Orlicz function F
and the fundamental function of the corresponding Orlicz space LF .
Proposition 2.3. Suppose that p ∈ [1,∞) and let F be an Orlicz function such that
both F and its complementary function F ∗ satisfy the ∆2-condition for large u. Then the
following conditions are equivalent:
(a) There exists a constant C′ > 0 such that for any N ∈ N there exists τ (N) ∈ (0, 1]
with
F (u)
F (uN−1/p)
C ′
≈ N for all u ≥ F −1(1/τ (N)).
(4)
(b) There exists a constant C > 0 such that for any N ∈ N the fundamental function
ϕLF satisfies condition (3) with the same τ (N).
3 w–decomposable Banach lattices
Later on C will denote a constant whose value may be different in its different appear-
ances.
The following notion was introduced in paper [37] and it will be very important for
us. Let X be a Banach lattice on (Ω, Σ, µ) and w be a weight on Ω. We say that X
is w-decomposable if there exists C > 0 such that for any n ∈ N and for all x1, . . . , xn,
y1, . . . , yn in X satisfying the conditions:
kxikX = kyikX , i = 1, 2, . . . , n,
(5)
and
inf w(supp xi ∪ supp yi) ≥ 2 sup w(supp xi+1 ∪ supp yi+1), i = 1, 2, . . . , n − 1,
(6)
we have that
k
n
Xi=1
xikX
C
≈ k
n
Xi=1
yikX.
7
(7)
To clarify the meaning of condition (6), consider the following example:
let X be a
Banach lattice of Lebesgue measurable functions on [0, 1] and w(t) = 1/t (0 < t ≤ 1).
Then (6) is equivalent to the following inequality
2 sup(supp xi ∪ supp yi) ≤ inf(supp xi+1 ∪ supp yi+1), i = 1, 2, . . . , n − 1.
In other words, there are some intervals [ai, bi] ⊂ [0, 1] (depending on xi, yi) such that
2bi ≤ ai+1 (i = 1, 2, . . . , n − 1), supp xi ⊂ [ai, bi] and supp yi ⊂ [ai, bi] (i = 1, 2, . . . , n).
It is not hard to see that 1/t-decomposability is equivalent to 1/tq-decomposability
and, more generally, w-decomposability and wq-decomposability are equivalent for any
weight w and any q > 0 (see [38], Corollary 2.2 on page 61).
It turns out that the w-decomposability of a Banach lattice X guarantees the K–
monotonicity of the weighted couple (X, X(w)). More precisely, Tikhomirov in [37] ob-
tained the following generalization of Kalton's results from [18].
Theorem 3.1. Suppose X is a Banach lattice on a σ–finite measure space (Ω, Σ, µ) with
supp X = Ω which has the Fatou property and w is a (non-trivial) weight on Ω. Then the
Banach couple (X, X(w)) is K–monotone if and only if X is w-decomposable.
In the case of symmetric spaces on [0, 1] the notion of w-decomposability can be
clarified by using the well–known Krivine theorem.
Proposition 3.2. Let w be a weight on [0, 1]. A symmetric space X on [0, 1] is w-
decomposable if and only if there exist C > 0 and 1 ≤ p ≤ ∞ such that for any n ∈ N
and for all x1, x2, . . . , xn ∈ X satisfying the conditions
inf w(supp xi) ≥ 2 sup w(supp xi+1), 1 ≤ i ≤ n − 1,
we have that
(8)
(9)
C
≈ n
Xi=1
X!1/p
kxikp
,
n
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
xi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X
where, as usual, in the case p = ∞ the right hand side should be replaced by max1≤i≤n kxikX.
Proof. By Krivine's theorem (see [22, Theorem 2.b.6] or [31]), there exists p ∈ [1/βX, 1/αX]
such that for every m ∈ N there are pairwise disjoint equimeasurable functions y1, y2, . . . , ym
∈ X,kykkX = 1 (k = 1, 2, . . . , m), such that for any αk ∈ R (k = 1, 2, . . . , m) we have
1
2k(αk)kp ≤(cid:13)(cid:13)(cid:13)
m
Xk=1
αkyk(cid:13)(cid:13)(cid:13)X ≤ 2k(αk)kp.
Obviously, the support of each function yk has measure not greater than 1/m.
(10)
Suppose that a symmetric space X is w-decomposable and that, for some n ∈ N,
functions x1, . . . , xn in X satisfy condition (8). Without loss of generality we may assume
that xi 6= 0 for each i = 1, 2, . . . , n. We choose m ∈ N sufficiently large so that the support
of each xi has measure greater than 1/m (and so of course we also have m ≥ n). For this
choice of m we consider the disjoint measurable functions y1, y2, . . . , ym,kykkX = 1 (k =
1, 2, . . . , m), obtained as it is described in the previous paragraph. In fact, we will only
8
need the first n of these functions, and we will only need special case of (10) for sequences
(αk) which satisfy αk = 0 for k > n. We may assume without loss of generality, that the
support of yi is contained in the support of xi for each i = 1, 2, . . . , n. (If not, since X is
symmetric, we can simply replace each yi by an equimeasurable function which has this
property and the above mentioned special case of (10) will remain valid.) Thus condition
(8) implies that condition (6) is satisfied and therefore, applying w-decomposability (see
(7)) and then the special case of (10), we obtain that
αi
n
Xi=1
(cid:13)(cid:13)(cid:13)
xi
kxikX(cid:13)(cid:13)(cid:13)X ≈(cid:13)(cid:13)(cid:13)
n
Xi=1
αiyi(cid:13)(cid:13)(cid:13)X ≈ k(αk)n
i=1kp
for all choices of real numbers αi. In particular, when αi = kxikX we obtain (9). Since
the reverse implication is obvious, the proof is complete.
For a given weight w consider the sets
Mk := {t ∈ [0, 1] : w(t) ∈ [2k, 2k+1)}, k ∈ Z.
Let (wr)∞r=1 be the non-increasing rearrangement of the sequence (m(Mk))+∞k=−∞
weight w is non-trivial it follows that wr > 0 for all r = 1, 2, . . . .
. Since the
For some fixed n ∈ N, let x1, x2, . . . , xn be functions in X. Suppose first that these n
functions satisfy condition (8). Then it is easy to see that
card{i : Mk ∩ supp xi 6= ∅} ≤ 1 for each k ∈ Z.
Alternatively, more or less conversely, suppose that the functions xi satisfy
card{k : Mk ∩ supp xi 6= ∅} ≤ 1 for each i ∈ {1, 2, . . . , n},
i.e., for each i, there exists a unique ki ∈ Z for which supp xi ⊂ Mki. Furthermore,
suppose k1 < k2 < . . . < kn. While this is not sufficient to imply that the collection of
functions x1, x2, . . . , xn satisfies condition (8), it does imply that (after relabelling) the
collection of functions x1, x3, x5, . . . satisfies (8) and so does the collection x2, x4, . . . .
By {M r}∞r=1 we will denote any rearrangement of the sets Mk (k = 0,±1,±2, . . .) such
that m(M r) = wr, r = 1, 2, . . .. Thus, by Proposition 3.2, we obtain the following result.
Theorem 3.3. Suppose w is a non-trivial weight on [0, 1]. A symmetric space X on [0, 1]
is w-decomposable if and only if there exist C > 0 and 1 ≤ p ≤ ∞ such that for any n ∈ N
and for all x1, x2, . . . , xn ∈ X satisfying the condition
supp xi ⊂ M i, 1 ≤ i ≤ n,
(11)
we have (9).
Next, we will need some corollaries of Theorem 3.3. Firstly, using the symmetry of
the norm in X, we get
9
Corollary 3.4. Let w be a non-trivial weight on [0, 1]. A symmetric space X on [0, 1] is
w-decomposable if and only if there exist C > 0 and 1 ≤ p ≤ ∞ such that for any n ∈ N
and for all pairwise disjoint x1, x2, . . . , xn ∈ X satisfying the condition
m(supp xi) ≤ wi, 1 ≤ i ≤ n,
(12)
we have (9).
Corollary 3.5. A symmetric space X on [0, 1] is w-decomposable for some non-trivial
weight w on [0, 1] if and only if there exist C > 0, 1 ≤ p ≤ ∞, and a sequence of disjoint
intervals {∆k}∞k=1 from [0, 1] such that for any n ∈ N and for all x1, x2, . . . , xn ∈ X
satisfying the condition supp xi ⊂ ∆i (1 ≤ i ≤ n) we have (9).
Corollary 3.6. Let w be a non-trivial weight on [0, 1] and let the sequence (wr)∞r=1 be
as above. Suppose that X is a w-decomposable symmetric space X on [0, 1] with the
fundamental function ϕ. Then there exist some C > 0 and p ∈ [1,∞] such that, for every
sequence of reals (τr)∞r=1 satisfying 0 < τr ≤ wr (r ∈ N), we have
∞
C
≈ ∞
Xr=1
ϕp(τr)!1/p
Xr=1
ϕ(cid:0)
τr(cid:1)
with the natural modification for p = ∞.
Corollary 3.7. Let w be a non-trivial weight on [0, 1] such that a symmetric space X on
[0, 1] is w-decomposable. Then there exist C > 0 and 1 ≤ p ≤ ∞ such that condition (3)
is fulfilled with τ (N) = wN (N ∈ N). In particular, the fundamental function ϕ of X is
equivalent to a regularly varying function at zero of order p and αX = γϕ = δϕ = βX = 1/p.
(13)
Proof. First we note that condition (3) is an immediate consequence of (13). Moreover,
it is well known that the assertion of Krivine's theorem holds for both p = 1/αX and
p = 1/βX (see [22, p. 141], [31] and [3]). Therefore, coincidence of the Boyd indices
and dilation indices follows from an inspection of the proof of Proposition 3.2 and the
inequalities αX ≤ γϕ ≤ δϕ ≤ βX (cf. [21, p. 102] and [24, p. 28]).
Let us show that, conversely, (13) can be derived from (3) with τ (N) = wN for a large
class of weights w.
Theorem 3.8. Let w be a weight on [0, 1] such that qwr+1 ≤ wr (r = 1, 2, . . . ) for some
q > 1 and let ϕ be a quasi–concave function on [0, 1]. Suppose there exist C > 0 and
1 ≤ p ≤ ∞ such that ϕ satisfies (3) with τ (N) = wN (N = 1, 2, . . . ). Then, for any
sequence of reals (τr)∞r=1 such that 0 < τr ≤ wr (r = 1, 2, . . . ), estimate (13) holds.
Proof. We present the proof for 1 ≤ p < ∞ since the case p = ∞ needs only minor
changes.
Firstly, it is easy to see that condition (3) can be extended as follows: we can find a
(possibly different) constant C > 0 such that for every real z ≥ 1 and τ (z) := τ ([z]) we
have
ϕ(zt)
ϕ(t)
C
≈ z1/p if 0 < t ≤ τ (z).
10
(14)
Let us show that for every m ∈ N there is a constant C(m) > 0 such that for all even
N ∈ N satisfying the inequality N m ≤ qN/2 and all z ∈ [1, N] we have
ϕ(zmt)
ϕ(t)
C(m)
≈ zm/p if 0 < t ≤ τ (N).
(15)
In fact, by the assumption, τ (N/2) ≥ qN/2τ (N), whence
zkt ≤ zmt ≤ N mτ (N) ≤ qN/2τ (N) ≤ τ (N/2) ≤ 1 (k = 0, 1, . . . , m)
provided that t ≤ τ (N). Therefore, using the quasi–concavity of ϕ and equivalence (14)
for max(1, z/2) we obtain that
ϕ(zkt)
ϕ(zk−1t) ≈
ϕ(max(1, z/2)zk−1t)
ϕ(zk−1t)
≈ z1/p if 0 < t ≤ τ (N),
with a constant of equivalence depending on p. Multiplying these relations for all k =
1, 2, . . . , m, we come to (15).
Next, let
ϕ0(s) = lim sup
t→0+
ϕ(ts)
ϕ(t)
for s > 0.
Clearly, condition (14) implies ϕ0(s) ≈ s1/p(s > 0). On the other hand, in view of Boyd's
result [8] (see also [24, Theorem 2.2]) ϕ0(s) ≥ sγϕ if 0 < s ≤ 1 and ϕ0(s) ≥ sδϕ if s > 1.
Since γϕ ≤ δϕ it follows that γϕ = δϕ = 1
p > 0. Therefore, there exist A > 0 and κ > 0
such that
(16)
ϕ(st)
ϕ(s) ≤ Atκ for all 0 ≤ t ≤ 1.
sup
0<s≤1
Let us prove that (13) is a consequence of (15) and (16). Take a natural number
m0 ≥ 2 such that κ m0 > 1 and consider an arbitrary sequence (τr)∞r=1 satisfying τr ≤
wr, r = 1, 2, . . . . Since the non-increasing rearrangement (τ∗r )∞r=1 of this sequence also
satisfies τ∗r ≤ wr for r = 1, 2, . . . we can assume without loss of generality that the
sequence (τr)∞r=1 is itself non-increasing. Further, set I = {r ∈ N : τr rm0 ≥ τ1}, J = N\ I.
Clearly, 1 ∈ I. By (16) and the choice of m0,
rm0(cid:17) ≤ A(cid:16)
ϕ(τ1) ≤ C1ϕ(τ1).
τr(cid:17) ≤ ϕ(cid:16)
r−m0(cid:17)κ
ϕ(cid:16)Xr∈J
∞
Xr=2
τ1
∞
Xr=2
Analogously,
ϕp(τr) ≤
Xr∈J
∞
Xr=2
ϕp(τ1/rm0) ≤ Ap
∞
Xr=2
r−p κ m0ϕp(τ1) ≤ C2 ϕp(τ1).
Thus, it is sufficient to prove equivalence (13) for (τr)r∈I.
If card I < ∞ then there is nothing to prove. So, assume that card I = ∞. Choose
a positive integer i0 ∈ I, i0 ≥ 2 such that for N = 2 [i0/2] we have N m0 ≤ qN/2. Denote
δr = (τr/τi0)1/m0 for r ∈ I ∩{1, 2, . . . , i0}. Then, by the definition of I, δr ≤ (τ1/τi0)1/m0 ≤
11
i0 ≤ 2N. Applying (15) in the case m = m0, z = max(1, δr/2) for all r ∈ I, r ≤ i0, we get
with a constant of equivalence depending on m0 and p. The last formula implies that
ϕ(τr)
ϕ(τi0)
=
ϕ(δm0
r τi0)
ϕ(τi0) ≈ δm0/p
r
Xr∈I∩{1,2,...,i0}
ϕp(τr) ≈
ϕp(τi0)
τi0
,
τr.
τi0(cid:19)1/p
=(cid:18) τr
Xr∈I∩{1,2,...,i0}
τr/τi0(cid:17)1/(m0+1)
≤ i0.
we get
On the other hand, setting δ :=(cid:16)Pr∈I∩{1,2,...,i0}
δ ≤(cid:16) Xr∈I∩{1,2,...,i0}
τ1/τi0(cid:17)1/(m0+1)
Therefore, again by (15), we obtain
ϕp(τi0)
τi0
Xr∈I∩{1,2,...,i0}
τr = δm0+1ϕp(τi0) ≈ ϕp(δm0+1τi0) = ϕp(cid:16) Xr∈I∩{1,2,...,i0}
τr(cid:17),
with a constant depending on m0 and p. Combining the above formulas and noting that
i0 can be arbitrarily large, we conclude that equivalence (13) holds and the proof is
complete.
Theorem 3.8 allows us to construct non-trivial quasi–concave functions satisfying con-
dition (13), for a large class of weights. For example, let w(t) = 1/t (0 < t ≤ 1). In
this case wr = 2−r, r = 1, 2, . . . Define ϕ(t) = t log e
t (0 < t ≤ 1). Obviously, ϕ is
quasi–concave. Elementary calculations show that (3) is fulfilled for ϕ with p = 1 and
τ (N) = wN = 2−N (N = 1, 2, . . . ). Thus, by Theorem 3.8, ϕ satisfies (13).
4 w–decomposable Lorentz and Marcinkiewicz spaces
For 1 ≤ p < ∞ and any increasing concave function ϕ, ϕ(0) = 0, the Lorentz space
Λp,ϕ consists of all classes of measurable functions x on [0, 1] such that
kxkΛp,ϕ =(cid:18)Z 1
0
[x∗(t)ϕ(t)]p dt
t (cid:19)1/p
< ∞.
The space Λp,ϕ was investigated by Sharpley [34] and Raynaud [30], who proved that if
0 < γϕ ≤ δϕ < 1, then Λp,ϕ is a symmetric space on [0, 1] with an equivalent norm
Λp,ϕ =(cid:18)Z 1
kxk⋆
0 x∗(s) ds (cf.
0
[x∗∗(t)ϕ(t)]p dt
t (cid:19)1/p
,
where x∗∗(t) = 1
[34], Lemma 3.1). Moreover, if γϕ > 0, then applying
Corollary 3 on page 57 of [21] to the function ψ = ϕp (1 ≤ p < ∞) (see also [24, Theorem
6.4(a)]), we obtain that there exists a constant K = K(p) ≥ 1 such that
t R t
K−1ϕp(t) ≤Z t
0
ϕp(s)
s
ds ≤ Kϕp(t) (0 < t ≤ 1).
(17)
12
Therefore, the fundamental function ϕΛp,ϕ(t) is equivalent to ϕ(t). Inequalities (17) imply
also that, if γϕ > 0, then the space Λ1,ϕ coincides with the Lorentz space Λϕ with the
norm
kxkΛϕ :=Z 1
0
x∗(t)dϕ(t).
Recall also that the Kothe dual of the Lorentz space Λϕ is isometric to the Marcinkiewicz
space M ϕ with ϕ(t) = t
ϕ(t) and its norm is
kxkM ϕ = sup
0<t≤1
ϕ(t)x∗∗(t) = sup
0<t≤1
(cf. [21], Theorem 5.2 on page 112).
1
ϕ(t)Z t
0
x∗(s)ds
We will prove that condition (13) is necessary and sufficient for Lorentz and Marcin-
kiewicz spaces to be w-decomposable. We start by proving a specific geometric property
of Lorentz spaces.
Proposition 4.1. Let ϕ be an increasing non-negative concave function on [0, 1] such
that γϕ > 0, and let 1 ≤ p < ∞. Then for arbitrary b > 1 there exists a constant
C = C(b, ϕ, p) > 0 with the following property: for any two-sided non-decreasing sequence
(aj)+∞j=−∞
b−jχ(aj−1,aj ] belongs to
Λp,ϕ, we have
of reals from [0, 1] such that the function x = P+∞j=−∞
kxkp
Λp,ϕ
C
≈
+∞
Xj=−∞
b−pjϕp(aj − aj−1).
(18)
Proof. Since γϕ > 0, there exist κ > 0 and A > 0 such that inequality (16) holds. Choose
a constant C1 = C1(ϕ) > 1 satisfying the inequality
(C1 + 1)κ
A
≥ 2 K 2,
(19)
where K is the constant from (17), and denote by I the set of all indices j ∈ Z such that
aj − aj−1 ≥ C1 aj−1. We prove the following equivalences:
and, if b > C1 + 1,
Z aj
aj−1
ϕp(t)
t
dt ≈ ϕp(aj − aj−1), j ∈ I
ϕp(t)
t
dt,
kxkp
Λp,ϕ ≈Xj∈I
b−pjZ aj
b−pjϕp(aj − aj−1) ≈ Xj∈I
aj−1
+∞
Xj=−∞
b−pjϕp(aj − aj−1),
(20)
(21)
(22)
with constants which depend only on b, ϕ and p.
At first, if j ∈ I then, by (16) and (19),
ϕ(aj) ≥ ϕ((C1 + 1)aj−1) ≥
(C1 + 1)κ
A
13
ϕ(aj−1) ≥ 2K 2 ϕ(aj−1).
Combining this with (17) and the inequality
ϕ(aj) ≤ ϕ(aj − aj−1) + ϕ(aj−1) ≤ 2 ϕ(aj − aj−1),
(23)
we obtain
1
2K
ϕp(aj − aj−1) ≤
1
2K
ϕp(aj) ≤
1
2K
[2 ϕp(aj) − (2K 2)p ϕp(aj−1)]
ϕp(aj)
K − K ϕp(aj−1)
1
2K
≤
≤ Z aj
≤ Z aj
0
0
[2 ϕp(aj) − 2K 2 ϕp(aj−1)] =
ϕp(t)
ϕp(t)
ϕp(t)
dt −Z aj−1
dt ≤ Kϕp(aj) ≤ 2pKϕp(aj − aj−1),
dt =Z aj
aj−1
t
t
0
dt
t
ϕp(t)
t
which implies (20).
Now, assuming b > C1 + 1, we show that the set I is unbounded from below. In fact,
otherwise there is j0 ∈ Z such that aj − aj−1 < C1aj−1 for all j ≤ j0. Then, we have
aj0 ≤ (C1 + 1)j0−j aj (j ≤ j0) and by (17) and the concavity of ϕ,
kxkp
Λp,ϕ ≥ sup
j≤j0
1
K
≥
sup
j≤j0
ϕp(t)
b−pjZ aj
0
dt ≥
(C1 + 1)p(j−j0)
t
1
K
sup
j≤j0
b−pjϕp(aj)
bpj
ϕp(aj0) = ∞.
Therefore, for a given i /∈ I we can find k = max{j < i : j ∈ I}. Further, from the
definition of I it follows that ai < (C1 + 1)i−k ak. Since ϕ is concave and 2ak−1 ≤ ak, we
get
Z ai
ai−1
ϕp(t)
t
dt ≤ Z (C1+1)i−kak
ak
ϕp(t)
t
dt
and so
b−piZ ai
ai−1
ϕ(t)
t
dt
)Z (C1+1)i−kak
ak
ϕ(t)
t
dt
ak
) ak
ak
2
≤ ϕp−1((C1 + 1)i−kak)Z (C1+1)i−kak
≤ 2p−1(C1 + 1)(p−1)(i−k) ϕp−1(
ak
≤ 2p−1(C1 + 1)p(i−k) ϕp−1(
2
ak
≤ 2p(C1 + 1)p(i−k)ϕp−1(
2
≤ 2p(C1 + 1)p(i−k) Z ak
ϕp(t)
≤ 2p(C1 + 1)p(i−k) Z ak
dt ≤ 2p(cid:16) C1 + 1
(cid:17)p(i−k)
) Z ak
ϕp(t)
ϕp(t)
ak−1
ak/2
ak/2
dt
dt
b
t
t
t
b−pk Z ak
ak−1
ϕ(ak)
ak
ϕ(t)
t
14
dt
ϕp(t)
t
dt.
Since b > C1 + 1 we obtain (21).
In a similar way, applying (23) for j = k, we get
b−piϕp(ai − ai−1) ≤ b−pi(C1 + 1)p(i−k)ϕp(ak)
≤ 2p(cid:18)C1 + 1
b (cid:19)p(i−k)
b−pkϕp(ak − ak−1),
which implies (22).
Relations (20)–(22) imply (18), so we proved the statement for b > C1 + 1. To extend
this result to all b > 1 it suffices to prove the following: whenever (18) holds for some b > 1
and arbitrary non-decreasing sequence (aj)∞j=−∞
with a constant C, it is automatically
fulfilled for b1/2 with a constant not exceeding 2pbpC. Indeed, if
y =
+∞
Xj=−∞
b−j/2χ(aj−1,aj ] and z =
b−jχ(a2j−2,a2j ],
+∞
Xj=−∞
then
kykp
Λp,ϕ =
=
bp/2
≈
On the other hand,
+∞
+∞
Xj=−∞
Xj=−∞
Xj=−∞
+∞
dt
t
aj−1
ϕp(t)
b−pj/2 Z aj
b−p(2j−1)/2 Z a2j−1
b−pj Z a2j
ϕp(t)
a2j−2
a2j−2
ϕp(t)
dt
t
+
+∞
Xj=−∞
b−pj Z a2j
a2j−1
ϕp(t)
dt
t
dt
t
= kzkp
Λp,ϕ.
b−pjϕp(a2j − a2j−2)
2pbp/2
≈ b−pjϕp(a2j − a2j−1) + b−p(2j−1)/2ϕp(a2j−1 − a2j−2),
so we get an analog of (18) for y and b1/2 and the proof is complete.
Remark 4.2. For the space Λ1,ϕ = Λϕ the result can be proved also by using the following
well-known formula (cf. formula 5.1 in [21] on page 108)
kxkΛϕ =
+∞
Xj=−∞
(b−j − b−j−1)ϕ(aj).
Let, as above, for a given weight w, Mk = {t ∈ [0, 1] : w(t) ∈ [2k, 2k+1)} (k ∈ Z) and
(wr)∞r=1 be the non-increasing rearrangement of the sequence (m(Mk))+∞k=−∞
Theorem 4.3. Let ϕ be an increasing concave function on [0, 1] such that γϕ > 0, 1 ≤ p <
∞ and let w be a weight on [0, 1]. Then the Lorentz space X := Λp,ϕ is w-decomposable
if and only if ϕ satisfies condition (13).
.
15
Proof. If X = Λp,ϕ is w-decomposable then, by Corollary 3.6, the relation (13) holds for
the fundamental function ϕX. Since, as it was mentioned above, ϕ ≈ ϕX, then (13) is
fulfilled for ϕ as well.
Conversely, suppose that ϕ satisfies (13). Let n ∈ N and x1, x2, . . . , xn be non-negative
functions from X satisfying (12). Evidently, there exist x′1, x′2, . . . , x′n ∈ X taking their
values from the set {2−k}∞k=−∞ ∪ {0} and such that xi(t)
≈ x′i(t) (0 < t ≤ 1). Clearly,
m(supp x′i) = m(supp xi) ≤ wi (1 ≤ i ≤ n) and
x′i(t) = 2−ko =
Xi=1
m{t : x′i(t) = 2−k}
mnt :
Xi=1
n
n
2
for all integer k. Therefore, applying (13), we get that
n
Xi=1
ϕp(m{t : x′i(t) = 2−k}) ≈ ϕp(m{t :
n
Xi=1
x′i(t) = 2−k}) (k ∈ Z).
(24)
(25)
(26)
On the other hand, Proposition 4.1 yields
kx′ikp
X ≈
+∞
Xk=−∞
n
and
2−pkϕp(m{t : x′i(t) = 2−k}) (1 ≤ i ≤ n)
x′ikp
X ≈
k
Xi=1
+∞
Xk=−∞
2−pkϕp(m{t :
x′i(t) = 2−k})
n
Xi=1
with a constant which depends only on ϕ and p. Combining relations (25) and (26) with
(24), we obtain (9) for x′i and so for xi. The proof is complete.
In particular, from the above theorem and a remark after Theorem 3.8 it follows that
t is 1/t-decomposable and
the Lorentz space Λϕ generated by the function ϕ(t) = t log e
therefore the Banach couple (Λ(ϕ), Λ(ϕ)( 1
t )) is K-monotone.
Theorem 4.4. Suppose that ϕ is an increasing concave function on [0, 1] such that γϕ > 0
and 1 ≤ p < ∞. The following conditions are equivalent:
(a) there exists a weight w on [0, 1] such that the Lorentz space Λp,ϕ is w-decomposable;
(b) ϕ is equivalent to a regularly varying function at zero of order p.
Proof. First, if X := Λp,ϕ is w-decomposable for some weight w on [0, 1], then, by Corol-
lary 3.7, as in the proof of the previous theorem, we conclude that ϕ is equivalent to a
regularly varying function at zero of order p.
Conversely, suppose that ϕ is equivalent to a function that varies regularly at zero with
order p, that is, ϕ satisfies (3) for some τ (N) (N = 1, 2, . . . ). Consider a family (MN )∞N =1
of pairwise disjoint measurable subsets of [0, 1] such that m(M2) = min(τ (2), 1/4),
m(MN ) = min(τ (N),
m(MN−1)
2
), N > 2,
and let M1 := [0, 1] \S∞N =2 MN . Set w(t) := 2N for all t ∈ MN and N ∈ N. Clearly,
m(MN +1) ≤ m(MN )/2 (N ∈ N). Therefore, by Theorem 3.8, ϕ satisfies (13) for any
sequence (τN )∞N =1 majorized by the sequence (m(MN ))∞N =1. To complete the proof it
remains to apply Theorem 4.3.
16
It is obvious that Lp-spaces (1 ≤ p ≤ ∞) are w-decomposable for every weight w.
On the other hand, we show that for an arbitrary weight w there exist w-decomposable
Lorentz spaces Λϕ different from L1.
Theorem 4.5. Let w be an arbitrary weight on [0, 1]. Then there exists an increasing
concave function ϕ such that the space Λϕ is w-decomposable and Λϕ 6= L1.
Proof. As above, Mk = {t ∈ [0, 1] : w(t) ∈ [2k, 2k+1)} for k ∈ Z and (wr)∞r=1 is the
non-increasing rearrangement of the sequence (m(Mk))+∞k=−∞
. Define
G(α) :=
min{α, wr}, α ≥ 0.
∞
Xr=1
it follows that
Then G(1) = 1, G(0) = 0 and G is increasing and continuous at zero.
Let (tk)∞k=0 be a sequence from (0, 1] such that t0 = 1, 0 < tk < tk−1/3 for k ≥ 1 and
(27)
G(tk+1) ≤ 2−k tk, k = 0, 1, . . . .
Then we set ϕ′k(t) = maxi=0,1,...,k{2iχ[0,ti](t)}, k = 0, 1, . . . and ϕ′(t) = limk→∞ ϕ′k(t) (0 <
t ≤ 1). It is easy to see that ϕ′k and ϕ′ are non-increasing functions on(0, 1]. Moreover,
since
tkϕ′(tk) = tk 2k ≤
Z 1
Xk=0
∞
2
3
tk−1 2k−1 =
2
3
tk−1ϕ′(tk−1)
3(cid:19)k
Xk=0(cid:18)2
< ∞.
∞
0
ϕ′(tk)tk ≤
ϕ′(t)dt ≤
Therefore, the function ϕ(t) :=R t
In view of Theorem 4.3, it suffices to show that for some constant C ≥ 1 and for any
sequence of reals (dr)∞r=1 such that 0 < dr ≤ wr (r = 1, 2, . . . ) we have
0 ϕ′(s) ds is well-defined, increasing and concave on (0, 1].
We shall prove that the Lorentz space Λϕ is w-decomposable.
∞
Xr=1
dr! .
dr! ≤
ϕ(dr) ≤ Cϕ ∞
Xr=1
Note that the left hand side of this inequality is an immediate consequence of the concavity
ϕ ∞
Xr=1
of ϕ. Further, since ϕk(t) := R t
0 ϕ′k(s) ds ↑ ϕ(t), then limk→∞P∞r=1 ϕk(dr) = P∞r=1 ϕ(dr).
P∞r=1 ϕk(dr)
ϕk (P∞r=1 dr) ≤ 3, k ≥ 0.
Therefore, it is enough to prove that
(28)
Noting that P∞r=1 dr ≤ t0 = 1, we set
k0 := maxnk = 0, 1, 2,· · · :
∞
Xr=1
dr ≤ tko.
17
From the definition of ϕk it follows that
ϕk ∞
Xr=1
dr! = 2k
∞
Xr=1
dr =
∞
Xr=1
ϕk(dr) if 0 ≤ k ≤ k0.
Since tk0+1 <
∞
Pr=1
Let k > k0 be arbitrary. The inequality
dr > tk implies that
dr ≤ tk0, then, again by the definition of ϕk,
dr ≤ 2ϕk0+1 ∞
Xr=1
ϕk0+1(dr) ≤ 2k0+1
Xr=1
∞
∞
dr! .
∞
Xr=1
Pr=1
dr! > ϕk(tk) = 2ktk.
ϕk ∞
Xr=1
Moreover, since
we obtain
ϕk+1(dr) =(2k+1dr = 2 ϕk(dr),
2ktk+1 + ϕk(dr),
if dr ≤ tk+1,
if dr > tk+1,
(29)
(30)
(31)
Hence, for any k > k0, by (31) and (27), we obtain
∞
∞
∞
ϕk(dr) =
Xr=1
ϕk+1(dr) −
Xr=1
Xr=1
P∞r=1 ϕk+1(dr)
ϕk+1 (P∞r=1 dr) ≤ P∞r=1 ϕk(dr)
ϕk (P∞r=1 dr)
≤ P∞r=1 ϕk(dr)
ϕk (P∞r=1 dr)
min(2ktk+1, 2kdr) ≤ 2kG(tk+1).
+ P∞r=1 ϕk+1(dr) −P∞r=1 ϕk(dr)
G(tk+1)
+
tk
ϕk (P∞r=1 dr)
≤ P∞r=1 ϕk(dr)
ϕk (P∞r=1 dr)
+ 2−k.
Applying the last estimate together with (29) and (30), we obtain (28). It is easy to see
that ϕ(t) is not equivalent to t, and therefore Λϕ 6= L1. The proof is complete.
Remark 4.6. Theorem 4.5 can be easily extended to the spaces Λp,ψ with p ∈ (1,∞).
Indeed, let w be an arbitrary weight on [0, 1] and ϕ be the function from the proof of
Theorem 4.5. Set ψ := ϕ1/p. Clearly, ψ is an increasing concave function not equivalent
to the function t1/p. Therefore, Λp,ψ 6= Lp. Since relation (13) is fulfilled for ψ as well,
then, by Theorem 4.3, the space Λp,ψ is w-decomposable.
Our next goal is to prove analogous results for Marcinkiewicz spaces Mϕ. To make use
of the duality of Lorentz and Marcinkiewicz spaces we will need the following statement
which is of interest in its own right.
18
Theorem 4.7. Let X be a Banach lattice on a σ-finite measure space (Ω, Σ, µ) with
suppX = Ω which has the Fatou property and w be a non-trivial weight on Ω. Then the
couple (X, X(w)) is K-monotone if and only if (X′, X′(w)) is K-monotone, where X′ is
the Kothe dual of X.
The proof follows from Theorem 3.1 proved in [37] and the following result.
Theorem 4.8. Let X be a Banach lattice on a σ-finite measure space (Ω, Σ, µ) with
suppX = Ω which has the Fatou property and w be a non-trivial weight on Ω. Then X is
w-decomposable if and only if its Kothe dual X′ is w-decomposable.
Proof. Suppose that X is w-decomposable. Let n ∈ N and the functions x′1, x′2, . . . , x′n,
y′1, y′2, . . . , y′n ∈ X′ satisfy (5) (with the norm from X′) and (6). Take a function x ∈ X,
kxkX = 1, such that supp x ⊂
supp x′i and
n
n
Si=1
x′ikX ′ ≤ 2ZΩ
Xi=1
k
n
Xi=1
x′i(t)x(t) dµ.
Now, consider yi ∈ X such that supp yi ⊂ supp y′i, kyikX = kxχsupp x′
kyikX ZΩ y′i(t)yi(t) dµ, 1 ≤ i ≤ n.
ky′ikX ′ ≤
2
ikX and
Then, according to the hypothesis,
k
n
Xi=1
yikX ≤ Ck
n
Xi=1
xχsupp x′
ikX = C,
and, therefore,
k
n
Xi=1
y′ikX ′ ≥
≥
≥
y′i(t)yi(t) dµ
y′j(t) dµ =
n
n
1
n
yi(t)
1
2C
C ZΩ
Xi=1
Xj=1
Xi=1
ky′ikX ′kyikX =
Xi=1 ZΩ x′i(t)x(t)χsupp x′
1
2C
1
2C
n
i
n
n
1
C
Xi=1 ZΩ
kx′ikX ′kxχsupp x′
Xi=1
1
4Ck
Xi=1
(t) dµ ≥
n
ikX
x′ikX ′.
Certainly, the same argument can be applied to get the opposite estimate. The proof is
complete.
Since M′ϕ = Λ ϕ (cf. [21], p. 117) and δϕ + γ ϕ = 1 for any increasing concave function
ϕ on [0, 1] (cf. [21], Theorem 4.12 on page 107 or [24], p. 28), then by Theorems 4.3, 4.4
and 4.8 we immediately obtain the following statements.
19
Corollary 4.9. Let ϕ be an increasing concave function on [0, 1] such that δϕ < 1 and let
w be a weight on [0, 1]. Then the Marcinkiewicz space Mϕ is w-decomposable if and only
if ϕ(t) = t/ϕ(t) satisfies (13) with p = 1.
Corollary 4.10. If ϕ is an increasing concave function on [0, 1] such that δϕ < 1, then
the space Mϕ is w-decomposable for some weight w on [0, 1] if and only if ϕ is equivalent
to a regularly varying function at zero of order ∞.
In the paper [18], Kalton proved that if X and Y are symmetric sequence spaces with
the Fatou property such that the couple (X, Y (w)) is K-monotone for some non-trivial
weight w, then X = lp and Y = lq with 1 ≤ p, q ≤ ∞. The results in this section and
Theorem 3.1 show that in the case of symmetric function spaces on [0, 1] the situation
is completely different. The following theorems present new examples of K-monotone
Banach couples of weighted Lorentz and Marcinkiewicz function spaces. The first theorem
follows from Theorem 3.1, Theorem 4.4, Theorem 4.5 and Remark 2 and the second one
from Theorem 3.1, Theorem 4.8 on the duality and Corollary 4.10.
Theorem 4.11. If ϕ is an increasing concave function on [0, 1] such that γϕ > 0 and
1 ≤ p < ∞, then the weighted couple (Λp,ϕ, Λp,ϕ(w)) is K-monotone for some (non-trivial)
weight w on [0, 1] if and only if ϕ is equivalent to a regularly varying function at zero of
order p. On the other hand, for arbitrary weight w on [0, 1] and 1 ≤ p < ∞ there exists an
increasing concave function ϕ on [0, 1] such that the couple (Λp,ϕ, Λp,ϕ(w)) is K-monotone
and Λp,ϕ 6= Lp.
Theorem 4.12. If ϕ is an increasing concave function on [0, 1] such that δϕ < 1, then
the weighted couple (Mϕ, Mϕ(w)) is K-monotone for some (non-trivial) weight w on [0, 1]
if and only if ϕ is equivalent to a regularly varying function at zero of order ∞.
5 w–decomposable Orlicz spaces
As we have seen in the previous section, in order to check the property of w-decomposa-
bility for Lorentz spaces, it is enough to consider only characteristic functions (Theorem
4.3).
In this section we will prove that in the case of Orlicz spaces it is sufficient to
examine scalar multiples of characteristic functions.
As above, for a weight w on [0, 1] let Mk := {t ∈ [0, 1] : w(t) ∈ [2k, 2k+1)} (k ∈
and { ¯Mr}∞r=1
Z), (wr)∞r=1 be the non-increasing rearrangement of the sequence (m(Mk))+∞k=−∞
denote any rearrangement of the sets Mk such that m( ¯Mr) = wr, r = 1, 2, . . .
Theorem 5.1. Let an Orlicz function F satisfy the ∆2-condition for large u and let w
be a weight on [0, 1]. Then, the Orlicz space LF = LF [0, 1] is w-decomposable if and only
if there exists p ∈ [1,∞) such that for any n ∈ N, all measurable sets Ak ⊂ M k and reals
ck (1 ≤ k ≤ n) we have
n
with a constant independent of ck, Ak (1 ≤ k ≤ n) and n ∈ N. If, in addition, the com-
plementary function F ∗ satisfies the ∆2-condition for large u, then the w-decomposability
of LF implies that F is equivalent to a regularly varying Orlicz function at ∞ of order p.
ckχAkkp
LF ≈
kckχAkkp
LF
(32)
k
Xk=1
n
Xk=1
20
Proof. Suppose, first, that LF is w-decomposable. By Proposition 3.2, there is p ∈ [1,∞]
such that (9) holds for X = LF , which implies (32). Since F satisfies the ∆2-condition
for large u > 0, then αX > 0. Therefore, by Corollary 3.7, p < ∞.
Conversely, let n ∈ N and yk ∈ LF , supp yk ⊂ ¯Mk, 1 ≤ k ≤ n. We may (and will)
assume that yk are positive bounded functions and
Taking into account Theorem 3.3, we need to show that
n
Xk=1
kykkp
LF
= 1.
k
n
Xk=1
ykkp
LF ≈ 1,
with a constant independent from n and yk. For each 1 ≤ k ≤ n we set
(33)
(34)
and
ck =
kykkLF
2ϕLF (m(supp yk))
yk(t) :=(yk(t),
0,
if yk(t) ≥ ck,
if yk(t) < ck.
Applying (32) to the functions ckχsuppyk and taking into account the definition of ck and
(33) we get
k
n
Xk=1
ckχsuppykkp
LF ≤ C1
= C1
n
n
Xk=1
Xk=1
cp
kϕLF (m(suppyk))p
2−p kykkp
LF
= 2−pC1.
Up to equivalence of norms the Orlicz space LF = LF [0, 1] depends only on the behaviour
of F for large enough u > 0. Therefore, we may assume that F (2u) ≤ C2F (u) for all
u > 0. Then, from the last inequality it follows that
n
Xk=1
m(suppyk)F (ck) ≤ C3,
(35)
where C3 is a constant independent of n and yk. Moreover, from the definition of ck and
yk we have
kykkLF ≤ kykkLF and kykkLF ≥ kykkLF − kckχsupp ykkLF =
1
2kykkLF .
Next, let us show that there is rk ∈ [ck, supt yk(t)] such that
F (rk) = F (cid:18) rk
kykkLF(cid:19)Z 1
0
F (yk(t))dt.
21
(36)
(37)
In fact, consider the function
Hk(t) :=
From the equality R 1
0 F (
yk(t)
kyk(t)kLF
F (yk(t))
F (cid:16) yk(t)
kyk(t)kLF(cid:17)
, t ∈ supp yk.
)dt = 1 it follows that
Hk(t) ≤Z 1
0
inf
F [yk(t)]dt ≤ sup
t∈supp yk
Hk(t).
t∈supp yk
yk(t) ≥ ck, by the continuity of F , equality (37) holds for some rk
Thus, since
inf
t∈supp yk
from the interval [ck, supt yk(t)].
Next, define dk ∈ [0, 1] (k = 1, 2, . . . , n) as follows:
dk =(ϕ−1
LF (cid:16)kykkLF
rk (cid:17) ,
m(supp yk),
if kykkLF ≤ rkϕLF (m(supp yk)),
if kykkLF > rkϕLF (m(supp yk)).
Clearly, by the definition of dk,
rkϕLF (dk) ≤ kykkLF .
On the other hand, since rk ≥ ck, we obtain
rkϕLF (dk) ≥
1
2kykkLF ,
(38)
(39)
LF
whence dk ≥ ϕ−1
(kykkLF /(2rk)). Hence, taking into account that F satisfies the ∆2-
condition with constant C2 for all u > 0, the formula ϕLF (t) = 1/F −1(1/t) (see formula
(9.23) in [20] on page 79 of the English version or Corollary 5 in [25] on page 58) and
(37), we have
dkF (rk) ≥
F (rk)
kykkLF(cid:17) ≥
F (cid:16) 2rk
1
C2
F (rk)
M(cid:16) rk
kykkLF(cid:17)
=
1
C2
1
Z0
F [yk(t)] dt.
(40)
Conversely, from the equality 1/dk = F (1/ϕLF (dk)), (38) and (37) it follows that
dkF (rk) =
F (rk)
ϕLF (dk)) ≤
1
F (
F (rk)
F (cid:16) rk
kykkLF(cid:17)
=
1
Z0
F [yk(t)] dt.
(41)
Now, by the definition of dk, we have dk ≤ m(supp yk). Therefore, we can define the
scalar multiples of characteristic functions fk(t) := rkχBk(t), where Bk ⊂ supp yk and
m(Bk) = dk. According to (38), (39) and (36), we have
1
4kykkLF ≤ kfkkLF ≤ kykkLF , k = 1, 2, . . . , n.
22
Therefore, in view of (32) and (33), we obtain
k
n
Xk=1
fkkp
LF ≈
n
Xk=1
kfkkp
LF ≈
n
Xk=1
kykkp
LF
= 1,
with constants which depend only on p. Hence, taking into account that F satisfies the
∆2-condition, we conclude that (34) will be proved once we show that
k
n
Xk=1
ykkLF ≈ k
n
Xk=1
fkkLF
with constants independent of n and yk. Since the functions fk (respectively, yk) are
pairwise disjoint, in view of estimate (41), we find that
Z 1
0
n
F [
Xk=1
n
fk(t)] dt =
Xk=1
≤ Z 1
0
Conversely, by (40) and (35), we get
n
Xk=1Z 1
0
F (yk(t)) dt
n
dkF (rk) ≤
Xk=1
F [
yk(t)] dt.
Z 1
0
F [
n
Xk=1
yk(t)] dt ≤
F [yk(t)] dt +
m(supp yk)F (ck)
n
Xk=1
n
0
Xk=1Z 1
≤ C2Z 1
0
n
Xk=1
F [
fk(t)] dt + C3,
and we come to the desired result.
In order to obtain the second assertion of the theorem it is sufficient to apply Corollary
3.6, Lemmas 2.1 and 2.2, Proposition 2.3 and the elementary observation that condition
(a) in that proposition implies the equivalence of F to an Orlicz function which is regularly
varying at ∞ of order p.
Remark 5.2. Arguing in the same way as in the proof of Theorem 5.1 we may obtain
the following result: Let an Orlicz function F satisfy the ∆2-condition for large u and
1 < p, q < ∞. The Orlicz space LF [0, 1] satisfies the upper p-estimate, respectively the
lower q-estimate, if and only if there exsists a constant C > 0 such that for any n ∈ N,
all pairwise disjoint measurable sets Ak and reals ck we have
respectively,
k
n
Xk=1
ckχAkkLF ≤ C(
n
Xk=1
kckχAkkp
LF
)1/p,
n
(
Xk=1
kckχAkkq
LF
)1/q ≤ Ck
n
Xk=1
ckχAkkLF .
23
However, an inspection of the proof of results from [19] (pages 120-121 and 124) shows
that the first of these inequalities is equivalent to either of the following conditions: the
Orlicz space LF [0, 1] is p-convex or LF [0, 1] satisfies the upper p-estimate or there exists
an Orlicz function F1 equivalent to F for large arguments such that F1(u1/p) is a convex
function on [0,∞). At the same time, the second of them is equivalent to either of the
following conditions: the Orlicz space LF [0, 1] is q-concave or LF [0, 1] satisfies the lower
q-estimate or there exists an Orlicz function F1 equivalent to F for large arguments such
that F1(u1/q) is a concave function on [0,∞).
The following result is analogous to Theorem 4.4 for Lorentz spaces.
Theorem 5.3. Let F be an Orlicz function equivalent to an Orlicz function which is
regularly varying at ∞ of order p ∈ [1,∞). Then there is a weight w on [0, 1] such
that the Orlicz space LF is w-decomposable and, consequently, the couple (LF , LF (w)) is
K-monotone.
Proof. By Corollary 3.5, it is sufficient to find a sequence of pairwise disjoint intervals
{∆k}∞k=1 from [0, 1] such that for any n ∈ N and x1, x2, . . . , xn ∈ X satisfying the condition
supp xi ⊂ ∆i (1 ≤ i ≤ n), relation (9) holds.
First, since F is equivalent to a regularly varying Orlicz function at ∞ of order p, then
Lemma 1 and a simple compactness argument (see also [18, Lemma 6.1]) show that there
exists a constant C1 > 1 such that for every k ∈ N there is vk > 0 such that for all v ≥ vk
and u ∈ [k−2/8, 1] we have that
F (uv)
C1≈ upF (v).
(42)
Let v > 0, ε > 0 be arbitrary and ∆ be an interval from [0, 1] such that m(∆) ≤
ε/F (v). Moreover, suppose that z ∈ LF , z ≥ 0 and supp z ⊂ ∆. Then
Z{t∈∆: z(t)≤v}
F [z(t)] dt ≤ F (v) m(∆) ≤ ε.
Let {∆k}∞k=1 be a sequence of disjoint intervals from [0, 1] such that
m(∆k) ≤ 2−k−1(F (vk))−1 (k = 1, 2, . . . ).
Then, as it was noted above, for every z ∈ LF such that z ≥ 0 and supp z ⊂ ∆k, we have
(43)
F [z(t)] dt ≤ 2−k−1 (k = 1, 2, . . . ).
Z{t∈∆k: z(t)≤vk}
Suppose that {xk}∞k=1 is an arbitrary sequence from LF such that xk ≥ 0 and supp xk ⊂ ∆k
(k = 1, 2, . . . ). To prove (9) we assume that
or, equivalently,
= 1,
xi(cid:13)(cid:13)(cid:13)LF
n
(cid:13)(cid:13)(cid:13)
Xi=1
Xi=1 Z∆i
n
F [xi(t)] dt = 1.
(44)
24
If λi := kxikLF (i = 1, 2, . . . ), then 0 ≤ λi ≤ 1 and
Z∆i
Fhxi(t)
λi i dt = 1 (i = 1, 2, . . . ).
Denote I1 := {i = 1, 2, . . . , n : λi ≤ i−2/8}, I2 := {1, 2, . . . , n} \ I1. Then
λp
i ≤
Xi∈I1
1
8Xi∈I1
i−2p ≤
1
4
.
Now, let i ∈ I2, i.e., λi ≥ i−2/8. Then, if xi(t) ≥ λivi, from (42) it follows that
C−1
1 λp
i Fhxi(t)
λi i ≤ F [xi(t)] ≤ C1λp
i Fh xi(t)
λi i.
Moreover, by (43) and (45), we have
(45)
(46)
(47)
Z{t∈∆i: xi(t)>λivi}
Fh xi(t)
λi i dt = 1 −Z{t∈∆i: xi(t)≤λivi}
Fhxi(t)
λi i dt
≥ 1 − 2−i−1 ≥
3
4
,
whence, taking into account the left hand side of (47), we obtain
Z∆i
F [xi(t)] dt ≥ C−1
1 λp
i Z{t∈∆i: xi(t)>λivi}
Fhxi(t)
λi i dt ≥
3
4
C−1
1 λp
i ,
i ∈ I2.
Combining this with (44) and (46), we get
n
Xi=1
λp
i =Xi∈I1
λp
i +Xi∈I2
λp
i ≤
1
4
+
4
3
C1
n
Xi=1 Z∆i
F [xi(t)] dt ≤ 2C1,
and the first inequality in (9) is proved.
On the other hand, using the right hand side of (47) and (45), we infer that
Xi∈I2Z{t∈∆i: xi(t)>λivi}
F [xi(t)] dt ≤ C1Xi∈I2
Xi=1
≤ C1
n
λp
i Z{t∈∆i: xi(t)>λivi}
Fh xi(t)
λi i dt
(48)
λp
i .
At the same time, by (43) and the convexity of F , we obtain
Xi∈I2Z{t∈∆i: xi(t)≤λivi}
λiZ{t∈∆i: xi(t)≤λivi}
Fhxi(t)
λi i dt
2−i−1 =
1
2
F [xi(t)] dt ≤ Xi∈I2
Xi=1
≤
∞
25
and, by (45) and the definition of I1,
Hence, taking into account (44), we get
Xi∈I1Z∆i
Xi∈I2Z{t∈∆i: xi(t)>λivi}
1
4
.
λiZ∆i
Fh xi(t)
λi i dt ≤
F [xi(t)] dt ≤Xi∈I1
F [xi(t)] dt = 1 −Xi∈I2Z{t∈∆i: xi(t)≤λivi}
F [xi(t)] dt ≥
F [xi(t)] dt
− Xi∈I1Z∆i
i ≥ 1/(4C1), and so the proof of (9) is complete.
1
4
.
From this and (48) it follows thatPn
i=1 λp
6 Ultrasymmetric Orlicz spaces and w–decomposa-
bility
In the previous sections we have examined the problem of the K-monotonicity of
weighted couples generated by Lorentz, Marcinkiewicz and Orlicz spaces. We have seen
that the central role in the question is played by the notion of w-decomposibility. It turns
out that studying the last property in a natural way leads to the so-called ultrasymmetric
Orlicz spaces.
Recall that a symmetric space X on [0, 1] is ultrasymmetric if X is an interpolation
space between the Lorentz space ΛϕX and the Marcinkiewicz space MϕX . These spaces
were studied by Pustylnik [27], who proved that they embrace all possible generalizations
of Lorentz-Zygmund spaces and have a simple analytical description. Moreover, one could
substitute ultrasymmetric spaces into almost all results concerning classical spaces such
as Lorentz-Zygmund spaces, and so they are very useful in many applications (see, for
example, Pustylnik [28] and [29]).
Pustylnik asked about a description of ultrasymmetric Orlicz spaces (see [27], p. 172).
In the case of reflexive Orlicz spaces this problem was solved in [4]: such a space is
ultrasymmetric if and only if it coincides (up to equivalence of norms) with a Lorentz
space Λp,ϕ for some 1 < p < ∞ and some increasing concave function ϕ on [0, 1].
As it was said above, the class of w–decomposable symmetric spaces is closely related
to the class of ultrasymmetric Orlicz spaces. Our next theorem shows that in the case
when a weight w changes sufficiently fast any w–decomposable symmetric space with
non-trivial Boyd indices is an ultrasymmetric Orlicz space.
Again, as above, for a weight w defined on [0, 1], let Mk := {t ∈ [0, 1] : w(t) ∈
[2k, 2k+1)} (k ∈ Z) and (wk)∞k=1 be the non-increasing rearrangement of the sequence
(m(Mk))+∞k=−∞
Theorem 6.1. Let X be a symmetric space on [0, 1] with non-trivial Boyd indices and w
be a weight on [0, 1] satisfying the condition:
there are k0 ∈ N and c0 > 0 such that wk2k ≥ c0 for k ≥ k0.
(49)
26
(a) If X is w–decomposable, then X is an ultrasymmetric Orlicz space.
(b) If X has the Fatou property and (X, X(w)) is a K-monotone couple, then X is an
ultrasymmetric Orlicz space.
Proof. (a) Firstly, taking into account the boundedness of the dilation operator and The-
orem 3.3, a symmetric space X is w–decomposable if and only if it is v–decomposable,
where v(u) = w(cu) for some c > 0. Therefore, we may assume that c0 = 1. Denote
Ik := [2−k, 2−k+1), ¯χIk := χIk/ϕ(2−k) (k = 1, 2, . . . ), where ϕ is the fundamental function
of X. From (49) it follows that m(supp ¯χIk) ≤ wk for all k ≥ k0. Applying Corollary 3.4
to scalar multiples of ¯χIk (k ≥ k0), we get that ( ¯χIk)∞k=k0 spans lp for some p ∈ [1,∞)
(p 6= ∞ because the Boyd indices of X are non-trivial). Obviously, replacing ( ¯χIk)∞k=k0
with ( ¯χIk)∞k=1 does not change this property, so for all ak ∈ R (k = 1, 2, . . . )
Then, taking into account [4, Proposition 2], we get
∞
Xk=1
(cid:13)(cid:13)(cid:13)
ak ¯χIk(cid:13)(cid:13)(cid:13)X ≈ k(ak)klp.
X = (L1, L∞)K
lp((ϕ(2−k )2−k)∞
k=1).
By Corollary 3.7, δϕ = βX < 1. Therefore, limt→∞ kσtkX→X/t = 0, and we can apply [21,
Theorem II.6.6, p. 137] in the case when A is the identity operator, to obtain
kxkX ≈ k(ϕ(2−k) x∗∗(2−k))∞k=1klp ≈ k(ϕ(2−k) x∗(2−k))∞k=1klp
t (cid:19)1/p
≈ (cid:18)Z 1
[x∗(t)ϕ(t)]p dt
0
,
and we conclude that
Next, denote
F (u) =Z u
0
F (t)
t
X = Λp,ϕ.
(50)
dt, where F (t) =( t
1
ϕ−1(1)
ϕ−1( 1
t )
if 0 ≤ t ≤ 1,
if t ≥ 1.
Since F (t)/t is increasing on (0,∞), then F (u) is a convex function and for u > 0 we
have that
F (u/2) ≤Z u
u/2
F (t)
t
dt ≤ F (u) ≤ F (u).
Moreover, by Corollary 3.7, we have that γϕ = αX > 0, which implies that F satisfies the
∆2–condition for all u > 0. Therefore, for all u > 0
F (u) ≥ F (u/2) ≥ c F (u),
that is, the functions F and F are equivalent on (0,∞).
27
Now, we recall the following definition due to Kalton [18] (see also [4], where the
notion is used): For an Orlicz function F and 1 ≤ p < ∞, define the function Ψ∞F,p(u, C)
for 0 < u ≤ 1, C > 1 to be the supremum (possibly ∞) of all N such that there exist
ak−1 ≥ 2 for k = 2, . . . , N such that for all k either Fak (u) ≥ Cup
1 ≤ a1 < a2 < . . . < aN ,
or up ≥ CFak (u), where Fa(u) := F (au)
have that
To complete the proof it suffices to verify that for some C0 > 0, C1 > 0 and r > 0 we
for a, u > 0.
F (a)
ak
Ψ∞F,p(u, C0) ≤ C1u−r
for all u ∈ (0, 1].
Indeed, once it is done, we can apply Theorem 1 from [4] to conclude that the Orlicz
space LF is ultrasymmetric and that it coincides with a Lorentz space Λp,ψ generated by
some increasing concave function ψ. Since the fundamental function of LF is equivalent
to ϕ, then LF = Λp,ϕ, and, in view of (50), the proof is complete.
Since the functions F and F are equivalent, then, by [4, Lemma 1], it is sufficient to
prove the inequality for F , i.e., to prove that for some C0 > 0, C1 > 0 and r > 0 we have
Ψ∞F ,p(u, C0) ≤ C1u−r
for all u ∈ (0, 1].
(51)
In view of w-decomposability, Corollary 3.7, Lemma 2.2 and the inequality wk ≥ 2−k,
there is a constant C > 0 such that for any l = 1, 2, . . .
ϕ(lt)
ϕ(t)
C
≈ l1/p
if 0 < t ≤ 2−l.
Since 0 < αX ≤ βX < 1 it follows that 0 < γϕ ≤ δϕ < 1. Therefore, from the definition of
F it follows that both F and its complementary function satisfy the ∆2-condition. Hence,
by Proposition 2.3 and by the definition of F once more, we obtain that there exists a
constant C1 > 0 such that, for any l ∈ N and for all x ≥ F −1(2l), we have
1
C1l ≤
F (xl−1/p)
F (x) ≤
C1
l
.
By standard arguments, there are constants C2 > 0 and C3 > 0 such that
2 up ≤
C−1
F (ua)
F (a) ≤ C2up,
(52)
for all 0 < u ≤ 1 and any a satisfying F (a) ≥ C32u−p.
Suppose that 1 ≤ a1 < a2 < . . . < aN ,
ak
ak−1 ≥ 2 for k = 2, . . . , N such that for all k
either
F (uak)
F (ak) ≥ 2C2up or
F (uak)
F (ak) ≤
1
2C2
up.
Then, by (52), we have that F (aN ) ≤ C32u−p, which implies F (a12N−1) ≤ C32u−p. Hence,
N ≤ C4u−p, that is, Ψ∞F ,p(u, 2C2) ≤ C4u−p (0 < u ≤ 1), and (51) is proved.
(b) This part follows immediately from (a) and Theorem 3.1.
28
Using equality (50) from the proof of Theorem 6.1, we obtain the following corollary.
Corollary 6.2. Let X be a symmetric space on [0, 1] and w be a weight on [0, 1] satisfying
the condition (49). Assume that either X is w–decomposable or X has the Fatou property
and (X, X(w)) is K-monotone couple. If ϕX (t) = t1/p for some 1 < p < ∞, then X = Lp.
Remark 6.3. Using Krivine's theorem and the arguments from the beginning of the proof
of Theorem 6.1, the last assertion can be proved for p = 1 and p = ∞ as well.
Remark 6.4. It is well known that there is a regularly varying at ∞ Orlicz function
F such that the corresponding Orlicz space LF is not ultrasymmetric (see [18]). Thus,
Theorems 4.4 and 5.3 show that condition (49) on the weight w from Theorem 6.1 and
Corollary 6.2 is essential.
Remark 6.5. Conversely, if LF is an ultrasymmetric reflexive Orlicz space on [0, 1], then
there is a weight w on [0, 1] such that LF is w-decomposable and, equivalently, the Banach
couple (LF , LF (w)) is K-monotone. In fact, in that case F is regularly varying at ∞ of
order p ∈ (1,∞) (cf. [4]) and we can apply Theorem 5.3.
Examples. Theorem 5.3 guarantees that a weighted couple of Orlicz spaces (LF , LF (w))
on [0, 1] is K-monotone for some weight w on [0, 1] if F is equivalent to an Orlicz function
which is regularly varying at ∞ of order p ∈ [1,∞). We present some examples of such
Orlicz functions below.
1. The function F (u) = up(1 + ln u) for p ≥ (3 + √5)/2 is an Orlicz function on
(0,∞) which is regularly varying at ∞ of order p (cf. [24, Example 4]).
2. The function F (u) = up[1 + c sin(p ln u)] for 0 < c < 1/√2 and p ≥ (1 −
√2c√1−2c2 )−1
is an Orlicz function on (0,∞) which is not regularly varying but it is equivalent to up
and 1
4up ≤ F (u) ≤ 2up for all u > 0 (cf. [24, Example 10] and [25], Example 5 on p. 93
with c = 1/√5 and p ≥ 6).
3. Let an Orlicz function F be equivalent for large u to the function
F (u) = up(ln u)q1(ln ln u)q2 . . . (ln . . . ln u)qn,
where p ∈ (1,∞) and q1, . . . , qn are arbitrary real numbers. It is easy to see that F is
equivalent to a regularly varying function at ∞ of order p (in fact, the corresponding
Orlicz space LF is even ultrasymmetric [4]).
4. Some more examples of Orlicz functions that are equivalent to some regularly
varying functions at ∞ of order p are given by Kalton [18].
References
[1] E. V. Abakumov and A. A. Mekler, A concave regularly varying leader for equi–concave
functions, J. Math. Anal. Appl. 187 (1994), no. 3, 943–951.
[2] J. Arazy and M. Cwikel, A new characterization of the interpolation spaces between Lp and
Lq, Math. Scand. 55 (1984), no. 2, 253–270.
[3] S. V. Astashkin, On finite representability of lp–spaces in symmetric spaces, Algebra i
Analiz 23 (2011), no. 2, 77–101 (Russian); English transl. in St. Petersburg Math. J. 23
(2012), no. 2, 257–273.
29
[4] S. V. Astashkin and L. Maligranda, Ultrasymmetric Orlicz spaces, J. Math. Anal. Appl.
347 (2008), no. 2, 273–285.
[5] S. V. Astashkin and K. E. Tikhomirov, On stability of K-monotonicity of Banach couples,
Rev. Mat. Complut. 23 (2010), no. 1, 113–137.
[6] S. V. Astashkin and K. E. Tikhomirov, On stably K-monotone Banach couples, Funktsional.
Anal. i Prilozhen. 44 (2010), no. 3, 65–69 (Russian); English transl. in Funct. Anal. Appl.
44 (2010), no. 3, 212–215.
[7] C. Bennett and R. Sharpley, Interpolation of Operators, Academic Press, Boston 1988.
[8] D. W. Boyd, Indices for the Orlicz spaces, Pacific J. Math. 65 (1971), 325–335.
[9] Ju. A. Brudnyı and N. Ya. Krugljak, Interpolation Functors and Interpolation Spaces. I,
North-Holland, Amsterdam 1991.
[10] A. P. Calder´on, Spaces between L1 and L∞ and the theorem of Marcinkiewicz, Studia Math.
26 (1966), no. 3, 273–299.
[11] M. Cwikel, Monotonicity properties of interpolation spaces, Ark. Mat. 14 (1976), no. 2,
213–236.
[12] M. Cwikel, K-divisibility of the K-functional and Calder´on couples, Ark. Mat. 22 (1984),
no. 1, 39–62.
[13] M. Cwikel and I. Kozlov, Interpolation of weighted L1 spaces-a new proof of the Sedaev-
Semenov theorem, Illinois J. Math. 46 (2002), no. 2, 405–419.
[14] M. Cwikel and P. G. Nilsson, The coincidence of real and complex interpolation methods for
couples of weighted Banach lattices, in: Interpolation Spaces and Allied Topics in Analysis
(Lund, 1983), Lecture Notes in Math. 1070, Springer, Berlin 1984, 54–65.
[15] M. Cwikel and P. G. Nilsson, Interpolation of weighted Banach lattices, Mem. Amer. Math.
Soc. 165 (2003), 1–105.
[16] M. Cwikel, P. G. Nilsson and G. Schechtman, A characterization of relatively decomposable
Banach lattices, Mem. Amer. Math. Soc. 165 (2003), 106–127.
[17] V. I. Dmitriev, On the interpolation of operators in Lp spaces, Dokl. Akad. Nauk SSSR 260
(1981), no. 5, 1051–1054; English transl. in Soviet Math. Dokl. 24 (1981), no. 2, 373–376
(1982).
[18] N. J. Kalton, Calder´on couples of rearrangement invariant spaces, Studia Math. 106 (1993),
no. 3, 233–277.
[19] A. Kami´nska, L. Maligranda and L. E. Persson, Type, cotype and convexity properties
of Orlicz spaces, Publicaciones del Departamento de Analysis Matematico, Facultad de
Matematicas, Universidad Complutense de Madrid, No. 42, 1996-97, 113-126.
[20] M. A. Krasnoselskiı and Ja. B. Rutickiı, Convex Functions and Orlicz Spaces, Fizmatgiz,
Moscow 1958 (Russian); English transl. Noordhoff, Groningen 1961.
[21] S. G. Krein, Yu. I. Petunin and E. M. Semenov, Interpolation of Linear Operators, Amer.
Math. Soc., Providence 1982.
[22] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces. II. Function Spaces, Springer-
Verlag, Berlin-New York 1979.
[23] G. G. Lorentz, Relations between function spaces, Proc. Amer. Math. Soc. 12 (1961), 127–
132.
30
[24] L. Maligranda, Indices and interpolation, Dissertationes Math. (Rozprawy Mat.) 234
(1985), 1–49.
[25] L. Maligranda, Orlicz Spaces and Interpolation, Seminars in Mathematics 5, University of
Campinas, Campinas SP, Brazil 1989.
[26] B. S. Mitjagin, An interpolation theorem for modular spaces, Mat. Sb. (N.S.) 66 (1965),
473–482 (Russian); English transl. in Lecture Notes in Math. 1070, Springer, Berlin 1984,
10–23.
[27] E. Pustylnik, Ultrasymmetric spaces, J. London Math. Soc. (2) 68 (2003), no. 1, 165–182.
[28] E. Pustylnik, Sobolev type inequalities in ultrasymmetric spaces with applications to Orlicz-
Sobolev embeddings, J. Funct. Spaces Appl. 3 (2005), no. 2, 183–208.
[29] E. Pustylnik, Ultrasymmetric sequence spaces in approximation theory, Collect. Math. 57
(2006), no. 3, 257–277.
[30] Y. Raynaud, On Lorentz-Sharpley spaces, Israel Math. Conf. Proc., Vol. 5, Bar-Ilan Univ.,
Ramat Gan 1992, 207–228.
[31] H. P. Rosenthal, On a theorem of J. L. Krivine concerning block finite representability of
lp in general Banach spaces, J. Funct. Anal. 28 (1978), no. 2, 197–225.
[32] A. A. Sedaev, A description of the interpolation spaces of the couple (Lp
a1), and certain
related questions, Dokl. Akad. Nauk SSSR 209 (1973), 798–800; English transl. in Soviet
Math. Dokl. 14 (1973), 538–541.
a0 , Lp
[33] A. A. Sedaev and E. M. Semenov, The possibility of describing interpolation spaces in terms
of Peetre's K-method, Optimizacija 4(21) (1971), 98–114 (Russian).
[34] R. Sharpley, Spaces Λα(X) and interpolation, J. Funct. Anal. 11 (1972), 479–513.
[35] G. Sparr, Interpolation des espaces Lp
w, C. R. Acad. Sci. Paris S´er. A 278 (1974), 491–492.
[36] G. Sparr, Interpolation of weighted Lp-spaces, Studia Math. 62 (1978), no. 3, 229–271.
[37] K. E. Tikhomirov, K–monotone weighted couples of Banach lattices, Sibirsk. Math. Zh. 52
(2011), no. 1, 187–200 (Russian); English transl. in Siberian Math. J. 52 (2011), no. 1,
147–158.
[38] K. E. Tikhomirov, K–monotone weighted couples of Banach lattices, Candidate of Sciences
Dissertation, Samara State University, Samara 2011, 115 pp. (Russian).
Sergey V. Astashkin and Konstantin E. Tikhomirov
Department of Mathematics and Mechanics, Samara State University
Acad. Pavlova 1, 443011 Samara, Russia
E-mail addresses: [email protected], [email protected]
Lech Maligranda
Department of Engineering Sciences and Mathematics
Lulea University of Technology
SE-971 87 Lulea, Sweden
E-mail address: [email protected]
31
|
1807.03061 | 1 | 1807 | 2018-07-09T12:02:27 | On the norm-continuity for evolution family arising from non-autonomous forms | [
"math.FA"
] | We consider evolution equations of the form \begin{equation*}\label{Abstract equation} \dot u(t)+ A(t)u(t)=0,\ \ t\in[0,T],\ \ u(0)=u_0, \end{equation*} where $A(t),\ t\in [0,T],$ are associated with a non-autonomous sesquilinear form $\mathfrak a(t,\cdot,\cdot)$ on a Hilbert space $H$ with constant domain $V\subset H.$ In this note we continue the study of fundamental operator theoretical properties of the solutions. We give a sufficient condition for norm-continuity of evolution families on each spaces $V, H$ and on the dual space $V'$ of $V.$ The abstract results are applied to a class of equations governed by time dependent Robin boundary conditions on exterior domains and by Schr\"odinger operator with time dependent potentials. | math.FA | math |
ON THE NORM-CONTINUITY FOR EVOLUTION FAMILY ARISING FROM
NON-AUTONOMOUS FORMS∗
OMAR EL-MENNAOUI AND HAFIDA LAASRI
Abstract. We consider evolution equations of the form
u(t) + A(t)u(t) = 0,
t ∈ [0, T ], u(0) = u0,
where A(t), t ∈ [0, T ], are associated with a non-autonomous sesquilinear form a(t, ·, ·) on a Hilbert
space H with constant domain V ⊂ H. In this note we continue the study of fundamental operator
theoretical properties of the solutions. We give a sufficient condition for norm-continuity of evolution
families on each spaces V, H and on the dual space V ′ of V. The abstract results are applied to a
class of equations governed by time dependent Robin boundary conditions on exterior domains and by
Schrödinger operator with time dependent potentials.
Introduction
Throughout this paper H, V are two separable Hilbert spaces over C such that V is densely and continu-
ously embedded into H (we write V ֒→
H). We denote by (· ·)V the scalar product and k · kV the norm
d
on V and by (· ·)H , k · kH the corresponding quantities in H. Let V ′ be the antidual of V and denote by
h., .i the duality between V ′ and V. As usual, by identifying H with H ′ we have V ֒→
V ′ see
d
H ∼= H ′ ֒→
d
e.g., [5].
Let a : [0, T ] × V × V → C be a non-autonomous sesquilinear form, i.e., a(t; ·, ·) is for each t ∈ [0, T ] a
sesquilinear form,
(1)
such that
a(·; u, v) is measurable for all u, v ∈ V,
(2)
a(t, u, v) ≤ M kukV kvkV and Re a(t, u, u) ≥ αkuk2
V
(t, s ∈ [0, T ], u, v ∈ V ),
for some constants M, α > 0 that are independent of t, u, v. Under these assumptions there exists for each
t ∈ [0, T ] an isomorphism A(t) : V → V ′ such that hA(t)u, vi = a(t, u, v) for all u, v ∈ V. It is well known
that −A(t), seen as unbounded operator with domain V, generates an analytic C0-semigroup on V ′. The
operator A(t) is usually called the operator associated with a(t, ·, ·) on V ′. Moreover, we associate an
operator A(t) with a(t; ·, ·) on H as follows
D(A(t)) = {u ∈ V ∃f ∈ H such that a(t; u, v) = (f v)H for all v ∈ V }
A(t)u = f.
It is not difficult to see that A(t) is the part of A(t) in H. In fact, we have D(A(t)) = {u ∈ V : A(t)u ∈ H}
and A(t)u = A(t)u. Furthermore, −A(t) with domain D(A(t)) generates a holomorphic C0-semigroup on
H which is the restriction to H of that generated by −A(t). For all this results see e.g. [22, Chapter 2]
or [3, Lecture 7].
We now assume that there exist 0 < γ < 1 and a continuous function ω : [0, T ] −→ [0, +∞) such that
(3)
a(t, u, v) − a(s, u, v) ≤ ω(t − s)kukVγ kvkVγ
(t, s ∈ [0, T ], u, v ∈ V ),
∗THIS WORK IS SUPPORTED BY DEUTSCHE FORSCHUNGSGEMEINSCHAFT DFG (GRANT LA
4197/8-1)
1
2
with
(4)
OMAR EL-MENNAOUI AND HAFIDA LAASRI
sup
t∈[0,T ]
ω(t)
tγ/2 < ∞ and Z T
0
ω(t)
t1+γ/2 < ∞
where Vγ := [H, V ] is the complex interpolation space.
square root property for some t0 ∈ [0, T ] (and then for all t ∈ [0, T ] by [4, Proposition 2.5] ), i.e.,
In addition we assume that a(t0; ·, ·) has the
(5)
D(A
1
2 (t0)) = V.
Recall that for symmetric forms, i.e., if a(t; u, v) = a(t; v, u) for all t, u, v, then the square root property
is satisfied.
Under the assumptions (1)-(5) it is known that for each x0 ∈ V the non-autonomous homogeneous
Cauchy problem
(6)
( u(t) + A(t)u(t) = 0
u(0) = x0,
a.e. on [0, T ],
has a unique solution u ∈ MR (V, H) := L2(0, T ; V ) ∩ H 1(0, T ; H) such that u ∈ C([0, T ]; V ). This result
has been proved by Arendt and Monniaux [4] (see also [10]) when the form a satisfies the weaker condition
(7)
a(t, u, v) − a(s, u, v) ≤ ω(t − s)kukV kvkVγ
(t, s ∈ [0, T ], u, v ∈ V ).
In this paper we continue to investigate further regularity of the solution of (6). For this it is necessary
to associate to the Cauchy problem (6) an evolution family
which means that:
U :=nU (t, s) : 0 ≤ s ≤ t ≤ To ⊂ L(H)
(i) U (t, t) = I and U (t, s) = U (t, r)U (r, s) for every 0 ≤ r ≤ s ≤ t ≤ T,
(ii) for every x ∈ X the function (t, s) 7→ U (t, s)x is continuous into H for 0 ≤ s ≤ t ≤ T
(iii) for each x0 ∈ H, U (·, s)x0 is the unique solution of (6).
Definition 0.1. Let Y ⊆ H be a subspace. An evolution family U ⊂ L(H) is said to be norm continuous
in Y if U ⊂ L(Y ) and the map (t, s) 7→ U (t, s) is norm continuous with value in L(Y ) for 0 ≤ s < t ≤ T.
If the non-autonomous form a satisfies the weaker condition (7) then it is known that (6) is governed
by an evolution family which is norm continuous in V [15, Theorem 2.6], and norm continuous in H
if in addition V ֒→ H is compact [15, Theorem 3.4]. However, for many boundary value problem the
compactness embedding fails.
In this paper we prove that the compactness assumption can be omitted provided a satisfies (3) instead
of (7). This will allow us to consider a large class of examples of applications. One of the main ingredient
used here is the non-autonomous returned adjoint form a∗
r : [0, T ] × V × V −→ C defined by
(8)
a
∗
r(t, u, v) := a(T − t, v, u)
(t, s ∈ [0, T ], u, v ∈ V ).
The concept of returned adjoint forms appeared in the work of D. Daners [7] but for different interest.
Furthermore, [15, Theorem 2.6] cited above will be also needed to prove our main result.
We note that the study of regularity properties of the evolution family with respect to (t, s) in general
Banach spaces has been investigated in the case of constant domains by Komatsu [14] and Lunardi [17],
and by Acquistapace [1] for time-dependent domains.
We illustrate our abstract results by two relevant examples. The first one concerns the Laplacian with
non-autonomous Robin boundary conditions on a unbounded Lipschitz domain. The second one traits a
class of Schrödinger operators with time dependent potential.
ON THE NORM-CONTINUITY FOR EVOLUTION FAMILY ARISING FROM NON-AUTONOMOUS FORMS∗1
3
1. Preliminary results
Let a : [0, T ] × V × V → C a non-autonomous sesquilinear form satisfying (1) and (2). Then the
following well known result regarding L2-maximal regularity in V ′ is due to J. L. Lions
Theorem 1.1 (Lions, 1961). For each given s ∈ [0, T ) and x0 ∈ H the homogenuous Cauchy problems
(9)
( u(t) + A(t)u(t) = 0
u(s) = x,
a.e. on [s, T ],
has a unique solution u ∈ MR (V, V ′) := MR (s, T ; V, V ′) := L2(s, T ; V ) ∩ H 1(s, T ; V ′).
Recall that the maximal regularity space MR (V, V ′) is continuously embedded into C([s, T ], H) [21,
page 106]. A proof of Theorem 1.1 using a representation theorem of linear functionals, known in the
literature as Lions's representation Theorem can be found in [21, page 112] or [8, XVIII, Chpater 3, page
513].
Furthermore, we consider the non-autonomous adjoint form a∗ : [0, T ] × V × V −→ C of a defined by
∗(t; u, v) := a(t; v, u)
a
for all t ∈ [0, t] and u, v ∈ V. Finally, we will need to consider the returned adjoin form a∗
V −→ C given by
r : [0, T ] × V ×
a
∗
r(t, u, v) := a
∗(T − t, u, v).
Clearly, the adjoint form is a non-autonomous sesquilinear form and satisfies (1) and (2) with the same
constant M, α. Moreover, the adjoint operators A∗(t), t ∈ [0, T ] of A(t), t ∈ [0, T ] coincide with the
operators associated with a∗ on H. Thus applying Theorem 1.1 to the returned adjoint form we obtain
that the Cauchy problem associated with A∗
r(t) := A∗(T − t)
(10)
( v(t) + A∗
v(s) = x,
r(t)v(t) = 0
a.e. on [s, T ],
has for each x ∈ H a unique solution v ∈ M R(V, V ′). Accordingly, for every (t, s) ∈ ∆ := {(t, s) ∈ [0, T ]2 :
t ≤ s} and every x ∈ H we can define the following family of linear operators
(11)
U (t, s)x := u(t)
and U ∗
r (t, s)x := v(t),
(t, s) ∈ ∆} and {U ∗
where u and v are the unique solutions in M R(V, V ′) respectively of (9) and (10). Thus each family
(t, s) ∈ ∆} yields a contractive, strongly continuous evolution
{U(t, s) :
family on H [15, Proposition].
In the autonomous case, i.e., if a(t, ·, ·) = a(·, ·) for all t ∈ [0, T ], then one knows that −A0, the operator
associated with a0 in H, generates a C0-semigroup (T (t))t≥0 in H. In this case U (t, s) := T (t − s) yields
a strongly continuous evolution family on H. Moreover, we have
r (t, s) :
(12)
U (t, s)′ = T (t − s)′ = T ∗(t − s) = U ∗(t, s) = U ∗
r (t, s).
Here, T (·)′ denote the adjoint of T (·) which coincides with the C0-semigroup (T ∗(t))t≥0 associated with
the adjoint form a∗. In the non-autonomous setting however, (12) fails in general even in the finite
dimensional case, see [7, Remark 2.7]. Nevertheless, Proposition 1.2 below shows that the evolution
families U and U ∗
r can be related in a similar way. This formula appeared in [7, Theorem 2.6].
Proposition 1.2. Let U and U ∗
r be given by (11). Then we have
(13)
for all x ∈ H and (t, s) ∈ ∆.
x = U (T − s, T − t)x
r (t, s)(cid:3)′
(cid:2)U ∗
The equality (13) will play a crucial role in the proof of our main result. We include here a new proof
for the sake of completeness.
4
OMAR EL-MENNAOUI AND HAFIDA LAASRI
Proof. (of Proposition 1.2) Let Λ = (0 = λ0 < λ1 < ... < λn+1 = T ) be a subdivision of [0, T ]. Let
ak : V × V → C for k = 0, 1, ..., n be given by
ak(u, v) := ak,Λ(u, v) :=
a(r; u, v)dr for u, v ∈ V.
1
λk+1 − λk Z λk+1
λk
All these forms satisfy (2) with the same constants α, M. The associated operators in V ′ are denoted by
Ak ∈ L(V, V ′) and are given for all u ∈ V and k = 0, 1, ..., n by
(14)
Aku := Ak,Λ :=
1
λk+1 − λk Z λk+1
λk
A(r)udr.
Consider the non-autonomous form aΛ : [0, T ] × V × V → C defined by
(15)
aΛ(t; ·, ·) :=(ak(·, ·)
an(·, ·)
if t ∈ [λk, λk+1)
if t = T .
Its associated time dependent operator AΛ(·) : [0, T ] ⊂ L(V, V ′) is given by
(16)
AΛ(t) :=(Ak
An
if t ∈ [λk, λk+1)
if t = T .
Next denote by Tk the C0−semigroup associated with ak in H for all k = 0, 1...n. Then applying Theorem
1.1) to the form aΛ we obtain that in this case the associated evolution family UΛ(t, s) is given explicitly
for λm−1 ≤ s < λm < ... < λl−1 ≤ t < λl by
(17)
UΛ(t, s) := Tl−1(t − λl−1)Tl−2(λl−1 − λl−2)...Tm(λm+1 − λm)Tm−1(λm − s),
and for λl−1 ≤ a ≤ b < λl by
(18)
UΛ(t, s) := Tl−1(t − s).
By [20, Theorem 3.2] we know that (UΛ)Λ converges weakly in M R(V, V ′) as Λ → 0 and
lim
Λ→0
kUΛ − U kMR(V,V ′) = 0
The continuous embedding of M R(V, V ′) into C([0, T ]; H) implies that lim
Λ→0
UΛ = U in the weak operator
topology of L(H).
Now, let (t, s) ∈ ∆ with λm−1 ≤ s < λm < ... < λl−1 ≤ t < λl be fixed. Applying the above
approximation argument to a∗
r one obtains that
(19)
(20)
U ∗
Λ,r(t, s) = T ∗
= T ′
l−1,r(t − λl−1)T ∗
l−1,r(t − λl−1)T ′
l−2,r(λl−1 − λl−2)...T ∗
l−2,r(λl−1 − λl−2)...T ′
m,r(λm+1 − λm)T ∗
m,r(λm+1 − λm)T ′
m−1,r(λm − s)
m−1,r(λm − s)
where Tk,r and T ∗
k,r are the C0-semigroups associated with
(21)
ak,r(u, v) :=
1
λk+1 − λk Z λk+1
λk
a(T − r; u, v)dr =
1
λk+1 − λk Z T −λk
T −λk+1
a(r; u, v)dr
k,r, respectively. Recall that T ∗
and its adjoint a∗
On the other hand, the last equality in (21) implies that Tk,r coincides with the semigroup associated
with ak,ΛT where ΛT is the subdivision ΛT := (0 = T − λn+1 < T − λn < ... < T − λ1 < T − λ0 = T ). It
follows from (17)-(18) and (19)-(20) that
k,r = T ′
k,r.
Λ,r(t, s)i′
hU ∗
= Tm−1,r(λm − s)Tm,r(λm+1 − λm)...Tl−2,r(λl−1 − λl−2)Tl−1,r(t − λl−1)
= Tm−1(cid:16)(T − s) − (T − λm)(cid:17)Tm(cid:16)(T − λm) − (T − λm+1)(cid:17)...Tl−1(cid:16)(T − λl−1) − (T − t)(cid:17)
= UΛT (T − s, T − t)
Finally, the desired equality (13) follows by passing to the limit as Λ = ΛT → 0.
(cid:3)
ON THE NORM-CONTINUITY FOR EVOLUTION FAMILY ARISING FROM NON-AUTONOMOUS FORMS∗2
5
Remark 1.3. The coerciveness assumption in (2) may be replaced with
(22)
Re a(t, u, u) + ωkuk2
H ≥ αkuk2
V
(t ∈ [0, T ], u ∈ V )
for some ω ∈ R. In fact, a satisfies (22) if and only if the form aω given by aω(t; ·, ·) := a(t; ·, ·) + ω(· ·)
satisfies the second inequality in (2). Moreover, if u ∈ M R(V, V ′) and v := e−w.u, then v ∈ M R(V, V ′)
and u satisfies (9) if and only if v satisfies
v(t) + (ω + A(t))v(t) = 0 t−a.e. on [s, T ],
v(s) = x.
2. Norm continuous evolution family
In this section we assume that the non-autonomous form a satisfies (2)-(5). Thus as mentioned in the
introduction, under theses assumptions the Cauchy problem (9) has L2-maximal regularity in H. Thus
for each x ∈ V,
U (·, s)x ∈ MR (V, H) := MR (s, T ; V, H) := L2(s, T ; V ) ∩ H 1(s, T ; H).
Moreover, U (·, s)x ∈ C[s, T ]; V ) by [4, Theorem 4.2]. From [15, Theorem 2.7] we known that the re-
striction of U to V defines an evolution family which norm continuous. The same is also true for the
Cauchy problem (10) and the assocaited evolution family U ∗
r inherits
all properties of a.
In the following we establish that U can be extended to a strongly continuous evolution family on V ′.
r since the returned adjoint form a∗
Proposition 2.1. Let a be a non-autonomous sesquilinear form satisfying (2)-(5). Then U can be
extended to a strongly continuous evolution family on V ′, which we still denote U.
Proof. Let x ∈ H and (t, s) ∈ ∆. Then using Proposition 1.2 and the fact that U and U ∗
strongly continuous evolution families on V and H we obtain that
r define both
kU (t, s)xkV ′ = sup
< U (t, s)x, v >
kvkV =1
v∈V
= sup
(U (t, s)xv)H = sup
(xU (t, s)′v)H
kvkV =1
v∈V
= sup
kvkV =1
v∈V
= sup
kvkV =1
v∈V
kvkV =1
v∈V
(xU ∗
r (T − s, T − t)v)H
< x, U ∗
r (T − s, T − t)v >
≤ kxkV ′kU ∗
≤ ckxkV ′
r (T − s, T − t)kL(V )
where c > 0 is such that sup
t,s∈∆
kU ∗
r (t, s)kL(V ) ≤ c. Thus, the claim follows since H is dense in V ′.
(cid:3)
Let ∆ := {(t, s) ∈ ∆ t ≥ s}. The following theorem is the main result of this paper
Theorem 2.2. Let a be a non-autonomous sesquilinear form satisfying (2)-(5). Let {U (t, s) : (t, s) ∈ ∆}
given by (11). Then the function (t, s) 7→ U (t, s) is norm continuous on ∆ into L(X) for X = V, H and
V ′.
Proof. The norm continuity for U in the case where X = V follows from [15, Theorem 2.7]. On the other
hand, applying [15, Theorem 2.7] to a∗
r is also norm continuous on ∆ with values in
L(V ). Using Proposition 1.2, we obtain by similar arguments as in the proof of Lemma 2.1
r we obtain that U ∗
(23)
kU (t, s) − U (t′, s′)kL(V ′) ≤ kU ∗
r (T − s, T − t)x − U ∗
r (T − s′, T − t′)xkL(V )
for all x ∈ V ′ and (t, s), (t′, s′) ∈ ∆. This implies that U is norm continuous on ∆ with values in L(V ′).
Finally, the norm continuity in H follows then by interpolation.
(cid:3)
6
OMAR EL-MENNAOUI AND HAFIDA LAASRI
3. Examples
This section is devoted to some relevant examples illustrating the theory developed in the previous
sections. We refer to [4] and [19] and the references therein for further examples.
(i) Laplacian with time dependent Robin boundary conditions on exterior domain Let Ω
be a bounded domain of Rd with Lipschitz boundary Γ. Denote by σ the (d − 1)-dimensional Hausdorff
measure on Γ. Let Ωext denote the exterior domain of Ω, i.e., Ωext := Rd \ Ω. Let T > 0 and α > 1/4.
Let β : [0, T ] × Γ −→ R be a bounded measurable function such that
β(t, ξ) − β(t, ξ) ≤ ct − sα
for some constant c > 0 and every t, s ∈ [0, T ], ξ ∈ Γ. We consider the from a : [0, T ] × H 1(Ωext) ×
H 1(Ωext) −→ C defined by
a(t; u, v) :=ZΩext
∇u · ∇vdξ +ZΩext
β(t, ·)uΓ¯vΓdσ
where u → uΓ : H 1(Ωext) −→ L2(Γ, σ) is the trace operator which is bounded [2, Theorem 5.36]. The
operator A(t) associated with a(t; ·, ·) on H := L2(Ωext) is minus the Laplacian with time dependent
Robin boundary conditions
Here ∂ν is the weak normal derivative. Thus the domain of A(t) is the set
∂ν u(t) + β(t, ·)u = 0 on Γ.
D(A(t)) =nu ∈ H 1(Ωext) △u ∈ L2(Ωext), ∂νu(t) + β(t, ·)uΓ = 0o
and for u ∈ D(A(t)), A(t)u := − △u. Thus similarly as in [4, Section 5] one obtains that a satisfies
(2)-(5) with γ := r0 + 1/2 and ω(t) = tα where r0 ∈ (0, 1/2) such that r0 + 1/2 < 2α. We note that [4,
Section 5] the authors considered the Robin Laplacian on the bounded Lipschitz domain Ω. The main
ingredient used there is that the trace operators are bounded from H s(Ω) with value in H s−1/2(Γ, σ) for
all 1/2 < s < 3/4. This boundary trace embedding theorem holds also for unbounded Lipschitz domain,
and thus for Ωext, see [18, Theorem 3.38] or [6, Lemma 3.6].
Thus applying [4, Theorem 4.1] and Theorem 2.2 we obtain that the non-autonomous Cauchy problem
(24)
(
u(t) − △u(t) = 0, u(0) = x ∈ H 1(Ωext)
∂νu(t) + β(t, ·)u = 0 on Γ
has L2-maximal regularity in L2(Ωext) and its solution is governed by an evolution family U (·, ·) that is
norm continuous on each space V, L2(Ωext) and V ′.
3.1. Non-autonomous Schrödinger operators. Let m0, m1 ∈ L1
be a measurable function such that there exist positive constants α1, α2 and κ such that
Loc(Rd) and m : [0, T ] × Rd −→ R
α1m0(ξ) ≤ m(t, ξ) ≤ α2m0(ξ),
m(t, ξ) − m(s, ξ) ≤ κt − sm1(ξ)
for almost every ξ ∈ Rd and every t, s ∈ [0, T ]. Assume moreover that there exist a constant c > 0 and
s ∈ [0, 1] such that for u ∈ C∞
c (Rd)
(25)
m1(ξ)u(ξ)2dξ ≤ ckukHs(Rd).
ZRd
Consider the non-autonomous Cauchy problem
(26)
( u(t)− △u(t) + m(t, ·)u(t) = 0,
u(0) = x ∈ V.
Here A(t) = − △ + m(t, ·) is associated with the non-autonomous form a : [0, T ] × V × V −→ C given by
V :=(cid:26)u ∈ H 1(Rd) :ZRd
m0(ξ)u(ξ)2dξ < ∞(cid:27)
ON THE NORM-CONTINUITY FOR EVOLUTION FAMILY ARISING FROM NON-AUTONOMOUS FORMS∗3
7
and
a(t; u, v) =ZRd
∇u · ∇vdξ +ZRd
m(t, ξ)u(ξ)2dξ.
The form a satisfies also (2)-(5) with γ := s and ω(t) = tα for α > s
2 and s ∈ [0, 1].
This example is taken from [19, Example 3.1]. Using our Theorem 2.2 we have that the solution of
Cauchy problem (26) is governed by a norm continuous evolution family on L2(Rd), V and V ′.
References
[1] P. Acquistapace. Evolution operators and strong solutions of abstract linear parabolic equations. Differential Integral
Equations 1 (1988), no. 4, 433-457.
[2] R. A. Adams, J. J. F. Fournier. Sobolev spaces. Second edition. Pure and Applied Mathematics (Amsterdam), 140.
Elsevier/Academic Press, Amsterdam, 2003.
[3] W. Arendt. Heat
kernels.
9th
Internet Seminar
(ISEM)
2005/2006. Available
at. https://www.uni-
ulm.de/mawi/iaa/members/professors/arendt.html
[4] W. Arendt, S. Monniaux. Maximal regularity for non-autonomous Robin boundary conditions. Math. Nachr. 1-
16(2016) /DOI: 10.1002/mana.201400319
[5] H. Brézis. Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer, Berlin 2011.
[6] M. Costabel. Boundary integral operators on Lipschitz domains: elementary results. SIAM J. Math. Anal. 19 (1988),
no. 3, 613-626.
[7] D. Daners. Heat kernel estimates for operators with boundary conditions. Math. Nachr. 217 (2000), 13-41.
[8] R. Dautray and J.L. Lions. Analyse Mathématique et Calcul Numérique pour les Sciences et les Techniques. Vol. 8,
Masson, Paris, 1988.
[9] O. El-Mennaoui, H. Laasri. Stability for non-autonomous linear evolution equations with Lp− maximal regularity.
Czechoslovak Mathematical Journal. 63 (138) 2013.
[10] O. El-Mennaoui, H. Laasri. On evolution equations governed by non-autonomous forms. Archiv der Mathematik
(2016), 1-15, DOI 10.1007/s00013-016-0903-5
[11] K. J. Engel, R. Nagel. One-Parameter Semigroups for Linear Evolution Equations, Springer-Verlag, 2000.
[12] S. Fackler J.-L. Lions' problem concerning maximal regularity of equations governed by non-autonomous forms. Ann.
Inst. H. Poincaré Anal. Non Linéaire 34 (2017)
[13] T. Kato. Perturbation theory for linear operators. Springer-Verlag, Berlin 1992.
[14] H. Komatsu, Abstract analyticity in time and unique continuation property of solutions of a parabolic equation, J.
Fac. Sci. Univ. Tokyo, Sect. 1 9 (1961), 1-11.
[15] H. Laasri. Regularity properties for evolution family governed by non-autonomous forms. Archiv der Mathematik
(2018), https://doi.org/10.1007/s00013-018-1175-z.
[16] J.L. Lions. Equations Différentielles Opérationnelles et Problèmes aux Limites. Springer-Verlag, Berlin, Göttingen,
Heidelberg, 1961.
[17] A. Lunardi. Differentiability with respect to (t, s) of the parabolic evolution operator. Israel J. Math. 68 (1989), no.
2, 161-184.
[18] W. Mclean. Strongly elliptic systems and boundary integral equations. Cambridge University Press, Cambridge, 2000.
[19] E. M. Ouhabaz. Maximal regularity for non-autonomous evolution equations governed by forms having less regularity.
Arch. Math. 105 (2015), 79-91.
Princeton Univ. Press 2005.
[20] A. Sani, H. Laasri, Evolution Equations governed by Lipschitz Continuous Non-autonomous Forms. Czechoslovak
Mathematical Journal. 65 (140) (2015), 475-491.
[21] R. E. Showalter. Monotone Operators in Banach Space and Nonlinear Partial Differential Equations. Mathematical
Surveys and Monographs. American Mathematical Society, Providence, RI, 1997.
[22] H. Tanabe. Equations of Evolution. Pitman 1979.
University of Wuppertal School of Mathematics and Natural Sciences Arbeitsgruppe Funktionalanalysis,
Gaussstrasse 20 D-42119 Wuppertal, Germany
E-mail address: [email protected]
Ibn Zohr University, Faculty of Sciences Departement of mathematics, Agadir, Morocco
E-mail address: [email protected]
|
1211.3511 | 2 | 1211 | 2013-04-20T13:53:54 | On Kadison-Schwarz type quantum quadratic operators on $\bm_2(\mathbb{C})$ | [
"math.FA",
"math.DS"
] | In the present paper we study description of Kadison-Schwarz type quantum quadratic operators acting from $\bm_2(\mathbb{C})$ into $\bm_2(\mathbb{C})\o\bm_2(\mathbb{C})$. Note that such kind of operator is a generalization of quantum convolution. By means of such a description we provide an example of q.q.o. which is not a Kadision-Schwartz operator. Moreover, we study dynamics of an associated nonlinear (i.e. quadratic) operators acting on the state space of $\bm_2(\mathbb{C})$. | math.FA | math | ON KADISON-SCHWARZ TYPE QUANTUM QUADRATIC OPERATORS
ON M2(C)
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
Abstract. In the present paper we study description of Kadison-Schwarz type quantum qua-
dratic operators acting from M2(C) into M2(C) ⊗ M2(C). Note that such kind of operator
is a generalization of quantum convolution. By means of such a description we provide an
example of q.q.o. which is not a Kadision-Schwartz operator. Moreover, we study dynamics of
an associated nonlinear (i.e. quadratic) operators acting on the state space of M2(C).
Mathematics Subject Classification: 46L35, 46L55, 46A37. 60J99.
Key words: quantum quadratic operators; Kadison-Schwarz operator.
3
1
0
2
r
p
A
0
2
]
.
A
F
h
t
a
m
[
2
v
1
1
5
3
.
1
1
2
1
:
v
i
X
r
a
1. Introduction
It is known that one of the main problems of quantum information is characterization of
positive and completely positive maps on C∗-algebras. There are many papers devoted to this
problem (see for example [5, 14, 24, 25]). In the literature the completely positive maps have
proved to be of great importance in the structure theory of C∗-algebras. However, general
positive (order-preserving) linear maps are very intractable[14, 13]. It is therefore of interest
to study conditions stronger than positivity, but weaker than complete positivity. Such a
condition is called Kadison-Schwarz property, i.e a map φ satisfies the Kadison-Schwarz property
if φ(a)∗φ(a) ≤ φ(a∗a) holds for every a. Note that every unital completely positive map satisfies
this inequality, and a famous result of Kadison states that any positive unital map satisfies the
inequality for self-adjoint elements a. In [23] relations between n-positivity of a map φ and the
Kadison-Schwarz property of certain map is established. Certain relations between complete
positivity, positivite and the Kadison-Schwarz property have been considered in [1],[3],[2]. Some
spectral and ergodic properties of Kadison-Schwarz maps were investigated in [9, 10, 22].
In [20] we have studied quantum quadratic operators (q.q.o.), i.e. maps from M2(C) into
M2(C)⊗ M2(C), with the Kadison-Schwarz property. It is found some necessary conditions for
the trace-preserving quadratic operators to be the Kadison-Schwarz ones. Since trace-preserving
maps arise naturally in quantum information theory (see e.g. [21]) and other situations in which
one wishes to restrict attention to a quantum system that should properly be considered a
subsystem of a larger system with which it interacts. Note that in [6, 7] quantum quadratic op-
erators acting on a von Neumann algebra were defined and studied. Certain ergodic properties
of such operators were studied in [17, 18] (see for review [8]). In the present paper we continue
our investigation, i.e. we are going to study further properties of q.q.o. with Kadison-Schwarz
property. We will provide an example of q.q.o. which is not a Kadision-Schwarz operator, and
study its dynamics. We should stress that q.q.o.
is a generalization of quantum convolution
(see [26]). Some dynamical properties of quantum convolutions were investigated in [11].
Note that a description of bistochastic Kadison-Schwarz mappings from M2(C) into M2(C)
has been provided in [19].
1
2
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
2. Preliminaries
In what follows, by M2(C) we denote an algebra of 2 × 2 matrices over complex filed C. By
M2(C)⊗ M2(C) we mean tensor product of M2(C) into itself. We note that such a product can
be considered as an algebra of 4 × 4 matrices M4(C) over C. In the sequel 1I means an identity
matrix, i.e. 1I = (cid:18) 1 0
0 1 (cid:19). By S(M2(C)) we denote the set of all states (i.e.
linear positive
functionals which take value 1 at 1I) defined on M2(C).
Definition 2.1. A linear operator ∆ : M2(C) → M2(C) ⊗ M2(C) is said to be
(a) -- a quantum quadratic operator (q.q.o.) if it satisfies the following conditions:
(i) unital, i.e. ∆1I = 1I ⊗ 1I;
(ii) ∆ is positive, i.e. ∆x ≥ 0 whenever x ≥ 0;
(b) -- a Kadison-Schwarz operator (KS) if it satisfies
(2.1)
∆(x∗x) ≥ ∆(x∗)∆(x) for all x ∈ M2(C).
One can see that if ∆ is unital and KS operator, then it is a q.q.o. A state h ∈ S(M2(C)) is
called a Haar state for a q.q.o. ∆ if for every x ∈ M2(C) one has
(2.2)
Remark 2.2. Note that if a quantum convolution ∆ on M2(C) becomes a ∗-homomorphic map
with a condition
(h ⊗ id) ◦ ∆(x) = (id ⊗ h) ◦ ∆(x) = h(x)1I.
Lin((1I ⊗ M2(C))∆(M2(C))) = Lin((M2(C) ⊗ 1I)∆(M2(C))) = M2(C) ⊗ M2(C)
then a pair (M2(C), ∆) is called a compact quantum group [26]. It is known [26] that for given
any compact quantum group there exists a unique Haar state w.r.t. ∆.
Remark 2.3. Let U : M2(C) ⊗ M2(C) → M2(C) ⊗ M2(C) be a linear operator such that
U(x ⊗ y) = y ⊗ x for all x, y ∈ M2(C). If a q.q.o. ∆ satisfies U∆ = ∆, then ∆ is called a
quantum quadratic stochastic operator. Such a kind of operators were studied and investigated
in [17].
Each q.q.o. ∆ defines a conjugate operator ∆∗ : (M2(C) ⊗ M2(C))∗ → M2(C)∗ by
∆∗(f )(x) = f (∆x), f ∈ (M2(C) ⊗ M2(C))∗, x ∈ M2(C).
(2.3)
(2.4)
One can define an operator V∆ by
which is called a quadratic operator (q.c.). Thanks to the conditions (a) (i),(ii) of Def. 2.1 the
operator V∆ maps S(M2(C)) to S(M2(C)).
V∆(ϕ) = ∆∗(ϕ ⊗ ϕ), ϕ ∈ S(M2(C)),
3. Quantum quadratic operators with Kadison-Schwarz property on M2(C)
In this section we are going to describe quantum quadratic operators on M2(C) as well as
find necessary conditions for such operators to satisfy the Kadison-Schwarz property.
Recall [4] that the identity and Pauli matrices {1I, σ1, σ2, σ3} form a basis for M2(C), where
σ1 =(cid:18) 0 1
1 0 (cid:19) σ2 =(cid:18) 0 −i
i
0 (cid:19) σ3 =(cid:18) 1
0 −1 (cid:19) .
0
In this basis every matrix x ∈ M2(C) can be written as x = w01I + wσ with w0 ∈ C,
w = (w1, w2, w3) ∈ C3, here wσ = w1σ1 + w2σ2 + w3σ3.
ON DYNAMICS OF QUADRATIC OPERATORS
3
Lemma 3.1. [24] The following assertions hold true:
(a) x is self-adjoint iff w0, w are reals;
(b) Tr(x) = 1 iff w0 = 0.5, here Tr is the trace of a matrix x;
(c) x > 0 iff kwk ≤ w0, where kwk =pw12 + w22 + w32.
Note that any state ϕ ∈ S(M2(C)) can be represented by
ϕ(w01I + wσ) = w0 + hw, fi,
(3.1)
where f = (f1, f2, f3) ∈ R3 with kfk ≤ 1. Here as before h·,·i stands for the scalar product in
C3. Therefore, in the sequel we will identify a state ϕ with a vector f ∈ R3.
In what follows by τ we denote a normalized trace, i.e. τ (x) = 1
2 Tr(x), x ∈ M2(C),
Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a q.q.o. with a Haar state τ . Then one has
τ ⊗ τ (∆x) = τ (τ ⊗ id)(∆(x)) = τ (x)τ (1I) = τ (x),
x ∈ M2(C),
which means that τ is an invariant state for ∆.
Let us write the operator ∆ in terms of a basis in M2(C) ⊗ M2(C) formed by the Pauli
matrices. Namely,
∆1I = 1I ⊗ 1I;
∆(σi) = bi(1I ⊗ 1I) +
Xj=1
b(1)
ji (1I ⊗ σj) +
Xj=1
ij , bijk ∈ C, (i, j, k ∈ {1, 2, 3}).
where bi, b(1)
ij , b(2)
One can prove the following
3
3
3
b(2)
ji (σj ⊗ 1I) +
bml,i(σm ⊗ σl),
Xm,l=1
i = 1, 2, 3,
Theorem 3.2. [20, Proposition 3.2] Let ∆ : M2(C) → M2(C)⊗ M2(C) be a q.q.o. with a Haar
state τ , then it has the following form:
(3.2)
∆(x) = w01I ⊗ 1I +
3
hbml, wiσm ⊗ σl,
Xm,l=1
where x = w0 + wσ, bml = (bml,1, bml,2, bml,3) ∈ R3, m, n, k ∈ {1, 2, 3}.
Let us turn to the positivity of ∆. Given vector f = (f1, f2, f3) ∈ R3 put
(3.3)
β(f)ij =
bki,jfk.
3
Xk=1
Define a matrix B(f) = (β(f)ij)3
ij=1.
By kB(f)k we denote a norm of the matrix B(f) associated with Euclidean norm in C3. Put
S = {p = (p1, p2, p3) ∈ R3 : p2
1 + p2
2 + p2
3 ≤ 1}
and denote
kBk = sup
f∈S kB(f)k.
Proposition 3.3. [20, Proposition 3.3] Let ∆ be a q.q.o. with a Haar state τ , then kBk ≤ 1.
4
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a liner operator with a Haar state τ . Then due to
Theorem 3.2 ∆ has a form (3.2). Take arbitrary states ϕ, ψ ∈ S(M2(C)) and f, p ∈ S be the
corresponding vectors (see (3.1)). Then one finds that
3
(3.4)
∆∗(ϕ ⊗ ψ)(σk) =
bij,kfipj,
k = 1, 2, 3.
Xi,j=1
Thanks to Lemma 3.1 the functional ∆∗(ϕ ⊗ ψ) is a state if and only if the vector
f∆∗(ϕ,ψ) =(cid:18) 3
Xi,j=1
bij,1fipj,
3
3
bij,2fipj,
Xi,j=1
Xi,j=1
bij,3fipj(cid:19).
satisfies kf∆∗(ϕ,ψ)k ≤ 1.
So, we have the following
Proposition 3.4. [20, Proposition 4.1] Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a liner operator
with a Haar state τ . Then ∆∗(ϕ ⊗ ψ) ∈ S(M2(C)) for any ϕ, ψ ∈ S(M2(C)) if and only if one
holds
(3.5)
≤ 1
for all
f, p ∈ S.
2
3
3
Xk=1(cid:12)(cid:12)(cid:12)(cid:12)
bij,kfipj(cid:12)(cid:12)(cid:12)(cid:12)
Xi,j=1
From the proof of Proposition 3.3 and the last proposition we conclude that kBk ≤ 1 holds
if and only if (3.5) is satisfied.
Remark 3.5. Note that characterizations of positive maps defined on M2(C) were considered
in [15] (see also [12]). Characterization of completely positive mappings from M2(C) into itself
with invariant state τ was established in [24] (see also [16]).
Next we would like to recall (see [20]) some conditions for q.q.o. to be the Kadison-Schwarz
ones.
xm =(cid:0)hbm1, wi,hbm2, wi,hbm3, wi(cid:1), fm = ϕ(σm),
αml = hxm, xli − hxl, xmi,
Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a linear operator with a Haar state τ , then it has a
form (3.2). Now we are going to find some conditions to the coefficients {bml,k} when ∆ is a
Kadison-Schwarz operator. Given x = w0 + wσ and state ϕ ∈ S(M2(C)) let us denote
(3.6)
(3.7)
where m, l = 1, 2, 3. Here and what follows [·,·] stands for the usual cross product in C3. Note
that here the numbers αml are skew-symmetric, i.e. αml = −αml. By π we shall denote mapping
{1, 2, 3, 4} to {1, 2, 3} defined by π(1) = 2, π(2) = 3, π(3) = 1, π(4) = π(1).
q(f, w) =(cid:0)hβ(f)1, [w, w]i,hβ(f)2, [w, w]i,hβ(f)3, [w, w]i(cid:1),
where β(f)m =(cid:0)β(f)m1, β(f)m2, β(f)m3(cid:1) (see (3.3))
Theorem 3.6. [20, Theorem 3.6] Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a Kadison-Schwarz
operator with a Haar state τ , then it has the form (3.2) and the coefficients {bml,k} satisfy the
following conditions
γml = [xm, xl] + [xm, xl],
Denote
(3.8)
(3.9)
3
kwk2 ≥ i
Xm=1
fmαπ(m),π(m+1) +
3
kxmk2
Xm=1
q(f, w) − i
(cid:13)(cid:13)(cid:13)(cid:13)
3
Xm=1
(3.10)
ON DYNAMICS OF QUADRATIC OPERATORS
5
fmγπ(m),π(m+1) − [xm, xm](cid:13)(cid:13)(cid:13)(cid:13)
3
fkαπ(k),π(k+1)
Xk=1
≤ kwk2 − i
Xm=1
kxmk2.
−
3
for all f ∈ S, w ∈ C3. Here as before xm =(cid:0)hbm1, wi,hbm2, wi,hbm3, wi(cid:1), bml = (bml,1, bml,2, bml,3)
and q(f, w), αml and γml are defined in (3.8),(3.6),(3.7), respectively.
Remark 3.7. The provided characterization with [14, 24] allows us to construct examples of
positive or Kadison-Schwarz operators which are not completely positive (see next section).
Now we are going to give a general characterization of KS-operators. Let us first give some
notations. For a given mapping ∆ : M2(C) → M2(C) ⊗ M2(C), by ∆(σ) we denote the vector
(∆(σ1), ∆(σ2), ∆(σ3)), and by w∆(σ) we mean the following
(3.11)
w∆(σ) = w1∆(σ1) + w2∆(σ2) + w3∆(σ3),
where w ∈ C3. Note that the last equality (3.11), due to the linearity of ∆, also can be written
as w∆(σ) = ∆(wσ).
Theorem 3.8. Let ∆ : M2(C) → M2(C) ⊗ M2(C) be a unital ∗-preserving linear mapping.
Then ∆ is a KS-operator if and only if one has
(3.12)
i[w, w]∆(σ) + (w∆(σ))(w∆(σ)) ≤ 1I ⊗ 1I,
for all w ∈ C3 with kwk = 1.
Proof. Let x ∈ M2(C) be an arbitrary element, i.e. x = w01I + wσ. Then x∗ = w01I + wσ.
Therefore
Consequently, we have
(3.13)
(3.14)
(3.15)
x∗x =(cid:0)w02 + kwk2(cid:1)1I +(cid:0)w0w + w0w − i(cid:2)w, w(cid:3)(cid:1)σ
∆(x) = w01I ⊗ 1I + w∆(σ), ∆(x∗) = w01I ⊗ 1I + w∆(σ)
∆(x∗x) =(cid:0)w02 + kwk2(cid:1)1I ⊗ 1I +(cid:0)w0w + w0w − i(cid:2)w, w(cid:3)(cid:1)∆(σ)
∆(x)∗∆(x) = w021I ⊗ 1I +(cid:0)w0w + w0w)∆(σ) + (w∆(σ))(w∆(σ))
∆(x∗x) − ∆(x)∗∆(x) = kwk21I ⊗ 1I − i(cid:2)w, w(cid:3)∆(σ) − (w∆(σ))(w∆(σ)).
From (3.14)-(3.15) one gets
So, the positivity of the last equality implies that
kwk21I ⊗ 1I − i(cid:2)w, w(cid:3)∆(σ) − (w∆(σ))(w∆(σ)) ≥ 0.
Now dividing both sides by kwk2 we get the required inequality. Hence, this completes the
proof.
(cid:3)
6
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
4. An Example of q.q.o. which is not Kadision-Schwarz one
In this section we are going to study dynamics of (5.2) for a special class of quadratic
operators. Such a class operators associated with the following matrix {bij,k} given by:
b11,1 = ε;
b12,1 = 0;
b13,1 = 0;
b22,1 = 0;
b23,1 = ε;
b33,1 = 0;
b11,2 = 0;
b12,2 = 0;
b13,2 = ε;
b22,2 = ε;
b23,2 = 0;
b33,2 = 0;
b11,3 = 0;
b12,3 = ε;
b13,3 = 0;
b22,3 = 0;
b23,3 = 0;
b33,3 = ε;
and bij,k = bji,k.
Via (3.2) we define a liner operator ∆ε, for which τ is a Haar state. In the sequel we would
like to find some conditions to ε which ensures positivity of ∆ε.
It is easy that for given {bijk} one can find a form of ∆ε as follows
(4.1)
∆ε(x) = w01I ⊗ 1I + εω1σ1 ⊗ σ1 + εω3σ1 ⊗ σ2 + εω2σ1 ⊗ σ3
+εω3σ2 ⊗ σ1 + εω2σ2 ⊗ σ2 + εω1σ2 ⊗ σ3
+εω2σ3 ⊗ σ1 + εω1σ3 ⊗ σ2 + εω3σ3 ⊗ σ3,
where as before x = w01I + wσ.
Theorem 4.1. A linear operator ∆ε given by (4.1) is a q.q.o. if and only if ε ≤ 1
3 .
Proof. Let x = w01I + wσ be a positive element from M2(C). Let us show positivity of the
matrix ∆ε(x). To do it, we rewrite (4.1) as follows ∆ε(x) = w01I + εB, here
ω1 − 2iω3 − ω2
ω3
B =
ω2 + iω1
ω2 + iω1
ω2 − iω1
−ω3
ω1 + ω2
ω2 − iω1
ω1 + ω2
−ω3
ω1 + 2iω3 − ω2 −ω2 − iω1 −ω2 − iω1
−ω2 + iω1
−ω2 + iω1
ω3
,
where positivity of x yields that w0, ω1, ω2, ω3 are real numbers. In what follows, without loss
of generality, we may assume that w0 = 1, and therefore kwk ≤ 1. It is known that positivity
of ∆ε(x) is equivalent to positivity of the eigenvalues of ∆ε(x).
Let us first examine eigenvalues of B. Simple algebra shows us that all eigenvalues of B can
be written as follows
λ1(w) = ω1 + ω2 + ω3 + 2qω2
λ2(w) = ω1 + ω2 + ω3 − 2qω2
λ3(w) = λ4(w) = −ω1 − ω2 − ω3
1 + ω2
2 + ω2
3 − ω1ω2 − ω1ω3 − ω2ω3
1 + ω2
2 + ω2
3 − ω1ω2 − ω1ω3 − ω2ω3
Now examine maximum and minimum values of the functions λ1(w), λ2(w), λ3(w), λ4(w) on
the ball kwk ≤ 1.
One can see that
(4.2)
λ3(w) = λ4(w) ≤
3
Xk=1
ωk ≤
√3
3
Xk=1
√3
ωk2 ≤
ON DYNAMICS OF QUADRATIC OPERATORS
7
Note that the functions λ3,λ4 can reach values ±√3 at ±(1/√3, 1/√3, 1/√3).
Now let us rewrite λ1(w) and λ2(w) as follows
(4.3)
(4.4)
λ1(w) = ω1 + ω2 + ω3 +
λ2(w) = ω1 + ω2 + ω3 −
One can see that
2
√2q3(ω2
√2q3(ω2
2
1 + ω2
2 + ω2
3) − (ω1 + ω2 + ω3)2
1 + ω2
2 + ω2
3) − (ω1 + ω2 + ω3)2
(4.5)
(4.6)
λk(hω1, hω2, hω3) = hλk(ω1, ω2, w3),
λ1(hω1, hω2, hω3) = hλ2(ω1, ω2, w3),
if h ≥ 0,
if h ≤ 0.
where k = 1, 2. Therefore, the functions λk(w), k = 1, 2 reach their maximum and minumum
on the sphere ω2
3 = 1 (i.e. kwk = 1). Hence, denoting t = ω1 + ω2 + ω3 from (4.5)
and (4.4) we introduce the following functions
1 + ω2
2 + ω2
g1(t) = t +
2
√2
√3 − t2,
g2(t) = t −
2
√2
√3 − t2
where t ≤ √3.
One can find that the critical values of g1 are t = ±1, and the critical value of g2 is t = −1.
Consequently, extremal values of g1 and g2 on t ≤ √3 are the following:
g1(t) = −
√3, max
t≤√3
g2(t) = −3, max
t≤√3
min
t≤√3
min
t≤√3
g1(t) = 3,
g2(t) = √3.
Therefore, from (4.5),(4.6) we conclude that
(4.7)
− 3 ≤ λk(w) ≤ 3,
for any kwk ≤ 1, k = 1, 2.
It is known that for the spectrum of 1I + εB one has
Therefore,
So, if
Sp(1I + εB) = 1 + εSp(B)
Sp(1I + εB) = {1 + ελk(w) : k = 1, 4}
ε ≤
1
max
kwk≤1λk(w)
,
k = 1, 4
then one can see 1 + ελk(w) ≥ 0 for all kwk ≤ 1 k = 1, 4. This implies that the matrix 1I + εB
is positive for all w with kwk ≤ 1.
Now assume that ∆ε is positive. Then ∆ε(x) is positive whenever x is positive. This means
that 1 + ελk(w) ≥ 0 for all kwk ≤ 1 (k = 1, 4). From (4.2),(4.7) we conclude that ε ≤ 1/3.
This completes the proof.
(cid:3)
Theorem 4.2. Let ε = 1
3 then the corresponding q.q.o. ∆ε is not KS-operator.
8
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
Proof. It is enough to show the dissatisfaction of (3.10) at some values of w (kwk ≤ 1) and
f = (f1, f1, f2).
Assume that f = (1, 0, 0), then a little algebra shows that (3.10) reduces to the following one
(4.8)
where
√A + B + C ≤ D
A = ε(ω2ω3 − ω3ω2) − iε2(2ω2ω3 − 2ω12 − ω2ω1 + ω1ω2 − ω1ω3 + ω3ω1)2
B = ε(ω1ω2 − ω2ω1) − iε2(2ω1ω2 − 2ω32 − ω1ω3 + ω3ω1 − ω3ω2 + ω2ω3)2
C = ε(ω3ω1 − ω1ω3) − iε2(2ω3ω1 − 2ω22 − ω3ω2 + ω2ω3 − ω2ω1 + ω1ω2)2
D = (1 − 3ε2)(ω12 + ω22 + ω32)
−iε2(ω3w2 − ω2ω3 + ω2ω1 − ω1ω2 + ω1ω3 − ω3ω1)
Now choose w as follows:
Then calculations show that
ω1 = −
1
9
;
ω2 =
5
36
;
ω3 =
5i
27
A =
C =
Hence, we find
9594
19131876
;
B =
1625
3779136
;
D =
19625
;
86093442
589
17496
.
r 9594
19131876
+
19625
86093442
+
1625
3779136
>
589
17496
which means that (4.8) is not satisfied. Hence, ∆ε is not a KS-operator at ε = 1/3.
(cid:3)
Recall that a linear operator T : Mk(C) → Mm(C) is completely positive if for any positive
matrix (cid:0)aij)(cid:1)n
i,j=1 is positive for all n ∈ N. Now we
are interested when the operator ∆ε is completely positive. It is known [5] that the complete
positivity of ∆ε is equivalent to the positivity of the following matrix
i,j=1 ∈ Mk(Mn(C)) the matrix (cid:0)T (aij)(cid:1)n
∆ε =(cid:18) ∆ε(e11) ∆ε(e12)
∆ε(e21) ∆ε(e22) (cid:19)
here eij (i, j = 1, 2) are the standard matrix units in M2(C).
From (4.1) one can calculate that
∆ε(e11) =
1
2
1I ⊗ 1I + εB11, ∆ε(e22) =
1
2
1I ⊗ 1I − εB11
∆ε(e12) = εB12, ∆ε(e21) = εB∗12
where
B11 =
1
2
0
0 −i
0 − 1
0
0
0 − 1
0
0
2
1
0
0
i
2
2
, B12 =
0
1+i
0
0
1+i
0
i
2
0
i
1−i
2 −i −i
2
1−i
2
0
0
0
ON DYNAMICS OF QUADRATIC OPERATORS
9
Hence, we find
where 1I8 is the unit matrix in M8(C) and
2 ∆ε = 1I8 + εB
B =
1
0
0
2i
0
0
0
1 + i
0
0
−1
0
0
−1
0
0
−2i −2i
0
1 − i
1 − i
0
0
0
1 + i
0
0
0
1 + i
−2i
0
0
2i
0
2i
0
1
1 − i −2i −2i
0
1 + i −1
0
0
2i
0
2i
1
0
−2i
0
0
1
0
0
1 − i
0
0
0
2i
0
0
−1
So, the matrix ∆ε is positive if and only if
ε ≤
1
λmax(B)
,
where λmax(B) = max
λ∈Sp(B)λ.
One can easily calculate that λmax(B) = 3√3. Therefore, we have the following
Theorem 4.3. Let ∆ε : M2(C) → M2(C) ⊗ M2(C) be given by (4.1). Then ∆ε is completely
positive if and only if ε ≤ 1
3√3
.
5. Dynamics of ∆ε
Let ∆ be a q.q.o. on M2(C). Let us consider the corresponding quadratic operator defined by
V∆(ϕ) = ∆∗(ϕ ⊗ ϕ), ϕ ∈ S(M2(C)). From Theorem 3.2 one can see that the defined operator
V∆ maps S(M2(C)) into itself if and only if kBk ≤ 1 or equivalently (3.5) holds. From (3.4)
we find that
3
(5.1)
Xi,j=1
Here as before S = {f = (f1, f2f p3) ∈ R3 : f 2
V∆(ϕ)(σk) =
bij,kfifj,
f ∈ S.
1 + f 2
2 + f 2
3 ≤ 1}.
So, (5.1) suggests us to consider of the following nonlinear operator V : S → S defined by
3
(5.2)
V (f)k =
bij,kfifj,
Xi,j=1
k = 1, 2, 3.
where f = (f1, f2, f3) ∈ S.
by (5.2) was investigated in [20, Theorem 4.4].
It is worth to mention that uniqueness of the fixed point (i.e. (0, 0, 0)) of the operator given
In this section, we are going to study dynamics of the quadratic operator Vε corresponding
to ∆ε (see (4.1)), which has the following form
(5.3)
Let us first find some condition on ε which ensures (3.5).
Vε(f )1 = ε(f 2
Vε(f )2 = ε(f 2
Vε(f )3 = ε(f 2
1 + 2f2f3)
2 + 2f1f3)
3 + 2f1f2)
10
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
Lemma 5.1. Let Vε be given by (5.3). Then Vε maps S into itself if and only if ε ≤ 1√3
satisfied.
is
Proof. "If" part. Assume that Vε maps S into itself. Then (3.5) is satisfied. Take f =
(1/√3, 1/√3, 1/√3), p = f. Then from (3.5) one finds
3
Xi,j=1
3
Xk=1(cid:12)(cid:12)(cid:12)(cid:12)
2
bij,kfipj(cid:12)(cid:12)(cid:12)(cid:12)
= 3ε2 ≤ 1
one finds
which yields ε ≤ 1/√3.
"only if" part. Assume that ε ≤ 1/√3. Take any f = (f1, f2, f3), p = (p1, p2, p3) ∈ S. Then
Xk=1(cid:12)(cid:12)(cid:12)(cid:12)
bij,kfipj(cid:12)(cid:12)(cid:12)(cid:12)
= ε2(f1p1 + f3p2 + f2p32 + f3p1 + f2p2 + f1p32
Xi,j=1
3
3
2
≤ ε2((f 2
+(p2
+f2p1 + f1p2 + f3p32)
1 + p2
2 + f 2
≤ ε2(1 + 1 + 1) = 3ε2 ≤ 1.
2 + f 2
2 + p2
1 + f 2
1 + p2
3 )(p2
3)(f 2
3) + (f 2
2 + p2
1 + f 2
3 ))
3 + f 2
2 + f 2
1 )(p2
1 + p2
2 + p2
3)
This completes the proof.
(cid:3)
Remark 5.2. We stress that condition (3.5) is necessary for ∆ to be a positive operator. Namely,
from Theorem 4.1 and Lemma 5.1 we conclude that if ε ∈ ( 1
] then the operator ∆ε is not
positive, while (3.5) is satisfied.
3, 1√3
(5.4)
In what follows, to study dynamics of Vε we assume ε ≤ 1√3
. Recall that a vector f ∈ S is a
fixed point of Vε if Vε(f) = f. Clearly (0, 0, 0) is a fixed point of Vε. Let us find others. To do
it, we need to solve the following equation
ε(f 2
ε(f 2
ε(f 2
then Vε has a unique fixed point (0, 0, 0) in S. If ε = 1√3
Proposition 5.3. If ε < 1√3
Vε has the following fixed points (0, 0, 0) and (± 1√3
Proof. It is clear that (0, 0, 0) is a fixed point of Vε. If fk = 0, for some k ∈ {1, 2, 3} then due to
ε ≤ 1√3
, one can see that only solution of (5.4) belonging to S is f1 = f2 = f3 = 0. Therefore,
we assume that fk 6= 0 (k = 1, 2, 3). So, from (5.4) one finds
1 + 2f2f3) = f1
2 + 2f1f3) = f2
3 + 2f1f2) = f3
We have the following
,± 1√3
,± 1√3
) in S.
then
(5.5)
Denoting
(5.6)
f1
f2
f1
f3
f2
f3
2+2f2f3
2+2f1f3
2+2f2f3
2+2f1f2
2+2f1f3
2+2f1f2
= f1
f2
= f1
f3
= f2
f3
x =
f1
f2
, y =
f1
f3
, z =
f2
f3
ON DYNAMICS OF QUADRATIC OPERATORS
11
From (5.5) it follows that
(5.7)
According to our assumption x, y, z are nonzero, so from (5.7) one gets
1+ 2x
x(cid:18) x(cid:0)1+ 2
xy(cid:1)
z − 1(cid:19) = 0
xy(cid:1)
y(cid:18) y(cid:0)1+ 2
1+2yz − 1(cid:19) = 0
z(cid:18) z(cid:0)1+ 2x
z(cid:1)
1+2yz − 1(cid:19) = 0
x(cid:0)1+ 2
xy(cid:1)
y(cid:0)1+ 2
xy(cid:1)
z(cid:0)1+ 2x
z(cid:1)
y(cid:0)1 + 2x
z (cid:1)
1+2yz = 1
1+2yz = 1
x(1 + 2yz)
= 1
= 1
1+ 2x
z
(5.8)
(5.9)
where 2x 6= −z and 2yz 6= −1.
Dividing the second equality of (5.8) to the first one of (5.8) we find
which with xz = y yields
Simplifying the last equality one gets
y + 2x2 = x + 2y2.
This means that either y = x or x + y = 1
2 .
(y − x)(1 − 2(y + x)) = 0.
Assume that x = y. Then from xz = y, one finds z = 1. Moreover, from the second equality
y = 1 + 2y. So, y2 + y − 2 = 0 therefore, the solutions of the last one are
2 − y. We note that y 6= 1/2, since x 6= 0. So, from
of (5.8) we have y + 2
y1 = 1, y2 = −2. Hence, x1 = 1, x2 = −2.
the second equality of (5.8) we find
Now suppose that x + y = 1
2 , then x = 1
y +
= 1 +
4
1 − 2y
4y2
1 − 2y
.
So, 2y2 − y − 1 = 0 which yields the solutions y3 = − 1
z3 = − 1
Consequently, solutions of (5.8) are the following ones
2 and x4 = − 1
2 , y4 = 1. Therefore, we obtain x3 = 1,
Now owing to (5.6) and we need to solve the following equations
2 , z4 = −2.
(1, 1, 1), (cid:0)1,−
( f1
f2
f2
f3
1
2
,−
1
2
1
2(cid:1), (cid:0) −
, 1,−2(cid:1), (−2,−2, 1).
= xk,
= zk,
k = 1, 4.
According to our assumption fk 6= 0, therefore we consider cases when xkzk 6= 0.
Now let us start to consider several cases:
case 1. Let x2 = 1, z2 = 1. Then from (5.9) one gets f1 = f2 = f3. So, from (5.4) we
3ε2 ≤ 1.
3ε . Now taking into account f1
2 ≤ 1 one gets
2 = f1, i.e. f1 = 1
find 3εf1
2 + f2
2 + f3
1
12
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
. Due to Lemma 5.1 the operator Vε is well defined
From the last inequality we have ε ≥ 1√3
iff ε ≤ 1√3
the last ones to (5.4) we get f1 + 3f1
, therefore, one gets ε = 1√3
;± 1√3(cid:1).
case 2. Let x2 = 1, z2 = −1/2. Then from (5.9) one finds f1 = f2, 2f2 = −f3. Substituting
3ε . Taking
into account f1
3, due to
Lemma 5.1 in this case the operator Vε is not well defined, therefore, we conclude that there is
not a fixed point of Vε belonging to S.
. Hence, in this case a solution is (cid:0) ± 1√3
3ε , f2 = − 1
9ε2 ≤ 1. This means ε ≥ q 2
2ε = 0. Then, we have f1 = − 1
;± 1√3
3ε , f3 = 2
2 ≤ 1 we find 1
9ε2 + 4
9ε2 + 1
2 + f3
2 + f2
Using the same argument for the rest cases we conclude the absence of solutions. This shows
that if ε < 1/√3 the operator Vε has unique fixed point in S. If ε = 1/√3, then Vε has three
fixed points belonging to S. This completes the proof.
(cid:3)
Now we are going to study dynamics of operator Vε.
Theorem 5.4. Let Vε be given by (5.3). Then the following assertions hold true:
(i) if ε < 1/√3, then for any f ∈ S one has V n
(ii) if ε = 1/√3, then for any f ∈ S with f /∈(cid:8)(± 1√3
as n → ∞.
ε (f) → (0, 0, 0) as n → ∞.
)(cid:9) one has V n
,± 1√3
,± 1√3
ε (f) → (0, 0, 0)
Proof. Let us consider the following function ρ(f) = f 2
1 + f 2
2 + f 2
3 . Then we have
ρ(Vε(f)) = ε2(cid:0)(f 2
≤ ε2(cid:0)f 2
≤ ε2(cid:0)f 2
= 3ε2(f 2
1 + 2f2f3)2 + (f 2
1 + 2f2f3 + f 2
3 + f 2
1 + f 2
1 + f 2
3 ) = 3ε2ρ(f)
2 + f 2
2 + f 2
2 + 2f1f3)2 + (f 2
2 + 2f1f3 + f 2
2 + f 2
3 + f 2
1 + f 2
3 + 2f1f2)2(cid:1)
3 + 2f1f2(cid:1)
1 + f 2
2 )
3 + f 2
This means
(5.10)
ρ(Vε(f)) ≤ 3ε2ρ(f).
Due to ε2 ≤ 1
3 from (5.10) one finds
ρ(V n+1
ε
(f)) ≤ ρ(V n
ε (f)),
which yields that the sequence {ρ(V n
of {ρ(V n
(i). First we assume that ε < 1√3
ε (f))}.
ε (f))} is convergent. Next we would like to find the limit
, then from (5.10) we obtain
ρ(V n
ε (f)) ≤ 3ε2ρ(V n−1
ε
(f)) ≤ · · · ≤ (3ε2)nρ(f).
This yields that ρ(V n
(ii). Now let ε = 1√3
1 + f 2
Case (a). Let f 2
ε (f)) → 0 as n → ∞, for all f ∈ S.
. Then consider two distinct subcases.
1 + f 2
2 + f 2
3 < 1 and denote d = f 2
1 + 2f2f3)2 + (f 2
3 )2 + (f 2
1 + f 2
2 + f 2
ρ(Vε(f)) ≤ ε2(cid:0)(f 2
≤ ε2(cid:0)(f 2
= 3ε2d2 = dd = dρ(f).
Hance, we have ρ(Vε(f)) ≤ dρ(f). This means ρ(V n
n → ∞.
2 + f 2
3 . Then one gets
1 + f 2
2 + 2f1f3)2 + (f 2
3 )2 + (f 2
2 + f 2
3 + 2f1f2)2(cid:1)
2 )2(cid:1)
ε (f)) ≤ dnρ(f) → 0. Hence, V n
1 + f 2
3 + f 2
ε (f) → 0 as
ON DYNAMICS OF QUADRATIC OPERATORS
13
Case (b). Now take f 2
3 = 1 and assume that f is not a fixed point. Therefore, we
may assume that fi 6= fj for some i 6= j, otherwise from Lemma 5.3 one concludes that f is a
fixed point. Hence, from (5.3) one finds
2 + f 2
1 + f 2
Vε(f)1 = ε(f 2
1 + 2f2f3) = ε(1 − f 2
2 − f 2
3 + 2f2f3) = ε(1 − (f2 − f3)2).
Similarly, one gets
Vε(f)2 = ε(1 − (f1 − f3)2),
Vε(f)3 = ε(1 − (f1 − f2)2).
It is clear that Vε(f)k ≤ ε (k = 1, 2, 3). According to our assumption fi 6= fj (i 6= j) we
conclude that one of Vε(f)k is strictly less than 1√3
3 < 1.
Therefore, from the case (a), one gets that V n
, this means Vε(f)2
2 + Vε(f)2
1 + Vε(f)2
ε (f) → 0 as n → ∞.
(cid:3)
Acknowledgement
The authors acknowledges the MOHE Grant FRGS11-022-0170. The first named author
thanks the Junior Associate scheme of the Abdus Salam International Centre for Theoretical
Physics, Trieste, Italy. The authors would like to thank to an anonymous referee whose useful
suggestions and comments improve the content of the paper.
References
[1] S.J. Bhatt, Stinespring representability and Kadison's Schwarz inequality in non-unital Banach star algebras
and applications, Proc. Indian Acad. Sci. (Math. Sci.), 108 (1998), 283 -- 303.
[2] R. Bhatia, C. Davies, More Operator Versions of the Schwarz Inequality, Commun. Math. Phys. 215 (2000),
239 -- 244.
[3] R. Bhatia, R. Sharma, Some inequalities for positive linear maps, Linear Alg. Appl. 436 (2012), 1562 -- 1571.
[4] O. Bratteli and D. W. Robertson, Operator algebras and quantum statistical mechanics. I, Springer, New
York-Heidelberg-Berlin 1979.
[5] M.-D. Choi, Completely Positive Maps on Complex Matrices, Linear Algebra Appl. 10 (1975), 285-290.
[6] N. N. Ganikhodzhaev, F. M. Mukhamedov, On quantum quadratic stochastic processes, and some ergodic
theorems for such processes, Uzb. Matem. Zh. 1997, no. 3, 820. (Russian)
[7] N.N. Ganikhodzhaev, F. M. Mukhamedov, Ergodic properties of quantum quadratic stochastic processes,
Izv. Math. 65 (2000), 873890.
[8] R. Ganikhodzhaev, F. Mukhamedov, U. Rozikov, Quadratic stochastic operators and processes: results
and open problems, Inf. Dim. Anal. Quantum Probab. and Related Topics 14(2011), 279-335.
[9] U. Groh, The peripheral point spectrum of Schwarz operator on C ∗-algebras, Math. Z. 176 (1981), 311 -- 318.
[10] U. Groh, Uniform ergodic theorems for identity preserving Schwarz maps on W ∗-algebras, J. Operator
Theory 11 (1984), 395 -- 404.
[11] U. Franz, A. Skalski, On ergodic properties of convolution operators associated with compact quantum
groups, Colloq. Math. 113 (2008), 13 -- 23.
[12] A. Kossakowski, A Class of Linear Positive Maps in Matrix Algebras, Open Sys. & Information Dyn.
10(2003) 213 -- 220.
[13] W.A. Majewski, On non-completely positive quantum dynamical maps on spin chains, J. Phys. A: Math.
Gen. 40 (2007) 11539-11545.
[14] W.A. Majewski, M. Marciniak, On a characterization of positive maps, J. Phys. A: Math. Gen. 34 (2001)
5863-5874.
[15] W.A. Majewski, M. Marciniak, On the structure of positive maps between matrix algebras, In book:
Noncommutative harmonic analysis with applications to probability, 249 -- 263, Banach Center Publ., 78,
Polish Acad. Sci. Inst. Math., Warsaw, 2007.
[16] W.A. Majewski, M. Marciniak, Decomposability of extremal positive unital maps on M2(C). In book: M.
Bozejko, (ed.) et al., Quantum probability, Banach Center Publications 73(2006), 347 -- 356.
14
FARRUKH MUKHAMEDOV AND ABDUAZIZ ABDUGANIEV
[17] F.M. Mukhamedov, On ergodic properties of discrete quadratic dynamical system on C ∗-algebras. Method
of Funct. Anal. and Topology, 7(2001), No.1, 63 -- 75.
[18] F.M. Mukhamedov, On decomposition of quantum quadratic stochastic processes into layer-Markov pro-
cesses defned on von Neumann algebras, Izvestiya Math. 68(2004), No.5. 1009 -- 1024.
[19] F. Mukhamedov, A. Abduganiev, On description of bistochastic Kadison-Schwarz operators on M2(C),
Open Sys. & Inform. Dynam. 17(2010), 245-253.
[20] F. Mukhamedov, H. Akin, S. Temir, A. Abduganiev, On quantum quadratic operators on M2(C) and their
dynamics, Jour. Math. Anal. Appl. 376(2011), 641 -- 655.
[21] M.A. Nielsen, I.L. Chuang, Quantum Computation and Quantum Information, Cambridge University Press,
Cambridge, 2000.
[22] A. G. Robertson, A Korovkin theorem for Schwarz maps on C ∗-algebras, Math. Z. 156(1977), 205 -- 206.
[23] A. G. Robertson, Schwarz inequalities and the decomposition of positive maps on C ∗-algebras Math. Proc.
Camb. Philos. Soc. 94(1983), 291 -- 296.
[24] M.B. Ruskai, S. Szarek, E. Werner, An analysis of completely positive trace-preserving maps on M2, Lin.
Alg. Appl. 347 (2002) 159-187.
[25] E. Stormer, Positive linear maps of operator algebras, Acta Math. 110(1963), 233 -- 278.
[26] S.L. Woronowicz, Compact matrix pseudogroups. Comm. Math. Phys. 111 (1987), 613-665.
Farrukh Mukhamedov, Department of Computational & Theoretical Sciences, Faculty of
Science, International Islamic University Malaysia, P.O. Box, 141, 25710, Kuantan, Pahang,
Malaysia
E-mail address: [email protected], farrukh [email protected]
Abduaziz Abduganiev, Department of Computational & Theoretical Sciences, Faculty of
Science, International Islamic University Malaysia, P.O. Box, 141, 25710, Kuantan, Pahang,
Malaysia
E-mail address: [email protected]
|
1703.09677 | 2 | 1703 | 2018-06-11T16:15:22 | A multiplier algebra functional calculus | [
"math.FA"
] | This paper generalizes the classical Sz.-Nagy--Foias $H^{\infty}(\mathbb{D})$ functional calculus for Hilbert space contractions. In particular, we replace the single contraction $T$ with a tuple $T=(T_1, \dots, T_d)$ of commuting bounded operators on a Hilbert space and replace $H^{\infty}(\mathbb{D})$ with a large class of multiplier algebras of Hilbert function spaces on the unit ball in $\mathbb C^d$. | math.FA | math |
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
KELLY BICKEL, MICHAEL HARTZ, AND JOHN E. MCCARTHY
Abstract. This paper generalizes the classical Sz.-Nagy–Foias H ∞(D) functional
calculus for Hilbert space contractions.
In particular, we replace the single con-
traction T with a tuple T = (T1, . . . , Td) of commuting bounded operators on a
Hilbert space and replace H ∞(D) with a large class of multiplier algebras of Hilbert
function spaces on the unit ball in Cd.
1. Introduction
1.1. Overview. One seminal result connecting operator theory and complex function
theory is the classical Sz.-Nagy–Foias H ∞(D) functional calculus [29], which says that
for a completely non-unitary contraction T on a Hilbert space K, the polynomial
functional calculus
(1)
ΦT : C[z] → B(K),
T 7→ p(T ),
extends to a weak-∗ continuous, contractive algebra homomorphism on H ∞(D) (with
related statements for general contractions). This allows one to make sense of f (T )
for fairly general functions f and operators T and has both illuminated the structure
of C0 contractions and related operators, see [8, 10, 28, 29], and led to breakthroughs
related to the invariant subspace problem, see [11, 12].
In this paper, we generalize the Sz.-Nagy–Foias H ∞(D) functional calculus in two
nontrivial ways: first, we replace the single contraction T ∈ B(K) with a tuple T =
(T1, . . . , Td) of commuting (but not necessarily contractive) bounded operators on K
and secondly, we replace H ∞(D) with a multiplier algebra associated to any of a large
class of reproducing kernel Hilbert spaces H on the open unit ball Bd ⊆ Cd.
Results of this type were proved by J. Eschmeier, who established an H ∞(Bd)-
functional calculus for completely non-unitary tuples of commuting operators sat-
isfying von Neumann's inequality over the unit ball [19] and by R. Clouâtre and
K. Davidson, who studied the Drury-Arveson space H 2
d on Bd and its multiplier
algebra Mult(H 2
d ). They proved that completely non-unitary commuting row con-
tractions admit a Mult(H 2
d )-functional calculus [13]. This current paper generalizes
these earlier investigations and provides new arguments that avoid technical analyses
of specific multiplier algebras.
2010 Mathematics Subject Classification. Primary 47A60; Secondary 47A13, 46E22.
Key words and phrases. Functional calculus, multiplier algebra, unit ball.
K.B. was partially supported by National Science Foundation Grant DMS 1448846.
M.H. was partially supported by a Feodor Lynen Fellowship.
J.M. was partially supported by National Science Foundation Grant DMS 1565243.
1
2
K. BICKEL, M. HARTZ, AND J. MCCARTHY
1.2. Main Result. Here is our setting: Let T = (T1, . . . , Td) be a tuple of commuting
operators on a Hilbert space K and let U = (U1, . . . , Ud) be a spherical unitary on K,
i = I. We say T is completely non-unitary if it has
i.e. each Ui is normal andP UiU ∗
no non-zero reducing subspace M with T M a spherical unitary.
Let H be a reproducing kernel Hilbert space on Bd whose multiplier algebra,
Mult(H), contains the polynomials C[z1, . . . , zd]. Thus, the polynomial functional
calculus for T ,
(2)
ΦT : C[z1, . . . , zd] → B(K),
T 7→ p(T ),
is defined on a subalgebra of Mult(H).
We are interested in well-behaved extensions of the polynomial functional calculus
ΦT , namely ways to make sense of f (T ) for a larger class of functions. Let A(H)
denote the norm closure of the polynomials in Mult(H). Then, we say T admits an
A(H)-functional calculus if ΦT extends to a completely contractive algebra homomor-
phism A(H) → B(K); this loosely says T admits an A(H) version of von Neumann's
inequality. Moreover, we say T is Mult(H)-absolutely continuous if in addition ΦT
extends to a weak-∗ continuous algebra homomorphism Mult(H) → B(K). Thus, the
following two questions naturally arise:
(Q1) When does T admit an A(H)-functional calculus?
(Q2) If T admits an A(H)-functional calculus, is T Mult(H)-absolutely continuous?
If H is a complete Nevanlinna-Pick space, then (Q1) has a concrete answer in terms
of the reproducing kernel of H (see Section 5). However, even in the case when H is
the Hardy space on the unit ball, so that A(H) is the ball algebra, we are not aware of
an explicit answer to (Q1). Nevertheless, Arveson's dilation theorem shows that one
can rephrase (Q1) in terms of dilations, and the resulting dilation has a very explicit
form if H satisfies some additional assumptions, which we now describe. We say that
H is a unitarily invariant space on Bd if the reproducing kernel K of H is of the form
∞
K(z, w) =
anhz, win with a0 = 1 and an > 0 for all n ∈ N.
Xn=0
If in addition limn→∞ an/an+1 = 1, we say that H is a regular unitarily invariant
space. A discussion of this condition can be found in Subsection 2.1. For now, we
merely mention that this class of spaces includes many well-known spaces on Bd such
as the Bergman space, the Hardy space, the Dirichlet space and the Drury-Arveson
space.
Then our main result is the following answer to (Q2) for completely non-unitary
tuples T :
Theorem 1.1. Let H be a regular unitarily invariant space on Bd. Let T = (T1, . . . , Td)
be a tuple of commuting operators on a Hilbert space that admits an A(H)-functional
calculus. If T is completely non-unitary, then T is Mult(H)-absolutely continuous.
This loosely says that if T satisfies an A(H) version of von Neumann's inequality
and is completely non-unitary, then we can make sense of f (T ) for all f ∈ Mult(H).
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
3
This theorem partially extends Corollary 1.7 in [19], which establishes an H ∞(Bd)-
functional calculus for completely non-unitary tuples which satisfy von Neumann's
inequality over the unit ball. Moreover, if H = H 2
d , the Drury-Arveson space, then
a theorem of Müller–Vasilescu [24] and Arveson [6] shows that a tuple of commuting
operators T admits an A(H)-functional calculus if and only if T is a row contraction,
i ≤ I. Thus, in this setting, we recover Theorem 4.3 in [13]. We remark
that in this last case, our proof does not require the detailed description of the dual
of A(H 2
d ) from [14], which was used in [13].
i=1 TiT ∗
i.e. Pd
Even in the one variable setting, Theorem 1.1 appears to be new. For example,
consider the one-variable Dirichlet space D. Because the Dirichlet space has an ir-
reducible complete Nevanlinna-Pick kernel, there is an explicit characterization (see
Theorem 5.1) for when T admits an A(D)-functional calculus, and this characteriza-
tion does not require T to be a contraction. Then for such T which are completely
non-unitary, Theorem 1.1 says that T is Mult(D)-absolutely continuous.
1.3. Outline of Paper. Recall that the Sz.-Nagy–Foias H ∞(D) functional calculus
requires two structural results about contractions: first, each contraction T decom-
poses as T = Tcnu ⊕ U, with Tcnu completely non-unitary and U unitary; and second,
Sz.-Nagy's dilation theorem, namely each contraction admits a unitary dilation. If
T = U is unitary, the spectral theorem implies that U has a weak-∗ continuous H ∞(D)
functional calculus if and only if its scalar spectral measure is absolutely continuous
with respect to Lebesgue measure. If T = Tcnu is completely non-unitary, then the
spectral measure of its minimal unitary dilation is absolutely continuous with respect
to Lebesgue measure, from which the general calculus follows [29].
This paper follows the classical outline.
In Section 2, we recall some necessary
background material on reproducing kernel Hilbert spaces and multivariable operator
theory. In particular, Proposition 2.2 notes that every tuple of commuting operators
T splits as T = Tcnu ⊕ U, for Tcnu completely non-unitary and U a spherical unitary.
Using standard dilation theoretic arguments, we show in Theorem 2.4 that (Q1) can
be rephrased in terms of concrete dilations of T for regular unitarily invariant spaces
H.
Section 3 develops some properties of Mult(H)-Henkin measures, the analogue of
measures which are absolutely continuous with respect to Lebesgue measure in our
setting. This section closely follows part of the work of Clouâtre and Davidson on Ad-
Henkin measures [14]. In particular, adapting their arguments, we extend in Lemma
3.3 a theorem of Henkin by showing that Mult(H)-Henkin measures form a band.
Moreover, Lemma 3.4 gives the following answer to (Q2) for spherical unitaries: a
spherical unitary U is Mult(H)-absolutely continuous if and only if its scalar spectral
measure is Mult(H)-Henkin. In Proposition 3.5, we also provide an answer to (Q2)
for operator tuples T in terms of a minimal dilation of T .
Section 4 proves our main result, Theorem 1.1. The proof requires the dilation the-
orem (Theorem 2.4) as well as Lemma 4.1, which says analytic polynomials are weak-∗
dense in L∞(µ), for µ a totally singular measure. Then Theorem 4.3 also answers
(Q2) for commuting operator tuples which may have spherical unitary summands.
4
K. BICKEL, M. HARTZ, AND J. MCCARTHY
In Section 5, we consider the case when H is a complete Nevanlinna-Pick space.
In this setting, Corollary 5.2 characterizes those operator tuples for which (Q1) and
(Q2) have affirmative answers.
2. Preliminaries
2.1. Unitarily invariant spaces. We begin by discussing unitarily invariant repro-
ducing kernel Hilbert spaces on the unit ball Bd; in what follows, we only consider
the situation d < ∞. Background material on reproducing kernel Hilbert spaces can
be found in [3] and [26]. Recall from the introduction that a unitarily invariant space
on Bd is a reproducing kernel Hilbert space on Bd whose reproducing kernel K is of
the form
∞
K(z, w) =
anhz, win with a0 = 1 and an > 0 for all n ∈ N.
of convergence of the power series P∞
Moreover, we say that H is a regular unitarily invariant space if in addition, we have
limn→∞ an/an+1 = 1. Since H is a space on Bd, it is natural to assume that the radius
n=0 antn is 1, so that the limit limn→∞ an/an+1,
if it exists, is necessarily equal to 1. Thus, we regard this condition as a regularity
condition on the sequence (an). If H is a regular unitarily invariant space, then the
polynomials are automatically multipliers of H (see, for example [20, Corollary 4.4
(1)]). Regularity is very useful to us since it guarantees that the tuple (Mz1, . . . , Mzd)
is a spherical unitary tuple modulo the compacts (see the proof of Theorem 2.4 below).
Examples of regular unitarily invariant spaces are the spaces with reproducing
kernels
K(z, w) =
1
(1 − hz, wi)α ,
where α ∈ (0, ∞), which includes in particular the Drury-Arveson space (α = 1), the
Hardy space (α = d) and the Bergman space (α = d + 1). A related scale of spaces
is given by the kernels
∞
Xn=0
K(z, w) =
(n + 1)shz, win,
Xn=0
where s ∈ R. Here, the Drury-Arveson space corresponds to the choice s = 0. If
s = −1, then we obtain the Dirichlet space on Bd.
2.2. The weak-∗ topology on the multiplier algebra. Let H be a unitarily
invariant space on Bd. Since 1 ∈ H, every multiplier ϕ ∈ Mult(H) is uniquely
determined by its associated multiplication operator Mϕ. Thus, the assignment ϕ 7→
Mϕ allows us to regard Mult(H) as a unital subalgebra of B(H). It is not hard to
see that Mult(H) is closed in the weak operator topology and a fortiori in the weak-∗
topology. Therefore, Mult(H) is a dual space in its own right, namely the dual of
B(H)∗/ Mult(H)⊥. We endow Mult(H) with the resulting weak-∗ topology.
Since unitarily invariant spaces are separable, so is B(H)∗/ Mult(H)⊥, hence the
weak-∗ topology on bounded subsets of Mult(H) is metrizable. Moreover, one easily
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
5
verifies that on bounded subsets of Mult(H), the weak-∗ topology coincides with the
topology of pointwise convergence on Bd.
2.3. Decomposition of operator tuples. The proof of [13, Theorem 4.1] implicitly
contains the fact that every commuting operator tuple T = (T1, . . . , Td) is the direct
sum of a completely non-unitary tuple and a spherical unitary. The authors of [13]
attribute the argument to Jörg Eschmeier. For completeness, we repeat the relevant
part of the argument, beginning with this lemma:
Lemma 2.1. Let U = (U1, . . . , Ud) be a tuple of commuting operators on a Hilbert
space. Then U is a spherical unitary if and only if
(3)
UkU ∗
k =
d
Xk=1
d
Xk=1
U ∗
k Uk = I.
Proof. It is obvious that every spherical unitary satisfies (3). Conversely, suppose that
U satisfies (3). Then U and U ∗ are spherical isometries. By a theorem of Athavale
[7], U therefore extends to a spherical unitary V = (V1, . . . , Vd), say
Vk =(cid:20)Uk Ak
0 Bk(cid:21) .
k=1 VkV ∗
From Pd
k=1 UkU ∗
k = I and Pd
0 I(cid:21) =(cid:20)I +Pd
(cid:20)I 0
k = I, we deduce
k=1 AkA∗
∗
k ∗
∗(cid:21) ,
so that Ak = 0 for k = 1, . . . , d. Hence, U is the restriction of the spherical unitary
V to a reducing subspace, so U itself is a spherical unitary.
(cid:3)
The desired decomposition of operator tuples follows easily. Note that there are no
contractivity assumptions on the operator tuple.
Proposition 2.2. Let T be a tuple of commuting operators on a Hilbert space K.
Then there exist unique complementary reducing subspaces Kcnu and Ku for T such
is a spherical unitary.
is completely non-unitary and U := T(cid:12)(cid:12)Ku
is a spherical unitary and let Ku be the closed linear span of all elements of U. Then
is completely
that Tcnu := T(cid:12)(cid:12)Kcnu
Proof. Let U be the set of all reducing subspaces M of T with the property that T(cid:12)(cid:12)M
Ku is reducing for T , and if we define Kcnu = K ⊖ Ku, then T(cid:12)(cid:12)Kcnu
Let U = T(cid:12)(cid:12)Ku
. Observe that each M ∈ U is contained in the space
non-unitary by definition.
d
ker I −
UkU ∗
Xk=1
U ∗
k Uk!
d
k! ∩ ker I −
Xk=1
k =Pd
k=1 UkU ∗
and hence, so is Ku. This shows that Pd
spherical unitary by Lemma 2.1.
k=1 U ∗
k Uk = IKu, so that U is a
6
K. BICKEL, M. HARTZ, AND J. MCCARTHY
u ⊕ K′
Finally, if K = K′
cnu is another decomposition as in the statement, then by
definition, K′
cnu ∩ Ku, the tuple T
is both spherical unitary and completely non-unitary. Hence, this subspace is trivial,
so K′
(cid:3)
u ⊂ Ku. On the reducing subspace Ku ⊖ K′
u = Ku and thus, K′
cnu = Kcnu.
u = K′
Remark 2.3. According to our definition, an operator tuple T is completely non-
unitary if and only if T has no non-zero reducing subspace M with T M a spherical
unitary. If T is a row contraction, then it is not hard to check that if T is completely
non-unitary in our sense, then it does not even have a non-zero invariant subspace
M with T M a spherical unitary. This is the definition used in [13].
2.4. Dilations. If T = (T1, . . . , Td) is a tuple of commuting operators on a Hilbert
space K, and if S = (S1, . . . , Sd) is a tuple of commuting operators on a larger Hilbert
space L ⊃ K, then we say that T dilates to S if L decomposes as L = L− ⊕ K ⊕ L+
such that with respect to this decomposition,
Sk =
0
∗
0
∗ Tk 0
∗
∗
∗
,
for k = 1, . . . , d.
Recall that the polynomials are multipliers of every regular unitarily invariant space
H. We let Mz denote the d-shift on H, namely the tuple of multiplication operators
(Mz1, . . . , Mzd). Abstract dilation theory in the form of Arveson's dilation theorem
yields the following result (cf. [6, Section 8]).
Theorem 2.4. Let H be a regular unitarily invariant space on Bd and let T be a tuple
of commuting operators on a Hilbert space K. Then the following are equivalent:
(i) T admits an A(H)-functional calculus.
(ii) T dilates to M κ
z ⊕ U for some cardinal κ and spherical unitary U.
Moreover, if K is separable, then the dilation M κ
space.
z ⊕ U can be realized on a separable
Proof. The implication (ii) ⇒ (i) is clear, since M κ
z tautologically admits an A(H)-
functional calculus, and U admits an A(H)-functional calculus since the multiplier
norm dominates the supremum norm.
Conversely, suppose that (i) holds. By Arveson's dilation theorem (see, for example,
[25, Corollary 7.7]), the A(H)-functional calculus for T dilates to a ∗-representation
π : C ∗(A(H)) → B(L), for some Hilbert space L. A result of Sarason (see, for example,
[3, Lemma 10.2]) implies that L decomposes as L = L− ⊕ K ⊕ L+ such that
π(Mzk) =
0
∗
0
∗ Tk 0
∗
∗
∗
for k = 1, . . . , d. It remains to show that π(Mz) is unitarily equivalent to M κ
z ⊕ U.
Since limn→∞ an/an+1 = 1, an application of [20, Theorem 4.6] shows that there
exists a short exact sequence of C ∗-algebras
0 −→ K(H) −→ C ∗(A(H)) −→ C(∂Bd) −→ 0,
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
7
where K(H) is the ideal of all compact operators on H, the first map is the inclusion
map and the second map sends Mzk to zk for k = 1, . . . , d. In this setting, general re-
sults about representations of C ∗-algebras (see, for example, the discussion preceding
Theorem 1.3.4 and Corollary 2 in Section 1.4 in [5]) show that π splits as an orthog-
onal direct sum π = π1 ⊕ π2, where π1 is unitarily equivalent to a multiple of the
identity representation and π2 annihilates the compact operators and hence can be
regarded as a representation of C(∂Bd). Consequently, π1(Mz) is unitarily equivalent
to M κ
z for some cardinal κ, and π2(Mz) is a spherical unitary tuple.
The additional claim follows because if H is separable, then the dilation π, which
is obtained from Stinespring's dilation theorem, can be realized on a separable space
(see the discussion following [25, Theorem 4.1]).
(cid:3)
In the setting of Theorem 2.4, we say that M κ
z ⊕ U ∈ B(L)d is a minimal dilation
of T if L is the smallest reducing subspace for M κ
z ⊕ U that contains K. If T acts
on a separable Hilbert space, then so does every minimal dilation of T . Note that
minimial dilations of this type are generally not unique up to unitary equivalence.
Indeed, if H = H 2, the Hardy space on the unit disc, and if T = Mz ∈ B(H 2), then
Mz itself and the bilateral shift are both minimal dilations of Mz of the form Mz ⊕ U,
but they are not unitarily equivalent.
2.5. Normal tuples. We briefly recall a few notions from the spectral theory for
commuting normal operators. Let N = (N1, . . . , Nd) be a commuting tuple of normal
operators on a Hilbert space K. By the Putnam-Fuglede theorem, the unital C ∗-
algebra C ∗(N) = C ∗(N1, . . . , Nd) is commutative. Moreover, the maximal ideal space
X of C ∗(N) can be identified with a compact subset of Cd via
X → Cd,
ρ 7→ (ρ(N1), . . . , ρ(Nd)).
The range of this map is called the joint spectrum of N, denoted by σ(N). By the
Gelfand-Naimark theorem, C ∗(N) is isomorphic to C(σ(N)) via a ∗-homomorphism
that sends Ni to zi for each i = 1, . . . , d.
A well-known result about representations of algebras of continuous functions (see
[16, Theorem IX.1.14]) shows that there exists a projection-valued spectral measure
E on σ(N) such that
p(N, N ∗) =Zσ(N )
p(z, z) dE
for every p ∈ C[z1, . . . , zd, z1, . . . , zd]. If K is separable, then W ∗(N), the von Neu-
mann algebra generated by N, admits a separating vector x ∈ K (see [16, Corollary
IX.7.9]), so the scalar-valued measure µ = hE(·)x, xi has the same null sets as E. Any
scalar-valued measure with this property is called a (scalar-valued) spectral measure
for N.
It is unique up to mutual absolute continuity. Moreover, if µ is a scalar-
valued spectral measure for N, then the isomorphism between C ∗(N) and C(σ(N))
extends to a ∗-isomorphism between W ∗(N) and L∞(µ) which is a weak-∗-weak-∗
homeomorphism (see [18, Theorem II.2.5]).
8
K. BICKEL, M. HARTZ, AND J. MCCARTHY
3. Henkin measures
We begin by recalling a few definitions of measures on the unit sphere which are
related to the ball algebra. Background material on this topic can be found in [27,
Chapter 9]. Let M(∂Bd) denote the space of complex regular Borel measures on ∂Bd,
which can be identified with the dual space of C(∂Bd), the set of continuous functions
on ∂Bd. A probability measure µ ∈ M(∂Bd) is called a representing measure for the
origin if
p(0) =Z∂Bd
p dµ
holds for every polynomial p (or, equivalently, for every function in the ball algebra).
A Borel set E is said to be totally null if it is null for every representing measure of
the origin. A measure µ ∈ M(∂Bd) is said to be totally singular if it is singular with
respect to every representing measure of the origin.
Let H be a unitarily invariant space on Bd such that the polynomials are multipliers
of H. Let A(H) denote the norm closure of the polynomials. Since the supremum
norm is dominated by the multiplier norm, A(H) is contained in the ball algebra and
so, every function in A(H) extends uniquely to a continuous function on Bd.
Then, we say that a functional ρ ∈ A(H)∗ is Mult(H)-Henkin if it extends to a
weak-∗ continuous functional on Mult(H). We say that a measure µ ∈ M(∂Bd) is a
Mult(H)-Henkin measure if the functional ρµ ∈ A(H)∗ defined by
ρµ(f ) =Z∂Bd
f dµ,
f ∈ A(H)
is Mult(H)-Henkin.
The classical Henkin measures are precisely the H ∞(Bd)-Henkin measures in our
terminology. Moreover, Mult(H 2
d )-Henkin measures are called Ad-Henkin in [13].
The following lemma is a straightforward generalization of [19, Lemma 1.1].
In
particular, specializing to the case K = C, we see that the notion of Mult(H)-Henkin
measures (or functionals) only depends on Mult(H) and not on the particular choice
of Hilbert space H.
Lemma 3.1. Let H be a unitarily invariant space on Bd such that the polynomials
are multipliers of H, let K be a Hilbert space and let Φ : A(H) → B(K) be a bounded
linear map. Then the following are equivalent:
(i) Φ extends to a weak-∗ continuous linear map from Mult(H) to B(K).
(ii) Whenever (pn) is a sequence of polynomials such that pnMult(H) ≤ 1 for all
n ∈ N and limn→∞ pn(z) = 0 for all z ∈ Bd, the sequence (Φ(pn)) converges
to 0 in the weak-∗ topology of B(K).
If Φ is multiplicative, then so is its weak-∗ continuous extension to Mult(H). More-
over, the set of Mult(H)-Henkin functionals forms a norm closed subspace of A(H)∗.
Proof. The implication (i) ⇒ (ii) is trivial. For the proof of the reverse implication,
we first observe that unitary invariance of H implies that there exists a strongly con-
tinuous one-parameter unitary group (Ut)t∈R on H defined by (Utf )(z) = f (eitz) for
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
9
f ∈ H and z ∈ Bd. If ϕ ∈ Mult(H), then UtMϕU ∗
t = Mϕ(t), where ϕ(t)(z) = ϕ(eitz).
Consequently, if ϕ ∈ Mult(H), then ϕ(t) belongs to Mult(H) with ϕ(t)Mult(H) =
ϕMult(H), and the map t 7→ ϕ(t) is continuous in the strong operator topology. Us-
ing these two facts, a routine argument involving the Fejér kernel (cf. [23, Lemma I
n=0 ϕn is the homogeneous expansion
of ϕ as an analytic function on Bd, then the sequence of polynomials (ψn) defined by
2.5]) shows that if ϕ ∈ Mult(H) and if ϕ = P∞
Xj=0
Xk=0
n + 1
ψn =
n
k
1
ϕj
satisfies ψnMult(H) ≤ ϕMult(H) and converges to ϕ in the strong operator topology.
With this observation in hand, the proof of [19, Lemma 1.1] carries over to the present
setting. Moreover, the first additional claim follows from the fact that multiplication
is separately continuous in the weak-∗ topology.
To show that the set of Mult(H)-Henkin functionals is norm closed inside of A(H)∗,
observe that the argument in the first part of the proof shows that the unit ball of
A(H) is weak-∗ dense in the unit ball of Mult(H). Thus, the contraction
Mult(H)∗ → A(H)∗,
ρ 7→ ρ(cid:12)(cid:12)A(H),
restricts to an isometry between the predual of Mult(H) and the set of Mult(H)-
Henkin functionals. Since the predual of Mult(H) is complete, it follows that the set
of Mult(H)-Henkin functionals is closed in A(H)∗. (Alternatively, this can also be
seen by using the characterization (ii) of Mult(H)-Henkin functionals.)
(cid:3)
Remark 3.2. Since the multiplier norm dominates the supremum norm over Bd, it also
follows from Lemma 3.1 that every classical Henkin measure is a Mult(H)-Henkin
measure. The converse is false in general. Indeed, there are Mult(H 2
2 )-Henkin mea-
sures that are not classical Henkin measures [21].
According to Henkin's theorem [22], see also [27, Theorem 9.3.1], classical Henkin
measures form a band. Clouâtre and Davidson showed that Mult(H 2
d )-Henkin mea-
sures form a band as well [14, Theorem 5.4]. Their proof generalizes to Mult(H)-
Henkin measures. If ν, µ ∈ M(∂Bd) are complex measures, then we write ν ≪ µ if
ν ≪ µ.
Lemma 3.3. Let H be a unitarily invariant space on Bd such that the polynomials
are multipliers of H. Then the Mult(H)-Henkin measures form a band. That is, if
µ, ν ∈ M(∂Bd) such that µ is Mult(H)-Henkin and ν ≪ µ, then ν is Mult(H)-Henkin.
Proof. While the relevant part of the proof of [14, Theorem 5.4] does generalize to
this setting, other parts do not easily generalize. Thus, for clarity, we include the
proof here. Given a subset E of ∂Bd, let kf kE denote the value supz∈E f (z).
By the Glicksberg-König-Seever decomposition theorem and Henkin's theorem (see
Theorem 9.4.4, Theorem 9.3.1 and Section 9.8 in [27]), we can decompose µ = µa +µs,
where µa is H ∞(Bd)-Henkin and µs is totally singular. Since every H ∞(Bd)-Henkin
measure is Mult(H)-Henkin, the assumption implies that µs = µ − µa is Mult(H)-
Henkin. Since ν ≪ µ, by the Radon-Nikodym theorem, there is a function h ∈ L1(µ)
10
K. BICKEL, M. HARTZ, AND J. MCCARTHY
such that ν = hµ = hµa + hµs. By Henkin's theorem [27, Theorem 9.3.1], hµa is
H ∞(Bd)-Henkin and hence, Mult(H)-Henkin. It therefore suffices to show that hµs
is Mult(H)-Henkin.
Let ε > 0 and let g be a continuous function on ∂Bd such that g − hL1(µs) < ε.
Since µs is totally singular, Rainwater's lemma (see [27, Lemma 9.4.3]) shows that
µs is concentrated on an Fσ set which is totally null. Thus, there exists a totally null
compact set K ⊂ ∂Bd such that µs − µs(cid:12)(cid:12)KM (∂Bd) < ε/(g∂Bd + 1). By Bishop's
theorem (see [27, Theorem 10.1.2]), there exists a function in the ball algebra which
agrees with g on K and is bounded by g∂Bd. Thus, there is a polynomial p such
that p − gK < ε and p∂Bd ≤ g∂Bd. Clearly, pµs is Mult(H)-Henkin since µs is
Mult(H)-Henkin. Moreover,
pµs − hµsA(H)∗ ≤ pµs − hµsM (∂Bd)
≤ (p − g)µsKM (∂Bd) + (p − g)(µs − µsK)M (∂Bd)
+ gµs − hµsM (∂Bd)
≤ µs(K) p − gK + p − g∂Bdµs − µsKM (∂Bd)
+ g − hL1(µs)
≤ ε(3 + µsM (∂Bd)).
Since ε > 0 was arbitrary and since the set of Mult(H)-Henkin functionals is norm
closed in A(H)∗ by Lemma 3.1, it follows that integration against hµs is Mult(H)-
Henkin, as desired.
(cid:3)
The spectral measure of a commuting tuple of normal operators (see Subsection 2.5)
is only unique up to mutual absolute continuity. Nevertheless, Lemma 3.3 shows that
it is meaningful to say that the spectral measure of a commuting spherical unitary is
Mult(H)-Henkin. Using Lemma 3.1 and Lemma 3.3, the following result can now be
proved just like [13, Lemma 3.1].
Lemma 3.4. Let H be a unitarily invariant space on Bd such that the polynomials
are multipliers of H, and let U be a spherical unitary on a separable Hilbert space.
Then U is Mult(H)-absolutely continuous if and only if the spectral measure of U is
Mult(H)-Henkin.
(cid:3)
We also obtain an answer to (Q2) in terms of minimal dilations of an operator
tuple. Recall from the discussion at the end of Subsection 2.4 that minimal dilations
are not necessarily unique. The following result generalizes [13, Theorem 3.2], at least
in the separable case. Since we are dealing with dilations as opposed to co-extensions,
the proof is somewhat more complicated.
Proposition 3.5. Let H be a regular unitarily invariant space on Bd. Let T =
(T1, . . . , Td) be a tuple of commuting operators on a separable Hilbert space K and
z ⊕ U ∈ B(Hκ ⊕ L) be a minimal dilation of T , where U is spherical unitary.
let M κ
Then T is Mult(H)-absolutely continuous if and only if the spectral measure of U is
Mult(H)-Henkin.
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
11
Proof. If the spectral measure of U is Mult(H)-Henkin, then U is Mult(H)-absolutely
continuous by Lemma 3.4. Since M κ
is Mult(H)-absolutely continuous and since
z
p(T ) = PKp(M κ
continuous as well.
z ⊕U)(cid:12)(cid:12)K for every polynomial p, it follows that T is Mult(H)-absolutely
Conversely, suppose that T is Mult(H)-absolutely continuous. Let E denote the
projection-valued spectral measure of U and let
M = {(x, y) ∈ L ⊕ L : hE(·)x, yi is Mult(H)-Henkin}.
We have to show that M = L ⊕ L. For y ∈ L, let
Since the map
My = {x ∈ L : (x, y) ∈ M}.
L → A(H)∗,
x 7→ hE(·)x, yi,
is linear and continuous, and since the set of all Mult(H)-Henkin functionals forms
a norm closed subspace of A(H)∗ by Lemma 3.1, we see that each My is a closed
subspace of L.
We claim that each My is reducing for U. To this end, note that
Z∂Bd
p dhE(·)Uix, yi = hp(U)Uix, yi =Z∂Bd
pzidhE(·)x, yi
and
Z∂Bd
p dhE(·)U ∗
i x, yi = hp(U)U ∗
i x, yi =Z∂Bd
pzidhE(·)x, yi
holds for every polynomial p. Clearly, the measures zihE(·)x, yi and zihE(·)x, yi
are absolutely continuous with respect to hE(·)x, yi, hence Lemma 3.3 shows that if
x ∈ My, then also Uix ∈ My and U ∗
i x ∈ My. Thus, each My is reducing for U.
Next, we show that if x, y ∈ K, then (PLx, PLy) ∈ M. Let p be a polynomial. Then
Z∂Bd
p dhE(·)PLx, PLyi = hp(U)PLx, PLyi = hp(M κ
z ⊕ U)x, yi − hp(M κ
z )PHκx, PHκyi
= hp(T )x, yi − hp(M κ
z )PHκx, PHκyi.
Since T and M κ
is Mult(H)-Henkin, so that (PLx, PLy) ∈ M.
z are both Mult(H)-absolutely continuous, we see that hE(·)PLx, PLyi
Finally, since M κ
z ⊕ U is a minimal dilation of T , it follows in particular that
L is the smallest reducing subspace for U that contains PLK. From the preceding
two paragraphs, we therefore deduce that if y ∈ PLK, then My = L. Another
application of Lemma 3.3 shows that if (x, y) ∈ M, then also (y, x) ∈ M, since
hE(·)x, yi = hE(·)y, xi. Thus, for every y ∈ L, the space My contains PLK. Using
minimiality of the dilation again, we conclude that My = L for every y ∈ L, so that
M = L ⊕ L, as desired.
(cid:3)
12
K. BICKEL, M. HARTZ, AND J. MCCARTHY
4. Mult(H)-absolute continuity
The goal of this section is to prove our main result, Theorem 1.1. Before proceeding
with the proof, we require a lemma about totally singular measures on ∂Bd. It is
a well-known consequence of the classical F. and M. Riesz theorem that if µ is a
measure on the unit circle which is singular with respect to Lebesgue measure, then
the analytic polynomials are weak-∗ dense in L∞(µ) (see, for example, [17, Exercise
VI.7.10]). The following lemma is a generalization of this fact to higher dimensions.
Lemma 4.1. Let µ be a positive totally singular measure on ∂Bd. Then the analytic
polynomials are weak-∗ dense in L∞(µ).
Proof. Let f ∈ L1(µ) such that
pf dµ = 0
Z∂Bd
for all analytic polynomials p. By the Hahn-Banach theorem, we need to show that
f = 0 ∈ L1(µ). To this end, observe that the measure f µ annihilates the ball algebra,
hence the Cole-Range theorem (see [27, Theorem 9.6.1]) shows that there exists a
representing measure ρ for the origin such that f µ ≪ ρ. On the other hand, since µ
is totally singular, µ and hence f µ is singular with respect to ρ. Thus, f µ = 0, as
desired.
(cid:3)
Let H be a regular unitarily invariant space on Bd. To prove Theorem 1.1, we will
use the dilation of Theorem 2.4 and so, require information about operator tuples of
z ⊕ U, where Mz = (Mz1, . . . , Mzd) ∈ B(H)d and U ∈ B(L)d is a spherical
the form M κ
unitary. Let us assume that L is separable, and let µ ∈ M(∂Bd) be a scalar valued
spectral measure of U. By the Glicksberg-König-Seever decomposition theorem [27,
Theorem 9.4.4] and Henkin's theorem [27, Theorem 9.3.1] (see also the discussion in
[27, Section 9.8]), we can decompose µ = µa + µs, where µa is H ∞(Bd)-Henkin and µs
is totally singular. Correspondingly, U = Ua ⊕ Us, where µa is a spectral measure of
Ua and µs is a spectral measure of Us. Let La, Ls denote the spaces on which Ua, Us
act, respectively.
The following lemma plays a key role in the proof of Theorem 1.1.
Lemma 4.2. In the setting of the preceding paragraph, let M be the unital weak-∗
closed subalgebra of B(Hκ ⊕ La ⊕ Ls) generated by M κ
z ⊕ Ua ⊕ Us, and let
Ms = {N ∈ W ∗(Us) : 0 ⊕ 0 ⊕ N ∈ M}.
Then Ms is a weak-∗ closed ideal of the von Neumann algebra W ∗(Us).
Proof. It is clear that Ms is a weak-∗ closed subspace of W ∗(Us). To show that it is
an ideal, let N ∈ Ms. For p ∈ C[z1, . . . , zd],
0 ⊕ 0 ⊕ p(Us)N = p(M κ
z ⊕ Ua ⊕ Us)(0 ⊕ 0 ⊕ N)
belongs to M and hence, p(Us)N ∈ Ms. The spectral measure µs of Us is totally
singular and so, Lemma 4.1 implies that the analytic polynomials are weak-∗ dense
in L∞(µs). Recall that W ∗(Us) and L∞(µs) are isomorphic as von Neumann algebras
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
13
via an isomorphism that sends the tuple Us to (z1, . . . , zd). Thus, operators of the
form p(Us), where p is an analytic polynomial, are weak-∗ dense in W ∗(Us). Since
Ms is weak-∗ closed, we conclude that RN ∈ Ms for all R ∈ W ∗(Us), and by the
commutativity of W ∗(Us), NR is also in Ms. Thus, Ms is an ideal.
(cid:3)
We are now in a position to prove Theorem 1.1.
Proof of Theorem 1.1. Let T = (T1, . . . , Td) ∈ B(K)d be a commuting operator tuple
on a Hilbert space K that admits an A(H)-functional calculus. Suppose that T is
not Mult(H)-absolutely continuous. We will show that T has a spherical unitary
summand.
In a first step, we observe that it suffices to consider the case where K is separable.
Indeed, since T is not Mult(H)-absolutely continuous and since the weak-∗ and weak
operator topologies coincide on bounded subsets of B(K), there exist by Lemma 3.1
a sequence of polynomials (pn) in the unit ball of Mult(H) that converges to 0 in the
weak-∗ topology of Mult(H) and x, y ∈ K such that the sequence (hpn(T )x, yi) does
not converge to 0. Let K0 be the smallest reducing subspace for T that contains x and
is not Mult(H)-absolutely continuous. Moreover,
has a spherical unitary summand, then so does T . Thus, we may assume that
y. Then K0 is separable, and T(cid:12)(cid:12)K0
if T(cid:12)(cid:12)K0
K is separable.
By Theorem 2.4, the tuple T dilates to M κ
z ⊕U, where κ is a countable cardinal and
U is a spherical unitary on a separable Hilbert space. As in the discussion preceding
Lemma 4.2, we can decompose U = Ua ⊕ Us. Since T is not Mult(H)-absolutely
continuous, Lemma 3.1 guarantees the existence of a sequence of polynomials (pn)
in the unit ball of Mult(H) such that (pn) converges to 0 in the weak-∗ topology of
Mult(H) but the bounded sequence (pn(T )) does not converge to 0 in the weak-∗
topology of B(K). Since K is separable, the weak-∗ topology on the unit ball of
B(K) is metrizable. Then, by passing to a subsequence, we can assume that (pn(T ))
actually converges to a non-zero operator A ∈ B(K).
Observe that (pn(M κ
z )) converges to 0 weak-∗. Moreover, (pn) is bounded in the
supremum norm on Bd and converges to 0 pointwise in Bd. Hence, as the spectral
measure of Ua is H ∞(Bd)-Henkin, the sequence (pn(Ua)) converges to 0 weak-∗ as
well. From
pn(T ) = PKpn(M κ
z ⊕ Ua ⊕ Us)(cid:12)(cid:12)K
and the fact that (pn(T )) converges to the non-zero operator A, we deduce that the
bounded sequence (pn(Us)) does not converge to 0 weak-∗. By passing to another
subsequence, we can assume that (pn(Us)) converges to a non-zero contraction R ∈
W ∗(Us), so that A = PK(0 ⊕ 0 ⊕ R)(cid:12)(cid:12)K.
Let M and Ms be defined as in Lemma 4.2. By our previous arguments, R belongs
to Ms. By Lemma 4.2, Ms is a weak-∗ closed ideal of the von Neumann algebra
W ∗(Us). Thus, there exists an orthogonal projection P ∈ Ms such that P S = S for
all S ∈ Ms (see, for example, [9, Section III.1.1.13]). Since M κ
z ⊕ Ua ⊕ Us dilates T ,
the space K is semi-invariant for M κ
z ⊕ Ua ⊕ Us and hence for every operator in M.
14
K. BICKEL, M. HARTZ, AND J. MCCARTHY
Therefore, the map
M → B(K), S 7→ PKS(cid:12)(cid:12)K,
is multiplicative, which we will use repeatedly. Since K is semi-invariant and thus
reducing for the self-adjoint operator 0⊕0⊕P ∈ M, we see that Q = PK0⊕0⊕P(cid:12)(cid:12)K is
an orthogonal projection. Recall that M is commutative and so, Q and Tk commute
for 1 ≤ k ≤ d. Moreover, from P R = R, we deduce that QA = A, which implies
Q 6= 0.
We finish the proof by showing that the restriction of T to the range of Q is a
spherical unitary. Applying Lemma 4.2 again, we see that P (Us)k and P (Us)∗
k belong
to Ms for 1 ≤ k ≤ d, hence K is reducing for each operator 0 ⊕ 0 ⊕ P (Us)k. Moreover,
for 1 ≤ k ≤ d. In particular, each QTk is the compression of a normal operator to a
reducing subspace and thus, is normal. Moreover,
QTk = PK(0 ⊕ 0 ⊕ P (Us)k)(cid:12)(cid:12)K
d
which completes the proof.
Xk=1
QTk(QTk)∗ = PK(0 ⊕ 0 ⊕ P )(cid:12)(cid:12)K = Q,
(cid:3)
Moreover, in the presence of a spherical unitary summand, we obtain the following
characterization of Mult(H)-absolute continuity.
Theorem 4.3. Let H be a regular unitarily invariant space on Bd, and let T =
(T1, . . . , Td) be a commuting tuple of operators on a separable Hilbert space that admits
an A(H)-functional calculus. Let T = Tcnu ⊕ U be the decomposition of T from
Proposition 2.2. Then T is Mult(H)-absolutely continuous if and only if the spectral
measure of U is Mult(H)-Henkin.
Proof. This follows immediately from Theorem 1.1 and Lemma 3.4.
(cid:3)
5. Complete Nevanlinna-Pick Kernels
In this section, we obtain a refined version of Theorem 1.1 when H is a complete
Nevanlinna-Pick space. Specifically, we use the fact that in this case, there exists a
more checkable condition for T to admit an A(H)-functional calculus. Background
material on Nevanlinna-Pick spaces can be found in the book [3].
Let H be a regular unitarily invariant space on Bd with reproducing kernel K(z, w) =
n=0 anhz, win. A straightforward generalization of [3, Theorem 7.33] shows that H
is an irreducible complete Nevanlinna-Pick space if and only if the sequence (bn) given
by
P∞
(4)
satisfies bn ≥ 0 for all n ≥ 1.
bntn = 1 −
∞
Xn=1
1
n=0 antn
P∞
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
15
For example, the class of regular unitarily invariant complete Nevanlinna-Pick
spaces on Bd includes the spaces with reproducing kernels
K(z, w) =
1
(1 − hz, wi)α
for α ∈ (0, 1] and the spaces with kernels
∞
K(z, w) =
(n + 1)shz, win
Xn=0
for s ≤ 0, and hence in particular the Drury-Arveson space and the Dirichlet space
(see Subsection 2.1).
Let H be a regular unitarily invariant complete Nevanlinna-Pick space on Bd with
kernel K and let (bn) be the sequence of Equation (4). We write 1/K(T, T ∗) ≥ 0 if
N
Xn=1
bn Xα=n(cid:18)n
α(cid:19)T α(T ∗)α ≤ I
for all N ∈ N. This definition goes back to work of Agler [1]. For more discussion,
see [15, Section 5]. The following result is [15, Theorem 5.4].
Theorem 5.1. Let H be a regular unitarily invariant complete Nevanlinna-Pick space
on Bd with kernel K. Let T = (T1, . . . , Td) be a tuple of commuting operators on a
Hilbert space. Then the following are equivalent:
(i) The tuple T satisfies 1/K(T, T ∗) ≥ 0.
(ii) The tuple T admits an A(H)-functional calculus.
We remark that if H is not a complete Nevanlinna-Pick space, then one can often
still make sense of the condition 1/K(T, T ∗) ≥ 0, see [2, 4]. In general, however, not
every tuple which admits an A(H)-functional calculus satisfies 1/K(T, T ∗) ≥ 0, as the
example of the Bergman space and the unilateral shift shows. Rather, the condition
1/K(T, T ∗) ≥ 0 is related to the existence of co-extensions, whereas admitting an
A(H)-functional calculus is equivalent to the existence of dilations (see Theorem
2.4). In the complete Nevanlinna-Pick setting, there is no difference, see [15] for more
discussion.
The following refinement of Theorem 4.3 in the complete Nevanlinna-Pick setting
is almost immediate.
Corollary 5.2. Let H be a regular unitarily invariant complete Nevanlinna-Pick space
on Bd with kernel K. Let T = (T1, . . . , Td) be a commuting tuple of operators on a
separable Hilbert space and let T = Tcnu ⊕ U be the decomposition of Proposition 2.2.
Then the following are equivalent:
(i) The tuple T admits an A(H)-functional calculus and is Mult(H)-absolutely
continuous.
(ii) The tuple T satisfies 1/K(T, T ∗) ≥ 0 and the spectral measure of U is Mult(H)-
Henkin.
Proof. This follows from Theorem 5.1 and Theorem 4.3.
(cid:3)
16
K. BICKEL, M. HARTZ, AND J. MCCARTHY
References
1. Jim Agler, The Arveson extension theorem and coanalytic models, Integral Equations Operator
Theory 5 (1982), no. 5, 608–631.
2. Jim Agler and John E. McCarthy, Nevanlinna-Pick kernels and localization, Operator theoretical
methods (Timişoara, 1998), Theta Found., Bucharest, 2000, pp. 1–20.
3.
, Pick interpolation and Hilbert function spaces, Graduate Studies in Mathematics,
vol. 44, American Mathematical Society, Providence, RI, 2002.
4. C.-G. Ambrozie, M. Engliš, and V. Müller, Operator tuples and analytic models over general
domains in Cn, J. Operator Theory 47 (2002), no. 2, 287–302.
5. William Arveson, An invitation to C ∗-algebras, Springer-Verlag, New York-Heidelberg, 1976,
Graduate Texts in Mathematics, No. 39.
6.
, Subalgebras of C ∗-algebras. III. Multivariable operator theory, Acta Math. 181 (1998),
no. 2, 159–228.
7. Ameer Athavale, On the intertwining of joint isometries, J. Operator Theory 23 (1990), no. 2,
339–350.
8. Hari Bercovici, Operator theory and arithmetic in H ∞, Mathematical Surveys and Monographs,
vol. 26, American Mathematical Society, Providence, RI, 1988.
9. Bruce Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122, Springer-
Verlag, Berlin, 2006, Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and
Non-commutative Geometry, III.
10. Scott W. Brown and Bernard Chevreau, Toute contraction à calcul fonctionnel isométrique est
réflexive, C. R. Acad. Sci. Paris Sér. I Math. 307 (1988), no. 5, 185–188.
11. Scott W. Brown, Bernard Chevreau, and Carl M. Pearcy, Contractions with rich spectrum have
invariant subspaces, J. Operator Theory 1 (1979), no. 1, 123–136.
12.
13. Raphaël Clouâtre and Kenneth R. Davidson, Absolute continuity for commuting row contrac-
, On the structure of contraction operators II, J. Funct. Anal. 76 (1988), no. 1, 30–55.
tions, J. Funct. Anal. 271 (2016), no. 3, 620–641.
14.
, Duality, convexity and peak interpolation in the Drury-Arveson space, Adv. Math. 295
(2016), 90–149.
15. Raphaël Clouâtre and Michael Hartz, Multiplier algebras of complete Nevanlinna-Pick spaces:
Dilations, boundary representations and hyperrigidity, arXiv:1612.03936 (2016).
16. John B. Conway, A course in functional analysis, second ed., Graduate Texts in Mathematics,
vol. 96, Springer-Verlag, New York, 1990.
17.
, The theory of subnormal operators, Mathematical Surveys and Monographs, vol. 36,
American Mathematical Society, Providence, RI, 1991.
18. Kenneth R. Davidson, C ∗-algebras by example, Fields Institute Monographs, vol. 6, American
Mathematical Society, Providence, RI, 1996.
19. Jörg Eschmeier, Invariant subspaces for spherical contractions, Proc. London Math. Soc. (3) 75
(1997), no. 1, 157–176.
20. Kunyu Guo, Junyun Hu, and Xianmin Xu, Toeplitz algebras, subnormal tuples and rigidity on
reproducing C[z1, . . . , zd]-modules, J. Funct. Anal. 210 (2004), no. 1, 214–247.
21. Michael Hartz, Henkin measures for the Drury-Arveson space, Michigan Math. J. (to appear),
arXiv:1701.07777.
22. G. M. Henkin, The Banach spaces of analytic functions in a ball and in a bicylinder are non-
isomorphic, Funkcional. Anal. i Priložen. 2 (1968), no. 4, 82–91.
23. Yitzhak Katznelson, An introduction to harmonic analysis, third ed., Cambridge Mathematical
Library, Cambridge University Press, Cambridge, 2004.
24. V. Müller and F.-H. Vasilescu, Standard models for some commuting multioperators, Proc. Amer.
Math. Soc. 117 (1993), no. 4, 979–989.
A MULTIPLIER ALGEBRA FUNCTIONAL CALCULUS
17
25. Vern Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced
Mathematics, vol. 78, Cambridge University Press, Cambridge, 2002.
26. Vern I. Paulsen and Mrinal Raghupathi, An introduction to the theory of reproducing kernel
Hilbert spaces, Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2016.
27. Walter Rudin, Function theory in the unit ball of Cn, Classics in Mathematics, Springer-Verlag,
Berlin, 2008, Reprint of the 1980 edition.
28. Béla Sz.-Nagy and Ciprian Foias, Sur les contractions de l'espace de Hilbert. vii. triangulations
canoniques. fonction minimum, Acta Sci. Math. 25 (1964), 12–37.
29. Béla Sz.-Nagy, Ciprian Foias, Hari Bercovici, and László Kérchy, Harmonic analysis of operators
on Hilbert space, second ed., Cambridge Mathematical Library, Universitext Springer, New York,
2010.
Department of Mathematics, Bucknell University, 701 Moore Ave, Lewisburg, PA
17837, USA
E-mail address: [email protected]
Department of Mathematics, Washington University in St. Louis, One Brookings
Drive, St. Louis, MO 63130, USA
E-mail address: [email protected]
Department of Mathematics, Washington University in St. Louis, One Brookings
Drive, St. Louis, MO 63130, USA
E-mail address: [email protected]
|
1112.2329 | 1 | 1112 | 2011-12-11T07:38:42 | To the Many-Hilbert-Space Theory of Quantum Measurements | [
"math.FA"
] | In this work, a connection between some spectral properties of direct integral of operators in the direct integral of Hilbert spaces and their coordinate operators has been investigated. | math.FA | math | To the Many-Hilbert-Space Theory
of Quantum Measurements
by
Z.I.ISMAILOV and E.OTKUN ÇEVIK
Karadeniz Technical University, Faculty of Sciences, Department of Mathematics
61080 Trabzon, TURKEY
e-mail address : [email protected]
Abstract: In this work, a connection between some spectral properties of direct integral of operators
in the direct integral of Hilbert spaces and their coordinate operators has been investigated.
Keywords: Direct integral of Hilbert spaces and operators; spectrum and resolvent sets; compact
operators;Schatten-von Neumann operator classes; power and polynomially bounded operators.
2000 AMS Subject Classification: 47A10;81Q35.
1.Introduction
It is known that the general theory of linear closed operators in Hilbert spaces and its
applications to physical problems have been investigated by many researchers(for
example,see ).But many physical problems require studying the theory of linear
operators in direct sums or in general direct integrals of Hilbert spaces. The concepts of
direct integral of Hilbert spaces and direct integral of operators as a generalization of the
concept of direct sum of Hilbert spaces and direct sum of operators were introduced to
mathematics and developed in 1949 by John von Neumann in his work .These subjects
were incorporated in several works(see .A spectral theory of some operators on a
finite sum of Hilbert spaces was investigated by N.Dunford .Note that,in terms of
application ,there are some results in papers in the finite sum cases .Also in the
infinite direct sum cases there is a work ,in which some spectral and compactness
properties are surveyed.
Furthermore,some spectral investigations of the direct integral of operators in the direct
integral of Hilbert spaces have been provided by T.R.Chow ,T.R.Chow,F.Gilfeather ,
E.A.Azoff L.A.Fialkow .
It must be noted that the theory of direct integral of Hilbert spaces and operators on the
these spaces has important role in the representation theory of locally compact groups,in
the theory of decomposition rings of operators to factors,invariant measures,reduction
theory,von Neumann algebras and ets.On the other hand, many physical problems of today
1
arising in the modelling of processes of multiparticle quantum mechanics , quantum field theory and
in the physics of rigid bodies support to study a theory of direct integral of operators in the direct
integral of Hilbert spaces (see [19] and references in it).
Numereous scientific investigations have been done to explain quantum measurements .Dealing
with these subjects, S.Machida and M.Namiki (see, also have offered many-Hilbert-
space theory lately.Also,they find that decoherence of the wave function is only necessity to
formulate quantum mechanical measurements. A direct integral space of continiously many Hilbert
spaces and a continious superselection rule is starring in their theory.The direct integral space
structure is assigned to the measurement devise which reflects its macroscopic features.On the other
hand,observed system is left to be defined by a single Hilbert space.Furthermore,they have examined
double slit experiments and negative result experiments in the framework of the many-Hilbert-space
theory.Note that a direct integral space of continiously many Hilbert spaces often arises in the
quantum version of Lax-Phillips theory .In this investigation the direct integral space is introduced
in order to the allow the generator of motion to have a spectrum in real axis ,which is a necessary
condition for the application of the Lax-Phillips theory.
In second section of this paper a connection between parts of spectrum, resolvent set of direct
integral of operators defined in the direct integral of Hilbert spaces and parts of the spectrum of
‘’coordinate operators ‘’has been established . Note that the another approach to analogous problem
has been used in work . In this present paper sharp formulas for the connection are given. In
third section the compactness properties of these operators have been examined.Finally,in fourth
section in special case the analogous questions for the power and polynomially bounded operators
have been provided.
In the special case of direct sum of Hilbert spaces, these questions have been investigated in
Along this paper the triplet be any measure space and the Hilbert spaces are
looked at will become infinite dimensional. In addition ,the space of compact operators and
Schatten-von Neumann classes in any Hilbert space will be denoted by and
respectively.
2. On the spectrum of direct integral of operators
In this section, the relationship between the spectrum and resolvent sets of the direct integral of
operators and its coordinate operators will be investigated.
Before of all prove the following result.
Theorem 2.1.For the operator
are true
,
in the Hilbert space =
2
Proof. For any there exist element such that and
.Then almost everywhere with respect to measure µ it is true
Since ,then there exist which satisfy the above equality and .
This means that Hence
and from this it is obtained that
The proof of the second proposition is clear.
Actually ,in one special case the following stronger assertions are true.
Theorem 2.3. Assume that every one-point set is measurable and its measure is positive. For
the
parts
of
spectrum
and
resolvent
sets
of
the
operator
in
Hilbert
space
=
;
the following claims are true
;
;
;
Proof.Here only the first and second relations of theorem will be prove. The validity of other
propositions can be proved by the similar ideas.
such that .So
Assumed that . Then there exist
.Hence
)
for some
for every
and
.Then
On the contrary,assumed that
and from this
is true
at least one index is hold ,i.e. for some
.In this case for the element and
we
have .Consequently, .
Now we prove the second relation on the continuous spectrum.Let
. In this case by the
is a one-to-one operator,
definition of continuous spectrum
and
is dense in . From this and definition of direct integral it implies that for every
and
an operator is a one-to-one operator in ,
is dense in
,then
. Since
or
.
This means that
3
It is easy to prove the inverse implication.
On the other hand the simple calculations show that the following relations are true.
Corollary 2.4. Under the assumptions of last theorem we have
,
.
Corollary 2.5.Let Λ be any countable set , = and be
any measure with property for every point .In this case the formulas
,
,
,
.
are true.
Note that when is counting measure the analogous results have been
established in work
3. SOME COMPACTNESS PROPERTIES of DIRECT INTEGRAL of OPERATORS
In this section the compactness and spectral properties between direct integral of operators and
their ‘’coordinate operators’’have been established.In general ,there is not any relation between
mentioned operators in compactness means.
Example3.1. -counting measure,
.
In this case for every ) ,but
Example 3.2.In some cases from the relations no implies ) for every .
Indeed, from the definition of direct integral of operators on the set having null µ-measure the
coordinate operators may be defined by arbitrary way.
But in certain situations there are concrete results.
4
Theorem 3.3. Let Λ be any countable set , = and be
any measure with property for every point .Then
,then for every , .
(1) If
(2)Let Λ infinite countable set and for every .In this case
if and only if
This theorem is proved by analogous scheme of the proof in theorem 4.6 in .
Now give one characterizating theorem on the point spectrum of compact direct integral of
operators.
Theorem 3.4.Assumed that
in the Hilbert space =
, and .In this case there exist countable
subset of Λ such that the set is minimal and
.
From the definition of singular number s(.) (or characteristic numbers ) of any compact
operator in any Hilbert space and Theorems 2.1 and 3.4 it is easy to prove the validity of
the following result.
Theorem 3.5. Assumed that
in the Hilbert space =
, and .In this case there exist countable
subset of Λ such that
;
(2) İf then for every ,
;
(3) Let
.Then if and only if the series =1 =1 ( ) converges.
(4) If
and the series
is convergent,then
(5) If
and sup ,then for every ( ).
(6) If
, sup and for some , ,then for every
.
Proof.The validity of the claims (1) and (2) are clear.Prove third assertion of theorem.If the operator
,then the series
is convergent.In this case by the first proposition of
this theorem and important theorem on the convergence of the rearrangement series it is
obtained that the series
is convergent.
5
On the contrary, if the series
is convergent ,then the series
being a rearrangement of the above series,is also convergent.So .
Now prove (4).If for every then from the inequality
and first claim the validity of this assertion is clear.Now consider the general case.In this
case the operator can be written in form
.
Then On the other hand ,since and
,
then from the (3) of this theorem it implies that with .
Therefore,by the important theorem of the operator theory .
Furthermore ,by using proposition (2) of this theorem it is easy to prove the claim(5).On the
other hand, the claim(6) is one of the corollary of (5).
Remark 3.6.Note that for the some in representation
Theorem 3.4. may be hold
in
In these situations corresponding conditions for such index in the Theorem 3.5(3-6) may be
omitted ,for example,as in the following assertion.
Theorem 3.7. Assumed that
in the Hilbert space =
, and .In this case there exist countable
subset of Λ such that
If
,
,
and
,then .
6
4.Power and Polynomially Boundednessity of the Direct Sum Operators
In this section let us and is the counting measure .Here a connection of power
(and polynomially ) boundedness property of the direct sum operators in the direct sum Hilbert
spaces and its coordinate operators were established.In advance,give some necessary definitions for
the later.
Definition 4.1 .Let ℋ be any Hilbert space.
(1).An operator ℋ is called power bounded ( ℋ if there exist a constant
such that for any is satisfied (3.1)
(2)Operator ℋ polynomially bounded ℋ ,if there exist a constant
such that for any polynomial is satisfied , (3.2)
where =sup .
(3) The smallest number satisfying (3.1) (resp.(3.2)) is called the power bound (resp.polynomial
bound) of the operator and will be denoted by (resp. .
Before of all note that the following theorem is true.
, then for every , .
Theorem 4.2. If
This result is a one of the corollary of following equality
.
In general,the inverse of last assertion may be not true.
Example 4.3. Let
,
, ,
, . Consequently,
It is easy to see that and
=
.
Hence,for any ,but
Example 4.4. Let
,
, and
, ,i.e.for every , ,but
In this case and
=
Therefore, .
Actually,it is true the following result .
Theorem 4.5. if and only if and .
Proof.If ,then from the following relation
7
for each .From this it is determined that for any
it is implies that
.
On the other hand,it is clear that for each
Therefore, .
and
On the contrary,if for any ,
,then from the equality
it is obtained that .
Now polynomially boundness property of the direct sum operators will be investigated.In
advance,note that the following proposition is true.
,
and ,then for every ,
Theorem 4.6. If
.
Unfortunately,the inverse of last theorem may be not true in general.
Example 4.7. Let
, ,
It is known that and is a nilpotent operator with power of nilpotency 2 ,for
we have
any .In this case for any polynomial function
.
In the other words ,for every , .Unfortunately,for the polynomial
we have
i.e. .
But in general case the following result is true.
,
and .In order to the
Theorem 4.8. Let
necessary and sufficient conditions are and .
Proof. Assumed that for every polynomial p(.) .and
. In this case since for every polynomial function
),then
.From last relation it is obtained that
.Hence .
Now let us ,i.e.for any and polynomial p(.) it is valid
8
.
Then it is clear that , .In the other hand , from last equality it implies that
for every is hold .Hence, .This completes the
proof of the theorem.
Acknowledgment
The authors are grateful to G.İsmailov(student of Trabzon Kanuni Anadolu High School) for his
helping suggestion to english version and other technical discussion.
REFERENCES
some
sum
of
Hilbert
spaces and
1. Dunford, N., Schwartz, J. T., Linear Operators, I , II,Interscience, New York, 1958,1963.
2. Rofe-Beketov F.S.,Kholkin A.M.,Spectral Analysis of Differential Equations,First ed.,World
Scientific Monograph Series in Mathematics,New Jersey,Vol.7,2005.
3. von Neumann,J.,On rings of operators.Reduction theory,Ann.Math.,50,1949,pp.401-485.
4. Dixmier,J.,Les algebres d’operateurs dans l’espace Hilbertien,Geauther-Villars,Paris,1957.
5. Schwartz,J.T., algebras,Gordon and Breach,New York ,1967.
6. Mackey,G.W.,The theory of unitary group representations,University of Chicago
Press,1976.
7. Naimark,M.A.,Fomin,S.V.,Continuous direct
sum of Hilbert
applications,Uspehi Mat.Nauk,10,2(64),1955,111-142(in Russian).
8. Dunford,N.,Spectral
operators
in
a
direct
spaces,Mathematics,v.50,1963,p.1041-1043.
9. Dunford,N.,A spectral theory for certain operators on direct sum
spaces,Mat.Ann,162,1966,pp.294-330.
10. Sokolov,M.S.,An abstract approach to some spectral problems of direct sum differential
operators,Electr.J.Diff.Equat.,v.2003(2003),no.75,pp.1-6.
11. Zettl, A., Sturm-Liouville Theory, Amer. Math. Soc., Math. Survey and Monographs vol.
121, USA, 2005.
12. Ismailov, Z.I., Multipoint Normal Differential Operators for First Order, Opusc. Math., 29, 4,
(2009), 399-414.
13. Ismailov Z.I.,Otkun Çevik E.,Unluyol E.,Compact inverses of multipoint normal differential
operators for first order,Electr.J.Diff.Equat.,v.2011(2011),no.89,pp.1-11.
14. Chow
,T.R.,A
spectral
theory
for
the
direct
sum
operators,Math.Ann.188(1970),285-303.
15. Chow
of
,T.R.,Gilfeather,Functions
operators,Proc.Amer.Math.Soc.,v.29,2,1971 ,pp.325-330.
16. Azoff,E.A.,Spectrum and direct integral,Trans.Amer.Math.Soc.,197(1974),211-223.
17. Azoff,E.A.,Clancey,K.F.,Spectral multiplicity
of
integral
direct
for
operators,J.Operator theory,3,1980,pp.213-235.
of Hilbert
integral
of
integral
direct
of
normal
9
of
in Quantum
quasinilpotent
sum
direct
of
note
18. Fialkow,L.A.,A
operators,Proc.Amer.Math.Soc.,v.48.1,1975,pp.125-131.
19. Albeverio,S.,Gesztesy,F.,Hoegh-Krohn,R.and Holden,H.,Solvable Models
Mechanics,New-York,Springer,1988.
20. Machida ,S.,Namiki,M.,Theory of measurements in quantum-mechnics-mechanism of
reduction of the wave packet,Prog.Theor.Phys.63(5),1980,pp.1457-1473.
21. Namiki,M.,Many-Hilbert-Space Theory of quantum measurements,Foundations of
physics,v.18,1,1988,pp,29-55.
22. Namiki.M.,Pascazio,S.,Superselection rules and fluctuations in the Many-Hilbert-Space
approach to quantum measurements ,Found.Phys.Lett.,4(3),1991,pp.203-???.
23. Wheeler
J.A.,Zurek,W.H.,Quantum
theory and measurements,Princeton University
Press,1993.
24. Lax,P.D.,Phillips,R.S.,Scattering theory,Academic Press,New York,1967.
25. Halmos,P.R..,A Hilbert space problem book,Springer-Verlag,New York,1982.
26. Ionascu,E.J.,On power bounded operators,Proc.Amer.Math.Soc.,125,5,1997,pp.1435-1441.
10
|
1705.09861 | 2 | 1705 | 2018-06-25T21:59:21 | Binary frames with prescribed dot products and frame operator | [
"math.FA"
] | This paper extends three results from classical finite frame theory over real or complex numbers to binary frames for the vector space ${\mathbb Z}_2^d$. Without the notion of inner products or order, we provide an analog of the "fundamental inequality" of tight frames. In addition, we prove the binary analog of the characterization of dual frames with given inner products and of general frames with prescribed norms and frame operator. | math.FA | math |
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND
FRAME OPERATOR
VERONIKA FURST AND ERIC P. SMITH
Abstract. This paper extends three results from classical finite frame theory over
real or complex numbers to binary frames for the vector space Zd
2. Without the no-
tion of inner products or order, we provide an analog of the "fundamental inequality"
of tight frames. In addition, we prove the binary analog of the characterization of
dual frames with given inner products and of general frames with prescribed norms
and frame operator.
1. Introduction
Due to applications in signal and image processing, data compression, sampling
theory, and other problems in engineering and computer science, frames in finite-
dimensional spaces have received much attention from pure and applied mathemati-
cians alike, over the past thirty years (see, for example, Chapter 1 of [6]). The
redundant representation of vectors inherent to frame theory is central to the idea of
efficient data storage and transmission that is robust to noise and erasures.
Frames for Cd and Rd have been extensively studied (see [9] for a standard intro-
duction to frame theory, [16] for applications, and [13] for an exposition at the un-
dergraduate level). Noting the similarity between frames and error-correcting codes,
Bodmann, Le, Reza, Tobin, and Tomforde ([3]) introduced the concept of binary
frames, that is, finite frames for the vector space Zd
2. Binary Parseval frames robust
to erasures were characterized in [2], and their Gramian matrices were studied in [1].
A more generalized approach to binary frames was taken in [15].
We begin with a brief introduction to classical frame theory terminology. Let Hd
denote the Hilbert space Cd or Rd.
Definition 1.1. A (finite) frame for a Hilbert space Hd is a collection {xj}K
j=1 of
vectors in Hd for which there exist finite constants A, B > 0 such that for every
y ∈ Hd,
Akyk2 ≤
KXj=1
hy, xji2 ≤ Bkyk2.
The constants A and B are known as frame bounds. An A-tight frame is one for
which A = B, and a Parseval frame is one for which A = B = 1.
The vectors xj in the above definition need not be orthogonal or even linearly
independent. An orthonormal basis is most closely resembled by a Parseval frame,
for which we have the (not necessarily unique) reconstruction formula:
1
2
V. FURST AND E. P. SMITH
Proposition 1.2. A collection of vectors {xj}K
Hd is a Parseval frame for Hd if and only if
j=1 in a finite-dimensional Hilbert space
y =
KXj=1
hy, xji xj
for each y ∈ Hd.
Definition 1.3. Let {xj}K
The corresponding frame operator S : Hd → Hd is defined by
j=1 be a frame for the finite-dimensional Hilbert space Hd.
S(x) =
KXj=1
hx, xji xj.
It can be seen as the composition S = ΘΘ∗ of the synthesis operator Θ : CK → Hd
and its adjoint, the analysis operator Θ∗ : Hd → CK, given by the formulas
Θ
c1
c2
...
cK
=
KXj=1
respectively.
cjxj
and
Θ∗(x) =
hx, x1i
hx, x2i
...
hx, xKi
,
The frame operator is a bounded, invertible, self-adjoint operator satisfying AId ≤
S ≤ BId. Here and in what follows, we use Id to denote the d × d identity matrix
and 0d to denote the d × d zero matrix. A frame is Parseval if and only if its frame
operator is the identity operator.
From both a pure and an applied point-of-view, construction of frames with desired
properties has been a central question ([4]). In particular, much attention has been
paid to tight frames with prescribed norms and general frames with both prescribed
norms and frame operator.
In the case of tight frames, the answer, the so-called
"fundamental frame inequality," was provided by Casazza, Fickus, Kovacevi´c, Leon,
and Tremain:
Theorem 1.4 ([7], Corollary 4.11). Given real numbers a1 ≥ a2 ≥ . . . ≥ aK > 0,
K ≥ d, there exists a λ-tight frame {xj}K
j=1 for a d-dimensional Hilbert space Hd with
prescribed norms kxjk2 = aj for 1 ≤ j ≤ K if and only if
λ =
1
d
KXj=1
aj ≥ a1.
Casazza and Leon generalized this result to frames with prescribed frame operators
(the classical case is when S = λId):
Theorem 1.5 ([8], Theorem 2.1). Let S be a positive self-adjoint operator on a d-
dimensional Hilbert space Hd with eigenvalues λ1 ≥ λ2 ≥ . . . ≥ λd > 0. Given real
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
3
numbers a1 ≥ a2 ≥ . . . ≥ aK > 0, K ≥ d, there is a frame {xj}K
operator S and kxjk2 = aj for all 1 ≤ j ≤ K if and only if
j=1 for Hd with frame
KXj=1
aj =
dXj=1
λj
and
kXj=1
aj ≤
kXj=1
λj
for every 1 ≤ k ≤ d.
This can be seen as a consequence of the classical Schur-Horn theorem ([4]). In [5],
Cahill, Fickus, Mixon, Poteet, and Strawn introduce a so-called eigenstep method for
constructing all frames with a given frame operator and set of norms (see also [12],
and [4] for a survey of the topic).
A different approach was taken by Christensen, Powell, and Xiao in [10], extending
j=1, a sequence
Theorem 1.4 to the setting of dual frame pairs. Given a frame {xj}K
{yj}K
j=1 is called a dual frame if for every y ∈ Hd,
y =
KXj=1
hy, xji yj =
KXj=1
hy, yji xj.
Theorem 1.6 ([10], Theorem 3.1). Given a sequence of numbers {aj}K
K > d, the following are equivalent:
(1) There exist dual frames {xj}K
j=1 and {yj}K
j=1 for Hd such that hxj, yji = aj for
j=1 ⊂ H with
all 1 ≤ j ≤ K.
(2) There exists a tight frame {xj}K
j=1 and dual frame {yj}K
j=1 for Hd such that
hxj, yji = aj for all 1 ≤ j ≤ K.
(3) PK
j=1 aj = d.
The goal of this paper is to extend the theory of frames with prescribed norms (or
inner products) from the classical Hilbert spaces of Cd and Rd to the binary space Zd
2.
We provide analogs of Theorems 1.4, 1.5, and 1.6 for binary frames. The challenge, of
course, is the lack of an inner product, positive elements, and guaranteed eigenvalues.
Section 2 contains background material on binary frames. In Section 3, we explore
dual binary frames and prove the binary version of Theorems 1.6 and 1.4. In Section
4, we construct binary frames with prescribed "norms" and frame operator, as an
analog to Theorem 1.5. We conclude in Section 5 with examples and a catalog.
2. Binary Frames
In [3], Bodmann, Le, Reza, Tobin, and Tomforde introduce a theory of frames over
the d-dimensional binary space Zd
2 is the direct product Z2 ⊕ · · · ⊕ Z2 having
d ≥ 1 copies of Z2. The main trouble in defining frames in a binary space stems from
the lack of an ordering on Z2. Without an order, there can be no inner product
defined for binary space. In spite of this, [3] establishes the dot product as the analog
of the inner product on Rd and Cd.
2 where Zd
4
V. FURST AND E. P. SMITH
Definition 2.1 ([3]). The dot product on Zd
given by
2 is defined as the map (·, ·) : Zd
2×Zd
2 → Z2
(a, b) =
dXn=1
a[n]b[n].
Due to the degenerate nature of the dot product (note that (a, a) = 0 need not
imply a = 0), it fails to help define a frame in the manner of Definition 1.1. However,
when working over finite-dimensional spaces in the classical case, a frame is merely
a spanning sequence of vectors. This motivates the definition of a frame in binary
space.
Definition 2.2 ([3]). A frame is a sequence of vectors F = {fj}K
span(F ) = Zd
2.
j=1 in Zd
2 such that
The synthesis, analysis, and frame operators of F are defined similarly as in Defi-
nition 1.3 and are denoted ΘF , Θ∗
F , and SF , respectively.
Definition 2.3 ([3]). The synthesis operator of a frame F = {fj}K
matrix whose ith column is the ith vector in F . The analysis operator Θ∗
transpose of the synthesis operator. Explicitly,
j=1 is the d × K
F is the
ΘF =
f1
· · ·
fK
and
Θ∗
F =
- f ∗
1 -
...
- f ∗
K -
.
The frame operator is SF = ΘF Θ∗
F .
It is demonstrated in [3] that the spanning property of F is necessary and sufficient
for F to have a reconstruction identity with a dual family G. This fact is summed
up in the following theorem and is shown by choosing a basis consisting of d vectors
in F (without loss of generality, assumed to be f1, . . . , fd) and applying the Riesz
Representation Theorem to the linear functionals γi defined by γi(fj) = δij.
Theorem 2.4 ([3], Theorem 2.4). The family F = {fj}K
only if there exist vectors G = {gj}K
j=1 such that for all y ∈ Zd
2,
j=1 in Zd
2 is a frame if and
(1)
y =
KXj=1
(y, gj)fj.
In the proof, gi is defined as the unique vector satisfying γi(y) = (y, gi) for every y
for 1 ≤ i ≤ d, and gi = 0 for d < i ≤ K. Equation (1) can be rewritten as
ΘF Θ∗
G = Id,
which is equivalent to ΘGΘ∗
F = Id. Consequently, G is a dual frame to F . We will refer
to the dual frame G as a natural dual to F . Note that this definition is unrelated to the
usual definition of the canonical dual in Cd or Rd as {S −1(fj)}, where S = ΘΘ∗ is the
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
5
frame operator from Definition 1.3. Although SF is no longer necessarily invertible,
we still have
dXj=1
dXj=1
(gi, fj)fj =
δijfj = fi
for i ≤ d.
Propositions 2.5 and 2.6 make clear that the natural dual frame is unique, up to
permutation, if and only if K = d.
Proposition 2.5. Let F = {fj}K
Then H is a dual frame of F if and only if Θ∗
satisfying ΘF C = 0d.
j=1 be a frame for Zd
H = Θ∗
2 with a natural dual frame G.
G + C for some K × d matrix C
Proof. Given the existence of a matrix C with Θ∗
immediate that ΘF Θ∗
letting C = Θ∗
H − Θ∗
H = ΘF Θ∗
G gives ΘF C = ΘF Θ∗
G + C and ΘF C = 0d, it is
G = Id. Conversely, if H is a dual frame of F , then
(cid:3)
G = Id − Id = 0d.
H − ΘF Θ∗
H = Θ∗
As in Rd and Cd (see [13], Proposition 6.3), the following result still holds in Zd
2;
however, since the proof in [13] uses the invertibility of the frame operator, we provide
a modified proof here.
Proposition 2.6. Let F = {fj}K
frame if and only if F is a basis.
j=1 be a frame for Zd
2. Then F has a unique dual
Proof. Since a frame is a spanning set, F is a basis if and only if the vectors fj are
linearly independent and K = d. This is equivalent to the only K × d matrix C
satisfying ΘF C = 0d being the zero matrix. By Proposition 2.5, this happens if and
only if the (unique choice of) natural dual G is the only dual frame of F .
(cid:3)
The diagonal of the Gramian matrix Θ∗
when F and H are a dual frame pair, the diagonal of the cross-Gramian matrix Θ∗
is the vector whose ith element is (hi, fi).
Definition 2.7 ([3]). A Parseval frame for Zd
Zd
2 is a sequence of vectors F = {fj}K
j=1 ⊂
F ΘF is the vector whose ith entry is (fi, fi);
F ΘH
2 such that
y =
KXj=1
(y, fj)fj
for all y ∈ Zd
2.
Note that a binary Parseval frame must be a binary frame. In matrix notation,
F is a Parseval frame for Zd
If a collection of vectors
{xj} ⊂ Zd
2 satisfies (xi, xj) = 0 for all i 6= j and (xi, xi) = 1 for all i, we say, through
a slight abuse of terminology, that {xj} is an orthonormal set. An easy, matrix
theoretical consequence of the definitions of frame and Parseval frame is the following
proposition:
2 if and only if ΘF Θ∗
F = Id.
Proposition 2.8. Let F = {fj}K
j=1 be a sequence of vectors in Zd
2.
(1) The rows of ΘF are linearly independent if and only if F is a frame.
(2) The rows of ΘF are orthonormal if and only if F is a Parseval frame.
6
V. FURST AND E. P. SMITH
In the remainder of this paper, unless otherwise noted, all vectors are elements of
the binary vector spaces Zd
2 . All operations are performed modulo 2; for exam-
ple, Tr(A) represents the usual trace of a matrix A, but Tr(AB) ≡ Tr(BA) (mod 2)
for two binary matrices A and B. All frames refer to binary frames. Throughout, we
denote the standard orthonormal basis in Zd
2 by {ε1, ε2, . . . , εd}.
2 or ZK
3. Dual and Parseval Binary Frames
If F = {fj}K
j=1 is a frame for Zd
2 and K = d, then the vectors {fj} must be
linearly independent and hence a basis with a unique (natural) dual G. In this case,
(fj, gj) = γj(fj) = 1 for all 1 ≤ j ≤ K. In this section, we are largely concerned with
the question of which sequences α ∈ ZK
2 satisfy α[j] = (fj, hj) for a dual frame pair
(F , H); so we assume K > d. We use kαk0 to denote the number of nonzero entries
in a vector α, the parity of which will be fundamental in this paper.
j=1 be a frame for Zd
Lemma 3.1. Let F = {fj}K
2 and let π be a permutation of the set
{1, 2, . . . , K}. Denote by Fπ the frame {fπ(j)}K
j=1. Then a frame H is a dual frame
of F if and only if Hπ is a dual frame of Fπ. Furthermore, if α is a sequence in ZK
2 ,
then the dual frame pair (F , H) satisfies (fj, hj) = α[j] for every j if and only if the
dual frame pair (Fπ, Hπ) satisfies (fπ(j), hπ(j)) = α[π(j)] for every j.
Proof. Suppose ΘF Θ∗
H = Id and diag(Θ∗
ΘFπΘ∗
fπ(1)
· · ·
fπ(1)
· · ·
fπ(K)
HΘF ) = α. Then
- h∗
π(1) -
...
π(K) -
- h∗
π(1) -
...
π(K) -
P ∗P
- h∗
- h∗
fπ(K)
· · ·
fK
- h∗
1 -
...
- h∗
K -
Hπ =
=
=
f1
= ΘF Θ∗
= Id
H
where
P =
- ε∗
π−1(1) -
...
- ε∗
π−1(K) -
,
and here we use εi to indicate the ith standard basis vector in ZK
2 . Thus Fπ and Hπ
are dual frames and (fj, hj) = α[j] for each j implies (fπ(j), hπ(j)) = α[π(j)]. For the
converse statements, let σ = π−1.
(cid:3)
The next theorem and corollary are the analog of Theorem 1.6 in binary space.
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
7
Theorem 3.2. Given α ∈ ZK
(fi, hi) = α[i] for every i if and only if kαk0 ≡ d (mod 2).
2 , there exists a dual frame pair (F , H) for Zd
2 with
Proof. Suppose (F , H) is a dual frame pair for Zd
i. Then
2 such that (fi, hi) = α[i] for every
kαk0 ≡ Tr(Θ∗
HΘF ) ≡ Tr(ΘF Θ∗
H) ≡ Tr(Id) ≡ d (mod 2).
Conversely, suppose kαk0 ≡ d (mod 2). We consider three cases.
Case 1: kαk0 = d. Let fj = εj for 1 ≤ j ≤ d and let fj be arbitrary for d < j ≤ K.
A natural dual frame is then given by gj = εj for 1 ≤ j ≤ d and gj = 0 for d < j ≤ K.
Define β ∈ ZK
2 by β[j] = 1 for 1 ≤ j ≤ d and β[j] = 0 for d < j ≤ K, and let π be a
permutation of {1, 2, . . . , K} such that β[j] = α[π(j)]. It follows that
diag(Θ∗
GΘF ) = β.
By Lemma 3.1, (Fπ−1, Gπ−1) is the desired dual frame pair with (fπ−1(j), gπ−1(j)) =
β[π−1(j)] = α[j] for each j.
Case 2: kαk0 = d + 2t for some positive integer t ≤ K −d
2 . Consider the frame F
defined in Case 1 above, but set fj = ε1 for all d + 1 ≤ j ≤ d + 2t. A natural dual
frame G of F is the same as that defined in Case 1 above. Let C be the K × d matrix
whose top d × d block is 0d and rows d + 1 through d + 2t are equal to ε∗
1. The
remaining rows of C are zeros. Due to the introduction of an even number of ε∗
1's, we
see that ΘF C = 0d, and hence H given by the rows of Θ∗
G + C is a dual frame
of F , by Proposition 2.5. Since diag(Θ∗
HΘF ) is the vector composed of d + 2t ones
followed by K − (d + 2t) zeros, Lemma 3.1 again implies the existence of the desired
dual frame pair.
H = Θ∗
Case 3: kαk0 = d − 2t for some positive integer t ≤ d
2. Consider again the frame F
defined in Case 1, except set fd+1 = εd + εd−1 + · · · + εd−2t+2 + εd−2t+1. We again take
the same natural dual frame G as in Case 1 above. Let C be the K × d matrix whose
top d − 2t rows are zeros, row d − 2t + 1 through row d + 1 are f ∗
d+1, and the remaining
rows are zeros. Then ΘF C = 0d, and hence H defined by Θ∗
H = Θ∗
G + C is a dual
frame of F . Due to the presence of an even number of ones in fd+1, (fd+1, fd+1) = 0
while (fj, hj) = 0 for d − 2t + 1 ≤ j ≤ d. Consequently, diag(Θ∗
HΘF ) consist of ones
in its first d − 2t entries followed by K − (d − 2t) zeros. Lemma 3.1 again completes
the proof.
(cid:3)
Corollary 3.3. Given α ∈ ZK
dual frame H for Zd
2 , there exists a Parseval frame F and a corresponding
2 with (fi, hi) = α[i] for every i if and only if kαk0 ≡ d (mod 2).
Proof. The necessity of the condition on kαk0 follows immediately from Theorem 3.2.
The sufficiency depends on slight modifications of the frame F constructed in the
proof of Theorem 3.2. In Case 1, instead of letting fj for d < j ≤ K be arbitrary,
set each of those vectors to be the zero vector, ~0, in Zd
2. Proposition 2.8 implies F is
a Parseval frame. Similarly, the frame built in Case 2 is a Parseval frame if we set
fj = ~0 for 2t+1 ≤ j ≤ K. The frame built in Case 3 is not a Parseval frame; however,
consider instead the frame F ′ defined as f ′
d+2 = fd+1, and
j = ~0 for d + 3 ≤ j ≤ K. By Proposition 2.8, F ′ is a Parseval frame. Note that each
f ′
column of the matrix C constructed in Case 3 is still a (possibly trivial) dependence
j = fj for 1 ≤ j ≤ d + 1, f ′
8
V. FURST AND E. P. SMITH
relation among the columns of ΘF ′, which implies ΘF ′C = 0d. Since the natural dual
G of F constructed in Case 3 is still a natural dual of F ′, the frame H with analysis
operator Θ∗
j, hj) = (fj, hj) for all j
since hd+2 = 0.
(cid:3)
G + C is a dual frame of F ′; moreover, (f ′
H = Θ∗
Remark 3.4. The Parseval frames built in Cases 1 and 2 of the above corollary, in
fact, satisfy (fj, fj) = α[j] for each j after a suitable permutation, as allowed by
Lemma 3.1. However, this is not true in Case 3. By constructing a Parseval frame
that satisfies (fj, fj) = α[j] for each j in the case when kαk0 = d−2t for some positive
integer t ≤ d
2 , we will prove the binary analog of Theorem 1.4.
Theorem 3.5. Given nonzero α ∈ ZK
(fi, fi) = α[i] for every i if and only if kαk0 ≡ d (mod 2).
2 , there exists a Parseval frame F for Zd
2 with
Proof. Since a Parseval frame is self-dual, the necessity of the condition on kαk0
follows immediately from Theorem 3.2. For sufficiency, Remark 3.4 implies that we
need only construct Parseval frames satisfying (fi, fi) = α[i] for every i for kαk0 < d.
Since kαk0 ≡ d (mod 2) and kαk0 6= 0, we must have d ≥ 3.
For d = 3 and kαk0 = 1, we build the Parseval frame
2 that, after suitable permutation and inclusion of copies of ~0, satisfies any
for Z4
kαk0 = 2.
2
Given any odd dimension d, suppose we have constructed, without zero vectors,
corresponding to kαk0 = 1, 3, . . . , d −
the Parseval frames F 1, F 3, . . . , F d−4 for Zd−2
Zd
with two zero entries and unioning the augmented vectors with εd−1, εd ∈ Zd
2. Then
2 corresponding to kαk0 = 3, 5, . . . , d − 2.
4. For each odd n, create the collection eF n+2 by augmenting each vector of F n
eF 3, eF 5, . . . , eF d−2 are Parseval frames for Zd
Let eF 1 = {~1 + ε1,~1 + ε2, . . . ,~1 + εd,~1} where ~1 represents the vector with d ones in
2. Then eF1 is a Parseval frame for Zd
If the dimension d is even, we create the Parseval frames eF n+1 from eF n for each
eF n with a last entry of 0 and insert the vector εd ∈ Zd
In both cases, permutation of the vectors and possible inclusion of copies of ~0
(cid:3)
n = 1, 3, . . . , d − 3, corresponding to kαk0 = 2, 4, . . . , d − 2 : augment each vector of
2 corresponding to kαk0 = 1.
2.
finishes the proof.
and note that we may permute the vectors as needed. Moreover, we may insert any
number of copies of ~0 to satisfy any K > 4. By augmenting each vector with a last
entry of 0 and inserting the vector ε4 ∈ Z4
2, we construct the Parseval frame
0
1
1
,
1
0
1
,
1
1
0
,
1
1
1
,
0
1
1
0
,
1
0
1
0
,
1
1
0
0
,
1
1
1
0
,
0
0
0
1
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
9
4. Binary Frames with Prescribed Frame Operator
In the previous section, we gave a necessary and sufficient condition on α ∈ ZK
2 for
the existence of a Parseval frame F for Zd
2 with (fj, fj) = α[j] for every j. In classical
frame theory over R or C, the characterization has been broadened to frames with
a given frame operator and specified values for kfjk (the case of a Parseval frame
is when S = I), as in Theorem 1.5. In the classical case, the frame operator is a
symmetric, invertible, positive definite matrix. For a binary frame F , SF = ΘF Θ∗
F is
not necessarily invertible; for example, the zero matrix is the frame operator of any
frame in which every vector occurs twice. Consequently, we must first characterize
those binary symmetric matrices that are frame operators of binary frames.
Throughout this section, we rely heavily on the idea of vector parity in Zd
2.
Definition 4.1. Describe a vector v ∈ Zd
is even if kvk0 ≡ 0 (mod 2). If a vector is not even, then it is odd.
2 as even if (v, v) = 0. Equivalently, a vector
Lemma 4.2.
(1) The sum of two even vectors is an even vector.
(2) The sum of two odd vectors is an even vector.
(3) The sum of an odd vector and an even vector is an odd vector.
Proof. This follows from the above definition and the observation that if u, v ∈ Zd
2,
then
(u + v, u + v) = (u, u) + (u, v) + (v, u) + (v, v) = (u, u) + (v, v).
(cid:3)
As a consequence of this lemma, we note that a collection of only even vectors
cannot span Zd
2.
Given a d × d symmetric matrix S, we call A a factor of S if S = AA∗. We say A
is a minimal factor if it has the minimum number of columns over all factors of S.
Minimal factorization of symmetric binary matrices also arises in the computation of
the covering radius of Reed-Muller codes ([11]).
Theorem 4.3 ([17], Theorem 1). Every binary symmetric matrix S can be factorized
as S = AA∗ for some binary matrix A. The number of columns of a minimal factor
of S is rank(S) if diag(S) 6= ~0 and rank(S) + 1 if diag(S) = ~0.
Proposition 4.4. If S = AA∗ for some d × m matrix A where m = rank(S) or
m = rank(S) + 1, then rank(A) = rank(S).
Proof. Frobenius's rank inequality and the fact that, for properly sized matrices C
and D, rank(CD) ≤ min{rank(C), rank(D)} (see, for example, [14]) imply that
rank(A) + rank(A∗) ≤ m + rank(S) ≤ m + rank(A).
If m = rank(S), the above inequalities simplify to
rank(A) ≤ rank(S) ≤ rank(A),
and hence rank(A) = rank(S). If m = rank(S) + 1, we instead have
2 rank(A) ≤ 2 rank(S) + 1 ≤ 2 rank(A) + 1.
Since 2 rank(A) = 2 rank(S) + 1 is impossible, we must have rank(A) = rank(S). (cid:3)
10
V. FURST AND E. P. SMITH
We can use factors of a matrix to construct frames with a given frame operator.
Minimal factors correspond to minimal frames, that is, frames with the fewest number
of elements. In the previous section, we disregarded frames {fj}K
j=1 that were bases
since they corresponded to unique duals {gj}K
j=1 with predetermined values for the
dot products (fj, gj). In this section, however, we are concerned with (fj, fj), so we
do not rule the case K = d out of consideration.
Theorem 4.5. Let S be a d × d symmetric matrix with rank(S) = d. There exists
a (minimal) d-element frame F (i.e., basis) such that S = ΘF Θ∗
F if and only if
diag(S) 6= ~0. If diag(S) = ~0, then there exists a (minimal) (d + 1)-element frame F
such that S = ΘF Θ∗
F .
F . If diag(S) = ~0,
Proof. Suppose there exists a d-element frame F such that S = ΘF Θ∗
then ΘF is a d×d square matrix with all even rows, which cannot span Zd
2. Hence, the
columns of ΘF cannot span Zd
2, contradicting F being a frame. Conversely, suppose
diag(S) 6= ~0. By Theorem 4.3, there exists a d × k matrix A such that S = AA∗ and
k = rank(S). By Proposition 4.4, rank(A) = rank(S) = k = d. Thus, A is a d × d
matrix whose columns span Zd
2. Defining F by ΘF = A constructs the d-element
frame.
Now suppose diag(S) = ~0. By Theorem 4.3, there exists a minimal d × k factor A
such that S = AA∗ and k = rank(S)+1. By Proposition 4.4, rank(A) = rank(S) = d.
Therefore, the columns of A span Zd
2, and we can construct a (d + 1)-element frame
by taking ΘF = A.
(cid:3)
Remark 4.6. The columns of ΘF can be augmented by copies of the zero vector
without affecting S, so non-minimal frames can always be constructed from minimal
frames by including any number of copies of the zero vector.
Next we construct minimal frames whose prescribed frame operators are not of full
rank. Let r(A) and c(A) denote the number of rows and columns of a matrix A,
respectively.
Lemma 4.7 ([17], Section 4). Let S be a d × d symmetric matrix of rank(S) < d.
There exists a permutation matrix P and a nonsingular matrix T such that
S = P ∗(cid:20) L M
M ∗ K (cid:21) P = P ∗T (cid:20) L 0
0 0 (cid:21) T ∗P
where L is a symmetric matrix with r(L) = rank(L) = rank(S).
Corollary 4.8. A d × d symmetric matrix S with diag(S) = ~0 must have even rank.
Proof. It is known that if rank(S) = d, then d must be even (see, for example, [11]
Section 9.3). Suppose rank(S) < d. The rank(S) × rank(S) symmetric matrix L
constructed in Lemma 4.7 has rank(L) = rank(S). Since the diagonal elements of S
are the diagonal elements of L and K, diag(L) = ~0. So rank(S) must be even.
(cid:3)
Theorem 4.9. Let S be a d × d symmetric matrix with rank(S) < d. There exists a
k-element frame F such that S = ΘF Θ∗
F if and only if k ≥ 2d − rank(S).
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
11
Proof. Necessity follows from Frobenius's rank inequality
rank(ΘF ) + rank(Θ∗
F ) ≤ k + rank(S)
since frames are spanning sets.
Conversely, let k be an integer such that k ≥ 2d − rank(S). Let L, P, T be the
matrices guaranteed by Lemma 4.7, and let V = P ∗T . Suppose diag(L) 6= ~0. By
Theorem 4.3 there exists a factor H of L such that
and r(H) = c(H) = rank(L). Consider the augmented matrix
(cid:20) L 0
0 0 (cid:21) =(cid:20) H
0 (cid:21)(cid:2) H ∗ 0 (cid:3)
A =(cid:20) H 0
0 B (cid:21)
where the columns of B are the standard basis vectors of Zd−rank(S)
twice. Then r(A) = d, c(A) = 2d − rank(S), and AA∗ =(cid:20) L 0
0 0 (cid:21). By construction,
, each repeated
2
rank(A) = d. Since V is nonsingular, V A is a d × (2d − rank(S)) matrix of rank d
such that S = V AA∗V ∗. If k = 2d − rank(S), let a minimal frame F be the columns
of V A; if k > 2d − rank(S), augment the columns of V A with the necessary number
of zero vectors.
Now suppose diag(L) = ~0. By Corollary 4.8, rank(L) = rank(S) must be even. As
above (by Theorem 4.3) we can factorize L with a matrix H, but now r(H) = rank(L)
and c(H) = rank(L) + 1. In this case, we build the following augmented matrix:
eA =
0
r1
H (cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12) r2
(cid:12)(cid:12)(cid:12)(cid:12) eB
0
where r1 = [1 1 1 · · · 1] is a vector of length c(H), r2 = [1 0 0 0 · · · 0] has length
2
basis vectors of Zd−(rank(S)+1)
, each repeated twice. Since the (i, j) entry of the product
2(d − rank(S)) − 1, and the columns of eB are the zero vector followed by the standard
eAeA∗ can be viewed as the dot product of the ith and jth rows of eA, it is easy to
see that eAeA∗ = (cid:20) L 0
0 0 (cid:21).
vector. Since c(H) is odd, the vector [r1 r2] is even, as is each row of eB. By
construction, rank(eA) = d, r(eA) = r(H) + 1 + d − rank(S) − 1 = d, and c(eA) =
of V eA if k = 2d − rank(S) or these columns together with k − 2d + rank(S) copies of
c(H) + 1 + 2(d − rank(S) − 1) = 2d − rank(S). As above let F consist of the columns
Indeed, since diag(L) = ~0, each row of H is an even
the zero vector if k > 2d − rank(S).
(cid:3)
Theorems 4.5 and 4.9 provide minimal (and non-minimal) frames with frame oper-
ator S, subject only to restrictions based on rank(S). In what follows, sometimes we
12
V. FURST AND E. P. SMITH
will make additional assumptions on S, which allow the construction of frames with
frame operator S in different, and sometimes more intuitive, ways.
Definition 4.10. A d × d symmetric matrix S is said to be parity indicative if, for
every 1 ≤ i ≤ d, the diagonal entry Sii = 1 if and only if the ith row of S is odd.
Lemma 4.11. Let S be a d × d symmetric matrix, and suppose S = AA∗ for some
d × m matrix A. If every column of A is odd, then S is parity indicative. Conversely,
if S is parity indicative and the columns of A are linearly independent, then every
column of A must be odd.
Proof. Assume S = AA∗ =Pm
i where each column ai of A is odd. The ith row
i=1 aia∗
j satisfying aj[i] = 1. Suppose the ith row of S is odd.
of S equals the sum of those a∗
Then this sum must be composed of an odd number of nonzero terms by Lemma 4.2.
That is, there is an odd number of indices j having aj[i] = 1. Consequently, the ith
j=1 aj[i]aj[i] = 1. On the other hand, if the ith
row of S is even, then Lemma 4.2 implies that an even number of aj have aj[i] = 1,
resulting in Sii = 0.
row of A is an odd vector, and Sii =Pm
To show the converse, assume that the columns of A are linearly independent.
Suppose some of the columns of A are even; denote the odd columns by {oj} and the
even columns by {ej}. If for every index l, ej[l] = 1 for an even number of the vectors
{ej}, then P ej = ~0, contradicting the linear independence of the columns of A. So,
assume that there exists an index i such that the number of the vectors {ej} that
satisfy ej[i] = 1 is odd. If an odd number of the vectors {oj} are such that oj[i] = 1,
then the ith row of A is even, so Sii = 0; on the other hand, the ith row of S equals
the sum of an odd number of rows e∗
j , which is odd,
by Lemma 4.2. If there are an even number of the vectors {oj} with oj[i] = 1, then
the ith row of A is odd, so Sii = 1; on the other hand, the ith row of S equals the
sum of an odd number of e∗
j , which is even, by Lemma 4.2.
Therefore, S is not parity indicative.
(cid:3)
j plus an odd number of rows o∗
j plus an even number of o∗
Lemma 4.12. Let S be a d × d symmetric matrix, and suppose S = AA∗ for some
d × m matrix A. If S is parity indicative, diag(S) = ~0, and c(A) = rank(A) + 1, then
either every column of A is even or every column of A is odd.
Proof. Denote the odd columns and even columns of A by {oj} and {ej}, respectively,
and assume both sets are nonempty. Since each row of S is even, for every i, an even
number of the vectors {oj} must have oj[i] = 1, by Lemma 4.2.
It follows that
P oj = ~0; that is, Ax = ~0 where x[j] = 1 if j is the index of an odd column and
x[j] = 0 if j is the index of an even column. But since every row of A is even, A~1 = ~0.
Hence Ay = ~0 for y = ~1 + x. Since the nonzero linearly independent vectors x and y
are both contained in the null space of A, the rank-nullity theorem implies
rank(A) ≤ c(A) − 2 = rank(A) + 1 − 2 = rank(A) − 1,
a contradiction. Therefore, either {oj} or {ej} must be empty.
(cid:3)
One additional useful fact is required before we state our main result.
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
13
Lemma 4.13.
(1) Suppose e, o1, o2, o3 ∈ Zd
2 are four vectors such that e is even
and o1, o2, o3 are odd. Then there exists three even vectors f1, f2, f3 and an
odd vector p such that
ee∗ + o1o∗
1 + o2o∗
2 + o3o∗
3 = f1f ∗
1 + f2f ∗
2 + f3f ∗
3 + pp∗,
and span{e, o1, o2, o3} = span{f1, f2, f3, p}.
(2) Suppose e1, e2, e3, o ∈ Zd
2 are four vectors such that e1, e2, e3 are even and o is
odd. Then there exists an even vector f and odd vectors p1, p2, p3 such that
e1e∗
1 + e2e∗
2 + e3e∗
3 + oo∗ = f f ∗ + p1p∗
1 + p2p∗
2 + p3p∗
3,
and span{e1, e2, e3, o} = span{f, p1, p2, p3}.
Proof. For part (1), let f1 = e + o1 + o2, f2 = e + o1 + o3, f3 = e + o2 + o3, and
p = o1 + o2 + o3. By Lemma 4.2, f1, f2 and f3 are even and p is odd. For part (2),
let f = e1 + e2 + e3, and p1 = e1 + e2 + o, p2 = e1 + e3 + o, p3 = e2 + e3 + o. Lemma
4.2 implies f is even and p1, p2, p3 are odd. Easy computations show that the given
equalities are satisfied.
(cid:3)
We are now ready for the binary analog of Theorem 1.5: necessary and sufficient
conditions on pairs of matrices S and vectors α such that S is the frame operator of
a frame with vector "norms" determined by α. The necessary condition is easy.
Theorem 4.14. Let F = {fi}K
vector in ZK
2 defined by α[i] = (fi, fi) for each i. Then kαk0 ≡ Tr(S) (mod 2).
i=1 be a frame such that S = ΘF Θ∗
F . Let α be the
Proof.
kαk0 ≡ Tr(Θ∗
F ΘF ) ≡ Tr(ΘF Θ∗
F ) ≡ Tr(S) (mod 2).
(cid:3)
Sufficiency breaks down into three possible scenarios.
If S is parity indicative,
then a minimal frame F with frame operator S must consist of only odd vectors
or can attain any nonzero vector α with kαk0 ≡ Tr(S) (mod 2) in the sense that
(fi, fi) = α[i] for each i; if S is not parity indicative, a minimal frame must contain at
least one even vector. This is shown in Theorems 4.15 and 4.17. Nonminimal frames
can be constructed to correspond to any nonzero α with kαk0 ≡ Tr(S) (mod 2) if S
is parity indicative or to any such α with at least one zero entry if S is not parity
indicative (Corollary 4.16 and Theorem 4.17).
The frame elements can be permuted in any way without affecting the frame op-
F = S and Θ eF = ΘF P ∗ for some permutation matrix P ,
Indeed, if ΘF Θ∗
erator.
then
Θ eF Θ∗
eF = ΘF P ∗P Θ∗
F = ΘF Θ∗
F = S.
Therefore, in what follows, we need only construct frames with the correct number of
odd elements, corresponding to kαk0, in order to attain the dot products prescribed
by α.
Theorem 4.15. Let S be a d × d parity indicative symmetric matrix. Let K =
2d − rank(S), and let α ∈ ZK
(1) Suppose diag(S) 6= ~0.
2 be a nonzero vector with kαk0 ≡ Tr(S) (mod 2).
14
V. FURST AND E. P. SMITH
(a) If rank(S) = d, there exists a (minimal) K-element frame F such that
S = ΘF Θ∗
F and (fi, fi) = α[i] for every i only if kαk0 = K.
(b) If rank(S) < d, there exists a (minimal) K-element frame F such that
S = ΘF Θ∗
F and (fi, fi) = α[i] for every i.
(2) Suppose diag(S) = ~0. Then rank(S) < d.
(a) If rank(S) = d − 1, there exists a (minimal) K-element frame F such
that S = ΘF Θ∗
F and (fi, fi) = α[i] for every i only if kαk0 = K.
(b) If rank(S) < d − 1, there exists a (minimal) K-element frame F such
that S = ΘF Θ∗
F and (fi, fi) = α[i] for every i.
Proof. In case (1a), Theorem 4.5 implies the existence of a K = d - element frame
F whose frame operator is S. The d columns of ΘF must be linearly independent,
so every fi must be odd, by Lemma 4.11. (As a corollary of Theorem 4.14, we note
that d ≡ Tr(S) (mod 2) for any d × d, full rank, parity indicative symmetric matrix
S with diag(S) 6= ~0.)
For case (1b), instead of using the result of Theorem 4.9, we rely directly on The-
orem 4.3 to construct a d × rank(S) matrix A with rank(A) = rank(S) such that
AA∗ = S. Since the columns of A are linearly independent, Lemma 4.11 implies that
they are all odd. As in the proof of Theorem 4.9, we consider an augmented matrix
ΘF =(cid:2) A B (cid:3)
but with more care taken in the choice of B. By letting the columns of B be d−rank(S)
of the standard basis vectors not in the column space of A, each repeated twice, we
construct a frame F for kαk0 = 2d−rank(S). Replacing any identical pair of columns
of B, say {εl, εl}, with {εl +εn, εl +εn} for any other basis vector εn 6= εl, the columns
of ΘF still span, ΘF Θ∗
F is still equal to S, but now F contains two fewer odd vectors.
In this way, we are able to construct a frame F with dot products (fi, fi) satisfying any
kαk0 = rank(S) + 2m for 0 ≤ m ≤ d − rank(S). (Note that, by the proof of Theorem
4.14, rank(S) ≡ Tr(S) (mod 2).) By Lemma 4.13, any four vectors consisting of three
odds and one even can be substituted by four vectors consisting of three evens and
one odd, having the same span and no effect on ΘF Θ∗
F . Each substitution allows
us to increase the number of even vectors by two, until only two odd vectors remain
in F if rank(S) is even or one odd vector remains if rank(S) is odd. Therefore, we
can build a frame with 2d − rank(S) elements, corresponding to any nonzero α with
kαk0 ≡ rank(S) ≡ Tr(S) (mod 2).
Now let S be parity indicative with diag(S) = ~0. Since every row of S is even,
rank(S) < d. In case (2a), Theorem 4.9 implies the existence of a (d + 1)-element
frame F whose frame operator is S. Since F is a spanning set, it must contain an odd
vector. By Lemma 4.12, every vector in F must be odd. (Note that, by Corollary
4.8, case (2a) can only occur if rank(S) is even, hence d is odd.)
Lastly, we assume in case (2b) that rank(S) ≤ d − 2. By Theorem 4.3 and Proposi-
tion 4.4, there exists a d × (rank(S) + 1) matrix A with rank(A) = rank(S) such that
AA∗ = S. By Lemma 4.12, either every column of A is even or every column is odd.
Since Sii = 0 for every i, every row of A is even, hence the sum of the columns of A
is ~0. Moreover, since diag(S) = ~0, we know that rank(S) must be even, by Corollary
4.8. If every column of A were odd, then the sum of all rank(S) + 1 columns would
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
15
have to be odd, by Lemma 4.2, yielding a contradiction. So every column of A must
be even.
n ∈ span{b∗
j : b∗
2
n of B such that b∗
Augment A with a column of zeros and call the resulting matrix B. Then each
column of B is even, each row of B is even, and B has an even number of columns.
Consider a row b∗
j is a row of B and j 6= n}. Replace
n by its complement (that is, add ~1∗ for ~1 ∈ Zrank(S)+2
b∗
n), and call the resulting
matrix C. Then CC ∗ = S, rank(C) = rank(S) + 1, and C is composed of rank(S) + 2
odd columns. As in case (1b), we now augment C with d−(rank(S)+1) of the standard
basis vectors not in the column space of C, each repeated twice, to construct ΘF . In
doing so, we construct a frame F consisting of rank(S) + 2 + 2(d − (rank(S) + 1)) =
2d − rank(S) vectors, with frame operator S, such that every element of F is odd.
By Theorem 4.14, kαk0 must be even. As in case (1b), we can replace pairs of odd
elements of F by even vectors until only two odd vectors remain.
(cid:3)
to b∗
Corollary 4.16. Let S be a d × d parity indicative symmetric matrix. Let K >
2d − rank(S). Let α ∈ ZK
2 be a nonzero vector such that kαk0 ≡ Tr(S) (mod 2). Then
there exists a K-element frame F such that S = ΘF Θ∗
F and (fi, fi) = α[i] for every
i.
Proof. Since S is parity indicative, a K-element frame with K > 2d − rank(S) is
necessarily non-minimal and can be constructed by augmenting the minimal frames of
the previous theorem. Consider first the minimal frame F guaranteed by case (1a) of
Theorem 4.15. Adding the zero vector to F allows us to apply Lemma 4.13 and create
frames satisfying (fi, fi) = α[i] for any kαk0 ≡ Tr(S) (mod 2) with 0 < kαk0 < d.
Similarly, for case (2a), including the zero vector allows the construction of a frame
corresponding to any kαk0 = 2, 4, 6, . . . d + 1.
In either case, the addition of two
identical copies of odd vectors or two identical copies of even vectors provides frames
for any kαk0 ≡ Tr(S) (mod 2) when kαk0 ≥ d + 2 (case (1a)) or kαk0 ≥ d + 3 (case
(2a)). Similarly in cases (1b) and (2b), the addition of two identical copies of an
odd vector or two identical copies of an even vector yield frames for 2d − rank(S) <
kαk0.
(cid:3)
Theorem 4.17. Let S be a d × d symmetric matrix that is not parity indicative. Let
K ≥ 2d − rank(S) or if rank(S) = d and diag(S) = ~0, let K ≥ d + 1 . Let α ∈ ZK
2
be a nonzero vector such that kαk0 ≡ Tr(S) (mod 2). Then there exists a K-element
frame F such that S = ΘF Θ∗
F and (fi, fi) = α[i] for every i only if kαk0 6= K.
Proof. We use Theorem 4.5 and Remark 4.6 or Theorem 4.9 to construct a K-element
frame F such that ΘF Θ∗
F = S. By Lemma 4.11, F must contain an even vector. Of
course, F must also contain an odd vector, in order to span. Let m ≡ Tr(S) (mod 2)
represent the number of odd elements of F and K −m be the number of even elements.
By Lemma 4.13, F may be replaced by a frame with two more or two fewer odd
vectors. Through repeated applications, we can construct a frame F corresponding
to any kαk0 = 1, 3, 5, . . . , K −1 if m is odd and K is even, any kαk0 = 1, 3, 5, . . . , K −2
if m is odd and K is odd, any kαk0 = 2, 4, 6, . . . , K − 1 if m (≥ 2) is even and K
(> m) is odd, or any kαk0 = 2, 4, 6, . . . , K − 2 if m (≥ 2) is even and K (> m) is
even.
(cid:3)
16
V. FURST AND E. P. SMITH
5. Examples and Data
5.1. Examples. In this subsection we consider two symmetric matrices S and build
frames with various α's to illustrate the main result of Section 4. The algorithm for
factoring a matrix as S = AA∗ and for reducing A into a minimal factor can be found
in [17].
Example 5.1. Consider the identity matrix
S = I4 =
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
and note that it is a symmetric, parity indicative, full rank matrix. Any frame with
frame operator S is a Parseval frame. By Theorem 4.15, a minimal 4-element such
Parseval frame must satisfy kαk0 = 4 where α[i] = (fi, fi) for each i; clearly, this
follows from the Parseval frame necessarily being an orthonormal basis. Corollary
4.16 guarantees Parseval frames in Z4
2 of length K = 5 with either 2 or 4 odd vectors
corresponding to any α ∈ Z5
2 with kαk0 = 2, 4. To begin the construction, factor S
as I4I ∗
4 . Appending the zero-column to the left factor yields the matrix
ΘF =
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
,
the columns of which constitute a frame with frame operator S and α = (1, 1, 1, 1, 0).
To obtain any other α ∈ Z5
2 with kαk0 = 4, simply permute the columns.
We utilize Lemma 4.13 to reduce the number of odd vectors by 2. Let e be the
zero-column and o1, o2, o3 be the first, second, and third columns of ΘF , respectively.
Replacing e, o1, o2, o3 with their counterparts constructed in Lemma 4.13 results in
ΘF ′ =
1 1 0 0 1
1 0 1 0 1
0 1 1 0 1
0 0 0 1 0
.
Taking the columns of ΘF ′ as frame vectors builds the frame F ′ satisfying α =
(0, 0, 0, 1, 1). Again, the columns of F ′ can be permuted to acquire any α ∈ Z5
2 with
kαk0 = 2. Notice that a permutation of F ′ appears in the proof of Theorem 3.5.
Example 5.2. Suppose we wish to find a frame for Z3
2 of length 7 with frame operator
S =
0 1 0
1 1 1
0 1 0
.
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
17
Since this rank 2, symmetric matrix is not parity indicative, we apply Theorem 4.17.
In doing so, we follow the proof of Theorem 4.9 and factor S as
S = P ∗T(cid:20) L 0
0 0 (cid:21) T ∗P
where P is the 3 × 3 identity matrix,
L =(cid:20) 0 1
1 1 (cid:21) ,
and
T =
1 0 0
0 1 0
1 0 1
.
Then
A =(cid:20) H 0
0 B (cid:21) =
1 1 0 0
1 0 0 0
0 0 1 1
,
and we append three zero-columns to T A to build the frame F :
ΘF =hT A ~0 ~0 ~0i =
1 1 0 0 0 0 0
1 0 0 0 0 0 0
1 1 1 1 0 0 0
.
Letting (fi, fi) = α[i] for each i, we see that F satisfies α = (1, 0, 1, 1, 0, 0, 0). We in-
crease or decrease the number of odd vectors as desired, by applying Lemma 4.13 first
to {f1, f3, f4, f5} and then to {f4, f5, f6, f7}. We obtain frames F1 and F2 satisfying
ΘF1 =
ΘF2 =
1 1 1 0 1 0 0
1 0 1 0 1 0 0
0 1 0 0 1 0 0
1 1 0 0 0 0 0
1 0 0 0 0 0 0
1 1 1 0 1 1 1
, α1 = (0, 0, 0, 0, 1, 0, 0);
, α2 = (1, 0, 1, 0, 1, 1, 1).
By Theorem 4.17, kαk0 = 7 is unattainable.
j=1 in Zd
5.2. Data. An exhaustive search for frame operators ΘF Θ∗
F and kαk0 associated
with F = {fj}K
2 was performed, using Python 3.6, for various dimensions
and frame lengths (i.e. various d's and K's). The tables contained in this subsection
hold information about the number of symmetric matrices that are frame operators
and the set of kαk0 that accompany them. We include summaries for dimensions
d = 2, . . . , 5. Because every frame in Zd
2 must have at least d vectors, and because 2d
is the minimum number of vectors needed to ensure every symmetric matrix is a frame
operator (Theorems 4.5, 4.9), the computations were performed for K = d, . . . , 2d.
For d = 2, . . . , 5, in the table containing information about the d-dimensional binary
space, the entry in the row labeled {αmin, αmin + 2, . . . , αmin + 2t} and column labeled
K = k0 shows the number of symmetric matrices S in d-dimensional space such that
for each α ∈ {αmin, αmin + 2, . . . , αmin + 2t} there exists a frame F = {fj}k0
j=1 such
that S = ΘF Θ∗
F and α[i] = (fi, fi) for 1 ≤ i ≤ k0.
18
V. FURST AND E. P. SMITH
In each table, the sum of the entries of the last column represents all possible
symmetric d × d binary matrices. There are 2d(d+1)/2 such matrices, which becomes
prohibitively large as the dimension d increases.
{kαk0}
{1}
{2}
{1, 3}
{2, 4}
K
2 3 4
2 2 0
1 3 2
0 2 4
0 0 2
Table 1. Number of attainable frame
operators of frames for Z2
2
K
{kαk0}
5
3
4
12 0
0
{1}
12 21 0
{2}
4
0
{3}
0
0
{4}
0
{1, 3}
0
{2, 4}
{1, 3, 5} 0
{2, 4, 6} 0
6
0
0
0
0
1
0
28 24 0
6
31 24
32
8
0
0
0
8
Table 2. Number of attainable frame
operators of frames for Z3
2
{kαk0}
K
6
5
4
0
168 0
{2}
28
0
0
{4}
224 392 0
{1, 3}
0
{2, 4}
0
{1, 3, 5}
{2, 4, 6}
0
{1, 3, 5, 7} 0
{2, 4, 6, 8} 0
7
0
0
0
420 441 0
56
0
0
0
8
0
0
0
0
504 448 0
63
0
0
511 448
64
512
64
0
Table 3. Number of attainable frame operators of frames for Z4
2
{kαk0}
7
6
5
0
0
448
{5}
0
0
28
{6}
6720 0
0
{1, 3}
6720 13020 0
{2, 4}
0
{1, 3, 5}
0
{2, 4, 6}
0
{1, 3, 5, 7}
0
{2, 4, 6, 8}
{1, 3, 5, 7, 9}
0
{2, 4, 6, 8, 10} 0
K
8
0
0
0
0
13888 15120 0
840
0
0
0
0
9
0
0
0
0
0
15988 15345 0
1008
0
0
0
10
0
0
0
0
0
0
16368 15360 0
1023
0
0
16383 15360
16384
1024
0
1024
Table 4. Number of attainable frame operators of frames for Z5
2
BINARY FRAMES WITH PRESCRIBED DOT PRODUCTS AND FRAME OPERATOR
19
We thank Erich McAlister for his very valuable feedback.
Acknowledgement
References
[1] Z. Baker, B. Bodmann, S. Branum, M. Bullock, and J. McLaney, What is odd about binary
Parseval frames?, Involve (2017), in press.
[2] B. Bodmann, B. Camp, and D. Mahoney, Binary frames, graphs, and erasures, Involve 7:2
(2014), 151–169.
[3] B. Bodmann, M. Le, L. Reza, M. Tobin, and M. Tomforde, Frame theory for binary vector
spaces, Involve 2:5 (2009), 589–602.
[4] M. Bownik and J. Jasper, Existence of frames with prescribed norms and frame operator, pp.
103–117 in Excursions in harmonic analysis, Vol 4, Appl. Numer. Harmon. Anal. New York:
Birkhauser/Springer, 2015.
[5] J. Cahill, M. Fickus, D. Mixon, M. Poteet, and N. Strawn, Constructing finite frames of a given
spectrum and set of lengths, Appl. Comput. Harmon. Anal. 35:1 (2013), 52–73.
[6] P. Casazza and G. Kutyniok, ed. Finite frames: theory and applications, Appl. Numer. Harmon.
Anal. New York: Birkhauser/Springer, 2013.
[7] P. Casazza, M. Fickus, J. Kovacevi´c, M. Leon, and J. Tremain, A physical interpretation of
tight frames, pp. 51–76 in Harmonic analysis and applications, Appl. Numer. Harmon. Anal.
Boston: Birkhauser, 2006.
[8] P. Casazza and M. Leon, Existence and construction of finite frames with a given frame operator,
Int. J. Pure Appl. Math. 63:2 (2010), 149–157.
[9] O. Christensen, An introduction to frames and Riesz bases, Appl. Numer. Harmon. Anal.
Boston: Birkhauser, 2003.
[10] O. Christensen, A. Powell, and X. Xiao, A note on finite dual frame pairs, Proc. Amer. Math.
Soc. 140:11 (2012), 3921–3930.
[11] G. Cohen, I. Honkala, S. Litsyn, and A. Lobstein, Covering Codes, North-Holland Math. Li-
brary. New York: Elsevier, 1997.
[12] M. Fickus, D. Mixon, M. Poteet, and N. Strawn, Constructing all self-adjoint matrices with
prescribed spectrum and diagonal, Adv. Comp. Math. 39:3-4 (2013), 585–609.
[13] D. Han, K. Kornelson, D. Larson, and E. Weber, Frames for undergraduates, Student Math.
Library. Providence: American Mathematical Society, 2007.
[14] R. Horn and C. Johnson, Matrix Analysis. New York: Cambridge University Press, 1985.
[15] R. Hotovy, D. Larson, and S. Scholze, Binary frames, Houston J. Math. 41:3 (2015), 875–899.
[16] J. Kovacevi´c and A. Chebira, Life Beyond Bases: The Advent of Frames (Part II). IEEE SP
Mag. 24:5 (2007), 115–125.
[17] A. Lempel, Matrix factorization over GF (2) and trace-orthogonal bases of GF (2n)∗, SIAM J.
Comput. 4:2 (1975), 175–186.
Veronika Furst, Department of Mathematics, Fort Lewis College, Durango, Col-
orado 81301, USA
E-mail address: furst [email protected]
Eric P. Smith, Department of Mathematics, Fort Lewis College, Durango, Col-
orado 81301, USA
E-mail address: [email protected]
|
1705.02625 | 1 | 1705 | 2017-05-07T14:01:50 | Strongly extreme points and approximation properties | [
"math.FA"
] | We show that if $x$ is a strongly extreme point of a bounded closed convex subset of a Banach space and the identity has a geometrically and topologically good enough local approximation at $x$, then $x$ is already a denting point. It turns out that such an approximation of the identity exists at any strongly extreme point of the unit ball of a Banach space with the unconditional compact approximation property. We also prove that every Banach space with a Schauder basis can be equivalently renormed to satisfy the sufficient conditions mentioned. In contrast to the above results we also construct a non-symmetric norm on $c_0$ for which all points on the unit sphere are strongly extreme, but none of these points are denting. | math.FA | math | STRONGLY EXTREME POINTS AND
APPROXIMATION PROPERTIES
TROND A. ABRAHAMSEN, PETR H ´AJEK, OLAV NYGAARD,
AND STANIMIR TROYANSKI
Abstract. We show that if x is a strongly extreme point of a
bounded closed convex subset of a Banach space and the identity
has a geometrically and topologically good enough local approxi-
mation at x, then x is already a denting point. It turns out that
such an approximation of the identity exists at any strongly ex-
treme point of the unit ball of a Banach space with the uncondi-
tional compact approximation property. We also prove that every
Banach space with a Schauder basis can be equivalently renormed
to satisfy the sufficient conditions mentioned. In contrast to the
above results we also construct a non-symmetric norm on c0 for
which all points on the unit sphere are strongly extreme, but none
of these points are denting.
1. Introduction
Let X be a (real) Banach space and denote by BX its unit ball, SX
its unit sphere, and X ∗ its topological dual. Let A be a non-empty set
in X. By a slice of A we mean a subset of A of the form
S(A, x∗, ε) := {x ∈ A : x∗(x) > M − ε}
where ε > 0, x∗ ∈ X ∗ with x∗ 6= 0, and M = supx∈A x∗(x). We will
simply write S(x∗, ε) for a slice of a set when it is clear from the setting
what set we are considering slices of.
Definition 1.1. Let B be a non-empty bounded closed convex set in
a Banach space X and let x ∈ B. Then x
a) is an extreme point of B if for any y, z in B we have
x =
y + z
2 ⇒ y = z = x.
7
1
0
2
y
a
M
7
]
.
A
F
h
t
a
m
[
1
v
5
2
6
2
0
.
5
0
7
1
:
v
i
X
r
a
b) is a strongly extreme point of B if for any sequences (yn)∞
n=1, (zn)∞
n=1
in B we have
n kx −
lim
yn + zn
2
k = 0 ⇒ lim
n kyn − znk = 0.
2010 Mathematics Subject Classification. Primary: 46B20; Secondary: 46B04.
Key words and phrases. denting point, strongly extreme point.
The second author was financially supported by GACR 16-073785 and RVO:
67985840. The fourth author was partially supported by MTM2014-54182-P and
the Bulgarian National Scientific Fund under Grant DFNI-I02/10.
1
2
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
When B is the unit ball, the above condition can be replaced
by
n kx ± xnk = 1 ⇒ lim
lim
n kxnk = 0.
In this case we say that the norm is midpoint locally uniformly
rotund (MLUR) at x.
c) is a point of continuity for the map Φ : B → X if Φ is weak to
norm continuous at x. When Φ is the identity mapping we just
say that x is a point of continuity (PC).
d) is a denting point of B if for every ε > 0 and δ > 0 there exists
a slice S(x∗, δ) of B with diameter less than ε.
e) is a locally uniformly rotund (LUR) point of BX if for any
sequence (xn)∞
n kx + xnk = 2 lim
lim
n=1 we have
n kxnk = 2kxk = 2 ⇒ lim
n kx − xnk = 0.
It is well known that LUR points are denting points and that denting
points are strongly extreme points [KR82]. Trivially strongly extreme
points are extreme points.
The importance of denting points became clear in the sixties when
the Radon-Nikod´ym Property (RNP) got its geometric description. In
particular, it became clear that extreme points in many cases are al-
ready denting as every bounded closed convex set in a space with the
RNP has at least one denting point. The "extra" an extreme point
needs to become a denting points is precisely described in the following
Theorem 1.2. [LLT88] Let x be an extreme point of continuity of a
bounded closed convex set C in X. Then x is a denting point of C.
It is well known that all points of the unit sphere of ℓ1 are points of
continuity for the unit ball Bℓ1. So from Theorem 1.2 we get that every
extreme point of the unit ball of any subspace of ℓ1 automatically gets
the "extra" to become denting.
However, despite the theoretical elegance of Theorem 1.2, it is not
always easy to to check whether the identity mapping is weak to norm
continuous at a certain point of a bounded closed convex set. For this
reason it is natural to look for geometrical conditions which ensure weak
to norm continuity of the identity operator at x when we approximate
it strongly by maps that are weak to norm continuous at x.
One such idea could be to assume that x is strongly extreme (not just
extreme as in Theorem 1.2) and that the identity map is approximated
strongly by finite rank operators. But this is not enough to give the
extreme point the "extra" needed to be denting: Consider x = x(t) ≡
1 ∈ BC[0,1]. Then x is strongly extreme in BC[0,1], but the identity map
I : BC[0,1] → BC[0,1] is not weak to norm continuous there (see the next
paragraph), and limn kPnx − xk = 0, where (Pn) are the projections
corresponding to the Schauder basis in C[0, 1]. Clearly Pn is weak to
norm continuous at any point of BC[0,1] (as any compact operator is).
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
3
Actually C[0, 1] belongs to the class LD2 of Banach spaces where all
slices of the unit ball have diameter 2. Naturally, in such spaces no
point of the unit sphere can be a PC point. See e.g. the references in
[AHN+16] for more information about the class LD2.
Assuming x is strongly extreme, we need to make stronger assump-
tions of the approximating sequence of the identity. One such condition
which we impose is related to the behaviour of the approximating map-
pings close to the point x. In particular we obtain as a corollary that
in Banach spaces with the unconditional compact approximation prop-
erty (UKAP) (see Definition 2.8), every strongly extreme point in the
unit ball is PC and therefore denting. In particular, we have that this
conclusion holds for Banach spaces with an unconditional basis with
unconditionally basis constant 1. Further we show that every Banach
space with a Schauder basis can be renormed to satisfy the conditions
of Theorem 2.1.
Nevertheless we construct, in Section 3, a closed convex body in c0
where all boundary points are strongly extreme, but none of them is
denting. This body is not symmetric; we refer to its gauge as a non-
symmetric equivalent norm on c0. Thus, we construct a non-symmetric
equivalent norm on c0 for which all points on the unit sphere are
strongly extreme points, but none of these points is denting. In fact
every slice of the unit ball of c0 with this non-symmetric norm has
diameter at least 1/√2.
In Section 4 we investigate when equivalent norms conserve strongly
extreme and denting points of the corresponding unit balls.
The notation and conventions we use are standard and follow [JL01].
When considered necessary, notation and concepts are explained as the
text proceeds.
2. Weak to norm continuity of the identity map
Our most general result on how to force a strongly extreme point x
to be denting in terms of approximating the identity map I : X → X
at x is the following
Theorem 2.1. Let x be a strongly extreme point of a non-empty bounded
closed convex set C in a Banach space X. Let x be a point of continuity
for a sequence Φn : C → X, n = 1, 2, . . . of maps such that
(1)
n kΦnx − xk = 0
lim
and
(2)
where
lim
n
lim
ε→0+
fn(ε) = 0,
fn(ε) = sup{dist((1 + λ)Φny − λy, C) : y ∈ C,kΦnx − Φnyk ≤ ε},
for some λ ∈ (0, 1]. Then x is a denting point of C.
4
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
The proof follows from Theorem 1.2 and the next proposition which
is an interplay between weak and norm topology. With B(x, ρ) we
denote the ball with center at x and radius ρ.
Proposition 2.2. Let x be a strongly extreme point of a convex set C
of a normed space X and let 0 < λ ≤ 1. Assume that for every η > 0
there exist a weak neighbourhood W of x and a map Φ : W ∩ C → X
such that
(3)
and
(4)
Then x is PC.
Φ(W ∩ C) ⊂ B(x, η)
sup
w∈W ∩C
dist((1 + λ)Φw − λw, C) < η.
Proof. Since x is strongly extreme point, for every ε > 0 we can find
δ > 0 such that
kx −
u, v ∈ C ⇒ ku − vk < λε.
(5)
u + v
2 k < δ,
Set η = min{δ, λε}/2. There is a weak neighbourhood W of x and a
map Φ satisfying (3) and (4). Set Ψ = I − Φ and pick an arbitrary
w ∈ W ∩ C. Put y+ = (1 − λ)x + λw. Since x, w ∈ C we get y+ ∈ C
by convexity. Since
Φw + λΨw − y+ = (1 − λ)(Φw − x),
we have from (3)
kΦw + λΨw − y+k ≤ (1 − λ)η < η.
(6)
Having in mind (4) we can find y− ∈ C such that
k(Φw − λΨw) − y−k < η.
(7)
This and (6) imply
kx −
y+ + y−
2
1
2k(Φw + λΨw − y+) + (Φw − λΨw − y−)k
k ≤ kx − Φwk +
< kx − Φwk + η ≤ 2η.
From (5) we get
On the other hand, using again (6) and (7), we get
ky+ − y−k < λε.
ky+ − y−k = ky+ − (Φw + λΨw) − y− + (Φw − λΨw) + 2λΨwk
> 2λkΨwk − 2η.
Hence
2λkΨwk < ky+ − y−k + 2η < λε + 2λε = 3λε.
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
5
This and (3) imply
kw − xk ≤ kΦw − xk + kΨwk < 2ε.
Since w is an arbitrary element of W ∩C we get that W ∩C ⊂ B(x, 2ε).
(cid:3)
Remark 2.3. If x is PC for C we get that x satisfies the hypotheses of
Proposition 2.2 just taking Φ = I, λ ∈ (0, 1].
Proof of Theorem 2.1. Let {εn} be a sequence of positive numbers tend-
ing to 0. Since Φn : C → X, n = 1, 2, . . . is weak to norm continuous
at x there is weak neighbourhood Vn of x such that
Φn(Vn ∩ C) ⊂ B(x, εn), n = 1, 2, . . .
Thus the conditions of Theorem 2.1 imply that for every η > 0 we can
find n = n(η) such that (3) and (4) hold for W = Vn and Φ = Φn. (cid:3)
Recall that every linear compact operator is weak to norm continuous
on bounded sets. This together with Theorem 2.1 gives
Corollary 2.4. Let (X,k · k) be a Banach space and let · be an
equivalent (not necessarily symmetric) norm on X with corresponding
unit ball C. Let x be a strongly extreme point of C. Let Tn : X → X
be linear compact operators such that
(8)
(9)
n kTnx − xk = 0,
lim
lim
n
lim
ε→0+
sup{(1 + λ)Tny − λy : y ≤ 1,kTn(x − y)k ≤ ε} = 1,
for some λ ∈ (0, 1]. In particular the above is satisfied if
(10)
n (1 + λ)Tn − λI = 1.
lim
Then x is a denting point of C.
Proof. It is enough to prove that (9) implies (2). Indeed, since there
exists k > 0 such that k · k ≤ k·, then for every u ∈ X \ C we have
dist(u, C) = inf{ku − vk : v ∈ C} ≤ k inf{u − v : v ∈ C}
u −
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
≤ k(cid:12)(cid:12)(cid:12)(cid:12)
u
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
u(cid:12)(cid:12)(cid:12)(cid:12)
= k(u − 1).
(cid:3)
Remark 2.5. The functions fn defined in Theorem 2.1 can be discon-
tinuous at 0. Indeed, X be a Banach space, e ∈ BX , and e∗ ∈ SX ∗ be
such that e∗(e) = kek = ke∗k = 1. Define a (norm one) projection P
on X by P x = e∗(x)e and put
f (ε) = sup{kP y − Ryk : kyk ≤ 1,kP (e − y)k ≤ ε},
6
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
where R = I − P . Now, if the norm k · k on X is either strictly
convex or Gateaux differentiable at e∗, then f is discontinuous at 0.
Indeed, let ε = 0,kyk ≤ 1, and P (e − y) = 0. We get e∗(e − y)e = 0.
Hence e∗(y) = e∗(e) = 1. By the strict convexity of the norm or the
Gateaux differentiability of the norm at e∗, we have y = e. This implies
Ry = Re = 0, so f (0) = 1. In order to prove that f is discontinuous
at 0 we simply apply Corollary 2.4 with Tn = P . Since e is strongly
extreme, but not denting, we get limε→0+ f (ε) > 1.
Remark 2.5 shows that one cannot replace the limit condition (9) of
Corollary 2.4, by
lim
n
sup{(1 + λ)Tny − λy : y ≤ 1, Tnx = Tny} = 1.
The conditions in Corollary 2.4 are essential. Let us illustrate this
by examples.
Example 2.6. Consider the space c of convergent sequences endowed
with its natural norm. Let e = (1, 1, . . .) ∈ Sc and Pn the projection
on c which projects vectors onto their n first coordinates. Clearly e
is a strongly extreme point of Bc which is not denting. Moreover,
it is evident that the condition limn kPne − ek = 0 fails and that the
condition (9) (moreover (8)) holds for λ = 1 (and thus for all λ ∈ (0, 1]).
It follows that the approximation condition in Corollary 2.4 is essential.
Example 2.7. Consider again c endowed with its natural norm. Let
e ∈ c be is as in Example 2.6. Define a projection P on c by P x =
limn x(n)e and put Pn = P for all n. By construction Pne = e. For
z = (0, 1, 1, . . .) we have Pnz = P z = e. Now, for any λ ∈ (0, 1] we have
k(1 + λ)Pnz − λzk = k(1 + λ)e − λzk
= 1 + λ.
Thus
lim
n
lim
ε→0+
sup{k(1 + λ)Pny − λyk : y ∈ BX,kPn(e − y)k < εk} ≥ 1 + λ.
It follows that the condition (2) in Theorem 2.1 is essential.
We now present our results in terms of an approximation property
introduced and studied by Godefroy, Kalton and Saphar.
Definition 2.8. [GKS93] A Banach space X is said to have the un-
conditional compact approximation property (UKAP) if there exists a
sequence (Tn) of linear compact operators on X such that limn kTnx −
xk = 0 for every x ∈ X and limn kI − 2Tnk = 1.
Clearly Banach spaces X with the UKAP satisfy condition (10) for
λ = 1. Clearly also Banach spaces with an unconditional basis with
basis constant 1 have the UKAP (simply put Tn = Pn the projection
onto the n first vectors of the basis). Thus we immediately have the
following corollary.
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
7
Corollary 2.9. If X has the UKAP, in particular if X has an un-
conditional basis with unconditional basis constant 1, then all strongly
extreme points in BX are denting points.
Let us mention that the global condition (10) is much stronger than
the local condition (9), even in the case when it holds for all x in SX .
This will be clear from the discussion below and in particular from
Example 2.12 which shows that the condition (10) is strictly stronger
than (9). For that example we will use the following result.
Proposition 2.10. Let X be a Banach space and x a locally uniformly
rotund (LUR) point in SX. Let (Tn) be a sequence of linear bounded
operators on X, with limn kTnk = 1, and which satisfies condition (8)
in Corollary 2.4 for x ∈ SX . Then condition (9) holds for λ = 1 (and
thus for all λ ∈ (0, 1]).
Proof. Pick an arbitrary sequence (εn) with εn > 0 and limn εn = 0.
First we show that
(11)
lim
n
diamDn = 0,
where Dn = {y ∈ BX : kTn(x − y)k < εn}. To this end note that it
suffices to show that
(12)
yn ∈ BX,
n kTn(x − yn)k = 0 ⇒ lim
lim
n kx − ynk = 0.
Indeed,
kTnkkx + ynk ≥ kTn(x + yn)k
= k2Tnx + Tn(yn − x)k
≥ 2kTnxk − kTn(yn − x)k.
Hence lim inf n kx + ynk ≥ 2. Since kynk ≤ 1 we get limn kx + ynk = 2.
Since x is a LUR point, we get that (12) holds, and thus (11) holds. In
order to prove (9) for λ = 1, it is enough to show that
lim
n
dn = 1,
where dn = sup{kTnx − Rnyk : y ∈ Dn}, Rn = I − Tn. Since x ∈ Dn,
we have dn ≥ kTnx − Rnxk. So we get from (8) that lim inf dn ≥ 1.
Now, pick arbitrary y ∈ Dn. Then we have
kTnx − Rnyk ≤ kTnxk + kRnyk
≤ kTnxk + kRnxk + kRn(y − x)k
≤ kTnxk + kRnxk + kRnkk(y − x)k
≤ kTnxk + kRnxk + (kTnk + 1)diamDn.
Hence lim sup dn ≤ 1.
(cid:3)
8
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
Proposition 2.11. Let (Tn) be a bounded sequence of linear compact
operators on X, Rn = I − Tn, and (fn) ⊂ SX ∗ a separating family for
X. Then the norm
u = (cid:18) ∞
Xn=1
2−n(kRnuk2 + f 2
n(u))(cid:19)
1
2
is LUR at x ∈ X provided limn kRnxk = 0. Moreover, if the operators
(Tn) commute and limn kTnk = 1, then limn Tn = 1.
Proof. Pick a sequence (xk) ⊂ X with limk xk + x = x = xk.
By convex arguments ([DGZ93, Fact 2.3 p. 45]) we have
(13)
(14)
k kRnxkk = kRnxk, n = 1, 2, . . .
lim
fn(xk) = fn(x), n = 1, 2, . . . .
lim
k
First we show that (xk) is norm compact. Given ε > 0 we can find
n with kRnxk < ε. Using (13) we can find kε such that kRnxkk < ε for
k < kε. The set K = {x1, x2, . . . , xkε} ∪ kxkTn(BX) is norm compact.
We show that K is an ε-net for (xk). Indeed, pick xk, k > kε. Then
kxk − Tnxkk = kRnxkk < ε and Tnxk ∈ kxkTn(BX). So (xk) is norm
compact. Since (fn) is total, we get from (14) that limk kxk − xk = 0.
Thus the norm · is LUR at the point x ∈ X.
Now, let us prove the moreover part. As (Tn) commute, we have
∞
Tmu2 =
2−n(kRnTmuk2 + f 2
n(Tmu))
∞
=
Xn=1
Xn=1
Xn=1
≤
= kTmk2u2.
∞
2−n(kTmRnuk2 + (T ∗
mfn(u))2)
2−n(kTmk2kRnuk2 + kT ∗
mk2f 2
n(u))
Hence Tm2 ≤ kTmk2 for all m = 1, 2, . . . , so lim supm Tm ≤
lim supm kTmk. Since limm kTmx − xk = 0, we get lim inf m Tm ≥ 1,
and so limm Tm = 1 provided limm kTmk = 1.
(cid:3)
It is now easy to give the announced example which shows that
condition (10) can fail as condition (9) holds for every x in SX.
Example 2.12. Consider c0 endowed with the norm k · k defined by
kxk = supi,j≥1(x(i) − x(j)) where x = (x(k)) ∈ c0. Clearly k · k is
equivalent to the canonical norm on c0. Let Pn be the projection onto
the n first vectors in the canonical basis (ek) of c0 and let · be the
norm on c0 given in Proposition 2.11 where fn = 0 for every n. Then
(c0,·) fulfils the conditions of Proposition 2.10 and thus satisfies
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
9
condition (9) for every x in Sc0. Nevertheless we have Pk − λRk > 1
for any λ ∈ (0, 1], so condition (10) fails. For the latter, just consider
(Pk − λRk)(Pk+1
From the two preceding propositions we also get
i=1 ei).
Corollary 2.13. Let X be a Banach space with a Schauder basis. Then
there exists an equivalent norm k · k on X for which the sequence of
projections Pn onto the first n vectors of the basis, satisfy (9) for λ = 1.
On the other hand we have
Proposition 2.14. There exists an equivalent norm k · k on C[0, 1]
such that (9) does not hold for any λ > 0 and any sequence (Tn) of
compact linear operators on Xwhen x ∈ C[0, 1] with kxk = 1 and
limn kTnx − xk∞ = 0.
Proof. The norm on C[0, 1] constructed in [AHN+16, Theorem 2.4] is
midpoint locally uniformly rotund and has the diameter two property,
i.e. all non-empty relatively weakly open subsets of the unit ball have
diameter 2. In particular, in this norm all points on the unit sphere are
strongly extreme, but none are denting. Thus the conclusion follows
from Theorem 2.1.
(cid:3)
3. An MLUR non-symmetric norm on c0 without denting
points on the unit sphere
We will now construct a non-symmetric equivalent norm on c0 for
which all points on the unit sphere are strongly extreme, but none of
these points are denting.
Proposition 3.1. There exists in c0 an equivalent (non-symmetric)
MLUR norm · with no denting points. Moreover, every open slice
of the unit ball corresponding to · has diameter ≥ 1/√2.
Proof. With J we denote the set of all finite strictly increasing se-
quences (jk)k≥0 of natural numbers. For x = (x(k))k≥1 ∈ ℓ∞ and j ∈ J
we put
Q(x, j) = x(j0) +Xk≥1
2−kx+(jk), x+ = max{x, 0}
2−kx2(k))1/2, kxk = sup{Q(x, j) : j ∈ J},
q(x) = (Xk≥1
x = (kxk2 + q2(x))1/2.
Clearly
(15)
kxk∞ ≤ kxk ≤ x ≤ (kxk2 + kxk2
∞)1/2 ≤
√2kxk ≤ 3kxk∞.
10
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
Claim 3.2. For every x ∈ c0 and every ε > 0, there exists m =
m(x, δ) ∈ N and δ = δ(x, ε) > 0 such that
(16)
max{kx + yk,kx − yk} ≥ kxk + δ
whenever
(17)
y ∈ ℓ∞,kRmyk∞ > ε,
where Pm is the projection onto the first m vectors of the canonical
basis of c0 and Rm = I − Pm.
Proof. Pick m such that
(18)
and put δ = ε/2m+3. Using the definition of k·k we can find j = (ji)p
such that
kRmxk∞ < ε/8,
k=0
kxk − δ/2 < Q(x, j).
(19)
We may assume that p ≥ m. Choose i ≥ 1 such that
(20)
and put j1 = (jk)i−1
(21)
Pick y satisfying (17). There is r > m with y(r) > ε. We have
(22)
k=0. Using (20) we get
Q(x, j) ≤ Q(x, j1) + 2−i+1kRmxk∞.
ji−1 ≤ m < ji
(x(r) + y(r))+ + (x(r) − y(r))+ ≥ y(r) − x(r) > ε − kRmxk∞.
Put j2 = (j0, j1, . . . , ji−1, r). Since r > m ≥ ji−1 we have j2 ∈ J. So
kx ± yk ≥ Q(x ± y, j2) = Q(x±, j1) + 2−i(x(r) ± y(r))+.
Since i ≤ ji−1 we get from (20), (21), and (22)
(kx + yk + kx − yk)/2 ≥ Q(x, j1) + 2−i−1(ε − kRmk∞)
≥ Q(x, j) + 2−i−1(ε − 5kRmxk∞)
≥ Q(x, j) + 2−m−1(ε − 5kRmxk∞).
This and (18), (19) imply (16).
(cid:3)
Let x ∈ c0, xn ∈ ℓ∞, and limn x ± xn = x. By convex arguments
as in [DGZ93, p. 45] we have
(23)
and
(24)
n kx ± xnk = kxk,
lim
lim
n
xn(k) = 0, k = 0, 1, . . . .
We have to show that limn kxnk∞ = 0. Assume the contrary. Then we
may assume that kxnk∞ ≥ 2ε > 0 for all n = 1, 2, . . . . From (24) it
follows that there exists m such that kPmxnk∞ < ε for all n which are
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
11
sufficiently big. But, this contradicts (23) and Claim 3.2. Hence · is
MLUR.
Now, put C = {u ∈ c0 : u ≤ 1}. We will show that every non-void
slice of C has k · k∞ diameter ≥ 1/√2. To this end, let f ∈ ℓ1, a ∈
R, H = {u ∈ c0 : f (u) > a}, and S = H ∩ C 6= ∅.
Claim 3.3. For every x ∈ H with x < 1, there is y ∈ S with
x − y ≥ kxk.
Proof. Choose δ > 0 such that
(kxk + 2δ)2 + q2(x) + δ < 1, f (x) − δ > a.
There exists a natural number m such that
kRmxk∞ < δ,
(26)
0 < kxk − x(m) < √2mδ,
(25)
(27)
and
f (em) < δ
(28)
where em is the standard basis vector number m in c0. Put y = x −
kxkem. Pick arbitrary j ∈ J. We will show that
(29)
Q(y, j) ≤ kxk + 2δ.
If m < j0, then y(jk) = x(jk) for k = 0, 1, 2, . . . . So
Q(y, j) = Q(x, j) ≤ kxk.
If m > j0, we get Q(y, j) ≤ Q(x, j) ≤ kxk since y+(m) = 0. If m = j0
Q(y, j) ≤ kxk − x(m) + kRmxk∞.
This and (26) imply (29). From (27) we get q2(y) < q2(x) + δ. From
this, (25), and (29) we have y < 1. Since kxk < 1 we get from (28)
that y ∈ H. Thus y ∈ S, and kx − yk∞ = kxk. Using (15) we get
x − y ≥ kx − yk∞ = kxk, and the claim follows.
Since x ∈ S can be chosen with ·-norm arbitrarily close to 1 we
get from (15) that the diameter of S ≥ 1/√2.
(cid:3)
(cid:3)
4. Weak continuous perturbation of the norm
It is natural to expect that weak continuous perturbations of the
norm would preserve points of continuity. We will show that simi-
lar perturbations of the norm preserve strongly extreme points of the
corresponding ball.
Theorem 4.1. Let X be a Banach space and let B be the unit ball
corresponding to an equivalent norm · on X. Assume that the re-
striction ·SX of · to SX is continuous at x ∈ SX
12
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
a) with respect to the weak topology. Then x is a point of continuity
for BX provided x is a point of continuity for B.
b) with respect to a topology σ = σ(X, Y ), Y ⊂ X ∗ with the prop-
erty that for every bounded sequence in X there exists a σ-
Cauchy subsequence, and let x be an extreme point of BX. Then
x is a strongly extreme point of BX provided it is a strongly ex-
treme point for B.
Proof. From the assumption x = 1.
a). Pick ε > 0. Since x is a point of continuity for B, there is a
weakly open set W containing x such that diam(W ∩ B) < ε. Since ·
is uniformly norm continuous, we can find δ > 0 such that diam(W ∩
(1 + 2δ)B) < 2ε. Since ·SX is weakly continuous at x, we can find a
weakly open set V containing x such that V ∩ SX ⊂ (1 + δ)B. Since
k · k is uniformly norm continuous, we can find η ∈ (0, 1) such that
V ∩ (BX \ ηBX ) ⊂ (1 + 2δ)B. Finally there is a weakly open set U
containing x such that U ∩ ηBX = ∅. Hence x ∈ U ∩ V ∩ W and
diam(U ∩ V ∩ W ∩ BX ) < 2ε.
b). Pick an arbitrary sequence (un) ⊂ X with limn kx ± unk = 1.
In order to prove that limkunk = 0, it is enough to show that there
exists a subsequence of (un) converging to zero. To this end we assume
without loss of generality that
σ − lim
(um − un) = 0.
m,n
Taking into account that
kx ±
um − un
2
k ≤ kx ± umk + kx − unk
2
,
we get
m,n kx ±
lim
um − un
2
k = 1.
Put um,n = (um−un)/2, α±
and σ−limm,n(x±um,n)/(1+α±
Since ·SX is σ-continuous at x, we get
m,n = kx±um,nk−1. Clearly limm,n α±
m,n) = x, and (x±um,n)/(1+α±
m,n = 0
m,n) ∈ SX .
lim
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
m,n(cid:12)(cid:12)(cid:12)(cid:12)
= x = 1.
x ± um,n
1 + α±
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
m,n(cid:12)(cid:12)(cid:12)(cid:12)
This implies limm,n x ± um,n = 1. Having in mind that x is a strongly
extreme point of B, we get limm,n um,n = 0. Hence (un) is a norm
Cauchy sequence. It follows that there is u ∈ X with limn kun−uk = 0.
Thus kx ± uk = 1, which implies that u = 0 as x is an extreme point.
Now limn kunk = 0 which finishes the proof.
Remark 4.2. The requirement that x is an extreme point of BX is
essential for part b). Indeed, in finite dimensional Banach spaces, all
(cid:3)
STRONGLY EXTREME POINTS AND APPROXIMATION PROPERTIES
13
norms are weak(= norm) continuous. Clearly a point x can be extreme
for the ball B and not extreme for BX.
We end with a result that follows directly by combining Theorems
1.2 and 4.1.
Corollary 4.3. Let k·kW be a weakly continuous semi norm on bounded
sets in a Banach space (X,k · k) and let · be a lattice norm on R2.
Let x = (kxk,kxkW ). Let B be the unit ball of (X,·). Then the
strongly extreme points of B and BX coincide, and the denting points
of B and BX coincide.
References
[AHN+16] T. A. Abrahamsen, P. H´ajek, O. Nygaard, J. Talponen, and S. Troy-
anski, Diameter 2 properties and convexity, Studia Math. 232 (2016),
no. 3, 227–242. MR 3499106
[DGZ93] R. Deville, G. Godefroy, and V. Zizler, Smoothness and renormings in
Banach spaces, Pitman Monographs and Surveys in Pure and Applied
Mathematics, vol. 64, Longman Scientific & Technical, Harlow; copub-
lished in the United States with John Wiley & Sons, Inc., New York,
1993. MR 1211634 (94d:46012)
[JL01]
[KR82]
[GKS93] G. Godefroy, N. J. Kalton, and P. D. Saphar, Unconditional ideals in Ba-
nach spaces, Studia Math. 104 (1993), 13–59. MR 1208038 (94k:46024)
W. B. Johnson and J. Lindenstrauss (eds.), Handbook of the geometry
of Banach spaces. Vol. I, North-Holland Publishing Co., Amsterdam,
2001. MR 1863688
K. Kunen and H. Rosenthal, Martingale proofs of some geometrical re-
sults in Banach space theory, Pacific J. Math. 100 (1982), no. 1, 153–
175. MR 661446
B. L. Lin, P. K. Lin, and S. L. Troyanski, Characterizations of denting
points, Proc. Amer. Math. Soc. 102 (1988), 526–528. MR 0928972
(89e:46016)
[LLT88]
14
T. A. ABRAHAMSEN, P. H ´AJEK, O. NYGAARD, AND S. TROYANSKI
Department of Mathematics, University of Agder, Postboks 422,
4604 Kristiansand, Norway.
E-mail address: [email protected]
URL: http://home.uia.no/trondaa/index.php3
Mathematical Institute, Czech Academy of Science, Zitn´a 25, 115
67 Praha 1, Czech Republic and Department of Mathematics, Faculty
of Electrical Engineering, Czech Technical University in Prague,
Zikova 4, 160 00, Prague, Czech Republic
E-mail address: [email protected]
Department of Mathematics, University of Agder, Postboks 422,
4604 Kristiansand, Norway
E-mail address: [email protected]
URL: http://home.hia.no/olavn/
Institute of Mathematics and Informatics, Bulgarian Academy of
Science, bl.8, acad. G. Bonchev str. 1113 Sofia, Bulgaria and De-
partamento de Matem´aticas, Universidad de Murcia, Campus de Es-
pinardo, 30100 Espinardo (Murcia), Spain
E-mail address: [email protected]
|
0909.5672 | 5 | 0909 | 2010-06-02T12:55:33 | The Cauchy problem for Schr\"odinger-type partial differential operators with generalized functions in the principal part and as data | [
"math.FA",
"math.AP"
] | We set-up and solve the Cauchy problem for Schr\"odinger-type differential operators with generalized functions as coefficients, in particular, allowing for distributional coefficients in the principal part. Equations involving such kind of operators appeared in models of deep earth seismology. We prove existence and uniqueness of Colombeau generalized solutions and analyze the relations with classical and distributional solutions. Furthermore, we provide a construction of generalized initial values that may serve as square roots of arbitrary probability measures. | math.FA | math |
THE CAUCHY PROBLEM FOR SCHR ODINGER-TYPE PARTIAL
DIFFERENTIAL OPERATORS WITH GENERALIZED FUNCTIONS IN THE
PRINCIPAL PART AND AS DATA
G UNTHER H ORMANN
Dedicated to Peter Michor on the occasion of his 60th birthday
celebrated at the Central European Seminar in Mikulov, Czech Republic, May 2009
Abstract. We set-up and solve the Cauchy problem for Schrodinger-type differential operators
with generalized functions as coefficients, in particular, allowing for distributional coefficients in
the principal part. Equations involving such kind of operators appeared in models of deep earth
seismology. We prove existence and uniqueness of Colombeau generalized solutions and analyze
the relations with classical and distributional solutions. Furthermore, we provide a construction
of generalized initial values that may serve as square roots of arbitrary probability measures.
1. Introduction
Partial (and pseudo-) differential operators of Schrodinger-type with variable coefficients in the
principal part arise in the form of so-called paraxial equations in models of wave propagation.
These are based on narrow-angle symbol approximations of wave operators and have been applied
in integrated optics, underwater acoustic tomography, reflection seismic imaging, time-reversal
mirror experiments (cf. [2, 5, 41]), and recently in [12] to seismic wave propagation near the core-
mantle boundary inside the Earth (at approximately 2800 km depth). Since paraxial equations
are used to split the wave fields according to a prescribed principal direction of propagation they
are also called one-way wave equations. The leading-order approximation leads to model equations
of Schrodinger-type, where the material properties are still reflected by the regularity structure
of the coefficients in the principal part. Under strong smoothness conditions on the wave speed
function (i.e., the coefficients in the original wave operators) well-posedness of the one-way wave
Cauchy problems has been discussed in [20, 39].
In [12] this smoothness assumption has been
considerably relaxed by allowing the coefficients to be of Holder- or Sobolev-type regularity below
log-Lipschitz continuity. Recall that in general existence of distributional solutions to the second-
order wave equation may fail below log-Lipschitz regularity of the coefficients (cf. [10]). Beyond
such coefficient regularity barrier unique solvability of the Cauchy problem holds in the sense of
Colombeau generalized functions (cf. [17]), even in a covariant setting.
The relevance of coefficients with low Holder or Sobolev regularity has been shown in a variety of
geophysical applications, e.g. in the study of phase transitions in Earth's lowermost mantle ([40])
or in exploration geophysics ([13, 21, 22, 24, 32, 36, 42]). The model analyzed in [12] describes
the paraxial approximation to seismic wave propagation near the core-mantle boundary by the
following Schrodinger-type equation with depth z as evolution variable, the 2-dimensional lateral
x-variable, and a pseudodifferential time-frequency dependence
(1)
∂zu − i(cid:0)∂x1 (c1∂x1u) + ∂x2(c2∂x2 u)(cid:1) = 0,
Date: November 19, 2018.
2000 Mathematics Subject Classification. Primary: 46F30; Secondary: 35D99.
Research supported by the FWF START grant Y237.
1
where c1(z, x, τ ) and c2(z, x, τ ) are strictly positive symbols of order −1 in τ , continuously differen-
tiable with respect to z, but a non-Lipschitz dependence regarding the x-regularity, e.g. of Sobolev
type H r+1 or Holder continuous with exponent r, where r < 1. Roughly speaking, the main result
of [12] is the unique solvability of the corresponding Cauchy problem for the wave component u
in C([0,∞[, S ′(R; H 2(R2)) (i.e., continuous dependence on z, arbitrary S ′ quality with respect
to time or frequency, but lateral x-regularity of class H 2), given initial data in L1(R, H 2(R2)) (at
z = 0) and a right-hand side in C1([0,∞[, L1(R, L2(R2))). The emphasis in this result is on the
(Holder or) Sobolev regularity of the solution (with respect to x) in relation to the initial data
regularity under lowest possible regularity assumptions on the coefficient (cf. the brief discussion
on inverse analysis of medium regularity at the end of [12]).
In the current paper we strive for a widening of the possible range of applications and thereby
also simply for a more unified general mathematical set-up. Thus, we allow for discontinuous or
distributional coefficients, initial data, and right-hand sides. In fact, a natural method is then to
place the whole Cauchy problem into the context of nonlinear theories of generalized functions,
in particular the differential algebras of Colombeau generalized functions (cf. [7, 8, 9, 19, 29]). In
view of models from geophysics this amounts to including discontinuous material properties (e.g.
at fault zones and geological boundaries or cracks) and Dirac-type data such as strong impulsive
sources (e.g. explosions or earth quakes). A quantum field theoretic application of Colombeau
generalized functions in terms of regularizations of Wightman distributions has been given in
[18]. In the simpler context of quantum mechanics one will of course be interested in allowing
for a singular zero order term in the operator to represents a potential and non-smooth data
corresponding to an initial probability distribution. The square of the modulus of a solution
to the standard Schrodinger equation is usually interpreted as (time evolution of a) probability
density. Why should one not turn this around in the sense of allowing now for generalized initial
data which represent a "square root of a given arbitrary probability measure"?1 We will show
below how to construct a Colombeau generalized function whose square is associated with a given
probability measure in the sense of distributional shadows. The issue of squares of distribution
theoretic objects or generalized functions which contain or model measures (in particular, Dirac-
type terms) also arises in general relativity theory, e.g. in the pseudo-Riemannian metric associated
with impulsive pp-waves (cf. [19, Section 5.3] and [23, 34, 35]) and in the stress energy tensor
corresponding to ultrarelativistic Reissner-Nordstrøm fields (cf. [33]). A different regularization
approach for powers of Dirac measures as initial value appears in [27] in the context of (semilinear)
heat equations with singular potentials.
In our analysis of the Cauchy problem we will return to denoting the evolution variable by t and
furthermore incorporate also first-order terms in the differential operator. Note that unlike in the
specific geophysical model situation of [12] we neglect additional pseudodifferential aspects in the
symbol, since these played only the role of an external parameter upon a partial Fourier transform.
In summary of the basic structure, let T > 0 be arbitrary. We consider the Cauchy problem for a
generalized function u on Rn × [0, T ] in the form
(2)
∂tu − i
nXk=1
∂xk (ck∂xk u) − iV u = f
u t=0 = g,
(3)
where ck (k = 1, . . . , n), V , and f are generalized functions on Rn × [0, T ] and g is a general-
ized function on Rn. We note that basics of a theory for abstract variational problems in the
context of Colombeau-spaces have been presented in [16]. As indicated in [16, Section 8] the
Lax-Milgram-type theorem established there provides solutions to Dirichlet-problems involving
differential operators similar to the spatial derivatives appearing in the left-hand side of our above
differential equation. However, this does not directly extend to our kind of evolution problem and
1My thanks go to Gebhard Grubl and Michael Oberguggenberger who brought up this viewpoint in joint
discussions at the University of Innsbruck in April 2002.
2
furthermore involves less strict solution spaces than we are willing to accept here, namely GH 1 as
compared to GH∞ . Colombeau-generalized solutions to linear and nonlinear Schrodinger equations
with constant coefficient principal part have been constructed previously in [4, 37, 38].
We may point out that our current paper has both focus and aim different from the enormous
literature on the prominent questions concerning properties of Lp-solutions and spectral theory for
Schrodinger equations with standard principal part (i.e., with constant or smoooth coefficients)
and singular potentials. Instead we investigate here the feasibility of extending the solution concept
and the basic analysis to the case of generalized functions and, in particular, non-smooth or highly
singular principal parts. The plan of the paper is as follows. In Section 2 we briefly review the
regularization approach to generalized functions in the sense of Colombeau and show how square
roots of probability measures can be implemented in this framework. Section 3 presents the main
result of unique existence of generalized solutions to our Schrodinger-type Cauchy problem (2-
3). Furthermore, we discuss the relation of Colombeau generalized solutions with classical and
distributional solution concepts.
2. Regularization of coefficients and data
The basic idea of regularization methods is to replace non-smooth data by approximating nets
of smooth functions, e.g. instead of g consider C∞ ∋ gε → g (ε → 0). More generally, we may
replace g by a net (gε)0<ε≤1 of C∞ functions, convergent or not, but with moderate asymptotics
with respect to ε and identify regularizing nets whose differences compared to the moderateness
scale are negligible. For a modern introduction to Colombeau algebras we refer to [19]. Here we
will also make use of constructions and notations from [15].
Construction of generalized functions based on a locally convex topological vector space E:
Let E be a locally convex topological vector space whose topology is given by the family of
seminorms {pj}j∈J . The elements of
ME := {(uε)ε ∈ E(0,1] : ∀j ∈ J ∃N ∈ N pj(uε) = O(ε−N ) as ε → 0}
and
NE := {(uε)ε ∈ E(0,1] : ∀j ∈ J ∀q ∈ N pj(uε) = O(εq) as ε → 0},
are called E-moderate and E-negligible, respectively. With operations defined componentwise, e.g.
(uε) + (vε) := (uε + vε) etc., NE becomes a vector subspace of ME. We define the generalized
functions based on E as the factor space GE := ME/NE. If E is a differential algebra then NE is
an ideal in ME and GE becomes a differential algebra too.
By particular choices of E we reproduce several standard Colombeau algebras of generalized
functions. For example, E = C with the absolute value as norm yields the generalized complex
numbers GE = eC and for any Ω ⊆ Rd open, E = C∞(Ω) with the topology of compact uniform
convergence of all derivatives provides the so-called special Colombeau algebra GE = G(Ω). Recall
that Ω 7→ G(Ω) is a fine sheaf, thus, in particular, the restriction uB of u ∈ G(Ω) to an arbitrary
open subset B ⊆ Ω is well-defined and yields uB ∈ G(B). Moreover, we may embed D′(Ω) into
G(Ω) by appropriate localization and convolution regularization.
In case E ⊆ D′(Ω) certain generalized functions can be projected into the space of distributions
by taking weak limits: we say that u ∈ GE is associated with w ∈ D′(Ω), if uε → w in D′(Ω) as
ε → 0 holds for any (hence every) representative (uε) of u. This fact is also denoted by u ≈ w.
In the current paper we will consider open strips of the form ΩT = Rn× ]0, T [⊆ Rn+1 (with T > 0
arbitrary) and employ the spaces E = H ∞(ΩT ) = {h ∈ C∞(ΩT ) : ∂αh ∈ L2(ΩT ) ∀α ∈ Nn+1}
3
with the family of (semi-)norms
khkHk =(cid:16) Xα≤k
k∂αhk2
L2(cid:17)1/2
(k ∈ N),
as well as E = W ∞,∞(ΩT ) = {h ∈ C∞(ΩT ) : ∂αh ∈ L∞(ΩT ) ∀α ∈ Nn+1} with the family of
(semi-)norms
khkW k,∞ = max
α≤k k∂αhkL∞ (k ∈ N).
(Note that ΩT clearly satisfies the strong local Lipschitz property [1, Chapter IV, 4.6, p. 66] and
therefore the Sobolev embedding theorem [1, Chapter V, Theorem 5.4, Part II, p. 98] implies that
every element of H ∞(ΩT ) and W ∞,∞(ΩT ) belongs to C∞(ΩT ).)
To avoid overloaded subscripts we shall make use of the following (abuses of) notation in the
sequel
GL2 (Rn × [0, T ]) := GH∞(ΩT )
and
GL∞ (Rn × [0, T ]) := GW ∞,∞(ΩT ).
For example and in explicit terms we will represent a generalized initial value g ∈ GL2(Rn × [0, T ])
by a net (gε) with the moderateness property
and similarly for the right-hand side and the coefficients. Asymptotically negligible errors in the
∀k ∃m :
kgεkHk = O(ε−m)
(ε → 0)
initial data, e.g. by using a representative (egε) instead, are expressed by estimates of the form
∀k ∀p :
kgε − egεkHk = O(εp)
(ε → 0).
Similar constructions and notations will be used in case of E = H ∞(Rn) and E = W ∞,∞(Rn).
Note that by Young's inequality ([14, Proposition 8.9.(a)]) any standard convolution regularization
with a scaled mollifier of Schwartz class provides embeddings L2 ֒→ GL2 and Lp ֒→ GL∞ (1 ≤ p ≤
∞).
As an example of a detailed regularization model we construct Colombeau generalized positive
square roots of arbitrary probability measures, which can serve as initial values in the Cauchy
problem analyzed below.
Proposition 2.1. Let µ be a (Borel) probability measure on Rn. Choose ρ ∈ L1(Rn)∩W ∞,∞(Rn)
to be positive with R ρ = 1 and satisfying ρ(x) ≥ x−m0 when x ≥ 1 with some m0 > n. Set
ρε(x) = 1
(1) hε is positive and setting φε := √hε the net (φε)ε∈ ]0,1] represents an element φ ∈ G(Rn)
such that φ2 ≈ µ;
(2) there exists a generalized function g ∈ GL2(Rn) such that g2 ≈ µ and the class of (gεΩ)ε∈ ]0,1]
is equal to φΩ in G(Ω), or by slight abuse of notation gΩ = φΩ, for every bounded open
subset Ω ⊆ Rn.
ε ) and hε := µ ∗ ρε, then we have
εn ρ( x
Proof. (1) The Borel measure µ is regular since it is finite and Rn is locally compact and second
countable ([6, Proposition 7.2.3] or also [14, Theorem 7.8]). Thus µ(Rn) = 1 implies that we
can find a compact subset A ⊆ Rn such that µ(A) ≥ 1/2. A variant of Young's inequality for
measures (cf. [14, Proposition 8.49]) applied to ∂αhε = µ ∗ ∂αρε directly implies that the net
(hε) is C∞(Rn)-moderate, even with global L∞-norms, since kµ ∗ ∂αρεkL∞ ≤ kµkvar k∂αρεkL∞ =
k∂αρkL∞ε−n−α.
(Here k kvar denotes the total variation norm on the space of finite Radon
measures on Rn, which gives kµkvar = µ(Rn) = 1 in case of the positive probability measure µ).
Furthermore, for any x ∈ Rn and ε > 0 we have
ρε(x − y)dµ(y) ≥Z
hε(x) = µ ∗ ρε(x) =Z
ρε(x − y)dµ(y) ≥ µ(A) · min
ρε(z) > 0
z∈{x}−A
Rn
A
4
and hence that φε = √hε ∈ C∞(Rn) (0 < ε ≤ 1). Moreover, if K ⊆ Rn is compact then we have2
with r(K) := max{x : x ∈ K − A} > 0 the estimate
(∗)
inf
x∈K
hε(x) = inf
x∈K
µ ∗ ρε(x) ≥ µ(A) · inf
x∈K
min
z∈{x}−A
1
2εn min
z≤r(K)
≥
ρ(z/ε) ≥
1
2εn
r(K)m0
ρε(z) ≥
εm0
=
1
2 · min
z∈K−A
εm0−n
2 r(K)m0
ρ(z/ε)
εn
(0 < ε < 1/r(K)).
If α ∈ Nn is arbitrary then ∂αφε is a linear combination of terms of the form ∂β1hε ··· ∂βk hε/hl/2
with appropriate β1, . . . , βk ∈ Nn and l ∈ N. Hence C∞(Rn)-moderateness of (hε) together with
the lower bounds for inf x∈K hε(x) obtained above prove C∞(Rn)-moderateness of φε.
Finally, we have that by construction φ2 is represented by hε = µ ∗ ρε and thus clearly converges
to µ as ε → 0 in the sense of distributions.
(2) For j ∈ N let Kj be the closed ball of radius 2j around 0 in Rn. Let χ0 ∈ D(Rn), 0 ≤ χ0 ≤ 1
with χ0(x) = 1 when x ≤ 1 and χ0(x) = 0 when x > 2 and put χj(x) := χ0(2−jx) (j ≥ 1).
Thus we have χj(x) = 1 when x ≤ 2j, χj(x) = 0 when x > 2j+1, and for every γ ∈ Nn
ε
k∂γχjkL2 = 2(−γ+n/2)jk∂γχ0kL2.
We define the net (gε)ε∈ ]0,1] of smooth functions on Rn by
gε := χj φε, if 2−j−1 < ε ≤ 2−j
(j ∈ N).
For every ε ∈ ]0, 1] the function gε is smooth and has compact support, hence belongs to H ∞(Rn).
We have to show that (gε) is an H ∞-moderate net.
Let α ∈ Nn arbitrary and 2−j−1 < ε ≤ 2−j. By the Leibniz rule we have that ∂αgε is a linear
combination of terms of the form ∂βφε · ∂α−βχj (β ≤ α). Therefore we have by the triangle
inequality that k∂αgεkL2 is bounded above by a linear combination of terms of the form
k∂βφε · ∂α−βχjkL2 ≤ k∂α−βχjkL2k∂βφεkL∞(Kj+1)
≤ 2nj/2 kχ0kHαk∂βφεkL∞(Kj+1) ≤ ε−n/2 kχ0kHαk∂βφεkL∞(Kj+1)
and it remains to prove ε-moderate bounds for k∂βφεkL∞(Kj+1) when β ≤ α (note that the coupling
of j with ε has to be taken into account here). As already noted in (1) the latter is in turn bounded
above by a linear combination of terms of the form
sup
x∈Kj+1
∂β1hε(x)··· ∂βk hε(x)
hε(x)l/2
with suitable β1, . . . , βk ∈ Nn and l ∈ N. Thanks to the estimate (∗) in the proof of (1) we
have 1/hε(x)l/2 ≤ 2 r(Kj+1)m0 ε−(m0−n) uniformly with respect to x ∈ Kj+1, where r(Kj+1) =
O(2j+1) = O(1/ε). Thus it remains to consider the factors ∂βq hε(x) in the above supremum.
As has already been noted in the first paragraph of the proof of (1) we have global L∞-estimates
with moderate ε-dependence. Thus in summary, we have shown the H ∞-moderateness of (gε).
Finally, let Ω ⊆ Rn be bounded and open. Choose j sufficiently large so that Ω ⊆ Kj. Then
we have for 0 < ε < 2−j by construction of gε that gεΩ = φεΩ and hence equality of the
corresponding classes in G(Ω). Moreover, the latter also implies that g2 ≈ µ, since the support of
any test function in D(Rn) is contained in some open bounded subset.
(cid:3)
Remark 2.2. (i) Property (∗) established in course of the proof shows in fact that the class of
(hε) in G(Rn) as well as φ are strictly positive generalized functions (cf. [26, Theorem 3.4]). In
this sense, φ provides a strictly positive square root of the positive measure µ.
2We may exclude the case A = K = {x0} without loss of generality
5
(ii) For specific choices of ρ in L1(Rn) ∩ H ∞(Rn) such that √ρ ∈ H ∞(Rn) we could obtain that
(φε) is also H ∞-moderate and directly defines a square root in GL2(Rn) without having to undergo
the cut-off procedure in part (2) of Proposition 2.1 (which, on the other hand, cannot be avoided
for general ρ ∈ H ∞). For example, putting ρ(x) = c(1+x2)−(n+1)/2 with a suitable normalization
constant c > 0 provides such a mollifier. However, the above Proposition leaves considerably more
flexibility in adapting the regularization to particular applications.
3. Generalized function solutions and coherence properties
We come now to the main existence and uniqueness result for generalized solutions to the Cauchy
problem (2-3). We recall that a regularization of an arbitrary finite-order distribution which meets
the log-type conditions on the coefficients ck and V in the following statement is easily achieved
by employing a re-scaled mollification process as described in [28].
Theorem 3.1. Let ck (k = 1, . . . , n) as well as V be real3 generalized functions in GL∞(Rn×[0, T ]),
f in GL2(Rn × [0, T ]), and g be in GL2(Rn). Assume that the following log-type conditions hold
for some (hence all) representatives (ckε)ε∈]0,1] of ck (k = 1 . . . , n) and (Vε) of V :
k∂tckεkL∞ = O(log(
1
ε
))
and k∂tVεkL∞ = O(log(
1
ε
))
(ε → 0).
In addition let the positivity conditions ckε(x, t) ≥ c0 for all (x, t) ∈ Rn × [0, T ], ε ∈ ]0, 1], k =
1, . . . , n with some constant c0 > 0 be met.4
Then the Cauchy problem
(4)
∂tu − i
nXk=1
∂xk (ck∂xk u) − iV u = f,
u t=0 = g,
(5)
has a unique solution u ∈ GL2(Rn × [0, T ]).
Proof. In terms of representatives (4-5) means that we have a family of Cauchy problems parametrized
by ε ∈ ]0, 1]:
(6)
∂tuε − i
nXk=1
∂xk (ckε∂xk uε) − iVεuε = fε
uε t=0 = gε
on ΩT = Rn× ]0, T [,
on Rn,
(7)
where ckε (k = 1, . . . , n) and Vε (are real-valued and) belong to W ∞,∞(ΩT ), ckε satisfies the
log-type and the positivity condition as stated above, gε ∈ H ∞(Rn), and fε ∈ H ∞(ΩT ).
Our strategy is to solve the corresponding problem at fixed, but arbitrary, parameter value ε
and thereby produce a solution candidate (uε). The substance of the proof lies in the efforts to
show that each uε belongs to H ∞(Rn×]0, T [) and in addition satisfies moderateness estimates in
every derivative with respect to the L2-norm. We will obtain basic energy estimates by standard
variational methods as discussed by Dautray-Lions in [11, Chapter XVIII, §7, Section 1], but we
will need to perform an additional analysis of the dependence of the constants in these estimates
on the various norms of the coefficient functions.
Step 1 (basic estimates and regularity): We apply the set-up and constructions in [11, Chapter
XVIII, §7, Section 1] to the Hilbert space triple H 1(Rn) ⊂ L2(Rn) ⊂ H −1(Rn) and the sesquilinear
form
aε(t; ϕ, ψ) :=
nXk=1
hckε(t) ∂xk ϕ∂xk ψiL2(Rn) + hVε(t)ϕψi
(ϕ, ψ ∈ H 1(Rn)),
3In the sense that they possess representating nets of real-valued functions.
4For any other representative the conditions thus hold for sufficiently small ε with c0/2 instead.
6
where we have used the short-hand notation v(t) for a function x 7→ v(x, t). The basic conditions
[11, (7.1) and (7.2), p. 621] on the quadratic form aε are easily seen to be met: continuous
differentiability of the map t 7→ aε(t; ϕ, ψ) follows from smoothness of ckε (k = 1, . . . , n) and Vε,
the form aε(t; ., .) is hermitian since each ckε and Vε is real-valued, and the positivity conditions
on the coefficients ckε (k = 1, . . . , n) in the statement of Theorem 3.1 yield the following coercivity
estimate
aε(t; ϕ, ϕ) + λεkϕk2
L2 ≥ c0kϕk2
H 1
∀ϕ ∈ H 1(Rn),
where λε = c0+kVεkL∞ (thus we have λ = λε and α = c0 in [11, (7.2), ii) on p. 621] ). Furthermore,
the hypotheses [11, (7.5) and (7.6), pp. 621-622] on the initial values hold by our assumptions on gε
and fε. We claim that [11, Theorem 1, pp. 621-622] in combination with the remark on additional
regularity [11, Remark 3, p. 625] implies that we obtain a unique solution
uε ∈ C([0, T ], H 1(Rn)) ∩ C1([0, T ], H −1(Rn))
(S1)
to the Cauchy problem (6-7) for every ε ∈ ]0, 1]. In fact, the basic theorem gives existence and
uniqueness with uε ∈ L2([0, T ], H 1(Rn)) and u′
ε ∈ L2([0, T ], H −1(Rn)) only. However, as men-
tioned in [11, Remark 3, p. 625] the existence proof given in Dautray-Lions directly shows that the
solution as well as its t-derivative are L∞-functions of t with values in H 1 or H −1, respectively.
The same remark states that even continuity with respect to t holds (an earlier reference for this
fact is [25, Remark 10.2, p. 302], which also indicates a proof based on convolution regulariza-
tion). Finally, we note that thanks to C([0, T ], H 1) ⊆ C([0, T ], H −1) an application of [11, Chapter
XVIII, §1, Section 2, Proposition 7, p. 477] shows that the solution belongs to C1([0, T ], H −1).
Moreover, a careful inspection of all constants appearing in the derivation of the a priori estimate
[11, (7.16) and (7.17) on p. 624] shows that we may deduce the following basic energy estimate
uniformly for all t ∈ [0, T ]
(8)
H 1 ≤ C2ε eC1ε
kuε(t)k2
H 1 +
kgεk2
TZ
0
(cid:16)kfε(τ )k2
L2 + k∂tfε(τ )k2
H−1(cid:17) dτ
,
where the specifics on the constants C1ε and C2ε according to the derivation in [11, pp. 623-624]
are as follows: C2ε = O(T (c0 + kVεkL∞)) = O(ε−p) (ε → 0) for some p ∈ N, and
1≤k≤n k∂tckεkL∞ + k∂tVεkL∞ )(cid:17) (ε → 0),
C1ε = O(cid:16) T
c0 · ( max
which by the log-type conditions on (ckε) and (Vε) imply moderateness estimates for the first-order
spatial derivatives in the form
∃N1 ∈ N :
k∂xluεkL2(Rn×[0,T ]) ≤
√T sup
t∈[0,T ]kuε(t)kH 1 = O(ε−N1 )
(ε → 0, l = 1, . . . , n).
As a preparation for the procedure in Step 2 we claim that additional regularity properties for uε
hold in the form
(∗)
(at fixed ε without precise asymptotic estimates of the norms). These can be obtained from the
concept of mild solutions and evolution systems on L2(Rn) (cf. [30, Chapter 5]) for the self-adjoint
k=1 ∂xk (ckε(x, t)∂xk v)(x) (v ∈ H 2(Rn))
familiy of operators (Aε(t))t∈[0,T ], where Aε(t)v(x) := Pn
with common (i.e., t-independent) domain H 2(Rn): In fact, Equation (4) has the form
uε ∈ L∞([0, T ], H 2(Rn)) ∩ W 1,∞([0, T ], H 1(Rn))
∂tuε − iAε(t)uε = fε + iVεuε =: Fε,
where we already know from (S1) that Fε ∈ C([0, T ], L2(Rn)) ∩ C1([0, T ], H −2(Rn)). Let the
strongly continuous evolution system corresponding to (Aε(t))t∈[0,T ] be denoted by (Uε(t, s))0≤s≤t≤T ,
then we necessarily have by Duhamel's formula
uε(t) = Uε(t, 0)gε +
tZ
0
Uε(t, τ )Fε(τ ) dτ ∈ C1
w([0, T ], L2(Rn)),
7
where C1
Rn with values in L2(Rn) for every t. Writing the differential equation now in the form
w([0, T ], L2(Rn)) means weakly continuously differentiable as t-dependent distribution on
−iAε(t)uε = −∂tuε + Fε ∈ Cw([0, T ], L2(Rn))
we may employ elliptic regularity (spatial, with respect to x) for the second-order operator Aε(t)
and deduce uε ∈ Cw([0, T ], H 2(Rn)) ⊆ L∞([0, T ], H 2(Rn)). Moreover, we may now state in
addition that Fε ∈ Cw([0, T ], H 1(Rn)) and use Duhamel's formula again to show that also uε ∈
w([0, T ], H 1(Rn)) ⊆ W 1,∞([0, T ], H 1(Rn)) as claimed.
C1
Step 2 (higher x-derivatives): We take the partial xj-derivative on both sides in Equations (6)
and (7) to obtain a similar differential equation for ∂xj uε in the form
nXk=1
nXk=1
∂xk (ckε∂xk ∂xj uε) − iVε∂xj uε = i
∂t∂xj uε − i
and the initial condition ∂xj uε(0) = ∂xj gε. Thanks to (∗) we have f1ε ∈ L2([0, T ], L2(Rn)) and
∂tf1ε ∈ L2([0, T ], H −1(Rn)), which corresponds to conditions (7.5-6) of [11, Chapter XVIII, §7,
Section 1, Theorem 1]. Thus we may apply (S1) and (8) with ∂xj uε, ∂xj gε, f1ε replacing uε, gε,
fε, respectively, but with the same constants C1 and C2. Collecting the results for j = 1, . . . , n
we arrive at
∂xk (∂xj ckε∂xk uε) + i∂xj Vεuε + ∂xj fε =: f1ε
(S2)
uε ∈ C([0, T ], H 2(Rn)) ∩ C1([0, T ], L2(Rn))
k∂xl∂xj uεkL2(Rn×[0,T ]) ≤ n√T sup
and
∃N2 ∈ N :
(ε → 0, j, l = 1, . . . , n).
Similarly, taking higher partial x-derivatives in Equations (6) and (7) produces the same kind of
differential equations and initial conditions for these higher derivatives, where we may always con-
sider the already estimated lower order derivatives of uε as part of the right-hand side. Therefore
we obtain successively for arbitrary m ∈ N
(Sm)
t∈[0,T ] kuε(t)kH 2 = O(ε−N2 )
uε ∈ C([0, T ], H m(Rn)) ∩ C1([0, T ], H m−1(Rn))
x uεkL2(Rn×[0,T ]) = O(ε−Nm )
0 ,α ≤ m :
k∂α
∃Nm ∈ N∀α ∈ Nn
(ε → 0).
and
Step 3 (t-derivatives): We first note that due to (Sm) we have ∂tuε(0) ∈ H ∞(Rn) and Equation
(6) evaluated at t = 0 provides moderate ε-asymptotics of its Sobolev norms. Taking the partial
t-derivative in Equation (6) now yields again the same kind of differential equation for ∂tuε as
above when interpreted in the following way:
∂t∂tuε − i
nXk=1
∂xk (ckε∂xk ∂tuε) − iVε∂tuε = i
nXk=1
∂xk (∂tckε∂xk uε) + i∂tVεuε + ∂tfε =: ef1ε.
0 , α ≤ m,
By property (Sm) (now replacing (∗) in a similar argument in Step 2) we deduce the required
regularity conditions on ef1ε and ∂tef1ε ∈ L2([0, T ], H −1(Rn)) to apply (S1) and (8) now with ∂tuε,
∂tuε(0), ef1ε replacing uε, gε, fε, respectively. We obtain for α ∈ Nn
0 , α ≤ 1,
∃M1 ∈ N : k∂t∂α
uε ∈ C1([0, T ], H 1(Rn)) ∩ C2([0, T ], H −1(Rn)),
x uεkL2(Rn×[0,T ]) = O(ε−M1 ).
Furthermore, inserting here the same procedure as in Step 1 now yields for any m ∈ N and for all
α ∈ Nn
x uεkL2(Rn×[0,T ]) = O(ε−Mm ).
uε ∈ C1([0, T ], H m(Rn)) ∩ C2([0, T ], H m−1(Rn)),
Thus in turn we may now deduce from the above differential equation for ∂tuε that ∂2
t uε(0) ∈
H ∞(Rm) holds with moderate spatial Sobolev norms. Hence we may now play the same game
again with ∂2
t uε etc. and finally arrive at the statements that for arbitrary d, m ∈ N and for all
α ∈ Nn
uε ∈ Cd([0, T ], H m(Rn))∩Cd+2([0, T ], H m−1(Rn)), ∃Mmd ∈ N : k∂d
x uεkL2(Rn×[0,T ]) = O(ε−Mmd ).
0 , α ≤ m, we have
∃Mm ∈ N : k∂t∂α
t ∂α
8
Therefore we may conclude that (uε) is moderate net and its class GL2 (Rn × [0, T ]) defines a
solution u to the Cauchy problem (4-5).
The proof of uniqueness requires to show negligibility of any solution (uε) assuming that the
right-hand side (fε) and the initial data (gε) are negligible. But this follows successively from
the corresponding variant of the energy estimate (8) applied at each step of the sequence of
differentiated differential equations used in course of the existence proof.
(cid:3)
We note that in case of smooth coefficients an easy integration by parts argument shows that any
solution to the Cauchy problem obtained from the variational method as in [11, Chapter XVIII,
§7, Section 1]) also provides a solution in the sense of distributions. The following statement
ensures in addition the coherence with the Colombeau generalized solution.
Corollary 3.2. Let V and ck (k = 1, . . . , n) belong to C∞(ΩT ) ∩ L∞(ΩT ) with bounded time
derivatives of first-order, g0 ∈ H 1(Rn), and f0 ∈ C1([0, T ], L2(Rn)). Let u denote the unique
Colombeau generalized solution to the Cauchy problem (4-5), where g, f denote standard embed-
dings of g0, f0, respectively. Then u ≈ w, where w ∈ C([0, T ], H 1(Rn)) is the unique distributional
solution obtained from the variational method.
Proof. Let (uε) ∈ u be as in the proof of Theorem 3.1, then vε := uε − w satisfies the following
Cauchy problem
∂tvε − i
nXk=1
∂xk (ck∂xk vε) − iV vε = fε − f0,
vε t=0= gε − g0.
Moreover, we have the energy estimate (8), where now both C1ε =: C1 and C2ε =: C2 both are
independent of ε, in the form
sup
t∈[0,T ]kvε(t)k2
H 1 ≤ C2eC1(cid:16)kgε − g0k2
H 1 +
TZ
0
(cid:16)kfε(τ ) − f0(τ )k2
L2 + k∂tfε(τ ) − ∂tf0(τ )k2
H−1(cid:17) dτ(cid:17).
Since the smoothing process via mollification ensures convergence to 0 of the norms on the right-
hand side of the above estimate we obtain even norm convergence of uε to w.
(cid:3)
Remark 3.3. To compare our main result with the evolution systems solution constructed in
[12] for the case of Sobolev coefficients and spatial dimension n = 2 we drop the additional
pseudodifferential aspects in that model. Then we obtain also coherence of the Colombeau concept
with the functional analytic solution as we sketch in the following discussion:
Let 0 < r < 1 and V = 0, c10, c20 ∈ C1([0, T ], H r+1(R2)), g0 ∈ H 2(R2), f0 ∈ C1([0, T ], L2(R2)).
Let v denote the unique H 2-valued solution to the Cauchy problem corresponding to (4-5) with
initial value g0 and right-hand side f0 (and coefficients cj0 replacing cj) according to [12, Theorem
4.2]. Let g, f , c1, c2 denote the embeddings of g0, f0, c10, c20 into corresponding Colombeau
spaces, respectively, and let u ∈ GL2(R2 × [0, T ]) be the unique Colombeau solution to (4-5).
Subtracting the equations satisfied by a representative (uε) of u and v we deduce the following
equations for the difference hε := uε − v
which imply an energy estimate of the form (8) for hε and with suitably adapted initial value
and right-hand side efε instead. As noted in the proof of Corollary 3.2 we have appropriate norm
convergence of the standard regularizations gε, fε of the data g0,f0. In addition, we now have to call
on uniform boundedness of k∂tcεkL∞(R2×[0,T ]) and on the C1([0, T ], H r+1(R2))-norm convergence
ckε − ck0 → 0 (ε → 0). Using the continuity of H r+1(R2) · H 2(R2) → H r+1(R2) we obtain efε → 0
in C1([0, T ], H r+1(R2)). Thus we finally obtain the convergence uε → v in C([0, T ], H 1(R2)), and
in particular we have u ≈ v.
9
∂thε − i
∂xk (ckε∂xk hε) = fε − f0 + i
2Xk=1
2Xk=1
∂xk ((ckε − ck0)∂xk v) =: efε,
hε t=0= gε − g0,
Example 3.4. Consider the strictly positive square root of δ represented by (√ρε)ε∈ ]0,1], where
ρ is a mollifier as in Proposition 2.1. (Note that we obtain hε = δ ∗ ρε = ρε in this case.) To
simplify technical matters, assume in addition that ρ satisfies the conditions discussed in Remark
2.2(ii) and moreover √ρ ∈ L1(Rn). (For example, any suitably normalized function of the form
x 7→ (1 + x2)−m/2 with m > 2n would do.) Then we may consider the class g ∈ GL2(Rn), given
directly by (√ρε)ε∈ ]0,1], as a square root of δ.
We have g2 ≈ δ, but g ≈ 0, which is easily seen by action on a test function ϕ upon substituting
y = x/ε inRpρε(x) ϕ(x)dx = εn/2Rpρ(y) ϕ(εy)dy and applying dominated convergence (thereby
using that √ρ ∈ L1). In essence, this effect has already been observed earlier in the generalized
function model of ultrarelativistic Reissner-Nordstrøm fields in [33, Equations (15) and (17)].
The Cauchy problem (4-5) with generalized initial value g, right-hand side f = 0, constant coeffi-
cients ck = 1 (k = 1, . . . , n) and V = 0, written out for representatives then reads
∂tuε = i∆uε,
uεt=0 = √ρε.
µt
uε(t, x)2 dx =Z
(Ut√ρε)(x) · (Ut√ρε)(x) dx =Z
ε denote the positive measure on Rn with density function uε(t, .)2 for the
pρε(x)2 dx =Z
ε : t ∈ R, ε ∈ ]0, 1]} is a family probability measures on Rn and kµt
The solution is given by the action of the strongly continuous unitary group Ut := exp(it∆) (t ∈ R)
of operators on L2(Rn) in the form uε(t, x) = (Ut√ρε)(x).
Let t ∈ R and µt
Lebesgue measure. By unitarity of Ut we obtain
ε(Rn) =Z
hence {µt
ε ∈ ]0, 1]) holds in the Banach space of finite measures.
We claim that for any t 6= 0 the net (µt
on finite measures (cf. [3, §30]): Since √ρε ∈ L1(Rn) we obtain from the L1-L∞-estimate for
the Schrodinger propagator ([31, §4.4, Theorem 1]) that kuε(t, .)kL∞ ≤ k√ρεkL1/(4πt)n/2 and
therefore for any ψ ∈ Cc(Rn)
ε)ε∈ ]0,1] converges to 0 with respect to the vague topology
εkvar = 1 (t ∈ R,
ρε(x) dx = 1,
Rn
Rn
Rn
Rn
ε, ψi ≤Z
hµt
Rn
uε(t, x)2ψ(x) dx ≤ (4πt)−n kψkL1k√ρεkL1
= (4πt)−n kψkL1k√ρkL1 εn/2 → 0
(ε → 0).
ε → 0 in S ′(Rn) (when t 6= 0), whereas
ε)ε∈ ]0,1] can certainly not be weakly convergent as net of finite measures (i.e., in the weak-∗
ε(Rn) =
The same obviously holds with ψ ∈ S (Rn), hence also µt
(µt
topology in the dual of the space of bounded continuous functions on Rn) since hµt
1 6→ 0 as ε → 0 (cf. also [3, Theorem 30.8]).
ε, 1i = µt
Acknowledgement: The author wishes to express his sincere thanks to two anonymous referees.
In particular, we thank the one for hints to additional references of related work, and we are
deeply grateful to the second referee for patient comments and clarifications that lead to crucial
corrections in Theorem 3.1 and its proof.
References
[1] R. Adams. Sobolev Spaces. Academic Press, New York, 1975.
[2] G. Bal, G. Papanicolaou, and L. Ryzhik. Self-averaging in time reversal for the parabolic wave equation.
Stochastics and Dynamics, 2:507 -- 531, 2002.
[3] H. Bauer. Measure and integration theory, volume 26 of de Gruyter Studies in Mathematics. Walter de Gruyter
& Co., Berlin, 2001.
[4] C. Bu. Generalized solutions to the cubic Schrodinger equation. Nonlinear Anal., 27(7):769 -- 774, 1996.
[5] J. F. Claerbout. Imaging the Earth's interior. Blackwell Scientific Publications, Palo Alto, 1985.
[6] D. L. Cohn. Measure theory. Birkhauser, Boston, 1980.
10
[7] J. F. Colombeau. Une multiplication g´en´erale des distributions. C. R. Acad. Sci. Paris, S´er.I 296:357 -- 360,
1983.
[8] J. F. Colombeau. New generalized functions and multiplication of distributions. North-Holland, Amsterdam,
1984.
[9] J. F. Colombeau. Multiplication of distributions. A tool in mathematics, numerical engineering and theoretical
physics. Number 1532 in Lecture Notes in Math. Springer-Verlag, New York, 1992.
[10] F. Colombini and N. Lerner. Hyperbolic operators with non-Lipschitz coefficients. Duke Math. J., 77(3):657 --
698, 1995.
[11] R. Dautray and J.-L. Lions. Mathematical analysis and numerical methods for science and technology. Vol.
5. Springer-Verlag, Berlin, 1992.
[12] M. V. de Hoop, G. Hormann, and M. Oberguggenberger. Evolution systems for paraxial wave equations of
Schrodinger-type with non-smooth coefficients. J. Differential Equations, 245(6):1413 -- 1432, 2008.
[13] S. Dolan, C. Bean, and B. Riollet. The broad-band fractal nature of heterogeneity in the upper crust from
petrophysical logs. Geophys. Jour. Int., 132:489 -- 507, 1998.
[14] G. B. Folland. Real Analysis. John Wiley and Sons, New York, 1999.
[15] C. Garetto. Topological structures in Colombeau algebras: topological C-modules and duality theory. Acta
Appl. Math., 88(1):81 -- 123, 2005.
[16] C. Garetto and H. Vernaeve. Hilberte-modules: Structural properties and applications to variational problems.
Trans. Amer. Math. Soc., 2009. to appear.
[17] J. D. E. Grant, E. Mayerhofer, and R. Steinbauer. The wave equation on singular space-times. Comm. Math.
Phys., 285(2):399 -- 420, 2009.
[18] H. Grosse, M. Oberguggenberger, and I. T. Todorov. Generalized functions for quantum fields obeying quadratic
exchange relations. Tr. Mat. Inst. Steklova, 228(Probl. Sovrem. Mat. Fiz.):90 -- 100, 2000.
[19] M. Grosser, M. Kunzinger, M. Oberguggenberger, and R. Steinbauer. Geometric theory of generalized func-
tions. Kluwer, Dordrecht, 2001.
[20] L. Halpern and L. N. Trefethen. Wide-angle one-way wave equations. J. Acoust. Soc. Amer., 84:1397 -- 1404,
1988.
[21] F. J. Herrmann. A scaling medium representation. A discussion on well-logs, fractals and waves. PhD thesis,
Technische Universiteit Delft, 1997.
[22] G. Hormann and M. V. de Hoop. Geophysical modeling with Colombeau functions: Microlocal properties and
Zygmund regularity. In A. Delcroix, M. Hasler, J.-A. Marti, and V. Valmorin, editors, Nonlinear Algebraic
Analysis. Cambridge Scientific Publishers, 2004.
[23] M. Kunzinger and R. Steinbauer. A note on the Penrose junction conditions. Classical Quantum Gravity,
16(4):1255 -- 1264, 1999.
[24] X. Li and J. Haury. Characterization of heterogeneities from sonic velocity measurements using the wavelet
transform. Expanded Abstracts Soc. Expl. Geophys., 1995:488 -- 491, 1995.
[25] J. L. Lions and E. Magenes. Non-homogeneous boundary value problems and applications, volume 1. Springer-
Verlag, Berlin, New York, 1972.
[26] E. Mayerhofer. On Lorentz geometry in algebras of generalized functions. Proc. Roy. Soc. Edinburgh Sect. A,
138(4):843 -- 871, 2008.
[27] M. Nedeljkov, S. Pilipovi´c, and D. Rajter- ´Ciri´c. Heat equation with singular potential and singular data. Proc.
Roy. Soc. Edinburgh Sect. A, 135(4):863 -- 886, 2005.
[28] M. Oberguggenberger. Hyperbolic systems with discontinuous coefficients: generalized solutions and a trans-
mission problem in acoustics. J. Math. Anal. Appl., 142:452 -- 467, 1989.
[29] M. Oberguggenberger. Multiplication of distributions and applications to partial differential equations. Long-
man Scientific & Technical, 1992.
[30] A. Pazy. Semigroups of linear operators and applications to partial differential equations. Springer-Verlag, New
York, 1983.
[31] J. Rauch. Partial differential equations, volume 128 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 1991.
[32] A. Saucier and J. Muller. Use of multifractal analysis in the characterization of geological information. Fractals,
1(3):617 -- 628, 1993.
[33] R. Steinbauer. The ultrarelativistic Reissner-Nordstrøm field in the Colombeau algebra. J. Math. Phys.,
38(3):1614 -- 1622, 1997.
[34] R. Steinbauer. Geodesics and geodesic deviation for impulsive gravitational waves. J. Math. Phys., 39(4):2201 --
2212, 1998.
[35] R. Steinbauer and J. A. Vickers. The use of generalized functions and distributions in general relativity.
Classical Quantum Gravity, 23(10):R91 -- R114, 2006.
[36] B. Z. Steinberg and J. J. McCoy. Toward local effective parameter theories using multiresolution decomposition.
Jour. Acoustic Soc. Am., 96(2):1130 -- 1143, 1994.
[37] M. Stojanovi´c. Nonlinear Schrodinger equation with singular potential and initial data. Nonlinear Anal.,
64(7):1460 -- 1474, 2006.
[38] M. Stojanovi´c. Perturbed Schrodinger equation with singular potential and initial data. Commun. Contemp.
Math., 8(4):433 -- 452, 2006.
11
[39] L. N. Trefethen and L. Halpern. Well-posedness of one-way wave equations and absorbing boundary conditions.
Math. Comp., 47:421 -- 435, 1986.
[40] R. van der Hilst, M. V. de Hoop, P. Wang, S.-H. Shim, P. Ma, and L. Tenorio. Seismo-stratigraphy and thermal
structure of earth's core-mantle boundary region. Science, 315:1813 -- 1817, 2007.
[41] S. C. Wales and J. J. McCoy. A comparison of parabolic wave theories for linearly elastic solids. Wave Motion,
5:99 -- 113, 1983.
[42] K. Wapenaar. Seismic reflection and transmission coefficients of a self-similar interface. Geophys. J. Int.,
135:585 -- 594, 1998.
Fakultat fur Mathematik, Universitat Wien
E-mail address: [email protected]
12
|
1801.08885 | 6 | 1801 | 2018-07-25T07:18:12 | Fractional powers and singular perturbations of quantum differential Hamiltonians | [
"math.FA",
"math-ph",
"math-ph"
] | We consider the fractional powers of singular (point-like) perturbations of the Laplacian, and the singular perturbations of fractional powers of the Laplacian, and we compare such two constructions focusing on their perturbative structure for resolvents and on the local singularity structure of their domains. In application to the linear and non-linear Schr\"{o}dinger equations for the corresponding operators we outline a programme of relevant questions that deserve being investigated. | math.FA | math |
FRACTIONAL POWERS AND SINGULAR PERTURBATIONS OF
QUANTUM DIFFERENTIAL HAMILTONIANS
ALESSANDRO MICHELANGELI, ANDREA OTTOLINI, AND RAFFAELE SCANDONE
Dedicated to Gianfausto Dell'Antonio on the occasion of his 85th birthday
Abstract. We consider the fractional powers of singular (point-like) pertur-
bations of the Laplacian, and the singular perturbations of fractional powers of
the Laplacian, and we compare such two constructions focusing on their per-
turbative structure for resolvents and on the local singularity structure of their
domains. In application to the linear and non-linear Schrodinger equations for
the corresponding operators we outline a programme of relevant questions that
deserve being investigated.
1. Background: at the edge of fractional quantum mechanics and
zero-range interactions
At the edge of the theory of quantum Hamiltonians with zero-range interactions,
the theory of partial differential operators, and the recent mainstream of fractional
quantum mechanics, there are two constructions, each of which is classical in nature,
the combination of which has been receiving an increasing attention in the recent
times.
The first is the construction of fractional powers of a differential operator with
non-negative symbol and more generally the fractional power of a non-negative
self-adjoint operator on L2(Rd). For concreteness, let us focus on the negative
Laplacian
−∆ = − d(cid:88)
j=1
∂2
∂x2
j
,
x ≡ (x1, . . . , xd) ∈ Rd ,
that we shall simply call the Laplacian. In this case the definition of (−∆)s/2 for
s ∈ R is obvious in terms of the corresponding power of the Fourier multiplier for
−∆:
((−∆)s/2f )(x) = (ps(cid:98)f )∨(x) ,
2(cid:82)
where (Ff )(p) ≡ (cid:98)f (p) = (2π)− d
Rd e−ixpf (x)dx is the present convention for
In fact, F(−∆)F∗ gives an explicit multiplication form
the Fourier transform.
and hence an explicit spectral decomposition for −∆ as a non-negative self-adjoint
operator on L2(Rd), thus (−∆)s/2 given by the identity above coincides with the
construction with functional calculus.
The second construction is that kind of perturbation of a given pseudo-differential
operator that heuristically amounts to add to it a potential supported at one point
only, whence the jargon of singular perturbation [2, 3]. This is a typical restriction-
extension construction, where first one restricts the initial operator to sufficiently
smooth functions vanishing in neighbourhoods of a given point x0 ∈ Rd, and then
one builds an operator extension of such restriction that is self-adjoint on L2(Rd).
Date: July 26, 2018.
Key words and phrases. Fractional Quantum Mechanics, fractional Laplacian, singular per-
turbations of differential operators, Kreın-Visik-Birman self-adjoint extension theory.
1
2
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
This procedure, when the initial operator is −∆, is known to be equivalent to the
somewhat more concrete scheme of taking the limit ε ↓ 0 in Schrodinger operators
of the form −∆ + Vε, where Vε is a regular potential essentially supported at a scale
ε−1 around x0 and magnitude diverging with ε, which shrink to a delta-like profile
centred at x0 [2, 3].
Both constructions are well known and relevant in several contexts: Sobolev
spaces, fractional Sobolev norms, high and low regularity theory for non-linear
PDE, etc., concerning the former; solvable yet realistic models with computable
spectral features (eigenvalues, scattering matrix,...) in quantum mechanics, chem-
istry, biology, acoustics, concerning the latter.
Recently, especially for the solution theory of non-linear Schrodinger equations
whose linear part is governed by singular Hamiltonians of point interactions[8, 12],
as well as for linear Schrodinger-like equations with singular perturbations of frac-
tional powers of the Laplacian [14, 16, 4, 11, 18, 20, 10, 15, 17], the interest has
increased around various ways of combining the two constructions above.
The goal of this work is to set up the problem in a systematic way for two
operations that in a sense do not commute, to make the rigorous constructions of
the operators of interest, and to make qualitative and quantitative comparisons.
We shall consider, in the appropriate sense that we are going to specify, the class
of singular perturbations of the d-dimensional Laplacian, supported at a point x0 ∈
Rd, in the homogeneous and inhomogeneous case, namely (informally speaking),
(1.1)
hτ = −∆ + singular perturbation at x0
hτ + 1 = −∆ + 1 + singular perturbation at x0 ,
and then we shall combine this construction with that of fractional powers, thus
considering on the one hand the operators
(1.2)
τ = (−∆ + singular perturbation at x0)s/2
hs/2
(hτ + 1)s/2 = (−∆ + 1 + singular perturbation at x0)s/2
and on the other hand the operators
(1.3)
k(s/2)
τ
d(s/2)
τ
= (−∆)s/2 + singular perturbation at x0
= (−∆ + 1)s/2 + singular perturbation at x0 .
All operators in (1.1) -- (1.3) above are meant to be self-adjoint.
Let us remark that hs/2
τ
and (hτ + 1)s/2 are going to be genuine fractional powers
of a non-negative self-adjoint operator on L2(R3) (to this aim one has to consider
non-restrictively only those singular perturbations that produce non-negative oper-
ators), whereas the different notation for the superscript s/2 in k(s/2)
and in d(s/2)
is to indicate that the latter operators are instead singular perturbations of s/2-th
powers (and not fractional powers of singular perturbations).
, for some J ∈ N, is going to be a parameter
that qualifies one element in the infinite family of self-adjoint realisations of the
considered singular perturbation, with the customary convention that 'τ = ∞'
denotes the absence of perturbation.
In all cases τ ∈ (R ∪ {∞})J 2
τ
τ
As for the choice of the point x0, there is no loss of generality in choosing x0 = 0,
which we shall do henceforth.
The knowledge of singular quantum Hamiltonians of the type (1.1) is well es-
tablished in the literature and we review them in Section 2. The study of their
fractional powers, hence of the operators of the type (1.2), has only started re-
cently and in the second part of Section 2 we give an account of the main known
facts about them.
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS
3
In comparison to (1.1) and (1.2), we shall then discuss the rigorous construction
of operators of the type (1.3) (Sections 3 -- 7). Our presentation will have two main
focuses, which reflect into the formulation of our results:
1. to qualify the nature of the perturbation in the resolvent sense (finite rank
vs infinite-rank perturbations);
2. to qualify the natural decomposition of the domain of the considered oper-
ators into a regular component and a singular component, and to determine
the boundary condition constraining such two components.
The first issue is central for deducing an amount of properties from the unperturbed
to the perturbed operators. The second issue also arises naturally, as one can see
heuristically that the considered operators must act in an ordinary way on those
functions supported away from the perturbation centre, and therefore their domains
must contain a subspace of H s-regular functions, where s is the considered power,
next to a more singular component that is the signature of the perturbation.
In this respect we are going to highlight profound differences between two con-
structions, say,
(−∆ + 1 + singular perturbation at x0)s/2
(−∆ + 1)s/2 + singular perturbation at x0 ,
that are therefore 'non-commutative'.
We organized the material as follows:
nians is presented in Section 2;
• the analysis of operators of fractional power of point interaction Hamilto-
• on the other hand, the construction of singular perturbations of fractional
Laplacians is presented in Sections 3 through 7, with the general set-up in
Section 3, the detailed discussion of the paradigmatic scenario of rank-one
perturbations in Sections 4 and 5 (homogeneous case) and in Section 6 (in-
homogeneous case), and an outlook on high-rank perturbations in Section
7;
• last, in Section 8 we outline an amount of relevant questions that deserve
being investigated in application to the linear and non-linear Schrodinger
equations governed by the operators constructed in this work.
2. Singular perturbations of the Laplacian and their fractional
powers
In this Section we qualify the operators hτ , hs/2
, and (hτ + 1)s/2 of (1.1)-(1.2).
Whereas the former is well known in all dimensions in which it is not trivial, namely
d = 1, 2, 3, the latter two operators have been studied mainly in three dimensions,
thus in this Section we choose d = 3 for our presentation.
τ
(cid:16)
(cid:17)∨
λx
e−√
4πx =
1
1
(2π) 3
2
p2 + λ
(x) ,
x, p ∈ R3 ,
For chosen λ > 0 we set
(2.1)
Gλ(x) :=
whence
(−∆ + λ)Gλ = δ(x)
(2.2)
as a distributional identity on R3. We also set
h := −∆ (cid:22) C∞
(2.3)
as an operator closure with respect to the Hilbert space L2(R3).
0 (R3\{0})
4
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
As well known[13], h is densely defined, closed, and symmetric on L2(R3), with
(2.4)
D(h) =
f ∈ H 2(R3)
hf = −∆f .
Its Friedrichs extension given by the self-adjoint Laplacian on L2(R3) with domain
H 2(R3), and its adjoint is the operator
η
(p2 + λ)2 +
η, ξ ∈ C
ξ
p2 + λ
f λ ∈ D(h) ,
(cid:110)
,
R3
(cid:12)(cid:12)(cid:12)(cid:90)
(cid:111)
(cid:98)f (p) dp = 0
g ∈ L2(R3)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:98)g(p) =(cid:99)f λ(p) +
(cid:17)
(cid:16)(cid:99)f λ(p) +
η
.
= H 2(R3) (cid:117) span{Gλ}
(p2 + λ)2
D(h∗) =
(2.5)
(cid:92)
(h∗ + λ)g) (p) = (p2 + λ)
(
The structure in Eq. (2.5) is typical of a well-known decomposition (see, e.g.,
Eq. (2.5) and (2.10) in [7]). We observe that D(h∗) does not depend on λ, only the
splitting of g ∈ D(h∗) into a H 2-function plus a less regular component does. The
last identity in (2.5) may be re-interpreted distributionally as
h∗g = −∆g + (2π)
3
2 ξ δ(x)g ,
where neither −∆g nor (2π) 3
2 ξ δ(x)g belongs to L2(R3), but their sum does.
The fact that ker(h∗ + λ1) = span{Gλ} indicates that h has deficiency index 1
and hence admits a one-(real-)parameter family of self-adjoint extensions. They can
be classified in terms of the standard parametrisation of the Kreın-Visik-Birman
self-adjoint extension theory (see, e.g.,[7, Sec. 3]), identifying each of them as a
restriction of h∗ by imposing in (2.5) Birman's self-adjointness condition η = τ ξ
for some τ ∈ R ∪ {∞} (see, e.g., [13, Theorem 1 and Corollary 1]), whence the
following Theorem.
Theorem 2.1. Let λ > 0.
(i) The self-adjoint extensions in L2(R3) of the operator h form the family
(hτ )τ∈R∪{∞}, where h∞ = hF , the Friedrichs extension, and all other
(proper) extensions are given by hτ = h∗ (cid:22) D(hτ ), where
g ∈ L2(R3)
(cid:110)
g = F λ + 8π
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:98)g(p) =(cid:99)f λ(p) +
ξ ∈ C ,
√
τ F λ(0) Gλ
λ
ξ
p2 + λ
τ ξ
(p2 + λ)2 +
f λ ∈ D(h)
(cid:12)(cid:12)(cid:12) F λ ∈ H 2(R3)
(cid:111)
.
(2.6)
D(hτ ) :=
=
(ii) Each extension is semi-bounded from below and
(2.7)
inf σ(hτ + λ1) (cid:62) 0 ⇔ τ (cid:62) 0
inf σ(hτ + λ1) > 0 ⇔ τ > 0
(hτ + λ1) is invertible ⇔ τ (cid:54)= 0 .
(iii) For each τ ∈ R the quadratic form of the extension hτ is given by
(2.8)
(2.9)
D[hτ ] = H 1(R3) (cid:117) span{Gλ}
hτ [F λ + κλGλ] = (cid:107)∇F λ(cid:107)2
√
τ
8
L2(R3) − λ(cid:107)F λ + κλGλ(cid:107)2
κλ2
for any F λ ∈ H 1(R3) and κλ ∈ C.
+
λ
L2(R3) + λ(cid:107)F λ(cid:107)2
L2(R3)
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS
5
Remark 2.2. The τ -parametrisation depends on the initial choice of the λ-shift
and thus the same extension is described by infinitely many different pairs (λ, τ ) --
of course with a unique τ once λ is chosen. This is clear by inspecting the boundary
condition at x = 0 between regular and singular component of a generic g ∈ D(hτ ):
for any two pairs (λ, τ ) and (λ(cid:48), τ(cid:48)) identifying the same extension hτ , owing to
(2.6) one has
(p2+λ)(p2+λ(cid:48)) ,
ξ
√
ξ(cid:48)
ξ(cid:48)
p2 + λ
:= 8π
λ
τ(cid:48) (2π)3
p2 + λ(cid:48)
(cid:98)g = (cid:99)F λ +
(cid:82)
R3 (cid:99)F λdp = τ (2π)3
(cid:82)
R3(cid:100)F λ(cid:48)
(cid:0)1 + 2
τ − 2λ√
= (cid:100)F λ(cid:48)
(cid:0)(cid:82)
R3 (cid:99)F λdp(cid:1) , (cid:100)F λ(cid:48)
:= (cid:99)F λ + ξ(cid:48)(λ(cid:48)−λ)
(cid:0)1 + 2
(cid:1) ,
(cid:1), or equivalently
τ(cid:48) − 2λ(cid:48)√
λ(cid:48)
√
dp = τ(cid:48) (2π)3
λ(cid:48) ξ(cid:48) · τ
λ(cid:48)
√
τ(cid:48)√
λ
λ(cid:48)−√
√
√
λ(cid:48)
τ(cid:48)√
λ
λ(cid:48)−√
8π
λ
√
√
λ(
√
λ(
τ
=
√
ξ
λ)
τ
8π
+
λ)
λ
.
with
and also
whence necessarily 1 = τ
(2.10)
Thus, when referring to the extension hτ , we shall always implicitly declare the
choice of λ, and any other λ(cid:48) and τ(cid:48) satisfying (2.10) actually identify the same
extension.
In the literature it is customary to find the second expression of (2.6) with the
alternative extension parameter α defined by
τ − 2λ
√
λ
8π
α :=
(2.11)
thus re-writing (hτ )τ∈R∪{∞} as (hα)α∈R∪{∞} with hα := h∗ (cid:22) D(hα) (see [13, Re-
mark 3] and [2, Eq. (I.1.1.26)]). Of course, owing to Remark 2.2, in particular to
formula (2.10), the parameter α identifies unambiguously an extension irrespec-
tively of the shift λ chosen for the explicit domain decomposition. From this and a
bit of further spectral analysis [2, Sec. I.1.1] one then deduces the following.
,
Theorem 2.3.
(i) The self-adjoint extensions in L2(R3) of the operator h form the family
(hα)α∈R∪{∞}, where h∞ = hF , the Friedrichs extension, and all other
(proper) extensions are given by hα := h∗ (cid:22) D(hα), where, for arbitrary
λ > 0
D(hα) =
g = F λ + (α +
√
4π )−1F λ(0) Gλ
λ
(cid:12)(cid:12)(cid:12) F λ ∈ H 2(R3)
(cid:111)
(2.12)
(hα + λ) g = (−∆ + λ) F λ .
(ii) The spectrum of hα is given by
σess(hα) = σac(hα) = [0, +∞) ,
σsc(hα) = ∅ ,
(2.13)
σp(hα) =
∅
{−(4πα)2}
if α ∈ [0, +∞]
if α ∈ (−∞, 0) .
(cid:110)
(cid:40)
The negative eigenvalue −(4πα)2, when it exists, is non-degenerate and the
corresponding eigenfunction is x−1e−4πα x. Thus, α (cid:62) 0 corresponds to
a non-confining, 'repulsive' contact interaction.
6
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
(iii) For each α ∈ R the quadratic form of the extension hα is given by
D[hα] = H 1(R3) (cid:117) span{Gλ}
(2.14)
hα[F λ + κλ Gλ] = −λ(cid:107)F λ + κλ Gλ(cid:107)2
L2(R3)
+ (cid:107)∇F λ(cid:107)2
L2(R3) + λ(cid:107)F λ(cid:107)2
L2(R3) +(cid:0)α +
(cid:1)κλ2
√
λ
4π
for arbitrary λ > 0.
(iv) The resolvent of hα is given by
(2.15)
(hα + λ1)−1 = (−∆ + +λ1)−1 + (α +
√
4π )−1Gλ(cid:105)(cid:104)Gλ
λ
for arbitrary λ > 0.
For the operators hα + λ1 with λ sufficiently large, and the operators hα with
α (cid:62) 0, the self-adjoint functional calculus defines the powers (hα + λ1)s/2 or hs/2
for s ∈ R.
A surely relevant regime is surely s ∈ [0, 2]: the integer powers s = 2, 1, 0
correspond, respectively, to the considered operator, its square root whose domain
is then the form domain of the considered operator, and the identity; the fractional
powers in between are of interest when one needs to discuss the corresponding linear
or non-linear dynamics in spaces of convenient fractional regularity.
α
From the thorough analysis of such fractional powers made in [8], one has the
following.
Theorem 2.4. Let α (cid:62) 0, λ > 0, and s ∈ [0, 2], and set
(i) One has
α ) = D(hα + λ1)s/2
(cid:101)H s
α(R3) := D(hs/2
(cid:107)g(cid:107)(cid:101)H s
α(R3) := (cid:107)(hα + λ1)s/2g(cid:107)L2(R3) .
(cid:101)H s
α(R3) = H s(R3)
(cid:107)ψ(cid:107)(cid:101)H s
α(R3) ≈ (cid:107)ψ(cid:107)H s(R3)
(cid:101)H s
α(R3) = H s(R3) (cid:117) span{Gλ}
(cid:110)
(cid:107)F λ + κλ Gλ(cid:107)(cid:101)H s
g = F λ +(cid:0)α +
(cid:1)−1
α(R3) ≈ (cid:107)F λ(cid:107)H s(R3) + (1 + α)κλ
(cid:101)H s
α(R3) =
(cid:107)F λ + κλ Gλ(cid:107)(cid:101)H s
α(R3) ≈ (cid:107)F λ(cid:107)H s(R3)
where here '≈' denotes the equivalence of norms.
F λ(0) Gλ
λ
4π
√
(2.16)
(2.17)
(2.18)
(2.19)
if s ∈ [0, 1
2 ) ,
(cid:12)(cid:12)(cid:12) F λ ∈ H s(R3)
(cid:111)
if s ∈ ( 1
2 , 3
2 ) ,
if s ∈ ( 3
2 , 2) ,
(ii) For g ∈ D(hs/2
α ),
(hα + λ1)s/2g = (−∆ + λ1)s/2g
(2.20)
− 4 sin sπ
2
(iii) One has the resolvent identity
(2.21)
(hα + λ1)−s/2 = (−∆ + λ1)−s/2 −
(cid:90) +∞
0
dt
(cid:90) +∞
0
ts/2 (cid:104)Gλ+t, g(cid:105)L2(R3)
√
4πα +
λ + t
Gλ+t .
− s
4t
4πα+
2 sin
√
sπ
2
λ+t
Gλ+t(cid:105)(cid:104)Gλ+t .
dt
2 and s = 3
The transition cases s = 1
2 too can be qualified, however with less
explicit formulas -- see [8, Prop. 8.1 and 8.2].
α(R3) decomposes into a regular
H s-component and a singular component with local singularity x−1, precisely as
the domain of hα itself, and for s > 3
2 regular and singular parts are constrained by
Thus, as described in Theorem 2.4, for s > 1
2 , (cid:101)H s
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS
7
a local boundary condition among them of the same type as in (2.12); low powers
s < 1
2 , instead, only produce domains where no leading singularity can be singled
out.
It is remarkable that one has such an explicit and clean knowledge of the frac-
tional powers of the singular perturbations of the Laplacian: the singular pertur-
bation yields an operator hα that is not a (pseudo-)differential operator any longer
and its powers are not simply recovered as Fourier multipliers, nor is it a priori ob-
vious how the fractional power affects the local boundary condition between regular
and singular components of elements in D(hα).
α , and (hα + λ1)s/2 for s ∈
It is also worth remarking that whereas hα, hs/2
2 , 2) share the same singular behaviour x−1 of the functions in their domain, one
( 1
noticeable difference is given by equations (2.15) and (2.21): indeed, (hα + λ1)−1
is a rank-one perturbation of (−∆ + λ1)−1, while (hα + λ1)−s/2 is not a finite rank
perturbation of (−∆ + λ1)−s/2.
3. Singular perturbations of the fractional Laplacian: general
setting for the homogeneous case
τ
In comparison to hs/2
, the fractional power of a singular perturbation of the
Laplacian, we start discussing in this Section the rigorous construction of operators
of the type k(s/2)
, as introduced informally in (1.3), namely the self-adjoint singular
perturbations of the fractional power of the Laplacian. Then, in Section 4 we shall
consider the concrete cases of dimension d = 3, and in Section 5 the case d = 1.
τ
For chosen d ∈ N, λ > 0, and s ∈ R we set
(3.1)
Gs,λ(x) :=
1
(2π) d
2
1
ps + λ
(x) ,
x, p ∈ Rd ,
(cid:16)
(cid:17)∨
whence
(3.2)
((−∆)s/2 + λ) Gs,λ = δ(x)
as a distributional identity on Rd. We also set
(3.3)
k(s/2) := (−∆)s/2 (cid:22) C∞
0 (Rd\{0})
as an operator closure with respect to the Hilbert space L2(Rd). Thus, in compar-
ison to Section 2, when d = 3 one has G2,λ = Gλ and k(1) = h.
The domain of k(s/2), as a consequence of the operator closure in (3.3), is a space
of functions with H s-regularity and vanishing conditions at x = 0 for each function
and its partial derivatives. The amount of vanishing conditions depends on d and
s, to classify which we introduce the intervals
(cid:40)
(3.4)
I (d)
n
:=
(0, d
2 )
2 + n − 1, d
( d
n = 0
2 + n) n = 1, 2, . . .
For our purposes it is convenient to use momentum coordinates to express the
vanishing conditions that qualify the domain of k(s/2): thus, with the notation
p ≡ (p1, . . . , pd) ∈ Rd, by means of an approximation argument (see Appendix A)
8
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
we find
(3.5)
D(k(s/2)) = H s
=
0 (Rd\{0}) = C∞
H s(Rd)
0 (Rd\{0})
(cid:107) (cid:107)Hs
f ∈ H s(Rd) such that
Rd pγ1
(cid:82)
d (cid:98)f (p) dp = 0
γ1, . . . , γd ∈ N0 , (cid:80)d
d (cid:98)f (p) dp = 0 is the same as(cid:0) ∂γ1
j=1 γj (cid:54) n − 1
1 ··· pγd
0
if s ∈ I (d)
if s ∈ I (d)
··· ∂γd
∂x
γd
d
∂xγ1
1
Clearly,(cid:82)
Rd pγ1
1 ··· pγd
notation x ≡ (x1, . . . , xd) ∈ Rd.
n , n = 1, 2, . . .
f(cid:1)(0) = 0, with the
The expression of D(k(s/2)) for the endpoint values s = d
2 + n require an amount
of extra analysis besides the arguments proof of (3.5): we do not discuss it here,
an omission that does not affect the conceptual structure of our presentation.
Being densely defined, closed, and positive, either the symmetric operator k(s/2)
is already self-adjoint on L2(Rd), or it admits infinitely many self-adjoint extensions.
As well known, the infinite multiplicity of such extensions is quantified by the
deficiency index of k(s/2), which is the quantity
J (s, d) := dim ker(cid:0)(k(s/2))∗ + λ1(cid:1)
(3.6)
for one, and hence for all λ > 0. The self-adjointness of k(s/2) is equivalent to
J (s, d) = 0.
It is not difficult to compute J (s, d) for generic values of d and s and to identify
a natural basis of the J (s, d)-dimensional space ker(cid:0)(k(s/2))∗ + λ1(cid:1).
Lemma 3.1. For given d ∈ N and s > 0,
(3.7)
s ∈ I (d)
n ⇒ J (s, d) =
In particular, when s ∈ I (d)
(cid:110)
ker(cid:0)(k(s/2))∗ + λ1(cid:1) = span
(3.8)
n
for some n ∈ N, then
uλ
γ1,...,γd
(cid:18)d + n − 1
(cid:19)
(cid:12)(cid:12)(cid:12) γ1, . . . , γd ∈ N0 ,
d
.
(cid:111)
,
γj (cid:54) n − 1
d(cid:88)
j=1
where
(3.9)
1 ··· pγd
pγ1
ps + λ
It is worth noticing, comparing (3.1) and (3.9), that
γ1,...,γd (p) :=
d
.
(cid:99)uλ
s ∈ I (d)
to
uλ
0,...,0 = (2π)
(3.10)
Proof of Lemma 3.1. When s ∈ I (d)
then ker(cid:0)(k(s/2))∗ + λ1(cid:1) is trivial and J (s, d) = 0, consistently with (3.7). When
n , n = 1, 2, . . . , then u ∈ ker(cid:0)(k(s/2))∗ +λ1(cid:1) = ran(cid:0)k(s/2) +λ1(cid:1)⊥
0 , we see from (3.5) that k(s/2) is self-adjoint:
is equivalent
2 Gs,λ .
d
and one argues from (3.5) that ker(cid:0)(k(s/2))∗ + λ1(cid:1) is spanned by linearly indepen-
∀f ∈ D(k(s/2))
0 =
dent functions of the form uλ
j=1 γj (cid:54) n − 1. Such functions are as
(cid:90)
R3(cid:98)u(p)(ps + λ)(cid:98)f (p)dp
, with(cid:80)d
γ1,...,γd
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS
9
(cid:18)d + n − 1
(cid:19)
d
many as the linearly independent monomials in d variables with degree at most
equal to n − 1, and therefore their number equals
(cid:3)
.
The knowledge of ker(cid:0)(k(s/2))∗+λ1(cid:1) and of the inverse of the Friedrichs extension
of k(s/2) are the two inputs for the Kreın-Visik-Birman extension theory (see, e.g.,
[7, Sec. 3]), by means of which we can produce the whole family of self-adjoint
extensions of k(s/2).
Such a construction is particularly clean in the case, relevant in applications, of
deficiency index one: the comprehension of this case is instructive to understand
the case of higher deficiency index. Moreover, as we shall see, in this case the self-
adjoint extensions of k(s/2) turns out to be rank-one perturbations, in the resolvent
sense: we will use the jargon J = 1 or 'rank one' interchangeably.
the J (s, d) = 1 scenario when s ∈ I (d)
high-J scenario. This corresponds to analysing the regimes s ∈ ( 1
s ∈ (1, 2) when d = 2, s ∈ ( 3
Therefore, in this work we make the presentational choice to discuss in detail
1 , deferring to Section 7 an outlook on the
2 ) when d = 1,
2 , 3
2 ) when d = 3, etc.
Indeed, when s ∈ I (d)
(3.10) that ker(cid:0)(k(s/2))∗ +λ1(cid:1) = span{Gs,λ}, and the function Gs,λ may or may not
The construction of the self-adjoint extensions of k(s/2) in any such regimes is
technically the very same, irrespectively of d, except for a noticeable peculiarity
when d = 1, as opposite to d = 2, 3, . . .
and hence J (s, d) = 1, we know from Lemma 3.1 and
have a local singularity as x → 0. As follows from the d-dimensional distributional
identity
2 , 5
1
2
s
2 Γ( s
2 )
1
ps = 2
d−s
2 Γ( d−s
2 )
,
s ∈ (0, d) ,
(cid:17)
(cid:92)(cid:16)
1
xd−s
Gs,0 has a singularity ∼ x−(d−s) when s < d, it has a logarithmic singularity when
s = d, and it is continuous at x = 0 when s > d, with asymptotics
(3.11)
xd−s Gs,λ(x)
Gs,λ(x)
:=
x→0−−−−→ Λ(d)
x→0−−−−→ Gs,λ(0) = (cid:0)2d−1π
(2π) d
s
Γ( d−s
2 )
2 2s− d
,
2 Γ( d
2 )
2 −1Γ( d
2 ) λ
d
(cid:1)−1
s ∈ (0, d)
,
s > d .
s−d
s s sin πd
s
Now, all the considered regimes s ∈ (1, 2) when d = 2, s ∈ ( 3
2 ) when d = 3,
etc. lie below the transition value s = d between the local singular and the local
regular behaviour of Gs,λ, whereas the regime s ∈ ( 1
2 ) when d = 1 lies across the
transition value s = 1.
of ker(cid:0)(k(s/2))∗ + λ1(cid:1) for higher deficiency index J (s, d).
The same type of distinction clearly occurs for the spanning functions (3.8)-(3.9)
2 , 3
2 , 5
In the present context, the peculiarity described above when d = 1 results in
certain different steps of the construction of the self-adjoint extensions of k(s/2) and
ultimately in the type of parametrisation of such extensions, as we shall see.
Therefore, we articulate our discussion on the extensions of k(s/2) when the
deficiency index is one discussing first the three-dimensional case (Section 4) and
then the one-dimensional case (Section 5). As commented already, for generic d (cid:62) 2
the discussion and the final results are completely analogous to d = 3.
10
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
4. Rank-one singular perturbations of the fractional Laplacian:
homogeneous case in dimension three
In terms of the general discussion of Sec. 3, we consider here the operator k(s/2)
2 ). k(s/2) acts as the fractional Laplacian (−∆)s/2 on the
2 , 5
on L2(R3) when s ∈ ( 3
domain
(cid:110)
(cid:12)(cid:12)(cid:12)(cid:90)
R3
(cid:111)
(cid:98)f (p) dp = 0
(4.1)
D(k(s/2)) =
f ∈ H s(R3)
and its deficiency index is 1.
One has the following construction.
Theorem 4.1. Let s ∈ ( 3
2 , 5
2 ) and λ > 0.
(i) The self-adjoint extensions in L2(R3) of the operator k(s/2) form the family
(k(s/2)
)τ∈R∪{∞}, where k(s/2)∞ is its Friedrichs extension, namely the self-
adjoint fractional Laplacian (−∆)s/2, and all other extensions are given
by
τ
D(cid:0)k(s/2)
(cid:1) :=
τ
(4.2)
(4.3)
and
=
(cid:0)k(s/2)
τ
τ ξ
g ∈ L2(R3)
(cid:110)
+ λ1(cid:1)g := F−1(cid:16)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:98)g(p) =(cid:99)f λ(p) +
ξ ∈ C , f λ ∈ H s(R3) , (cid:82)
(cid:16)(cid:99)f λ +
g = F λ + 2πs2 sin( 3π
τ (s−3)
(ps + λ)
F λ(0) Gs,λ
s )λ2− 3
τ ξ
s
(ps + λ)2 +
ξ
ps + λ
R3(cid:99)f λ(p) dp = 0
(cid:12)(cid:12)(cid:12) F λ ∈ H s(R3)
(cid:111)
(cid:17)(cid:17)
,
(ps + λ)2
.
(ii) Each extension is semi-bounded from below and
τ
inf σ(k(s/2)
inf σ(k(s/2)
+ λ1) (cid:62) 0 ⇔ τ (cid:62) 0
+ λ1) > 0 ⇔ τ > 0
+ λ1) is invertible ⇔ τ (cid:54)= 0 .
τ
(k(s/2)
τ
(4.4)
(iii) For each τ ∈ R the quadratic form of the extension k(s/2)
τ
is given by
(4.5)
(4.6)
k(s/2)
τ
D[k(s/2)
τ
] = H
s
2 (R3) (cid:117) span{Gs,λ}
[F λ + κλGs,λ] = (cid:107)∇ s
L2(R3) − λ(cid:107)F λ + κλGs,λ(cid:107)2
2 F λ(cid:107)2
+λ(cid:107)F λ(cid:107)2
τ (s − 3)
L2(R3)
κλ2
L2(R3) +
2πs2λ2− 3
s sin( 3π
s )
for any F λ ∈ H s/2(R3) and κλ ∈ C.
(iv) For τ (cid:54)= 0, one has the resolvent identity
+ λ1)−1 = ((−∆)s/2 + λ1)−1 +
(k(s/2)
(4.7)
s )λ2− 3
s
2πs2 sin( 3π
τ (s − 3)
Gs,λ(cid:105)(cid:104)Gs,λ .
τ
Proof. The whole construction is based upon the Kreın-Visik-Birman self-adjoint
extension of k(s/2) + λ1 is the Fourier multiplier (ps + λ), one has the following
formula for the adjoint (see, e.g., [7, Theorem 2.2]):
extension scheme. Since ker(cid:0)(k(s/2))∗ + λ1(cid:1) = span{Gs,λ} and the Friedrichs
g ∈ L2(R3)
D(cid:0)(k(s/2))∗(cid:1) =
(cid:0)(k(s/2))∗ + λ1(cid:1)g = F−1(cid:16)
(cid:98)g(p) =(cid:99)f λ(p) +
η, ξ ∈ C , f λ ∈ H s(R3) , (cid:82)
(cid:16)(cid:99)f λ +
(ps + λ)2 +
ps + λ
R3(cid:99)f λ(p) dp = 0
(ps + λ)
(cid:17)(cid:17)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
η
η
ξ
.
(ps + λ)2
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 11
transform reads
Each element of the one-parameter family of self-adjoint extensions of k(s/2) is
identified (see, e.g., [7, Theorem 3.4]) by the Birman self-adjointness condition
η = τ ξ for some τ ∈ R ∪ {∞}. This establishes the first line of (4.2). Setting
(cid:99)F λ := (cid:99)f λ + (ps + λ)−2τ ξ, the boundary condition between F λ and ξ in Fourier
(cid:99)F λ(p) dp = ξ
2(cid:82)
R3 (cid:99)F λdp, and using (3.1) with d = 3, the second line
4π2τ (s−3)
s sin( 3π
s )
is a restriction of (k(s/2))∗, from the above action of
R3
Then, from F λ(0) = (2π)− 3
of (4.2) follows. Since k(s/2)
the adjoint one deduces (4.3). This completes the proof of part (i).
s2λ2− 3
(cid:90)
(*)
.
τ
Part (ii) lists standard facts of the Kreın-Visik-Birman theory -- see [7, Theorems
The quadratic form is characterised in the extension theory [7, Theorem 3.6] by
3.5 and 5.1].
the formulas D[k(s/2)
whence (4.5), and (k(s/2)
whence (4.6). The proof of part (iii) is completed.
] (cid:117) ker(cid:0)(k(s/2))∗ + λ1(cid:1) ('F ' stands for Friedrichs),
] = D[k(s/2)
τ +λ1)[F λ+κλGs,λ] = ((−∆)s/2+λ1)[F λ]+τκλ2(cid:107)Gs,λ(cid:107)2
F
τ
L2(R3),
Kreın's resolvent formula for deficiency index 1 [7, Theorem 6.6] prescribes
f (k(s/2)
τ
+ λ1)−1 = ((−∆)s/2 + λ1)−1 + βλ,τ Gs,λ(cid:105)(cid:104)Gs,λ
for some scalar βλ,τ to be determined, whenever (k(s/2)
τ (cid:54)= 0. Thus, for a generic h ∈ L2(R3), the element g := (k(s/2)
reads, in view of (4.2) and of the resolvent formula above,
τ
+λ1) is invertible, hence for
τ +λ1)−1h ∈ D(k(s/2)
)
τ
(cid:98)g(p) = (cid:99)F λ(p) +
(cid:98)h(p)
(cid:99)F λ(p) :=
ξλ
ps + λ
(cid:90)
(cid:98)h(q)
βλ,τ
(2π)3
,
R3
dq .
ξλ :=
ps + λ
τ (s−3)
2πs2 sin( 3π
qs + λ
The boundary condition (*) for F λ and ξλ then implies 1 = βλ,τ
,
s )λ2− 3
which determines βλ,τ and proves (4.7), thus completing also the proof of (iv). (cid:3)
In analogy to what argued in Remark 2.2, the τ -parametrisation of the family
(k(s/2)
)τ∈R∪{∞} depends on the initially chosen shift λ > 0, meaning that with a
different choice λ(cid:48) > 0 the same self-adjoint realisation previously identified by τ
with shift λ is now selected by a different extension parameter τ(cid:48). In certain contexts
it is more convenient to switch onto a natural parametrisation that identifies one
extension irrespectively of the infinitely many different pairs (λ, τ ) attached to it by
the parametrisation of Theorem 4.1. We shall do it in the next Theorem: observe
that indeed, as compared to Theorem 4.1, here below λ > 0 is arbitrary.
τ
s
Theorem 4.2. Let s ∈ ( 3
2 , 5
2 ).
(i) The self-adjoint extensions in L2(R3) of the operator k(s/2) form the family
(k(s/2)
)α∈R∪{∞}, where k(s/2)∞ is its Friedrichs extension, namely the self-
adjoint fractional Laplacian (−∆)s/2, and all other extensions are given,
for arbitrary λ > 0, by
α
g = F λ +
D(k(s/2)
α
) =
(4.8)
F λ(0)
α − λ
3
s
−1
2πs sin( 3π
s )
Gs,λ
(k(s/2)
α
+ λ) g = ((−∆)s/2 + λ) F λ .
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) F λ ∈ H s(R3)
12
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
(ii) For each α ∈ R the quadratic form of the extension k(s/2)
α
is given by
(4.9)
(4.10)
k(s/2)
α
D[k(s/2)
α
] = H
s
2 (R3) (cid:117) span{Gs,λ}
[F λ + κλGs,λ] = (cid:107)∇ s
L2(R3) − λ(cid:107)F λ + κλGs,λ(cid:107)2
L2(R3)
2 F λ(cid:107)2
L2(R3) +(cid:0)α − λ
−1
3
s
2πs sin( 3π
s )
(cid:1)κλ2
+ λ(cid:107)F λ(cid:107)2
for arbitrary λ > 0.
(iii) The resolvent of k(s/2)
α
is given by
(4.11)
(k(s/2)
α
+ λ1)−1 = ((−∆)s/2 + λ1)−1
+(cid:0)α − λ
−1
3
s
2πs sin( 3π
s )
(cid:1)−1 Gs,λ(cid:105)(cid:104)Gs,λ
for arbitrary λ > 0.
(iv) Each extension is semi-bounded from below, and
(4.12)
(4.13)
σsc(k(s/2)
α
) = ∅ ,
σess(k(s/2)
α
σdisc(k(s/2)
α
) =
) = σac(k(s/2)
α
) = [0, +∞) ,
(cid:40) ∅
α = −(cid:0)2πα s sin(− 3π
s )(cid:1) s
if α (cid:62) 0
if α < 0 ,
{E(s)
α }
where the eigenvalue E(s)
α is non-degenerate and is given by
3−s ,
the (non-normalised) eigenfunction being Gs,λ=E(s)
α .
E(s)
Proof. Reasoning as in Remark 2.2, we seek for the relation τ = τ (λ) that ensures
that all the pairs (λ, τ (λ)), with λ > 0, preserve the decomposition (4.2)-(4.3) and
thus label the same element of the family of extensions.
For chosen λ and τ , a function g ∈ D(k(s/2)
τ
) decomposes uniquely as
ξ
ps + λ
,
F λ ∈ H s(R3)
ξ ∈ C
(cid:99)F λdp = ξ
4π2τ (s−3)
s sin( 3π
s )
s2λ2− 3
.
Let now λ(cid:48) > 0 and τ(cid:48) ∈ R be such that for the same function g in the domain of
the same self-adjoint realisation k(s/2)
one also has
τ
(cid:98)g = (cid:99)F λ +
(cid:98)g = (cid:100)F λ(cid:48)
+
(cid:90)
(cid:90)
R3
R3
(cid:100)F λ(cid:48)
,
,
ξ(cid:48)
F λ(cid:48) ∈ H s(R3)
ξ(cid:48) ∈ C
The new splitting of g is equivalent to
ps + λ(cid:48) ,
dp = ξ(cid:48)
4π2τ(cid:48)(s−3)
(cid:48)2− 3
s sin( 3π
s )
s2λ
.
ξ(cid:48) = ξ ,
F λ(cid:48)
and the boundary condition for F λ(cid:48)
ξ(cid:48)
(cid:16)
(cid:90)
ξ
4π2τ (s−3)
s sin( 3π
s )
s2λ2− 3
+
ps + λ
R3
ξ
−
ps + λ
= F λ +
and ξ(cid:48) is equivalent to
dp = ξ(cid:48)
(cid:17)
−
ξ(cid:48)
ps + λ(cid:48)
(*)
ξ(cid:48)
ps + λ(cid:48) ,
4π2τ(cid:48)(s−3)
(cid:48)2− 3
s sin( 3π
s )
s2λ
.
Let us analyze the integral in (*). Both summands in the integrand diverge,
with two identical divergences that cancel out. Thus, by means of the identity
r2(rs + λ)−1 = r2−s − λr2−s(rs + λ)−1, one has
(cid:90)
(cid:16)
(cid:17)
1
ps + λ
−
1
ps + λ(cid:48)
R3
dp = 4π lim
R→+∞
(cid:16)(cid:90) R
(cid:16)(cid:90) R
0
(cid:90) R
(cid:90) R
0
−
r2
dr −
rs + λ
λ(cid:48) dr
(cid:17)
r2
rs + λ(cid:48) dr
λ dr
rs−2(rs + λ)
(cid:17)
= 4π lim
0
R→+∞
4π2
s s sin( 3π
s )
λ 1− 3
rs−2(rs + λ(cid:48))
−
λ(cid:48) 1− 3
4π2
s s sin( 3π
s )
0
.
=
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 13
Plugging the result of the above computation into (*) yields
τ (3 − s) − sλ
τ(cid:48)(3 − s) − sλ(cid:48)
λ2− 3
s
=
λ(cid:48)2− 3
s
,
which shows, in complete analogy to (2.10) when s = 2, that all pairs (λ, τ ) such
that
(**)
− τ (3 − s) − sλ
s )λ2− 3
2πs2 sin( 3π
s
=: α
indeed label the same extension (the pre-factor −2πs2 sin( 3π
s ) having being added
for convenience). Thus, α ∈ R∪{∞} defined in (**) is the natural parametrisation
we were aiming for (and the Friedrichs case τ → +∞ corresponds to α → +∞).
= (cid:0)α −
(cid:1)−1
Upon replacing 2πs2 sin( 3π
τ (s−3)
s )λ2− 3
s
1
2πs sin( 3π
s )λ1− 3
s
in the formulas of
Theorem 4.1 we deduce at once all formulas of parts (i), (ii), and (iii), together
of course with the certainty, proved above, that the decompositions are now λ-
independent.
Since the deficiency index is 1, and hence all extensions are a rank-one per-
turbation, in the resolvent sense, of the self-adjoint fractional Laplacian, then all
extensions have the same essential spectrum [0, +∞) of the latter, and additionally
may have at most one negative non-degenerate eigenvalue, in any case all exten-
sions are semi-bounded from below -- all these being general facts of the extension
theory, see, e.g., [7, Theorem 5.9 and Corollary 5.10]. This proves, in particular,
the first line in (4.12).
The occurrence of a negative eigenvalue Eα = −λ of an extension k(s/2)
, for
some λ > 0, can be read out from the resolvent formula (4.11) as the pole of
(k(s/2)
α
α
+ λ1)−1, that is, imposing
α −
i.e.,
= 0 ,
1
2πs sin( 3π
s
3−s
α = −λ
s )λ1− 3
s (cid:0)2πs sin(− 3π
s )(cid:1)−1
λ = (cid:0)2πα s sin(− 3π
s )(cid:1) s
3−s .
.
The identity above can be only satisfied by some λ > 0 when α < 0, because
sin(− 3π
s ) > 0, in which case
Alternatively, one can argue from (4.2)-(4.3) that the eigenvalue −λ must corre-
1ps+λ )∨, that is, an element of the domain with only
spond to the eigenfunction (
singular component, and to the parameter τ = 0, hence with f λ ≡ 0 in the notation
therein. Then, setting τ = 0 in (**) yields the same condition above on α and λ.
This proves (4.13) and the second line in (4.12) when α < 0, and it also qualifies
the eigenfunction.
When such a bound state is absent, and therefore when α (cid:62) 0, for what argued
) = [0, +∞). This proves the second line in
(cid:3)
before one has σ(k(s/2)
(4.12) when α (cid:62) 0, and completes the proof of part (iv).
) = σess(k(s/2)
α
α
Mirroring the observations made in the conclusions of Section 2, we see that
the elements of the domains of both the operators hs/2
(the fractional power of
the singular perturbation of (−∆)) and k(s/2)
(the singular perturbation of the
fractional power (−∆)s/2) split into a regular H s-component plus a singular com-
ponent; however, in the former case the local singularity is x−1 for all considered
2 , 2], whereas in the latter it is the singularity of Gs,λ, namely x−(3−s)
powers s ∈ ( 1
for all powers s ∈ ( 3
α
α
2 , 5
2 ).
14
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
In either case, a local boundary condition constrains regular and singular com-
ponents: working out the asymptotics as x → 0 in (2.12) and (4.8) by means of
(3.11) we find
(4.14)
g(x) ∼ (α + (4πx)−1)
g(x) ∼ (α + Λsx−(3−s))
s ∈ ( 3
2 , 2]
) , s ∈ ( 3
2 , 5
where Λs is defined in (3.11). (Observe that Λs=2 = (4π)−1, consistently.)
g ∈ D(hs/2
α ) ,
g ∈ D(k(s/2)
as x → 0 ,
as x → 0 ,
Furthermore, whereas (hα + λ1)−s/2 is not a finite rank perturbation of (−∆ +
+ λ1)−1 is indeed a rank-one perturbation of ((−∆)s/2 + λ1)−1.
λ1)−s/2, (k(s/2)
2 ) ,
α
α
5. Rank-one singular perturbations of the fractional Laplacian:
homogeneous case in dimension one
In terms of the general discussion of Sec. 3, we consider here now the operator
k(s/2) on L2(R) when s ∈ ( 1
2 )\{1}.
2 , 3
We start with identifying the Friedrichs extension k(s/2)
of k(s/2). Unlike the
depends on whether s < 1 or s > 1.
F
three-dimensional case, the structure of k(s/2)
Proposition 5.1. Let s ∈ ( 1
F
2 , 1) ∪ (1, 3
2 ).
(cid:40)
(i) The quadratic form of the Friedrichs extension k(s/2)
if s ∈ ( 1
2 , 1)
if s ∈ (1, 3
2 )
D[k(s/2)
H s/2(R)
H s/2
[f, g] = (cid:104)∇ s
(R\{0})
2 g(cid:105) .
2 f,∇ s
] =
F
F
0
k(s/2)
F
(5.1)
of k(s/2) is
(ii) When s ∈ ( 1
2 , 1), one has
F
D(k(s/2)
k(s/2)
F
) = H s(R)
f = (−∆)
s
2 f .
(5.2)
(5.3)
(iii) When s ∈ (1, 3
(cid:26)
2 ), for every λ > 0 one has
f = φ − φ(0)
Gs,λ(0)
D(k(s/2)
+ λ1)f = ((−∆)
s
2 + λ)φ .
) =
F
(k(s/2)
F
(cid:27)
(cid:12)(cid:12)(cid:12) φ ∈ H s(R)
Gs,λ
In particular, D(k(s/2)
λ1) has an everywhere defined and bounded inverse on L2(R) with
) ⊂ H s(R)(cid:117)span{Gs,λ}. In this regime of s, (k(s/2)
F +
F
(5.4)
(k(s/2)
F
+ λ1)−1 = ((−∆)
s
2 + λ1)−1 −
1
Gs,λ(0)
Gs,λ(cid:105)(cid:104)Gs,λ .
Proof. Following the standard form construction of the Friedrichs extension (see,
e.g., [7, Theorem A.2]), the Friedrichs form domain is the completion of D(k(s/2))
with respect to the H s
D[k(s/2)
F
] = H s
2 -norm, and therefore
0 (R\{0})
Hs/2(R) =
(cid:107) (cid:107)
(cid:40)
H s/2(R)
H s/2
(R\{0})
0
if s ∈ ( 1
2 , 1)
if s ∈ (1, 3
2 ) ,
last identity being proved precisely as (3.5). Moreover, k(s/2)
for all the sequences (fn)n, (gn)n of H s
whence k(s/2)
[f, g] = (cid:104)∇ s
2 f,∇ s
2 g(cid:105). This completes the proof of part (i).
[f, g] = lim
2 -approximants of f and g respectively,
F
F
n→+∞(cid:104)∇ s
2 fn,∇ s
2 gn(cid:105)
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 15
The self-adjoint operator associated with the form (5.1) is qualified by the for-
mulas
D(k(s/2)
F
+ λ1) =
(k(s/2)
F
+ λ1)f = uf
f ∈ D[k(s/2)
F
]
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:104)∇ s
∃ uf ∈ L2(R) such that
2 f,∇ s
2 g(cid:105) + λ(cid:104)f, g(cid:105) = (cid:104)uf , g(cid:105)
∀ g ∈ D[k(s/2)
]
F
valid for any λ > 0.
unique.) In particular, D(k(s/2)
is. When s ∈ ( 1
f ∈ H s(R) and uf = (−∆) s
When instead s ∈ (1, 3
(By density of H s/2
(R\{0}) in L2(R), the above uf is
) is independent of λ, only its internal decomposition
2 , 1) the condition identifying f and uf is clearly equivalent to
F
0
2 f , which yields part (ii) of the thesis.
(cid:90)
2(cid:98)f p s
p s
R
∀g such that
2 ) the above condition reads
2(cid:98)g dp + λ
(cid:90)
(cid:90)
R
(p s
R
(cid:90)
(cid:90)
(cid:98)f(cid:98)g dp =
R(cid:99)uf(cid:98)g dp
2 + 1)(cid:98)g(p)2dp < +∞ and
(cid:1)(cid:98)g dp = 0
(cid:0)(ps + λ)(cid:98)f −(cid:99)uf
R
(*)
that is,
(cid:90)
R(cid:98)g dp = 0 ,
for all the g's indicated in (*) and for some uf ∈ L2(R). It is easy to see that this
is the same as (cid:98)f (p) = (cid:99)uf (p) + c
ps + λ
,
(cid:99)uf (p) = (ps + λ)(cid:98)f (p) − c
for some c ∈ C. Now, the condition for f to vanish at x = 0 and belong to H s/2(R)
is equivalent to
(cid:90)
R
(cid:12)(cid:12)(cid:12)(p s
2 + 1)(cid:99)uf (p) + c
ps + λ
(cid:12)(cid:12)(cid:12)2
dp .
dp
and
+ ∞ >
The finiteness of the second integral above is guaranteed by s > 1 and uf ∈ L2(R),
whereas the vanishing of the first integral is the same as
(cid:90)
R
0 =
ps + λ
(cid:99)uf (p) + c
(cid:90)
(cid:99)uf (p)
ps + λ
R
(cid:90)
R
(cid:90)
(cid:82)
dp = −c
dp
ps + λ
= −2πc Gs,λ(0) ,
i.e.,
√
2π c = − 1
((−∆)
s
2 + λ)−1uf dx .
R
Gs,λ(0)
Therefore, as a distributional identity,
uf = ((−∆)
In turn, the condition uf ∈ L2(R) is equivalent to φ := ((−∆) s
in terms of such φ the previous identity reads
2π c δ = ((−∆)
2 + λ)f −
s
2 + λ)
(cid:16)
(cid:82)
f +
√
s
(cid:17)
R((−∆) s
2 + λ)−1uf
Gs,λ
.
Gs,λ(0)
2 + λ)−1uf ∈ H s(R):
e the condition f (0) = 0 reads(cid:82)
f = φ −
R φ dx
Gs,λ(0)
Gs,λ ,
R φ dx = φ(0). We have thus found
f = φ − φ(0)
Gs,λ(0)
Gs,λ ,
uf = ((−∆)
s
2 + λ)φ ,
16
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
which proves (5.3). The following inversion formula is then a straightforward con-
sequence:
(k(s/2)
F
+ λ1)−1u = φu − φu(0)
Gs,λ(0)
2 + λ)−1u ∈ H s(R) .
φu := ((−∆)
Gs,λ
s
On the other hand,
φu(0) =
1√
2π
(cid:90)
R
(cid:98)u(p)
ps + λ
dp = (cid:104)Gs,λ, u(cid:105) ,
whence (5.4) and the conclusion of the proof of part (iii).
(cid:3)
As a special case of (5.4) above,
(5.5)
hs,λ :=
=
+ λ1)−1 Gs,λ
(cid:17)
2π (k(s/2)
√
F
F−1(cid:16)
F−1(cid:16)
1
1
(ps + λ)2
(ps + λ)2 − s − 1
2 ) .
and (as follows from (3.11)) Gs,λ(0) = (λ1− 1
if s ∈ ( 1
2 , 1)
if s ∈ (1, 3
ps + λ
λ s
1
(cid:17)
s s sin π
s )−1,
Indeed, (cid:104)Gs,λ, Gs,λ(cid:105) =
whence (cid:107)Gs,λ(cid:107)2
s−1
s s2 sin π
s
λ2− 1
L2(R)/Gs,λ(0) = s−1
λ s .
We can now establish the following construction.
Theorem 5.2. Let s ∈ ( 1
2 ) and λ > 0. Set
(5.6)
ω(s) :=
,
θs :=
0
1
if s < 1
if s > 1 .
2 , 1) ∪ (1, 3
s2 sin ( π
s )
s − 1
(cid:40)
(cid:17)
(cid:1)
D(cid:0)k(s/2)
+ λ1(cid:1)g
(5.7)
τ
(5.8)
(cid:0)k(s/2)
τ
(5.9)
(i) The self-adjoint extensions in L2(R) of the operator k(s/2) form the family
)τ∈R∪{∞}, where k(s/2)∞ is its Friedrichs extension, already qualified
(k(s/2)
in Proposition 5.1, and all other extensions are given by
τ
(cid:41)
(cid:12)(cid:12)(cid:12) F λ ∈ H s(R)
− θs
Gs,λ(0)
F λ(0) Gs,λ
(cid:40)
(cid:16) ω(s)λ2− 1
s
g = F λ +
:=
:= ((−∆)
s
2 + λ)F λ .
τ
(ii) Each extension is semi-bounded from below and
τ
inf σ(k(s/2)
inf σ(k(s/2)
+ λ1) (cid:62) 0 ⇔ τ (cid:62) 0
+ λ1) > 0 ⇔ τ > 0
+ λ1) is invertible ⇔ τ (cid:54)= 0 .
τ
(k(s/2)
τ
(cid:40)
(iii) For each τ ∈ R the quadratic form of the extension k(s/2)
τ
is given by
(5.10)
k(s/2)
τ
(5.11)
D[k(s/2)
H s/2(R) (cid:117) span{Gs,λ}
H s/2
[F λ + κλGs,λ] = (cid:107)∇ s
(R\{0}) (cid:117) span{Gs,λ}
] =
0
τ
if s ∈ ( 1
2 , 1)
if s ∈ (1, 3
2 )
L2(R)
2 F λ(cid:107)2
+λ(cid:107)F λ(cid:107)2
L2(R) − λ(cid:107)F λ + κλGs,λ(cid:107)2
κλ2
L2(R) +
τ
ω(s)λ2− 1
s
for any F λ ∈ D[k(s/2)
F
] and κλ ∈ C.
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 17
(iv) For τ (cid:54)= 0, one has the resolvent identity
(5.12) (k(s/2)
τ
+ λ1)−1 = ((−∆)s/2 + λ1)−1 +
(cid:16) ω(s)λ2− 1
s
τ
(cid:17)Gs,λ(cid:105)(cid:104)Gs,λ .
− θs
Gs,λ(0)
Proof. We proceed along the line of the proof of Theorem 4.1, based upon the
Kreın-Visik-Birman self-adjoint extension scheme and Proposition 5.1.
(I)
The adjoint of k(s/2) is qualified by
D(cid:0)(k(s/2))∗(cid:1) =
g ∈ L2(R)
(cid:0)(k(s/2))∗ + λ1(cid:1)g = (k(s/2)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
F
+ λ1)(f λ + η hs,λ) ,
(cid:98)g(p) =(cid:99)f λ(p) + η(cid:100)hs,λ +
η, ξ ∈ C , f λ ∈ H s(R) , (cid:82)
ξ
ps + λ
R(cid:99)f λ(p) dp = 0
where hs,λ is the function (5.5), and the self-adjoint restrictions of (k(s/2))∗ are
qualified by the self-adjointness condition η = τ ξ for some τ ∈ R∪{∞}. For clarity
of presentation, let us split the discussion into the two regimes s < 1 and s > 1.
First case: s ∈ ( 1
Let F λ := f λ + τ ξ hs,λ. When f λ and ξ run over their possible domains, F λ
2 , 1).
spans the whole Friedrichs domain H s(R). Moreover,
(cid:90)
R
(cid:99)F λ(p) dp = ξ
2πτ
λ2− 1
s ω(s)
.
(II)
Thus, the first line in (I) and (II) yield (5.7). Owing to (5.2) and to the fact that
is a restriction of (k(s/2))∗, one deduces (5.8) from (I). Thus, part (i) is proved.
k(s/2)
τ
] =
+ λ1)[F λ +
] (cid:117) ker(cid:0)(k(s/2))∗ + λ1(cid:1) and (5.1) imply (5.10), whereas (k(s/2)
Parts (ii) and (iii) are proved as in Theorem 4.1:
in particular, D[k(s/2)
D[k(s/2)
κλGs,λ] = ((−∆)s/2 + λ1)[F λ] + τκλ2(cid:107)Gs,λ(cid:107)2
L2(R) and (5.1) imply (5.11).
F
τ
τ
Kreın's resolvent formula for deficiency index 1 and (5.2) prescribe
+ λ1)−1 = ((−∆)s/2 + λ1)−1 + βλ,τ Gs,λ(cid:105)(cid:104)Gs,λ
(k(s/2)
τ
for some scalar βλ,τ to be determined, whenever (k(s/2)
τ (cid:54)= 0. Thus, for a generic h ∈ L2(R), the element g := (k(s/2)
τ
reads, in view of (4.2) and of the resolvent formula above,
τ
+λ1) is invertible, hence for
+ λ1)−1h ∈ D(k(s/2)
)
τ
(cid:98)g(p) = (cid:99)F λ(p) +
ξλ
ps + λ
,
(cid:99)F λ(p) :=
(cid:98)h(p)
ps + λ
,
ξλ :=
βλ,τ
2π
qs + λ
R
dq .
(cid:90)
(cid:98)h(q)
The boundary condition (I) for F λ and ξλ then implies 1 = βλ,τ
ω(s)λ2− 1
determines βλ,τ and proves (5.12), thus completing also the proof of (iv).
s
τ
, which
Second case: s ∈ (1, 3
2 ).
Let F λ := f λ +τ ξ hs,λ. (This is for consistency with the first case, but such F λ is
not going to be the same as the F λ of the thesis: functions will be renamed later.)
When f λ and ξ run over their possible domains, F λ spans the whole Friedrichs
domain (5.3). In particular F λ(0) = 0, which shows that the boundary condition
in g between F λ and g − F λ cannot have the form (II) any longer.
Owing to (5.3), we can re-write
(III)
D(k(s/2)
F
) (cid:51) F λ = f λ + τ ξ hs,λ = φλ
τ,ξ − φλ
τ,ξ(0)
Gs,λ(0)
Gs,λ ,
as follows from (3.11) and (5.6), then (V) becomes
λ s
s − 1
Gs,λ(0)
(cid:16) ω(s) λ2− 1
s
τ
1
= ω(s) λ2− 1
s ,
(cid:17)
φλ
τ,ξ(0) Gs,λ .
−
1
Gs,λ(0)
g = φλ
τ,ξ +
(VI)
Moreover,
(VII)
18
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
τ,ξ running over the whole H s(R) when F λ runs over the whole D(k(s/2)
φλ
(5.5) this is the same as
F
). Using
s − 1
λ s
1
ps + λ
= (cid:100)φλ
τ,ξ − φλ
√
τ,ξ(0)
2π Gs,λ(0)
1
ps + λ
,
(cid:99)f λ +
whence the identification
τ ξ
(ps + λ)2 − τ ξ
(cid:100)φλ
τ,ξ = (cid:99)f λ +
τ ξ
(ps + λ)2 ,
√
φλ
τ,ξ(0)
2π Gs,λ(0)
= τ ξ
s − 1
λ s
.
(In fact, a straightforward computations confirms that assuming the first of (IV),
the second follows.)
From (I) (with η = τ ξ) we then see that a generic g ∈ D(k(s/2)
) has the form
(cid:98)g = (cid:99)F λ +
= (cid:100)φλ
τ,ξ +
√
2π ξ(cid:98)Gs,λ = (cid:100)φλ
(cid:17) φλ
(cid:16) 1
λ s
s − 1
τ
− 1
τ,ξ +
τ,ξ(0)
Gs,λ(0)
(cid:17)(cid:98)Gs,λ
(cid:16)√
2π ξ − φλ
(cid:98)Gs,λ ,
τ
τ,ξ(0)
Gs,λ(0)
where we used (III) in the second step and the second identity of (IV) in the third
step. Since
(IV)
(V)
(k(s/2)
τ
+ λ1)g = (k(s/2)
F
+ λ1)F λ = ((−∆)
s
2 + λ)φλ
τ,ξ ,
τ
the whole D(k(s/2)
Part (i) is proved.
Renaming φλ
having used the second line of (I) in the first identity and (5.3) in the second identity.
τ,ξ into F λ, now F λ runs over the whole H s(R) when g runs over
): this fact and (VI) then yield (5.7), whereas (VII) yields (5.8).
The proof of parts (ii) and (iii) follows the same scheme as in the case s ∈ ( 1
] (cid:117) ker(cid:0)(k(s/2))∗ + λ1(cid:1) and (5.1) imply (5.10), whereas
] = D[k(s/2)
+λ1)[F λ +κλGs,λ] = ((−∆)s/2 +λ1)[F λ]+τκλ2(cid:107)Gs,λ(cid:107)2
L2(R) and (5.1) imply
Concerning part (iv), Kreın's resolvent formula and (5.4) prescribe for all τ (cid:54)= 0
thus, D[k(s/2)
(k(s/2)
(5.11).
2 , 1):
F
τ
τ
(VIII)
(k(s/2)
τ
+ λ1)−1 = (k(s/2)
= ((−∆)
F
+ λ1)−1 + βλ,τ Gs,λ(cid:105)(cid:104)Gs,λ
2 + λ1)−1 +
1
βλ,τ −
s
for some scalar βλ,τ to be determined. Owing to (VIII) and to (5.7), in order for
g := (k(s/2)
τ
+ λ1)−1h = ((−∆)
s
2 + λ1)−1h +
βλ,τ −
1
τ
to belong to D(k(s/2)
((−∆) s
λ1)F λ(cid:105) = F λ(0), one must necessarily have βλ,τ = ω(s)λ2− 1
(5.12).
) for a generic h ∈ L2(R), keeping into account that F λ :=
2 +λ1)−1h is a generic function in H s(R) and that (cid:104)Gs,λ, h(cid:105) = (cid:104)Gs,λ, ((−∆) s
2 +
s /τ . Then (VIII) yields
(cid:3)
Analogously to the change of parametrisation from Theorem 4.1 to Theorem 4.2,
we deduce from Theorem 5.2 the following version.
(cid:16)
(cid:16)
(cid:17)Gs,λ(cid:105)(cid:104)Gs,λ
(cid:17)(cid:104)Gs,λ, h(cid:105) Gs,λ
Gs,λ(0)
Gs,λ(0)
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 19
Theorem 5.3. Let s ∈ ( 1
2 , 1) ∪ (1, 3
Θ(s, λ) := (cid:0)λ1− 1
,
λ > 0 .
(i) The self-adjoint extensions in L2(R) of the operator k(s/2) form the family
(5.13)
(k(s/2)
α
(5.14)
2 ) and
s s sin ( π
s )(cid:1)−1
g = F λ +
)α∈R∪{∞}, where for arbitrary λ > 0
D(k(s/2)
) =
α
F λ ∈ H s(R)
+ λ) g = ((−∆)s/2 + λ) F λ .
(k(s/2)
α
F λ(0)
α − Θ(s, λ)
Gs,λ
The Friedrichs extension k(s/2)
, already qualified in Proposition 5.1, cor-
responds to α = ∞ when s ∈ ( 1
2 ). For
generic s, the extension with α = ∞ is the ordinary self-adjoint fractional
Laplacian (−∆)s/2 on L2(R).
2 , 1) and to α = 0 when s ∈ (1, 3
F
(ii) For each α ∈ R the quadratic form of the extension k(s/2)
if s ∈ ( 1
2 , 1)
if s ∈ (1, 3
2 ) and α (cid:54)= 0
L2(R)
H s/2(R) (cid:117) span{Gs,λ}
H s/2
[F λ + κλGs,λ] = (cid:107)∇ s
L2(R) − λ(cid:107)F λ + κλGs,λ(cid:107)2
(R\{0}) (cid:117) span{Gs,λ}
D[k(s/2)
is given by
] =
α
α
0
(cid:40)
2 F λ(cid:107)2
+ λ(cid:107)F λ(cid:107)2
(cid:16)
L2(R) +
θs
Θ(s, λ)
+
1
α − Θ(s, λ)
(cid:17)−1κλ2
(5.15)
k(s/2)
α
(5.16)
for arbitrary λ > 0.
(iii) The resolvent of k(s/2)
α
is given by
(5.17)
(k(s/2)
α
+ λ1)−1 = ((−∆)s/2 + λ1)−1
+(cid:0)α − Θ(s, λ)(cid:1)−1 Gs,λ(cid:105)(cid:104)Gs,λ
for arbitrary λ > 0.
(iv) For each α ∈ R the extension k(s/2)
α
is semi-bounded from below, and
(5.18)
σess(k(s/2)
α
) = σac(k(s/2)
σsc(k(s/2)
α
) = ∅ ,
(5.19)
σdisc(k(s/2)
α
) =
where the eigenvalue −E(s)
α is non-degenerate and is given by
α
) = [0, +∞) ,
(cid:40) ∅
α = (cid:0)αs sin( π
s )(cid:1) s
α }
{−E(s)
1−s
E(s)
if (s − 1) α (cid:54) 0
if (s − 1) α > 0
(5.20)
the (non-normalised) eigenfunction being Gs,λ=E(s)
α .
Proof. For any two pairs (λ, τ ) and (λ(cid:48), τ(cid:48)) identifying the same self-adjoint reali-
sation k(s/2)
) decomposes as
τ
τ
, a function g ∈ D(k(s/2)
g = F λ + A(λ, τ )F λ(0)Gs,λ = F λ(cid:48)
− θs
ω(s)λ2− 1
A(λ, τ ) :=
s
τ
,
Gs,λ(0)
+ A(λ(cid:48), τ(cid:48))F λ(cid:48)
(0)Gs,λ(cid:48)
and the uniqueness of the decomposition implies
(I)
F λ(cid:48)
= F λ + A(λ, τ )F λ(0)Gs,λ − A(λ(cid:48), τ(cid:48))F λ(cid:48)
(0)Gs,λ(cid:48) .
20
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
In order for F λ(cid:48)
Gs,λ(cid:48) must cancel out, that is,
to belong to H s(R), the non-H s singularities at x = 0 of Gs,λ and
(II)
A(λ, τ )F λ(0) = A(λ(cid:48), τ(cid:48))F λ(cid:48)
(0) .
Plugging (II) into (I) and evaluating of (I) at x = 0 yields
(III)
A(λ, τ )F λ(0)
(0) = F λ(0) + A(λ, τ )F λ(0)
= F λ(cid:48)
F λ(cid:48)
(0)
(cid:90)
1
2π
(cid:16)
(cid:17)
dp .
1
ps + λ
−
1
ps + λ(cid:48)
A straightforward analysis of the integral above (exploiting the compensation of
singularities when s ∈ ( 1
2 , 1)) shows that
1
(cid:16)
(cid:17)
1−s
1
1
s (cid:1)
s − λ(cid:48) 1−s
−
ps + λ
R
ps + λ(cid:48)
dp =
(cid:90)
1
2π
(IV)
R
(cid:0)λ
s sin π
s
= Θ(s, λ) − Θ(s, λ(cid:48)) .
Combining (III) and (IV) together implies that (λ, τ ) and (λ(cid:48), τ(cid:48)) are linked by the
relation
i.e.,
(V)
α − Θ(s, λ) =
1
A(λ, τ )
+ Θ(s, λ) =
+ Θ(s, λ(cid:48)) =: α ,
1
A(λ(cid:48), τ(cid:48))
(cid:16) ω(s)λ2− 1
s
− θs
Θ(s, λ)
τ
(cid:17)−1
,
which gives the natural extension parameter α. It is immediate from (V) that the
Friedrichs case τ → +∞ corresponds to α → +∞ when s ∈ ( 1
2 , 1) and to α = 0
when s ∈ (1, 3
2 ).
Upon replacing (V) in the formulas of Theorem 4.1 we deduce parts (i), (ii),
and (iii). Moreover, arguing as in the analogous point of the proof of Theorem 4.2,
formulas (5.18) follow, and one also concludes that each k(s/2)
may have at most
one negative non-degenerate eigenvalue Eα = −λ, λ > 0.
α
The occurrence of Eα is read out from the resolvent formula (5.17) as the pole
+ λ1)−1, that is, imposing α − Θ(s, λ) = 0 and hence
of (k(s/2)
α
(VI)
When s < 1, (VI) can be only satisfied by some λ > 0 when α < 0, in which case
s )(cid:1)−1
.
s−1
s s sin( π
α =(cid:0)λ
λ =(cid:0)αs sin( π
s )(cid:1) s
λ =(cid:0)αs sin( π
s )(cid:1) s
1−s
1−s
(s < 1 , α < 0) .
(s > 1 , α > 0) .
When instead s > 1, a solution λ > 0 to (VI) exists only when α > 0, and is given
by
(cid:3)
Hence we proved also (5.19), which completes the proof of part (iv).
In the regime s ∈ (1, 3
α
2 ) Theorem 5.3(ii) can be re-phrased in the following even
can be equivalently qualified as
more natural formulation, which shows that k(s/2)
a form perturbation of (−∆)s/2.
Proposition 5.4. Let s ∈ (1, 3
2 ). The self-adjoint extensions in L2(R) of k(s/2)
form the family (k(s/2)
)α∈R∪{∞}, where α = 0 labels the Friedrichs extension given
by (5.1), α = ∞ labels the ordinary self-adjoint fractional Laplacian (−∆)s/2, and
for α ∈ R \ {0} one has
D[k(s/2)
α
] = H s/2
[g] = (cid:107)∇ s
0
α
k(s/2)
α
(R\{0}) (cid:117) span{Gs,λ} = H s/2(R)
2 g (cid:107)2
g(0)2
L2(R) − 1
α
(5.21)
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 21
for every λ > 0.
Proof. In view of Theorem 5.3(ii), we only need to prove the second line of (5.21) for
Θ(s,λ) −
a generic g ∈ D[k(s/2)
α−Θ(s,λ) )−1 for short and decompose
]. We set Σ := (
g = F λ + κλGs,λ for some F λ ∈ H s/2
(R\{0}) and κλ = g(0)/Gs,λ(0). Applying
(5.16), we find
α
1
1
0
k(s/2)
α
L2(R) + (cid:107)∇ s
[g] = −λ(cid:107)g(cid:107)2
= (cid:107)∇ s
2 g(cid:107)2
− 2 Re κλ
L2(R) + κλ2(cid:0)(cid:107)∇ s
(cid:0)(cid:104)∇ s
2 (g − κλGs,λ)(cid:107)2
2 Gs,λ(cid:107)2
2 Gs,λ(cid:105) + λ(cid:104)g, Gs,λ(cid:105)(cid:1) .
L2(R) + λ(cid:107)g − κλGs,λ(cid:107)2
L2(R) + λ(cid:107)Gs,λ(cid:107)2
L2(R) + Σ(cid:1)
L2(R) + Σκλ2
2 g,∇ s
L2(R) + λ(cid:107)Gs,λ(cid:107)2
2 g,∇ s
2 Gs,λ(cid:107)2
Since (cid:107)∇ s
Θ(s, λ) and analogously (cid:104)∇ s
2 g(cid:107)2
[g] = (cid:107)∇ s
k(s/2)
α
L2(R) +
2 + λ)Gs,λ(cid:105) = Gs,λ(0) =
L2(R) = (cid:104)Gs,λ, ((−∆) s
2 Gs,λ(cid:105) + λ(cid:104)g, Gs,λ(cid:105) = g(0), then
g(0)2
Θ(s, λ)2 (Θ(s, λ) + Σ) − 2
g(0)2
Θ(s, λ)
(cid:16)
The coefficient of the g(0)2-term above amounts to
1
Θ(s, λ)2
whence indeed k(s/2)
α
Θ(s, λ) +
[g] = (cid:107)∇ s
1
1
α−Θ(s,λ)
Θ(s,λ) −
2 g (cid:107)2
L2(R) − α−1g(0)2.
1
(cid:17) −
2
Θ(s, λ)
= − 1
α
,
.
(cid:3)
6. Rank-one singular perturbations of the fractional Laplacian:
inhomogeneous case
For completeness of presentation, in this Section we work out the inhomogeneous
version of the operator k(s/2)
of Section 4, for concreteness in dimension d = 3.
That is, instead of constructing a singular perturbation of (−∆)s/2, we consider
the singular perturbation of the fractional power (−∆ + 1)s/2. This is going to be
the operator d(s/2)
introduced informally in (1.3).
τ
τ
The conceptual scheme is the very same as in Sections 3 and 4, and only certain
explicit computations are modified in an easy way. Thus, we content ourselves to
state the main results without proofs.
For chosen λ > 0 and s ∈ R we set
x, p ∈ R3 ,
Gs,λ(x) :=
1
1
2
(x) ,
(2π) 3
(p2 + λ)s/2
(6.1)
whence (−∆ + λ1)s/2 Gs,λ = δ(x) distributionally. We also set
0 (R3\{0})
(6.2)
as an operator closure with respect to the Hilbert space L2(R3). Thus, in compar-
ison to Section 2, G2,λ = Gλ and d(1)
x→0−−−−→ Λs =
:= (−∆ + λ1)s/2 (cid:22) C∞
λ = h + λ1. Moreover,
x3−s Gs,λ(x)
s ∈ (0, 3)
d(s/2)
λ
,
(cid:17)∨
(cid:16)
(6.3)
Gs,λ(x)
x→0−−−−→ Gs,λ(0) =
,
s > 3 .
Γ( 3−s
2 )
2 2s− 3
(2π) 3
2 Γ( s
2 )
Γ( s−3
2 )
s−3
2 Γ( s
2 )
2 λ
8π 3
Reasoning as in (3.5) and in Lemma 3.1, we see that when s ∈ ( 3
index of d(s/2)
equals 1.
λ
2 , 5
2 ) the deficiency
One has the following construction.
Theorem 6.1. Let s ∈ ( 3
2 , 5
2 ) and λ > 0.
22
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
(i) The self-adjoint extensions in L2(R3) of the operator d(s/2)
form the family
(d(s/2)
λ,∞ is its Friedrichs extension, namely the self-
adjoint fractional shifted Laplacian (−∆ + λ1)s/2, and all other extensions
are given by
)τ∈R∪{∞}, where d(s/2)
λ,τ
λ
λ,τ
(cid:1) :=
D(cid:0)d(s/2)
g ∈ L2(R3)
(cid:110)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:98)g(p) =(cid:99)f λ(p) +
ξ ∈ C , f λ ∈ H s(R3) , (cid:82)
where (cid:99)F λ =(cid:99)f λ(p) + (p2 + λ)−sτ ξ, and
2 ) F λ(0)Gs,λ
g = F λ + 8π
2 λs− 3
τ Γ(s− 3
2 Γ(s)
τ ξ
=
3
(p2 + λ)s +
ξ
(p2 + λ)s/2
R3(cid:99)f λ(p) dp = 0
(cid:12)(cid:12)(cid:12) F λ ∈ H s(R3)
(cid:111)
,
(6.4)
(6.5)
λ,τ g := (−∆ + λ1)s/2F λ .
d(s/2)
(ii) Each extension is semi-bounded from below and
(6.6)
λ,τ
inf σ(d(s/2)
inf σ(d(s/2)
d(s/2)
λ,τ
) (cid:62) 0 ⇔ τ (cid:62) 0
) > 0 ⇔ τ > 0
is invertible ⇔ τ (cid:54)= 0 .
λ,τ
(iii) The eigenvalue zero of the extension d(s/2)
normalised) eigenfunction is Gs,λ. When τ < 0 the operator d(s/2)
one non-degenerate negative eigenvalue Eτ < τ .
λ,τ =0 is non-degenerate and the (non-
admits
(iv) For each τ ∈ R the quadratic form of the extension d(s/2)
is given by
λ,τ
λ,τ
(6.7)
(6.8)
D[d(s/2)
2 (R3) (cid:117) span{Gs,λ}
[F λ + κλGs,λ] = (cid:107)(−∆ + λ1)s/4F λ(cid:107)2
] = H
λ,τ
s
d(s/2)
λ,τ
for any F λ ∈ H s
2 (R3) and κλ ∈ C.
(v) For τ (cid:54)= 0, one has the resolvent identity
)−1 = (−∆ + λ)−s/2 + 8π
(d(s/2)
3
λ,τ
(6.9)
2 λs− 3
2 Γ(s)
τ Γ(s− 3
2 )
Gs,λ(cid:105)(cid:104)Gs,λ .
L2(R3) + τ Γ(s− 3
2 λs− 3
8π
2 )
2 Γ(s)
3
κλ2
or k(s/2)
It is clear that, as opposite to hs/2
τ
: in fact, the domain of d(s/2)
, the shift λ > 0 is inherent the very
is independent of
construction of the operator d(s/2)
λ > 0, but its action is not (the difference d(s/2)
trivial operator), thus also the adjoint (d(s/2)
λ-dependent (the adjoints (hs/2)∗ and (k(s/2))∗ are λ-independent, instead).
λ
is a bounded, yet non-
)∗ and its self-adjoint restrictions are
− d(s/2)
λ(cid:48)
λ,τ
λ
λ
τ
Let us also elaborate further on part (iii) of the Theorem. As in the previous
Sections, both statements are classical facts in the Kreın-Visik-Birman extension
scheme. Chosen λ > 0 and s ∈ ( 3
2 ), the negative eigenvalue Eτ of the extension
d(s/2)
for τ < 0 is obtained as follows. Let g be the corresponding eigenfunction
λ,τ
and decompose(cid:98)g =(cid:99)f λ + (p2 + λ)−sτ + (p2 + λ)−s/2 according to (6.4) (without loss
2 , 5
of generality we re-absorbed ξ in f λ). Then the condition d(s/2)
the property (6.5), reads
2 τ = Eτ(cid:99)f λ + (p2 + λ)−sτ Eτ + (p2 + λ)− s
2 Eτ ,
λ,τ g = Eτ g, owing to
(p2 + λ)
whence
s
2(cid:99)f λ + (p2 + λ)− s
(cid:98)f (p) =
(cid:17)
.
(cid:16)
1
(p2 + λ) s
2 − Eτ
τ Eτ
(p2 + λ)s − τ − Eτ
(p2 + λ) s
2
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 23
Figure 1. Plot of the negative eigenvalue Eτ of the extension
d(s/2)
vs τ for λ = 1 and s = 1.8 (blue curve). The reference
λ,τ
orange curve gives the corresponding value of τ . Indeed, Eτ < τ
and Eτ→0 = 0.
The fact that f λ ∈ H s(R3) is then obvious, whereas the condition (cid:82)
R3(cid:99)f λdp = 0
selects the value of Eτ in terms of τ (and of λ) -- a numerical example is provided
in Figure 1.
Let us conclude by remarking that two relevant features of the homogeneous case
are present in the inhomogeneous case too.
First, and most importantly, the operator d(s/2)
the resolvent sense, of (−∆ + λ)s/2, precisely as k(s/2)
of (−∆)s/2.
τ
λ,τ
is a rank-one perturbation, in
is a rank-one perturbation
Second, the elements of the domain of d(s/2)
decompose into a regular H s-part
and a singular part, constrained to the former by a local boundary condition, where
the local singularity when s ∈ ( 3
2 ) is of the form x3−s as x → 0.
2 , 5
λ,τ
7. High deficiency index (high fractional power) scenario
Let us outline in this Section how the previous constructions of the self-adjoint
extensions of the operators k(s/2) or d(s/2)
We recall from Lemma 3.1 that when s ∈ I (d)
λ
get modified when s > d
2 + 1.
(cid:110)
ker(cid:0)(k(s/2))∗ + λ1(cid:1) = span
(cid:110)
)∗ + λ1(cid:1) = span
ker(cid:0)(d(s/2)
λ
uλ
γ1,...,γd
vλ
γ1,...,γd
n ( d
2 + n − 1, d
(cid:12)(cid:12)(cid:12) γ1, . . . , γd ∈ N0 ,
(cid:12)(cid:12)(cid:12) γ1, . . . , γd ∈ N0 ,
2 + n), n ∈ N, one has
(cid:111)
d(cid:88)
γj (cid:54) n − 1
,
j=1
d(cid:88)
j=1
(cid:111)
γj (cid:54) n − 1
,
(7.1)
(7.2)
and analogously
having defined
(7.3)
(cid:99)vλ
γ1,...,γd (p) :=
1 ··· pγd
pγ1
d
(p2 + λ) s
2
.
operator T in some subspace D(T ) of ker(cid:0)(k(s/2))∗ + λ1(cid:1) ∼= CJ (s,d), hence labelled
The same extension scheme applied in Section 4 provides an analogous classifica-
tion of all the self-adjoint extensions in the case of generic deficiency index J (s, d),
where now each extension of k(s/2) is an operator k(s/2)
labelled by a self-adjoint
by some N × N hermitian matrix, 1 (cid:54) N (cid:54) J (s, d).
T
-10-9-8-7-6-5-4-3-2τ-5-10-15-20Eτ 24
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
Explicitly (see, e.g., [7, Theorem 3.4]),
g ∈ L2(R3)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) g = f + (k(s/2)
w ∈ ker(cid:0)(k(s/2))∗ + λ1(cid:1) ∩ D(T )⊥
+ λ1)−1(T u + w) + u
0 (R3\{0}) , u ∈ D(T ) ,
where f ∈ H s
F
D(k(s/2)
T
) =
(7.4)
(k(s/2)
T
+ λ1)g = (k(s/2)
+ λ1)Fλ
F
Fλ := f + (k(s/2)
F
+ λ1)−1(T u + w) ∈ H s(R3) ,
with k(s/2)
F
denoting the Friedrichs extension.
Analogously, the self-adjoint extensions of d(s/2)
form a family of operators d(s/2)
λ,T
λ
with
(7.5)
D(d(s/2)
λ,T ) =
g ∈ L2(R3)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where f ∈ H s
g = f + (d(s/2)
w ∈ ker(cid:0)d(s/2)
λ,F )−1(T v + w) + v
0 (R3\{0}) , v ∈ D(T ) ,
)∗ ∩ D(T )⊥
λ
d(s/2)
λ,T g = d(s/2)
λ,F Fλ
Fλ := f + (d(s/2)
λ,F )−1(T v + w) ∈ H s(R3) .
The theory provides also a counterpart classification of the the quadratic forms
of the extensions (see, e.g., [7, Theorem 3.6]).
The above formulas show that for high powers s the operators k(s/2) and d(s/2)
have a richer variety (a J (s, d)2-parameter family) of self-adjoint extensions. The
parametrising matrix T determines a more complicated set of 'boundary condition'
between the 'Friedrichs' part of a generic element of the extension domain, and the
remaining part: the resulting constraint involves the evaluation at x = 0 of some
number of partial derivatives of the former component.
λ
This construction produces finite-rank perturbations in the resolvent sense, hence
extensions that are all semi-bounded from below and may admit a (finite) number
of negative eigenvalues, up to J (s, d), counting the multiplicity.
Unlike the case of deficiency index 1, depending on the extension parameter T
the large-p vanishing behaviour in momentum space of the singular component may
be milder than that of the Green function, and therefore the local singularity of g
in position space may be more severe than the behaviour of the Green function as
x → 0.
function g ∈ D(k(s/2)
identical.
Let us comment on how the worst leading singularity at x = 0 of a generic
λ,T ) is
) depends on s and d -- the discussion for g ∈ D(d(s/2)
T
As expressed by (7.4), such a singularity is due to those functions of type uλ
γ1,...,γd
that span D(T ). When s ∈ I (d)
the worst local singularity occurs when such
functions decrease at infinity in momentum coordinates with the slowest possible
vanishing rate compatible with s and d, that is, when γ1 + ··· + γd = n − 1.
p−(s+1−n) as p → +∞. Then u(x) ≈ x−(d−1+n−s) as x → 0. Since the
map
Let u be any such most singular function, which then behaves as (cid:98)u(p) ≈
n
n (cid:51) s (cid:55)→ d − 1 + n − s
I (d)
2 − 1, d
is monotone decreasing and takes values in ( d
such that D(T ) (cid:51) u, then the functions in D(k(s/2)
ranges from x− d
2 to x− d−1
s increases in I (3)
1 .
T
2 as long as s increases in I (d)
2 ), if the extension k(s/2)
is
) display a local singularity that
n , precisely as (4.14) when
T
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 25
Noticeably, at the transition values s ∈ N + 1
2 the above picture undergoes a
discontinuity in s, due to the further control of one more derivative in D(k(s/2)), as a
consequence of Sobolev's Lemma, and consequently to emergence in ker(cid:0)(k(s/2))∗ +
λ1(cid:1) of elements that in momentum coordinates vanish more slowly at infinity.
8. Applications and perspectives
Besides the operator-theoretic and functional-analytic interest per se of the con-
structions of the operators (1.3), our discussion is profoundly inspired to an amount
of natural applications.
Singular perturbations model point-like impurities and more generally point-
like interactions. In the realm of the evolutive equations of relevance for quantum
mechanics, they naturally govern the evolution of systems subject to such 'singular
potentials'.
This includes the linear Schrodinger evolution
i∂tu = hαu
as well as the class of semi-linear Schrodinger equations
i∂tu = hαu + N (u)
of the free Laplacian plus a point-like perturbation, with physically relevant non-
linearities such as the power-law local non-linearity N (u) = uγ−1u or the Hartree
type non-local non-linearity N (u) = (w ∗ u2)u. The reconstruction of the linear
propagator from the resolvent of hα is already known in the literature [19, 1], as well
as the dispersive properties and space-time estimates of such a propagator [5, 9],
and the existence, completeness, and Lp-boundedness of the wave operators for the
pairs (hα,−∆) [6].
In addition, for the study of the non-linear problem in a suitable space (the
energy space in the first place, as well as other spaces of lower or higher regularity),
the knowledge is needed of the corresponding singular norms, namely the norms
α(R3) = (cid:107)(hα + 1)s/2u(cid:107)L2(R3) considered in Theorem 2.4 above and [8]. In this
(cid:107)u(cid:107)(cid:101)H s
respect, and with such tools, the study of certain non-linear Schrodinger equations
with singular potentials has already started [12].
An analogous systematic knowledge for k(s/2)
This is even more needed due to the relevance of various models of singular per-
turbations of fractional differential operators. A relevant example are the powers of
the quantum-mechanical semi-relativistic energy operator
perturbation of which yields precisely operators of the type d(s/2)
Section 6 or, in the case of zero rest energy, of the type k(s/2)
√−∆ + m2, the singular
m2,τ considered in
as in Section 4.
is by know lacking.
τ
τ
and d(s/2)
λ,τ
What one finds in the literature is an increasing amount of recent studies [14, 16,
4, 11, 18, 20, 10, 15] where the singular perturbation of the fractional Laplacian is
approached through Green's function methods (together with Wick-like rotations
to obtain the propagator from the resolvent) that have the virtue of highlighting the
singular structure carried over by what we denoted with Gs,λ and Gs,λ, however with
no specific concern to the multiplicity of self-adjoint realisations and the associated
local boundary conditions, or to the increase of the deficiency index with the power
s.
As above, for each extension one would like to qualify the linear propagator, its
space-time estimates, the fractional norms, and to use these tools for the associated
non-linear problems.
We trust that the research programme emerging from the above considerations
may be successfully addressed over the next future!
26
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
Appendix A. Characterisation of H s
0 (Rd\{0})
We show in this Appendix how to prove the characterisation (3.5) of the space
0 (Rd \ {0}).
H s
It is not restrictive to fix d = 3 and to discuss and compare the first two regimes
s ∈ I (3)
n , n = 2, 3, . . . , is completely
analogous.
1 . The argument for s ∈ I (3)
and s ∈ I (3)
0
Thus, let us prove the following property.
Lemma A.1.
0 (R3\{0}) =
H s
(cid:40)
H s(R3)
(cid:8)f ∈ H s(R3)(cid:12)(cid:12)(cid:82)
R3 (cid:98)f (p) dp = 0(cid:9) if s ∈ ( 3
if s ∈ [0, 3
2 )
2 , 5
2 ) .
Proof. We consider first the case s ∈ [0, 3
2 ). The inclusion
0 (R3\{0}) ⊂ H s(R3)
H s
is obvious. For the other inclusion, for any f ∈ H s(R3) and for arbitrary ε > 0
we want to find fε ∈ C∞
0 (R3\{0}) such that (cid:107)fε − f(cid:107)H s (cid:54) ε, and by means of a
standard density argument, it is not restrictive to assume further that f ∈ S(R3)
and (cid:98)f is compactly supported. Let φ ∈ C∞(R3) be such that
φ(x) = 0 for x (cid:54) 1 ,
φ(x) = 1
for x (cid:62) 2 ,
and let ψ := φ − 1, φn(x) := φ(nx), and ψn(x) := ψ(nx), for n ∈ N. Thus,
0 (R3) and φnf ∈ C∞(R3), with φnf vanishing in a neighbourhood of x = 0.
ψ ∈ C∞
2(cid:107)(cid:104)p(cid:105)s((cid:98)ψn ∗ (cid:98)f )(cid:107)L2 ,
Moreover,
where Ds := (1 − ∆)s/2 and (cid:104)p(cid:105) =(cid:112)1 + p2. Clearly, (cid:98)ψn(p) = 1
(cid:107)φnf − f(cid:107)H s = (cid:107)Ds(φnf − f )(cid:107)L2 = (cid:107)Ds(ψnf )(cid:107)L2 = (2π)
n3(cid:98)ψ( p
(cid:12)(cid:12)(cid:12)2
dq (cid:98)f (q)(cid:98)ψn(p − q)
(cid:1)(cid:12)(cid:12)2
dq (cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:98)ψ(cid:0) p−q
(cid:90)
(cid:54) (2π)3(cid:107)(cid:98)f(cid:107)2
dp(cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:12)(cid:90)
supp (cid:98)f
(cid:90)
(cid:90)
(cid:90)
supp (cid:98)f
dp(cid:104)p(cid:105)2s(cid:98)ψ(p − q)2
(cid:107)φnf − f(cid:107)2
n ). Therefore,
H s = (2π)3
(cid:46) n2s
(cid:90)
dp
R3
L2
n
3
The last step above follows from the continuity of q (cid:55)→(cid:82)
(cid:46) n2s−3
R3 dp(cid:104)p(cid:105)2s(cid:98)ψ(p − q)2. In
particular, we can choose n := n(ε) sufficiently large such that
R3
n6
supp (cid:98)f
R3
n→+∞
−−−−−−→ 0 .
dq
n
(cid:107)φnf − f(cid:107)H s (cid:54) ε
2
.
For such n, we can find a smooth function x (cid:55)→ χ(x) that produces a slow cut-off
at infinity so that
(cid:107)χφnf − φnf(cid:107)H s (cid:54) ε
.
2
0 (R\{0}) satisfying
We have thus identified a function fε := χφnf ∈ C∞
(cid:107)fε − f(cid:107)H s (cid:54) (cid:107)χφnf − φnf(cid:107)H s + (cid:107)φnf − f(cid:107)H s (cid:54) ε .
Let us discuss now the case s ∈ ( 3
2 , 5
continuous embedding H s(R3) (cid:44)→ C 0,s− 3
2 ). Owing to Sobolev's Lemma, one has the
2 (R3) and hence any limit in the H s-norm
FRACTIONAL POWERS, SINGULAR PERTURBATIONS, QUANTUM HAMILTONIANS 27
of elements in C∞
prove the inclusion
0 (R3\{0}) must vanish at the origin. Therefore, we only need to
,
(cid:12)(cid:12)(cid:12)(cid:90) (cid:98)f (p) dp = 0
(cid:111)
(cid:17) 1{p(cid:54)1}
(cid:12)(cid:12){p (cid:54) 1}(cid:12)(cid:12) ,
H s
0 (R3\{0}) ⊃ (cid:110)
(cid:99)fR := (cid:98)f · 1{p(cid:54)R} −(cid:16)(cid:90)
f ∈ H s(R3)
p(cid:54)R
R > 0 ,
H s = (2π)3
part of the proof,
(cid:107)φnf − f(cid:107)2
R → +∞, it is not restrictive to assume from the beginning that f ∈ S(R3) with
that is, for any f ∈ H s(R3) with f (0) = 0 and for arbitrary ε > 0 we want to find
fε ∈ C∞
0 (R3\{0}) such that (cid:107)fε − f(cid:107)H s (cid:54) ε. Since the function fR defined by
(cid:98)f (p)dp
has the obvious properties fR ∈ S(R3),(cid:82)
R3(cid:99)fR(p) dp = 0, and (cid:107)f − fR(cid:107)H s → 0 as
f (0) = 0 and with compactly supported (cid:98)f . With the same notation as in the first
dp(cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:12)(cid:90)
(cid:12)(cid:12)(cid:12)2
dq (cid:98)f (q)(cid:98)ψn(p − q)
dp(cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:12)(cid:90)
dq (cid:98)f (q)(cid:0)(cid:98)ψn(p − q) − (cid:98)ψn(p)(cid:1)(cid:12)(cid:12)(cid:12)2
supp (cid:98)f
supp (cid:98)f
(cid:90)
(cid:90)
(cid:1)(cid:12)(cid:12)2
dq (cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:98)ψ(cid:0) p−q
(cid:1) − (cid:98)ψ(cid:0) p
(cid:90)
dq (cid:104)p(cid:105)2s(cid:12)(cid:12)(cid:12)(cid:90) 1
supp (cid:98)f
(cid:12)(cid:12)(cid:12)2
dt (∇(cid:98)ψ)(p − tq) · q
(cid:90) 1
dt(∇(cid:98)ψ)(p − tq)2
(cid:90)
(cid:90)
(cid:54) (2π)3(cid:107)(cid:98)f(cid:107)2
(cid:90)
where we used the condition (cid:82)
n−1 for q ∈ supp (cid:98)f
q (cid:55)→(cid:82)
R3 (cid:98)f (q) dq = 0 in the second step, the bound q (cid:46)
0 dt(cid:104)p(cid:105)2s(∇(cid:98))ψ(p − tq)2 in the last step. In particular, we can choose
in the penultimate step, and the continuity of the function
supp (cid:98)f
R3
n→+∞
−−−−−−→ 0 ,
R3
(cid:46) n2s−2
R3 dp(cid:82) 1
(cid:46) n2s−5
dp(cid:104)p(cid:105)2s
(cid:46) n2s
supp (cid:98)f
n
= (2π)3
R3
R3
dp
R3
n
n
n6
dp
0
0
dq
n
L2(R3)
(cid:90)
(cid:90)
n
n := n(ε) sufficiently large such that
(cid:107)φnf − f(cid:107)H s (cid:54) ε
2
.
For such n, we can find a smooth function x (cid:55)→ χ(x) that produces a slow cut-off
at infinity so that
(cid:107)χφnf − φnf(cid:107)H s (cid:54) ε
.
2
0 (R\{0}) satisfying
We have thus identified a function fε := χφnf ∈ C∞
(cid:107)fε − f(cid:107)H s (cid:54) (cid:107)χφnf − φnf(cid:107)H s + (cid:107)φnf − f(cid:107)H s (cid:54) ε,
(cid:3)
which concludes the proof.
2 , n + 3
When n = 2, 3, . . . and s ∈ In = (n + 1
2 ), Sobolev's Lemma guarantees
0 (R3\{0}) comes with the vanishing at x = 0
that the closure in the H s-norm of C∞
of the the function and its first partial derivatives up to order n. Then one can
0 (R3 \{0}) by repeating an analogous argument
complete the characterisation of H s
as above, now replacing f with its partial derivatives. This yields the formula
3 (cid:98)f (p) dp = 0
f ∈ H s(R3) such that
R3 pγ1
0 (R3\{0}) =
H s
s ∈ (n + 1
2 , n + 3
1 pγ2
2 pγ3
(cid:82)
2 ) .
γ1, γ2, γ3 ∈ N0 , γ1 + γ2 + γ3 (cid:54) n − 1
,
28
A. MICHELANGELI, A. OTTOLINI, AND R. SCANDONE
References
[1] S. Albeverio, Z. Brze´zniak, and L. Dabrowski, Fundamental solution of the heat and
Schrodinger equations with point interaction, J. Funct. Anal., 130 (1995), pp. 220 -- 254.
[2] S. Albeverio, F. Gesztesy, R. Høegh-Krohn, and H. Holden, Solvable Models in Quan-
tum Mechanics, Texts and Monographs in Physics, Springer-Verlag, New York, 1988.
[3] S. Albeverio and P. Kurasov, Singular perturbations of differential operators, vol. 271 of
London Mathematical Society Lecture Note Series, Cambridge University Press, Cambridge,
2000. Solvable Schrodinger type operators.
[4] E. Capelas de Oliveira and J. J. Vaz, Tunneling in fractional quantum mechanics, J.
Phys. A, 44 (2011), pp. 185303, 17.
[5] P. D'Ancona, V. Pierfelice, and A. Teta, Dispersive estimate for the Schrodinger equa-
tion with point interactions, Math. Methods Appl. Sci., 29 (2006), pp. 309 -- 323.
[6] G. Dell'Antonio, A. Michelangeli, R. Scandone, and K. Yajima, Lp-Boundedness of
Wave Operators for the Three-Dimensional Multi-Centre Point Interaction, Ann. Henri
Poincar´e, 19 (2018), pp. 283 -- 322.
[7] M. Gallone, A. Michelangeli, and A. Ottolini, Kreın-Visik-Birman self-adjoint exten-
sion theory revisited, SISSA preprint 25/2017/MATE (2017).
[8] V. Georgiev, A. Michelangeli, and R. Scandone, On fractional powers of singular per-
turbations of the Laplacian, Journal of Functional Analysis, 275 (2018) 1551-1602.
[9] F. Iandoli and R. Scandone, Dispersive estimates for Schrodinger operators with point in-
teractions in R3, in Advances in Quantum Mechanics: Contemporary Trends and Open Prob-
lems, A. Michelangeli and G. Dell'Antonio, eds., Springer INdAM Series, vol. 18, Springer
International Publishing, pp. 187 -- 199.
[10] S. Jarosz and J. J. Vaz, Fractional Schrodinger equation with Riesz-Feller derivative for
delta potentials, J. Math. Phys., 57 (2016), pp. 123506, 16.
[11] E. K. Lenzi, H. V. Ribeiro, M. A. F. dos Santos, R. Rossato, and R. S. Mendes, Time
dependent solutions for a fractional Schrodinger equation with delta potentials, J. Math.
Phys., 54 (2013), pp. 082107, 8.
[12] A. Michelangeli, A. Olgiati, and R. Scandone, The singular Hartree equation in frac-
tional perturbed Sobolev spaces, Journal Nonlin. Math. Phys. 25, (2018) pp. 1-32, 4
[13] A. Michelangeli and A. Ottolini, On point interactions realised as Ter-Martirosyan-
Skornyakov Hamiltonians, Rep. Math. Phys., 79 (2017), pp. 215 -- 260.
[14] S. I. Muslih, Solutions of a particle with fractional δ-potential in a fractional dimensional
space, Internat. J. Theoret. Phys., 49 (2010), pp. 2095 -- 2104.
[15] M. M. Nayga and J. P. Esguerra, Green's functions and energy eigenvalues for delta-
perturbed space-fractional quantum systems, J. Math. Phys., 57 (2016), pp. 022103, 7.
[16] E. C. d. Oliveira, F. S. Costa, and J. J. Vaz, The fractional Schrodinger equation for
delta potentials, J. Math. Phys., 51 (2010), pp. 123517, 16.
[17] A. Sacchetti, Stationary solutions of a fractional Laplacian with singular perturbation,
arXiv:1801.01694 (2018).
[18] T. Sandev, I. Petreska, and E. K. Lenzi, Time-dependent Schrodinger-like equation with
nonlocal term, J. Math. Phys., 55 (2014), pp. 092105, 10.
[19] S. Scarlatti and A. Teta, Derivation of the time-dependent propagator for the three-
interaction, J. Phys. A, 23 (1990),
dimensional Schrodinger equation with one-point
pp. L1033 -- L1035.
[20] J. D. Tare and J. P. H. Esguerra, Bound states for multiple Dirac-δ wells in space-
fractional quantum mechanics, J. Math. Phys., 55 (2014), pp. 012106, 10.
(A. Michelangeli) International School for Advanced Studies -- SISSA, via Bonomea
265, 34136 Trieste (Italy).
E-mail address: [email protected]
(A. Ottolini) Department of Mathematics, Stanford University, 450 Serra Mall, Stan-
ford CA 94305 (USA).
E-mail address: [email protected]
(R. Scandone) International School for Advanced Studies -- SISSA, via Bonomea 265,
34136 Trieste (Italy).
E-mail address: [email protected]
|
1904.03592 | 2 | 1904 | 2019-04-09T08:53:07 | Tracial moment problems on hypercubes | [
"math.FA",
"math.OA"
] | In this paper we introduce the "tracial $K$-moment problem" and the "sequential matrix-valued $K$-moment problem" and show the equivalence of the solvability of these problems. Using a Haviland's theorem for matrix polynomials, we solve these $K$-moment problems for the case where $K$ is the hypercube $[-1,1]^n$. | math.FA | math |
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
CONG TRINH LE
Abstract. In this paper we introduce the tracial K-moment prob-
lem and the sequential matrix-valued K-moment problem and show the
equivalence of the solvability of these problems. Using a Haviland's the-
orem for matrix polynomials, we solve these K-moment problems for
the case where K is the hypercube [−1, 1]n.
1. Introduction
Let R[X] := R[X1, . . . , Xn] denote the algebra of polynomials in n vari-
ables X1, . . . , Xn with real coefficients.
For a real-valued linear functional L on R[X] and a closed subset K ⊆ Rn,
the K-moment problem asks when does there exist a positive Borel measure
σ on Rn supported on K such that
L(f ) = Z f (x)dσ(x), ∀f ∈ R[X]?
Each functional of this form is called a K-moment functional.
Another kind of K-moment problems for sequences of numbers is stated
as follows. For a multi-index sequence s = (sα)α∈Nn
0 , the K-moment problem
for sequences asks when does there exist a positive Borel measure σ on Rn
supported on K such that
sα = Z xαdσ(x), ∀α ∈ Nn
0 ,
where xα = xα1
sequence.
1 . . . xαn
n . Each sequence of this form is called a K-moment
For each sequence s = (sα)α∈Nn
0 , the Riesz functional Ls on R[X] associ-
ated to s is defined by
Ls(xα) = sα, ∀α ∈ Nn
0 .
It is easy to check that s is a K-moment sequence if and only if Ls is a
K-moment functional.
In the case K = R (resp. [0, +∞), [0, 1]), we go back to the Hamburger
(resp. Stieltjes, Hausdorff ) moment problem which was solved completely
by many authors (see, for example, a very excellent book of Schmudgen [10]
Date: April 10, 2019.
2010 Mathematics Subject Classification. 44A60; 30E05; 13J30; 47A57; 11E25.
Key words and phrases. Moment problem; matrix polynomial; moment functional; mo-
ment sequence; hypercube.
1
2
CONG TRINH LE
on the moment problem). If K is a compact basic closed semi-algebraic set
in Rn, the K-moment problem was solved by Schmudgen (1991, [9]).
One of the power tools to solve K-moment problems is the Haviland
theorem, which was proved by E.K. Haviland (1935) and stated as follows.
Given a linear functional L on R[X] and a closed subset K in Rn. Then
L is a K-moment functional if and only if L(f ) ≥ 0 for all f ≥ 0 on K.
Here the notation f ≥ 0 on K means that f (x) ≥ 0, ∀x ∈ K. A similar
meaning is defined for the notation f > 0 on K.
For a fix integer t > 0, we denote by Mt(R[X]) the algebra of t×t matrices
with entries in R[X], and by St(R[X]) the subalgebra of symmetric matrices.
Each element A ∈ Mt(R[X]) is a matrix whose entries are polynomials
in R[X], which is called a polynomial matrix. A is also called a matrix
polynomial, because it can be viewed as a polynomial in X1, . . . , Xn whose
coefficients come from Mt(R). Namely, we can write A as
A(X) =
d
Xα=0
AαX α,
where α = (α1, · · · , αn) ∈ Nn
. . . X αn
n ,
Aα ∈ Mt(R). To unify notation, throughout the paper each element of
Mt(R[X]) is called a matrix polynomial.
0 , α := α1 + . . . + αn, X α := X α1
1
Let A(X) =
d
Xα=0
AαX α be a matrix polynomial. Denote
Ad(X) := Xα=d
AαX α.
If Ad 6≡ 0, the matrix polynomial A is called of degree d. In particular, when
d
n = 1 and A(x) =
Aixi is a univariate matrix polynomial of degree d,
Xi=0
the non-zero matrix Ad is called the leading coefficient of A.
Recently, there are some researchs on (truncated) noncommutative K-
moment problems, e.g. Dette and Studden (2002, [4]), Burgdorf and Klep
(2011, [1]), Burgdorf, Klep and Povh (2013, [2]), Cimpric and Zalar (2013,
[3]), Kimsey and Woerdeman (2013, [6]), Kimsey (2011, [5]) and the refer-
ences therein. In particular, in [3], Cimpric and Zalar proved a version of
the Haviland theorem for matrix polynomials, and they applied this devel-
opment to solve the matrix-valued moment problem on compact basic closed
semi-algbraic sets.
In the case where K ⊆ Rn is a basic closed semi-algebraic set defined by
a subset G = {g1, . . . , gm} of R[X], almost all solutions of the matrix-valued
K-moment problems are obtained using the positivity of the given linear
functional on the quadratic module M (G)t in Mt(R[X]), where M (G) is the
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
3
quadratic module in R[X] generated by G and
r
M (G)t = {
Xi=1
miAT
i
Air ∈ N, mi ∈ M (G), Ai ∈ Mt(R[X])}.
The main aim of this paper is to introduce some kinds of K-moment
problems for matrix polynomials and solve these matrix-valued K-moment
problems in the case where K is the hypercube [−1, 1]n. The most advantage
to work on the hypercube [−1, 1]n (and in general, on convex, compact
polyhedra with non-empty interior) is that any matrix polynomial which is
positive definite on [−1, 1]n can be represented in terms of positive definite
scalar matrices (see Theorem 3.1 for the case n = 1 and Corollary 4.2 for
n ≥ 1). Then the solution of the matrix-valued [−1, 1]n-moment problem is
obtained basing on the positivity of the given linear functional on the cone
of positive definite scalar matrices, which is well described.
The paper is organized as follows. In Section 2 we introduce the defi-
nition of the tracial K-moment problem and the sequential matrix-valued
K-moment problem and establish an equivalence of the solvability of these
problems (Proposition 2.8). Moreover, we also prove in this section a version
of Haviland's theorem for matrix polynomials which was proved mainly by
Cimpric and Zalar in [3]. Next, in Section 3 we solve the tracial (resp. sequen-
tial matrix-valued) [−1, 1]-moment problem, which is established based on a
representation of univariate matrix polynomials positive definite on [−1, 1].
A solution of the tracial (resp. sequential matrix-valued) [0, 1]-moment prob-
lem is also given in this section. Finally, in Section 4 we solve the tracial
(resp. sequential matrix-valued) moment problem on the general hypercube
[−1, 1]n.
Notation. For any matrix A ∈ Mt(R[X]), the notation A ≥ 0 means A
is positive semidefinite, i.e. for each x ∈ Rn, vT A(x)v ≥ 0 for all v ∈ Rt;
A > 0 means A is positive definite, i.e. for each x ∈ Rn, vT A(x)v > 0 for
all v ∈ Rt \ {0}.
2. Haviland's theorem for matrix polynomials
2.1. Matrix-valued measures and integrals. In this section we recall
some basic notions of the matrix-valued measures and integrals on a closed
subset of Rn, which can be seen from the thesis of Kimsey [5].
Throughout this subsection, let X ⊆ Rn be a non-empty closed set.
Definition 2.1. Denote by B(X) the smallest σ-algebra generated from the
open (or equivalently closed) subsets of X. A measure σ defined on B(X)
is called a Borel measure. A Borel measure σ on B(X) is called finite if
σ(X) < +∞. Denote by m(X) the set of all finite measures on X.
A measure σ ∈ m(X) is called positive if σ(E) ≥ 0 for all E ∈ B(X).
The set of all finite positive Borel measures on X is denoted by m+(X).
For each σ ∈ m(X), the support of σ is defined by
supp(σ) := {E ∈ B(X) : σ(E) > 0},
4
CONG TRINH LE
where σ(E) := σ(E) for all E ∈ B(X).
Definition 2.2. Let t be a positive integer. Let σij ∈ m(X), i, j = 1, . . . , t.
Define the matrix-valued function E : B(X) −→ Mt(R) by
E(A) := (σij(A))i,j=1,...,t ∈ Mt(R), ∀A ∈ B(X).
The matrix-valued function E defined by this way is called a matrix-valued
measure on X. The set of all matrix-valued measure on X is denoted by
M (X). The set
supp(E) :=
t
[i,j=1
supp(σij)
is called the support of the matrix-valued measure E.
If σij = σji for all i, j = 1, . . . , t, we say that E is a symmetric measure.
In addition, if for all A ∈ B(X) and for all v ∈ Rt we have vT E(A)v ≥ 0,
then E is called a positive semidefinite matrix-valued measure. The set of all
positive semidefinite matrix-valued measures on X is denoted by M +(X).
Definition 2.3. Let E = (σij) ∈ M (X). A function f : X → R is called
E-measurable if f is σij-measurable for every i, j = 1, . . . , t.
Let E = (σij) ∈ M (X) and f : X → R be E-measurable. The matrix-
valued integral of f with respect to the matrix measure E is defined by
ZX
f (x)dE(x) := (cid:16)ZX
f (x)dσij(x)(cid:17)i,j=1,...,t
∈ Mt(R).
2.2. Tracial K-moment problems. Let L be a real-valued linear func-
tional on Mt(R[X]), where R[X] := R[X1, . . . , Xn] and K ⊆ Rn a closed
subset.
Definition 2.4 (cf. [3]). L is called a tracial K-moment functional
exists a positive semidefinite matrix-valued E ∈ M +(Rn) such that
if there
supp(E) ⊆ K and L (F) = Z tr(F(x)dE(x)), ∀F ∈ Mt(R[X]).
(2.1)
The matrix-valued measure E ∈ M +(Rn) satisfying (2.1) is called a repre-
senting measure of the tracial K-moment functional L .
Problem 2.5. Let L be a linear functional on Mt(R[X]) and K ⊆ Rn a
closed set. The tracial K-moment problem asks when does there exist
a matrix-valued measure E ∈ M +(Rn) satisfying the conditions (2.1)?
The following definition of matrix-valued K-moment sequences is learned
from Kimsey in [5].
Definition 2.6. Let S = (Sα)α∈Nn
0 be a multi-indexed sequence of symmet-
ric matrices in Mt(R) and K ⊆ Rn a closed subset. The sequence S is called
a matrix-valued K-moment sequence if there exists a positive semidefinite
matrix-valued E ∈ M +(Rn) such that
supp(E) ⊆ K and Sα = Z xαdE(x), ∀α ∈ Nn
0 ,
(2.2)
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
5
where xα = xα1
1 . . . xαn
n with α = (α1, . . . , αn) and x = (x1, . . . , xn) ∈ Rn.
The matrix-valued measure E ∈ M +(Rn) satisfying (2.2) is called a rep-
resenting measure of the matrix-valued K-moment sequence S.
Problem 2.7. Let S = (Sα)α∈Nn
0 be a multi-indexed sequence of symmetric
matrices in Mt(R) and K ⊆ Rn a closed set. The sequential matrix-
valued K-moment problem asks when does there exist a matrix-valued
measure E ∈ M +(Rn) satisfying the conditions (2.2)?
For each matrix-valued sequence S = (Sα)α∈Nn
0 , let us consider the real-
valued function LS on Mt(R[X]) defined as follows. For each F(X) =
d
Xα=0
FαX α ∈ Mt(R[X]), define
LS(F) :=
d
Xα=0
tr(FαSα).
It is easy to check that LS is linear, which is called the Riesz functional
associated to the sequence S.
Proposition 2.8. Let S = (Sα)α∈Nn
0 be a sequence of symmetric matrices
in Mt(R) and K ⊆ Rn a closed set. Then S is a matrix-valued K-moment
sequence if and only if LS is a tracial K-moment functional.
Proof. Assume that S is a matrix-valued K-moment sequence. Then there
exists a matrix-valued measure E in M +(Rn) supported in K and
Sα = Z xαdE(x), ∀α ∈ Nn
0 .
Then, for each F(X) =
have
d
Xα=0
FαX α ∈ Mt(R[X]), by definition of LS, we
LS(F) = Xα
tr(FαSα) = Xα
tr(cid:16)Z xαFαdE(x)(cid:17)
= Z tr(cid:16)Xα
FαxαdE(x)(cid:17) = Z tr(F(x)dE(x)).
This implies that LS is a tracial K-moment functional.
Conversely, assume that E is a matrix-valued measure in M +(Rn) satis-
fying
supp(E) ⊆ K and LS(F) = Z tr(F(x)dE(x)), ∀F ∈ Mt(R[X]).
6
CONG TRINH LE
In particular, for each coordinate matrix Wk,l of the algebra Mt(R) and
each α ∈ Nn
0 , we have
LS(X αWk,l) = Z xαtr(Wk,ldE(x)) = Z xαdEl,k.
Here we use the fact that for each k, l = 1, . . . , t and for each matrix A =
(Ai,j)i,j ∈ Mt(R), we have
tr(Wk,lA) = Al,k.
On the other hand, by definition of the associated Riesz functional, we have
LS(X αWk,l) = tr(Wk,lSα) = (Sα)l,k.
It follows that
ST
α = (cid:0)(Sα)l,k(cid:1)k,l = (cid:16)Z xαdEl,k(x)(cid:17)k,l
= (cid:16)Z xαdE(x)(cid:17)T
.
Equivalently, Sα = Z xαdE(x), i.e. S is a matrix-valued K-moment se-
(cid:3)
quence.
2.3. Haviland's theorem for matrix polynomials.
Theorem 2.9. Let L be a real-valued linear functional on Mt(R[X]) and
K ⊆ Rn a closed set. Then the following are equivalent.
(1) L is a tracial K-moment functional, i.e. there exists a positive semi-
definite matrix-valued measure E on Rn whose support is contained
in K such that
L (F) = Z tr(FdE).
(2) L (F) ≥ 0 for all F ≥ 0 on K.
(3) L (F + ǫI) ≥ 0 for all F ≥ 0 on K and for all ǫ > 0.
(4) L (F) ≥ 0 for all F > 0 on K.
Proof. The equivalence of (1) and (2) was proved by Cimpric and Zalar [3,
Theorem 3 and Remark 5].
(2) → (3) is obvious, since L (I) ≥ 0 by hypothesis. Now we check (3) → (2).
for n ∈ N. Then by the assumption of
Take F ≥ 0 on K. Consider ǫ =
(3),
1
n
0 ≤ L (F +
1
n
I) = L (F) +
1
n
L (I),
where L (I) ≥ 0 by hypothesis. Letting n → +∞, we get L (F) ≥ 0, i.e. we
have (2).
It is obvious that (2) → (4). Now we verify (4) → (3). Take F ≥ 0 on
K and ǫ > 0. Then F + ǫI > 0 on K. Hence by the assumption we have
L (F + ǫI) ≥ 0, i.e. we have (3). The proof is complete.
(cid:3)
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
7
3. Tracial moment problems on the intervals [−1, 1] and [0, 1]
Firstly we propose a version of the Bernstein theorem (cf. [10, Prop. 3.4])
for matrix polynomials, representing a matrix polynomial positive definite
on the interval [−1, 1]. This is a special case of the Handelman's Positivstel-
lensatz for matrix polynomials established in [7], but the proof given here is
special for the case of one-variable matrix polynomials.
Theorem 3.1. Let F ∈ Mt(R[x]) be a matrix polynomial of degree d > 0.
Assume F(x) > 0 for all x ∈ [−1, 1]. Then there exists a positive number
N ∈ N and positive definite matrices Gi, i = 1, . . . , N + d such that
F(x) =
N +d
Xi=0
Gi(1 + x)i(1 − x)N +d−i.
The main idea of the proof is learned from the proof [10, Prop. 3.4] for
polynomials, using the Goursat transform and the Pólya theorem for homo-
geneous matrix polynomials established in [8].
Let F be as in Theorem 3.1. The Goursat transform of F is defined by
F(x) := (1 + x)dF(
1 − x
1 + x
).
Lemma 3.2. deg( F) = d, Fd > 0, and F(t) > 0 for all t ∈ [0, +∞), where
Fd denotes the leading coefficient of the matrix polynomial F.
Proof. Let F(x) = Pd
Fixi, with Fd 6≡ 0. Then we can write F as
i=0
d
Xi=0
The leading coefficient of F is
F(x) =
Fi(1 − x)i(1 + x)d−i.
Fd :=
d
Xi=0
(−1)iFi = F(−1)
which is positive definite by the assumption of F.
For every t ∈ [0, +∞), x :=
∈ (−1, 1]. Observe that t =
1 − x
1 + x
, and
hence
1 − t
1 + t
F(t) = (1 + t)dF(x).
It follows from the assumption of F that F(t) > 0.
(cid:3)
Lemma 3.3. Let G ∈ Mt(R[x]) be a matrix polynomial of degree d with a
positive definite leading coefficient Gd. If G(x) > 0 for all x ∈ [0, +∞)
then there exists a positive integer N and positive definite matrices Hi,
i = 0, . . . , N + d such that
(1 + x)N G(x) =
N +d
Xi=0
Hixi.
8
CONG TRINH LE
Proof. Assume G(x) = Pd
R2
Gixi. Denote by ∆ the standard simplex in
i=0
∆ = {(x, y) ∈ R2x ≥ 0, y ≥ 0, x + y + 1}.
Let Gh(x, y) ∈ Mt(R[x, y]) be the homogenization of G with respect to the
new variable y, defined by
Gh(x, y) :=
d
Xi=0
Gixiyd−i.
For any (x, y) ∈ ∆,
if y = 0, we have x = 1, then
If y > 0, we have
Gh(1, 0) = Gd > 0.
Gh(x, y) = ydG(
x
y
) > 0,
since
x
y
∈ [0, +∞). In summary, we have Gh(x, y) > 0 for all (x, y) ∈ ∆.
It follows from the Pólya theorem for homogeneous matrix polynomials
proved by Scherer and Hol [8, Theorem 3] that there exist a positive integer
N and positive definite matrices Gk,l, with k, l ∈ N0, k + l = 0, . . . , N + d
such that
(x + y)N Gh(x, y) = Xk+l=N +d
Gk,lxkyl.
(3.1)
Substituting y = 1 in both sides of the expression (3.1), observing that
Gh(x, 1) = G(x), we get
(1 + x)N G(x) = Xk+l=N +d
Gk,lxk =
N +d
Xk=0
Gk,N +d−kxk.
For each k = 0, . . . , N + d, the matrices Hk := Gk,N +d−k satisfy conclusion
of the lemma.
(cid:3)
Now we are ready to prove Theorem 3.1.
Proof of Theorem 3.1. Let F satisfy the hypothesis of the theorem. It
follows from Lemma 3.2 that the Goursat transform F satisfy the following
properties
• deg( F) = d;
• Fd > 0;
• F(t) > 0 for all t ∈ [0, +∞).
Ir follows from Lemma 3.3 that there exist a positive integer N and positive
definite matrices Hi, i = 0, . . . , N + d such that
(1 + t)N F(t) =
N +d
Xi=0
Hiti, ∀t ∈ [0, +∞).
(3.2)
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
9
For any x ∈ (−1, 1], there exists a unique t ∈ [0, +∞) such that x =
Then
t =
1 − x
1 + x
, (1 + t)−1 =
1 + x
2
, and F(t) = (1 + t)dF(x).
Substituting into (3.2), we get
1 − t
1 + t
.
(1 + t)N (1 + t)dF(x) =
N +d
Xi=0
Hi(cid:0)
1 − x
1 + x(cid:1)i
.
It follows that
N +d
F(x) = (1 + t)−N +d
Hi(cid:0)
Xi=0
N +d
1 − x
1 + x(cid:1)i
Hi(cid:0)
1 − x
1 + x(cid:1)i
Xi=0
1
2N +d (1 + x)N +d
Xi=0 (cid:16) 1
2N +d
N +d
=
=
Hi(cid:17)(1 − x)i(1 + x)N +d−i.
Hence the matrices Gi :=
of the theorem.
1
2N +d
Hi, i = 0, . . . , N + d, satisfy the conclusion
(cid:3)
It follows from the Haviland theorem for matrix polynomials (Theorem
2.9) and Theorem 3.1 the following solution of the tracial [−1, 1]-moment
problem.
Corollary 3.4. Let L be a real-valued linear functional on Mt(R[x]). Then
the following are equivalent.
(1) L is a tracial [−1, 1]-moment functional.
(2) L ((1+x)k(1−x)lG) ≥ 0 for all positive definite matrices G ∈ Mt(R)
and k, l ∈ N0.
Proof. If (1) holds, for each G positive definite and k, l ∈ N0, since (1 +
x)k(1 − x)lG ≥ 0 on [−1, 1], it follows from Theorem 2.9, (1) ⇒ (2), that
L ((1 + x)k(1 − x)lG) ≥ 0. Hence we have (1) → (2).
Conversely, assume (2) holds. Then for each F ∈ St(R[x]) of degree d,
F > 0 on [−1, 1], it follows from Theorem 3.1 that
F(x) =
N +d
Xi=0
Gi(1 + x)i(1 − x)N +d−i
for some N > 0 and Gi > 0 for all i = 0, . . . , N + d. Then
L (F) =
N +d
Xi=0
L (Gi(1 + x)i(1 − x)N +d−i) ≥ 0.
10
CONG TRINH LE
It follows from Theorem 2.9, (4) ⇒ (1), that L is a tracial [−1, 1]-moment
functional, i.e. (2) → (1). The proof is complete.
(cid:3)
As a consequence of this result, we obtain the following solution of the
tracial [0, 1]-moment problem, which was also solved by Cimpric and Zalar
[3, Coro. 1], using sums of Hermitian squares.
Corollary 3.5. Let L be a real-valued linear functional on Mt(R[x]). Then
the following are equivalent.
(1) L is a tracial [0, 1]-moment functional.
(2) L (xk(1 − x)lG) ≥ 0 for all positive definite matrices G ∈ Mt(R)
and k, l ∈ N0.
Proof. Observe that there is a bijection from the interval [−1, 1] onto the
interval [0, 1] given by
x 7→
x + 1
2
, ∀x ∈ [−1, 1].
Using this bijection, it follows from Theorem 3.1 that for F ∈ St(R[x]) of
degree d, if F > 0 on [0, 1] then
F(x) =
N +d
Xi=0
Gixi(1 − x)N +d−i
for some N > 0 and Gi > 0 for all i = 0, . . . , N + d. Then the result follows
from this representation and Theorem 2.9.
(cid:3)
For the sequential matrix-valued [−1, 1]-problem, we have the following
solution.
Corollary 3.6. Let S = (Si)i∈N0 be a sequence of symmetric matrices in
Mt(R). Then S is a matrix-valued [−1, 1]-moment sequence if and only if
k
Xi=0
l
Xj=0
(−1)j(cid:18)k
i(cid:19)(cid:18)l
j(cid:19)tr(GSi+j) ≥ 0
for all positive definite matrices G ∈ Mt(R) and k, l ∈ N0.
Proof. Let LS be the Riesz functional on Mt(R[x]) associated to the se-
quence S. For each G ∈ St(R[x]), G > 0, and k, l ∈ N0, we have
k
l
(−1)j(cid:18)k
LS(cid:0)(1 + x)k(1 − x)lG(cid:1) = LS(cid:0)
Xi=0
Xj=0
Xj=0
Xi=0
(−1)j(cid:18)k
=
k
i(cid:19)(cid:18)l
i(cid:19)(cid:18)l
j(cid:19)xi+j G(cid:1)
j(cid:19)LS(xi+jG)
l
l
=
k
Xi=0
Xj=0
(−1)j(cid:18)k
i(cid:19)(cid:18)l
j(cid:19)tr(GSi+j).
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
11
By Proposition 2.8, S is a matrix-valued [−1, 1]-moment sequence if and
only if LS is a tracial [−1, 1]-moment functional. It follows from Corollary
3.4 that this is equivalent to LS(cid:0)(1 + x)k(1 − x)lG(cid:1) ≥ 0 for all positive
definite matrices G ∈ Mt(R) and k, l ∈ N0. This implies the result.
(cid:3)
Similarly, using Corollary 3.5, we have the following solution of the se-
quential matrix-valued [0, 1]-moment problem.
Corollary 3.7. Let S = (Si)i∈N0 be a sequence of symmetric matrices in
Mt(R). Then S is a matrix-valued [0, 1]-moment sequence if and only if
k
Xi=0
(−1)i(cid:18)k
i(cid:19)tr(GSi+l) ≥ 0
for all positive definite matrices G ∈ Mt(R) and k, l ∈ N0.
4. Tracial moment problems on the hypercube [−1, 1]n
In this section we consider a convex, compact polyhedron K ⊆ Rn with
non-empty interior. Assume that K is the basic closed semi-algebraic set
defined by linear polynomials L1, . . . , Lm ∈ R[X] := R[X1, . . . , Xn], i.e.
K = {x ∈ RnL1(x) ≥ 0, . . . , Lm(x) ≥ 0}.
The following version of Handelman's Positivstellensatz for matrix polyno-
mials was proved by the the authors in [7].
Theorem 4.1. Let F ∈ St(R[X]) be a matrix polynomial of degree d > 0.
If F(x) > 0 for all x ∈ K, then there exist a positive integer N and positive
definite matrices Gα ∈ Mt(R), α ∈ Nm
0 , α = 0, . . . , N + d, such that
F(X) =
N +d
Xα=0
GαLα1
1 . . . Lαm
m .
Applying Theorem 4.1 for the linear polynomials
l1 = 1 + X1, l2 = 1 − X1, . . . , l2n−1 = 1 + Xn, l2n = 1 − Xn
we obtain the following representation for matrix polynomials positive defi-
nite on the hypercube [−1, 1]n.
Corollary 4.2. Let F ∈ St(R[X]) be a matrix polynomial of degree d > 0.
If F(x) > 0 for all x ∈ [−1, 1]n, then there exist a positive integer N and
positive definite matrices Gα ∈ Mt(R), α ∈ N2n
0 , α = 0, . . . , N + d, such
that
N +d
F(X) =
Xα=0
Gαlα1
1 lα2
2 . . . lα2n−1
2n−1 lα2n
2n .
Now, applying the Haviland theorem for matrix polynomials (Theorem
2.9) and Corollary 4.2 we obtain the following solution of the tracial [−1, 1]n-
moment problem.
12
CONG TRINH LE
Corollary 4.3. Let L be a real-valued linear functional on Mt(R[X]). Then
the following are equivalent.
(1) L is a tracial [−1, 1]n-moment functional.
(2) L (lα1
2 . . . lα2n−1
2n−1 lα2n
2n
1 lα2
Mt(R) and α = (α1, . . . , α2n) ∈ N2n
0 .
G) ≥ 0 for all positive definite matrices G ∈
Observe that the solution of the tracial K-moment problem, where K is
a convex, compact polyhedron in Rn with non-empty interior, can be solved
by the same way, using the linear polynomial Li instead of lj in Corollary
4.3.
Finally, applying Proposition 2.8 and Corollary 4.3 we obtain the following
solution of the sequential matrix-valued [−1, 1]n-moment problem.
Corollary 4.4. Let S = (Sα)α∈Nn
0 be a sequence of symmetric matrices in
Mt(R). Then S is a matrix-valued [−1, 1]n-moment sequence if and only if
β(cid:19)tr(GSβ) ≥ 0
for all positive definite matrices G ∈ Mt(R) and α ∈ N2n
0 .
(−1)e(α)(cid:18)α
Xβ∈N2n
0 ,β≤α
Here, for α = (α1, . . . , α2n) and β = (β1, . . . , β2n), β ≤ α means βi ≤ αi
n
for all i = 1, . . . , 2n; (cid:0)α
β(cid:1) := (cid:0)α1
β1(cid:1) . . .(cid:0)α2n
β2n(cid:1); and e(α) :=
α2i.
Xi=1
Proof. The proof is similar to that of Corollary 3.6, using Proposition 2.8,
Corollary 4.3 and the identity
(1+X1)α1(1−X1)α2 . . . (1+Xn)α2n−1 (1−Xn)α2n = Xβ∈N2n
0 ,β≤α
(−1)e(α)(cid:18)α
β(cid:19)X β,
where α = (α1, . . . , α2n).
(cid:3)
Acknowledgements. The author would like to express his sincere grati-
tude to Prof. Konrad Schmudgen for his series of lectures on moment prob-
lems given at the Vietnam Institute for Advanced Study in Mathematics
(VIASM) which motivates the author to study the matrix-valued moment
problems. This paper was finished during the visit of the author at VIASM.
He thanks VIASM for financial support and hospitality.
References
[1] S. Burgdorf and I. Klep The truncated tracial moment problem, J. Oper. Theory
68(1)(2012), 141 -- 163.
[2] S. Burgdorf, I. Klep and J. Povh, The tracial moment problem and trace-
optimization of polynomials, Math. Prog. 137, Issue 1 -- 2 (2013), 557 -- 578.
[3] J.Cimpri c, A. Zalar, Moment problems for operator polynomials, J. Math. Anal.
Appl. 401(1) (2013), 307-316.
[4] H. Dette and W.J. Studden, Matrix measures, moment spaces and Favard's the-
orem for the interval [0, 1] and [0, +∞), Linear Algebra Appl. 345 (2002), 169 -- 193.
TRACIAL MOMENT PROBLEMS ON HYPERCUBES
13
[5] D. P. Kimsey, Matrix-valued moment problems, Ph.D. dissertation, Drexel Univer-
sity, Philadelphia, PA, 2011.
[6] D. P. Kimsey and H. J. Woerdeman, Truncated matrix-valued moment problems
on Rd, Cd, and Td, Trans. Amer. Math. Soc. 365(10)(2013), 5393-5430.
[7] C.-T. Lê, T.H.B. Du, Handelman's Positivstellensatz for polynomial matrices posi-
tive definite on polyhedra, Positivity 22 (3) (2018), 449-460.
[8] C.W. Scherer, C.W.J. Hol, Matrix sum-of-squares relaxations for robust semi-
definite programs, Math. Program. 107 no. 1-2, Ser. B (2006), 189 -- 211.
[9] K. Schm udgen, The K-moment problem for compact semi-algebraic sets, Math. Ann.
289(1) (1991), 203-206.
[10] K. Schm udgen, The moment problem, Springer, 2017.
Department of Mathematics, Quy Nhon University, 170 An Duong Vuong,
Quy Nhon, Binh Dinh, Vietnam
E-mail address: [email protected]
|
1507.06860 | 1 | 1507 | 2015-07-24T14:35:15 | What is odd about binary Parseval frames? | [
"math.FA"
] | This paper examines the construction and properties of binary Parseval frames. We address two questions: When does a binary Parseval frame have a complementary Parseval frame? Which binary symmetric idempotent matrices are Gram matrices of binary Parseval frames? In contrast to the case of real or complex Parseval frames, the answer to these questions is not always affirmative. The key to our understanding comes from an algorithm that constructs binary orthonormal sequences that span a given subspace, whenever possible. Special regard is given to binary frames whose Gram matrices are circulants. | math.FA | math |
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
ZACHERY J. BAKER, BERNHARD G. BODMANN, MICAH G. BULLOCK,
SAMANTHA N. BRANUM, AND JACOB E. MCLANEY
Abstract. This paper examines the construction and properties of bi-
nary Parseval frames. We address two questions: When does a binary
Parseval frame have a complementary Parseval frame? Which binary
symmetric idempotent matrices are Gram matrices of binary Parseval
frames? In contrast to the case of real or complex Parseval frames, the
answer to these questions is not always affirmative. The key to our un-
derstanding comes from an algorithm that constructs binary orthonor-
mal sequences that span a given subspace, whenever possible. Special
regard is given to binary frames whose Gram matrices are circulants.
1. Introduction
Much of the literature on frames, from its beginnings in non-harmonic
Fourier analysis [7] to comprehensive overviews of theory and applications
[6, 13, 14] assumes an underlying structure of a real or complex Hilbert space
to study approximate expansions of vectors. Indeed, the correspondence be-
tween vectors in Hilbert spaces and linear functionals given by the Riesz
representation theorem provides a convenient way to characterize Parseval
frames, sequences of vectors that behave in a way that is similar to orthonor-
mal bases without requiring the vectors to be linearly independent [6]. In-
corporating linear dependence relations is useful to permit more flexibility
for expansions and to suppress errors that may model faulty signal trans-
missions in applications [20, 21, 17, 18, 12, 16, 3].
The concept of frames has also been established even in vector spaces
without (definite) inner product [4, 11].
In fact, the well known theory
of binary codes could be seen as a form of frame theory, in which linear
dependence relations among binary vectors are examined [19, 10, 2]. Here,
binary vector spaces are defined over the finite field with two elements;
a frame for a finite dimensional binary vector space is simply a spanning
sequence [4].
In a preceding paper [5], the study of binary codes from a
frame theoretic perspective has lead to additional combinatorial insights in
the design of error-correcting codes.
Date: January 30, 2018.
2010 Mathematics Subject Classification. 42C15.
Key words and phrases. frames, Parseval frames, finite-dimensional vector spaces, bi-
nary numbers, orthogonal extension principle, switching equivalence, Gram matrices.
This research was supported by NSF grant DMS-1109545 and DMS-1412524.
1
2 Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
The present paper is concerned with binary Parseval frames. These binary
frames provide explicit expansions of binary vectors using a bilinear form
that resembles the dot product in Euclidean spaces.
In contrast to the
inner product on real or complex Hilbert spaces, there are many nonzero
vectors whose dot product with themselves vanishes. Such vectors have
special significance in our results. Counting the number of non-vanishing
entries motivates calling them even vectors, whereas the others are called
odd. As a consequence of the degeneracy of the bilinear form, there are some
striking differences with frame theory over real or complex Hilbert spaces.
In this paper, we explore the construction and properties of binary Parseval
frames, and compare them with real and complex ones. Our main results
are as follows:
In the real or complex case, it is known that each Parseval frame has a
Naimark complement [6]. The complementarity is most easily formulated
by stating that the Gram matrices of two complementary Parseval frames
sum to the identity. We show that in the binary case, not every Parseval
frame has a Naimark complement. In addition, we show that a necessary
and sufficient condition for its existence is that the Parseval frame contains
at least one even vector.
Moreover, we study the structure of Gram matrices. The Gram matrices
of real or complex Parseval frames are characterized as symmetric or hermit-
ian idempotent matrices. The binary case requires the additional condition
that at least one column vector of the matrix is odd.
The general results we obtain are illustrated with examples. Special re-
gard is given to cyclic binary Parseval frames, whose Gram matrices are
circulants.
2. Preliminaries
We define the notions of a binary frame and a binary Parseval frame as
in a previous paper [4]. The vector space that these sequences of vectors
span is the direct sum Zn
2 = Z2 ⊕ · · · ⊕ Z2 of n copies of Z2 for some n ∈ N.
Here, Z2 is the field of binary numbers with the two elements 0 and 1, the
neutral element with respect to addition and the multiplicative identity.
Definition 2.1. A binary frame is a sequence F = {f1, . . . , fk} in a binary
vector space Zn
2 such that span F = Zn
2 .
A simple example of a frame is the canonical basis {e1, e2, . . . , en} for
2 . The ith vector has components (ei)j = δi,j, thus (ei)i = 1 is the only
i=1 is expanded
Zn
non-vanishing entry for ei. Consequently, a vector x = (xi)n
in terms of the canonical basis as x =Pn
Frames provide similar expansions of vectors in linear combinations of the
frame vectors. Parseval frames are especially convenient for this purpose,
because the linear combination can be determined with little effort. In the
real or complex case, this only requires computing values of inner products
between the vector to be expanded and the frame vectors. Although we
i=1 xiei.
With the help of this dot product, we define a Parseval frame for Zn
2 .
Definition 2.3. A binary Parseval frame is a sequence of vectors F = {f1, . . .,
fk} in Zn
2 , the sequence satisfies the reconstruction
identity
2 such that for all x ∈ Zn
k
(2.1)
x =
(x, fj)fj .
Xj=1
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
3
cannot introduce a non-degenerate inner product in the binary case, we
define Parseval frames using a bilinear form that resembles the dot product
on Rn. Other choices of bilinear forms and a more general theory of binary
frames have been investigated elsewhere, see [9].
Definition 2.2. The dot product on Zn
Z2 given by
2 is the bilinear map (·, ·) : Zn
2 × Zn
2 →
xn(cid:19) , y1
yn!! :=
(cid:18) x1
...
...
n
Xi=1
xiyi.
To keep track of the specifics of such a Parseval frame, we then also say that
F is a binary (k, n)-frame.
In the following, we use matrix algebra whenever it is convenient for
establishing properties of frames. We write A ∈ Mm,n(Z2) when A an m × n
2 to Zm
matrix with entries in Z2 and identify A with the linear map from Zn
2
induced by left multiplication of any (column) vector x ∈ Zn
2 with A. We
let A∗ denote the adjoint of A ∈ Mm,n(Z2); that is, (Ax, y) = (x, A∗y) for
all x ∈ Zn
2 and consequently, A∗ is the transpose of A.
2 , y ∈ Zm
Definition 2.4. Each frame F = {f1, . . . , fk} is associated with its anal-
ysis matrix ΘF , whose ith row is given by the ith frame vector for i ∈
{1, 2, . . . , k}. Its transpose Θ∗
F is called the synthesis matrix.
With the help of matrix multiplication, the reconstruction formula (2.1)
of a binary (k, n)-frame F with analysis matrix ΘF is simply expressed as
(2.2)
where In is the n × n identity matrix. We also note that for any x, y ∈ Zn
2 ,
then
F ΘF = In ,
Θ∗
hΘF x, ΘF yi = hx, yi
which motivates speaking of ΘF as an isometry, as in the case of real or
complex inner product spaces.
Another way to interpret identity (2.2) is in terms of the column vectors
of ΘF . Again borrowing a concept from Euclidean spaces, we introduce
orthonormality.
Definition 2.5. We say that a sequence of vectors {v1, v2, . . . , vr} in Zn
2 is
orthonormal if (vi, vj) = δi,j for i, j ∈ {1, 2, . . . , r}, that is, the dot product
of the pair vi and vj vanishes unless i = j, in which case it is equal to one.
4 Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
Inspecting the matrix identity (2.2), we see that a binary k × n matrix Θ
is the analysis matrix of a binary Parseval frame if and only if the columns
of Θ form an orthonormal sequence in Zk
2.
The orthogonality relations between the frame vectors are recorded in
the Gram matrix, whose entries consists of the dot products of all pairs of
vectors.
Definition 2.6. The Gram matrix of a binary frame F = {f1, f2, . . . , fk} for
Zn
2 is the k × k matrix G with entries Gi,j = (fj, fi).
It is straightforward to verify that the Gram matrix of F is expressed as
the composition of the analysis and synthesis matrices,
G = ΘF Θ∗
F .
The identity (2.2) implies that the Gram matrix of a Parseval frame satisfies
the equations
G = G∗ = G2 .
For frames over the real or complex numbers, these equations characterize
the set of all Gram matrices of Parseval frames as orthogonal projection
matrices. However, in the binary case, this is only a necessary condition, as
shown in the following proposition and the subsequent example.
Proposition 2.7. If M is binary matrix that satisfies M = M 2 = M ∗ and
it has only even column vectors, then M is not the Gram matrix of a binary
Parseval frame.
Proof. If G is the Gram matrix of a Parseval frame with analysis operator
Θ, then GΘ = ΘΘ∗Θ = Θ, and thus for each column ω of Θ, we obtain the
eigenvector equation Gω = ω. By the orthonormality of the columns of Θ,
each ω is odd.
On the other hand, if M has only even columns, then any eigenvector
corresponding to eigenvalue one is even, because it is a linear combination
of the column vectors of M . This means M cannot be the Gram matrix of
a binary Parseval frame.
(cid:3)
The following example shows that idempotent symmetric matrices that
are not Gram matrices of binary Parseval frames exist for any odd dimension
k ≥ 3.
Example 2.8. Let k ≥ 3 be odd and let M be the k × k matrix whose
entries are all equal to one except for vanishing entries on the diagonal,
Mi,j = 1 − δi,j , i, j ∈ {1, 2, . . . , k}. This matrix satisfies M = M 2 = M ∗, but
only has even columns and by the preceding proposition, it is not the Gram
matrix of a binary Parseval frame.
As shown in Section 4, having only even column vectors is the only way
a binary symmetric idempotent matrix can fail to be the Gram matrix of a
Parseval frame. The construction of Example 2.8 is intriguing, because the
alternative choice where k is odd and all entries of M are equal to one is the
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
5
Gram matrix of a binary Parseval frame. The relation between these two
alternatives can be interpreted as complementarity, which will be explored
in more detail in the next section.
3. Complementarity for binary Parseval frames
Over the real or complex numbers, each Parseval frame has a so-called
Naimark complement [6]; if G is the Gram matrix of a real or complex
Parseval frame, then it is an orthogonal projection matrix, and so is I − G,
which makes it the Gram matrix of a complementary Parseval frame.
We adopt the same definition for the binary case.
Definition 3.1. Two binary Parseval frames F and G having analysis oper-
ators ΘF ∈ Mk,n(Z2) and ΘG ∈ Mk,k−n(Z2) are complementary if
ΘF Θ∗
F + ΘGΘ∗
G = Ik .
We also say that F and G are Naimark complements of each other.
There is an equivalent statement of complementarity in terms of the block
matrix U = (ΘF ΘG) formed by adjoining ΘF and ΘG being orthogonal,
meaning U U ∗ = U ∗U = I, just as in the real case (or as U being unitary in
the complex case).
Proposition 3.2. Two binary Parseval frames F and G having analysis
operators ΘF ∈ Mk,n(Z2) and ΘG ∈ Mk,k−n(Z2) are complementary if and
only if the block matrix (ΘF ΘG) is an orthogonal k × k matrix.
Proof. In terms of the block matrix (ΘF ΘG), the complementarity is ex-
pressed as
(ΘF ΘG)(ΘF ΘG)∗ = Ik .
Since U = (ΘF ΘG) is a square matrix, U U ∗ = I is equivalent to U ∗ also
being a left inverse of U , meaning U U ∗ = U ∗U = Ik, U is orthogonal.
(cid:3)
In the binary case, not every Parseval frame has a Naimark complement.
For example, if k ≥ 3 is odd and n = 1, the frame consisting of k vectors
{1, 1, . . . , 1} in Z2 is Parseval, and the Gram matrix G is the k × k matrix
whose entries are all equal to one. However, I − G ≡ I + G is the matrix
M appearing in Example 2.8, which is not the Gram matrix of a binary
Parseval frame. This motivates the search for a condition that characterizes
the existence of complementary Parseval frames.
3.1. A simple condition for the existence of complementary Parse-
val frames. We observe that if F is a Parseval frame with analysis operator
ΘF that extends to an orthogonal matrix, then the column vectors of ΘF are
a subset of a set of n orthonormal vectors. This is true in the binary as well
as the real or complex case. Thus, one could try to relate the construction
of a complementary Parseval frame to a Gram-Schmidt orthogonalization
strategy. Indeed, this idea allows us to formulate a concrete condition that
6 Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
characterizes when F has a complementary Parseval frame. We prepare this
result with a lemma about extending orthonormal sequences.
Lemma 3.3. A binary orthonormal sequence Y = {v1, v2, . . . , vr} in Zk
2
with r ≤ k − 1 extends to an orthonormal sequence {v1, v2, . . . , vk} if and
2 whose entries are all equal
i=1 vi 6= ιk, where ιk is the vector in Zk
only if Pr
to one.
Proof. If the sequence extends, then {v1, v2, . . . , vk} forms a Parseval frame
for Zk
i=1(ιk, vi)vi = ιk. On the
other hand, the orthonormality forces the set {v1, v2, . . . , vk} to be linearly
independent, so ιk cannot be expressed as a linear combination of a proper
subset.
2, and by the orthonormality, Pk
i=1 vi = Pk
To show the converse, we use an inductive proof. Let V be the analysis
operator associated with an orthonormal sequence {v1, v2, . . . , vs}, r ≤ s ≤
i=1 vi 6= ιk. To extend the sequence by one vector, we
need to find vs+1 with (vs+1, vs+1) = (vs+1, ιk) = 1 and with (vj , vs+1) = 0
for all 1 ≤ j ≤ s. Using block matrices this is summarized in the equation
k − 1 satisfying Ps
(3.1)
(cid:18) V
ι∗
k (cid:19) vs+1 =(cid:18) 0s
1 (cid:19) ,
where 0s is the zero vector in Zs
2.
the orthonormal sequence. This is all that is needed if s = k − 1.
In order to verify that this equation is consistent, we note that by the
orthonormality of the sequence {v1, v2, . . . , vs}, the vector ιk is a linear com-
i=1 vi = ιk. Thus, there exists vs+1 which extends
bination if and only if Ps
chosen so that Ps+1
there is one choice of vs+1 such that Ps+1
Next, we need to show that if s ≤ k − 2, then a solution vs+1 can be
i=1 vi 6= ιk, so that the iterative extension procedure can
be continued. The solution set of equation (3.1) forms an affine subspace of
2 having dimension k − (s + 1), thus contains 2k−s−1 elements. If s ≤ k − 2,
Zk
then there are at least two elements in this affine subspace. Consequently,
(cid:3)
We are ready to characterize the complementarity property for binary
Parseval frames. The condition that determines the existence of a Naimark
complement is whether at least one frame vector is even, that is, its entries
sum to zero.
i=1 vi 6= ιk.
Theorem 3.4. A binary (k, n)-frame F with n < k has a complementary
Parseval frame if and only if at least one frame vector is even.
Proof. We first rewrite the condition on the frame vectors in the equivalent
form ΘF ιn 6= ιk. Re-expressed in terms of the column vectors {ω1, ω2, . . . , ωn}
of ΘF , we claim that a complementary Parseval frame exists if and only if
Pn
i=1 ωn 6= ιk.
On the other hand, the existence of a complementary Parseval frame is
equivalent to the sequence of column vectors having an extension to an
orthonormal sequence of k elements.
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
7
The preceding lemma provides the existence of such an extension if and
(cid:3)
i=1 ωn 6= ιk, which finishes the proof.
only if Pn
3.2. A catalog of binary Parseval frames with the complementarity
property. A previous work contained a catalogue of binary Parseval frames
for Zn
2 when n was small [4]. Here, we wish to compile a list of the binary
Parseval frames that have a complementary Parseval frame. For notational
convenience, we consider ΘF instead of the sequence of frame vectors. By
Proposition 3.2, every such ΘF is obtained by a selection of columns from
a binary orthogonal matrix, so we could simply list the set of all orthogonal
matrices for small k. However, such a list quickly becomes extensive as
k increases. To reduce the number of orthogonal matrices, we note that
although the frame depends on the order in which the columns are selected
to form ΘF , the Gram matrix does not.
Identifying frames whose Gram
matrices coincide has already been used to avoid repeating information when
examining real or complex frames [1] and binary frames [4]. We consider an
even coarser underlying equivalence relation [8, 12, 3] that has also appeared
in the context of binary frames [4].
Definition 3.5. Two families F = {f1, f2, . . . , fk} and G = {g1, g2, . . . gk} in
Zn
2 are called switching equivalent if there is an orthogonal n × n matrix U
and a permutation π of the set {1, 2, . . . k} such that
fj = U gπ(j) for all j ∈ {1, 2, . . . } .
Representing the permutation π by the associated permutation matrix P
with entries Pi,j = δi,π(j) gives that if F and G are switching equivalent, then
ΘF = P ΘGU with an orthogonal n × n matrix U and a k × k permutation
matrix P . Alternatively, switching equivalence is stated in the form of an
identity for the corresponding Gram matrices.
Theorem 3.6 ([4]). Two binary (k, n)-frames F and G are switching equiv-
alent if and only if their Gram matrices are related by conjugation with a
k × k permutation matrix P ,
GF = P GG P ∗ .
We deduce a consequence for switching equivalence and Naimark comple-
ments, which is inferred from the role of the Gram matrices in the definition
of complementarity.
Corollary 3.7. If F and G are switching equivalent binary (k, n)-frames,
then F has a Naimark complement if and only if G does.
Thus, to provide an exhaustive list, we only need to ensure that at least
one representative of each switching equivalence class appears as a selection
of columns in the orthogonal matrices we include. To reduce the number of
representatives, we identify matrices up to row and column permutations.
8 Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
Definition 3.8. Two matrices A, B ∈ Mk,k(Z2) are called permutation equiv-
alent if there are two permutation matrices P1, P2 ∈ Mk,k(Z2) such that
A = P1BP ∗
2 .
Proposition 3.9. If U1 and U2 are permutation equivalent binary orthogo-
nal matrices, then each (k, n)-frame F formed by a sequence of n columns
of U1 is switching equivalent to a (k, n)-frame G formed with columns of U2.
Proof. Without loss of generality, we can assume that the analysis matrix
ΘF is formed by the first n columns of U1. By the equivalence of U1 and U2,
U1P2 = P1U2 with permutation matrices P1 and P2. The right multiplication
of U1 with P2 gives a column permutation, which identifies a sequence of
columns in P1U2 that is identical to the first n columns of U1. If G is obtained
with the corresponding columns in U2, then the Gram matrices of F and G
are related by GF = P1GGP ∗
1 , which proves the switching equivalence. (cid:3)
A list of permutation-inequivalent orthogonal k × k matrices allows us to
obtain the Gram matrix of each binary (k, n)-frame with a Naimark com-
plement by selecting an appropriate choice of n columns from an orthogonal
k × k matrix to form Θ and then by applying a permutation matrix P to
obtain GF = P ΘΘ∗P ∗.
Accordingly, each representative of an equivalence class of orthogonal ma-
trices can be chosen so that the columns are in lexicographical order. Table 1
contains a complete list of representatives of binary orthogonal matrices for
k ∈ {3, 4, 5, 6} from each permutation-equivalence class. Each column vec-
tor in our list is recorded by the integer obtained from the binary expansion
with the entries of the vector. For example, if a frame vector in Z4
2 is
f1 = (1, 0, 1, 1), then it is represented by the integer 20 + 22 + 23 = 13. Ac-
cordingly, in Z4
2, the standard basis is recorded as the sequence of numbers
1, 2, 4, 8.
4. Gram matrices of binary Parseval frames
The preceding section on complementarity hinged on the problem that
if G is the Gram matrix of a binary Parseval frame, then I − G may not,
although it is symmetric and idempotent. Again, there is a simple condi-
tion that needs to be added; Gram matrices of binary Parseval frames are
symmetric and idempotent and have at least one odd column, that is, a
column whose entries sum to one. Because of the identity G2 = G, having
an odd column is equivalent to having a non-zero diagonal entry. Indeed, it
has been shown that for any binary symmetric matrix G without vanishing
diagonal, there is a factor Θ such that G = ΘΘ∗ and the rank of Θ is equal
to that of G [15]. The assumptions needed for our proof are stronger, but
our algorithm for producing Θ appears to be more straightforward than the
factorization procedure for general symmetric binary matrices.
Theorem 4.1. A binary symmetric idempotent matrix M is the Gram ma-
trix of a Parseval frame if and only if it has at least one odd column.
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
9
8
4
4
4
Orthogonal k × k Matrices, 3 ≤ k ≤ 6
k Integer corresponding to binary column vectors
2
3 1
2
8
4 1
11 13 14
7
2
16
8
5 1
11 19 25 26
4
8
19 21 22
7
11 13 14 16
7
4
16 32
2
6 1
8
19 35 49 50
4
11 16 35 41 42
4
11 19 25 26 32
4
16 35 37 38
8
7
8
19 21 22 32
7
7
11 13 14 16 32
13 14 28 44 55 59
21 22 28 47 52 59
25 26 28 47 55 56
31 37 38 44 52 59
31 41 42 44 55 56
31 47 49 50 52 56
31 47 55 59 61 62
of permutation equivalence
Table 1. Representatives
classes of orthogonal matrices. Up to switching equivalence,
the Gram matrix of each binary (k, n)-frame with a Naimark
complement is obtained by selecting appropriate columns in
one of the listed k × k orthogonal matrices.
Proof. First, we re-express the condition on the columns of a symmetric
k × k matrix M in the equivalent form of the matrix Ik + M having at
least one even column or row. This, in turn, is equivalent to the inequality
(Ik + M )ιk 6= ιk.
Next, we recall that both M and Ik + M are assumed to be idempotent.
2 is in the range of an idempotent P if
We observe that any vector y ∈ Zk
and only if P y = y if and only if y is in the kernel of Ik + P .
Assuming M is the Gram matrix of a Parseval frame, then M = ΘΘ∗
where Θ has orthonormal columns and and (M + Ik)Θ = 0. Combining the
two properties gives
(cid:18) M + Ik
ι∗
k
(cid:19) Θ =(cid:18) 0k,n
ιn (cid:19) .
This is inconsistent if and only if ιk is in the span of the columns of the
idempotent M + Ik, which is equivalent to (M + Ik)ιk = ιk.
10Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
Conversely, assuming that M is symmetric and idempotent and that the
range of Ik + M does not contain ιk, we construct a matrix Θ with orthonor-
mal columns such that M = ΘΘ∗. The assumption on M is equivalent to
M ιk 6= 0, so at least one row or column of M is odd. Let this column be
ω1, then the fact that M is idempotent gives M ω1 = ω1.
j ΘΘ∗ωi for each i, j ∈ {1, 2, . . . , n} imply M = ΘΘ∗.
Next, we follow an inductive strategy similar to an earlier proof. What
we need is an orthonormal sequence {ω1, . . . , ωn} such that n is the rank of
M and M ωi = ωi for all i ∈ {1, 2, . . . , n}. In that case, the range of M is
the span of the sequence, and so is the range of M ∗. Thus, the identities
j M ωi = δi,j = ω∗
ω∗
Proceeding inductively, we need to extend a given orthonormal sequence
{ω1, . . . , ωs} in the kernel of Ik+M by one vector if s ≤ n−1 and if ιk is not in
the span of the columns of Ik + M combined with the orthonormal sequence.
Let V be a matrix formed by a maximal set of linearly independent rows
in Ik + M , then if M has rank n, rank nullity gives that V has k − n rows.
Letting Y be the analysis matrix of the orthonormal sequence {ω1, . . . , ωs},
then extending it by one vector requires solving the equation
(4.1)
V
Y
ι∗
k
ωs+1 =
0k−n
0s
1
.
step, we need to show that if s ≤ n − 2, then ι∗
rows of the matrix formed by V , Y and ω∗
by the fact that V Y ∗ = 0, so if ιk =Ps+1
Moreover, in order to guarantee the induction assumption at the next
k is not in the span of the
s+1. As before, this is obtained
i=1 ciωi + v with v being in the span
of the columns of V ∗, then Y v = 0 and orthonormality forces ci = 1 for
all i ∈ {1, 2, . . . , s + 1}. The solutions of the equation (4.1) form an affine
subspace of dimension k − (k − n) − s − 1 = n − s − 1, so if s ≤ n − 2,
then there are at least two solutions, one of which does not satisfy the sum
(cid:3)
identity ιk =Ps+1
i=1 ciωi + v.
5. Binary cyclic frames and circulant Gram matrices
Next, we examine a special type of frame whose Gram matrices are cir-
culants. We recall that a cyclic subspace V of Zk
2 has the property that it is
closed under cyclic shifts, that is, the cyclic shift S, which is characterized
by Sej = ej+1 (mod) k, leaves V invariant.
Definition 5.1. A frame F = {f1, f2, . . . , fk} for Zn
2 is called a binary cyclic
frame if the range of the analysis operator is invariant under the cyclic shift
S. If F is also Parseval, then we say that is a binary cyclic (k, n)-frame.
Since the range of the Gram matrix G belonging to a Parseval frame is
identical to the set of eigenvectors corresponding to eigenvalue one, we have
a simple characterization of Gram matrices of binary cyclic Parseval frames.
Theorem 5.2. A binary frame F = {f1, f2, . . . , fk} for Zn
2 is a cyclic Par-
seval frame if and only if its Gram matrix GF is a symmetric idempotent
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
11
circulant matrix, that is, GF = G∗
column vectors.
F = G2
F and SGF S∗ = GF , with only odd
F = G2
Proof. If GF is the Gram matrix of a binary cyclic Parseval frame, then from
the Parseval property, we know that GF = G∗
F . Moreover, by the
cyclicity of the frame, the eigenspace corresponding to eigenvalue one of GF
is invariant under S, and thus if x = GF x, then Sx = SGF x = GF Sx. Using
this identity repeatedly and writing y = Sk−1x = S∗x gives y = SGF S∗y
for all y in the range of GF . By the symmetry of GF , the range of GF
is identical to that of G∗
F , so hGF x, yi = hSGF S∗x, yi for all x, y in the
range of GF establishes the circulant property GF = SGF S∗.
If GF is
a circulant, then each column vector generates all the others by applying
powers of the cyclic shift to it. Thus, if one column vector is odd, so are
all the other column vectors. Applying Theorem 4.1 then yields that the
Gram matrices of binary cyclic Parseval frames are symmetric idempotent
circulant matrices with only odd column vectors.
Conversely, if G is a symmetric idempotent circulant and each column
vector is odd, then Theorem 4.1 again yields that it is the Gram matrix of
a binary Parseval frame F with G = ΘF Θ∗
F . Moreover, the range of G is
invariant under the cyclic shift, because one column vector generates all the
others by applying powers of the cyclic shift to it. Since the range of G is
identical to that of ΘF , F is a cyclic binary Parseval frame.
(cid:3)
Since adding the identity matrix changes odd columns of G to even
columns, we conclude that complementary Parseval frames do not exist for
binary cyclic Parseval frames.
Corollary 5.3. If F is a binary cyclic Parseval frame, then it has no com-
plementary Parseval frame.
In Table 2, we provide an exhaustive list of the Gram matrices of cyclic
binary Parseval frames with 3 ≤ k ≤ 20. Factoring these into the cor-
responding analysis and synthesis matrices shows that many of these ex-
amples contain repeated frame vectors. In an earlier paper, such repeated
vectors have been associated with a trivial form of redundancy incorporated
in the analysis matrix ΘF . Table 3 lists the circulant Gram matrices of rank
n < k ≤ 20 paired with k × n analysis matrices, for which no repetition of
frame vectors occurs.
References
[1] Balan, R., Equivalence relations and distances between Hilbert frames. Proceedings
AMS 127 (8) (1999), 2353-2366.
[2] Betten, A., Braun, M., Fripertinger, H., Kerber, A., Kohnert, A. and Wassermann,
A., Error-correcting linear codes, in: Algorithms and Computation in Mathematics
18, Springer-Verlag, Berlin, 2006, xl+798 pp.
[3] Bodmann, B. G. and Paulsen, V. I., Frames, graphs and erasures, Linear Algebra
Appl. 404 (2005), 118–146.
12Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
[4] Bodmann, B. G. Le, M., Reza, L., Tobin, M. and Tomforde, M., Frame theory for
binary vector spaces, Involve, 2 (2009), no. 5, 589-602.
[5] Bodmann, B. G., Camp, B. and Mahoney, D., Binary frames, graphs and erasures,
Involve 7, 151-169 (2014).
[6] Christensen, O., An Introduction to Frames and Riesz Bases, Birkhauser, Boston,
(2003).
[7] Duffin, R. J. and Schaeffer, A. C., A class of nonharmonic Fourier series, Trans.
Amer. Soc. 72 (1952), 341-366.
[8] Goyal, V. K., Kovacevi´c, J., and Kelner, J. A., Quantized frame expansions with
erasures, Appl. Comp. Harm. Anal., vol. 10, pp. 203–233, 2001.
[9] Hotovy, R., Larson, D. and Scholze, S., Binary frames, Houston Journal of Mathe-
matics, to appear.
[10] Haemers, W. H., Peeters, R., and van Rijckevorsel, J. M., Binary codes of strongly
regular graphs, Des. Codes Cryptogr. 17, 187–209 (1999).
[11] Han, D., Kornelson, K., Larson, D. and Weber, E., Frames for Undergraduates, Stu-
dent Mathematical Library, 40, American Mathematical Society, Providence, RI,
2007. xiv+295 pp.
[12] Holmes, R. B., Paulsen, V.I., Optimal frames for erasures. Linear Algebra Appl. 377
(2004), 31–51. DOI 10.1016/j.laa.2003.07.012.
[13] Kovacevi´c, J. and Chebira, A., Life beyond bases: The advent of frames (Part I),
IEEE Signal Processing Magazine 24 (2007), no. 4, 86-104.
[14] Kovacevi´c, J. and Chebira, A., Life beyond bases: The advent of frames (Part II),
IEEE Signal Processing Magazine 24 (2007), no. 5, 15-125.
[15] Lempel, A., Matrix factorization over GF (2) and trace-orthogonal bases of GF (2n),
SIAM J. Comput. 4 (1975), 175–186.
[16] Puschel, M., Kovacevi´c, J., Real, tight frames with maximal robustness to erasures.
In: Data Compression Conference, 2005. Proceedings. DCC 2005, pp. 63–72 (2005).
[17] Rath, G. and Guillemot, C., Performance analysis and recursive syndrome decoding
of DFT codes for bursty erasure recovery, IEEE Trans. on Signal Processing, vol. 51
(2003), no. 5, 1335-1350.
[18] Rath, G. and Guillemot, C., Frame-theoretic analysis of DFT codes with erasures,
IEEE Transactions on Signal Processing, vol. 52 (2004), no. 2, 447-460.
[19] MacWilliams, F. J. and Sloane, N. J., The Theory of Error-Correcting Codes, North
Holland, Amsterdam, 1977.
[20] Marshall, T., J., Coding of real-number sequences for error correction: A digital signal
processing problem. IEEE Journal on Selected Areas in Communications 2 (1984),
no. 2, 381–392.
[21] Marshall, T., Fourier transform convolutional error-correcting codes. In: Signals,
Systems and Computers, 1989. Twenty-Third Asilomar Conference on, vol. 2, pp.
658–662 (1989).
Department of Mathematics, University of Houston, Houston, TX 77204-
3008, USA
E-mail address: [email protected]
E-mail address: [email protected]
E-mail address: [email protected]
E-mail address: [email protected]
E-mail address: [email protected]
WHAT IS ODD ABOUT BINARY PARSEVAL FRAMES?
13
k
3
4
5
6
7
8
9
First row of circulant k × k Gram matrix, 3 ≤ k ≤ 20
1 0 0
1 1 1
1 0 0 0
1 0 0 0 0
1 1 1 1 1
1 0 0 0 0 0
1 0 1 0 1 0
1 0 0 0 0 0 0
1 1 1 1 1 1 1
1 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0
1 0 0 1 0 0 1 0 0
1 1 1 0 1 1 0 1 1
1 1 1 1 1 1 1 1 1
10 1 0 0 0 0 0 0 0 0 0
1 0 1 0 1 0 1 0 1 0
11 1 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 1 1 1
12 1 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 1 0 0 0 1 0 0 0
13 1 0 0 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 1 1 1 1 1
14 1 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 1 0 1 0 1 0 1 0 1 0 1 0
15 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 0 1 0 0 0 0 1 0 0 0 0
1 0 0 1 0 0 1 0 0 1 0 0 1 0 0
1 0 0 1 0 1 1 0 0 1 1 0 1 0 0
1 1 1 0 1 0 0 1 1 0 0 1 0 1 1
1 1 1 0 1 1 0 1 1 0 1 1 0 1 1
1 1 1 1 1 0 1 1 1 1 0 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
16 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
17 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 0 1 0 1 1 1 0 0 1 1 1 0 1 0 0
1 1 1 0 1 0 0 0 1 1 0 0 0 1 0 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
18 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0
1 0 1 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0
1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0
19 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
20 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0
Table 2. Circulant Gram matrices of binary cyclic (k, n)-
frames, first row shown.
14Z. J. BAKER, B. G. BODMANN, M. G. BULLOCK, S. N. BRANUM, AND J. MCLANEY
k = 9
Circulant Gram matrix
1 1 1 0 1 1 0 1 1
1 1 1 1 0 1 1 0 1
1 1 1 1 1 0 1 1 0
0 1 1 1 1 1 0 1 1
1 0 1 1 1 1 1 0 1
1 1 0 1 1 1 1 1 0
0 1 1 0 1 1 1 1 1
1 0 1 1 0 1 1 1 1
1 1 0 1 1 0 1 1 1
1 1 1 0 1 0 0 1 1 0 0 1 0 1 1
1 1 1 1 0 1 0 0 1 1 0 0 1 0 1
1 1 1 1 1 0 1 0 0 1 1 0 0 1 0
0 1 1 1 1 1 0 1 0 0 1 1 0 0 1
1 0 1 1 1 1 1 0 1 0 0 1 1 0 0
0 1 0 1 1 1 1 1 0 1 0 0 1 1 0
0 0 1 0 1 1 1 1 1 0 1 0 0 1 1
1 0 0 1 0 1 1 1 1 1 0 1 0 0 1
1 1 0 0 1 0 1 1 1 1 1 0 1 0 0
0 1 1 0 0 1 0 1 1 1 1 1 0 1 0
0 0 1 1 0 0 1 0 1 1 1 1 1 0 1
1 0 0 1 1 0 0 1 0 1 1 1 1 1 0
0 1 0 0 1 1 0 0 1 0 1 1 1 1 1
1 0 1 0 0 1 1 0 0 1 0 1 1 1 1
1 1 0 1 0 0 1 1 0 0 1 0 1 1 1
1 0 0 1 0 1 1 0 0 1 1 0 1 0 0
0 1 0 0 1 0 1 1 0 0 1 1 0 1 0
0 0 1 0 0 1 0 1 1 0 0 1 1 0 1
1 0 0 1 0 0 1 0 1 1 0 0 1 1 0
0 1 0 0 1 0 0 1 0 1 1 0 0 1 1
1 0 1 0 0 1 0 0 1 0 1 1 0 0 1
1 1 0 1 0 0 1 0 0 1 0 1 1 0 0
0 1 1 0 1 0 0 1 0 0 1 0 1 1 0
0 0 1 1 0 1 0 0 1 0 0 1 0 1 1
1 0 0 1 1 0 1 0 0 1 0 0 1 0 1
1 1 0 0 1 1 0 1 0 0 1 0 0 1 0
0 1 1 0 0 1 1 0 1 0 0 1 0 0 1
1 0 1 1 0 0 1 1 0 1 0 0 1 0 0
0 1 0 1 1 0 0 1 1 0 1 0 0 1 0
0 0 1 0 1 1 0 0 1 1 0 1 0 0 1
1 1 1 1 1 0 1 1 1 1 0 1 1 1 1
1 1 1 1 1 1 0 1 1 1 1 0 1 1 1
1 1 1 1 1 1 1 0 1 1 1 1 0 1 1
1 1 1 1 1 1 1 1 0 1 1 1 1 0 1
1 1 1 1 1 1 1 1 1 0 1 1 1 1 0
0 1 1 1 1 1 1 1 1 1 0 1 1 1 1
1 0 1 1 1 1 1 1 1 1 1 0 1 1 1
1 1 0 1 1 1 1 1 1 1 1 1 0 1 1
1 1 1 0 1 1 1 1 1 1 1 1 1 0 1
1 1 1 1 0 1 1 1 1 1 1 1 1 1 0
0 1 1 1 1 0 1 1 1 1 1 1 1 1 1
1 0 1 1 1 1 0 1 1 1 1 1 1 1 1
1 1 0 1 1 1 1 0 1 1 1 1 1 1 1
1 1 1 0 1 1 1 1 0 1 1 1 1 1 1
1 1 1 1 0 1 1 1 1 0 1 1 1 1 1
1 1 1 0 1 1 0 1 1 0 1 1 0 1 1
1 1 1 1 0 1 1 0 1 1 0 1 1 0 1
1 1 1 1 1 0 1 1 0 1 1 0 1 1 0
0 1 1 1 1 1 0 1 1 0 1 1 0 1 1
1 0 1 1 1 1 1 0 1 1 0 1 1 0 1
1 1 0 1 1 1 1 1 0 1 1 0 1 1 0
0 1 1 0 1 1 1 1 1 0 1 1 0 1 1
1 0 1 1 0 1 1 1 1 1 0 1 1 0 1
1 1 0 1 1 0 1 1 1 1 1 0 1 1 0
0 1 1 0 1 1 0 1 1 1 1 1 0 1 1
1 0 1 1 0 1 1 0 1 1 1 1 1 0 1
1 1 0 1 1 0 1 1 0 1 1 1 1 1 0
0 1 1 0 1 1 0 1 1 0 1 1 1 1 1
1 0 1 1 0 1 1 0 1 1 0 1 1 1 1
1 1 0 1 1 0 1 1 0 1 1 0 1 1 1
Corresponding
analysis matrix
1 0 0 1 1 1 1
1 1 1 0 0 1 1
1 1 1 1 1 0 0
0 1 0 1 1 1 1
0 0 0 1 0 1 1
0 0 0 0 0 1 0
0 0 1 1 1 1 1
0 0 0 0 1 1 1
0 0 0 0 0 0 1
1 0 1 0 1 1 1
1 1 0 1 1 0 1
0 1 0 1 1 0 0
1 0 1 1 0 1 1
1 1 0 1 1 1 0
1 1 0 0 1 0 0
0 1 1 0 0 0 1
1 0 0 0 0 1 1
1 1 0 1 0 0 0
0 1 1 0 0 1 0
0 0 0 1 0 1 1
0 0 0 0 0 1 0
0 0 1 1 1 1 1
0 0 0 0 1 1 1
0 0 0 0 0 0 1
n = 7
n = 9
1 0 1 0 0 0 1 0 0
0 1 0 0 1 0 0 1 0
1 0 0 0 0 0 1 0 1
1 1 1 0 0 0 1 1 0
1 1 1 1 1 0 0 1 1
0 1 1 0 1 1 0 0 1
1 0 1 1 1 0 1 0 0
0 1 1 1 1 1 1 1 0
0 0 0 1 1 1 0 1 1
0 0 0 0 0 1 1 0 1
0 0 1 1 0 0 0 1 0
0 0 0 0 1 1 0 0 1
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1
1 0 0 1 1 1 1 1 1 1 1
1 1 1 0 0 1 1 1 1 1 1
1 1 1 1 1 0 0 1 1 1 1
1 1 1 1 1 1 1 0 0 1 1
1 1 1 1 1 1 1 1 1 0 0
0 1 0 1 1 1 1 1 1 1 1
0 0 0 1 0 1 1 1 1 1 1
0 0 0 0 0 1 0 1 1 1 1
0 0 0 0 0 0 0 1 0 1 1
0 0 0 0 0 0 0 0 0 1 0
0 0 1 1 1 1 1 1 1 1 1
0 0 0 0 1 1 1 1 1 1 1
0 0 0 0 0 0 1 1 1 1 1
0 0 0 0 0 0 0 0 1 1 1
0 0 0 0 0 0 0 0 0 0 1
1 0 0 1 1 1 1 0 0 1 1 1 1
1 1 1 0 0 1 1 1 1 0 0 1 1
1 1 1 1 1 0 0 1 1 1 1 0 0
0 1 0 1 1 1 1 0 0 1 1 1 1
0 0 0 1 0 1 1 1 1 0 0 1 1
0 0 0 0 0 1 0 1 1 1 1 0 0
0 0 1 1 1 1 1 0 0 1 1 1 1
0 0 0 0 1 1 1 1 1 0 0 1 1
0 0 0 0 0 0 1 1 1 1 1 0 0
0 0 0 0 0 0 0 1 0 1 1 1 1
0 0 0 0 0 0 0 0 0 1 0 1 1
0 0 0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 1 1 1
0 0 0 0 0 0 0 0 0 0 0 0 1
n = 11
n = 13
Table 3. Gram and analysis matrices of binary cyclic (k, n)-
frames, k > n, whose vectors do not repeat.
k = 15
|
1906.03021 | 1 | 1906 | 2019-06-07T11:33:15 | Convergence in variation for the multidimensional generalized sampling series and applications to smoothing for digital image processing | [
"math.FA"
] | In this paper we study the problem of the convergence in variation for the generalized sampling series based upon averaged-type kernels in the multidimensional setting. As a crucial tool, we introduce a family of operators of sampling-Kantorovich type for which we prove convergence in L^p on a subspace of L^p(R^N): therefore we obtain the convergence in variation for the multidimensional generalized sampling series by means of a relation between the partial derivatives of such operators acting on an absolutely continuous function f and the sampling-Kantorovich type operators acting on the partial derivatives of f. Applications to digital image processing are also furnished. | math.FA | math |
Convergence in variation for the
multidimensional generalized sampling series
and applications to smoothing
for digital image processing
Laura Angeloni, Danilo Costarelli, Gianluca Vinti
Department of Mathematics and Computer Science
University of Perugia, Via Vanvitelli 1, 06123, Perugia, Italy
[email protected]
-
[email protected]
[email protected]
Abstract
In this paper we study the problem of the convergence in variation
for the generalized sampling series based upon averaged-type kernels in
the multidimensional setting. As a crucial tool, we introduce a family of
operators of sampling-Kantorovich type for which we prove convergence
in Lp on a subspace of Lp(RN ): therefore we obtain the convergence in
variation for the multidimensional generalized sampling series by means
of a relation between the partial derivatives of such operators acting on
an absolutely continuous function f and the sampling-Kantorovich type
operators acting on the partial derivatives of f. Applications to digital
image processing are also furnished.
AMS 2010 Mathematics Subject Classification: 41A30, 41A05
Key words and phrases: convergence in variation; multidimensional general-
ized sampling series; sampling-Kantorovich operators; variation diminishing type
property; smoothing in digital image processing
1
1
Introduction
In this paper we present approximation results in BV-spaces for the generalized
sampling series in the multidimensional frame.
The above sampling series, defined as
(I)
(Swf )(t) :=
f
χ(wt − k),
t ∈ R, w > 0,
(cid:18) k
(cid:19)
w
(cid:88)
k∈Z
has been introduced by P.L. Butzer (see, e.g., [24]) and revealed to be very
interesting for both the theoretical and applicative aspects.
Indeed, several
approximation results have been obtained in the last forty years with respect
to different notions of convergence, such as uniform, Lp, and modular conver-
gence ([18, 37, 16]). Moreover, they have many important applications to Signal
Theory since they provide an approximate sampling formula, which allows to
reconstruct not necessarily band-limited signals. By its multivariate generaliza-
tion, that is,
(II)
(Swf )(t) :=
f
χ(wt − k), t ∈ RN , w > 0,
(cid:18) k
(cid:19)
w
(cid:88)
k∈ZN
several problems of image processing can be treated (see, e.g., [17]).
A natural setting to study some issues involving digital images is furnished
by the spaces of functions of bounded variation ([36, 44, 47]). In this direction,
it is interesting to have results about estimates and convergence in variation
for the above discrete operators. In the case of functions of one-variable, the
problem has been faced in [6] for the operators (I) based upon averaged type
kernels.
Here, also with the aim to consider applicative aspects, we deal with the
multivariate case, namely we study results about approximation in variation
by means of the operators (II) based upon multidimensional product kernels of
averaged type, namely
where
¯χi,m(t) :=
1
m
¯χm(t) :=
¯χi,m(ti)
χi(t + v) dv,
for some m ∈ N, and χi : R −→ R are suitable one-dimensional kernels. We
will use the multidimensional generalization of the classical Jordan variation
N(cid:89)
(cid:90) m
i=1
2
− m
2
2
introduced by Tonelli (see, e.g., [45]) and later generalized by T. Radó ([40])
and C. Vinti ([48]) to the case of functions of N-variables.
In particular, we establish a variation diminishing-type property (Proposi-
tion 1) and a convergence theorem (Theorem 1).
Due to form of the above discrete sampling type operators, to obtain results
about approximation in variation is a very delicate problem. The strategy we
propose is to introduce a new family of Kantorovich type operators defined as
w
(cid:88)
k∈ZN
(cid:90) kj +1
w
kj
w
(cid:18) k1
w
f
(cid:19)
χ(wt − k),
(III)
(Kw,jf )(t) :=
, . . . , u . . . ,
kN
w
du
t ∈ RN, w > 0, j = 1, . . . , N, for which we establish Lp-convergence in a
subspace of Lp(RN ) (Theorem 4), therefore obtaining a convergence result for
the new class of operators (III). We notice that the above result is obtained
through the use of the τ−modulus of continuity (see [34, 33, 42]) which seems
to be the most suitable approach in this setting.
The crucial point in order to reach the convergence in variation is to find a
link between the two classes of operators: to this aim we prove (Proposition 2)
a relation between the gradient of the multivariate generalized sampling series
of a function f and the family of Kantorovich-type discrete operators acting on
the partial derivatives of f.
Some of the results of the present paper may have an applicative interpre-
tation: in particular, the variation diminishing type property (Proposition 1),
ensuring that the generalized sampling series (II) have smaller variation than the
function on which they act, can be viewed as a smoothing procedure. Indeed,
if the function f is a gray scale digital image, the action of the operators and
the consequent variation diminishing corresponds to reducing the jumps of gray
levels with respect to the original image, hence producing a smoothing effect.
In Section 6 we discuss such digital image processing applications in details,
furnishing some numerical examples.
2 Notations and preliminaries
Our results will be set in the frame of the space of multivariate functions of
bounded variation: in particular we will consider the concept of variation intro-
duced by Tonelli ([45]) for two variables, extended to the general case of RN by
3
Radó and C. Vinti ([40, 48]). Here we will recall it.
For a function f : RN → R and x = (x1, . . . , xN ) ∈ RN, we will use the
notation
j = (x1, . . . , xj−1, xj+1, . . . , xN ) ∈ RN−1, x = (x
(cid:48)
(cid:48)
(cid:48)
j, xj),
j, xj), f (x) = f (x
x
an N−dimensional interval I =(cid:81)N
if we are interested in the j−th coordinate of x, j = 1, . . . , N. Moreover, given
j] we will denote the
(N − 1)−dimensional interval obtained deleting by I the j−th coordinate, i.e.,
i=1[ai, bi], by I(cid:48)
j = [a(cid:48)
j, b(cid:48)
I = [a
(cid:48)
j, b
j] × [aj, bj],
(cid:48)
j = 1, . . . , N.
α =(cid:0) x1
Given a vector x ∈ RN and α ∈ R, we will use the usual notation for products
and quotients, i.e., αx = (αx1, . . . , αxN ) and, for α (cid:54)= 0, x
By Rloc(RN ) we will denote the space of locally Riemann integrable functions
on RN, while M (RN ) will denote the space of all the measurable and bounded
functions f : RN → R.
Definition 1 A function f ∈ M (RN ) is said to be of bounded variation (f ∈
j,·)] (the usual Jordan one-dimensional variation of the j−th
BV (RN )) if VR[f (x(cid:48)
section of f) is finite for a.e. x(cid:48)
α , . . . , xN
α
(cid:1).
(cid:90)
j ∈ RN−1 and
j < +∞,
j,·)] dx
(cid:48)
(cid:48)
VR[f (x
for every j = 1, . . . , N.
RN−1
For more details about BV -spaces, see, e.g., [9, 5, 10, 11, 4, 13].
The first step is to consider, for I = (cid:81)N
Let us now recall how to compute the multidimensional Tonelli variation.
i=1[ai, bi] and j = 1, . . . , N the (N −
1)−dimensional integrals
(cid:90)
Φj(f, I) :=
[a(cid:48)
j ,b(cid:48)
j ]
V[aj ,bj ][f (x
j,·)]dx
(cid:48)
(cid:48)
j,
j,·)] is the usual one-dimensional (Jordan) variation of the
where V[aj ,bj ][f (x(cid:48)
j − th section of f. Let now Φ(f, I) be the Euclidean norm of the vector
(Φ1(f, I), . . . , ΦN (f, I)), namely
Φ(f, I) :=
Φ2
j (f, I)
1
2
,
N(cid:88)
j=1
4
where Φ(f, I) = +∞ if Φj(f, I) = +∞ for some j = 1, . . . , N.
Then the variation of f on I ⊂ RN is defined as
m(cid:88)
VI [f ] := sup
Φ(f, Jk),
k=1
where the supremum is taken over all the finite families of N−dimensional in-
tervals {J1, . . . , Jm} which form partitions of I.
Passing to the supremum over all the intervals I ⊂ RN, we obtain the
variation of f over the whole RN, i.e.,
V [f ] := sup
I⊂RN
VI [f ].
It is well known that, for every f ∈ BV (RN ), ∇f exists a.e. in RN and ∂f
L1(RN ), for every j = 1, . . . , N (see e.g. [40, 48]).
∂xj
∈
We now recall the notion of absolute continuity in the sense of Tonelli.
sense of Tonelli (f ∈ ACloc(RN )) if, for every interval I =(cid:81)N
Definition 2 A function f : RN → R is locally absolutely continuous in the
i=1[ai, bi] and for
j,·) : [aj, bj] → R is absolutely
every j = 1, 2, . . . , N, the j−th section of f, f (x(cid:48)
continuous for almost every x(cid:48)
j ∈ [a(cid:48)
j, b(cid:48)
j].
It is a well known result that, if f ∈ BV (RN ) ∩ ACloc(RN ), then
(cid:90)
V [f ] =
RN
∇f (x) dx
(see [40, 48, 35, 12]), that is, an integral representation for the variation of f
holds.
We will therefore denote by AC(RN ) the space of all the functions f ∈
BV (RN ) ∩ ACloc(RN ) (absolutely continuous functions).
We will study approximation results for the multivariate generalized sampling
series, namely a family of discrete operators defined as
(Swf )(t) :=
f
χ(wt − k), t ∈ RN , w > 0 :
(cid:18) k
(cid:19)
w
(cid:88)
k∈ZN
such operators are the multidimensional version of the generalized sampling
series (see, e.g., [21, 22, 17]).
Here χ is a kernel, that is, a function χ : RN → R that satisfies the following
assumptions:
5
(χ1) χ ∈ L1(RN ) is such that(cid:80)
k∈ZN χ(u − k) = 1, for every u ∈ RN;
(cid:80)
k∈ZN χ(u − k) < +∞, where the convergence of the
(χ2) Aχ := supu∈RN
series is uniform on the compact sets of RN.
The above assumptions are quite standard working with discrete families of
operators: see e.g., [21]. We point out that the operators (Swf )w>0 are well-
defined, for example, for every f ∈ BV (RN ): indeed in this case f is bounded
and since f (t) ≤ M, for some M > 0 and for every t ∈ RN, we have
(Swf )(t) ≤ M
by (χ2).
χ(wt − k) ≤ M Aχ,
In particular, in the present paper we will study the convergence in variation
for the multivariate generalized sampling series with product kernels of averaged
type, that is kernels of the form
where
¯χi,m(t) :=
1
m
¯χm(t) :=
¯χi,m(ti)
(1)
χi(t + v) dv,
for some m ∈ N, and χi : R −→ R is a (one-dimensional) kernel for every
i = 1, . . . , N (i.e., satisfying (χ1) and (χ2) with N = 1).
Notice that it is easy to see that ¯χm is a kernel itself and moreover, by the
Fubini-Tonelli theorem,
(cid:88)
k∈ZN
N(cid:89)
(cid:90) m
i=1
2
− m
2
(cid:107) ¯χi,m(cid:107)L1(R) =
χi(t + v) dv
χi(t + v) dt dv = (cid:107)χi(cid:107)L1(R)
(2)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dt
(cid:90)
R
≤ 1
m
2
2
m
− m
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:90) m
(cid:90)
(cid:90) m
(cid:107) ¯χm(cid:107)1 ≤ N(cid:89)
− m
R
2
2
(cid:107)χi(cid:107)L1(R).
and, by this,
We point out that is it easy to give examples of product kernels of averaged
type (see Section 6) to which our results can be applied.
i=1
6
The corresponding multivariate generalized sampling series (associated to
the averaged product kernel ¯χm) will be denoted as
( ¯Sm
w f )(t) :=
¯χm(wt − k), t ∈ RN , w > 0.
It is easy to see that ¯χm is differentiable and, taking into account that,
(cid:18) k
(cid:19)
w
f
(cid:88)
k∈ZN
(cid:104)
(cid:16)
obviously,
∂ ¯χm
∂tj
(t) =
1
m
(cid:17)(cid:105)
(cid:16)
(cid:17) − χj
(cid:16)
(cid:104)
¯χi,m(ti)
χj
tj +
m
2
tj − m
2
, t ∈ RN ,
it is possible to write, for every j = 1, . . . , N, t ∈ RN, w > 0,
∂ ¯Sm
wtj − kj +
w f
∂tj
¯χi,m(wti−ki)
(cid:88)
(t) =
w
m
χj
w
f
(cid:19)(cid:89)
m
2
k∈ZN
i(cid:54)=j
(cid:12)(cid:12)(cid:12)(cid:12) ∂ ¯Sm
w f
∂tj
(cid:12)(cid:12)(cid:12)(cid:12) ≤ w
m
(t)
M
Notice that, since f is bounded, ∂ ¯Sm
w f
∂tj
(t) exists for every t ∈ RN since
(cid:89)
¯χi,m(wti − ki)(cid:104)(cid:12)(cid:12)(cid:12)χj
(cid:16)
wtj − kj +
m
2
(cid:89)
(cid:18) k
i(cid:54)=j
(cid:88)
N(cid:89)
i=1
k∈ZN
i(cid:54)=j
≤ 2w
m
M
Aχi,
(cid:17) − χj
(cid:12)(cid:12)(cid:12)χj
(cid:17)(cid:12)(cid:12)(cid:12) +
(cid:16)
(cid:17)(cid:105)
.
wtj − kj − m
2
(3)
(cid:16)
wtj − kj − m
2
(cid:17)(cid:12)(cid:12)(cid:12)(cid:105)
(4)
again by (χ2) for each one-dimensional kernel χi.
One of the main ideas of this paper is to establish a relation between the
gradient of the multivariate generalized sampling series of a function f and a
family of Kantorovich-type discrete operators, that we now introduce, acting
on the partial derivatives of f. This generalizes the analogous result, in the
one-dimensional case, obtained in [6], similarly to what happens between the
Bernstein polynomials and their Kantorovich version (see, e.g., [1]).
We therefore introduce the following family of multidimensional discrete op-
erators of sampling-Kantorovich type:
w
(cid:88)
k∈ZN
(cid:90) kj +1
w
kj
w
(cid:18) k1
w
f
(cid:19)
χ(wt−k), t ∈ RN , w > 0,
, . . . , u . . . ,
kN
w
du
(Kw,jf )(t) :=
j = 1, . . . , N. Notice that, similarly to the case of (Swf )w>0, it is easy to see
that the operators (Kw,jf )w>0, j = 1, . . . , N, are well-defined if, for example
f ∈ BV (RN ), since f is in particular bounded.
7
3 Results for the multidimensional generalized
sampling series
w f )w>0 map BV (RN ) into AC(RN ).
We will first prove that the operators ( ¯Sm
Such result is a kind of "asymptotic" variation diminishing property:
indeed,
choosing m ∈ N sufficiently large, we obtain the classical variation diminishing
property for the multidimensional generalized sampling series.
Proposition 1 Let f ∈ BV (RN ). Then ¯Sm
m ∈ N, and
w f ∈ AC(RN ), for every w > 0,
V [ ¯Sm
w f ] ≤ 1
m
(cid:107)χi(cid:107)1V [f ].
(5)
N(cid:89)
i=1
Proof. Let us fix w > 0 and m ∈ N. By (3) and (4), the partial derivatives of
w f exist for every t ∈ RN, j = 1 . . . , N, and are bounded: this implies that
¯Sm
the sections of the function f are locally absolutely continuous, namely f ∈
[ai, bi] ⊂ RN and let {J1, . . . , Jp} be a partition
ACloc(RN ). Let us fix I =
N(cid:89)
of I, with Jk =
i=1
[(k)aj,(k) bj], k = 1, . . . p. For every fixed j = 1, . . . N,
N(cid:89)
j=1
k = 1, . . . p, similarly to the proof of Proposition 1 of [6], it is possible to show
that, since ( ¯Sm
j,·) ∈ ACloc(R), a.e. t(cid:48)
w f )(t(cid:48)
j ∈ [(k)a(cid:48)
j,(k) b(cid:48)
j],
j,·)]
(cid:48)
V[(k)aj ,(k)bj ][f (t
(cid:89)
i(cid:54)=j
(cid:107) ¯χi,m(cid:107)1(cid:107)χj(cid:107)1
V[(k)aj ,(k)bj ][( ¯Sm
w f )(t
j,·)] ≤ 1
(cid:48)
m
and therefore, by (2),
Φj( ¯Sm
w f, Jk) =
(cid:90)
V[(k)aj ,(k)bj ][( ¯Sm
w f )(t
[(k)a(cid:48)
j ,(k)b(cid:48)
j ]
(cid:90)
(cid:48)
j
j,·)] dt
(cid:48)
N(cid:89)
j,·)] dt
(cid:48)
(cid:48)
j
(cid:107)χi(cid:107)1
≤ 1
m
=
1
m
V[(k)aj ,(k)bj ][f (t
[(k)a(cid:48)
j ,(k)b(cid:48)
j ]
i=1
(cid:107)χi(cid:107)1.
Φj(f, Jk)
N(cid:89)
i=1
8
This implies that, for every k = 1, . . . , p,
N(cid:88)
N(cid:88)
1
m
j=1
j=1
Φ( ¯Sm
w f, Jk) =
=
[Φj( ¯Sm
w f, Jk)]2
1
m2 [Φj(f, Jk)]2
N(cid:88)
j=1
2
≤
1
1
2 N(cid:89)
i=1
N(cid:89)
i=1
[Φj(f, Jk)]2
(cid:107)χi(cid:107)1 =
1
m
(cid:107)χi(cid:107)1Φ(f, Jk).
N(cid:89)
i=1
(cid:107)χi(cid:107)2
1
1
2
Summing now over k = 1, . . . , p and passing to the supremum over all the
possible partitions {J1, . . . , Jp} of I, we conclude that
and hence, by the arbitrariness of I ⊂ RN,
N(cid:89)
i=1
N(cid:89)
i=1
(cid:107)χi(cid:107)1VI [f ]
(cid:107)χi(cid:107)1V [f ].
VI [ ¯Sm
w f ] ≤ 1
m
V [ ¯Sm
w f ] ≤ 1
m
This implies that ¯Sm
w f ∈ BV (RN ) and hence ¯Sm
w f ∈ AC(RN ).
We will now give the result that establishes a relation between the partial
derivatives of the multidimensional sampling series and the multidimensional
sampling-Kantorovich type operators acting on the partial derivatives of the
function.
Proposition 2 If f ∈ AC(RN ), then for every t ∈ RN, j = 1, . . . , N,
(cid:18)
m(cid:88)
i=1
(cid:19)(cid:18)
j, tj − m − 2(i − 1)
(cid:48)
t
2w
(cid:19)
,
∂ ¯Sm
w f
∂tj
w > 0, m ∈ N.
(t) =
1
m
Kw,j
∂f
∂tj
9
Proof. By (3), since f is locally absolutely continuous, we have that
∂ ¯Sm
w f
∂tj
(t) =
w
m
¯χi,m(wti − ki)
wtj − kj +
m
2
(cid:104)
(cid:16)
(cid:18) k(cid:48)
χj
j
w
(cid:19)
(cid:17) − χj
(cid:16)
(cid:17)(cid:105)
wtj − kj − m
2
(cid:104)
(cid:16)
, u
du + f
, 0
¯χi,m(wti − ki)
χj
wtj − kj +
m
2
i(cid:54)=j
(cid:19)(cid:89)
(cid:19)(cid:89)
(cid:19)(cid:89)
i(cid:54)=j
=
w
m
− χj
w
∂f
∂tj
k∈ZN
wtj − kj − m
2
0
k∈ZN
(cid:88)
(cid:88)
(cid:16)
(cid:88)
(cid:88)
j
w
i(cid:54)=j
(cid:19)(cid:89)
(cid:18) k(cid:48)
(cid:17)(cid:105)
(cid:18) k(cid:48)
(cid:18) k(cid:48)
∂f
∂tj
f
w
(cid:18) k
(cid:90) kj
(cid:90) kj
(cid:90) kj
0
w
w
(cid:17)
(cid:17)
(cid:17)
.
wtj − kj +
m
2
wtj − kj − m
2
(cid:17)
(cid:16)
wtj −(cid:101)kj − m
(cid:17)
2
wtj − kj − m
2
(cid:17)
=
, 0
, u
j
w
j
w
w
m
k∈ZN
du + f
(cid:19)
(cid:19)
(cid:18) k(cid:48)
(cid:18) k(cid:48)
− w
m
¯χi,m(wti − ki)χj
(cid:16)
(cid:16)
Now putting in the first series(cid:101)kj = kj − m and(cid:101)ki = ki for every i (cid:54)= j, there
¯χi,m(wti −(cid:101)ki)χj
(cid:16)
¯χi,m(wti − ki)χj
∂ ¯Sm
w f
∂tj
∂f
∂tj
(cid:33)
holds
du + f
(t) =
k∈ZN
w
m
i(cid:54)=j
i(cid:54)=j
j
w
j
w
j
w
, u
, u
, 0
, 0
0
w
w
¯χi,m(wti − ki)χj
0
0
k∈ZN
, 0
, u
j
w
j
w
j
w
du + f
du + f
∂f
∂tj
∂f
∂tj
(cid:88)
(cid:101)k∈ZN
(cid:88)
(cid:88)
(cid:88)
(cid:26)(cid:18)
m(cid:88)
(cid:90) (cid:101)kj +m
(cid:90) kj
(cid:90) kj +m
m(cid:88)
(cid:18)
(cid:33)(cid:89)
(cid:32)(cid:101)k(cid:48)
(cid:19)(cid:89)
(cid:18) k(cid:48)
(cid:89)
(cid:19)
(cid:32)(cid:101)k(cid:48)
(cid:19)
(cid:18) k(cid:48)
(cid:19)
(cid:18) k(cid:48)
∂f
(cid:18) k(cid:48)
(cid:90) kj +i
(cid:89)
(cid:19)(cid:16)
(cid:18)
(cid:17)
∂tj
j, tj − m
(cid:19)(cid:18)
(cid:48)
2w
j, tj − m − 2(i − 1)
¯χi,m(wti − ki)χj
+ . . . +
∂f
∂tj
∂f
∂tj
∂f
∂tj
(cid:19)
kj +i−1
k∈ZN
k∈ZN
Kw,j
Kw,j
Kw,j
i(cid:54)=j
i(cid:54)=j
i(cid:54)=j
j
w
j
w
(cid:48)
t
du
du
, u
, u
i=1
kj
w
t
.
w
w
w
2w
∂f
∂tj
i=1
− w
m
=
=
=
=
w
m
w
m
1
m
1
m
(cid:16)
wtj − kj − m
2
(cid:16)
¯χi,m(wti − ki)χj
(cid:19)(cid:18)
wtj − kj − m
2
j, tj − m − 2(m − 1)
(cid:48)
t
2w
(cid:17)
(cid:19)(cid:27)
10
4 Convergence in Lp for the multidimensional sampling-
Kantorovich type operators
We now study the problem of the convergence in Lp for the multidimensional
sampling-Kantorovich type operators that we introduced. Such result will be
also fundamental in order to prove the convergence in variation for ( ¯Swf )w>0.
We recall that convergence in Lp holds for the multidimensional general-
ized sampling series (Swf )w>0 (see [17]) assuming that f belongs to a suit-
able subspace of Lp(RN ), namely Λp. Since the definition of our operators
(Kw,jf )w>0 is very close to that one of (Swf )w>0, it is natural to expect that
the Lp−convergence holds within the same subspace, that is what we will now
prove.
Let us recall the definition of Λp (see [17]).
Let us consider an admissible partition over the i−th axis, i.e., Σi := (xi,ji)ji∈Z
such that
0 < ∆ := min
i=1,...,N
ji∈Z(xi,ji − xi,ji−1) ≤ max
inf
i=1,...,N
(xi,ji − xi,ji−1) =: ∆ < +∞.
sup
ji∈Z
We say that Σ = (xj)j∈ZN ⊂ RN, xj = (x1,j1, . . . , xN,jN ), j = (j1, . . . , jN ) ∈
ZN, is an admissible sequence if it is the cartesian product of admissible parti-
tions Σi = (xi,ji)ji∈Z.
For a fixed admissible sequence Σ, the lp(Σ)− norm of a function f : RN → R
is defined as
where Qj =(cid:81)N
(cid:107)f(cid:107)lp(Σ) :=
(cid:88)
1
i=1[xi,ji−1 − xi,ji[ and ∆j :=(cid:81)N
f (x)p∆j
p
sup
x∈Qj
j∈ZN
, 1 ≤ p < +∞
i=1(xi,ji − xi,ji−1) is the volume
of Qj. With such notations the subspace Λp is defined as
Λp := {f ∈ M (RN ) : (cid:107)f(cid:107)lp(Σ) < +∞, for every admissible sequence Σ}.
In [17] it is proved that Λp is a proper linear subspace of Lp(RN ), together with
other properties concerning such space. In particular, we recall the following
important result of convergence in Lp for the τ−modulus of smoothness (see
[34, 33, 42])
τr(f ; δ, M (RN ))p := (cid:107)ωr(f ;·, δ, M (RN ))(cid:107)p,
11
where
ωr(f ; x, δ, M (RN )) := sup
∆r
hf (t) :=(cid:80)r
(cid:40)
hf (t) : t, t + hr ∈ N(cid:89)
(cid:32)
(cid:33)
i=1
(cid:20)
xi − δr
2
, xi +
δr
2
(cid:21)(cid:41)
,
j=0(−1)r+j
r
δ > 0, and ∆r
order r at t ∈ RN with increment h ∈ RN.
Proposition 3 (Prop. 7 of [17]) If f ∈ Λp ∩ Rloc(RN ), 1 ≤ p < +∞, r ∈ N,
then
f (t + jh) denotes the differences of
j
τr(f ; δ, M (RN ))p = 0.
lim
δ→0+
We point out that, of course, Λp is a non trivial subspace since it contains,
for example, all the functions in M (RN ) with compact support.
We are now ready to state the result of convergence in Lp for the multidi-
mensional sampling-Kantorovich type operators.
Let us now assume that χ is a kernel with compact support, i.e., χ(t) = 0
if t (cid:54)∈ [−T, T ]N, T > 0, not necessarily of product type, namely χ satisfies the
assumption:
(χ) there exists T > 0 such that χ(t) = 0 for t (cid:54)∈ [−T, T ]N and(cid:80)
k∈ZN χ(u − k) =
1, for every u ∈ RN.
We point out that, since χ has compact support, obviously χ ∈ L1(RN ) (and
therefore (χ1) holds) and satisfies the condition (χ2) since χ is bounded and the
series reduces to a finite sum.
Proposition 4 Let f ∈ Λp ∩ Rloc(RN ), 1 ≤ p < +∞. Then, for every j =
1, . . . , N,
Proof. By assumption (χ) there holds, for w > 0, j = 1, . . . , N and t ∈ RN,
(Kw,jf )(t) − f (t) =
du χ(wt − k) − f (t)
w
w
f
lim
k∈ZN
w→+∞(cid:107)Kw,jf − f(cid:107)p = 0.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:90) kj +1
(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
w
(cid:90) kj +1
(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)w
(cid:90) kj +1
≤ (cid:88)
(cid:18) k1
(cid:18) k1
(cid:18) k1
k∈ZN
k∈ZN
kj
w
kj
w
kj
w
w
w
f
=
w
, . . . , u . . . ,
w
f
, . . . , u . . . ,
, . . . , u . . . ,
w
(cid:19)
kN
w
kN
w
(cid:19)
kN
w
(cid:19)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
χ(wt − k)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)χ(wt − k).
du − f (t)
du − f (t)
12
du − f (t)
f ; t,
2T
w
, M (RN )
,
(cid:19)
Since χ has support contained in [−T, T ]N, χ(wt − k) = 0 if wt − k > T , and
therefore the series reduces to a finite sum over the indexes k ∈ ZN such that
w , for every i = 1, . . . , N: for such k
(cid:12)(cid:12) ≤ T
w
w
kj
w
w
f
, . . . , u . . . ,
we have that
(cid:18) k1
(cid:90) kj +1
(wt − k) ∈ [−T, T ]N, namely(cid:12)(cid:12)ti − ki
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)w
(cid:19)
(Kw,jf )(t) − f (t) ≤ (cid:88)
(cid:18)
(cid:18)
(wt−k)∈[−T,T ]N
≤ ω1
kN
w
2T
w
f ; t,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)w
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ ω1
(cid:18) k1
(cid:90) kj +1
(cid:19) (cid:88)
(cid:19)
kj
w
w
f
w
(wt−k)∈[−T,T ]N
by the definition of the modulus of smoothness. Therefore
(cid:19)
kN
w
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)χ(wt − k)
du − f (t)
, . . . , u, . . . ,
, M (RN )
χ(wt − k)
≤ Aχω1
f ; t,
2T
w
, M (RN )
,
by (χ2). Passing to the Lp−norm we obtain
(cid:107)Kw,jf − f(cid:107)p ≤ Aχτ1
and the thesis follows by Proposition 3.
(cid:18)
f ;
(cid:19)
,
p
, M (RN )
2T
w
5 Convergence in variation for the multidimen-
sional generalized sampling series
We are now ready to prove the main result of the paper, that is, the convergence
in variation for the multidimensional generalized sampling series with product
kernels of averaged type. Since we will use results of the previous section, we
assume here that χ is a kernel with compact support, that is, χ : RN → R
satisfies assumption (χ).
Theorem 1 Let f ∈ AC(RN ) be such that ∂f
j = 1, . . . , N. Then, for every m ∈ N,
∈ Λ1 ∩ Rloc(RN ), for every
∂xj
w→+∞ V [ ¯Sm
lim
w f − f ] = 0.
13
Proof. Since f ∈ AC(RN ), by Proposition 1, ¯Sm
and m ∈ N: then
w f ∈ AC(RN ), for every w > 0
(cid:90)
∇ ¯Sm
w f (t) − ∇f (t) dt.
(cid:12)(cid:12)(cid:12)(cid:12) dt
(t)
(cid:48)
j, tj − m − 2(i − 1)
j, tj − m − 2(i − 1)
2w
(cid:48)
t
t
2w
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dt
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dt
− ∂f
∂tj
(t)
(cid:19)
(cid:19)
Now, by Proposition 2,
V [ ¯Sm
w f − f ] ≤
=
RN
j=1
i=1
RN
i=1
1
m
RN
j=1
Kj,w
RN
m
Kj,w
V [ ¯Sm
∂f
∂tj
w f − f ] =
N(cid:88)
(cid:90)
( ¯Sm
∂tj
(cid:90)
N(cid:88)
≤ N(cid:88)
− m(cid:88)
(cid:90)
N(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)Kw,j
≤ N(cid:88)
N(cid:88)
m(cid:88)
N(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12) ∂
w f )(t) − ∂f
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:19)(cid:18)
(cid:18)
m(cid:88)
∂tj
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) m(cid:88)
(cid:18)
(cid:19)(cid:18)
(cid:90)
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) m(cid:88)
(cid:18)
(cid:13)(cid:13)(cid:13)(cid:13)1
(cid:12)(cid:12)(cid:12)(cid:12) ∂f
(cid:18)
(cid:90)
m(cid:88)
N(cid:88)
j, tj − m − 2(i − 1)
− ∂f
∂tj
∂f
∂tj
∂f
∂tj
∂f
∂tj
∂f
∂tj
RN
∂tj
1
m
+
1
m
RN
i=1
j=1
i=1
(cid:48)
t
2w
+
j=1
j=1
i=1
j=1
+
:=
Lj +
1
m
j=1
j=1
i=1
I i
j.
j, tj − m − 2(i − 1)
(cid:48)
t
2w
j, tj − m − 2(i − 1)
(cid:48)
t
2w
(cid:19)
(cid:19)
− ∂f
∂tj
− ∂f
∂tj
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dt
(cid:12)(cid:12)(cid:12)(cid:12) dt
(t)
(t)
Now, Lj → 0 as w → +∞ by Proposition 4, while I i
j → 0 as w → +∞, for every
i = 1, . . . , m, j = 1, . . . , N, by the continuity in L1 of the translation operator.
Therefore the theorem is proved.
6 Examples and applications to smoothing for
digital image processing
In this section, we provide some basic examples of kernels for which the above
results hold, and we discuss some applications of the variation diminishing-type
property (5) to smoothing in digital image processing.
14
Figure 1: On the left: the Fejér kernel F . On the right: the averaged Fejér
kernel with m = 1.
As stated in Section 2, in this paper we consider product kernels ¯χm of
averaged type.
As a first example, we can consider the multivariate product kernel of the
averaged type generated by the well-known Fejér kernel (see, e.g., Fig. 1 left
and [27]), defined by:
F (x) :=
sinc2(x/2),
x ∈ R,
where the sinc-function (see, e.g., [26, 7]) is of the form:
x (cid:54)= 0,
x = 0.
sinc(x) :=
sin(πx)/πx,
1,
1
2
(cid:40)
Now, the one-dimensional averaged Fejér kernel (see, e.g., Fig. 1, right) is the
following:
(cid:90) m/2
(cid:18) t + v
(cid:19)
2
¯Fm(t) :=
1
sinc2
2 m
−m/2
dv,
t ∈ R, m ∈ N,
and the corresponding multivariate version (see, e.g., Fig. 2 for the case N = 2)
is:
N(cid:89)
Fm(t) :=
¯Fm(ti),
t ∈ RN .
i=1
It is well-known that the Fejér kernel has unbounded support and satisfies as-
sumptions (χ1) and (χ2) (see e.g., [31, 8]), then for the multivariate generalized
sampling series based upon Fm holds the variation diminishing type property
established in Proposition 1, when multivariate signals of bounded variation are
considered.
15
Figure 2: The bivariate averaged Fejér kernel with m = 1.
Moreover, also the relation established in Proposition 2 holds, where the
j-th first order partial derivatives of the generalized sampling series of a given
absolutely continuous function f is related with Kw,j
, i.e., the corresponding
sampling series of the Kantorovich type of the j-th first order partial derivatives
of f.
∂f
∂tj
However, the convergence results proved in Section 4 can not be applied to
the generalized sampling series based upon Fm since the kernels Fm do not have
compact support, hence assumption (χ) is not satisfied.
Other examples of one-dimensional kernels with unbounded support that
can be used to define product averaged type kernels can be found, e.g., in
[38, 28, 29, 30].
In order to recall examples of kernels of one variable such that the corre-
sponding multivariate averaged type versions also satisfy assumption (χ), we
recall the definition of the well-known central B-spline of order n ∈ N (see, e.g.,
Fig. 3 and [46, 25, 2, 3, 20]), defined by:
n(cid:88)
i=0
(cid:18)n
(cid:19)(cid:16) n
i
2
(cid:17)n−1
+
Mn(x) :=
1
(n − 1)!
(−1)i
+ x − i
x ∈ R,
,
(6)
where (x)+ := max{x, 0} denotes "the positive part" of x ∈ R (see e.g., [41,
16
Figure 3: On the left: the central B-spline M2. On the right: the central
B-spline M3 which corresponds to ¯M2,1.
32]). The functions Mn(x) are non-negative, continuous with compact support
contained in [−n/2, n/2], and satisfy conditions (χ1) and (χ2).
Now, let us denote by
¯Mn,m(t) := m−1
(cid:90) m/2
−m/2
Mn(t + v) dv,
t ∈ R,
the averaged B-spline kernel of order n ∈ N. Recalling the following well-known
property:
M(cid:48)
n(t) = Mn−1(t + 1/2) − Mn−1(t − 1/2),
t ∈ R,
(n ≥ 2)
for m = 1, we have:
¯M(cid:48)
n,1(t) = Mn(t + 1/2) − Mn(t − 1/2) = M(cid:48)
n+1(t),
t ∈ R,
(n ≥ 1),
i.e., ¯Mn,1(t) = Mn+1(t) + k, k ∈ R. Now, since ¯Mn,1 belongs to L1(R) (see [6]),
we must have k = 0 and therefore we conclude that
¯Mn,1(t) = Mn+1(t),
t ∈ R,
for every n ∈ N, namely, the averaged kernel with m = 1 generated by a central
B-spline of order n is a B-spline itself of order n + 1.
In view of the above remark, we can explicitly state that the multivariate
averaged type kernel with m = 1 and generated by Mn (see, e.g., Fig. 4 for the
case n = 2 in two space dimension) is the following:
Mn
1 (t) :=
¯Mn,1(ti) =
Mn+1(ti),
t ∈ RN .
N(cid:89)
N(cid:89)
i=1
i=1
17
Figure 4: The bivariate averaged B-spline kernel of order 2 with m = 1 which
coincides with M2
i=1 M3(ti).
1(t) =(cid:81)2
In practice, in the latter case the multivariate generalized sampling operators
based upon averaged B-spline Mn with m = 1 coincide with the usual general-
ized sampling series based upon the multivariate central B-spline of order n + 1.
Generally speaking, we can also define the multivariate average central B-
spline kernels as follows:
N(cid:89)
i=1
(cid:90) m/2
−m/2
N(cid:89)
i=1
Mn
m(t) :=
¯Mn,m(ti) = m−1
Mn(ti + v) dv,
t ∈ RN .
Since Mn
m has compact support it satisfies also assumption (χ); thus the above
sampling series fulfills the results of both Section 3 and Section 5. Further, for
the latter examples of kernels also Proposition 4 and Theorem 1 can be applied.
Now, at the end of this section we consider some applications to smoothing
in digital processing. For basic facts concerning this numerical tool for imaging,
see e.g., [19, 39, 43].
It is well-known that any static gray scale image is a bivariate signal with
18
compact support; then it can be (naturally) modeled as follows:
m(cid:88)
m(cid:88)
IA(x, y) :=
aij · 1ij(x, y)
((x, y) ∈ R2),
i=1
j=1
for every image (matrix) A = (aij)ij, i, j = 1, 2, ..., m, where 1ij(x, y), i, j =
1, 2, ..., m, is the characteristic function of the sets (i − 1, i] × (j − 1, j] (i.e.
1ij(x, y) = 1, for (x, y) ∈ (i − 1, i] × (j − 1, j] and 1ij(x, y) = 0 otherwise).
Note that the above function IA(x, y) is defined in such a way that to every
pixel (i, j) it is associated the corresponding gray level aij.
Moreover, by the above representation IA of the image A it turns out that
IA ∈ BV (R2), hence one can consider approximations of A by means of the
bivariate generalized sampling series based upon the averaged type kernels. The
main advantage that can be achieved by the above procedure is expressed by
w IA have no-bigger variation than IA, for sufficiently large
Proposition 1: ¯Sm
m ∈ N, i.e., the operators provide an approximated version of the original
image A, but regularized. In this sense, the generalized sampling series can be
used for smoothing of images.
Clearly, in order to visualize an approximation (new image) of the original
w IA, we need to sample the operators, for w > 0,
image A by means of ¯Sm
with a fixed sampling rate concordant with the dimension of A. Obviously, the
sampling rate is chosen arbitrarily hence one can also consider different (high)
sampling rates.
The effect of the proposed procedure can be strongly noticed at the edges
of the figures, where the jumps of gray levels are reduced with respect to the
corresponding ones in the original image.
Now we can give the following practical examples of image reconstruction in
order to show the smoothing capabilities of the above operators. An optimized
version of the above described algorithm for image reconstruction and smooth-
ing can be implemented by means of the MATLAB programming language,
following the indications outlined in [14, 15] in case of the so-called sampling
Kantorovich algorithm for image enhancement.
For the numerical experiments, we consider the well-known images of Lena
and Baboon with 150 × 150 pixel resolution (see Fig. 5).
First of all, we reconstruct the original images in Fig. 5 by using the bivariate
averaged Fejér kernel with m = 4. According to Proposition 1, and observing
19
Figure 5: On the left: Lena. On the right: Baboon.
Figure 6: On the left: the reconstruction of Lena (of 150 × 150 pixel) by means
of the operator ¯Sm
w , w = 4, based upon the bivariate averaged Fejér kernel with
m = 4. On the right: the reconstruction of Baboon made as for Lena.
20
Figure 7: On the left: the reconstruction of Lena (of 150 × 150 pixel) by means
of the operator ¯Sm
w , w = 10, based upon the bivariate averaged Fejér kernel
with m = 4. On the right: the reconstruction of Baboon made as for Lena.
that (cid:107)F(cid:107)1 = 1, we have:
V [ ¯S4
wf ] ≤ 1
4
V [f ],
(7)
i.e., the total variation of any reconstructed images is, at least, 4 times smaller
than that of the original ones, producing the smoothing effect. In Fig. 6 we
have the reconstruction of Lena and Baboon (of 150 × 150 pixel resolution) by
w , w = 4, based upon the bivariate averaged Fejér
means of the operator ¯Sm
kernel with m = 4. By detailed analysis of the edges (especially in case of
Lena, at the contours of the hat) it is possible to observe the smoothing of the
analyzed images. In Fig. 7 we have the reconstruction of Lena and Baboon (of
150 × 150 pixel resolution) by means of the operator ¯Sm
w , w = 10, based upon
the bivariate averaged Fejér kernel with m = 4.
The main differences that can be observed between the images in Fig. 6 and
Fig. 7 are that, for big values of w the images are closer to the original and
consequently the edges tend to be more clear. Finally, in Fig. 8 we have the
reconstruction of Lena and Baboon (of 150 × 150 pixel resolution) by means of
the operator ¯Sm
w , w = 6, based upon the bivariate central B-spline of order 3
with m = 4. Note that, also for the latter case one can state a inequality as
that given in (7) since also (cid:107)M3(cid:107)1 = 1.
Acknowledgments
The authors are members of the Gruppo Nazionale per l'Analisi Matematica, la
Probabilitá e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta
21
Figure 8: On the left: the reconstruction of Lena (of 150 × 150 pixel) by means
of the operator ¯Sm
w , w = 6, based upon the bivariate averaged central B-spline
of order 3 with m = 4. On the right: the reconstruction of Baboon made as for
Lena.
Matematica (INdAM).
The authors are partially supported by the "Department of Mathematics and
Computer Science" of the University of Perugia (Italy). The first and the third
author have been partially supported within the project "Metodi di teoria degli
operatori e di Analisi Reale per problemi di approssimazione ed applicazioni",
funded by the 2017 basic research fund of the University of Perugia. Finally,
the second author of the paper has been partially supported within a 2018
GNAMPA-INdAM Project:
"Dinamiche non autonome, analisi reale e appli-
cazioni".
References
[1] O. Agratini, An approximation process of Kantorovich type, Math. Notes,
Miskolc, 2 (1) (2001), 3-10.
[2] G. Allasia, R. Cavoretto, A. De Rossi, A class of spline functions for
landmark-based image registration, Math. Methods Appl. Sci., 35 (2012),
923 -- 934.
[3] G. Allasia, R. Cavoretto, A. De Rossi, Lobachevsky spline functions and
interpolation to scattered data, Comput. Appl. Math., 32 (2013), 71 -- 87.
22
[4] L. Ambrosio, S. Di Marino, Equivalent definitions of BV space and of total
variation on metric measure spaces, Journal of Functional Analysis, 266
(7) (2014), 4150 -- 4188.
[5] L. Angeloni, Approximation results with respect to multidimensional ϕ-
variation for nonlinear integral operators, Z. Anal. Anwendungen, 32 (1)
(2013), 103 -- 128.
[6] L. Angeloni, D. Costarelli, G. Vinti, A characterization of the convergence
in variation for the generalized sampling series, Ann. Acad. Sci. Fenn.
Math., 43 (2018), 755-767.
[7] L. Angeloni, D. Costarelli, G. Vinti, A characterization of the absolute con-
tinuity in terms of convergence in variation for the sampling Kantorovich
operators, Mediterr. J. Math., 16 (2019), 44. DOI:10.1007/s00009-019-
1315-0
[8] L. Angeloni, D. Costarelli, G. Vinti, Quantitative estimates for sampling
type operators with respect to the Jordan variation, submitted (2019).
[9] L. Angeloni, G. Vinti, Convergence and rate of approximation for linear
integral operators in BV ϕ−spaces in multidimensional setting, J. Math.
Anal. Appl., 349 (2009), 317 -- 334.
[10] L. Angeloni, G. Vinti, Approximation in variation by homothetic operators
in multidimensional setting, Differential Integral Equation, 26 (5-6) (2013),
655 -- 674.
[11] L. Angeloni, G. Vinti, Convergence and rate of approximation in BV ϕ(RN
+ )
for a class of Mellin integral operators, Atti della Accademia Nazionale dei
Lincei, Classe di Scienze Fisiche, Matematiche e Naturali, Rendiconti Lincei
Matematica e Applicazioni, 25 (3) (2014), 217 -- 232.
[12] L. Angeloni, G. Vinti, Convergence in variation and a characterization of
the absolute continuity, Integral Transforms Spec. Funct., 26 (10) (2015),
829 -- 844.
[13] J. Appell, J. Banaś, N. Merentes, Bounded variation and around, 17 De
Gruyter Series in Nonlinear Analysis and Applications, De Gruyter, Berlin,
Germany, (2014).
23
[14] F. Asdrubali, G. Baldinelli, F. Bianchi, D. Costarelli, A. Rotili, M. Seracini,
G. Vinti, Detection of thermal bridges from thermographic images by means
of image processing approximation algorithms, Applied Mathematics and
Computation, 317 (2018) 160 -- 171.
[15] F. Asdrubali, G. Baldinelli, F. Bianchi, D. Costarelli, L. Evangelisti, A.
Rotili, M. Seracini, G. Vinti, A model for the improvement of thermal
bridges quantitative assessment by infrared thermography, Applied Energy,
211 (2018) 854 -- 864.
[16] C. Bardaro, P.L. Butzer, R.L. Stens, G. Vinti, Prediction by samples from
the past with error estimates covering discontinuous signals, IEEE Trans.
Inform. Theory, 56 (1) (2010), 614 -- 633.
[17] C. Bardaro, I. Mantellini, R. Stens, J. Vautz, G. Vinti, Generalized sam-
pling approximation for multivariate discontinuous signals and application
to image processing, New Perspectives on Approximation and Sampling
Theory-Festschrift in honor of Paul Butzer's 85th birthday, Birkhauser
(2014), 87-114.
[18] C. Bardaro, G. Vinti, A general approach to the convergence theorems of
generalized sampling series, Appl. Anal., 64 (1997), 203 -- 217.
[19] S. Bi, X. Han, Y. Yu, An L1 image transform for edge-preserving smoothing
and scene-level intrinsic decomposition, ACM Transactions on Graphics, 34
(4) (2015) No. 78.
[20] M. Bozzini, L. Lenarduzzi, M. Rossini, R. Schaback, Interpolation with
variably scaled kernels, IMA Journal of Numerical Analysis, 35 (1) (2015)
199 -- 219.
[21] P.L. Butzer, A. Fisher, R.L. Stens, Generalized sampling approximation of
multivariate signals: theory and applications, Note di Matematica, 10 (1)
(1990), 173 -- 191.
[22] P.L. Butzer, A. Fisher, R.L. Stens, Generalized sampling approximation of
multivariate signals: general theory, Atti Sem. Mat. Fis. Univ. Modena, 41
(1) (1993), 17 -- 37.
24
[23] P.L. Butzer, R.J. Nessel, Fourier Analysis and Approximation I, Academic
Press, New York-London, 1971.
[24] P.L. Butzer, W. Splettstösser, R.L. Stens, The sampling theorem and linear
prediction, Jahresberichte Dt. Math.-Verein, 90 (1988), 1 -- 70.
[25] M. Charina, C. Conti, K. Jetter, G. Zimmermann, Scalar multivariate sub-
division schemes and box splines, Computer Aided Geometric Design, 28
(5) (2011), 285-306.
[26] L. Coroianu, S.G. Gal, Lp- approximation by truncated max-product sam-
pling operators of Kantorovich-type based on Fejer kernel, Journal of Inte-
gral Equations and Applications, 29 (2) (2017) 349 -- 364.
[27] D. Costarelli, A.M. Minotti, G. Vinti, Approximation of discontinuous sig-
nals by sampling Kantorovich series, Journal of Mathematical Analysis and
Applications, 450 (2) (2017), 1083 -- 1103.
[28] D. Costarelli, G. Vinti, Approximation by max-product neural network oper-
ators of Kantorovich type, Results in Mathematics, 69 (3) (2016), 505 -- 519.
[29] D. Costarelli, G. Vinti, Max-product neural network and quasi-interpolation
operators activated by sigmoidal functions, J. Approx. Theory, 209 (2016),
1 -- 22.
[30] D. Costarelli, G. Vinti, Pointwise and uniform approximation by multivari-
ate neural network operators of the max-product type, Neural Networks, 81
(2016), 81 -- 90.
[31] D. Costarelli, G. Vinti, An inverse result of approximation by sampling
Kantorovich series, Proceedings of the Edinburgh Mathematical Society,
62 (1) (2019), 265-280.
[32] L. D'Amore, R. Campagna, A. Galletti, L. Marcellino, A. Murli, A smooth-
ing spline that approximates Laplace transform functions only known on
measurements on the real axis, Inverse Problems 28 (2) (2012), 025007.
[33] E.P. Dolženko, E.A. Sevast'janov, Approximations of functions in the Haus-
dorff metric by means of piecewise monotone (in particular, rational) func-
tions (Russian), Mat. Sb. (N.S.), 101 (143) no. 4 (1976), 508 -- 541.
25
[34] P.P. Korovkin, Attempt at an axiomatic construction in certain problems in
the theory of approximation of functions of one variable (Russian), Kalinin.
Gos. Ped. Inst. Učen. Zap., 69 (1969), 91 -- 109.
[35] E. Giusti, Minimal Surfaces and Functions of Bounded Variation, Mono-
graphs in Mathematics, vol. 80, Birkhäuser Verlag, Basel, 1984.
[36] Y. Gousseau, J.M. Morel, Are Natural Images of Bounded Variation?,
SIAM J. Math. Anal., 33 (3) (2001), 634 -- 648.
[37] I. Mantellini, G. Vinti, Approximation results for nonlinear integral opera-
tors in modular spaces and applications, Ann. Polon. Math., 81 (1) (2003),
55 -- 71.
[38] O. Orlova, G. Tamberg, On approximation properties of generalized
Kantorovich-type sampling operators, Journal of Approximation Theory,
201 (2016), 73 -- 86.
[39] J. Pan, X. Yang, H. Cai, B. Mu, Image noise smoothing using a modified
Kalman filter, Neurocomputing, 173 (3) (2016), 1625 -- 1629
[40] T. Radó, Length and Area, American Mathematical Society Colloquium
Publications, vol. 30, American Mathematical Society, New York, 1948.
[41] L. Romani, M.A. Sabin, The conversion matrix between uniform B-spline
and Bézier representations, Computer aided geometric design, 21 (6)
(2004), 549 -- 560.
[42] B. Sendov, V.A. Popov, The Averaged Moduli of Smoothness, Wiley, Chich-
ester, U.K., 1988.
[43] M. Sonka, V. Hlavac, R. Boyle, Image processing, analysis, and machine
vision, Cengage Learning, Australia - Brazil - Japan, (2014).
[44] D. Strong, T. Chan, Edge-preserving and scale-dependent properties of total
variation regularization, Inverse Problems, 19 (6) (2003).
[45] L. Tonelli, Su alcuni concetti dell'analisi moderna, Ann. Scuola Norm. Su-
per. Pisa. (2), 11 (1942), 107 -- 118.
[46] M. Unser, Ten good reasons for using spline wavelets, Wavelets Applications
in Signal and Image Processing, 3169 (5) (1997), 422 -- 431.
26
[47] T. Valkonen, K. Bredies, F. Knoll, Total Generalized Variation in Diffusion
Tensor Imaging, SIAM J. Imaging Sci., 6 (1) (2013), 487 -- 525.
[48] C. Vinti, Perimetro -- variazione, Ann. Scuola Norm. Sup. Pisa (3), 18
(1964), 201 -- 231.
27
|
1807.05599 | 3 | 1807 | 2018-12-07T20:37:11 | Inequalities for $L^p$-norms that sharpen the triangle inequality and complement Hanner's Inequality | [
"math.FA",
"math.CA"
] | In 2006 Carbery raised a question about an improvement on the na\"ive norm inequality $\|f+g\|_p^p \leq 2^{p-1}(\|f\|_p^p + \|g\|_p^p)$ for two functions in $L^p$ of any measure space. When $f=g$ this is an equality, but when the supports of $f$ and $g$ are disjoint the factor $2^{p-1}$ is not needed. Carbery's question concerns a proposed interpolation between the two situations for $p>2$. The interpolation parameter measuring the overlap is $\|fg\|_{p/2}$. We prove an inequality of this type that is stronger than the one Carbery proposed. Moreover, our stronger inequalities are valid for all $p$. | math.FA | math | INEQUALITIES FOR Lp-NORMS THAT SHARPEN THE TRIANGLE
INEQUALITY AND COMPLEMENT HANNER'S INEQUALITY
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
p + (cid:107)g(cid:107)p
p ≤ 2p−1((cid:107)f(cid:107)p
Abstract. In 2006 Carbery raised a question about an improvement on the naıve norm
inequality (cid:107)f + g(cid:107)p
p) for two functions in Lp of any measure space.
When f = g this is an equality, but when the supports of f and g are disjoint the factor
2p−1 is not needed. Carbery's question concerns a proposed interpolation between the two
situations for p > 2. The interpolation parameter measuring the overlap is (cid:107)f g(cid:107)p/2. We
prove an inequality of this type that is stronger than the one Carbery proposed. Moreover,
our stronger inequalities are valid for all p.
December 6, 2018
8
1
0
2
c
e
D
7
]
.
A
F
h
t
a
m
[
3
v
9
9
5
5
0
.
7
0
8
1
:
v
i
X
r
a
1. Introduction and main theorem
Since zp is a convex function of z for p ≥ 1, for any measure space, the Lp unit ball,
Bp := {f : (cid:82) fp ≤ 1}, is convex. One way to express this is with Minkowski's triangle
inequality (cid:107)f + g(cid:107)p ≤ (cid:107)f(cid:107)p + (cid:107)g(cid:107)p. Another is the inequality
p + (cid:107)g(cid:107)p
p ≤ 2p−1(cid:0)(cid:107)f(cid:107)p
(cid:107)f + g(cid:107)p
(cid:1) ,
p
(1.1)
valid for any functions f and g on any measure space. There is equality if and only if f = g
and, in Theorem 1.1, we improve (1.1) substantially when f and g are far from equal.
In 2006 Carbery proposed [3] several plausible refinements of (1.1) for p ≥ 2, of which the
strongest was
(cid:90)
(cid:18)
f + gp ≤
1 +
(cid:107)f g(cid:107)p/2
(cid:107)f(cid:107)p(cid:107)g(cid:107)p
(cid:19)p−1(cid:90)
(fp + gp) .
(1.2)
There is equality in (1.2) both when f = g and when f g = 0. Thus, (1.2), if true, can
be viewed as a refinement of (1.1) in which there is equality not only when f = g but also
when f g = 0.
The ratio Γ =
(cid:107)f g(cid:107)p/2
(cid:107)f(cid:107)p(cid:107)g(cid:107)p
varies between 1 and 2p−1, interpolating between the two cases of equality in (1.2).
varies between 0 and 1 and, therefore, the factor of (1 + Γ)p−1
c(cid:13) 2018 by the authors. This paper may be reproduced, in its entirety, for non-commercial purposes.
Work partially supported by NSF grants DMS -- 1501007 (E.A.C.), DMS -- 1363432 (R.L.F.), PHY -- 1265118
(E.H.L.).
1
2
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
We propose and prove a strengthening of (1.2) in which Γ is replaced by the quantity
(cid:18)(cid:107)f(cid:107)p
(cid:101)Γ := (cid:107)f g(cid:107)p/2
(cid:19)−2/p
p + (cid:107)g(cid:107)p
2
p
,
(1.3)
which is smaller by virtue of the arithmetic-geometric mean inequality.
Our improved inequalities are not restricted to p > 2, but are valid for all p ∈ R, as stated
in Theorem 1.1. There we write
(cid:107)f(cid:107)p :=
(cid:18)(cid:90)
(cid:19)1/p
fp
for all p (cid:54)= 0.
p and (cid:107)f(cid:107)p(cid:107)g(cid:107)p), while our inequality involves only two ((cid:107)f g(cid:107)p/2 and (cid:107)f(cid:107)p
We note that inequality (1.2) involves three kinds of quantities on the right side ((cid:107)f g(cid:107)p/2, (cid:107)f(cid:107)p
p+
(cid:107)g(cid:107)p
p), a
simplification that is essential for our proof.
Theorem 1.1 (Main Theorem). For all p ∈ (0, 1] ∪ [2,∞) and functions f and g on any
p + (cid:107)g(cid:107)p
measure space, (cid:90)
f + gp ≤
(cid:32)
(cid:33)p−1(cid:90)
1 +
22/p(cid:107)f g(cid:107)p/2
p + (cid:107)g(cid:107)p
((cid:107)f(cid:107)p
p)2/p
(fp + gp ) .
(1.4)
The inequality reverses if p ∈ (−∞, 0)∪ (1, 2), where, for p ∈ (1, 2), it is assumed that f and
g are positive almost everywhere.
For p > 2, (resp. for p ∈ (0, 1) ) the inequality is false if (cid:101)Γ is raised to any power q > 1,
For p < 0, (resp. for p ∈ [1, 2] ) the reversed inequality is false if (cid:101)Γ is raised to any power
(resp. q < 1).
q > 1, (resp. for q < 1).
For p > 0, p (cid:54)= 1, 2, (cid:107)f(cid:107)p,(cid:107)g(cid:107)p < ∞, there is equality in (1.4) if and only if f and
g have disjoint supports, up to a null set, or are equal almost everywhere. For p < 0,
(cid:107)f(cid:107)p,(cid:107)g(cid:107)p < ∞, there is equality in (1.4) if and only if f and g are equal almost everywhere.
We note that in proving the theorem, we may always assume that f and g are non-
negative. In fact, the right side of (1.4) only depends on f and g and the left side does
not decrease for p ≥ 0 and does not increase for p < 0 if f and g are replaced by f and
g. The latter follows since f + g ≤ f + g implies f + gp ≤ (f + g)p for p > 0 and
f + gp ≥ (f + g)p for p < 0.
Carbery proved that his proposed inequality is valid when f and g are characteristic
functions. Our theorem can also be easily proved in this special case.
Theorem 1.1 may be viewed as a refinement of Minkowski's inequality. Since (1.1), like
Minkowski's inequality, is a direct expression of the convexity of Bp, it is equivalent to
Minkowski's inequality. We recall the simple argument: For any unit vectors u, v ∈ Lp,
(1.1) says that (cid:107)(u + v)/2(cid:107)p ≤ 1, and then by continuity, (cid:107)λu + (1 − λ)v(cid:107)p ≤ 1 for all
HANNER COMPLEMENT -- December 6, 2018
3
λ ∈ (0, 1). Suppose 0 < (cid:107)f(cid:107)p,(cid:107)g(cid:107)p < ∞, and define λ = (cid:107)f(cid:107)p/((cid:107)fp(cid:107) + (cid:107)g(cid:107)p), u = (cid:107)f(cid:107)−1
and v = (cid:107)g(cid:107)−1
p f ,
p g. Then
(cid:107)f + g(cid:107)p
p = ((cid:107)f(cid:107)p + (cid:107)g(cid:107)p)p(cid:107)λu + (1 − λ)v(cid:107)p
p ≤ ((cid:107)f(cid:107)p + (cid:107)g(cid:107)p)p ,
which is Minkowski's inequality.
When p = 1 and f, g ≥ 0, (1.1) is an identity; otherwise when p > 1, there is equality
in (1.1) if and only if f = g. When the supports of f and g are disjoint, however, (1.1) is
far from an equality and the factor 2p−1 is not needed. There is equality in Minkowski's
inequality whenever f is a multiple of g or vice-versa. Hence although (1.1) is equivalent to
Minkowski's inequality, it becomes an equality in fewer circumstances.
There is another well-known refinement of Minkowski's inequality for 1 < p < ∞, namely
Hanner's inequality, [4, 2, 6] which gives the exact modulus of convexity of Bp, the unit ball
in Lp. For p ≥ 2, and unit vectors u and v, Hanner's inequality says that
(cid:13)(cid:13)(cid:13)(cid:13) u + v
2
(cid:13)(cid:13)(cid:13)(cid:13)p
p
(cid:13)(cid:13)(cid:13)(cid:13) u − v
2
(cid:13)(cid:13)(cid:13)(cid:13)p
p
+
≤ 1,
(1.5)
which is also a consequence of one of Clarkson's inequalities [1]. When u and v have disjoint
supports, (cid:107)u+v(cid:107)p
p = 2, and then the left hand side is 22−p, so that for unit vectors
u and v, the condition uv = 0, which yields equality in the inequality of Theorem 1.1, does
p = (cid:107)u−v(cid:107)p
not yield equality in Hanner's inequality. On the other hand, while one can derive a bound
on the modulus of convexity in Lp from (1.4), one does not obtain the sharp exact result
provided by Hanner's inequality. Both inequalities express a quantitative strict convexity
property of Bp, but neither implies the other; they provide complimentary information, with
the information provided by Theorem 1.1 being especially strong when f and g have small
overlap as measured by (cid:107)f g(cid:107)p/2.
Our proof of Theorem 1.1 consists of three parts:
Part A: We show how to reduce the inequality to a simpler one involving only one
function, namely α := f /(f + g) for f, g > 0, which takes values in [0, 1], and a reference
measure that is a probability measure. This exploits the fact that the only important
quantity is the ratio of f to g. This part is very easy.
Part B: In the second part, which is more difficult than Part A, we show that Theo-
rem 1.1 is true if it is true when the function α is constant. (This is the same as saying f
and g are proportional to each other.) When α := f /(f + g) is constant and the reference
measure is a probability measure, (1.4) yields the inequality for numbers α ∈ [0, 1] and
p ∈ [0, 1] ∪ [2,∞],
(cid:32)
1 +
1 ≤
(cid:33)2/pp−1
2αp/2(1 − α)p/2
αp + (1 − α)p
(αp + (1 − α)p) .
(1.6)
4
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
with the reverse inequality for p /∈ [0, 1] ∪ [2,∞].
2αp/2(1 − α)p/2
αp + (1 − α)p
Remark 1.2. The quantity R :=
decreases as q increases. Thus for p ≥ 2, the inequality
(cid:33)q(cid:33)p−1
(cid:32)
(cid:32)
1 ≤
1 +
2αp/2(1 − α)p/2
αp + (1 − α)p
lies in [0, 1] for all α and p. Therefore, Rq
(αp + (1 − α)p) ,
(1.7)
strengthens as q increases, and for p ∈ [0, 1], it strengthens as q decreases. Likewise, for
p ∈ [1, 2] the reverse of (1.7) is stronger for smaller q, and for p < 0, it is stronger for larger
q.
Part C: With Parts A and B complete, the proof reduces to a seemingly elementary
inequality, parametrized by p, for a number α ∈ [0, 1]. The proof of this is Part C. While the
validity of (1.6) appears to be a consequence of Theorem 1.1, one can also view Theorem
1.1 as a consequence of (1.6).
Theorem 1.3. For all numbers α ∈ [0, 1], inequality (1.6) is valid for all p ∈ (0, 1]∪ [2,∞),
and the reverse inequality is valid for all p ∈ (−∞, 0) ∪ (1, 2).
For p > 2, (resp. for p ∈ (0, 1) ) inequality (1.7) is false if q > 2/p, (resp. for q < 2/p).
for
For p < 0,
for p ∈ [1, 2] ) the reverse inequality is false if q > 2/p, (resp.
(resp.
q < 2/p).
For p > 0, p (cid:54)= 1, 2, there is equality if and only if α ∈ {0, 1/2, 1}. For p < 0, there is
equality if and only if α = 1/2.
1.1. Restatement of Theorem 1.3 in terms of means. Inequality (1.6) can be restated
in terms of qth power means [5]: For x, y > 0, define
if q ∈ R \ {0}
Mq(x.y) = ((xq + yq)/2)1/q
and
M0(x, y) =
√
xy .
Note that M0(x, y) is the geometric mean of x and y and M−1(x, y) is their harmonic mean.
Corollary 1.4. For all x, y > 0, and all p ∈ [0, 1] ∪ [2,∞]
(cid:18) Mp(x, y) + M−p(x, y)
(cid:19)p−1
2
1 (x, y) ≤
M p
Mp(x, y) ,
(1.8)
while the reverse inequality is valid for all p ∈ (−∞, 0) ∪ (1, 2).
M−p(x, y)
Proof. A simple calculation shows that for all p > 0,
Mp(x, y)
x = α and y = 1 − α, the inequality (1.6) can be written as
=
22/pxy
(xp + yp)2/p
. Thus, taking
(cid:18)
(cid:19)p−1
≤
1
2
M−p(α, 1 − α)
Mp(α, 1 − α)
1 +
p (α, 1 − α) ,
M p
Then by homogeneity and the fact that M1(α, 1 − α) = 1/2, (1.6) is equivalent to (1.8) (cid:3)
HANNER COMPLEMENT -- December 6, 2018
5
The following way to write our inequality sharpens and complements the arithmetic-
geometric mean inequality for any two numbers x, y > 0, provided one has information on
Mp(x, y).
Corollary 1.5 (Improved and complemented AGM inequality). For all x, y > 0, and all
p > 2,
1 −
(cid:18) A
(cid:19)p(cid:48)
Mp
(cid:32)
≥ 1
2
1 −
(cid:19)2(cid:33)
(cid:18) G
Mp
(cid:32)
1 −
(cid:19)2(cid:33)
(cid:18) G
Mp(cid:48)
(cid:18) A
(cid:19)p
Mp(cid:48)
≥ 1 −
(1.9)
≥ 1
2
√
(cid:80)n
where p(cid:48) = p/(p − 1), A = (x + y)/2 and G =
Remark 1.6. Since p, p(cid:48) ≥ 1, all of the quantities being compared in these inequalities are
non-negative.
xy.
Despite the classical appearance of (1.8), we have not been able to find it in the literature,
most of which concerns inequalities for means Mq(x1, . . . , xn) = ( 1
j )1/p of an n-tuple
n
of non-negative numbers, often with more general weights. The obvious generalization of
j=1 xp
(1.8) from two to three non-negative numbers x, y, and z is false as one sees by taking
z = 0: Then there is no help from M−p(x, y, z) on the right. A valid generalization to more
variables probably involves means over M−p(xj, xk) for the various pairs. In any case, as far
as we know, (1.8) is new.
A truly remarkable feature of the inequality (1.8) is that it is surprisingly close to equality
uniformly in the arguments. To see this, let f (α, p) denote the right hand side of (1.6).
Contour plots of this function for various ranges of p are shown in Figs. 1, 2 and 3 below.
Fig. 1
Fig. 2
Fig. 3
Fig. 1 is a contour plot of this function in [1/2, 1] × [2, 4]. The contours shown in Fig. 1
range from 1.00001 to 1.018. Note that the function f is identically 1 along three sides of
plot: α = 1/2, 1, and p = 2. The maximum value for 2 ≤ p ≤ 4, near 1.018, occurs towards
the middle of the segment at p = 4.
Fig. 2 is a contour plot of f on [1/2, 1]× [1, 2]. The contours range from 0.9961 (the small
closed contour) to 0.99999999 (close to the boundary). Amazingly, the function in (1.6)
is quite close -- within two percent -- to the constant 1 over the range p ≥ 1 and α ∈ [0, 1].
6
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
Moreover, the "landscape" is quite flat: The gradient has a small norm over the whole
domain.
Fig. 3 is a contour plot of f in the domain [0, 1/2] × [0, 1]. The contours in Fig. 3 range
from 1.0000001 to 1.06. Higher values are to the right. For p in this range, the maximum is
not so large -- about 1.06 -- but the landscape gets very "steep" near α = 1 and p = 0. The
proof of the inequality is especially delicate in this case.
For p < 0, there is equality only at α = 1/2, and the inequality is not so uniformly close
to an identity. The contour plot is less informative, and hence is not recorded here. This is
the case in which the inequality is easiest to prove.
It is possible to give a simple direct proof of the inequality for certain integer values of
p, as we discuss in Section 5. We also give a simple proof that for p > 2 and for p < 0,
validity of the inequality at p implies validity of the inequality at 2p, and we briefly discuss an
application of this to the problem in which functions are replaced by operators and integrals
are replaced by traces.
Remark 1.7. We close the introduction by briefly discussing one other way to write the
inequality (1.6). Introduce a new variable s ∈ (0, 1) through
√
s
α =
1 +
2
(cid:32)
(cid:33)
1 − s
2
p (s)
η
(1.10)
(1.11)
(1.12)
Rewriting (1.6), and taking the
1
p−1 root of both sides, we may rearrange terms to obtain.
2 ≤ η
1
p−1 (s)
1 +
1
p−1 (s) + (1 − s)η
= η
2−p
p(p−1) (s)
for 0 ≤ s ≤ 1, where
√
s)p + (1 − √
s)p
(1 +
2
.
η(s) :=
Taking the
1
p−1 eliminate the change of direction in the inequality at p = 1, and it now take
on a non-trivial form at p = 1: Define
p−1 (s) + (1 − s)η
for p (cid:54)= 1, and one easily computes the limit at p = 1:
fp(s) := η
1
2−p
p(p−1) (s) − 2 ,
f1(s) := (2 − s)(1 − √
1−√
s
2
s)
(1 +
√
s)
1+
√
2 − 2 .
s
Theorem 1.3 is equivalent to the assertion that for all s ∈ (0, 1)
fp(s) ≥ 0 for p ∈ (−∞, 0) ∪ (2,∞)
and fp(s) ≤ 0 for p ∈ (0, 2) .
(1.13)
In this form, the inequality is easy to check for some values of p. For example, for p = −1,
1−s and f−1(s) = (1 − s)1/2 + (1 − s)−1/2 − 2. which is clearly positive. One can
η(s) = 1
give simple proofs of (1.13) for other integer values of p, e.g., p = 3 and p = 4 along these
HANNER COMPLEMENT -- December 6, 2018
7
line, but this change of variables is not what we use to prove the general inequality. It is,
however, convenient for checking optimality of of the power 2/p in (1.6).
2. Part A. Reduction from two functions to one
While Theorem 1.1 involves two functions f and g one can use the arbitrariness of the
(cid:82) 1 = 1). We have already observed that it suffices to prove the inequality in the case where
measure to reduce the question to a single function defined on a probability space (that is,
f and g are both non-negative. For non-negative functions f and g, set
α = f /(f + g) ,
1 − α = g/(f + g) .
Replacing the underlying measure dx by the new measure (f + g)p dx/(cid:107)f + g(cid:107)p
p we see that it
suffices to prove the following inequality for p ∈ [0, 1] ∪ [2,∞], and also to prove the reverse
inequalities for p /∈ [0, 1] ∪ [2,∞] :
(cid:32)
(cid:33)p−1(cid:0)(cid:107)α(cid:107)p
for a single function 0 ≤ α ≤ 1 on a probability space, i.e.,(cid:82) 1 = 1.
22/p (cid:107)α(1 − α)(cid:107)p/2
((cid:107)α(cid:107)p
p )2/p
p + (cid:107)1 − α(cid:107)p
p + (cid:107)1 − α(cid:107)p
p
1 ≤
1 +
(2.1)
(cid:1)
3. Part B. Reduction to a constant function
In this section we prove the following.
Proposition 3.1. If p ∈ [0, 1] ∪ [2,∞), then inequality (2.1) is true for all functions α
(which is equivalent to (1.2) for all f, g) if and only if it is true for all constant functions,
that is, for all numbers α ∈ [0, 1],
(cid:32)
1 +
1 ≤
(cid:33)2/pp−1
2αp/2(1 − α)p/2
αp + (1 − α)p
(αp + (1 − α)p) .
(3.1)
If p /∈ [0, 1] ∪ [2,∞), then the reverse of inequality (2.1) is true for all functions α (which
is equivalent to the reverse of (1.2) for all f, g) if and only if it is true for all constant
functions, that is, for all numbers α ∈ [0, 1], the reverse of (3.1) holds.
Moreover, for p (cid:54)= 0, 1, 2, there is equality in (2.1) if and only if max{α(x), 1 − α(x)} is
constant almost everywhere.
To prove this Proposition we need a definition and a lemma.
Definition 3.2. Fix p ∈ R and for 0 ≤ a ≤ 1, let h(a) := ap/2(1 − a)p/2 and let b(a) :=
ap+(1−a)p. Clearly, b determines the unordered pair a and 1−a and, therefore, b determines
h. Thus, we can consider the function b (cid:55)→ H(b) := h(a−1(b)) (in which the dependence on
p is suppressed in the notation).
8
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
Lemma 3.3 (convex/concave H). The function b (cid:55)→ H(b) is strictly convex when p ∈ (2,∞)
and strictly concave when p ∈ (−∞, 2), p (cid:54)= 0, 1.
Proof. To prove this lemma we use the chain rule to compute the second derivative of
H. As a first step we define a useful reparametrization as follows: e2x := a/(1 − a). A
quick computation shows that h = (2 cosh x)−p and b = 2 cosh(px)(2 cosh x)−p. Thus, h =
b/(2 cosh(px)). By symmetry, we can restrict our attention to the half-line x ≥ 0.
We now compute the first two derivatives:
db/dx = 21−pp
sinh((p − 1)x)
(cosh x)p+1
dh/dx = −p
tanh x
( 2 cosh x )p
(dH/db)(x) =
dh/dx
db/dx
(d/dx)(dH/db)(x) = cosh(x)
sinh x
2 sinh((p − 1)x)
= −
(p − 1) tanh x − tanh((p − 1)x)
2 sinh((p − 1)x) tanh((p − 1)x)
(d2H/db2)(x) =
(d/dx)(dH/db)(x)
db/dx
(3.2)
(3.3)
(3.4)
(3.5)
(3.6)
Our goal is to show that (3.6) has the correct sign (depending on p) for all x ≥ 0.
Clearly, the quantity (3.2) is nonpositive for p ∈ (0, 1] and nonnegative elsewhere. We
claim that the quantity (3.5) is nonpositive for p ∈ (−∞, 0]∪[1, 2] and nonnegative elsewhere.
In fact, the denominator is always positive. For the numerator we write t = p − 1 and use
the fact that for all x > 0, t tanh(x) − tanh(tx) > 0 for t > 1 and for −1 < t < 0, while the
inequality reverses, and is strict for other values of t except t ∈ {−1, 0, 1}.
To see this, fix x > 0, and define f (t) := t tanh(x) − tanh(tx). Evidently f (t) = 0 for
t = −1, 0, 1. Then, since f(cid:48)(cid:48)(t) = 2x2 sinh(tx)/ cosh3(tx), f(cid:48)(cid:48)(t) > 0 for t > 0, and f(cid:48)(cid:48)(t) < 0
for t < 0. It follows that f (t) > 0 for −1 < t < 0 and t > 1, while f (t) < 0 for 0 < t < 1
and t < −1.
According to (3.6) the products of the signs of (3.2) and (3.5) yield the strict convex-
(cid:3)
ity/concavity properties of H(b) shown in rows 2 to 4 of the table below.
HANNER COMPLEMENT -- December 6, 2018
9
p < 0
0 < p < 1
1 < p < 2
p > 2
db
dx
(cid:0) dH
db
(cid:1)
d
dx
≥ 0
≤ 0
≤ 0
≥ 0
≥ 0
≤ 0
≥ 0
≥ 0
H(b)
concave
concave
concave
convex
p(p − 1)
Direction
≥ 0
≥
≤ 0
≤
≥ 0
≥
≥ 0
≤
Fig. 6: Table of signs determining the direction of the main inequality (1.4).
Lemma 3.4 (constant b(α(x))). For f, g ≥ 0, define α(x) = f (x)/(f (x) + g(x)). Then with
b(α(x)) is almost everywhere constant if max{α(x), 1− α(x)} is constant almost everywhere,
which is true if and only if only if either f and g have essentially disjoint support, or else
f (x) := max{f (x), g(x)} and g(x) := min{f (x), g(x)} are proportional.
Proof. Let F = {x : f (x) > 0} and G = {x : g(x) > 0}. Then α(x) = 1 on F\G, and
1 − α(x) = 1 on G\F . If the measure of F ∩ G is zero, b(α(x)) = 1 almost everywhere with
respect to (f (x) + g(x))dx. Conversely, since b(a) = 1 if and only if a ∈ {0, 1}, if b(α(x)) = 1
almost everywhere, then almost everywhere α(x) ∈ {0, 1}, which means that f and g have
essentially disjoint supports.
For b ∈ [21−p, 1), there is a unique a ∈ [1/2, 1) such that b(a) = b. Therefore, if b(α(x)) =
b ∈ [21−p, 1), there is a unique a ∈ [1/2, 1) such that α(x) ∈ {a, 1 − a} almost everywhere,
and this is the case if and only if max{α(x), 1 − α(x)} = f (x)/( f (x) + g(x)) = a almost
(cid:3)
everywhere.
Proof of Proposition 3.1. Consider the ratio in (2.1). The numerator is the integral(cid:82) H(b(α(x))).
By Jensen's inequality (recalling that(cid:82) 1 = 1) and the convexity/concavity of H in Lemma
3.3, this integral is bounded from below by H(B) in the convex case and from above in the
B :=
(αp + (1 − α)p) .
(3.7)
concave case, where
That is,
(cid:90)
(cid:82) H(b(y))
(3.8)
for p ≥ 2, while the reverse is true for p ≤ 2. Moreover, by the strict convexity/concavity of
H(b), the inequality in (3.8) is strict unless b(α(x)) is constant when p /∈ {0, 1, 2}. By the
first part of Lemma 3.4, b(α(x)) is constant if and only if max{α(x), 1− α(x)} is a constant,
B
≥ H(B)
B
10
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
necessarily belonging to [1/2, 1]. Then, taking into account the signs of 2/p and p − 1 in the
various ranges, (cid:32)
(cid:18) 2(cid:82) H(b(y))dy
(cid:19)2/p(cid:33)p−1
(cid:32)
(cid:18) 2H(B)
(cid:19)2/p(cid:33)p−1
1 +
B
≥
1 +
B
for p ∈ (0, 1]∪ [2,∞), with the reverse in equality for p ∈ (−∞, 0)∪ [1, 2]. The last two rows
in Fig. 6 summarize the interaction of the convexity/concavity properties of H(b) and the
signs of the exponents p/2 and p−1 in the direction of the inequality in (3.9) for the different
ranges of p, and taking into account the cases of equality discussed above, this yields the
result as stated.
Thus, it suffices for us to prove
1 ≤
(cid:32)
1 +
(cid:18) 2H(B)
(cid:19)2/p(cid:33)p−1
B
B
(3.9)
for p ∈ [0, 1] ∪ [2,∞] and the reverse inequality for p /∈ [0, 1] ∪ [2,∞]. We do not know what
the number B is, but that does not matter. In each case the range of b is an interval and,
therefore, the average value B lies in this same interval. Consequently, whatever B might
be, there is a number α such that B = αp + (1 − α)p. (Note that it is not claimed that this
(cid:3)
number α is related in any particular way to the function α(x).)
Proof of Theorem 1.1. This is immediate from Proposition 3.1 and Theorem 1.3.
(cid:3)
4. Part C. Proof of Theorem 1.3
4.1. Proof of the inequality. First we prove the inequality
(cid:32)
1 +
(αp + (1 − α)p)
2αp/2(1 − α)p/2
αp + (1 − α)p
(cid:33)2/pp−1
≥ 1
for all α ∈ (0, 1)
(4.1)
if p ∈ [0, 1] ∪ [2,∞), and the reverse inequality if p ∈ (−∞, 0] ∪ [1, 2].
For p > 0, there is evidently equality for α ∈ {0, 1
2 , 1}, and for p < 0, there is equality
for α = 1/2. Thus for the proof of (4.1) it suffices to consider α ∈ (1/2, 1) for p > 0, and
α ∈ (0, 1/2) if p < 0, and it is convenient to change variables
(cid:18) 1 − α
(cid:19)p ∈ (0, 1]
α
t :=
Moreover, for fixed c we introduce the function
f (t) := − 1
c
ln(1 + tc) + ln(1 + t) +
and
c := 1/p .
(cid:18)
ln
1 +
(cid:18) 4t
(t + 1)2
(cid:19)c(cid:19)
.
1 − c
c
HANNER COMPLEMENT -- December 6, 2018
11
By taking logarithms we see that the claimed inequality (4.1) is equivalent to
f (t) ≥ 0
for t ∈ (0, 1)
if p ∈ [0, 1] ∪ [2,∞) (that is, c ∈ (0, 1/2] ∪ [1,∞)), and the reverse inequality in (4.1) is
equivalent to the reverse inequality if p ∈ (−∞, 0]∪ [1, 2] (that is, c ∈ (−∞, 0]∪ [1/2, 1]). We
shall show that for c > 0 the derivative f(cid:48) has a unique sign change in (0, 1) and it changes
sign from + to − if c ∈ (0, 1/2)∪ (1,∞) and from − to + if c ∈ (1/2, 1). Moreover, for c < 0
we shall show that the derivative f(cid:48) is positive on (0, 1).
Since f (0) = f (1) = 0 for c > 0, this proves that f ≥ 0 if c ∈ (0, 1/2) ∪ (1,∞) and that
f ≤ 0 if c ∈ (1/2, 1). Moreover, since f (1) = 0 for c < 0, this proves that f ≤ 0 if c < 0.
Thus, we have reduced the proof of Theorem 1.3 to proving the above sign change properties
of f(cid:48).
In order to discuss the sign changes of f(cid:48) we compute
.
(4.2)
f(cid:48)(t) =
(1 − c)(1 − t)
t(1 + t)
tc − t
(1 − c)(tc + 1)(1 − t)
−
+ 1
(cid:16) (1+t)2
1
(cid:17)c
4t
Clearly, it suffices to consider the sign changes of the second factor and therefore to consider
the sign changes of
g(t) :=
(cid:18) (1 + t)2
(cid:19)c −
4t
(cid:18) (1 − c)(tc + 1)(1 − t)
(cid:19)
− 1
tc − t
.
(4.3)
We shall show that for c > 0, g has a unique sign change in (0, 1) and it changes sign from
− to + if c ∈ (0, 1/2) and from + to − if c ∈ (1/2,∞). Moreover, for c < 0 we shall show
that g is negative on (0, 1). Clearly, these properties of g imply the claimed properties of f(cid:48)
and therefore will conclude the proof.
We next observe that the second term in (4.3) is positive.
Lemma 4.1. For any c ∈ R \ {1} and t ∈ (0, 1),
(1 − c)(tc + 1)(1 − t)
tc − t
> 1 .
Proof. First, consider the case c ∈ [0, 1). Then concavity of the map t (cid:55)→ tc implies 1 − c +
ct − tc ≥ 0, therefore (1−c)(1−t)
tc−t ≥ 1, and the claim follows from tc + 1 > 1.
Next, for c > 1 the argument is similar using convexity of the map t (cid:55)→ tc.
Finally, for c < 0 convexity of t (cid:55)→ t1−c implies that
(1 − c)(1 + t−c)(1 − t)
(1 − c)(tc + 1)(1 − t)
tc − t
− 1 =
1 − t1−c
− 1 >
(1 − c)(1 − t)
1 − t1−c − 1 ≥ 0 .
This concludes the proof of the lemma.
(cid:3)
12
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
Because of Lemma 4.1, we can define
(cid:18) (1 + t)2
(cid:19)
4t
− ln
(cid:18) (1 − c)(tc + 1)(1 − t)
(cid:19)
− 1
tc − t
h(t) := c ln
.
(4.4)
We shall show that for c > 0, h has a unique sign change in (0, 1) and it changes sign from
− to + if c ∈ (0, 1/2) and from + to − if c ∈ (1/2,∞). Moreover, for c < 0 we shall show
that h is negative on (0, 1). Clearly, these properties of h imply the claimed properties of g
and therefore will conclude the proof.
We will prove this by investigating sign changes of h(cid:48). Namely, we shall show that for
c > 0, h(cid:48) has a unique sign change in (0, 1) and it changes sign from + to − if c ∈ (0, 1/2)
and from − to + if c ∈ (1/2,∞). Moreover, for c < 0 we shall show that h(cid:48) is positive on
(0, 1).
Let us show that this implies the claimed properties of h. Indeed, an elementary limiting
argument shows that
h(0) =
and
−∞
−2c ln 2 − ln(1 − c)
+∞
if c < 0 ,
if c ∈ (0, 1) ,
if c > 1 .
The function −2c ln 2− ln(1− c) is convex on (0, 1) and vanishes at c = 0 and c = 1/2. From
this we conclude that
h(1) = 0
for all c .
h(0) < 0 if c < 1/2 ,
h(0) = 0 if c = 1/2 ,
h(0) > 0 if c > 1/2 .
Because of this behavior of h(0) and h(1), the claimed properties of h(cid:48) imply the claimed
properties of h.
Therefore in order to complete the proof of Theorem 1.3 we need to discuss the sign
changes of h(cid:48). We compute
h(cid:48)(t) =
(1 + t)(tc − t)2(cid:16) (1−c)(tc+1)(1−t)
v(t)
tc−t
(cid:17)
− 1
with
v(t) := t(2c2 − 1) − t2c2 + 2c(1 − 2c)(tc − tc+1) + t2c(1 − 2c2) + (t1+2c − 1)(1 − c)2 + t2c−1c2 .
We shall show that for c > 0, v has a unique sign change in (0, 1) and it changes sign from
+ to − if c ∈ (0, 1/2) and from − to + if c ∈ (1/2,∞). Moreover, for c < 0 we shall show
that v is positive on (0, 1).
Since, by Lemma 4.1 the denominator in the above expression for h(cid:48) is positive, these
properties of v clearly imply those of h(cid:48) and therefore complete the proof of the theorem.
HANNER COMPLEMENT -- December 6, 2018
13
In order to prove the claimed properties of v we shall study the sign changes of v(cid:48)(cid:48). We
shall show that for c > 0, v(cid:48)(cid:48) has a unique sign change in (0, 1) and it changes sign from +
to − if c ∈ (0, 1/2) and from − to + if c ∈ (1/2,∞). Moreover, for c < 0 we shall show that
v(cid:48)(cid:48) is positive.
Let us now argue that these properties of v(cid:48)(cid:48) indeed imply the claimed properties of v.
We compute
v(cid:48)(t) = 2c2 − 1 − 2c2t + 2c(1 − 2c)(ctc−1 − (c + 1)tc) + 2c(1 − 2c2)t2c−1
+ (1 − c)2(1 + 2c)t2c + c2(2c − 1)t2c−2,
v(cid:48)(cid:48)(t) = 2c · [−c + c(1 − 2c)(c − 1)tc−2 − c(1 − 2c)(c + 1)tc−1 + (2c − 1)(1 − 2c2)t2c−2
+ (1 − c)2(1 + 2c)t2c−1 + c(2c − 1)(c − 1)t2c−3],
v(cid:48)(cid:48)(cid:48)(t) = 2c(1 − 2c)(c − 1)tc−3w(t)
(4.5)
(4.6)
and finally
with
w(t) := c(c − 2) − (c + 1)ct − 2tc(1 − 2c2) + (1 − c)(1 + 2c)tc+1 − c(2c − 3)tc−1 .
From these formulas we easily infer that
v(1) = v(cid:48)(1) = v(cid:48)(cid:48)(1) = 0 ,
v(cid:48)(cid:48)(cid:48)(1) = 2c(1 − 2c)(c − 1)2 .
In particular, v(cid:48)(cid:48)(cid:48)(1) > 0 if c ∈ (0, 1/2) and v(cid:48)(cid:48)(cid:48)(1) < 0 if c ∈ (−∞, 0) ∪ (1/2, 1) ∪ (1,∞).
This means that v is convex near t = 1 if c ∈ (−∞, 0) ∪ (1/2,∞) and concave near t = 1 if
c ∈ (0, 1/2).
Let us discuss the behavior near t = 0.
v(0) = +∞, and v(cid:48)(cid:48)(0) > 0. If c > 1/2, then v(0) = −(1 − c)2 and v(cid:48)(cid:48)(0) < 0.
If c < 1/2, then v(t) behaves like t2c−1c2, so
This behavior of v near 0 and 1, together with the claimed sign change properties of
v(cid:48)(cid:48), imply the claimed sign change properties of v and will therefore complete the proof of
Theorem 1.3. This is because, for example, if v is convex near t = 1 with v(1) = v(cid:48)(1) = 0,
and v has a single inflection point t0 ∈ (0, 1), then v is positive on [t0, 0), and v is concave
on (0, t0).
Thus, we are left with studying the sign changes of v(cid:48)(cid:48). In order to do so, we need to
distinguish several cases. For c < 1 we will argue via the sign changes of v(cid:48)(cid:48)(cid:48), while for c > 1
we will argue directly.
Case c ∈ (0, 1). We want to show that v(cid:48)(cid:48) changes sign from + to − if c ∈ (0, 1/2) and
from − to + if c ∈ (1/2, 1).
Since v(cid:48)(cid:48)(0) > 0 if c ∈ (0, 1/2), v(cid:48)(cid:48)(0) < 0 if c ∈ (1/2, 1), v(cid:48)(cid:48)(1) = 0, and v(cid:48)(cid:48)(cid:48)(1) > 0,
it suffices to show that v(cid:48)(cid:48)(cid:48) changes sign only once on (0, 1). Because of (4.6) this is the
14
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
same as showing that w changes sign only once on (0, 1). Notice that w(0) = +∞, and
w(1) = c − 1 < 0. Moreover,
w(cid:48)(cid:48)(t) = c(1 − c)tc−3p(t)
with
p(t) := t2(c + 1)(1 + 2c) + 2t(1 − 2c2) + 2c2 − 7c + 6 .
Indeed, when c ∈ (0, 1/2) this follows from the
The quadratic polynomial p is positive.
fact that all its coefficients are positive. When c ∈ (1/2, 1) we observe that the parabola
p is minimized on R at t = 2c2−1
, which is
positive for c ∈ (1/2, 1).
(c+1)(1+2c) , and its minimal value is (5−3c2)+c(11−8c2)
The fact that p is positive means that w is convex. Since w(0) = +∞ and w(1) < 0, we
(c+1)(1+2c)
conclude that w has only one root.
Case c ∈ (−∞, 0). We want to show that v(cid:48)(cid:48) is positive.
Since v(cid:48)(cid:48)(1) = 0, it suffices to show that v(cid:48)(cid:48)(cid:48) is negative which, by (4.6), is the same as
showing that w is negative. Clearly, w(0) = −∞, w(cid:48)(cid:48)(0) < 0, and w(1) = c − 1 < 0, and
w(cid:48)(1) = (3c− 1)(c− 1) > 0, so it suffices to show that w(cid:48)(cid:48) < 0 on (0, 1). For this it suffices to
show that p > 0 on (0, 1). We have p(0) > 0, and p(1) = 9−4c > 0. Thus if (1+c)(1+2c) ≤ 0
we have proved the claim. Consider the case when (1 + c)(1 + 2c) > 0. The vertex of the
parabola is t0 = 2c2−1
clearly
(c+1)(1+2c) > 1. If c ∈ (−1/2, 0), then
(c+1)(1+2c) . If c < −1 then clearly
2c2−1
2c2−1
(c+1)(1+2c) < 0.
Case c ∈ (1,∞). We want to show that v(cid:48)(cid:48) changes sign from − to +.
We begin with the case c ∈ (1, 2). We write (4.5) as v(cid:48)(cid:48)(t) = 2cq(t) with
q(t) := −c + c(1 − 2c)(c − 1)tc−2 − c(1 − 2c)(c + 1)tc−1 + (2c − 1)(1 − 2c2)t2c−2
+(1 − c)2(1 + 2c)t2c−1 + c(2c − 1)(c − 1)t2c−3.
Clearly q(0) = −∞ and q(1) = 0. It is enough to show that q(cid:48) changes sign from + to −.
We have
q(cid:48)(t) = t2c−4(2c − 1)(c − 1)m(t)
with
m(t) := c(2 − c)t1−c + c(c + 1)t2−c + 2(1 − 2c2)t + (c − 1)(1 + 2c)t2 + c(2c − 3) .
We shall show that m(t) changes sign only once from + to −. Clearly m(0) = +∞ and
m(cid:48)(cid:48)(0) > 0. Next, m(1) = 1 − c < 0, and m(cid:48)(cid:48)(1) = (c − 1)(c2 + 2c + 2) > 0. Thus it suffices
to show m(cid:48)(cid:48) > 0 on (0, 1). Since m(cid:48)(cid:48)(0) > 0, m(cid:48)(cid:48)(1) > 0, then m(cid:48)(cid:48) > 0 will follow from m(cid:48)(cid:48)(cid:48)
having the constant sign. We have
m(cid:48)(cid:48)(cid:48)(t) = t−c−2c2(c − 1)(c − 2)(c + 1)(1 − t) < 0.
HANNER COMPLEMENT -- December 6, 2018
15
This finishes the case c ∈ (1, 2).
If c = 2, then q(t) = (t − 1)(5t2 − 16t + 8), and we see that it changes sign only once.
In what follows we assume c > 2. Let us rewrite (4.5) as v(cid:48)(cid:48)(t) = 2ct2c−3u(t) with
u(t) := −ct3−2c + c(1 − 2c)(c − 1)t1−c − c(1 − 2c)(c + 1)t2−c + (2c − 1)(1 − 2c2)t
+ (1 − c)2(1 + 2c)t2 + c(2c − 1)(c − 1) .
We need to show that u changes sign only once. We have u(0) = −∞, and u(cid:48)(cid:48)(0) < 0. At the
point t = 1, we have u(1) = 0, u(cid:48)(1) = −(2c− 1)(c− 1)2 < 0, u(cid:48)(cid:48)(1) = −2(2c− 1)(c− 1)2 < 0.
It suffices to show that u(cid:48)(cid:48) < 0 on (0, 1). Since u(cid:48)(cid:48)(0) < 0, u(cid:48)(cid:48)(1) < 0, the latter claim will
follow from showing that u(cid:48)(cid:48)(cid:48) has a constant sign. We have
u(cid:48)(cid:48)(cid:48)(t) = t−2−cc(2c − 1)(c − 1)b(t)
with
b(t) := c3 − c − tc(c + 1)(c − 2) + 2t2−c(2c − 3) .
The factor b has the property that b(0) = +∞, b(1) = (c + 6)(c− 1) > 0. On the other hand,
b(cid:48)(t) = −c(c + 1)(c − 2)t1−c
is negative, so b is positive.
(cid:18)
tc−1 +
2(2c − 3)
c(c + 1)
(cid:19)
This concludes the proof of the inequality of Theorem 1.3.
4.2. Sharpness of the exponent 2/p. The sharpnes of the exponent 2/p is easily checked
using the variables introduced in Remark 1.7. If one rpelaces the power of 2/p in (1.6) and
kames the transforations described there, one is led to the function
gr,p(s) := η
1
p−1 (s)
1 +
− 2 .
(4.7)
(cid:32)
(cid:32)
(cid:33)r(cid:33)
1 − s
2
p (s)
η
instead of fp(s). A motivation for this reparametrization is that for fixed p, the function
on the right hand side of (1.6) is equal to 1 up to order O((α − 1/2)4) at α = 1/2. In the
variable s, the leading term in Taylor expansion in s will be second order, and we proves the
sharpness by an expansion at this point.
Proof of the second paragraph of Theorem 1.3. For fixed r > 0, define the function gr,p(s)
by (4.7). By the arithmetic-geometric mean inequality, (1− s)p/2 ≤ η(s) for all p, and hence
(1 − s)/η2/p(s) < 1 for p > 0, while (1 − s)/η2/p(s) > 1 for p < 0. Therefore, for fixed s and
p, gr,p(s) decreases as r increases for p > 0, and does the opposite for p < 0.
A Taylor expansion shows that
gr,p(s) = p(1 − r)s + o(s) .
16
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
It follows that gr,p(s) ≥ 0 on [0, 1] is false (near s = 0) for p > 2 and r > 1, and for p < 0
and r < 1. Likewise, it follows that gr,p(s) ≤ 0 on [0, 1] is false for p ∈ (0, 2) and r < 1.
Since the exponent q in (1.7) corresponds to r(2/p), this together with the reamrks leading
to (1.13) justifies the statements referring to q in Theorem 1.3.
Consideration of the argument shows that for p > 0, p (cid:54)= 1, 2, there is equality if and only
(cid:3)
if α ∈ {0, 1/2, 1} and for p < 0, if and only if α = 1/2.
The proof of Theorem 1.3 is now complete. By what has been explained above, the
inequality of Theorem 1.1 is proved. Concerning the cases of equality, we have seen in Section
3, that for all values of p under consideration, if there is equality then max{α(a), 1 − α(x)}
is constant, and then by what has just been proved here, and in Lemma 3.4 for p > 0, this
constant is either 1, in which case f and g have essentially disjoint support, or 1/2 in which
case = g. For p < 0, there is equality only in case f = g. Finally, it is evident that there is
equality in these cases.
5. Doubling arguments and a generalization to Schatten norms.
5.1. Doubling arguments. We begin this section with a simple proof showing that if the
inequality (1.4) is valid for some p ≥ 2 or some p < 0, then it is also valid for 2. Since
the inequality (1.4) holds as an identity for p = 2, and is simple to prove for p = −1 (see
Remark 1.7), this yields a simple proof of infinitely many cases of the inequality (1.4). The
proof is not only simple and elegant; it applies to certain non-commutative generalizations
of (1.4) for which the reductions in parts A and B of the proof we have just presented are
not applicable, as we discuss.
To introduce the doubling argument we present a direct proof of Theorem 1.1 for p = 4.
Direct proof of Theorem 1.1 for p = 4. Suppose f, g ≥ 0, By homogeneity, we may suppose
that (cid:107)f(cid:107)4
4 = 2. Define
4 + (cid:107)g(cid:107)4
X := f g ,
Y := f 2 + g2 , α := (cid:107)X(cid:107)2
and β := (cid:107)Y (cid:107)2 .
(5.1)
(cid:90)
(cid:90)
By the arithmetic-geometric mean inequality, X ≤ 1
2 Y , and hence
1
2
This yields α ≤ 1 and β ≤ 2. Then (f + g)2 = Y + 2X and hence
(f 4 + g4 + 2f 2g2)dµ =
X 2dµ ≤ 1
4
Y 2dµ =
+
(cid:90)
1
4
(cid:90)
1
2
X 2dµ .
It suffices to prove that β + 2α ≤ 21/2(1 + α)3/2. Note that β2 =(cid:82) (f 2 + g2)2dµ = 2 + 2α2,
4 = (cid:107)Y + 2X(cid:107)2 ≤ (cid:107)Y (cid:107)2 + 2(cid:107)X(cid:107)2 = β + 2α .
(cid:107)f + g(cid:107)2
(5.2)
and then since α ∈ [0, 1]. Thus it suffices to show that
(1 + α2)1/2 ≤ (1 + α)3/2 − 21/2α for all
0 ≤ α ≤ 1 .
(5.3)
HANNER COMPLEMENT -- December 6, 2018
17
Squaring both sides, this is equivalent to 1+α2 ≤ (1+α)3+2α2−23/2α(1+α)3/2. This reduces
to 23/2(1 + α)3/2 ≤ 3 + 4α + α2. Squaring both sides again, this reduces to (α2 − 1)2 ≥ 0,
(cid:3)
completing the proof.
What made this proof work is the fact that the inequality holds for p = 2 -- as an identity,
but that is unimportant. Then, using Minkowski's inequality, as in (5.2), together with the
numerical inequality (5.3) we arrive at the inequality for p = 4. This is a first instance of
the general doubling proposition, to be proved next. The inequality (5.3) is s special case of
the general inequality (5.4) proved below.
This strategy can be adapted to give direct proof of the inequality for other integer values
expansion of (f + g)p = f p + gp + mixed terms. Under the assumption that(cid:82) (f p + gp) = 2,
of p; e.g., p = 3. When p is an integer, and f and g are non-negative, one has the binomial
one is left with estimating the mixed terms, and one can use Holder for this. When p is not
an integer, there is no useful expression for (f + g)p − f p − gp.
Proposition 5.1 (A "doubling" argument). Suppose that for some p ≥ 2, (1.4) is valid for
all f, g ≥ 0. Then (1.4) is valid with p replaced by 2p for all f, g ≥ 0. Likewise, if for some
p < 0 the reverse of (1.4) is valid for all f, g > 0, then the reverse of (1.4) is valid with p
replaced by 2p for all f, g > 0.
The proof of Proposition 5.1 relies on the following lemma.
Lemma 5.2. For t ∈ R, define ψt on [0,∞) by
ψt(α) = (1 + α)1+t − (1 + α2)t − 2tα .
(5.4)
Then for t ∈ [0, 1], ψt(α) ≥ 0 on [0,∞), while for t > 1, ψt(α) ≤ 0 on [0,∞).
Proof. We write ψt(α) = (1 + α)t − (1 + α2)t − (2t − (1 + α)t)α. Therefore,
ψt(α)
α(1 − α)
=
(1 + α)t − (1 + α2)t
α(1 − α)
− 2t − (1 + α)t
1 − α
.
Defining a := 1 + α2, b := 1 + α and c := 2, and defining ϕ(α) := xt, the right hand side is
the same as
ϕ(b) − ϕ(a)
b − a
− ϕ(c) − ϕ(b)
c − b
.
For α ∈ [0, 1) we have a < b < c and therefore this quantity is positive when ϕ is concave,
and negative when ϕ is convex. For α ∈ (1,∞) we have a > b > c and therefore this quantity
(cid:3)
is negative when ϕ is concave, and positive when ϕ is convex.
Proof of Proposition 5.1. Let f, g ∈ L2p with (cid:107)f(cid:107)2p
2p + (cid:107)g(cid:107)2p
Y := f 2 + g2, and γ := (cid:107)X(cid:107)p and β := (cid:107)Y (cid:107)p. By the triangle inequality we have
2p = 2. Define X := f g and
(cid:107)f + g(cid:107)2
2p = (cid:107)Y + 2X(cid:107)p
≤ (cid:107)Y (cid:107)p + 2(cid:107)X(cid:107)p = β + 2γ
≥ (cid:107)Y (cid:107)p + 2(cid:107)X(cid:107)p = β + 2γ
if p ≥ 2 ,
if p < 0 .
18
ERIC A. CARLEN, RUPERT L. FRANK, PAATA IVANISVILI, AND ELLIOTT H. LIEB
(Note that the triangle inequality reverses for p < 0.) We now use the assumption that the
inequality (1.4) is valid for p. Applying the inequality with exponent p to the functions f 2
and g2, which satisfy (cid:107)f 2(cid:107)p
p = (cid:107)f(cid:107)2p
p + (cid:107)g2(cid:107)p
βp = (cid:107)f 2 + g2(cid:107)p
and similarly βp ≥ 2(cid:0)1 + γ2(cid:1)p−1 for p < 0. To summarize, we have shown that
2p + (cid:107)g(cid:107)2p
= 2(cid:0)1 + γ2(cid:1)p−1
p ≤ 2(cid:0)1 + (cid:107)f 2g2(cid:107)p/2
(cid:1)p−1
2p = 2, we obtain for p ≥ 2,
≤ 21/p(1 + γ2)1−1/p + 2γ
≥ 21/p(1 + γ2)1−1/p + 2γ
if p ≥ 2 ,
if p < 0 .
(cid:107)f + g(cid:107)2
2p
According to Lemma 5.2 (with t = 1− 1/p and α = γ) this is bounded from above for p ≥ 2
and from below for p < 0 by 21/p(1 + γ)2−1/p = 21/p(1 + (cid:107)f g(cid:107)p)2−1/p, which is the claimed
(cid:3)
inequality.
5.2. A generalization to Schatten norms. For p ∈ [1,∞), an operator A on some Hilbert
space belongs the Schatten p-class Sp in case (A∗A)p/2 is trace class, and the Schatten p
norm on Sp is defined by (cid:107)A(cid:107)p = (Tr[(A∗A)p/2])1/p. One possible non-commutative analog
of (part of) Theorem 1.1 would assert that for positive A, B ∈ Sp, p > 2.
(cid:32)
1 +
(cid:33)2/pp−1
Tr(A + B)p ≤
Tr[Bp/4Ap/2Bp/4]
2(cid:107)A(cid:107)p
2(cid:107)B(cid:107)p
1
p + 1
p
Tr ( Ap + Bp ) .
(5.5)
Note that for p = 2, (5.5) holds as an identity.
In this setting, it is not clear how to implement analogs of Parts A and B of our proof for
functions. However, the direct proofs sketched at the beginning of this section do allow us
to prove the valididty of (5.5) for all p = 2k, k ∈ N.
Theorem 5.3. If (5.5) is valid for some p ≥ 2 and all positive A, B ∈ Sp, then it is valid
for 2p and all A, B ∈ S2p. In particular, since (5.5) holds as an identity for p = 2, it is valid
for p = 2k for all k ∈ N.
Proof. Let A and B be positive operators in S2p, and assume that (cid:107)A(cid:107)2p
by homogeneity, entails no loss of generality. Define
2p +(cid:107)B(cid:107)2p
2p = 2, which,
Note that
X :=
1
2
(AB + BA)
and
Y = A2 + B2 .
(cid:107)X(cid:107)p ≤ 1
2
((cid:107)AB(cid:107)p + (cid:107)BA(cid:107)p) .
By definition, the Lieb -- Thirring inequality [7], and cyclicity of the trace,
(cid:107)AB(cid:107)p
p = Tr[(BA2B)p/2] ≤ Tr[Bp/2ApBp/2] = Tr[Ap/2BpAp/2] .
Define
β := (cid:107)Y (cid:107)p
and γ := (Tr[Bp/2ApBp/2])1/p .
HANNER COMPLEMENT -- December 6, 2018
19
2p = (cid:107)Y + 2X(cid:107)p ≤ (cid:107)Y (cid:107)p + 2(cid:107)X(cid:107)p ≤ β + 2γ. Since (cid:107)A2(cid:107)p
p + (cid:107)B2(cid:107)p
p = 2,
Therefore, (cid:107)A + B(cid:107)2
we can apply (5.5) to deduce that
βp = (cid:107)A2 + B2(cid:107)p
p ≤ 2
(cid:16)
1 + (Tr[B2p/4ApB2p/4])2/p(cid:17)p−1
= 2(cid:0)1 + γ2(cid:1)p−1
.
Altogether
(cid:107)A + B(cid:107)2
2p ≤ 21/p(1 + γ2)1−1/p + 2γ
and, by Lemma 5.2, the right side is bounded above by 21/p(1 + γ)2−1/p, which proves the
(cid:3)
inequality,
Acknowledgement We thank Anthony Carbery for useful correspondence.
References
[1] J. A. Clarkson, Uniformly convex spaces. Trans. Amer. Math. Soc., 40 (1936), 396 -- 414.
[2] K. Ball, E. A. Carlen, E. H. Lieb, Sharp uniform convexity and smoothness inequalities for trace norms.
Invent. Math. 115 (1994), no. 3, 463 -- 482.
[3] A. Carbery, Almost orthogonality in the Schatten -- von Neumann classes. J. Operator Theory 62 (2009),
no. 1, 151 -- 158.
[4] O. Hanner, On the uniform convexity of Lp and (cid:96)p. Ark. Math. 3 (1956), 239 -- 244.
[5] G. Hardy, J. E. Littlewood and G. Polya, Inequalities Cambridge Univ. Press, Cambridge, 1934.
[6] E. H. Lieb and M. Loss, Analysis, second edition, American Mathematical Society (2014).
[7] E. H. Lieb and W. Thirring, Inequalities for the moments of the eigenvalues of the Schrodinger hamilton-
ian and their relation to Sobolev inequalities, in Studies in Mathematical Physics, E.H. Lieb, B. Simon,
A. Wightman eds., Princeton University Press, 269-303 (1976).
(Eric A. Carlen) Department of Mathematics, Hill Center, Rutgers University, 110 Frel-
inghuysen Road, Piscataway NJ 08854-8019, USA
E-mail address: [email protected]
(Rupert L. Frank) Mathematisches Institut, Ludwig-Maximilans Universitat Munchen, There-
sienstr. 39, 80333 Munchen, Germany, and Mathematics 253-37, Caltech, Pasadena, CA 91125,
USA
E-mail address: [email protected]
(Paata Ivanisvili) Department of Mathematics, University of California, Irvine, CA 92617,
USA
E-mail address: [email protected]
(Elliott H. Lieb) Departments of Mathematics and Physics, Princeton University, Princeton,
NJ 08544, USA
E-mail address: [email protected]
|
1705.04342 | 2 | 1705 | 2017-09-21T19:02:06 | An invertibility criterion in a C*-algebra acting on the Hardy space with applications to composition operators | [
"math.FA",
"math.CV",
"math.OA"
] | In this paper we prove an invertibility criterion for certain operators which is given as a linear algebraic combination of Toeplitz operators and Fourier multipliers acting on the Hardy space of the unit disc. Very similar to the case of Toeplitz operators we prove that such operators are invertible if and only if they are Fredholm and their Fredholm index is zero. As an application we prove that for "quasi-parabolic" composition operators the spectra and the essential spectra are equal. | math.FA | math |
AN INVERTIBILITY CRITERION IN A C*-ALGEBRA ACTING
ON THE HARDY SPACE WITH APPLICATIONS TO
COMPOSITION OPERATORS
U GUR G UL AND BEYAZ BAS¸AK KOCA
Abstract. In this paper we prove an invertibility criterion for certain opera-
tors which is given as a linear algebraic combination of Toeplitz operators and
Fourier multipliers acting on the Hardy space of the unit disc. Very similar to
the case of Toeplitz operators we prove that such operators are invertible if and
only if they are Fredholm and their Fredholm index is zero. As an application
we prove that for "quasi-parabolic" composition operators the spectra and the
essential spectra are equal.
1. introduction
In this paper we investigate the invertibility of elements in the C*-algebra Ψ gen-
erated by Toeplitz operators and Fourier multipliers. The C*-algebra Ψ is defined
to be
Ψ = Ψ(QC, C([0,∞])) = C ∗({Tϕ : ϕ ∈ QC} ∪ {Dϑ : ϑ ∈ C([0,∞])})
the C*-algebra generated by Toeplitz operators with QC symbols and Fourier multi-
pliers with continuous symbols. This C*-algebra was introduced by the first author
in [3] in order to study the spectral properties of a class of composition operators.
In [3], the first author showed that Ψ/K(H 2) is a commutative C*-algebra with
identity and determined its maximal ideal space. The maximal ideal space M of
Ψ/K(H 2) is found to be homeomorphic to a certain subset of M (QC) × [0,∞]
which can be described as
M ∼= (M∞(QC(R)) × [0,∞]) ∪ (M (QC(R)) × {∞})
where M∞(QC) is the fiber of M (QC) at infinity.
In this paper we show that if T = P Tϕj Dϑj +P Dνj Tψj ∈ Ψ is written as a
finite sum or as an infinite sum converging in the operator norm where ψj, ϕj ∈ QC
and νj, ϑj ∈ C([0,∞]) then T is invertible if and only if T is Fredholm and has
Fredholm index zero. We do this through constructing a homotopy H : [0, 1] → Ψ
which is defined as
H(w) :=X Tϕj Dϑw
j +X Dνw
j Tψj
where νw
j (t) := νj(t − ln w), ϑw
j (t) := ϑj(t − ln w) and H(0) := Tϕ where ϕ :=
P λjϕj +P µjψj with λj := limt→∞ ϑj(t) and µj = limt→∞ νj(t). We observe that
this homotopy acts continuously on finite sums hence keeps on to act continuously
on infinite sums which converge in operator norm. We apply this result to show
2000 Mathematics Subject Classification. 47B33.
Key words and phrases. Composition Operators, Hardy Spaces, Essential Spectra.
Submitted May 2, 2017.
1
2
U. G UL, B. KOCA
that the class of composition operators that the first author studied in [3] have
spectra equal to their essential spectra. The class of composition operators that
was studied in [3] is the class of composition operators with symbols ϕ which have
upper half-plane re-incarnation
C−1 ◦ ϕ ◦ C(z) = z + ψ(z)
for a bounded analytic function ψ satisfying ℑ(ψ(z)) > ǫ > 0 for all z ∈ H. We call
this class of composition operators "quasi-parabolic".
2. preliminaries
In this section we fix the notation that we will use throughout and recall some
preliminary facts that will be used in the sequel.
Let S be a compact Hausdorff topological space. The space of all complex valued
continuous functions on S will be denoted by C(S). For any f ∈ C(S), k f k∞ will
denote the sup-norm of f , i.e.
k f k∞= sup{ f (s) : s ∈ S}.
For a Banach space X, K(X) will denote the space of all compact operators on X
and B(X) will denote the space of all bounded linear operators on X. The open
unit disc will be denoted by D, the open upper half-plane will be denoted by H, the
real line will be denoted by R and the complex plane will be denoted by C. The
one point compactification of R will be denoted by R which is homeomorphic to T.
For any z ∈ C, ℜ(z) will denote the real part, and ℑ(z) will denote the imaginary
part of z, respectively. For any subset S ⊂ B(H), where H is a Hilbert space, the
C*-algebra generated by S will be denoted by C ∗(S). The Cayley transform C will
be defined by
C(z) =
z − i
z + i
.
For any a ∈ L∞(R) (or a ∈ L∞(T)), Ma will be the multiplication operator on
L2(R) (or L2(T)) defined as
Ma(f )(x) = a(x)f (x).
For convenience, we remind the reader of the rudiments of the theory of Toeplitz
operators and commutative C*-algebras.
Let A be a commutative Banach algebra. Then its maximal ideal space M (A)
is defined as
M (A) = {x ∈ A∗ : x(ab) = x(a)x(b) ∀a, b ∈ A}
where A∗ is the dual space of A.
If A has identity then M (A) is a compact
Hausdorff topological space with the weak* topology. The Gelfand transform
Γ : A → C(M (A)) is defined as
Γ(a)(x) = x(a).
If A is a commutative C*-algebra with identity, then Γ is an isometric *-isomorphism
between A and C(M (A)). If A is a C*-algebra and I is a two-sided closed ideal of
A, then the quotient algebra A/I is also a C*-algebra (see [6] and [5]).
For a Banach algebra A, we denote by com(A) the closed ideal in A generated by
the commutators {a1a2−a2a1 : a1, a2 ∈ A}. It is an algebraic fact that the quotient
3
algebra A/com(A) is a commutative Banach algebra. For a ∈ A the spectrum σA(a)
of a on A is defined as
σA(a) = {λ ∈ C : λe − a is not invertible in A},
where e is the identity of A. In particular the spectrum σ(T ) of a linear bounded
operator T : X → X where X is a Banach space is defined as σ(T ) := σB(X)(T ).
Recall that a bounded linear operator T on a Hilbert space H is called Fredholm
if the range of T is closed, dim ker(T ) and dim ker(T ∗) are finite. The Fredholm
index ind is defined as
ind(T ) = dim(ker(T )) − dim(ker(T ∗))
It is a very well known fact that ([5]) when the set of Fredholm operators F ⊂
B(H) is equipped with operator norm topology and Z is equipped with discrete
topology, the index function ind : F → Z is continuous. The essential spectrum
σe(T ) of an operator T acting on a Banach space X is the spectrum of the coset
of T in the Calkin algebra B(X)/K(X), the algebra of bounded linear operators
modulo compact operators. The following Atkinson's characterization for Fredholm
operators is also well known:
Theorem 1. [5, p.28, Theorem 1.4.16] A bounded linear operator T on a Hilbert
space H is Fredholm if and only if T + K(H) is invertible in the quotient algebra
B(H)/K(H), where K(H) is the algebra of all compact operators on H.
the Hardy space of the upper half-plane will be denoted by H p(H).
For 1 ≤ p < ∞ the Hardy space of the unit disc will be denoted by H p(D) and
The two Hardy spaces H 2(D) and H 2(H) are isometrically isomorphic. An iso-
metric isomorphism Φ : H 2(D) −→ H 2(H) is given by
z + i(cid:19)
√π(z + i)(cid:19)g(cid:18) z − i
Φ(g)(z) =(cid:18)
1
(1)
The mapping Φ has an inverse Φ−1 : H 2(H) −→ H 2(D) given by
Φ−1(f )(z) =
e
1
2
iπ
2 (4π)
(1 − z)
1 − z (cid:19)
f(cid:18) i(1 + z)
Using the isometric isomorphism Φ, one may transfer Fatou's theorem in the
unit disc case to upper half-plane and may embed H 2(H) in L2(R) via f −→ f ∗
where f ∗(x) = limy→0 f (x + iy). This embedding is an isometry.
Throughout the paper, using Φ, we will go back and forth between H 2(D) and
H 2(H). We use the property that Φ preserves spectra, compactness and essential
spectra i.e. if T ∈ B(H 2(D)) then
σB(H2(D))(T ) = σB(H2(H))(Φ ◦ T ◦ Φ−1),
K ∈ K(H 2(D)) if and only if Φ ◦ K ◦ Φ−1 ∈ K(H 2(H)) and hence we have
(2)
We also note that T ∈ B(H 2(D)) is essentially normal if and only if Φ ◦ T ◦ Φ−1 ∈
B(H 2(H)) is essentially normal.
σe(T ) = σe(Φ ◦ T ◦ Φ−1).
The Toeplitz operator with symbol a is defined as
Ta = P MaH2 ,
4
U. G UL, B. KOCA
where P denotes the orthogonal projection of L2 onto H 2. A good reference about
Toeplitz operators on H 2 is Douglas' treatise ([2]). Although the Toeplitz operators
treated in [2] act on the Hardy space of the unit disc, the results can be transfered to
the upper half-plane case using the isometric isomorphism Φ introduced by equation
(1). In the sequel the following identity will be used:
Φ−1 ◦ Ta ◦ Φ = Ta◦C−1,
(3)
where a ∈ L∞(R). We also employ the fact
(4)
for any a ∈ L∞(R), which is a consequence of Theorem 7.11 of [2] (pp. 160–161) and
equation (3). For any subalgebra A ⊆ L∞(R) the Toeplitz C*-algebra generated
by symbols in A is defined to be
k Ta ke=k Ta k=k a k∞
It is a well-known result of Sarason (see [7]) that the set of functions
T (A) = C ∗({Ta : a ∈ A}).
H ∞ + C = {f1 + f2 : f1 ∈ H ∞(D), f2 ∈ C(T)}
is a closed subalgebra of L∞(T). The following theorem of Douglas [2] will be used
in the sequel.
Theorem 2 (Douglas' Theorem). Let a,b ∈ H ∞+C then the semi-commutators
Tab − TaTb ∈ K(H 2(D)),
Tab − TbTa ∈ K(H 2(D)),
and hence the commutator
is compact.
[Ta, Tb] = TaTb − TbTa ∈ K(H 2(D))
Let QC be the C*-algebra of functions in H ∞ + C whose complex conjugates
also belong to H ∞ + C. Let us also define the upper half-plane version of QC as
the following:
QC(R) = {ϕ ∈ L∞(R) : ϕ ◦ C−1 ∈ QC}.
Going back and forth with Cayley transform one can deduce that QC(R) is a closed
subalgebra of L∞(R).
By Douglas' theorem and equation (3), if a, b ∈ QC(R), then
TaTb − Tab ∈ K(H 2(H)).
Let scom(QC(R)) be the closed ideal in T (QC(R)) generated by the semi-commutators
{TaTb − Tab : a, b ∈ QC(R)}. Then we have
com(T (QC(R))) ⊆ scom(QC(R)) ⊆ K(H 2(H)).
By Proposition 7.12 of [2] and equation (3) we have
com(T (QC(R))) = scom(QC(R)) = K(H 2(H)).
Now consider the symbol map
(5)
defined as Σ(a) = Ta. This map is linear but not necessarily multiplicative; however
if we let q be the quotient map
Σ : QC(R) → T (QC(R))
q : T (QC(R)) → T (QC(R))/scom(QC(R)),
5
then q ◦ Σ is multiplicative; moreover by equations (4) and (5), we conclude that
q ◦ Σ is an isometric *-isomorphism from QC(R) onto T (QC(R))/K(H 2(H)). The
maximal ideal space M (QC(R)) is fibered over R in the following way: For any
x ∈ M (QC(R)) consider x = xC( R) then x ∈ M (C( R)) = R. Hence M (QC(R)) is
fibered over R, i.e.
where
M (QC(R)) = [t∈ R
Mt(QC),
Mt(QC) = {x ∈ M (QC(R)) : x = xC( R) = δt, δt(f ) = f (t)}.
We also remind the reader about the very important fact in the theory of Toeplitz
operators that any Toeplitz operator Tϕ with symbol ϕ ∈ L∞ is invertible if and
only if Tϕ is Fredholm and the Fredholm index ind(Tϕ) = 0 is zero. The proof of
this fact can also be found in [2]. This fact will be used in the proof of our main
result in this paper.
Let ϕ : D −→ D or ϕ : H −→ H be a holomorphic self-map of the unit disc or the
upper half-plane. The composition operator Cϕ on H p(D) or H p(H) with symbol
ϕ is defined by
Cϕ(g)(z) = g(ϕ(z)),
z ∈ D or
z ∈ H.
Composition operators of the unit disc are always bounded [1] whereas composition
operators of the upper half-plane are not always bounded. For the boundedness
problem of composition operators of the upper half-plane see [4]. The composition
operator Cϕ on H 2(D) is carried over to ( ϕ(z)+i
z+i )C ϕ on H 2(H) through Φ, where
ϕ = C ◦ ϕ ◦ C−1, i.e. we have
ΦCϕΦ−1 = T( ϕ(z)+i
z+i )C ϕ.
(6)
However this gives us the boundedness of Cϕ : H 2(H) → H 2(H) for
ϕ(z) = pz + ψ(z),
where p > 0, ψ ∈ H ∞ and ℑ(ψ(z)) > ǫ > 0 for all z ∈ H:
have
Let ϕ : D → D be an analytic self-map of D such that ϕ = C−1 ◦ ϕ ◦ C, then we
where
If
ΦC ϕΦ−1 = Tτ Cϕ
τ (z) =
ϕ(z) + i
z + i
.
ϕ(z) = pz + ψ(z)
with p > 0, ψ ∈ H ∞ and ℑ(ψ(z)) > ǫ > 0, then T 1
ΦC ϕΦ−1 is always bounded we conclude that Cϕ is bounded on H 2(H).
is a bounded operator. Since
τ
The Fourier transform F f of f ∈ S(R) (the Schwartz space, for a definition see
[6]) is defined by
The Fourier transform extends to an invertible isometry from L2(R) onto itself with
inverse
(F f )(t) =
1
√2π Z +∞
−∞
e−itxf (x)dx.
(F −1f )(t) =
1
√2π Z +∞
−∞
eitxf (x)dx.
6
U. G UL, B. KOCA
The following is a consequence of a theorem due to Paley and Wiener [6]. Let
1 < p < ∞. For f ∈ Lp(R), the following assertions are equivalent:
(i) f ∈ H p,
(ii) supp( f ) ⊆ [0,∞)
A reformulation of the Paley-Wiener theorem says that the image of H 2(H)
under the Fourier transform is L2((0,∞)).
By the Paley-Wiener theorem we observe that the operator
Dϑ = F −1MϑF
for ϑ ∈ C([0,∞]) maps H 2(H) into itself, where C([0,∞]) denotes the set of con-
tinuous functions on [0,∞) which have limits at infinity. Since F is unitary we also
observe that
Let F be defined as
k Dϑ k=k Mϑ k=k ϑ k∞
F = {Dϑ ∈ B(H 2(H)) : ϑ ∈ C([0,∞])}.
We observe that F is a commutative C*-algebra with identity and the map D :
C([0,∞]) → F given by
D(ϑ) = Dϑ
for a.e.
is an isometric *-isomorphism by equation above. Hence F is isometrically *-
isomorphic to C([0,∞]). The operator Dϑ is usually called a "Fourier Multiplier."
We will also need the fact that, under the Fourier transform the Beurling type
invariant subspace eiηxH 2(H), for η > 0, is mapped onto L2
η((0,∞)) := {f ∈
L2(0,∞) : f (t) = 0
t ∈ (0, η)} for all η > 0 and F TeiηxF −1 = Sη where
Sη : L2((0,∞)) → L2((0,∞)) is defined as Sηf (t) := f (t−η) if t ≥ η and Sηf (t) = 0
if 0 ≤ t < η. Similarly we have F Te−iηxF −1 = S ∗
η f (t) = f (t + η). Here
Teiηx and Te−iηx are Toeplitz operators with symbols eiηz and e−iηz respectively.
η((0,∞)) is dense in L2((0,∞)), ∪η>0eiηxH 2(H) is also dense in H 2(H).
Since ∪η>0L2
In [3] the first author studied the C*-algebra Ψ generated by Toeplitz operators
with QC symbols and Fourier multipliers with continuous symbols. He proved that
the commutator [Tϕ, Dϑ] = TϕDϑ − DϑTϕ ∈ K(H 2) of any Toeplitz operator with
QC symbol and a Fourier multiplier with continuous symbol is compact which im-
plies that the Calkin algebra Ψ/K(H 2) is a commutative C*-algebra with identity.
The maximal ideal space M of Ψ/K(H 2) is also studied in [3] and is found to be
η where S ∗
M ∼= (M∞(QC(R)) × [0,∞]) ∪ (M (QC(R)) × {∞}) ⊂ M (QC) × [0,∞]
where M∞(QC) := {x ∈ M (QC) : xC( R) = δ∞, δ∞(f ) = limt→∞ f (t)} is the fiber
of M (QC) at infinity. The Gelfand transform Γ of Ψ/K(H 2) looks like
Γ(cid:16)hX Tϕj Dϑji(cid:17) (x, t) =(P ϕj (x) ϑj (t)
P ϕj (x) ϑj (∞)
3. the main result
if x ∈ M∞(QC(R))
if
t = ∞
In this section we prove the main result of this paper which asserts that any sum
T = P Tϕj Dϑj +P Dνj Tψj ∈ Ψ convergent in the operator norm is invertible if
and only if T is Fredholm and has Fredholm index zero. This may be regarded as
a generalization of the fact that any Toeplitz operator Tϕ with a bounded symbol
ϕ ∈ L∞ is invertible if and only if Tϕ is Fredholm and has Fredholm index zero. In
7
proving this fact our main technical tool will be a homotopy H : [0, 1] → Ψ which
carries T to a Toeplitz operator. The homotopy H for T =P Tϕj Dϑj +P Dνj Tψj
is defined as follows:
H(w) :=X Tϕj Dϑw
j (t) = ϑj(t − ln w), νw
j +X Dνw
where w ∈ (0, 1], ϑw
j (t) = νj(t − ln w) and for w = 0,
H(0) := Tϕ, ϕ = P λjϕj +P µjψj and λj = limt→∞ ϑj(t), µj = limt→∞ νj(t).
It is easily seen that when T = P Tϕj Dϑj +P Dνj Tψj consists of finite sums, H
is continuous. And thus H keeps on being continuous when T consists of infinite
sums which are both convergent in operator norm. In the proof of this fact we will
always work with finite sums since it is enough to prove it for finite sums. In this
section, unless otherwise stated, H 2 will always be understood as H 2(H) and QC
will always be understood as QC(R).
j Tψj
Here is our main theorem:
Theorem 3. Let T =P Tϕj Dϑj +P Dνj Tψj ∈ Ψ be such that νj, ϑj ∈ C([0,∞]),
ψj, ϕj ∈ QC. Then
T ∈ Ψ is invertible ⇔ T is Fredholm and ind(T ) = 0
Proof. (⇒): trivial.
(⇐): Let T be Fredholm and ind(T ) = 0. Let H : [0, 1] → Ψ be the homotopy con-
structed above. Since σe(T ) = {P ϕj (x)ϑj (t) +P ψj(x)νj (t) : x ∈ M∞(QC), t ∈
[0,∞]}∪{P λj ϕj(x)+P µj ψj(x) : x ∈ M (QC)} and σe(H(w)) = {P ϕj(x)ϑj (t)+
P ψj(x)νj (t) : x ∈ M∞(QC), t ∈ [− ln w,∞]} ∪ {P λj ϕj(x) +P µj ψj(x) : x ∈
M (QC)}, we have σe(H(w)) ⊆ σe(T ) for all w ∈ [0, 1]. Hence if T is Fredholm
then 0 6∈ σe(T ) ⇒ 0 6∈ σe(H(w)) for all w ∈ [0, 1] which implies that H(w) is
Fredholm for all w ∈ [0, 1]. Since H is continuous and ind is continuous on the set
of Fredholm operators, we have
ind(H(0)) = ind(H(w)) = ind(H(1)) = ind(T ) = 0 ∀w ∈ [0, 1].
Hence ind(H(0) = ind(Tϕ) = 0 which implies that Tϕ is invertible since any
Toeplitz operator with L∞ symbol is invertible iff it is Fredholm with Fredholm
index zero. Since invertible elements in Ψ form an open subset and H is continu-
ous, there is a w0 ∈ (0, 1] such that H(w) is invertible for all 0 ≤ w < w0. Now
suppose that H(w0) is not invertible.
Sη where
Mg : L2((0,∞)) → L2((0,∞)), Mgf (t) := g(t)f (t) is the multiplication operator.
Hence we have
For η = ln(w0) − ln(w), where w0 > w, we have Mϑw
j = S ∗
η Mϑ
w0
j
Dϑw
j
= (F −1S ∗
ηF )Dϑ
w0
j
(F −1SηF ) = Te−iηx Dϑ
w0
j
Teiηx .
Since ψj, ϕj ∈ QC we have Tϕj Te−iηx − Te−iηxTϕj ∈ K(H 2) and Tψj Teiηx −
TeiηxTψj ∈ K(H 2) for all η > 0(see [2]). Hence we have
j Tψj =
j +X Dνw
Teiηx +X Te−iηx Dν
+X Dν
= Te−iηxH(w0)Teiηx + K(w, w0).
where K(w, w0) ∈ K(H 2) is a compact operator.
H(w) =X Tϕj Dϑw
X Tϕj Te−iηx Dϑ
= Te−iηx(X Tϕj Dϑ
w0
j
w0
j
w0
j
Teiηx Tψj
w0
j
Tψj )Teiηx + K(w, w0)
8
U. G UL, B. KOCA
Hence we have
H(w) = Te−iηxH(w0)Teiηx + K(w, w0)
for some compact operator K(w, w0) ∈ K(H 2). Since Fredholm index is stable
under compact perturbations this implies that the operator Te−iηx H(w0)Teiηx is
Fredholm with index 0 for all η > 0. Hence if H(w0) is non-invertible then
ker(Te−iηxH(w0)Teiηx ) 6= {0} for some η > 0 since Te−iηx H(w0)Teiηx is Fredholm
with index 0. If this is not the case i.e. if we have ker(Te−iηx H(w0)Teiηx ) = {0} for
all η > 0, since Te−iηx H(w0)Teiηx is Fredholm with index 0, Te−iηx H(w0)Teiηx is
invertible for all η > 0. And this implies that H(w0) maps each eiηxH 2 onto itself
in a one to one manner. Hence H(w0) maps ∪η>0eiηxH 2 onto itself in a one to one
manner. Since H(w0) is Fredholm, ran(H(w0)) ⊆ H 2 is closed in H 2 and we have
∪η>0eiηxH 2 ⊆ ran(H(w0)). Since ∪η>0eiηxH 2 is dense in H 2 and ind(H(w0)) = 0,
this implies that ran(H(w0)) = H 2 which in turn implies that ker(H(w0)) = {0}
and hence H(w0) is invertible which is a contradiction. Hence we should have
ker(Te−iηxH(w0)Teiηx ) 6= {0} for some η > 0. This implies that either ker(H(w0))∩
(eiηxH 2) 6= {0} or H(w0)(eiηxH 2) ∩ ker(Te−iηx ) 6= {0} should hold for some η > 0.
Now suppose that ψj, ϕj ∈ H ∞ for all j ∈ N. Since ψj, ϕj ∈ H ∞ for all j, we have
H(w0)(eiηxH 2) ⊆ (eiηxH 2) for all η > 0 and ker(Te−iηx ) = (eiηxH 2)⊥, we have
H(w0)(eiηxH 2) ∩ ker(Te−iηx ) = {0} for all η > 0. So if H(w0) is non-invertible we
should have ker(H(w0))∩ (eiηxH 2) 6= {0}. Since H(w0) is Fredholm, ker(H(w0)) is
finite dimensional an η0 := max{η > 0 : ker(H(w0)) ∩ (eiηxH 2) 6= {0}} exists. For
such η0 > 0, let H0 := Te−iη0xH(w0)Teiη0 x. Then since ker(H(w0)) ∩ (eiη0xH 2) 6=
{0}, H0 is non-invertible. But since ker(H(w0)) ∩ (ei(η0+δ)xH 2) = {0} for all δ > 0
and H(w0)(ei(η0+δ)xH 2) ⊆ ei(η0+δ)xH 2 we have ker(Te−i(η0 +δ)x H(w0)Tei(η0 +δ)x ) =
ker(Te−iδx H0Teiδx ) = {0} and this implies that Te−iδx H0Teiδx is invertible for all
δ > 0. This again implies that H0 maps eiδxH 2 onto itself in a one to one man-
ner which implies that H0 should be invertible. This contradicts our assumption
that H(w0) is non-invertible. Hence H(w0) should be invertible. Therefore, in
this case(where ψj, ϕj ∈ H ∞ for all j), if H(w) is invertible for all 0 ≤ w < w0
then H(w0) is also invertible. So by transfinite induction H(w) is invertible for all
w ∈ [0, 1] and in particular H(1) = T is invertible.
Now suppose that ϕj and ψj are continuous for all j. Since trigonometric poly-
nomials are dense in continuous functions, it is enough to prove the claim when
k=m bkzk are a trigonometric polynomials
ϕj(z) := Pm
k=−m akzk and ψj(z) = Pm
for all j. One can write ϕj and ψj in the form
ϕj(z) = z−m
m′
Xk=0
ajzk = z−mqj(z), ψj(z) = z−mpj(z)
where z = ( x−i
x+i ) and qj and pj are analytic polynomials. In this case we have
H(w) = Te−iηx H(w0)Teiηx + K(w, w0)
w0
j
= Te−iηx (X Tϕj Dϑ
= Te−iηx (X Tz−mTqj Dϑ
= Tz−mTe−iηx (X Tqj Dϑ
w0
j
w0
j
+X Dν
w0
j
+X Dν
+X Dν
w0
j
w0
j
Tψj )Teiηx + K(w, w0)
Tz−mTpj)Teiηx + K(w, w0)
Tpj )Teiηx + K0 + K(w, w0)
9
w0
j
w0
j
have
+P Dν
H(w) = Te−iηx Tz−m H(w0)Teiηx + K(w, w0)
for some K0 ∈ K(H 2) since Dνj Tz−m − Tz−mDνj ∈ K(H 2) ∀j, Tz−mTe−iηx =
Te−iηxTz−m and the sum is finite. Let H(w0) =P Tqj Dϑ
Tpj . Then we
where K(w, w0) = K0 + K(w, w0) ∈ K(H 2). Since Tz−mTe−iηx = Te−iηx Tz−m
we have ker( H(w0)) ∩ eiηxH 2 6= {0} for some η > 0. If this is not the case i.e.
if
ker( H(w0))∩eiηxH 2 = {0} ∀η > 0 then H(w)− K(w, w0) = Tz−mTe−iηx H(w0)Teiηx
is invertible for all η > 0. And this implies that H(w0)(eiηxH 2) = zm(eiηxH 2) ∀η >
0. Since ran( H(w0) ⊆ H 2 is closed, this implies that ran( H(w0)) = zmH 2 and
this in turn implies that H(w0) = Tz−m H(w0) is invertible which contradicts
our assumption. Since ker( H(w0) is finite dimensional, an η0 := max{η > 0 :
ker( H(w0)) ∩ (eiηxH 2) 6= {0}} exists. Now let H0 := Te−iη0 x H(w0)Teiη0x, then
ker( H0) 6= {0}. Since ker( H(w0)) ∩ (ei(η0+δ)xH 2) = {0} ∀δ > 0 we have
Tz−mTe−iδx H0Teiδx = Te−iδx Tz−m H0Teiδx
is invertible for all δ > 0. This implies that H0(eiδxH 2) = zm(eiδxH 2) for all
δ > 0. Since ran( H0) ⊆ H 2 is closed, this implies that ran(H0) = zmH 2. Since
ind(Tz−m H0) = 0 this implies that ind( H0) = −m which in turn implies that
ker( H0) = {0}. This contradiction implies that H(w0) is invertible.
Hence if ψj , ϕj ∈ H ∞ or ψj, ϕj are continuous for all j then T = P Tϕj Dϑj +
P Dνj Tψj is Fredholm with index zero implies that T is invertible. So if ψj, ϕj ∈
(H ∞ + C) ∩ QC = QC for all j then T =P Tϕj Dϑj +P Dνj Tψj is Fredholm with
index zero implies that T is invertible.
(cid:3)
A reinterpretation of this theorem would be relating the invertibility of a generic
element in Ψ to the invertibility of a related Toeplitz operator which is the corollary
below:
Corollary 4. For any T = P Tϕj Dϑj +P Dνj Tψj ∈ Ψ such that T is Fredholm
ψj ϕj ∈ QC and νj, ϑj ∈ C([0,∞]) ∀j ∈ N, T is invertible if and only if Tϕ ∈ Ψ is
invertible where ϕ =P λjϕj +P µjψj and µj = limt→∞ νj(t), λj = limt→∞ ϑj(t).
Proof. If T is invertible then T is Fredholm with ind(T ) = 0. Using the homotopy
H : [0, 1] → Ψ constructed in the beginning of this section, since ind is continuous
we have H(w) is Fredholm ∀w ∈ [0, 1] and ind(H(w)) = ind(T ) = 0 ∀w ∈ [0, 1]. In
particular H(0) = Tϕ is Fredholm and ind(Tϕ) = 0, since any Fredholm Toeplitz
operator Tϕ with ind(Tϕ) = 0 is invertible, Tϕ is invertible.
On the other hand if Tϕ is invertible, since T is Fredholm, by the proof of
Theorem 3, H(w) is invertible ∀w ∈ [0, 1], in particular H(1) = T is invertible. (cid:3)
4. Applications of the Main Results
An immediate application of Corollary 4 shows that the essential spectrum and
the spectrum of a quasi-parabolic composition operator coincide:
Theorem 5. Let ϕ : D → D and ϕ : H → H be such that ϕ(z) = 2iz+η(z)(1−z)
and ϕ(w) = w + ψ(w) where η ∈ QC(T) ∩ H ∞, ℑ(η(z)) > δ > 0 for all z ∈ D,
ψ ∈ QC(R) ∩ H ∞, ℑ(η(w)) > δ > 0 for all w ∈ H. Then C ϕ : H 2(H) → H 2(H) is
bounded and σe(C ϕ) = σ(C ϕ). We also have σe(Cϕ) = σ(Cϕ).
2i+η(z)(1−z)
10
U. G UL, B. KOCA
Proof. The boundedness of C ϕ was shown in [3]. In particular we have
∞
C ϕ =
Tτ nDϑn
Xn=0
where τ (x) = iα − ψ(x) and ϑn(t) = (−it)ne−αt
λ 6∈ σe(C ϕ) we have
n!
∞
for some α > 0. So for any
λ − C ϕ = λ −
Tτ nDϑn
Xn=0
where the series on the Right Hand Side converges in the operator norm. Since
λn := limt→∞ ϑn(t) = 0, by Corollary 1 we have λ − C ϕ is invertible if and only
if λ is invertible which is certainly the case since λ 6= 0. Hence λ 6∈ σ(C ϕ) ⇒
σ(C ϕ) ⊆ σe(C ϕ) ⇒ σ(C ϕ) = σe(C ϕ). The same argument applies to Cϕ since
∞
Φ ◦ Cϕ ◦ Φ−1 = T z+i+η◦C(z)
Xn=0
where Φ : H 2(D) → H 2(H) is the isometric isomorphism
√π(z + i)(cid:19)f(cid:18) z − i
z + i(cid:19),
Φ(f )(z) =(cid:18)
z+i
1
Tτ nDϑn
C is the Cayley transform and τ (x) = iα − η ◦ C(x). Hence we also have σ(Cϕ) =
σe(Cϕ).
(cid:3)
References
[1] C.C. Cowen, B.D. MacCluer Composition Operators on Spaces of Analytic Functions,
CRC Press, 1995.
[2] R.G. Douglas, Banach algebra techniques in Operator Theory, GTM Springer, 1998.
[3] U. Gul, Essential spectra of quasi-parabolic composition operators on Hardy spaces of ana-
lytic functions, J. Math. Anal. Appl. 377, (2011), 2, 771–791
[4] V. Matache, Composition operators on Hardy spaces of a half-plane, Proc. Amer. Math.
Soc. 127, (1999), no. 5, 1483–1491.
[5] G.J. Murphy, C ∗-algebras and Operator Theory, Academic Press, New York-London, 1990.
[6] W. Rudin, Functional Analysis, McGraw Hill Inc., 1973.
[7] D. Sarason, Function Theory on the Unit Circle, Lecture Notes for a Conference at Virginia
Polytechnic and State University, Blacksburg, Virginia, 1978.
ugur gul,
Hacettepe University, Department of Mathematics, 06800, Beytepe, Ankara, TURKEY
E-mail address: [email protected]
beyaz bas¸ak koca,
Istanbul University, Department of Mathematics, 34134, Vezneciler, Istanbul, TURKEY
E-mail address: [email protected]
|
1201.3531 | 2 | 1201 | 2012-11-06T08:06:17 | On the peripheral point spectrum and the asymptotic behavior of irreducible semigroups of Harris operators | [
"math.FA"
] | Given a positive, irreducible and bounded C_0-semigroup on a Banach lattice with order continuous norm, we prove that the peripheral point spectrum of its generator is trivial whenever one of its operators dominates a non-trivial compact or kernel operator. For a discrete semigroup, i.e. for powers of a single operator T, we show that the point spectrum of some power T^k intersects the unit circle at most in 1. As a consequence, we obtain a sufficient condition for strong convergence of the C_0-semigroup and for a subsequence of the powers of T, respectively. | math.FA | math |
ON THE PERIPHERAL POINT SPECTRUM AND THE
ASYMPTOTIC BEHAVIOR OF IRREDUCIBLE SEMIGROUPS OF
HARRIS OPERATORS
MORITZ GERLACH
Abstract. Given a positive, irreducible and bounded C0-semigroup on a Ba-
nach lattice with order continuous norm, we prove that the peripheral point
spectrum of its generator is trivial whenever one of its operators dominates
a non-trivial compact or kernel operator. For a discrete semigroup, i.e. for
powers of a single operator T , we show that the point spectrum of some power
T k intersects the unit circle at most in 1.
As a consequence, we obtain a sufficient condition for strong convergence
of the C0-semigroup and for a subsequence of the powers of T , respectively.
1. Introduction
E. B. Davies proved in [9] that σp(A)∩iR ⊆ {0} if A is the generator of a positive
and contractive semigroup on ℓp for some 1 ≤ p < ∞. This result was generalized
by V. Keicher in [14] to positive, bounded and strongly continuous semigroups on
atomic Banach lattices with order continuous norm. A further generalization to
(w)-solvable semigroups on super-atomic Banach lattices was found by M. Wolff
[19].
Independently, G. Greiner observed in [13, Thm. 3.2 and Kor. 3.11] that σp(A)∩
iR ⊆ {0} for the generator A of a positive, contractive and strongly continuous
semigroup on Lp(Ω, µ), 1 ≤ p < ∞, if the semigroup contains a kernel operator.
This covers in particular the atomic case because every regular operator on an order
complete atomic Banach lattice is a kernel operator. Greiner's result is also in the
center of a survey article by W. Arendt [4] that includes a generalization to Abel
bounded semigroups.
Under the assumption that the semigroup is irreducible, we generalize Greiner's
result in the present article in two respects. First, we consider general Banach
lattices with order continuous norm instead of Lp and secondly, we merely assume
that one operator of the semigroup dominates a non-trivial compact operator. Un-
der suitable assumptions, the latter holds in particular if the semigroups contains
a Harris operator, i.e. one operator dominates a non-trivial kernel operator. More
precisely, we prove the following theorem in Section 3.
Theorem. Let E be a Banach lattice with order continuous norm and let T =
(T (t))t∈[0,∞) be a positive, irreducible, bounded and strongly continuous semigroup
on E with generator A. Assume that there exists t0 > 0 such that T (t0) ≥ K > 0
where K is a compact or a kernel operator. Then σp(A) ∩ iR ⊆ {0}.
This result remains true for semigroups of Harris operators which are merely
Abel bounded, see Proposition 3.2.
2010 Mathematics Subject Classification. Primary 47A11; Secondary: 47B65, 47G10, 47D06.
Key words and phrases. peripheral point spectrum, irreducible, Harris operator, compact op-
erator, asymptotic behavior.
1
2
MORITZ GERLACH
We also consider discrete semigroups by which we mean a family T = (T n)n∈N0
for some bounded operator T . If T is positive, irreducible and power bounded such
that T m dominates a non-trivial compact or kernel operator for some m ∈ N, then
σp(T n) ∩ Γ ⊆ {1}
for some n ∈ N where Γ denotes the unit circle. This result, presented as The-
orem 3.4, is optimal in the sense that, under these assumptions, one can neither
expect σp(T ) ∩ Γ ⊆ {1} for the point spectrum of T nor σ(T n) ∩ Γ ⊆ {1} for the
spectrum of some operator T n, see Examples 3.3 and 3.6.
One important application of the above-mentioned results lies in the analysis of
the asymptotic behavior of a semigroup. Greiner deduced from his result that a
positive and contractive strongly continuous semigroup on Lp converges strongly
to an equilibrium if it admits a strictly positive fixed point and contains a kernel
operator [13, Kor. 3.11]. See also [4, Thm. 4.2] for a proof in the irreducible case.
In Section 4, we generalize this to bounded and irreducible semigroups of Harris
operators on Banach lattices with order continuous norm.
In addition, we obtain an analogous result for discrete semigroups in Theorem 4.3.
If T is a positive, power bounded and irreducible Harris operator with non-trivial
fixed space, then there exists some k ∈ N such that the subsequence T nk converges
strongly as n tends to infinity. If T is even strongly positive, i.e. T x is a quasi-
interior point of E+ for all x > 0, we show that the whole sequence T n converges
strongly.
The main tools for our study of the asymptotic behavior are a so-called zero-two
law by G. Greiner [13, Thm. 3.7] and a result by D. Axmann [7, Satz 3.5] on dis-
jointness of powers of an irreducible operator. In order to improve the accessibility
of both results we present their proofs in the appendices.
2. Notation and Tools
Let us fix some notation and recall some well-known facts from the theory of
Banach lattices and positive operators. The reader is referred to [16] for a more
detailed introduction.
Throughout, E denotes a (real) Banach lattice with order continuous norm. We
write E+ for the positive cone of E and x > 0 for an element x ∈ E+ different from
zero. For y ∈ E+ we denote by
Ey := {x ∈ E : x ≤ cy for some c > 0} = [c≥0
[−cy, cy]
the principal ideal generated by y. If Ey is dense in E, then y is called a quasi-
interior point of E+. We say that a positive functional x′ ∈ E′ is strictly positive
if hx′,xi > 0 for all x 6= 0. By Fix(T ) := ker(I − T ) we denote the fixed space of
a linear operator T : E → E.
2.1. Positive Discrete and Strongly Continuous Semigroups. For a conve-
nient simultaneous dealing with discrete and strongly continuous semigroups we
use the following definition.
Definition 2.1. Let R = N0 or R = [0,∞). A family T = (T (t))t∈R of bounded
linear operators on E is called a semigroup on E if T (t + s) = T (t)T (s) for all t, s ∈
R. A semigroup T = (T (t))t∈[0,∞) is called strongly continuous if the mappings
t 7→ T (t)x are continuous for all x ∈ E. A semigroup T = (T (t))t∈N0 is said to be
discrete.
A semigroup T is called positive if T (t) is positive for all t ∈ R and it is called
bounded if supt∈R kT (t)k < ∞.
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
3
Given a semigroup T = (T (t))t∈R, a subset A ⊆ E is said to be T -invariant
if T (t)A ⊆ A for all t ∈ R. The semigroup T is called irreducible if there are no
closed T -invariant ideals in E beside {0} and E. By
Fix(T ) := \t∈R
ker(I − T (t))
we denote the fixed space of a semigroup T .
It will be convenient to call a single operator T irreducible (power bounded) if
the discrete semigroup (T n)n∈N0 is irreducible (bounded).
By EC we denote the complexification of E, which is called a complex Banach
lattice (see [16, Sec. 2.2]), and by TC : EC → EC the complexification of a linear
operator T . If T = (T (t))t∈[0,∞) is a strongly continuous semigroup on E with
generator A, then so is TC = (T (t)C)t∈[0,∞) with generator AC.
We recall that a strongly continuous semigroup with generator A is called Abel
bounded if Reλ ≤ 0 for all λ ∈ σ(AC) and supλ>0 kλ(λ − AC)−1k < ∞. A discrete
semigroup T = (T n)n∈N0 is called Abel bounded if λ ≤ 1 for all λ ∈ σ(TC) and
supλ>1 k(λ − 1)(λ − TC)−1k < ∞.
Lemma 2.2. Let T = (T (t))t∈R be a positive and Abel bounded semigroup on E
which is strongly continuous or discrete. Assume that T (t)z ≥ z for some z ∈ E,
z > 0, and all t ∈ R. Then there exists 0 < x′ ∈ Fix(T ′).
If in addition T is irreducible, x′ is strictly positive and z is a quasi-interior
point of E+ with Fix(T ) = span{z}.
Proof. The existence of 0 < x′ ∈ Fix(T ′) follows from [17, Lem. V 4.8] in the dis-
crete case and from the proof of [6, Prop. 4.3.6] for a strongly continuous semigroup,
respectively.
Now, assume that T is irreducible. Since the absolute null ideal
N (x′) := {x ∈ E : hx′,xi = 0} 6= E
is closed and T -invariant, it follows that N (x′) = {0}, i.e. x′ is strictly positive.
Thus, hx′, T (t)z − zi = 0 for all t ∈ R implies that z ∈ Fix(T ). By the same
argument, we observe that T (t)x = x for all t ∈ R and x ∈ Fix(T ), i.e. Fix(T )
is a sublattice of E.
Let x ∈ Fix(T ). Then x+ := x ∨ 0 ∈ Fix(T ) and x− := x − x+ ∈ Fix(T )
and the principal ideals generated by x+ and x− are T -invariant and disjoint. The
irreducibility of T implies that either x+ is a quasi-interior point of E+ and x− = 0
or vice versa. Consequently, Fix(T ) is totally ordered and hence one-dimensional
by [17, Prop. II 3.4], i.e. Fix(T ) = span{z}.
(cid:3)
Next, we note a consequence of the well-known mean ergodic theorem.
Lemma 2.3. Let T : E → E be a positive and Abel bounded linear operator. If
there exists a quasi-interior point e ∈ Fix(T ) of E+, then T is mean ergodic, i.e.
E = Fix(T ) ⊕ (I − T )E.
Proof. For n ∈ N and x ∈ E let C(n)x := 1
averages of T .
M := supn∈NkC(n)k < ∞, see [10, 1.5]. Hence,
k=0 T kx denote the Cesàro
It is well-known that the Abel boundedness of T implies that
n Pn−1
(cid:13)(cid:13)(cid:13)(cid:13)
T n
n (cid:13)(cid:13)(cid:13)(cid:13)
≤ 2(cid:13)(cid:13)(cid:13)(cid:13)
T n
n + 1(cid:13)(cid:13)(cid:13)(cid:13)
for all n ∈ N. Now, it follows from lim 1
for all x ∈ E.
≤ 2kC(n + 1)k ≤ 2M
n T nx = 0 for all x ∈ Ee that lim 1
n T nx = 0
4
MORITZ GERLACH
Due to the order continuity of the norm on E, for every c > 0 the C(n)-invariant
order interval [−ce, ce] is σ(E, E′)-compact, see [16, Thm. 2.4.2]. Thus, the se-
quence (C(n)x) has a weak cluster point for every x ∈ Ee and the mean ergodic
theorem [15, §2 Thm. 1.1] yields that lim C(n)x exists for all x ∈ Ee. Consequently,
by the uniform boundedness of the operators C(n), lim C(n)x exists for all x ∈ E.
The desired ergodic decomposition E = Fix(T ) ⊕ (I − T )E now follows from [15,
§2 Thm. 1.3].
(cid:3)
We will use the following version of the famous splitting theorem by Jacobs, de
Leeuw and Glicksberg. Let Lσ(E) denote the space of all bounded linear operator
on E endowed with the weak operator topology.
Theorem 2.4 (Jacobs -- de Leeuw -- Glicksberg). Let T = (T (t))t∈R be a positive
and bounded semigroup on E and denote by S the closure of {T (t) : t ∈ R} in
Lσ(E). If there exists a quasi-interior point e ∈ Fix(T ) of E+, then S contains
a positive projection Q commuting with every operator in S such that the closed
subspace F := QE includes Fix(T ) and QS := {QT : T ∈ S } ⊆ S is a norm
bounded group of positive operators with neutral element Q.
If in addition T is irreducible, Q is strictly positive in the sense that Qy > 0 for
all y > 0.
Proof. It follows from the uniform boundedness principle that the closure of a norm
bounded set in Lσ(E) is bounded in norm, again. Moreover, the operators of S
are positive and they commute since T is Abelian.
Now, let x ∈ Ee, i.e. x ∈ [−ce, ce] for some c > 0. Then {T (t)x : t ∈ R} is a
subset of [−ce, ce] and hence, by the order continuity of the norm [16, Thm. 2.4.2],
relatively weakly compact. Since Ee is dense in E, {T (t) : t ∈ R} is relatively
compact in Lσ(E) [11, Cor. A 5]. Now, it follows from [15, §2 Thm 4.1] that there
exists a (positive) projection Q ∈ S that commutes with every operator in S and
QS is a (bounded) group with neutral element Q. Since Q is in the Lσ(E)-closure
of T , one has that Fix(T ) ⊆ Fix(Q) ⊆ F .
If in addition T is irreducible, then Q is strictly positive because the closed ideal
N (Q) := {y ∈ E : Qy = 0}
is T -invariant and distinct from E.
2.2. Atoms. We recall that an element a ∈ E+ is said to be an atom (of E) if the
generated ideal
(cid:3)
Ea := {x ∈ E : x ≤ ca for some c > 0}
is one-dimensional. If E contains no atoms, E is said to be diffuse.
Remark 2.5. Let a ∈ E+ be an atom and T : E → E be positive. Then clearly
T Ea ⊆ ET a. If in addition T is invertible with a positive inverse, i.e. T is a lattice
isomorphism, then
ET a = T T −1ET a ⊆ T ET −1T a = T Ea = span{T a}.
Hence, every lattice isomorphism on E maps atoms to atoms.
The lemma below is a more general formulation of [14, Prop. 3.5], followed by
an analogous assertion in the discrete case.
Lemma 2.6. Let T = (T (t))t∈R be a positive and bounded strongly continuous
group on E, i.e. T is a uniformly bounded family of positive linear operators on
E such that T (0) = I, T (t)T (s) = T (t + s) for all t, s ∈ R, and the mappings
t 7→ T (t)x are continuous from R to E for every x ∈ E. Then Fix(T ) contains
every atom of E.
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
5
Proof. Since T is a group, every operator T (t) is a lattice isomorphism and hence
maps atoms to atoms by Remark 2.5. Fix an atom a ∈ E+. Then for every disjoint
atom b ∈ E+ one has that a− b = a + b and hence ka− bk ≥ kak. Thus, it follows
from limt→0 T (t)a = a that T (t)a ∈ Ea for all t ∈ R. Therefore, T (t)a = exp(λt)a
for a λ ∈ R and all t ∈ R by the semigroup law. Since the group T is bounded, we
conclude that λ = 0 and hence a ∈ Fix(T ).
Lemma 2.7. Suppose that E has atoms and let T be an irreducible lattice isomor-
phism on E such that
(cid:3)
(2.1)
Then there exists n ∈ N such that T n = I.
Proof. Let a ∈ E+ be an atom. The ideal
sup{kT kk : k ∈ Z} < ∞.
J := {x ∈ E : x ≤ c(a + T a + ··· + T ma) for some c ≥ 0, m ∈ N}
is T -invariant and hence J = E. Since the norm on E is order continuous, J equals
J⊥⊥, the band generated by J. Thus, for every x > 0 we find some n ∈ N0 such
In particular, T −1a ∧ T n−1a > 0 for some n ∈ N. As T −1
that x ∧ T na > 0.
maps atoms to atoms by Remark 2.5, this implies that T n−1a = cT −1a for some
c > 0 and T na = ca. Now, it follows from assumption (2.1) that c = 1. Therefore,
T ka ∈ Fix(T n) for all k ∈ N0. Since every T ka is an atom, we conclude that
E = J = span{a, T a, . . . , T n−1a} ⊆ Fix(T n),
which completes the proof.
(cid:3)
2.3. Kernel and Harris Operators. First, we recall that a linear operator T :
E → E is called regular if it is the difference of two positive operators. Every
regular operator is bounded [16, Prop. 1.3.5] and admits a modulus. Moreover,
since E is order complete, the regular operators on E form a Banach lattice with
respect to the regular norm kTkr := kTkk [16, Prop. 1.3.6].
We denote by E′ ⊗ E the space of all finite rank operators on E. The elements
of (E′ ⊗ E)⊥⊥, the band generated by E′ ⊗ E in the space of all regular operators,
are called (regular) kernel operators. A regular operator T on E is called a Harris
operator if T n 6∈ (E′ ⊗ E)⊥ for some n ∈ N.
Note that a positive operator is a Harris operator if and only if some power
In the case where E = Lp, the band
dominates a non-trivial kernel operator.
(E′ ⊗ E)⊥⊥ consists precisely of those operators given by a measurable kernel (see
[17, Prop. IV 9.8] or [16, Thm. 3.3.7]).
The following well-known lemma states that the kernel operators form an algebra
ideal in the regular operators.
Lemma 2.8. Let K be a kernel operator. Then KT ∈ (E′ ⊗ E)⊥⊥ and T K ∈
(E′ ⊗ E)⊥⊥ for every regular operator T .
Proof. We may assume that K, T ≥ 0. Denote by J the (lattice) ideal of regular
operators generated by E′ ⊗ E and let A := [0, K] ∩ J. Then K = sup A by [16,
Prop. 1.2.6] and
Kx = sup{Ax : A ∈ A }
(x ∈ E+),
since A is upwards directed. Now, consider the directed set A T := {AT : A ∈ A }.
As every operator of A is dominated by a finite rank operator, so is every AT , i.e.
A T ⊆ J. Now, it follows from
KT x = sup{AT x : A ∈ A } (x ∈ E+)
6
MORITZ GERLACH
that KT = sup A T and hence KT ∈ (E′ ⊗ E)⊥⊥. By a similar argument, using in
addition the fact that T is order continuous because of the order continuity of the
norm, one observes that T K ∈ (E′ ⊗ E)⊥⊥.
(cid:3)
The following lemma shows that under certain conditions some power of a Harris
operator always dominates a non-trivial compact operator.
Lemma 2.9. Let T : E → E be a positive and power bounded Harris operator such
that Fix(T ) contains a quasi-interior point of E+ and that there exists a strictly
positive element in Fix(T ′). Then T n ≥ C > 0 for some n ∈ N and some compact
operator C.
Proof. After replacing T with a suitable power T m, we may assume that T ≥ K > 0
for some kernel operator K : E → E. Let S := T − K. Since the kernel operators
form an algebra ideal, see Proposition 2.8, Kn := T n − Sn ∈ (E′ ⊗ E)⊥⊥ for all
n ∈ N.
Let e ∈ Fix(T ) be a quasi-interior point of E+ and x′ ∈ Fix(T ′) be strictly
positive. It follows from e = T e = Ke + Se > Se that Sn+1e ≤ Sne for all n ∈ N.
By the order continuity of the norm, u := limn→∞ Sne ∈ Fix(S) exists and satisfies
0 ≤ u < e. Since x′ is strictly positive, we infer from T u = Ku + Su ≥ u and
hx′, T u − ui = 0 that u ∈ Fix(T ). Hence, v := e − u > 0 is a fixed point of T and
lim Snv = 0. As the operators Kn are uniformly bounded, it follows from
Knv = T nv − Snv → v
(n → ∞)
that lim K 2
nv = v. In particular, K 2
m 6= 0 for a suitable m ∈ N large enough.
that lim C2
k ≤ K 2
kC2
k v = K 2
mvk ≤ kCk(Ckv − Kmv)k + k(Ck − Km)Kmvk
It follows from [16, Prop. 1.2.6] that Km = sup([0, Km]∩ J) where J is the ideal
generated by E′ ⊗ E. Thus, there exists a (bounded) sequence (Ck) ⊆ [0, Km] ∩ J
such that lim Ckx = Kmx for x ∈ {v, Kmv}. Since every Ck is dominated by a finite
rank operator, C2
k is compact for every k ∈ N by [16, Cor. 3.7.15]. We conclude
from
kv − K 2
mv 6= 0. Therefore, 0 < C2
m ≤ T 2m for a suitable k ∈ N. (cid:3)
Arendt proved in [2, Kor. 1.26] that compact operators are disjoint from lattice
isomorphisms on a diffuse and order complete Banach lattice. This is an essential
tool for our spectral analysis in Section 3. In order to be more self-contained, we
give a proof of this result for a Banach lattice with order continuous norm.
Theorem 2.10. Let V, K : E → E be positive operators on a diffuse Banach lattice
E with order continuous norm. If V is a lattice isomorphism and K is compact,
then V ∧ K = 0.
Proof. First, we show that S := I ∧ K = 0. Aiming for a contradiction, we assume
that Sx > 0 for some x ∈ E+. Since E is Archimedean, we find some c > 0 such
that w := cSx − x is not negative, i.e. it has a non-trivial positive part. Denote by
P the band projection onto {w+}⊥⊥, the band generated by w+. Then
0 < w+ = P w = P (cSx − x) = cP Sx − P x.
It follows easily from S ≤ I that S is an orthomorphism, i.e. it commutes with
every band projection. Thus, SP x > 0 and hence P x > 0. Since E is diffuse,
we are able to construct a sequence (xn) ⊆ E of pairwise disjoint elements, each
satisfying 0 < xn < P x, by applying [16, Lem. 2.7.12] inductively. Let Pn be the
band projection onto {xn}⊥⊥ and define un := Pnx/kPnxk. As the orthomorphism
S is dominated by the compact operator K, S itself is compact by [1, Thm. 16.21].
Hence, after passing to a subsequence, (Sun) converges to some y ∈ E. By the order
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
7
continuity of the norm, the disjoint and order bounded sequence (Pny) converges
to zero [16, Thm. 2.4.2]. Now, it follows from
kPnxkun = Pnx = PnP x ≤ cPnSP x = cPnSx = ckPnxkSun
for all n ∈ N that
and hence
Pny = lim
m→∞
PnSum ≤ lim
m→∞
PnScSum = cSPny
kunk ≤ ckPnSunk ≤ ckPnSun − Pnyk + ckPnyk ≤ ckSun − yk + c2kSPnyk.
for all n ∈ N. The right-hand side tends to zero, which is a contradiction to
kunk = 1. Hence, S = I ∧ K = 0.
In order to complete the proof, we recall that the mapping T 7→ T V is a lattice
homomorphism on the lattice of all regular operators on E, see [1, Thm. 7.4].
Therefore, V ∧ K = (I ∧ KV −1)V and hence vanishes by the first part of the
proof since KV −1 is compact.
(cid:3)
3. Triviality of the Peripheral Point Spectrum
In this section we give the proofs of our main results, which are inspired by [4,
Thm. 3.1]. There, the semigroup is assumed to contain a kernel operator, which is
a less general hypothesis than ours.
Again, let E be a Banach lattice with order continuous norm.
Theorem 3.1. Let T = (T (t))t∈[0,∞) be a positive, irreducible and bounded strongly
continuous semigroup with generator A. If there exists t0 > 0 such that T (t0) ≥
K > 0 where K is a compact or a kernel operator, then σp(AC) ∩ iR ⊆ {0}.
Proof. Let α ∈ R such that iα ∈ σp(AC) ∩ iR. Then T (t)Cz = exp(iαt)z for some
z ∈ EC\{0}. Since 0 < z = T (t)Cz ≤ T (t)z for all t ≥ 0, by Lemma 2.2
there exists a strictly positive ϕ ∈ Fix(T ′) and Fix(T ) = span{e}, where e := z
is a quasi-interior point of E+. In view of Lemma 2.9, we may now assume that
T (t0) ≥ K > 0 for a compact operator K.
Denote by S the closure of {T (t) : t ≥ 0} in Lσ(E), the bounded linear operators
on E endowed with the weak operator topology. Then, by Theorem 2.4, S contains
a strictly positive projection Q commuting with every operator in S such that
Fix(T ) ⊆ F := QE and that QS := {QS : S ∈ S } ⊆ S is norm bounded group
of positive operators with neutral element Q. Hence, the restriction of the elements
of S to F defines a group of lattice isomorphisms.
If x ∈ F , then x = Qx ≤ Qx and since ϕ is also a strictly positive fixed
point of Q′, it follows that hϕ, Qx − xi = 0 and hence Qx = x. Thus, F is
a closed sublattice of E and hence a Banach lattice. Moreover, the norm on F is
order continuous because every monotone order bounded sequence in F converges
(see [16, Thm. 2.4.2]).
Since Q is strictly positive and the quasi-interior point e belongs to F , QKF e > 0.
Thus, T (t0)F = QT (t0)F dominates the non-trivial compact operator QKF . If
F were diffuse, compact operators would be disjoint from lattice isomorphisms by
Theorem 2.10. Thus, there exists at least one atom a ∈ F . Since a ∈ Fix(T ) =
span{e} by Lemma 2.6, we obtain that e is an atom of F and consequently,
F = Fe = span{e} = Fix(T ).
It remains to check that z ∈ FC = QCEC. Then z ∈ Fix(TC) and therefore α = 0.
Let w ∈ (EC)′ such that hT (t)Cz, wi = hz, wi = kzk > 0. Since QC is in the
closure of TC in Lσ(EC), there exists t ≥ 0 such that
hT (t)Cz − QCz, wi < kzk.
8
MORITZ GERLACH
This implies that
hQCz, wi ≥ hT (t)Cz, wi − hT (t)Cz − QCz, wi > 0
and hence QCz 6= 0. Now, choose w ∈ (EC)′ vanishing on z and let ε > 0. Again,
there exists t ≥ 0 such that
which implies that
hT (t)Cz − QCz, wi < ε
hQCz, wi ≤ hT (t)Cz − QCz, wi + hT (t)Cz, wi < ε.
Since ε > 0 was arbitrary, hQCz, wi = 0. Thus, QCz ∈ span{z} and consequently
QCz = z ∈ FC because QC is a projection.
(cid:3)
As we will see next, this result remains true for Abel bounded semigroups of
Harris operators.
Proposition 3.2. Let T = (T (t))t∈[0,∞) be a positive, irreducible and Abel bounded
strongly continuous semigroup with generator A. If T (t0) is a Harris operator for
some t0 > 0, then σp(AC) ∩ iR ⊆ {0}.
Proof. Assume that T (t)Cz = exp(iαt)z for some z ∈ EC\{0}, α ∈ R and all
t ≥ 0. Since z ≤ T (t)z for all t ≥ 0, it follows from Lemma 2.2 that there
exists a strictly positive linear form ϕ ∈ Fix(T ′). Now, we endow E with the order
continuous lattice norm
kxkϕ := hϕ,xi
(x ∈ E).
Let (F,k·kF ) be the completion of (E,k·kϕ), i.e. the closure of E in the bidual
(E,k·kϕ)′′. Then T uniquely extends to a positive strongly continuous contraction
semigroup T = ( T (t))t∈[0,∞) on F .
We show that T is irreducible. Let J be a closed and T -invariant ideal in
(F,k·kF ). Then J := J ∩ E is a closed and T -invariant ideal in (E,k·k) and hence
J = {0} or J = E. Since ϕ is order continuous, E is an ideal in F by [17, Lem.
IV 9.3]. This implies that J is the closure of J in F and hence either J = {0} or
J = F . Thus, T is irreducible.
Next, we verify that T (t0) is a Harris operator. To simplify notation, we assume
that T := T (t0) 6∈ (E′ ⊗ E)⊥; otherwise we replace t0 with nt0 for a suitable n ∈ N.
Hence, there exist w′ ∈ E′, w ∈ E and z ∈ E+ such that (T ∧ w′ ⊗ w)z > 0. In
view of Nakano's carrier theorem [16, Thm. 1.4.11], the strictly positive ϕ is a weak
order unit of E′. Since
w′ ⊗ w = sup
m∈N
((mϕ ∧ w′) ⊗ w) = sup
m∈N
((mϕ ⊗ w) ∧ (w′ ⊗ w)),
one has that
0 < T ∧ w′ ⊗ w = T ∧ (w′ ⊗ w)
= T ∧ sup{(mϕ ⊗ w) ∧ (w′ ⊗ w) : m ∈ N}
= sup{T ∧ (mϕ ⊗ w) ∧ (w′ ⊗ w) : m ∈ N}.
Hence, m(T ∧(ϕ⊗w)) ≥ T ∧(mϕ⊗w) > 0 for some m ∈ N. Since the extensions
ϕ ⊗ w and T leave the ideal E invariant, we conclude that
( T ∧ ( ϕ ⊗ w))z = inf
= inf
F { T (z − y) + ϕ(y)w : y ∈ F, 0 ≤ y ≤ z}
E {T (z − y) + ϕ(y)w : y ∈ E, 0 ≤ y ≤ z}
This shows that T = T (t0) 6∈ (F ′ ⊗ F )⊥.
= (T ∧ (ϕ ⊗ w))z > 0.
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
9
Finally, we conclude from Theorem 3.1 that σp(AC) ∩ iR ⊆ σp( AC) ∩ iR ⊆ {0}
where A is the generator of T .
(cid:3)
We continue with a consideration of discrete semigroups, i.e. powers of a single
operator. Let Γ := {z ∈ C : z = 1} denote the unit circle. The following
simple example shows that one cannot expect that σp(TC) ∩ Γ ⊆ {1} in analogy to
Theorem 3.1.
Example 3.3. On E = R2, consider the matrix
1
(3.1)
T := (cid:18)0
1
0(cid:19)
Then T is a positive, irreducible, contractive and compact operator but −1 ∈
σp(TC).
However, the following theorem states that the point spectrum of some power of
TC is trivial.
Theorem 3.4. Let T : E → E be a positive, irreducible and power bounded op-
erator. If there exists m ∈ N such that T m ≥ K > 0 for a compact or a kernel
operator K, then σp(T n
Proof. If σp(TC)∩Γ = ∅, then the assertion follows for n = 1. Thus, we may assume
that there exists α ∈ R and z ∈ EC such that z 6= 0 and TCz = exp(iα)z. Since
C ) ∩ Γ ⊆ {1} for some n ∈ N.
0 < z = exp(inα)z = T n
C z ≤ T nz
(n ∈ N),
Lemma 2.2 applied to the discrete semigroup T := (T n)n∈N0 yields that there
exists a strictly positive ϕ ∈ Fix(T ′) and Fix(T ) = span{e}, where e := z is
In view of Lemma 2.9, we may now assume that
a quasi-interior point of E+.
T m ≥ K > 0 for a compact operator K.
As in the proof of Theorem 3.1, we conclude from Theorem 2.4 that there ex-
ists a strictly positive projection Q : E → E commuting with T and having the
following properties: Its range F := QE is a closed sublattice of E with order con-
tinuous norm and includes Fix(T ). Moreover, the restriction of T to F is a lattice
isomorphism with sup{kT k
F dominates the
non-trivial compact operator QKF and it follows from Theorem 2.10 that F is not
diffuse.
Next, we show that TF is irreducible. Let J ⊆ F be a closed and T -invariant
ideal with corresponding band projection P : F → J and let e′ := P e. Since
T e′ ∈ J, we observe that
Fk : k ∈ Z} < ∞. On the other hand, T m
T e′ = T P e = P T P e ≤ P T e = P e = e′.
Now, it follows from hϕ, e′ − T e′i = 0 that e′ ∈ Fix(T ) = span{e} because ϕ is
strictly positive. Hence, e′ = 0 or e′ = e which implies that either J = {0} or
J = F .
Now, we conclude from Lemma 2.7, applied to the restriction of T to F , that
there exists some n ∈ N such that F ⊆ Fix(T n). Let exp(iβ) ∈ σp(T n
C ) and denote
by ξ1, . . . , ξn the nth roots of exp(iβ), i.e. (ξk)n = exp(iβ) for all 1 ≤ k ≤ n. Then
we infer from
exp(iβ) − T n
C = (ξ1 − TC)(ξ2 − TC) . . . (ξn − TC)
that ξk ∈ σp(TC) for at least one 1 ≤ k ≤ n. Pick y ∈ EC\{0} such that TCy = ξky.
Now, we observe that y ∈ FC = QCEC by the same arguments as in the proof of
Theorem 3.1. Indeed, since QC is in the closure of TC = (T k
C)k∈N0 in Lσ(EC), for
w ∈ (EC)′ satisfying hw, yi = kyk we find some j ∈ N0 such that
hT j
Cy − QCy, wi < kyk.
10
MORITZ GERLACH
Cy, wi = ξj
As hT j
khy, wi = kyk, it follows that hQCy, wi > 0 and hence QCy 6= 0.
On the other hand, for every w ∈ (EC)′ vanishing on y and for every ε > 0 there
exists j ∈ N0 such that hT j
Cy − QCy, wi < ε and hence hQCy, wi < ε. Thus,
QCy ∈ span{y} which implies that QCy = y ∈ FC since QC is a projection.
Altogether, we proved that
y = T n
C y = ξn
k y = exp(iβ)y
and hence exp(iβ) = 1.
(cid:3)
Remark 3.5. (a) The assertion of Theorem 3.4 remains true for every positive, ir-
reducible and Abel bounded Harris operator T : E → E. This can be proven
analogously to Proposition 3.2.
(b) It is well known that σp(TC) ∩ Γ ⊆ {1} for every power bounded operator
T : E → E which is strongly positive, i.e. T x is a quasi-interior point of E+
for all x > 0, c.f. [17, Prop. V 5.6]. This can also be obtained from the proof
of Theorem 3.4 by observing that such an operator T is irreducible and its
restriction to F is again strongly positive and therefore no lattice isomorphism
unless F contains atoms.
(c) The assumption T being irreducible cannot be dropped in Theorem 3.1 (and
Theorem 3.4). Indeed, let (T1(t))t∈[0,∞) be a positive and bounded strongly
continuous semigroup on E such that T1(t0) ≥ K > 0 for some compact op-
erator K and some t0 > 0. Now, let (T2(t))t∈[0,∞) be another positive and
bounded strongly continuous semigroup on E such that iα is an eigenvalue of
its generator for some α 6= 0. Then T (t) := T1(t) ⊕ T2(t) defines a positive and
bounded strongly continuous semigroup on E ⊕ E where T (t0) dominates the
compact operator K ⊕ 0 but iα is an eigenvalue of its generator.
(d) If A is the generator of a bounded and positive strongly continuous semigroup
T = (T (t))t∈[0,∞) such that T (t0) is compact for some t0 > 0, it is easy to see
that σ(AC) ∩ iR ⊆ {0}. Indeed, in this case T is eventually norm continuous
[11, Lem. II 4.22] and hence σ(AC)∩iR is bounded [11, Thm. II 4.18]. Since the
boundary spectrum is cyclic by [12, Thm. 2.4], it follows that σ(AC)∩iR ⊆ {0}.
It is natural to ask if even σ(T n
C ) ∩ Γ is trivial for some n ∈ N under the
assumptions of Theorem 3.4. The following example shows that this is not the
case.
Example 3.6. Fix 1 < p < ∞ and let E := ℓp. We construct a positive and
irreducible contraction T on E such that σp(TC) = ∅ and σ(T n
C ) ∩ Γ = Γ for all
n ∈ N. Since E is atomic, T is automatically a kernel operator and hence satisfies
the assumptions of Theorem 3.4.
Let (bn) ⊆ [0, 1] be a decreasing sequence with Q∞n=1(1 − bn) = 1
2 and let
0 ≤ an ≤ 1 − (1 − bn)p be small enough such that P∞n=1 an ≤ 1. For x = (xn) ∈ ℓp
we define
T x :=(cid:18) ∞
Xn=1
anxn, (1 − b1)x1, (1 − b2)x2, (1 − b3)x3, . . .(cid:19).
If x ≥ 0, then it follows from Jensen's inequality that
kT xkp =(cid:18) ∞
Xn=1
anxp
(1 − bn)pxp
anxn(cid:19)p
Xn=1
n ≤
∞
Xn=1
n +
∞
+
∞
Xn=1
(1 − bn)pxp
n ≤ kxkp
which shows that T is a contraction.
It is well-known that the complexification of E equals ℓp(C), the p-summable
sequences in C. Now, assume that there is λ ∈ σp(TC) and denote by z = (zn) ∈ EC,
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
11
z 6= 0, a corresponding eigenvector. Since TC is injective and contractive, 0 < λ ≤ 1
and it follows by induction that
zn+1 =
1
λn (1 − b1)(1 − b2) . . . (1 − bn)z1
for all n ∈ N. Hence, z is not a null sequence. This is impossible and therefore
σp(TC) = ∅.
In order to calculate the peripheral spectrum of TC, we first point out that the
discrete semigroup (T n
C )n∈N0 is not strongly stable. Indeed, for e1 = (1, 0, 0, . . . )
we have that
n
kT ne1k ≥
Yk=1
(1 − bk) ≥
1
2
for all n ∈ N. Thus, it follows from the characterization of strong stability by
Arendt, Batty, Lyubich and Vu [5, Thm. 5.1] that σ(TC) ∩ Γ is uncountable. Since
σ(TC) ∩ Γ is cyclic [17, Thm. V 4.9], it is dense in Γ and hence σ(TC) ∩ Γ = Γ. By
factorizing λ−T n
C )∩Γ = Γ
for all n ∈ N.
C as in the proof of Theorem 3.4 one observes that also σ(T n
4. Strong Convergence of the Semigroup
The asymptotic behavior of a bounded strongly continuous semigroup T =
(T (t))t∈[0,∞) on a complex Banach space X is highly related to the peripheral (point)
spectrum of its generator A. If, for instance, T is known to be asymptotically almost
periodic, i.e. X = X0 ⊕ XAP where
X0 := {x ∈ X : lim
t→∞kT (t)xk = 0}
and
XAP := span{x ∈ D(A) : Ax = iαx for some α ∈ R},
then σp(A) ∩ iR ⊆ {0} already implies that limt→∞ T (t)x exists for all x ∈ X.
This is the case if the generator A has compact resolvent [6, Prop. 5.4.7] or if the
peripheral spectrum σ(A)∩ iR is countable and T is totally ergodic [6, Thm. 5.5.5].
In the following, we prove a convergence result for semigroups of Harris operators
with a non-trivial fixed space by adjusting techniques developed by Greiner to
arbitrary operator semigroups that might by discrete or strongly continuous. Our
main tool is Greiner's zero-two law, see Theorem A.1 in the appendix.
Again, let E be a Banach lattice with order continuous norm. We start with the
following proposition, a generalized version of [13, Kor. 3.9], that yields sufficient
conditions for strong convergence of a semigroup.
Proposition 4.1. Let T = (T (t))t∈R be a positive, bounded and irreducible semi-
group on E which is strongly continuous or discrete such that Fix(T (t)) = Fix(T ) 6=
{0} for all t ∈ R and assume that there are r, s ∈ R, r > s, such that T (r)∧T (s) > 0.
Then
lim
t→∞
T (t)x = hx′, xie
(x ∈ E)
for some strictly positive x′ ∈ Fix(T ′) and a quasi-interior point e ∈ Fix(T ) of
E+.
Proof. By Lemma 2.2, there exists a strictly positive x′ ∈ Fix(T ′) and a quasi-
interior point e of E+ with Fix(T ) = span{e}. We may assume that hx′, ei = 1.
By assumption,
E2 := {y ∈ E : (T (t) ∧ T (t + τ ))y = 0 for all t ≥ 0} 6= E
12
MORITZ GERLACH
for τ := r − s. Since T is irreducible, it follows from Greiner's zero-two law,
Theorem A.1, that E2 = {0} and
lim
t→∞T (t) − T (t + τ )e = 0.
Hence,
T (t)(I − T (τ ))y ≤ T (t) − T (t + τ )e → 0 (t → ∞)
for all y ∈ [−e, e], i.e. lim T (t)z = 0 for all z ∈ D := (I − T (τ ))[−e, e]. As D is total
in (I − T (τ ))E and T is bounded, limt→∞ T (t)x = 0 for all x ∈ (I − T (τ ))E. By
Lemma 2.3, T (τ ) is mean ergodic, i.e.
E = Fix(T (τ )) ⊕ (I − T (τ ))E.
Since Fix(T (τ )) = span{e}, the corresponding mean ergodic projection is given by
x′ ⊗ e which completes the proof.
(cid:3)
It seems to be rather technical to assume that two operators T (r) and T (s) are
disjoint. However, by Theorem B.2 due to Axmann, this holds if the semigroup
contains an irreducible Harris operator. This leads us to the following theorems,
our main results in this section.
Theorem 4.2. Let T = (T (t))t∈[0,∞) be a positive, bounded, irreducible and
strongly continuous semigroup on E with generator A such that Fix(T ) 6= {0}.
If T (t0) is a Harris operator for some t0 > 0, then
lim
t→∞
T (t)x = hx′, xie
(x ∈ E)
for some strictly positive x′ ∈ Fix(T ′) and a quasi-interior point e ∈ Fix(T ) of
E+.
Proof. By Lemma 2.2, Fix(T ) = span{e} for a quasi-interior point e of E+ and
there exists a strictly positive element x′ ∈ Fix(T ′). Hence, Theorem 3.1 implies
that σp(AC) ∩ iR = {0} and therefore
(4.1)
(see [11, Cor. IV 3.8]). Next, we prove that T := T (t0) is irreducible. Let J ⊆ E be
a T -invariant closed ideal and P : E → J the corresponding band projection. Then
T P e ∈ J and hence T P e = P T P e ≤ P T e = P e. Since hx′, P e − T P ei = 0 and
x′ is strictly positive, it follows that P e ∈ Fix(T ) = span{e} which implies that
J = {0} or J = E. Therefore, it follows from Theorem B.2 that there exist some
natural numbers n < m such that T n ∧ T m > 0. Now, the assertion follows from
Proposition 4.1.
Fix(T (t)) = Fix(T ) = span{e} (t > 0)
(cid:3)
In the discrete case, we obtain strong convergence of a subsequence (T nk)k∈N
for a fixed n ∈ N which is optimal in view of Example 3.3. If the operator is not
only irreducible but even strongly positive, the sequence (T k)k∈N itself converges
strongly to a projection of rank one.
Theorem 4.3. Let T : E → E be a positive, power bounded and irreducible Harris
operator with Fix(T ) 6= {0}. Then there exists n ∈ N such that T nk converges
strongly as k tends to infinity.
If T is even strongly positive, i.e. T x is a quasi-interior point of E+ for all
x > 0, then limk→∞ T kx = hx′, xie for some strictly positive x′ ∈ Fix(T ′) and a
quasi-interior point e ∈ Fix(T ) of E+.
Proof. By Theorem B.2, there are natural numbers a < b such that T a ∧ T b > 0.
Let n := b − a. Now, we conclude as in the proof of Proposition 4.1 that
E = Fix(T n) ⊕ (I − T n)E
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
13
and limk→∞ T kx = 0 for all x ∈ (I − T n)E.
If T is strongly positive, then σp(TC) ∩ Γ ⊆ {1} by Remark 3.5 (b). There-
fore, Fix(T k) = Fix(T ) for all k ∈ N and the assertion follows immediately from
Proposition 4.1.
(cid:3)
Remark 4.4. It is natural to ask whether Theorem B.2 holds true for positive and
irreducible operators that merely dominate non-trivial compact operators, which
would allows us to generalize Theorem 4.2 and 4.3 accordingly.
This is not the case since there exists a positive, compact and irreducible operator
T on L2(Γ), the square-integrable functions on the unit circle endowed with the
Lebesgue measure, such that T n ∧ T m = 0 for all n 6= m. In fact, based on a work
of Varopoulos [18], Arendt constructed a self-adjoint and compact Markov operator
T on L2(Γ) such that T n ∧ T m = 0 whenever n 6= m [3, Ex. 3.7]. By Theorem 4.5
below, there exists a T -invariant band B in L2(Γ) such that the restriction of T to
B is irreducible and still compact. Since B is of the form {f ∈ L2(Γ) : f = 0 on A}
for some measurable A ⊆ Γ by [17, III §1 Ex.2] and Γ\A is not a nullset, B is in
turn isomorphic to L2(Γ) by [8, Cor. 6.6.7 and Thm. 9.2.2].
Theorem 4.5. Let T : E → E be a positive and compact operator. If there exist
a quasi-interior point e ∈ Fix(T ) of E+ and a strictly positive ϕ ∈ Fix(T ′), then
there are finitely many disjoint T -invariant bands B1, . . . , BN ⊆ E distinct from
{0} such that E = B1 ⊕ ··· ⊕ BN and the restriction of T to Bk is irreducible for
all k = 1, . . . , N .
Proof. We assume that there is no T -invariant band A ⊆ E except {0} such that
the restriction of T to A is irreducible. Then, in particular, T is not irreducible on
E and hence there is a T -invariant closed ideal A1 distinct from {0} and E. As the
norm on E is order continuous, every closed ideal is a projection band. Denote by
P1 : E → A1 the corresponding band projection. Since
T P1e = P1T P1e ≤ P1T e = P1e
and hP1e − T P1e, ϕi = 0, the strict positivity of ϕ implies that T P1e = P1e. By
linearity, T (I−P1)e = (I−P1)e and thus (I−P1)E = A⊥1 is a non-trivial T -invariant
band, too. We may assume that hϕ, ei = 1 and
hP1e, ϕi ≤
1
2he, ϕi;
otherwise we replace A1 with A⊥1 and P1 with I − P1.
A2 ⊆ A1 distinct from {0} and A1 with band projection P2 : E → A2 such that
By our assumption, TA1 is not irreducible and hence we find a T -invariant band
hP2e, ϕi ≤
1
2hP1e, ϕi.
Inductively, we obtain a decreasing sequence An+1 ⊆ An of T -invariant bands with
projections Pn : E → An satisfying
hPne, ϕi ≤ 2−nhe, ϕi
(n ∈ N).
Since ϕ is strictly positive and the norm on E is order continuous, this implies that
lim Pne = inf Pne = 0.
Now, consider the sequence xn := Pne/kPnek. The compactness of T yields
that a subsequence of (T xn) = (xn) converges to some x with kxk = 1. On the
other hand, x ∈ An for every n ∈ N because xk ∈ An whenever k ≥ n. Thus,
e = lim(I − Pn)e ∈ {x}⊥ which implies that x = 0, a contradiction.
We proved the existence of a T -invariant band {0} 6= B1 ⊆ E such that TB1 is
irreducible. If B1 6= E we may apply the same argument to the restriction of T to
B⊥1 to obtain a T -invariant band {0} 6= B2 ⊆ B⊥1 such that TB2 is irreducible.
14
MORITZ GERLACH
Continue this construction inductively as long as B1⊕···⊕Bn 6= E. Suppose that
this process does not terminate after finitely many steps, i.e. we obtain an infinite
sequence Bn of disjoint non-trivial bands such that TBn is irreducible. Denote by
Qn : E → Bn the corresponding band projections and let yn := Qne/kQnek. Then
a subsequence of T yn = yn converges to some y ∈ E with kyk = 1 since T is
compact. On the other hand, Qky = limn→∞ Qkyn = 0 for every k ∈ N implies
that y ∈ B⊥k for all k ∈ N. This shows that every yk is contained in {y}⊥ and so
is y. Hence, y = 0 contradicting kyk = 1.
We conclude that the process of constructing B1, B2, . . . ends after finitely many
steps, which completes the proof.
(cid:3)
Appendix A. Greiner's zero-two law
We present the proof of Greiner's zero-two law from [13] in a reformulation for
Banach lattices with order continuous norm and without any continuity condition
on the semigroup.
Throughout, let T = (T (t))t∈R be a positive and bounded semigroup on E, a
Banach lattice with order continuous norm, and fix τ > 0.
Theorem A.1 (Greiner's zero-two law). Assume that Fix(T ) contains a quasi-
interior point e of E+ and that there exists a strictly positive element in Fix(T ′).
Then
E2 := {y ∈ E : (T (t) ∧ T (t + τ ))y = 0 for all t ∈ R}
and E0 := E⊥2 are T -invariant bands. Moreover, if P denotes the band projection
onto E2, then
and
T (t) − T (t + τ )P e = 2P e for all t ∈ R
lim
t→∞T (t) − T (t + τ )(I − P )e = 0.
To simplify notation, for t ∈ R we define the positive operators
S(t) := T (t) ∧ T (t + τ )
and D(t) := T (t) − T (t + τ )
on E. It follows immediately that
(A.1)
S(t)x +
1
2
D(t)x = x
for all x ∈ Fix(T ) and t ∈ R. Further properties of S(t) and D(t) are provided by
the following lemma.
Lemma A.2. Let x′ ∈ Fix(T ′) be strictly positive and τ ∈ R. Then the following
assertions hold.
(a) D(t)T (s) ≥ D(t + s) and T (s)D(t) ≥ D(t + s) for all t, s ∈ R. Moreover, if
(b) S(t)T (s) ≤ S(t + s) and T (s)S(t) ≤ S(t + s) for all t, s ∈ R. Moreover, if
(c) If limt→∞ S(t)x > 0 for all 0 < x ∈ Fix(T ), then limt→∞ S(t)mx > 0 for all
x ∈ Fix(T ), then limt→∞ D(t)x ∈ Fix(T ).
x ∈ Fix(T ), then limt→∞ S(t)x ∈ Fix(T ).
m ∈ N and x ∈ Fix(T ), x > 0.
Proof. (a) For all t, s ∈ R one observes that
D(t)T (s) = T (t) − T (t + τ ) · T (s) ≥ (T (t) − T (t + τ ))T (s) = D(t + s)
and similarly that T (s)D(t) ≥ D(t + s). Let x ∈ Fix(T ), x ≥ 0. Then
D(t)x = D(t)T (s)x ≥ D(t + s)x
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
15
for all t, s ∈ R. Hence, by the order continuity of the norm, y := lim D(t)x exists
in E and
T (s)y = lim
t→∞
T (s)D(t)x ≥ lim
t→∞
D(t + s)x = y ≥ 0.
Finally, we conclude from hx′, T (s)y− yi = 0 that y ∈ Fix(T ) because x′ is strictly
positive.
(b) For all t, s ∈ R one observes that
S(t)T (s) =
≤
1
2
1
2
(T (t + s) + T (t + τ + s) − D(t)T (s))
(T (t + s) + T (t + τ + s) − D(t + s)) = S(t + s)
and similarly that T (s)S(t) ≤ S(t + s). Hence, S(t)x is increasing for all positive
x ∈ Fix(T ). Since 0 ≤ S(t) ≤ 1
2 (T (t) + T (t + τ )), we conclude as in the proof of
part (a) that lim S(t)x exists in Fix(T ) for all x ∈ Fix(T ).
(c) Let x ∈ Fix(T ), x > 0, and define recursively xk := limt→∞ S(t)xk−1 for all
k ∈ N where x0 := x. Then xk ∈ Fix(T ) by part (b) and xk > 0 by assumption.
It follows by induction that
S(t)mx − xm =
m
Xj=1
S(t)m−j(S(t)xj−1 − xj )
for all m ∈ N and t ∈ R. As kS(t)k ≤ supt∈R kT (t)k, we conclude that
S(t)mx = xm > 0
lim
t→∞
for all m ∈ N.
(cid:3)
The key for the proof of the zero-two law is the following combinatorial lemma.
2
√m
.
m
Xj=1
+ 2(cid:19) ≤
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m
j (cid:19) −(cid:18) m
Lemma A.3. For every m ∈ N
2−m(cid:18) m
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m
j (cid:19) −(cid:18) m
Xj=1
Proof. For k ∈ N and m = 2k − 1 we have
k−1
(cid:18)(cid:18)m
j (cid:19) −(cid:18) m
j − 1(cid:19)(cid:19) + 2
Xj=1
= 2(cid:18) m
k − 1(cid:19) = (cid:18)2k
k (cid:19)
(cid:18)(cid:18)m + 1
j (cid:19) −(cid:18)m + 1
j − 1(cid:19)(cid:19) + 1
Xj=1
(cid:12)(cid:12)(cid:12)(cid:12)
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m + 1
j (cid:19) −(cid:18)m + 1
+ 2 = 2
1
2
m+1
Xj=1
+ 1.
k
=
=
It follows from Stirling's formula that the central binomial coefficient can be esti-
mated by
(cid:18)2k
k (cid:19) ≤
22k
√2k
(k ∈ N).
16
MORITZ GERLACH
Thus, we obtain that
2−m(cid:18) m
Xj=1
(cid:12)(cid:12)(cid:12)(cid:12)
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m
j (cid:19) −(cid:18) m
which completes the proof.
+ 2(cid:19) = 2−(m+1)(cid:18) m+1
(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
k (cid:19) ≤
= 2−m(cid:18)2k
(cid:18)m + 1
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
j (cid:19) −(cid:18)m + 1
2
√m + 1 ≤
2
√m
,
+ 2(cid:19)
(cid:3)
Proof of Theorem A.1. By Lemma A.2 (b), we have
S(t)T (s)y ≤ S(t)T (s)y ≤ S(t + s)y = 0
for all y ∈ E2 and t, s ∈ R, which shows that E2 is T -invariant. Let x′ ∈ Fix(T ′)
be strictly positive. Since
0 ≤ T (t)P e = P T (t)P e ≤ P T (t)e = P e
and hx′, P e − T (t)P ei = 0 for all t ∈ R, it follows that P e ∈ Fix(T ). Define
e0 := (I − P )e ∈ E0 and e2 := P e ∈ E2. As e0 = e− e2 ∈ Fix(T ), we conclude that
E0, which equals the closure of the principle ideal generated by e0, is T -invariant.
It follows immediately from (A.1) that D(t)e2 = 2e2 for all t ∈ R. Hence, it
remains to show that lim D(t)e0 = 0. For simplicity, we omit the index 0 and write
E = E0, e = e0, T (t) = T (t)E0 and so on. Now, lim S(t)y > 0 for all y ∈ Fix(T ),
y > 0, by Lemma A.2 (b) and the definition of E0.
Assume that h := lim D(t)e > 0. As h ≤ 2e, there exists m ∈ N such that
k :=(cid:18)h −
2
√m
e(cid:19)+
> 0.
Moreover, k ∈ Fix(T ) since the fixed space is a sublattice. Indeed, if y ∈ Fix(T ),
it follows from T (t)y ≥ T (t)y = y and hx′, T (t)y − yi = 0 for all t ∈ R
that y ∈ Fix(T ) because x′ is strictly positive. Now, Lemma A.2 (c) yields that
S(t0)mk > 0 for some t0 ∈ R. Let t1 := m(t0 + τ ) and define the operator
(A.2)
U := T (t1) − 2−mS(t0)m(cid:18)I + T (τ )(cid:19)m
.
It follows from S(t0)(I + T (τ )) ≤ T (t0 + τ ) + T (t0)T (τ ) = 2T (t0 + τ ) that U is
positive. Moreover,
(A.3)
T (nt1) = U n + Rn2−m(cid:18)I + T (τ )(cid:19)m
for all n ∈ N where R1 := S(t0)m and Rn+1 := U nR1 + RnT (t1). We infer from
e = U ne + Rne that 0 ≤ Rne ≤ e and 0 ≤ U ne ≤ e for all n ∈ N. Now, by Lemma
A.2 (a) and Lemma A.3, we obtain that
h ≤ D(nt1)e = T (nt1)(I − T (τ ))e
Xj=0
e
m
m
j (cid:19)T j(τ )(I − T (τ ))(cid:12)(cid:12)(cid:12)(cid:12)
≤ 2U ne + Rn2−m(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m
j − 1(cid:19)T j(τ )(cid:12)(cid:12)(cid:12)(cid:12)
= 2U ne + Rn2−m(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18) m
(cid:18)m
j (cid:19)T j(τ ) −
Xj=1
Xj=0
(cid:12)(cid:12)(cid:12)(cid:12)
j − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)m
j (cid:19) −(cid:18) m
e(cid:19)
= 2U ne + Rn2−m(cid:18)2e +
m+1
m
e
≤ 2U ne + Rn
2
√m
Xj=1
e ≤ 2U ne +
2
√m
e
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
17
for every n ∈ N. Let y := limn→∞ U ne ≥ 0. Then h ≤ 2y + 2√m e and hence
0 < k =(cid:18)h −
2
√m
e(cid:19)+
≤ 2y.
Since y is a fixed point of U , equation (A.3) yields that T (nt1)y ≥ y ≥ 0 and we
conclude from hx′, T (nt1)y− yi = 0 that T (nt1)y = y for every n ∈ N. By equation
(A.2) we have
0 = S(t0)m(cid:18)I + T (τ )(cid:19)m
y ≥ S(t0)my ≥ 0.
Therefore, 0 < S(t0)mk ≤ 2S(t0)my = 0 which contradicts the preceded observation
that S(t0)mk > 0. Hence, h = lim D(t)e = 0.
(cid:3)
Appendix B. Axmann's theorem
We give a proof Axmann's theorem from [7, Satz 3.5] stating that not all powers
of an irreducible Harris operator can be disjoint. A proof of this for E = Lp can
be found in [4, Sec. 6].
We start with a version for L-spaces and reduce the general case to it in what
follows. Let us recall that a Banach lattice E is said to be a L-space if
kx + yk= kxk + kyk
holds for all x, y ∈ E+.
Proposition B.1. Let T be a positive and irreducible operator on a L-space E
such that T 6∈ (E′ ⊗ E)⊥. Then there is n ∈ N, n ≥ 2, such that T ∧ T n > 0.
Proof. Aiming for a contradiction, we assume that T ∧ T n = 0 for all n ≥ 2. Since
E is a L-space, we may identify E′ with C(K) for some compact space K by
Kakutani's theorem [16, Thm. 2.1.3]. For n, m ∈ N, n ≥ 2, define
An := {T ′h + T ′n(1 − h) : h ∈ C(K), 0 ≤ h ≤ 1} ⊆ C(K)
and
It follows from our assumption and Synnatschke's theorem [16, Prop. 1.4.17] that
On,m := {t ∈ K : h(t) < 1/m for some h ∈ An}.
inf An = (T ′ ∧ T ′n)1 = (T ∧ T n)′1 = 0
for all n ≥ 2. Now, we show that each of the open sets On,m is dense in K.
Assume the opposite. Then there exists a non-empty open set U ⊆ K\On,m for
some n, m ∈ N. By Urysohn's theorem, we can construct a continuous function
g : K → [0, 1
m for some t0 ∈ U . Hence,
g > 0 is a lower bound of An, which is impossible. Therefore, every On,m and, by
Baire's theorem, also G := ∩n,mOn,m is dense in K. Note that for all t ∈ G and
n ≥ 2 one has that
m ] vanishing on On,m such that g(t0) = 1
hT ′′δt ∧ T ′′nδt, 1i = inf{h(t) : h ∈ An} = 0
and consequently T ′′δt ∧ T ′′nδt = 0.
By assumption, there are x ∈ E+ and y′ ∈ E′+ such that T is not disjoint from
R = y′ ⊗ x. Then R′ = x ⊗ y′ corresponds to a rank-one operator µ ⊗ g on C(K)
for some µ ∈ C(K)′+ and g ∈ C(K)+. Since R′ ∧ T ′ ≥ (R ∧ T )′ > 0 there exists
some e ∈ C(K)+, such that := (R′ ∧ T ′)e > 0. For all 0 ≤ h ≤ e and t ∈ K it
follows from
that
= (R′ ∧ T ′)h + (R′ ∧ T ′)(e − h) ≤ R′h + T ′(e − h)
(t) ≤ hR′h, δti + he − h, T ′′δti = g(t)hµ, hi + he − h, T ′′δti.
(cid:3)
Finally, we consider w := T v. If w = 0, then the closed ideal Ev is T -invariant
for all n ∈ N because t ∈ G.
and non-trivial since T 6= 0. If w > 0, then the closure of the T -invariant ideal
J := {z ∈ E : z ≤ c(w + T w + ··· + T kw) for some c > 0 and m ∈ N}
6= 0. Then z′
is non-trivial because w ∈ J and v ∈ J⊥. In both cases, T cannot be irreducible.
Thus, we conclude that T ∧ T n > 0 for some n ≥ 2.
Theorem B.2. Let T be a positive and irreducible operator on E, a Banach lattice
with order continuous norm, such that T 6∈ (E′ ⊗ E)⊥. Then there is n ∈ N, n ≥ 2,
such that T ∧ T n > 0.
Proof. Fix λ > kTk and y′ ∈ E′+, y′
:= (λ − T ′)−1y′ satisfies
λz′ − T ′z′ = y′ and hence T ′z′ ≤ λz′. This implies that the closed ideal {x ∈ E :
hx, z′i = 0} is T -invariant and thus equal to {0}. Therefore, kxkz′ := hz′,xi
defines an order continuous lattice norm on E. Let (F,k·kF ) be the completion
of (E,k·kz′), i.e. the closure of E in (E,k·kz′)′′. Then F is an L-space and, since
kT xkF ≤ λkTkF for all x ∈ E, T uniquely extends to a positive operator T on F .
Now, it follows as in the proof of Proposition 3.2 that T is irreducible and
T 6∈ (F ′ ⊗ F )⊥.
by [17, Lem. IV 9.3]. Thus, for x ∈ E+ and n ∈ N we observe that
F { T (x − y) + T ny : y ∈ F, 0 ≤ y ≤ x}
E {T (x − y) + T ny : y ∈ E, 0 ≤ y ≤ x}
Since the norm on E is order continuous, so is z′ and hence E is an ideal in F
( T ∧ T n)x = inf
= inf
= (T ∧ T n)x.
18
MORITZ GERLACH
Taking the infimum over all 0 ≤ h ≤ e yields that (t) ≤ (g(t)µ ∧ T ′′δt)e for all
t ∈ K. Now, fix t ∈ {s ∈ K : (s) > 0} ∩ G which exists since G is dense in K.
Then ν := g(t)µ ∧ T ′′δt > 0 because (t) > 0. As ν is dominated by g(t)µ and µ
corresponds to x ∈ E, ν itself corresponds to a vector v ∈ E, v > 0, since E is an
ideal in E′′. This vector v satisfies
v ∧ T nv ≤ T ′′δt ∧ T ′′n+1δt = 0
As E+ is dense in F+, this shows that T ∧ T n = 0 if and only if T ∧ T n = 0. Thus,
the assertion follows from Proposition B.1.
(cid:3)
Acknowledgements
The author was supported by the graduate school Mathematical Analysis of
Evolution, Information and Complexity during the work on this article and he would
like to thank Wolfgang Arendt for many helpful discussions and the anonymous
referee for his/her constructive comments.
References
[1] C. D. Aliprantis and O. Burkinshaw. Positive operators, volume 119 of Pure and Applied
Mathematics. Academic Press Inc., Orlando, FL, 1985.
[2] W. Arendt. Über das Spektrum regulärer Operatoren. PhD thesis, Eberhard-Karls-Universität
zu Tübingen, 1979.
[3] W. Arendt. On the o-spectrum of regular operators and the spectrum of measures. Math. Z.,
178(2):271 -- 287, 1981.
[4] W. Arendt. Positive semigroups of kernel operators. Positivity, 12(1):25 -- 44, 2008.
[5] W. Arendt and C. J. K. Batty. Tauberian theorems and stability of one-parameter semigroups.
Trans. Amer. Math. Soc., 306(2):837 -- 852, 1988.
[6] W. Arendt, C. J. K. Batty, M. Hieber, and F. Neubrander. Vector-valued Laplace transforms
and Cauchy problems, volume 96 of Monographs in Mathematics. Birkhäuser Verlag, Basel,
2001.
IRREDUCIBLE SEMIGROUPS OF HARRIS OPERATORS
19
[7] D. Axmann. Struktur-und Ergodentheorie irreduzibler Operatoren auf Banachverbänden. PhD
thesis, Eberhard-Karls-Universität zu Tübingen, 1980.
[8] V. I. Bogachev. Measure theory. Vol. I, II. Springer-Verlag, Berlin, 2007.
[9] E. B. Davies. Triviality of the peripheral point spectrum. J. Evol. Equ., 5(3):407 -- 415, 2005.
[10] R. Émilion. Mean-bounded operators and mean ergodic theorems. J. Funct. Anal., 61(1):1 -- 14,
1985.
[11] K. Engel and R. Nagel. One-parameter semigroups for linear evolution equations, volume
194 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2000.
[12] G. Greiner. Zur Perron-Frobenius-Theorie stark stetiger Halbgruppen. Math. Z., 177(3):401 --
423, 1981.
[13] G. Greiner. Spektrum und Asymptotik stark stetiger Halbgruppen positiver Operatoren.
Sitzungsber. Heidelb. Akad. Wiss. Math.-Natur. Kl., pages 55 -- 80, 1982.
[14] V. Keicher. On the peripheral spectrum of bounded positive semigroups on atomic Banach
lattices. Arch. Math. (Basel), 87(4):359 -- 367, 2006.
[15] U. Krengel. Ergodic theorems, volume 6 of de Gruyter Studies in Mathematics. Walter de
Gruyter & Co., Berlin, 1985.
[16] P. Meyer-Nieberg. Banach lattices. Universitext. Springer-Verlag, Berlin, 1991.
[17] H. H. Schaefer. Banach lattices and positive operators. Springer-Verlag, New York, 1974. Die
Grundlehren der mathematischen Wissenschaften, Band 215.
[18] N. T. Varopoulos. Sets of multiplicity in locally compact abelian groups. Ann. Inst. Fourier
(Grenoble), 16(fasc. 2):123 -- 158, 1966.
[19] M. P. H. Wolff. Triviality of the peripheral point spectrum of positive semigroups on atomic
Banach lattices. Positivity, 12(1):185 -- 192, 2008.
University of Ulm, Institute of Applied Analysis, 89069 Ulm, Germany
E-mail address: [email protected]
|
1212.2817 | 1 | 1212 | 2012-12-12T13:54:17 | Covering $L^p$ spaces by balls | [
"math.FA",
"math.GN"
] | We prove that, given any covering of any separable infinite-dimensional uniformly rotund and uniformly smooth Banach space $X$ by closed balls each of positive radius, some point exists in $X$ which belongs to infinitely many balls. | math.FA | math | COVERING Lp SPACES BY BALLS
Vladimir P. Fonf ∗, Michael Levin †and Clemente Zanco ‡
2
1
0
2
c
e
D
2
1
]
.
A
F
h
t
a
m
[
1
v
7
1
8
2
.
2
1
2
1
:
v
i
X
r
a
Abstract
We prove that, given any covering of any separable infinite-dimensional uni-
formly rotund and uniformly smooth Banach space X by closed balls each of positive
radius, some point exists in X which belongs to infinitely many balls.
2000 Mathematics Subject Classification: Primary 46B20; Secondary 54D20.
Key words and phrases: point finite coverings, slices, uniformly rotund spaces, uni-
formly smooth spaces
∗Research of the first author was supported in part by Israel Science Foundation, Grant # 209/09 and
by the Istituto Nazionale di Alta Matematica of Italy.
†Research of the second author was supported in part by Israel Science Foundation, Grant # 836/08
‡Research of the third author was supported in part by the Ministero dell'Universit`a e della Ricerca
Scientifica e Tecnologica of Italy and by the Center for Advanced Studies in Mathematics at the Ben-
Gurion University of the Negev, Beer-Sheva, Israel.
1
1
Introduction and statement of the main result
In the present paper X always denotes a real separable infinite-dimensional Banach space;
by ball in X we mean a closed ball. Let A be a collection of subsets of X. We say that
A is point-finite if every point of X belongs to at most finitely many elements of A. A
point x of X is said to be a regular point for A if there is a neighborhood of x that meets
at most finitely many elements of A; x is said to be a singular point for A otherwise. If
every point of X is a regular point for A, we say that A is locally-finite (clearly, that is
equivalent to the requirement that no compact set in X meets infinitely many members
of A). A is said a covering of X if each point of X belongs to some member of A.
The aim of the present paper is to take a step forward to answering the following
Question 1.1 Which infinite-dimensional Banach spaces admit point-finite coverings by
balls (each of positive radius)?
In order to explain how such a question arises, we recall the following two results (the
first one is well known as "the Corson's Theorem").
Theorem 1.2 ([Co]) No (infinite-dimensional) reflexive Banach space admits locally fi-
nite coverings by bounded closed convex sets.
Theorem 1.3 ([MZ]) Any real Banach space X can be covered by bounded closed convex
sets, each with nonempty interior, in such a way that no point of X belongs to more than
two of them.
The family of sets exhibited in the general construction used to prove Theorem 1.3 is
very far from being a family of balls in the original norm of X. Moreover, some classical
Banach spaces admit point-finite coverings by balls. For instance, it is easy to check that
the covering of c0 that can be obtained by translating the unit ball without overlapping
interiors is even locally finite. V. Klee proved in [Kl] that the space l1(Γ) for suitable
(uncountable) Γ can be covered by translates of its unit ball without overlapping them
at all. So Question 1.1 seems to be very natural, though providing it with a complete
answer seems not to be an easy matter. A first step in that direction have been recently
made with the following Theorem, that excludes Hilbert spaces from the class of spaces
Question 1.1 asks for. Even if, looking for spaces outside that class, Hilbert spaces appear
as the simplest ones to be considered, up to now no elementary argument for getting such
exclusion seems to be available.
Theorem 1.4 ([FZ], Theorem 3.2) No covering by balls, each of positive radius, of the
infinite-dimensional separable Hilbert space can be point-finite.
We refer to [FZ] also for more details and references concerning the subjects involved in
the present Introduction. The goal of this paper is to extend Theorem 1.4 to a considerably
wider class of spaces. In fact we prove the following
2
Theorem 1.5 (Main result) No covering by balls, each of positive radius, of any infinite-
dimensional uniformly rotund and uniformly smooth separable Banach space can be point-
finite.
It is very well known that those Banach spaces that are uniformly rotund or uniformly
smooth are reflexive (in fact, super-reflexive). Moreover, if a Banach space X is uniformly
rotund or uniformly smooth, then an equivalent norm can be put on X under which X
is both uniformly rotund and uniformly smooth. Among those spaces that are both
uniformly rotund and uniformly smooth there are Lp(µ) spaces for any measure µ and
p ∈ (1, +∞), so in particular we claim
Corollary 1.6 No covering by balls, each of positive radius, of a separable infinite-
dimensional Lp(µ) space, 1 < p < ∞, µ any measure, can be point-finite.
The proof of Theorem 1.5 we provide here is based on a key result of [FZ]; however, after
that, it follows a completely different way than what was used in [FZ] to get Theorem
1.4.
In fact the argument there runs as follows. Separable polyhedral Banach spaces
are first characterized as those whose unit sphere under some equivalent norm admits
a point-finite covering by slices of the unit ball that do not contain the origin. (Recall
that a Banach space is called "polyhedral" if the unit ball of any its finite-dimensional
subspace is a polytope.) As a consequence, if the unit sphere of some separable Banach
space X admits such a covering, then X must be isomorphically polyhedral. It is well
known that no (infinite-dimensional) Hilbert space is isomorphically polyhedral. Next,
point-finite coverings of the Hilbert space by balls (if any) are easily reduced to point-
finite coverings of the unit sphere by balls that do not contain the origin. Finally, to get
a contradiction, these coverings are reduced to point-finite coverings of the unit sphere
by slices of the unit ball via the following observation: whenever two spheres in an inner
product space do not coincide and have nonempty intersection, such an intersection lies in
some hyperplane; this hyperplane splits each of the two balls determined by those spheres
in two complementary slices. Unfortunately such a situation characterizes inner product
spaces (see [Am] (15.17)), so the argument cannot be applied outside that class of spaces.
Our argument here in proving Theorem 1.5 has an essential topological component:
Corson's Theorem 1.2, which is based on Brouwer's fixed point Theorem, is now our basic
tool. We use it in connection with suitable considerations of geometrical nature.
Throughout the paper we use standard Geometry of Banach Spaces notation as in
In particular, for x ∈ X and r > 0, B(x, r) and S(x, r) respectively denote the
[JL].
closed ball and the sphere with center at x and radius r; moreover, B(0, 1) and S(0, 1)
are denoted in short respectively by BX and SX.
2 Proof of the main result
In order to prove Theorem 1.5, first of all we notice that, X being separable, we can
confine ourselves to prove our theorem for countable coverings. (In fact, let {xn} be any
3
sequence dense in X: since each ball has nonempty interior, for some n0 it must happen
that xn0 belongs to uncountably many balls.)
We start by borrowing from [FZ] the following Proposition. It describes a quite general
situation in which a sequence of slices of the unit ball of any separable Banach space cannot
be point-finite.
Proposition 2.1 ([FZ], Proposition 2.1) Let {fi}∞
functionals on X and {αi}∞
sequence {Si}∞
i=1 of slices of BX defined by
i=1 be a sequence of norm-one linear
i=1 a sequence in the interval (0, 1) converging to 0. Then the
Si = {x ∈ BX : fi(x) ≥ αi}, i = 1, 2, ...
is not point-finite.
Next we point out a very simple (probably well known) fact.
Roughly speaking, it simply states that any point of a sphere of any uniformly smooth
Banach space admits "almost flat" neighborhoods (relative to the sphere) of "big" dia-
mater provided that the radius of the sphere is "big". We make this sentence precise in
the following way.
Fact 2.2 Let X be uniformly smooth. For any ε > 0 there exists b > 0 such that, for any
R > b and x ∈ RSX, if Γx is the hyperplane supporting RBX at x, then
dist(y, Γx) ≤ ε ∀y ∈ RSX ∩ B(x, 2).
(1)
Proof. For t ∈ SX, denote by ft the (only) norm-one linear functional such that
ft(t) = 1. Fix ε > 0. By definition of uniform smoothness, for any ε > 0 there exists
δ = δ(ε) > 0 such that for any w, z ∈ SX with w − z ≤ δ, the following estimate holds
1 − fz(w) = z + (w − z) − 1 − fz(w − z) ≤ εw − z/2.
So for x = Rz, y = Rw ∈ RSX with y − x ≤ min{Rδ, 2} it is true that
dist(y, Γx) = R − fz(y) ≤ εy − x/2 ≤ ε.
By assuming b = 2/δ we are done. (cid:4)
The previous fact allows us to say that, in a uniformly smooth space, a "big" sphere
intersecting a "small" ball splits it in two parts that are not "too far" from being slices
of the small ball. So Proposition 2.1 leads us to the following Proposition, which will be
crucial for our purposes and may have interest by itself.
Proposition 2.3 Let B = {B(xn, Rn)}∞
n=1 be a countable collection of balls in a uniformly
smooth Banach space X such that Rn goes to infinity with n. If B is not locally finite,
then it is not point-finite.
4
Proof. Since B is not locally finite, there exist a point x ∈ X and a sequence {nk} of
integers such that
dist(x, B(xnk , Rnk) → 0 as k → ∞.
(2)
Without loss of generality we may assume x = 0 and {nk} = {k}. Moreover, we may
assume that 0 /∈ B(xn, Rn) for every n and that Rn > 2 and xn < Rn + 1. For any
n, let zn be the point at which the segment [0, xn] meets S(xn, Rn) and let Γn be the
hyperplane supporting B(xn, Rn) at zn. Let 0 < βn < 1 and fn ∈ SX ∗ be such that
Γn = {t ∈ X : fn(t) = βn}. Because of (2), clearly zn → 0 and βn → 0 as n → ∞.
By Fact 2.2, for any i ∈ N big enough there is ni such that dist(y, Γni) < 1/i for every
y ∈ S(xni, Rni) ∩ B(zni, 2): hence the set
Ti = {t ∈ B(zni, 2) : fni(t) ≥ βni + 1/i}
is a slice of B(zni, 2) contained in B(xni, Rni). We can choose ni+1 > ni for every i. Under
all our assumptions, for every i big enough (since BX ⊂ B(zni, 2)) the set
Si = {t ∈ BX : fni(t) ≥ βni + 1/i}
is a slice of BX that is contained in Ti. The sequence {Si} of such slices of BX satisfies
the assumptions of Proposition 2.1, so it is not point-finite. Since we have
we are done. (cid:4)
Si ⊂ Ti ⊂ B(xni, Rni)
The following Fact sounds, in some sense, as the converse of Fact 2.2.
Fact 2.4 Let X be uniformly rotund and b > 0. Let {Bn = B(xn, Rn)}∞
of balls in X such that Rn > b and x0 /∈ intBn for every n. Put
n=1 be a sequence
Fn = conv(Bn \ intB(x0, b))
and let
Then Rn → ∞ with n.
dist(x0, Fn) → 0 as n goes to ∞.
(3)
Proof. Let us recall that the following statement is equivalent to X being uniformly
rotund:
(UR)
for every β > 0 there exists α = α(β) > 0 such that, for any x ∈ SX and
f ∈ SX ∗ with f (x) = 1, the slice {t ∈ BX : f (t) ≥ 1 − α} of BX is contained in B(x, β).
Of course we may assume that function α(β) is increasing; it goes to 0 as β does.
Without loss of generality may assume x0 = 0.
5
Assume to the contrary that some subsequence {Rnk} of {Rn} is bounded: without
loss of generality we may assume that {Rn} itself is bounded. For any n, put
εn = xn − Rn ≥ 0 and Bn = B(xn, Rn + εn);
moreover, let fn ∈ SX ∗ be such that fn( Bn) ⊂ (−∞, 0]. Let αn = α(
α = α(β) is the number defined in condition (UR). We have
b
Rn + εn
), where
{t ∈ Bn : fn(t) ≥ −αn} ⊂ bBX
hence
(Bn \ intbBX ) ⊂ ( Bn \ intbBX ) ⊂ {t ∈ Bn) : fn(t) ≤ −αn}.
This last set is closed and convex, so
Fn = conv(Bn \ intbBX ) ⊂ {t ∈ Bn : fn(t) ≤ −αn}.
(4)
Since fn ∈ SX ∗, (4) implies dist(0, Fn) ≥ αn for every n: that contradicts (3) because
{Rn + εn} is bounded, so {αn} = {α(
)} is far away from 0. (cid:4)
b
Rn + εn
The following three Lemmas are of a technical nature. The first one is something like
a local version of Corson's Theorem 1.2.
Lemma 2.5 Let X be reflexive. Let x0 ∈ X, a > b > c > 0 and F a collection of closed
convex subsets of X contained in B(x0, a) \ intB(x0, c) such that F covers B(x0, a) \
intB(x0, b). Then F is not locally finite in X.
Proof. Without loss of generality we may assume x0 = 0. Suppose by contradiction that
F is locally finite. Set Fn = {(a/b)nF : F ∈ F }, n ∈ Z where αF = {αy : y ∈ F }.
Denote F ′ = ∪n∈ZFn. Then F ′ covers X \ {0} and 0 is the only singular point of F ′.
Split X into two closed half spaces X = X + ∪ X − such that X + ∩ X − is a hyperplane of
X, 0 /∈ X + ∩ X − and 0 ∈ X −. Let F + = {X + ∩ F : F ∈ F ′} and let F − be the covering
of X − which is a symmetric reflection of F + with respect to the hyperplane X + ∩ X −
(made via any line not contained in X + ∩ X −). Then F + ∪ F − is a locally finite covering
of X by bounded closed convex sets and we get a contradiction with Corson's Theorem
1.2. (cid:4)
Lemma 2.6 Let X be both uniformly rotund and uniformly smooth. Consider a closed
hyperplane X ′ ⊂ X and a ball B(x0, a), x0 ∈ X ′, a > 0. Assume that B = {Bn}∞
n=1 is
a countable point-finite collection of balls and F = {Fn}∞
n=1 is a countable collection of
closed convex sets such that F covers B(x0, a) ∩ X ′, Fn ⊂ Bn ∩ B(x0, a) and x0 /∈ intBn
for every n. Then there is a point y ∈ B(x0, a) ∩ X ′, y 6= x0, that is a singular point for
F .
6
Proof. Assume that every y ∈ B(x0, a) ∩ X ′, y 6= x0, is a regular point of F . Take any
n ∩ Fn ∩ X ′ and
b such that a > b > 0 and define B ′
F ′ = {F ′
n ∩ B(x0, c) = ∅
for every n and then our Lemma will follow from Lemma 2.5 applied to the collection F ′
with X being replaced by X ′.
n = B ′
n=1. We will show that there is c > 0 such that b > c and F ′
n = conv(Bn \ intB(x0, b)), F ′
n}∞
Assume that such c does not exist. Then any arbitrarily small neighborhood of x0
n for some n so, since x0 /∈ intBn, from Fact 2.4 we deduce that there is a
intersect B ′
subsequence of balls in B whose radii go to infinity for which x0 is a singular point. By
Proposition 2.3, this contradicts the pointwise finiteness of B. The Lemma is proved. (cid:4)
Lemma 2.7 Let X be both uniformly rotund and uniformly smooth. Let B = {Bn}∞
{Bn = B(xn, Rn)}∞
and B#
moreover, for every n we have that B#
B#.
n=1 =
1 = B1
n=1 is a covering of X;
n ⊂ Bn and any x0 ∈ intBn is a regular point for
n=1 be a countable point-finite covering of X by balls. Put B#
n+1 = conv((Bn+1 \ (B1 ∪ . . . ∪ Bn)). Then B# = {B#
n }∞
Proof. Assume that, for some n, x0 ∈ IntBn is a singular point for B#. Then for some
subsequence {B#
i=1 with ni > n it happens that x0 /∈ Bni for every i and B(x0, 1/i)
ni}∞
intersects the set B#
nj for every j ≥ i. Note that B#
ni ⊂ conv(Bni \ Bn) ⊂ Bni. Then
B(x0, 1/i) intersects conv(Bnj \ Bn) for j ≥ i. Let b > 0 such that B(x0, b) ⊂ Bn (we
assumed x0 ∈ intBn): of course
conv(Bnj \ B(x0, b) ⊃ conv(Bnj \ Bn)
hence, since and x0 /∈ Bni for every i, from Fact 2.4 we get that Rni −→ ∞ as i −→ ∞.
By Proposition 2.3 this contradicts the assumption that B is point-finite. (cid:4)
We are now ready to prove our main result.
Proof of Theorem 1.5. Assume that B = {Bn}∞
n=1 is a point-finite covering by balls
of a uniformly rotund and uniformly smooth Banach space X. Consider the covering B#
from Lemma 2.7 and let S be the set of the points that are singular for B#. By Theorem 1.2
we have S 6= ∅ and by Lemma 2.7 we have S ⊂ ∪n∂Bn. Clearly S is closed in X, hence by
the Baire cathegory theorem there are x0 ∈ S, a > 0 and Bm such that S∩B(x0, a) ⊂ ∂Bm.
Take a closed hyperplane X ′ in X passing through x0 and intersecting Bm only at x0.
Then, by applying Lemma 2.6 to the collection F = {B#
n ∈ B#} of closed
convex sets in X with respect to the hyperplane X ′, we get a contradiction. (cid:4)
n ∩ B(x0, a) : B#
7
References
[Am] D. Amir, Characterizations of inner product spaces, Operator Theory: Advances
and Applications, 20, Birkhauser Verlag, Basel, 1986.
[Co] H. H. Corson, Collections of convex sets which cover a Banach space, Fund. Math.
49 (1961), 143-145.
[FZ] V. P. Fonf and C. Zanco, Covering the unit sphere of certain Banach spaces by
sequences of slices and balls, to appear in Canad. Math. Bull., doi:10.4153/CMB-
2012-027-7.
[JL] W. B. Johnson and J. Lindenstrauss, Basic Concepts in the Geometry of Banach
Spaces, Handbook of the Geometry of Banach Spaces vol. 1, edited by W.B. Johnson
and J. Lindenstrauss, Elsevier Science B.V. (2001), 1-84.
[Kl] V. Klee, Dispersed Chebyshev sets and covering by balls, Math. Ann. 257 (1981),
251 -- 260. MR 84e:41036
[MZ] A. Marchese and C. Zanco, On a question by Corson about point-finite coverings,
Israel J. Math. 189 (2012), 55 -- 63.
Vladimir P. Fonf
Department of Mathematics
Ben-Gurion University of the Negev
84105 Beer-Sheva, Israel
E-mail address: [email protected]
Michael Levin
Department of Mathematics
Ben-Gurion University of the Negev
84105 Beer-Sheva, Israel
E-mail address: [email protected]
Clemente Zanco
Dipartimento di Matematica
Universit`a degli Studi
Via C. Saldini, 50
20133 Milano MI, Italy
E-mail address: [email protected]
ph. ++39 02 503 16164
fax ++39 02 503 16090
8
|
1610.01535 | 1 | 1610 | 2016-10-05T17:21:51 | A retract theorem for nilpotent Lie groups | [
"math.FA",
"math.OA"
] | Let $G= \exp(\g)$ be a connected, simply connected nilpotent Lie group. We show that for every $G$-invariant smooth sub-manifold $M$ of $g^*$, there exists an open relatively compact subset $\mathcal{M}$ of $M$ such that for any smooth adapted field of operators $(F(l))_{l\in M}$ supported in $G\cdot \mathcal{M}$ there exists a Schwartz function $f$ on $G$ such that $\pi_l(f)= \op_{F(l)}$ for all $l\in M$. This retract theorem can then be used to show that for every Lie group $\G$ of automorphisms of $G$ containing the inner automorphisms of $G$ with locally closed $\G$-orbits in $\g^*$, the proper $\G$-prime two-sided closed ideals of $L^1(G)$ are the kernels of $\G$-orbits in $\hat{G}$. | math.FA | math |
A retract theorem for nilpotent Lie groups
Ying-Fen Lin, Jean Ludwig and Carine Molitor-Braun
Abstract
1 Let G = exp(g) be a connected, simply connected nilpotent Lie group. We
show that for every G-invariant smooth sub-manifold M of g∗, there exists an open
relatively compact subset M of M such that for any smooth adapted field of operators
(F (l))l∈M supported in G · M there exists a Schwartz function f on G such that
πl(f ) = opF (l) for all l ∈ M . This retract theorem can then be used to show that
for every Lie group G of automorphisms of G containing the inner automorphisms
of G with locally closed G-orbits in g∗, the proper G-prime two-sided closed ideals
of L1(G) are the kernels of G-orbits in bG.
1
Introduction
For a connected, simply connected, nilpotent Lie group G, the description of its spectrum
and of the Fourier inversion theorem is due to Kirillov [3], who showed that the dual space
bG of G is in one-to-one correspondence with the space g∗/G of co-adjoint orbits of G.
R. Howe proved in [2] that for every irreducible unitary representation (π, Hπ) of G and
every smooth linear operator a on Hπ there exists a Schwartz function fa on G such that
π(fa) = a. He also showed that the mapping a 7→ fa is linear and continuous with respect
to the Fr´echet topology of the space B∞(Hπ) of smooth linear operators on Hπ and the
Fr´echet space S(G) of Schwartz functions on G.
In this paper, we study a version of the Fourier inversion theorem for nilpotent Lie
groups. More precisely, we generalise the result of R. Howe's mentioned above by con-
structing a continuous retract from the space of adapted smooth kernel functions defined
on a smooth G-invariant sub-manifold M of g∗ and supported in a subset G · M of M,
where M is a relatively compact open subset of M, into the space S(G). We will prove
this result, which we call the retract theorem, by proceeding an induction on the length
I of the largest index set I for which (B × g∗)I ∩ M 6= ∅. In order to do so we will apply
the variable group techniques developed in [11], which have already been used in [10].
Once we have the retract theorem, we can apply it to study the G-prime ideals of the
Banach algebra L1(G). Here G denotes a Lie subgroup of the automorphism group of
G with the property that the G-orbits in g∗ are all locally closed. The retract theorem
1keywords: nilpotent Lie group, irreducible representation, co-adjoint orbit, Fourier inversion, retract,
compact group action -- 2010 Mathematics Subject Classification: 22E30, 22E27, 43A20
1
implies that the Schwartz functions contained in the kernel of a G-orbit Ω in bG are dense
in the L1(G)-kernel of Ω. Using the methods in [7], it follows that every G-prime ideal
in L1(G) is the kernel of such an G-orbit Ω. One can for instance use this result for the
study of bounded irreducible representations (π, X) of a Lie group G on a Banach space
X. Restricting the representation π to the nilradical G, one obtains the kernel ker(πG) of
πG in the algebra L1(G). This ideal ker(πG) is then G-prime. If ker(πG) is given as the
kernel in L1(G) of a G-orbit G · π0 ⊂ bG for some π0 ∈ bG, then one use π0 to make an
analysis of π as Mackey did in the case of unitary representations.
For connected, simply connected, nilpotent Lie groups, J. Ludwig showed in [6] that
the closed prime ideals of L1(G) coincide with the kernels of the irreducible unitary rep-
resentations. In 1984, D. Poguntke studied the action of an abelian compact group K on
a nilpotent Lie group [12] and characterised the K-prime ideals as kernels of K-orbits. In
[4], R. Lahiani and C. Molitor-Braun identified the K-prime ideals with hull contained
in the generic part of the dual space of G for a general compact Lie subgroup K of the
automorphism group of G. In [7] and [8], it was shown that for an exponential Lie group
G, the G-prime ideals are also kernels of G-orbits. In this way the bounded irreducible
Banach space representations of an exponential Lie group could be determined.
The paper is organised in the following way:
in Section 2 we recall the definition of
induced representations and of kernel functions, we explain the notion of variable nilpotent
Lie groups and their Lie algebras, of index sets for co-adjoint orbits and of adapted kernel
functions on a G-invariant sub-manifold of g∗. In Section 3, we introduce the main theorem
of the paper, the Retract Theorem, and in Section 4 we present the proof of the theorem,
dividing it into several steps. As an application of the Retract Theorem, in the last section
(Section 5) we show that every G-prime ideal in L1(G) is the kernel of a G-orbit.
2 Notations and generalities
2.1 Representations and kernel functions
Let G = exp(g) be a connected, simply connected, nilpotent Lie group and g be its Lie alge-
bra. All the irreducible unitary representations of G (and hence of L1(G)) are obtained (up
to equivalence) in the following way: Let l ∈ g∗ and p = p(l) be an arbitrary polarisation
of l in g (a maximal isotropic subalgebra of g for the bilinear form (X, Y ) 7→ hl, [X, Y ]i).
Let P (l) = exp(p(l)). The induced representation denoted by πl := indG
P (l)χl on the Hilbert
space Hl,
Hl = L2(G/P (l), χl)
:= {ξ : G → C; ξ measurable , ξ(gp) = χl(p−1)ξ(g), g ∈ G, p ∈ P (l),
kξk2
2 =ZG/P (l)
ξ(g)2d g < ∞},
where d g is the invariant measure on G/P (l), is unitary and irreducible. Here χl is the
character defined on P (l) by χl(g) = e−ihl,log gi for all g ∈ P (l). Two different polarisations
2
for the same l give equivalent representations. The same is true for the case of two linear
forms l and l′ belonging to the same co-adjoint orbit.
of Zk, . . . , Zn and lk = lgk for all l ∈ g∗. The polarisation p(l)Z = p(l) := Pn
One particular way to obtain a polarisation is the following: Let {Z1, . . . , Zn} denote
a Jordan-Holder basis of g, for 1 ≤ k ≤ n, let gk := span{Zk, . . . , Zn} be the linear span
k=1 gk(lk) of
l in g, with gk(lk) := {U ∈ gk; hl, [U, gk]i ≡ 0}, is called the Vergne polarisation at l with
respect to the basis Z1, . . . , Zn. We refer to [1] for more details on the theory of irreducible
representations of nilpotent Lie groups.
Let πl = indG
P (l)χl. The corresponding representation of L1(G), also denoted by πl, is
obtained via the formula πl(f )ξ := RG f (x)(cid:0)πl(x)ξ(cid:1)dx, for all ξ ∈ Hl. If f ∈ L1(G), then
πl(f ) is a kernel operator, i.e. it is of the form
(cid:0)πl(f )ξ(cid:1)(g) =ZG/P (l)
F (l, g, u)ξ(u)du,
where F is the operator kernel given by
F (l, g, u) =ZP (l)
f (ghu−1)χl(h)dh for g, u ∈ G.
If f is a Schwartz function, then the kernel function F belongs to C∞ and satisfies the
covariance relation
F (l, gh, g′h′) = χl(h)χl(h′)F (l, g, g′)
for h, h′ ∈ P (l) and g, g′ ∈ G,
and is a Schwartz function on G/P (l) × G/P (l).
2.2 Group actions
Let G = exp(g) be a connected, simply connected, nilpotent Lie group and A be a Lie
subgroup of the automorphism group Aut(G) of G acting smoothly on G. This action will
be denoted by
A × G 7→ G
(a, x)
7→ a · x.
The action of A on G induces naturally actions of A on g, g∗, bG, L1(G), and on S(G).
These group actions will lead to examples for our retract theory and provide an important
application of retracts.
2.3 Variable Lie algebras and groups
We will prove our main theorem by induction; in our proofs, new parameters and new
variations will appear. This may be handled most easily by the concept of variable Lie
structures. Such structures were already considered in [5], [11], [10] and [9], among others.
3
Definition 2.3.1.
1. Let g be a real vector space of finite dimension n and B be an
arbitrary nonempty set. We say that (B, g) is a variable (nilpotent) Lie algebra if
(a) For every β ∈ B, there exists a Lie bracket [·, ·]β defined on g such that gβ :=
(g, [·, ·]β) is a nilpotent Lie algebra.
(b) There exists a fixed basis Z = Z 0 = {Z1 = Z 0
1 , . . . , Zn = Z 0
n} of g such that the
structure constants ak
ij(β) defined by
[Zi, Zj]β :=
nXk=1
ak
ij(β)Zk
satisfy the following property: For all β ∈ B and k ≤ max{i, j}, ak
This means that {Z1, . . . , Zn} is a Jordan-Holder basis for gβ = (g, [·, ·]β).
ij(β) = 0.
2. Assume that B is a smooth manifold. If the structure constants ak
ij(β) vary smoothly
on B, we say that (B, g) is a smooth variable (nilpotent) Lie algebra.
We will denote (B, g) = (g, [·, ·]β)β∈B for the variable Lie algebra.
For the rest of the paper we will assume that all variable Lie algebras are smooth. If
B is reduced to a singleton, we have in fact no dependency on β in B but a fixed Lie
algebra. To each variable Lie algebra, we associate a variable Lie group Gβ. The vari-
able Lie group G := (Gβ)β may be identified with the collection of Lie algebras (g, [·, ·]β)β
equipped with the corresponding Campbell-Baker-Hausdorff multiplications. If G = (Gβ)β
is a (smooth) variable Lie group endowed with a fixed Jordan-Holder basis, then the cor-
responding Vergne polarisations, induced representations and operator kernels all depend
on β ∈ B and l ∈ g∗.
2.4 Ludwig-Zahir indices
Let (B, g) be a smooth variable Lie algebra. We assume that g is equipped with a fixed basis
Z = Z 0 = {Z1 = Z 0
n}, which is a Jordan-Holder basis for every (g, [·, ·]β).
1 , . . . , Zn = Z 0
Let (β, l) ∈ B × g∗. The Ludwig-Zahir indices I(β, l) defined in [11] can be obtained
in the following way: Let gβ(l) := {U ∈ g; hl, [U, g]βi ≡ 0} be the stabiliser of l in gβ =
(g, [·, ·]β) and let aβ(l) be the maximal ideal contained in gβ(l). If aβ(l) = gβ(l) = g, then
χ(β,l)(x) := e−ihl,logβ xi is a character on Gβ and nothing has to be done. In this case, there
are no Ludwig-Zahir indices, i.e. I(β, l) = ∅. Otherwise, let
j1(β, l) = max{j ∈ {1, . . . , n}; Z 0
k1(β, l) = max{k ∈ {1, . . . , n}; hl, [Z 0
j 6∈ aβ(l)}, and
j1(β,l), Z 0
k]βi 6= 0}.
We let
X1(β, l)
Y1(β, l)
Z1(β, l)
c(β, l)
: = Z 0
k1(β,l),
: = Z 0
j1(β,l),
: = [Z 0
k1(β,l), Z 0
: = hl, Z1(β, l)i.
j1(β,l)]β, and
4
We then consider
g1(β, l) := {U ∈ g; hl, [U, Y1(β, l)]βi = 0}
(2.1)
which is an ideal of co-dimension one in gβ.
A Jordan-Holder basis of (g1(β, l), [·, ·]β) is given by Z 1(β, l) = {Z 1
i (β, l); i 6= k1(β, l)}
with
Z 1
i (β, l) := Z 0
i −
hl, [Z 0
i , Y1(β, l)]βi
c(β, l)
X1(β, l),
i 6= k1(β, l).
(2.2)
i (β, l) = Z 0
One sees that Z 1
i , if i > k1(β, l). As previously we may now compute the
indices j2(β, l), k2(β, l) of l1 := lg1(β,l) with respect to this new basis and construct the
corresponding subalgebra g2(β, l) with its associated basis {Z 2
i (β, l) ; i 6= k1(β, l), k2(β, l)}.
This procedure stops after a finite number r of steps. Let
IZ (β, l) = I(β, l) =(cid:0)(j1(β, l), k1(β, l)), . . . , (jr(β, l), kr(β, l))(cid:1),
which is called the Ludwig-Zahir index of l in gβ with respect to the basis {Z1, . . . , Zn}. The
construction in [11] shows that the final subalgebra gr(β, l) obtained by this construction
coincides with the Vergne polarisation of l in gβ with respect to the basis Z 0 (see also [10],
[9]). Note that the length I = 2r of the index set I = I(β, l) gives us the dimension of the
co-adjoint orbit Ad∗(Gβ)ℓ. The vectors Y1(β, l), · · · , Yr(β, l) together with the stabiliser
gβ(l) of l in gβ span the polarisation pβ(l) = gr(β, l) and
g = ⊕r
i=1
RXi(β, l) ⊕r
i=1
RYi(β, l) ⊕ gβ(l).
Let us introduce the following notations: For any index set I ∈ (N2)r ≡ N2r with
r = 0, · · · , dim(g)/2, we let
(B × g∗)I := {(β, l) ∈ B × g∗; I(β, l) = I}
and
(B × g∗)I ∩ ΣI := {(β, l) ∈ (B × g∗)I; l(Zji) = l(Zki) = 0 for 1 ≤ i ≤ r}.
This last line corresponds to the Pukanszky section associated to the index I. In fact, in
[9] it was proved that the indices js(β, l), ks(β, l) coincide with the Pukanszky indices of
the given layer (if one does not make any distinction between the j's and the k's). For
many I's, the subset (B × g∗)I is empty. Hence it is reasonable to define
I :=nI ∈
dim(g)/2[j=0
(N2)j; (B × g∗)I 6= ∅o,
and B × g∗ =
[I∈I
(B × g∗)I.
This gives a partition of B × g∗ into the different layers (B × g∗)I. The set I may be
r′)} ∈ I,
ordered lexicographically:
we say that I < I ′ if either 2r = I < I ′ = 2r′ or there exists a ∈ {1, . . . , r} such that
if I = {(j1, k1), · · · , (jr, kr)}, I ′ = {(j′
1), · · · , (j′
r′, k′
1, k′
(js, ks) = (j′
a),
s, k′
s) if s < a and (ja, ka) < (j′
a, k′
5
which means that
either ja < j′
a).
a or (ja = j′
a and ka < k′
This allows us to define
(B × g∗)≤I := {(β, l) ∈ (B × g∗)J ; J ≤ I} = [J≤I
(B × g∗)J .
By induction on the length of the index sets, it is easy to see that for every I ∈ I there
exists a smooth function PI on B × g∗, which is polynomial in l for fixed β ∈ B such that
(B × g∗)I = {(β, l); PI ′(β, l) = 0 for I ′ > I and PI(β, l) 6= 0}.
(2.3)
2.5 Co-adjoint orbits
For any index set I, we consider the subspace sI of g∗ which is given by
For each β ∈ B, let
sI = span{Z ∗
j ; j ∈ I}.
Σβ,I := {(β, l) ∈ ({β} × g∗)I; l ∈ sI }.
Then Σβ,I is locally closed in sI, since we have the smooth functions PI ′, I ′ ∈ I, defined
on B × g∗ as in (2.3).
Let d := I. For l ∈ g∗, let
Ωβ,l = {Ad∗
β(g)l; g ∈ G}
be the Gβ-orbit of l. Then
dim(Ωβ,l) = d for
l ∈ (B × g∗)I.
There exist functions pj : (B × g∗)I × Rd → R, j = 1, · · · , n, which are rational in l ∈ g∗
and polynomial in z ∈ Rd for fixed β ∈ B such that for every (β, l) ∈ (B × g∗)I,
Ωβ,l =n nXi=1
pi(β, l, z)Z ∗
i ; z ∈ Rdo.
Furthermore if we write I = {i1 < · · · < id}, then
pij (β, l, z) = zj
for j = 1, · · · , d,
and for i 6∈ I, we have
pi(β, l, z) = hl, Zii + p′
i(β, l, z1, · · · , zj),
ij < i < ij+1.
Definition 2.5.1. A subset M of B × g∗ is called G-invariant if for every (β, l) ∈ M the
element g · (β, l) := (β, Ad∗
β(g)l) is also contained in M.
6
2.6 Schwartz functions
Let r ∈ N \ {0}, we define the space of (generalised) Schwartz functions S(Rr, B, G) ≡
S(Rr, B, g) ≡ S(Rr, B, Rn) to be the set of all functions f from Rr × B × G to C such that
the function f defined by
f (α, β, (x1, . . . , xn)) := f (α, β, expβ(x1Z1 + · · · + xnZn))
for α ∈ Rr, β ∈ B
is smooth on Rr × B × Rn and that
k fkK,T1,...,Ts,A1,A2,B1,B2 =
β∈K;α∈Rr;x∈Rnh
sup
sup
αr1xs1
ri≤Ai;sj≤Bj;i,j∈{1,2}
T1T2 · · · Ts
∂r2
∂αr2
∂s2
∂xs2
< +∞,
f (α, β, (x1, . . . , xn))i
for any compact subset K of B, any finite collection T1, . . . Ts of smooth vector fields defined
on the manifold B, and any A1, A2, B1, B2 ∈ N. The function space S(Rr, B, G) is equipped
with the topology defined by the collection of all these semi-norms. One may of course also
use coordinates of the second kind to define the semi-norms on S(Rr, B, G). Note that the
space S(Rr, B, G) does not depend on the choice of the Jordan-Holder basis.
2.7 Kernel functions
Let S be a subset of B ×g∗ and L be a smooth manifold. We say that a mapping F : S → L
is smooth, if the restriction of F to any smooth manifold N contained in S is smooth.
Let B × g∗ be a smooth variable nilpotent Lie group with Jordan-Holder basis Z. For
any (β, l) ∈ B × g∗ denote the Vergne polarisation at (β, l) associated to Z. We put
π(β, l) := indG
P (β,l)χl, with P (β, l) := expβp(β, l), for the corresponding family of induced
unitary representations. Then the mapping (β, l) 7→ p(β, l) is smooth on each subset
(B × g∗)I. For each index set I with length dI and (β, l) ∈ B × g∗, choose a Malvev
basis R(β, l) = {R1(β, l), · · · , RdI (β, l)} of g relative to p(β, l), such that the mappings
(β, l) 7→ R(β, l) are smooth on the different layers (B × g∗)I.
Definition 2.7.1. Let M be any smooth G-invariant manifold of B × g∗ and let r ∈ N.
We denote by Dc
M,r the space of all functions F : Rr × M × G × G → C satisfying the
following conditions.
1. F satisfies the covariance condition for every (β, l) ∈ M with respect to p(β, l), i.e.
F (α, (β, l), x ·β p, y ·β q) = χl(p)χl(q)F (α, (β, l), x, y))
for all α ∈ Rr, p, q ∈ P (β, l) and x, y ∈ G.
7
2. The function F satisfies the following compatibility condition
F (α, (β, Ad∗
β(g)l), x, y) = F (α, (β, l), x ·β g, y ·β g),
for α ∈ Rr, (β, l) ∈ M and x, y, g ∈ G. This compatibility condition reflects the
unitary equivalence of the representations π(β,l) and π(β,Ad∗
β (g)l).
3. The support of F in (β, l) is compact modulo G, i.e. there exists a compact subset C
of M such that F (·, (β, l), ·, ·) is 0 outside the subset of G · C.
4. The function F has the Schwartz space property, i.e.
for any I ∈ I the function
FRr×M ∩(B×sI )×G×G is smooth and that
kF kD,A1,A2,B1,B2,C1,C2
:=
(β,l)∈M,α∈Rr,x,x′∈Rrh
sup
D(β,l)
∂r2
∂αr2
∂s2
∂xs2
∂t2
∂(x′)t2
where
sup
αr1xs1(x′)t1
ri≤Ai,sj≤Bj,tk≤Ck;i,j,k∈{1,2}
F (α, (β, l), x, x′)i < ∞,
(2.4)
F (α, (β, l), x, x′) := F (α, (β, l), expβ(x1R1) · · · expβ(xrRr), expβ(x′
1R1) · · · expβ(x′
rRr)),
for any smooth differential operator D = D(β,l) on the manifold M, and any A1, A2,
B1, B2, C1, C2 ∈ N.
The space Dc
M,r will be equipped with the topology defined by the collection of all these
semi-norms. This does of course not depend on the choice of the smooth Malcev basis of
g with respect to the smooth family of Vergne polarisations.
Definition 2.7.2. Let M ⊂ B × g∗. A field F = (F (β, l))(β,l)∈M of kernel functions is
called adapted if it satisfies the conditions in Definition 2.7.1.
For an adapted field of kernel functions F on M, denote by opF the field of smooth
operators defined through their kernel functions. For (β, l) ∈ M, the operator opF (β,l) acts
on the space L2(G/P (β, l), χ(β,l)) in the following way:
opF (β,l)ξ(g) =ZG/P (β,l)
F (β, l)(g, x)ξ(x)d x.
Remarks 2.7.3. a) If we impose the condition that the support of (β, l) be contained in
the set G · C0 for a fixed subset C0 of M, we will denote the space of kernel functions by
DC0
M .
b) One has a similar definition of the kernel functions if one takes another smooth family
of polarisations together with a smooth family of Malcev bases.
8
3 The retract theorem
In this section, we state the main theorem of the paper which will be proved in the next
section.
Theorem 3.1. Let B × G be a smooth variable nilpotent Lie group, I = {(j1, k1) < · · · <
(jr, kr)} be an index set and let M be a smooth G-invariant sub-manifold of B×g∗ contained
in (B × g∗)≤I such that MI := M ∩ (B × g∗)I 6= ∅. Let π(β, l) be defined as previously
from the smooth family of Vergne polarisations for (β, l) ∈ M. Then there exists an open
nonempty relatively compact subset M ⊂ MI with closure M contained in MI such that
the following holds: For any adapted kernel function F ∈ DM
M , there is a function f in
the Schwartz space S(Rr, B, G) such that π(β,l)(f (α, β, ·)) has F (α, (β, l), ·, ·) as an operator
kernel for all (α, (β, l)) ∈ Rr ×M. Moreover the mapping F 7→ f is continuous with respect
to the corresponding function space topologies.
If the variation is trivial, we get the following theorem.
Theorem 3.2. Let g be a nilpotent Lie algebra with Jordan-Holder basis Z. Let M be
a smooth G-invariant sub-manifold of g∗. Let I := max{J ∈ IZ : M ∩ g∗
J 6= ∅}. Let
πl = π(l) be defined as previously from the smooth family of Vergne polarisations for
l ∈ M. Then there exists an open, relatively compact nonempty subset M ⊂ g∗
I of M such
that M ⊂ M ⊂ MI , M is compact and that the following holds: For any kernel function
F ∈ DM
M , there is a function f in the Schwartz space S(G) such that πl(f ) has F (l, ·, ·) as
an operator kernel for all l ∈ M. Moreover, the Schwartz function f may be constructed
such that the mapping F 7→ f is continuous with respect to the corresponding function
space topologies.
Remark 3.2.1. If M is contained in g∗
we have the following (well known) result.
Imax, where Imax is the maximal index set in I, then
Theorem 3.3. Let B × G be a simply connected, connected smooth variable nilpotent Lie
group and M = (B × g∗)gen := (B × g∗)Imax be the space of generic co-adjoint orbits. Let
M be an open relatively compact subset of M such that M ⊂ M. For every adapted field
of kernel functions F ∈ DM
M , there exists a unique Schwartz function f = R(F ) : G → C
such that
π(β,l)(f ) = opF (β,l) for any (β, l) ∈ B × g∗,
and the mapping F 7→ R(F ) is continuous.
Proof. It suffices to apply the Fourier inversion formula. For each F ∈ DM
M , let
f (β, g) = R(F )(β, g) :=ZΣβ,Imax
tr(π(β,l)(g) ◦ opF (β,l))Pa(β, l)dl, g ∈ G,
where Pa(β, l) is the Pfaffian of the polynomial Q(l) = det (hl, [Zi, Zj]βii,j∈Imax). It follows
from [11] that the function f is Schwartz and the Fourier inversion theorem tells us that
π(β,l)(f ) = opF (β,l) for any (β, l) ∈ B × g∗.
9
4 Proof of the retract theorem
The proof of Theorem 3.1 proceeds by induction on the length I of the largest index set
I for which (B × g∗)I ∩ M 6= ∅ and it will be done in several steps.
4.0.1 The case I = ∅
Suppose that all the elements (β, l) ∈ M are characters of gβ, which means that their index
sets are empty.
Let us replace the variable group (B, G) by the group (C, G), where C = B as a manifold,
and the multiplications coming from C are abelian, i.e. [U, V ]γ = 0 for every U, V ∈ g, γ ∈
C. We identify now the group G with its Lie algebra and then U ·γ V = U + V for every
U, V ∈ g and γ ∈ C. This also means that χl is a character on Gγ = expγg, for all
(γ, l) ∈ C × g∗. Now take M = M. Let F ∈ S(Rr × M) be a kernel function with compact
support in the variables (γ, l). As Rr × M is a sub-manifold of Rr × C × g∗, the function
F may be extended to a Schwartz function eF (in the sense of Section 2.6 and 2.7) on
3 eF , where F −1
Rr × C × g∗ with compact support in the variables (γ, l). Let f := (2π)nF −1
denotes the partial inverse Fourier transform in the variable l which is the third variable
in Rr × C × g∗. Then f ∈ S(Rr × C × g∗). For all (α, (γ, l)) ∈ Rr × M, we have
3
π(γ,l)(cid:0)f (α, γ, ·)(cid:1) = bf 3(α, γ, l)
= (2π)nF3F −1
= F (α, (γ, l)).
3 F (α, (γ, l))
In particular, π(γ,l)(f (α, γ, ·)) = 0 if (α, (γ, l)) ∈ Rr × (M \ C). The continuity of the map
F 7→ f is obvious. This proves the first step in the induction procedure.
4.0.2 Reducing B
There are two cases where we can reduce the manifold B.
1. Suppose that there exists a smooth function ϕ : B → R+ which is not constant on the
subset BM := pB(M), where pB : B × g∗ → B is the canonical projection. Let β0 ∈ B
such that ϕ(β0) ∈]a, b[ for some b > a > 0 and let B0 := {β ∈ B; a
2 < ϕ(β) < 2b} and
M0 := {(β, l) ∈ M; β ∈ B0}.
Suppose that the theorem holds for the pair (B0, M0). Let us show that the result
remains true for the pair (B, M). Let M0 be an open relatively compact subset as
in the theorem for (B0, M0). We let M := {(β, l) ∈ M; a < ϕ(β) < b} ∩ M0. We will
show that M works for (B, M). Note that since M0 is open in M0, we have that M
is open in M.
Let F be a kernel function defined on Rr × M × G × G such that its support in (β, l)
is contained in G · M ⊂ M0. By assumption, there exists f ∈ S(Rr × B0 × G) such
that π(β,l)(f (·, β, ·)) admits F (·, (β, l), ·, ·) as an operator kernel if (β, l) ∈ M0.
In
10
c (R) with 0 ≤ ϑ ≤ 1, ϑ ≡ 1 on [a, b] and ϑ ≡ 0 on [0, a
particular, π(β,l)(f (·, β, ·)) = 0 if (β, l) ∈ M0 \ G · M0. As B0 is a sub-manifold of B,
we may extend f to a function in S(Rr × B × G) which we denote also by f . Choose
ϑ ∈ C∞
2 ] ∪ [2b, +∞[. We define
φ ∈ C∞(M) by φ(β, l) := ϑ(ϕ(β)). Then φ ≡ 1 on G · M0 and φ ≡ 0 on M \ G · M0.
By taking g := φ · f , we have that π(β,l)(g(·, β, ·)) = φ(β, l) · π(β,l)(f (·, β, ·)).
If (β, l) ∈ M ⊂ M0, then π(β,l)(g(·, β, ·)) = π(β,l)(f (·, β, ·)) and it admits F (·, (β, l), ·, ·)
If (β, l) ∈ M0 \ G · M0, then π(β,l)(f (·, β, ·)) = 0 and
as an operator kernel.
π(β,l)(g(·, β, ·)) = 0. If (β, l) ∈ M \M0, then ϕ(β) ∈ [0, a
2 ]∪[2b, +∞[, hence φ(β, l) = 0
and so π(β,l)(g(·, β, ·)) = 0. Hence the result is true for the function g.
2. If there exists a smooth sub-manifold B0 of B such that pB(M) ⊂ B0, then we can
apply our theorem to the pair (B0, M). Since every smooth function f0 on B0 × G
can be extended to a smooth function f on B × G, the retract theorem also holds for
(B, M).
Remark 4.0.3. Let B and M be given as in the statement of the theorem. Let
pB : M → B;
pB(β, l) = β,
If we denote by M max the subset of M consisting of all
be the canonical projection.
(β, l) ∈ M for which the rank of dpB(β, l) is maximal, then M max is open in M and the
subset pB(M max) of B is a smooth sub-manifold of B.
If pB(M max) contains at least 2
elements, by the reasoning in Subsection 4.0.2, using a non-constant smooth function ϕ0
on pB(M max), which can be extended to a smooth function ϕ of B, we reduce B to Bmax and
we can always assume in this way that pB(M) is a smooth sub-manifold of B. If pB(M max)
is a singleton {β0}, then M = M max and pB(M) is obviously a smooth sub-manifold of B.
4.1 Reducing to smoothly varying subspaces depending on B
Let M ⊂ B × g∗ be a smooth G-invariant sub-manifold of B × g∗. Let us fix the largest
index
IM = I =(cid:0)(j1, k1), · · · , (jr, kr)(cid:1) = (j1, k1) × I1,
where I1 = ((j2, k2), · · · , (jr, kr)) is the index set of (β, lg1(β,ℓ)), such that the open subset
MI := (B × g∗)I ∩ M of M is nonempty. Let pB : M → B; (β, l) 7→ β, be the projection
onto the first variable and set
BM := pB(M),
which is a smooth sub-manifold of B by Remark 4.0.3.
Let c1 := gj1+1 = span{Zj1+1, . . . , Zn} ⊂ g and let
n1
β := [g, c1]β + [Zj1, gk1+1]β ⊂ c1 for β ∈ B.
(4.5)
11
Then, by the definition of the indices (j1, k1), we have
n1
β ⊂ ker(l) ∩ c1 ⊂ aβ(l) if (β, l) ∈ (B × g∗)≤I.
(4.6)
It is easy to see that n1
β is an ideal in g. Let
Zβ := [Zk1, Zj1]β for β ∈ B.
We fix a scalar product h·, ·i on g such that {Z1, . . . , Zn} is an orthonormal basis and
r=j1+1 arZr of
1) with respect to the given
we identify c∗
c1. Denote by k · k2 the Euclidean norm on c1 (and hence on c∗
scalar product. We also identify
1 with c1 by identifying Pn
1 with the element Pn
r=j1+1 arZ ∗
r ∈ c∗
(n1
β)⊥ := {q ∈ c∗
1; hq, n1
βi = {0}}
with a subspace of c1. For all β ∈ B, we write c1 = n1
orthogonal projection of c1 onto (n1
by
β)⊥. For each β ∈ B, a generating subset of n1
β ⊕ (n1
β)⊥ and define pβ to be the
β is given
V (β) = {v1(β), · · · , vs(β)}
:= {[Za, Za′]β; a = 1, · · · , n, a′ = j1 + 1, · · · , n} ∪ {[Zb, Zj1]β; b = k1 + 1, · · · , n}.
Let
aj,j′(β) := hvj(β), vj′(β)i for 1 ≤ j, j′ ≤ s.
Fix 0 ≤ k ≤ s, let Ik = {J ⊂ {1, · · · , s}; J = k} and for β ∈ B, let
hk(β) := XJ∈Ik
det(cid:0)(cid:0)aj,j′(β)(cid:1)j,j′∈J(cid:1)2
.
It is easy to check that
hk(β) 6= 0 ⇔ v1(β), . . . , vs(β) have at least rank k,
hk(β) = 0 ⇔ v1(β), . . . , vs(β) have rank r < k.
Let n1 ∈ N and put f0 := hn1+1 and f1 := hn1. Let
B≤n1 = {β ∈ B; f0(β) = 0},
B≥n1 = {β ∈ B; f1(β) 6= 0},
Bn1 = {β ∈ B; f0(β) = 0 and f1(β) 6= 0}.
One sees that B≥n1 is open in B, and hence is a sub-manifold of B. Again, according to
the reduction argument in Subsection 4.0.2 we can assume that f1(β) 6= 0 for all β ∈ B.
On the other hand, let n1 := maxβ∈BM dim(n1
β), then we have
B≤n1
Bn1
B≥n1
:= {β ∈ B; dim(n1
:= {β ∈ B; dim(n1
:= {β ∈ B; dim(n1
β) ≤ n1},
β) = n1},
β) ≥ n1}.
12
Note that if we want n1
β to be of fixed dimension and to have n1
β)⊥ and pβ to
vary smoothly with respect to β, we must restrict to Bn1. But in general Bn1 is not a
sub-manifold of B. Therefore we must find a smooth sub-manifold inside Bn1 containing
an open subset of the smooth manifold BM = pB(M). We have to distinguish the following
two cases:
β, (n1
Case 1: If the differential df0 is not identically zero on BM , we may define
Bmax
M := {β ∈ BM ; df0(β) 6= 0} and Bmax := {β ∈ B; df0(β) 6= 0}.
By assumption, Bmax is a nonempty open subset of B.
Case 2: Assume that df0 is identically zero on BM .
If dim(BM ) < dim(B), we may build a function γ ∈ C∞(B) such that γ ≡ 0 on BM and
dγ is not identically zero on BM . We put ef0 = f0 + γ. Then ef0 ≡ 0 on BM and def0 is not
identically zero on BM . We then define
Bmax
M := {β ∈ BM ; def0(β) 6= 0}, Bmax := {β ∈ B; ef0(β) = 0, def0(β) 6= 0}.
By the construction of ef0, we have again that Bmax is an open subset of B and Bmax
M ⊂ Bmax.
If dim(BM ) = dim(B), then BM is open in B and we take a smooth function f 6= 0 in
B supported on BM . Let
Bmax := {β ∈ B; ef (β) 6= 0} and Bmax
M := {β ∈ BM ; ef (β) 6= 0}.
β vary smoothly on the smooth sub-manifold Bmax of B, since
In the two cases, the ideals n1
dim(n1
β) = n1 on Bmax
M . The projection pβ also varies smoothly on Bmax.
Remark 4.1.1. According to Remark 4.0.3, we can now assume that
B = Bmax.
2 is now smooth on B, we can take β0 ∈ B
Furthermore, since the function β 7→ kpβ(Zβ)k2
and 0 < δ < R < ∞ such that δ < kpβ0(Zβ0)k2 < R and by using the reduction argument,
we can then assume that the number kpβ(Zβ)k2 is contained in the interval [δ, R] for any
β ∈ B.
4.1.2 On the manifold M
Let us focus on the manifold M again. Let (β0, l0) ∈ M be fixed, but arbitrary. There
exist 0 < δ < R < ∞ such that
0 < δ < min{hl0, Zβ0i, kpβ0(Zβ0)k2} < max{hl0, Zβ0i, kpβ0(Zβ0)k2} < R.
This is due to the fact that M ⊂ (B × g∗)I. According to Remark 4.1.1 we can now assume
that
0 < δ < min{kpβ(Zβ)k2} < max{kpβ(Zβ)k2} < R
13
for all β ∈ B. We define
M δ,R = M red
:= {(β, l) ∈ M; 0 < δ < min{hl, Zβi} < max{hl, Zβi} < R}.
Obviously, M red is open in M and thus is a smooth sub-manifold of M. On the other
hand, we define
(B × g∗)≤I,R,δ = n(β, l) ∈ (B × g∗)≤I;
1
2
δ < min{hl, Zβi} < max{hl, Zβi} <
Ro.
3
2
4.1.3 Reducing M
Now we show that if the retract theorem holds for (B, M red), then it remains true for
(B, M).
Assume that the result is true for (B, M red). Let M be the open subset in M red given
by the assumption. We will show that one may take the same manifold M for (B, M) such
that the theorem remains true for (B, M). As M red is open in M, the set M also has a
nonempty interior in M. Moreover, pB(M) ⊂ pB(M red) ⊂ Bn1 ⊂ B≥n1. Let ∅ 6= C ⊂ M
be compact and let F be a kernel function defined on Rr × M × G × G whose support in
(β, l) is contained in G · C. The restriction of F to Rr × M red × G × G is a kernel function
for (B, M red).
By assumption, there exists f ∈ S(Rr × B × G) such that π(β,l)(f (·, β, ·)) admits
In particular, π(β,l)(f (·, β, ·)) = 0
F (·, (β, l), ·, ·) as an operator kernel if (β, l) ∈ M red.
if (β, l) ∈ M red \ C. As ∅ 6= C ⊂ M is compact, there exist δ1, R1 ∈ R+ such that
0 < δ < δ1 ≤ min{hl, Zβi}
≤ max{hl, Zβi}
≤ R1 < R
for all (β, l) ∈ C, as C ⊂ M ⊂ M ⊂ (B × g∗)I. Let u ∈ C∞
c (R) be odd such that
u ≡ 1 on [δ1, R1] and u ≡ 0 on [0, δ] ∪ [R, +∞[. There exists χ ∈ S(R) such that
For (β, l) ∈ (B × g∗)J with J ≤ I, we have Zβ ∈ aβ(l) ⊂ gβ(l) and π(β,l)RZβ (ψ(β, ·)) =
bχ = u. Let us define a function ψ on Sβ{β} × exp(RZβ) by ψ(β, exp(sZβ)) := χ(s).
bχ(hl, Zβi)IH(β,l) = u(hl, Zβi)IH(β,l).
We define a function g on Rr × B × G by
g(·, β, ·) := f (·, β, ·) ∗ ψ(β, ·).
This implies that
π(β,l)(g(·, β, ·)) = u(hl, Zβi)π(β,l)(f (·, β, ·)).
If (β, l) ∈ C, then u(hl, Zβi) = 1 and π(β,l)(g(·, β, ·)) = π(β,l)(f (·, β, ·)) admits F (·, (β, l), ·, ·)
as an operator kernel. If (β, l) ∈ M red \C, then π(β,l)(f (·, β, ·)) = 0, hence π(β,l)(g(·, β, ·)) =
14
0 and F (·, (β, l), ·, ·) = 0. If (β, l) ∈ M \ M red, then hl, Zβi /∈ [δ, R], i.e. u(hl, Zβi) = 0,
which implies that π(β,l)(g(·, β, ·)) = 0. Hence, the mapping F 7→ g satisfies the property
of the retract for (B, M).
4.2 Construction of a new variable group
4.2.1 The mapping α(β, l)
β)⊥ and
For (β, l) ∈ (B × g∗)≤I, we have seen in (4.6) that n1
hl, Zβi = hl, pβ(Zβ)i = hq, pβ(Zβ)i. For (β, l) ∈ (B×g∗)≤I,R,δ, we have that hl, [Zk1, Zj1]βi =
hl, Zβi > δ
2 > 0 implies hl, [Zk1, Zj1]βi 6= 0 and j1(β, l) = j1, k1(β, l) = k1.
β ⊂ ker(l). Let q := lc1 ∈ (n1
Take an odd function ϕ ∈ C ∞(R) with the properties that ϕ(s) = 0 for 0 ≤ s <
δ/4 and s > 2R, 1 > ϕ(s) > 0 for s ∈]δ/4, δ/2[∪]3R/2, 2R[ and ϕ(s) = 1 for 3R/2 ≥ s ≥
δ/2. For every (β, q) ∈ B × c∗
1, we construct the vector α(β, q) ∈ c1 ≃ (c1)∗ by
α(β, q)
:= ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)pβ(q)
+(cid:0)1 − ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)(cid:1)pβ(Zβ).
(4.7)
Then by the construction, α(β, q) ∈ (n1
other hand, for (β, l) ∈ (B × g∗)≤I,R,δ and q = lc1, we have that
β)⊥ ⊂ c∗
1 ≡ c1 for every (β, q) ∈ B × c∗
1. On the
α(β, q) = ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)pβ(q)
+(1 − ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i))pβ(Zβ)
= ±pβ(q) + (1 − 1)pβ(Zβ)
= ±q.
(4.8)
This is due to the fact that pβ(q) = q as n1
β ⊂ ker(q) for q = lc1, if (β, l) ∈ (B × g∗)≤I.
We will show that
hα(β, q), Zβi = ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)hpβ(q), Zβi
+(1 − ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i))kpβ(Zβ)k2
2 > 0
on B × c∗
and ϕ ≥ 0 on R+, we have
1. In fact, let us first notice that hpβ(q), Zβi = hq, pβ(Zβ)i. As ϕ is an odd function
A := ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)hpβ(q), Zβi ≥ 0.
Since 0 ≤ ϕ ≤ 1 on R+,
B :=(cid:0)1 − ϕ(kpβ(Zβ)k2)ϕ(hq, pβ(Zβ)i)(cid:1)kpβ(Zβ)k2
2 ≥ 0.
If none of the ϕ(·)'s is equal to zero and if hpβ(q), Zβi 6= 0, then A > 0. If hpβ(q), Zβi = 0,
then ϕ(hpβ(q), Zβi) = 0 and so B > 0, as now by Remark 4.1.1 kpβ(Zβ)k2 > 0. If one of
the ϕ(·)'s is equal to zero, then again B > 0.
15
For (β, q) ∈ B × c∗
1, let
We have that
and
δ(β, q) := ad∗
β(Zj1)α(β, q) ∈ g∗.
hδ(β, q), Zk1i = hα(β, q)), Zβi > 0
hδ(β, q), [g, g]βi = hα(β, q), [[g, g]β, Zj1]βi ⊂ hα(β, q), [g, [g, Zj1]β]βi
βi = {0},
⊂ hα(β, q), [g, c1]βi ⊂ hα(β, q), n1
by the definition of α(β, q) in (4.7). This means that δ(β, q) is an algebra homomorphism
of gβ = (g, [·, ·]β) which does not vanish at the vector Zk1. Hence the subspace
g1(β, q) := ker(δ(β, q))
is an ideal of gβ of co-dimension 1 and
g = RZk1 ⊕ g1(β, q).
(4.9)
Furthermore g1(β, q) contains c1 for any (β, q) ∈ B × c∗
1. In fact,
hδ(β, q), c1i = hα(β, q), [c1, Zj1]βi = 0
as α(β, q) ∈ (n1
β)⊥ and [c1, Zj1]β ⊂ n1
β.
4.2.2 The new variable group (B1, G1)
In order to construct a new variation in the induction procedure, we put
For (β, y, q) ∈ B1, we define a Jordan-Holder basis
B1 := B × R × c∗
1.
Z 1(β, y, q) = { Z 1
1 (β, q), · · · , Z 1
n−1(β, q)}
of g1(β, y, q) = ker(δ(β, q)) by
αβ,y,q
k = αβ,q
k
:=
hα(β, q), [Zk, Zj1]βi
hα(β, q), Zβi
and
eZ 1(β, y, q) = eZ 1(β, q)
:= {Z1 − αβ,q
= { Z 1
1 (β, q), · · · , Z 1
n−1(β, q)}.
1 Zk1, , · · · , Zk1−1 − αβ,q
k1−1Zk1, Zk1+1, · · · , Zn}
16
In particular, for (β, l) ∈ (B × g∗)≤I,R,δ we have by (2.1) that
g1(β, y, lc1) = g1(β, l) = g1(β, lc1).
In fact, in this case j1(β, l) = j1, k1(β, l) = k1 and
g1(β, y, lc1) = {U ∈ g hδ(β, lc1), Ui = 0} = {U ∈ g hl, [U, Zj1]βi = 0} = g1(β, l),
as α(β, lc1) = ε · lc1 with ε = ±1 if (β, l) ∈ (B × g∗)≤I,R,δ.
For each k, we also have that
αβ,q
k =
hl, [Zk, Zj1]βi
hl, Zβi
.
(4.10)
This new basis eZ 1(β, y, q) coincides then, up to normalisation, with the basis obtained in
Section 2.4, both procedures and bases generate the same indices. Furthermore by (4.10),
for (β, l) ∈ (B × g∗)≤I,R,δ, we have
Z 1(β, y, l1) = Z 1(β, l),
(4.11)
where Z 1(β, l) is defined in Section 2.4 and l1 = lc1.
For any (β, y, q) ∈ B1, let us write
[ Z 1
u(β, q), Z 1
v (β, q)] =
n−1Xw=1
γ(β, q)u,v
w
Z 1
w(β, q)
for u < v in {1, · · · , n − 1}.
We obtain in this way a new variable Lie algebra (B1, g1), where
g1 = Rn−1, B1 = B × R × c∗
1
and the canonical basis Z 1 = {Z 1
1 , · · · , Z 1
n−1} of g1 satisfies, by definition,
[Z 1
u, Z 1
v ](β,q) =
n−1Xw=1
γ(β, q)u,v
w Z 1
w,
for u < v in {1, · · · , n − 1}.
This means that the new variable Lie algebra (B1, g1) with g1 ≡ g1(β, q) is defined such
that (g1, [·, ·](β,q)) ≡ (g1(β, q), [·, ·](β,q)).
Given (β, l) ∈ B × g∗, let us define l1 ∈ g∗
i ) := l( Z 1
i ) = l(Zi+1) if i ≥ k1. We also define a map
1 by l1(Z 1
1}. One has l1(Z 1
i (β, q)) for all i ∈ {1, . . . , n −
ι1 : B × g∗ → B1 × g∗
1
(β, l)
7→ ((β, hl, Zk1i, α(β, lc1)), l1),
(4.12)
where l1 ≡ lg1(β,lc1 ). We see that ι1 is obviously smooth, injective and even a diffeomor-
phism onto its image.
17
Using (4.9) we can identify every l ∈ g∗ with the pair (v, l1) where v := hℓ, Zk1i and
l1 := lg1 ≡ lg1(β,lc1 ). We can then transfer the natural action of G on B × g∗ to B1 × g∗
using the mapping ι1. This gives us
1
g · ((β, v, q), l1) = ((β, v + hAd∗
β(g)l1).
β(g)l1, Zk1i, Ad∗
β(g)q), Ad∗
Then we have automatically the relation
ι1(g · (β, l)) = g · (ι1(β, l))
for any g ∈ G and (β, l) ∈ B × g∗.
Consider now the smooth manifold
(B × g∗)0
≤I,R,δ := {(β, l) ∈ (B × g∗)≤I,R,δ; hl, Zk1i = hl, Zj1i = 0}.
Then obviously the smooth manifold (B × g∗)≤I,R,δ is diffeomorphic with the manifold
R2 × (B × g∗)0
≤I,R,δ. The mapping
Φ : R2 × (B × g∗)0
≤I,R,δ → (B × g∗)≤I,R,δ
given by
Φ(s, t, (β, l)) :=(cid:16)β, Ad∗(exp(
s
hl, Zβi
Zj1)exp(
t
hl, Zβi
Zk1))l(cid:17)
is such a diffeomorphism. Hence every smooth G-invariant sub-manifold M of (B×g∗)≤I,R,δ
can be decomposed into a direct product of R2 with the smooth manifold M 0, where
M 0 := {(β, l) ∈ M; hl, Zk1i = hl, Zj1i = 0}.
For (β, l) ∈ B × g∗, one has l1(Z 1
i ) = l(Zi+1) if i ≥ k1. We
remark that for (β, l) and (β, l′) in M with ι1(β, l) = ι1(β, l′) we have that l and l′ have
the same restriction to g1(β, l) = g1(β, l′), so they are on the same co-adjoint orbit and
l′ = Ad∗(y)l for some y ∈ P (β, l) and hence
i ) = l(Zi) if i < k1 and l1(Z 1
F (β, l) = F (β, l′)
by the conditions on the operator fields defined over M given in Definition 2.7.1.
We denote the new variable Lie group by G1 = (B1, G1) where G1 = (expβ1g1)β1∈B1
and expβ1g1 is the connected, simply connected, nilpotent Lie group associated to the Lie
algebra (g1, [·, ·]β1).
4.3
Induction step
To simplify notations, from now on we will omit the subscript β in the notations of the mul-
tiplication and the exponential map, unless the subscript is crucial for the understanding.
There are two preliminary steps to check.
18
4.3.1
Induction hypothesis
In this subsection, we will prove the result for (B, M) using induction. Let M1 = ι1(M) be
as constructed in (4.12). Let us recall that l1 = l1(β, lc1) ≡ lg1(β,lc1 ) for (β, l) ∈ (B × g∗)≤I.
The Vergne polarisation p(β, l) for l in (g, [·, ·]β), obtained by the procedure of Ludwig-
Zahir (see [11], [9]), is also the Vergne polarisation for l1 in (g1, [·, ·]β). Let us denote by
P (β, l) = expβp(β, l)) the corresponding subgroup. The associated induced representations
∼=
will be denoted by π(β,l) := indG
indG
G1 π((β,lc1 ),l1), as usual.
Since M is G-invariant, the manifold M1 = ι1(M) is also G-invariant in B1 × g∗
P (β,l)χl, respectively, π((β,lc1 ),l1) := indG1
P (β,l)χl1. Then π(β,l)
1. Hence
we can write M1 as a direct product manifold R2 × M 0
1 , where
M 0
1 := {((β, 0, q), l1); hl1, Zj1i = 0}
is G1 invariant. Note that M 0
we have that ∞ > R > hl1, Zβi > δ > 0. The induction hypothesis in B1 × g∗
M 0
1 of M 0
the required properties of the theorem.
1)I1 and for every ((β, v, lc1), l1) ∈ M1
1 and
1 with
1)I1 gives us an open relatively compact non-empty subset M0
1 is contained in (B1 × g∗
1 ⊂ (B1 × g∗
We choose now a relatively compact open subset M1 of M1 such that M1 ⊂ M1 and
M1 is contained in G · M0
1. Let
M := ι−1
1 (M1)
and M0 := ι−1
1 (M0
1).
Then M is non-empty open with its closure M contained in M and M is contained
in G · M0. We take a kernel function F ∈ Dc
M such that its support is contained in
Rr × G · M × G × G.
Given the kernel function F , we will now define a kernel function for the variable
group (B1, G1). For simplicity, we will omit the subscripts β or (β, v, lc1) in the nota-
tions of the multiplication and the exponential map, and we will identify g1, g′
1 ∈ G1 =
≡ g1 with the corresponding elements in G1. In the following computa-
tions, the parameters β and (β, v, lc1) will indicate how to multiply group elements or
how to decompose smoothly the group elements, even if it is not marked explicitly. For
ι1(β, l) = ((β, hl, Zk1i, lc1), l1) ∈ M1, we put
(cid:0)(G1)(β1)(cid:1)β1∈B1
F1(α, u, t, ((β, hl, Zk1i, lc1), l1), g1, g′
1) :=
(2π)n−j1+1 · c(β, l) · F (α, (β, l), exp((u + t)X) · g1, exp(tX) · g′
1),
for α ∈ Rr, u, t ∈ R and g1, g′
function F1 has its support S1 := ι1(S) contained in G · M1, and belongs to Dc
operator field F1 is smooth on M1, since the mappings F and c are both smooth.
1 ∈ G1, where c(β, l) := hl, [Zk1, Zj1]βi 6= 0 and X = Zk1. This
M1. The
By the induction hypothesis, there exists h ∈ S(Rr+2, B1, G1) such that π((β1,lc1 ),l1)(cid:0)h(α, u, t, β1, ·)(cid:1)
admits F1(α, u, t, (β1, l1), ·, ·) as an operator kernel for all (β1, l1) ∈ M 0
of the retract function f will now be done in several steps.
1 . The construction
19
4.3.2 Definition of the retract function on the original group
For (β, v, q) ∈ B1, let us first define eh by
eh(α, u, t, (β, v, q), g1)
:= ZRZc1
h(α, u, t, (β, v, q), g1 · exp(yY ) · exp(Z))e−iq(Z)dZdy,
where Y = Zj1 and Z = Zβ = [X, Y ]β with X = Zk1. The integral converges, as h is
Schwartz in g1 (for fixed β1), and it is of rapidly decreased in q ∈ (c1)∗, because it is a
Fourier transform in Z. For all (β, v, q) ∈ B × R × c∗
1, we then define
ef (α, (β, v, q), exp(uX) · g1 · exp(yY ) · exp(Z)) = eiq(Z)ZReh(α, u, t, (β, v, q), g−t
1 )e−ityq([X,Y ]β)dt
with g = exp(uX) · g1 and g−t
1
Rr × (B × c∗
:= exp(−tX) · g1 · exp(tX). The function ef is smooth on
1 by construction, we may define f by
1) × G. As ef is of rapid decrease in q ∈ c∗
f (α, β, g) :=Z(c1)∗ ef (α, (β, 0, q), g)dq, α ∈ Rr, β ∈ B, g ∈ G.
One can see that f ∈ S(Rr, B, G) (in the sense of Section 2.6).
4.3.3 The retract property
Let us now compute π(β,l)(cid:0)f (α, β, ·)(cid:1) for (β, l) ∈ M. Since the manifold M is contained
in (B × g∗)≤I,R,δ we have that c1 ⊂ aβ(l) ⊂ gβ(l). If we identify exp(c1) and c1, as well as
exp(Z) and Z, for any function ξ(β) ∈ H(β,l) (the representation space of π(β,l)) and any
g ∈ G, we have that
(cid:16)π(β,l)(cid:0)f (α, β, ·)(cid:1)ξ(β)(cid:17)(g) = ZGβ /c1Zc1
= ZGβ /c1Zc1
= ZGβ /c1Zc1Z(c1)∗
= ZGβ /c1Zc1Z(c1)∗
2π(cid:17)n−j1ZGβ/c1
= (cid:16) 1
f (α, β, g · Z)(cid:16)π(β,l)(g)π(β,l)(Z)ξ(β)(cid:17)(g)dZd g
f (α, β, g · Z)e−il(Z)(cid:16)π(β,l)(g)ξ(β)(cid:17)(g)dZd g
f (α, (β, 0, q), g · Z)e−il(Z)(cid:16)π(β,l)(g)ξ(β)(cid:17)(g)dqdZd g
f (α, (β, 0, q), g)eiq(Z)e−il(Z)(cid:16)π(β,l)(g)ξ(β)(cid:17)(g)dqdZd g
f (α, (β, 0, lc1), g)(cid:16)π(β,l)(g)ξ(β)(cid:17)(g)d g.
We use the following smooth decomposition: X = Zk1, g1 = g1(β, l) which gives us
g = exp(sX) · g1 with s = s(g, β, lc1), g1 = g1(g, β, lc1).
20
We then obtain (using the fact that c1 ⊂ (aβ(l)) for all our (β, l)'s) that:
· exp((s − u)X) · g1)d g1du
· g1)d g1dr
f (α, (β, 0, lc1), exp(uX) · g1) ·
with gr
1 = exp(rX) · g1 · exp(−rX)
f (α, (β, 0, lc1), exp((s − r)X) · g1) ·
f (α, (β, 0, lc1), exp((s − r)X) · gr
1) · ξ(β, 0, lc1)(r)(g−1
1
f (α, (β, 0, lc1), exp(uX) · g1) · ξ(β)(g−1
1
with ξ(β, 0, lc1)(v)(g1) := ξ(β)(exp(vX) · g1)
f (α, (β, 0, lc1), g)(cid:16)π(β,l)(g)ξ(β)(cid:17)(exp(sX) · g1)d g
(2π)n−j1 ·(cid:16)π(β,l)(cid:0)f (α, β, ·)(cid:1)ξ(β)(cid:17)(exp(sX) · g1)
= ZGβ /c1
= ZRZG1/c1
= ZRZG1/c1
ξ(β, 0, lz)(s − u)(cid:16)(cid:0)exp(−(s − u)X) · g1 · exp((s − u)X)(cid:1)−1 · g1(cid:17)d g1du
= ZRZG1/c1
ξ(β, 0, lc1)(r)(cid:16)(cid:0)exp(−rX) · g1 · exp(rX)(cid:1)−1 · g1(cid:17)d g1dr with s − u = r
= ZRZG1/c1
= ZRZG1/c1
= ZRZG1/c1ZR
= ZRZG1/exp(RY )·c1ZRZR
(cid:16)π((β,0,lc1 ),l1)(w1) ξ(β, 0, lc1)(r)(cid:17)(g1)dtdyd w1dr (as l(Y ) = 0, for l1 ≡ lg1(β,0,lc1 ))
= ZRZG1/exp(RY )·c1ZRZR
(cid:16)π((β,0,lc1 ),l1)(w1) ξ(β, 0, lc1)(r)(cid:17)(g1)dtdyd w1dr
(as (cid:0)exp(yY )(cid:1)r−t
c(β, l)ZRZG1/exp(RY )·c1ZRZR
(cid:16)π((β,0,lc1 ),l1)(w1) ξ(β, 0, lc1)(r)(cid:17)(g1)dtdyd w1dr
= exp(yY + y(r − t)[Zk1, Zj1]β), by covariance of h, with
1)(cid:16)π((β,0,lc1 ),l1)(g1) ξ(β, 0, lc1)(r)(cid:17)(g1)d g1dr
)(cid:16)π((β,0,lc1 ),l1)(g1) ξ(β, 0, lc1)(r)(cid:17)(g1)dtd g1dr
h(α, s − r, t, (β, 0, lc1), wr−t
· (exp(yY )r−t)
h(α, s − r, t, (β, 0, lc1), wr−t
1
· exp(yY )) · ei(r−t)c(β,l)y
f (α, (β, 0, lc1), exp((s − r)X) · gr
c(β, l) = hl, [Zk1, Zj1]βi 6= 0 as before)
1
=
(with l1 ≡ lg1)
h(α, s − r, t, (β, 0, lc1), gr−t
1
h(α, s − r, t, (β, 0, lc1), wr−t
1
· exp(c(β, l)−1yY ))eir ye−ity
(with y = c(β, l)y)
1
21
=
=
=
=
1
h(α, s − r, t, (β, 0, lc1), wr−t
·exp(c(β, l)−1 yY ) · exp(y′Y ) · Z) · e−il(Z)eir ye−ity
c(β, l)ZRZG1/exp(RY )·c1ZRZRZRZc1
(cid:16)π((β,0,lc1 ),l1)(w1) ξ(β, 0, lc1)(r)(cid:17)(g1)dZdy′dtdyd w1dr
c(β, l)ZRZG1/exp(RY )·c1ZRZRZRZc1
·e−il(Z)eir ye−ity(cid:16)π(β1,l1)(w1) ξ(β, 0, lc1)(r)(cid:17)(g1)dZdy′′dtdyd w1dr (for y′′ = y′ + c(β, l)−1y)
h(α, s − r, t, (β, 0, lc1), wr−t
· exp(y′′Y ) · Z)
1
1
1
h(α, s − t, t, (β, 0, lc1), w1 · exp(y′′Y ) · Z)
1
2π
·
1
c(β, l)ZRZG1/exp(RY )·c1ZRZc1
·e−il(Z)(cid:16)π((β,0,lc1 ),l1)(w1) ξ(β, 0, lc1)(t)(cid:17)(g1)dZdy′′dtd w1
1
c(β, l)ZR
1
2π
·
π((β,0,lc1 ),l1)(cid:0)h(α, s − t, t, (β, 0, lc1), ·) ξ(β, 0, lc1)(t)(cid:1)(g1)dt
as e−il(Z)π((β,0,lc1 ),l1)(w1) = π((β,0,lc1 ),l1)(w1 · exp(y′′Y ) · Z).
Let us finish the computation for (β, l) ∈ M. It suffices to take (β, l) ∈ M 0. Then
((β, 0, lc1), l1) ∈ M 0
1 and by the induction hypothesis,
·
1
(cid:16)π(β,l)(cid:0)f (α, β, ·)(cid:1)ξ(β)(cid:17)(g)
= (cid:16)π(β,l)(cid:0)f (α, β, ·)(cid:1)ξ(β)(cid:17)(exp(sX) · g1)
c(β, l)ZRZG1/P (β,l)
= (cid:16) 1
2π(cid:17)n−j1+1
= ZRZG1/P (β,l)
= ZG/P (β,l)
ξ(β, 0, lc1)(t)(g1)d g1dt
F (α, (β, l), g, g)ξ(β)(g)dg.
F1(α, s − t, t, ((β, 0, lc1), l1), g1, g1)
F (α, (β, l), exp(sX) · g1, exp(tX) · g1)ξ(β)(exp(tX) · g1)dg1dt
Hence for every (β, l) ∈ M, we have the required result.
The algorithm used to build the retract function f respects the semi-norms defining
the topology of our function spaces. So the retract map F 7→ f is continuous.
5 G-prime ideals in L1(G)
In this section, we will study the structure of the A-prime ideals in L1(G) by using the
retract theorem.
22
5.1 A retract defined on closed orbits
Let G be a Lie group of automorphisms of a connected, simply connected, nilpotent Lie
group G = exp(g) containing the inner automorphisms of G. For instance take any simply
connected Lie group G and let G be the nilradical of G.
Let l0 ∈ g∗ be fixed, we consider the orbit Ω = Ωl0 := G · l0 in g∗, let O = Ol0 be the
G-orbit of l0. We assume that Ω is locally closed in g∗. In particular we can write
Ω = Ω ∩ U,
where Ω denotes the closure of Ω in g∗ and U is G-invariant open subset of g∗. It is then
a G-invariant smooth sub-manifold of g∗ diffeomorphic to the manifold G/Gl0, where Gl0
denotes the stabiliser Gl0 := {α ∈ G; α · l0 = l0}. The G-orbit G · (G · l0) in the orbit
space g∗/G is then locally closed and homeomorphic to the quotient G/GO, where GO is
the stabiliser of the set O in G. In fact, we have that GO = G · Gl0.
For a Jordan-Holder basis Z = {Z1, · · · , Zn} of g and g ∈ G, let
g · Z := {Ad(g)Z1, · · · , Ad(g)Zn},
which is again a Jordan-Holder basis of g. For every index set I, we have the following
relation (see [9]):
Ad∗(g)g∗
I,g·Z = g∗
I,Z,
g ∈ G.
(5.13)
For an index set I and a Jordan-Holder basis Z of g, recall that
sI :=Xi∈I
RZ ∗
i , ΣI,Z := sI ∩ g∗
I,Z,
and the mapping EI : Rd × ΣI,Z → g∗
I,Z is given by
EI(s1, t1, · · · , sr, tr; l) := Ad∗(exp(s1Zj1)exp(t1Zk1) · · · exp(srZjr)exp(trZkr ))l.
We have that EI is a bijection and EI(Rd × {l}) is the G-orbit of l. Let
Υ : g∗
I,Z → ΣI,Z; Υ(l) := Ad∗(G)l ∩ ΣI,Z = pΣI,Z (E−1
I (l)),
where pΣI,Z is the projection of Rd × ΣI,Z onto ΣI,Z.
For the orbit Ω, we need to construct a finite partition of unity (ψi)i∈Γ consisting of
smooth G-invariant functions ψi : Ω → R+ such that for every i ∈ Γ the support of each
function ψi is contained in an open subset of g∗
In order to do
that let ϕ : R → R+ be a smooth function with compact support which vanishes in a
neighbourhood of 0. We define a function ψ : g∗
I,gi·Z for some gi ∈ G.
≤I → R+ by
ψ(l) := ϕ(PI(Υ(l)) if l ∈ g∗
I
and ψ(l) := 0 if l ∈ g∗
<I,
23
where PI is a smooth function on B × g∗ defined in Section 2.4. We see that ψ is smooth
(since ϕ vanishes in a neighbourhood of 0) and is G-invariant by the construction. Let
UI,Z := {l ∈ Ω; ψ(l) 6= 0}.
Now assume that g∗
We have that Ω ∩ g∗
non-empty open subset of Ω contained in g∗
I = g∗
I,g·Z 6= ∅ and Ω ∩ g∗
I and
I,Z be the maximal layer with respect to Z such that Ω ∩ g∗
I,Z 6= ∅.
I ′,g·Z = ∅ for g ∈ G and I ′ > I. Moreover, UI,Z is a
Ω ⊂ [g∈G
Ad∗(g)UI,Z.
Let C be a compact subset of g∗ contained in Ω, then there exists a finite subset Γ ⊂ G
such that
C ⊂ [g∈Γ
Ad∗(g)UI,Z.
Hence there is a finite partition of unity (ψi)i∈Γ consisting of smooth G-invariant functions
ψi : Ω → R+ such that the support of each function ψi is contained in Ad∗(gi)UI,Z ⊂ g∗
for every gi ∈ Γ.
I,gi·Z
Suppose we have a smooth adapted operator field F on Ω supported on G · C, we can
write
F =Xi∈Γ
ψiF.
According to the retract theorem, for each i ∈ Γ there is a (retract) Schwartz function fi
on G such that
πl(fi) = opψiF (l)
for every l ∈ Ω. For f :=Pi∈Γ fi, we have that
πl(f ) = Xi∈Γ
= Xi∈Γ
πl(fi)
ψiopF (l)
= opF (l).
Hence for every smooth adapted kernel function supported on G·C, we have build a retract
function.
24
5.2 G-prime ideals
Let us first recall the definition of G-prime ideals.
Definition 5.2.1. A two-sided closed ideal I in L1(G) is called G-prime, if I is G-invariant
and if, for all G-invariant two-sided ideals I1 and I2 of L1(G), the following implication
holds
I1 ∗ I2 ⊂ I ⇒ I1 ⊂ I or I2 ⊂ I.
Denote by P rim∗(G) the collection of all the kernels of irreducible unitary representa-
tions of L1(G). For a closed subset C of P rim∗(G), let
ker(C) := \P ∈C
P.
For a subset I of L1(G), denote by h(I) the subset
h(I) := {P ∈ P rim∗(G); I ⊂ P }.
The set h(I) is then closed in P rim∗(G) with respect to the Fell topology.
We have the following result for G-prime ideals of L1(G) which can be viewed as an
application of the retract theorem.
Theorem 5.3. Let G be a simply connected, connected nilpotent Lie group and let G be a
Lie group of automorphisms of G containing the inner automorphisms, which acts smoothly
on the group G, such that every G-orbit in g∗ is locally closed. If I is a proper G-prime
ideal of L1(G), then there exists an G-orbit Ωl0 in g∗ such that
Moreover, all the kernels of G-orbits are G-prime ideals.
I = ker(Ωl0).
Proof. For any G-orbit Ω in g∗, the retract theorem tells us that the Schwartz functions
contained in ker(Ω) are dense in ker(Ω) (see [7, proof of Proposition 4.1] and [4]). From
the proof of [8, Theorem 1.2.12], it follows that the hull of a prime ideal I is the closure
of an G-orbit in P rim∗(G) ≃ bG. On the other hand, the density of S(G) ∩ ker(Ω) implies
that ker(Ω)N is contained in the minimal ideal J(Ω) with hull Ω for some N ∈ N. This
tells us that ker(Ω)N ⊂ J(Ω) ⊂ I, since the minimal ideal with hull Ω is contained in every
ideal with hull Ω. Since I is G-prime, we have that I = ker(Ω).
Obviously the ideal ker(Ω) is G-prime for any G-orbit Ω in g∗. To see this, let I1
and I2 be two G-invariant ideals of L1(G) such that I1 ∗ I2 ⊂ ker(Ω). This means that
I1 ∗ I2 ⊂ ker(Ω) ⊂ ker(πl) for some l ∈ Ω. We have then either I1 or I2 is contained
in ker(πl), since πl is irreducible. But if I1 is contained in ker(πl), it is also contained in
ker(πk·l) since I1 is G-invariant. Hence I1 ⊂ ker(Ω) and the proof is complete.
25
References
[1] L. Corwin and P. Greenleaf, Representations of nilpotent Lie groups and their
applications. Part 1: Basic theory and examples, Cambridge Studies in Ad-
vanced Mathematics, 18 (1990).
[2] R. Howe, On a connection between nilpotent groups and oscillatory integrals
associated to singularities, Pacific J. Math. 73 (1977), no. 2, 329-363.
[3] A. A. Kirillov, Unitary representations of nilpotent Lie groups, Uspekhi Mat.
Nauk. 17 (1962), 53-104.
[4] R. Lahiani and C. Molitor-Braun, Compact actions, retract theory and prime
ideals, Illinois J. Math. 55, no. 3 (2011), 1235-1266.
[5] H. Leptin, J. Ludwig, Unitary representation theory of exponential Lie groups,
De Gruyter Expositions in Mathematics 18, 1994.
[6] J. Ludwig, On Primary Ideals in the Group Algebra of a Nilpotent Lie Group,
Math. Ann. 262 (1983), 287-304.
[7] J. Ludwig and C. Molitor-Braun, Exponential actions, Orbits and their Ker-
nels, Bull. Austral. Math. Soc. 57 (1998), 497-513.
[8] J. Ludwig, C. Molitor-Braun, Repr´esentations irr´eductibles born´ees des
groupes de Lie exponentiels, Canad. J. Math. 53, no. 5 (2001), 944-978.
[9] J. Ludwig, C. Molitor-Braun and L. Scuto, On Fourier's inversion theorem
in the context of nilpotent Lie groups, Acta Sci. Math. (Szeged) 73 (2007),
547-591.
[10] J. Ludwig and D. Muller, Sub-Laplacians of Holomorphic Lp-type on Rank One
AN-Groups and Related Solvable Groups, J. Func. Anal. 170 (2000), 366-427.
[11] J. Ludwig and H. Zahir, On the nilpotent ∗-Fourier transform, Lett. Math.
Phys. 30 (1994), 23-34.
[12] D. Poguntke, Uber das Synthese-Problem fur nilpotente Liesche Gruppen,
Math. Ann. 269 (1984), 431-467.
Ying-Fen Lin, Pure Mathematics Research Centre, Queen's University Belfast, Belfast,
BT7 1NN, U.K. E-mail: [email protected]
Jean Ludwig, Universit´e de Lorraine, Institut Elie Cartan de Lorraine, UMR 7502, Metz,
F-57045, France. E-mail: [email protected]
Carine Molitor-Braun, Unit´e de Recherche en Math´ematiques, Universit´e du Luxembourg,
6, rue Coudenhove-Kalergi, L-1359 Luxembourg, Luxembourg. E-mail: [email protected]
26
|
1309.2393 | 2 | 1309 | 2013-11-22T18:20:02 | A Survey on Reproducing Kernel Krein Spaces | [
"math.FA"
] | This is a survey on reproducing kernel Krein spaces and their interplay with operator valued Hermitian kernels. Existence and uniqueness properties are carefully reviewed. The approach we follow in this survey uses a more abstract but very useful concept of linearization or Kolmogorov decomposition, as well as the underlying concept of Krein space induced by a selfadjoint operator and that of Krein space continuously embedded. The operator range feature of reproducing kernel spaces is emphasized. We include a careful presentation of Hermitian kernels on complex regions that points out a universality property of the Szego kernels with respect to reproducing kernel Krein spaces of holomorphic functions. | math.FA | math | A SURVEY ON REPRODUCING KERNEL KREIN SPACES
AURELIAN GHEONDEA
Abstract. This is a survey on reproducing kernel Kreın spaces and their interplay with
operator valued Hermitian kernels. Existence and uniqueness properties are carefully re-
viewed. The approach used in this survey involves the more abstract, but very useful,
concept of linearisation or Kolmogorov decomposition, as well as the underlying concepts
of Kreın space induced by a selfadjoint operator and that of Kreın space continuously
embedded. The operator range feature of reproducing kernel spaces is emphasised. A
careful presentation of Hermitian kernels on complex regions that points out a univer-
sality property of the Szego kernels with respect to reproducing kernel Kreın spaces of
holomorphic functions is included.
Acknowledgement. Work supported by a grant of the Romanian National Authority for
Scientific Research, CNCS UEFISCDI, project number PN-II-ID-PCE-2011-3-0119.
Key words and phrases. Kreın space, reproducing kernel, Hermitian kernel, linearisation,
Kolmogorov decomposition, induced space, holomorphy.
2010 Mathematics Subject Classification. 46C07, 46C20, 47B32, 47B50, 32A25.
3
1
0
2
v
o
N
2
2
]
.
A
F
h
t
a
m
[
2
v
3
9
3
2
.
9
0
3
1
:
v
i
X
r
a
1. Introduction
A kernel K on a set X and with operator entries is the analog of an abstract matrix
double indexed on X and for which each entry is a bounded linear operator acting between
appropriate spaces. The adjoint K ♯ of the kernel K can be defined by analogy with the
adjoint of matrices. A Hermitian kernel K is defined by the property K = K ♯. Roughly
speaking, a reproducing kernel Kreın space on a set X is a Kreın space R of functions on
X for which there exists a Hermitian kernel with the property that the evaluations of the
functions in R can be calculated in terms of the kernel K. One of the main problems is
that of associating a reproducing kernel Kreın space R to a given Hermitian kernel K.
The classical theory says that for any reproducing kernel Hilbert space R its reproducing
kernel K is positive semidefinite and uniquely determined by R. Conversely, to any
positive semidefinite kernel K there is associated a unique reproducing kernel Hilbert
space such that K is its reproducing kernel. These facts can be proven also for a slightly
more general situation relating in a similar fashion reproducing kernel Pontryagin spaces
to Hermitian kernels with finite negative (or positive) signatures.
For genuine indefinite Hermitian kernels the correspondence with reproducing kernel
Kreın spaces is much more complicated. There is a variety of characterisations of those
Hermitian kernels that produce reproducing kernel Kreın spaces in terms of boundedness
with respect to positive semidefinite kernels and in terms of decomposability as a difference
of two positive semidefinite kernels. Uniqueness of the reproducing kernel Kreın space,
provided that it exists, is also problematic. Characterisations of uniqueness are available in
terms of lateral spectral gaps as well as in terms of maximal uniformly positive subspaces.
1
2
A. GHEONDEA
When Hermitian kernels on regions in either a single complex variable or several complex
variables have certain holomorphy properties, existence of reproducing kernel Kreın spaces
is guaranteed. This fact is related to a certain universality property of the Szego kernel
that allows the construction of the corresponding reproducing kernel Kreın space inside
the Hardy space or, respectively, the Drury-Arveson space.
The approach used in this survey involves a more abstract, but very useful, concept
of linearisation or Kolmogorov decomposition, the underlying concept of Kreın spaces
induced by selfadjoint operators, as well as the concept of Kreın spaces continuously
embedded, pointing out the operator range feature of reproducing kernel spaces.
2. Kreın Spaces and their Linear Operators
A Kreın space K is a complex linear space on which it is defined an indefinite scalar
product [·, ·] such that K is decomposed in a direct sum
(2.1)
K = K+[ +]K−
in such a way that K± are Hilbert spaces with scalar products ±[·, ·], and the direct sum
in (2.1) is orthogonal with respect to the indefinite scalar product [·, ·], i.e. K+ ∩K− = {0}
and [x+, x−] = 0 for all x± ∈ K±. The decomposition (2.1) gives rise to a positive definite
scalar product h·, ·i by setting hx, yi := hx+, y+i−hx−, y−i, where x = x++x−, y = y++y−,
and x±, y± ∈ K±. The scalar product h·, ·i defines on K a structure of Hilbert space.
Subspaces K± are orthogonal with respect to the scalar product h·, ·i, too. One denotes
by P± the corresponding orthogonal projections onto K±, and let J = P+ − P−. The
operator J is a symmetry, i.e. a selfadjoint and unitary operator, J ∗J = JJ ∗ = J 2 = I.
Unless otherwise specified, any Kreın space is considered as a Banach space with the norm
given by an arbitrary norm induced by the positive definite inner product associated to
a fundamental decomposition. For two Kreın spaces K1 and K2, we denote by L(K1, K2)
the Banach space of all bounded linear operators T : K1 → K2.
Given a Kreın space (K, [·, ·]), the cardinal numbers
(2.2)
κ+(K) = dim(K+),
κ−(K) = dim(K−),
do not depend on the fundamental decomposition and they are called, respectively, the
geometric ranks of positivity/negativity of K.
The operator J is called a fundamental symmetry of the Kreın space K. Note that
[x, y] = hJx, yi, (x, y ∈ K). If T is a densely defined operator from a Kreın space K1 to
another Kreın space K2, the adjoint of T can be defined as an operator T ♯ defined on the
set of all y ∈ K2 for which there exists hy ∈ K1 such that [T x, y] = [x, hy], and T ♯y = hy.
In addition, T ♯ = J1T ∗J2, where T ∗ denotes the adjoint operator of T with respect to
the Hilbert spaces (K1, h·, ·iJ1) and (K2, h·, ·iJ2). The symbol ♯ denotes the adjoint when
at least of the spaces K1 or K2 is indefinite. In the case of an operator T defined on the
Kreın space K, T is called symmetric if T ⊆ T ♯, i.e. if the relation [T x, y] = [x, T y] holds
for each x, y ∈ Dom(T ), and T is called selfadjoint if T = T ♯.
In this survey, a bit of geometry of Kreın spaces is used. Thus, a (closed) subspace
L of a Kreın space K is called regular if K = L + L⊥, where L⊥ = {x ∈ K [x, y] =
0 for all y ∈ L}. Regular spaces of Kreın spaces are important since they are exactly the
analogue of Kreın subspaces, that is, if we want L be a Kreın space with the restricted
indefinite inner product and the same strong topology, then it should be regular.
REPRODUCING KERNEL KREIN SPACES
3
In addition, let us recall that, given a subspace L of a Kreın space, one calls L non-
negative (positive) if the inequality [x, x] ≥ 0 holds for x ∈ L (respectively, [x, x] > 0 for
all x ∈ L \ {0}). Similarly one defines non-positive and negative subspaces. A subspace
L is called degenerate if L ∩ L⊥ 6= {0}. Regular subspaces are non-degenerate. As
a consequence of the Schwarz inequality, if a subspace L is either positive or negative
it is nondegenerate. A remarkable class of subspaces are those regular spaces that are
either positive or negative, for which the terms uniformly positive, respectively, uniformly
negative are used. These notions can be defined for linear manifolds also, that is, without
assuming closedness.
A linear operator V defined from a subspace of a Kreın space K1 and valued into another
Kreın space K2 is called isometric if [V x, V y] = [x, y] for all x, y in the domain of V . Note
that isometric operators between genuine Kreın spaces may be unbounded and different
criteria of boundedness are available. However, in this presentation, a unitary operator
between Kreın spaces means that it is a bounded isometric operator that have bounded
inverse.
3. Hermitian Kernels
Let X be a nonempty set and let H = {Hx}x∈X be a family of Kreın spaces with
inner products denoted by [·, ·]Hx. A mapping K defined on X × X such that K(x, y) ∈
L(Hy, Hx) for all x, y ∈ X is called an H-kernel on X. In case Hx = H for all x ∈ X,
where H is some fixed Kreın space, one talks about an H-kernel on X, while, even more
particularly, if H = C, one talks about a scalar kernel on X, or simply a kernel on X.
To any H-kernel K one associates its adjoint K ♯ defined by K ♯(x, y) = K(y, x)♯, for all
x, y ∈ X. The H-kernel K is called Hermitian if K = K ♯, that is,
(3.1)
K(x, y) = K(y, x)♯,
x, y ∈ X.
Denote by F (H) the set of all H-vector fields f on X, that is, f = {fx}x∈X such that
fx ∈ Hx, for all x ∈ X, and let F0(H) denote the set of all f ∈ F (H) of finite support,
that is, the set supp(f ) = {x ∈ X fx 6= 0} is finite. Alternatively, one can view any
f ∈ F (H) as a function f : X →Sx∈X Hx such that f (x) ∈ Hx for all x ∈ X.
If K is a Hermitian H-kernel then one can introduce on F0(H) an inner product [·, ·]K
defined by
(3.2)
[f, g]K = Xx,y∈X
[K(x, y)f (y), g(x)]Hx,
f, g ∈ F0(H).
The H-kernel K is called positive semidefinite if
(3.3)
Xx,y∈X
[K(x, y)h(y), h(x)]Hx ≥ 0,
h ∈ F0(H).
Every positive semidefinite H-kernel is Hermitian. Also, a Hermitian H-kernel is positive
semidefinite if and only if the corresponding inner product in (3.2) is nonnegative.
Let us denote by Kh(H) the class of all Hermitian H-kernels and by K+(H) the subclass
of all positive semidefinite H-kernels. On Kh(H) one defines addition, subtraction and
multiplication with real numbers in a natural way. Moreover, on Kh(H) one has a natural
4
A. GHEONDEA
partial order defined as follows: if H, K ∈ Kh(H) then H ≤ K means [f, f ]H ≤ [f, f ]K,
for all f ∈ F0(H). With this definition one gets
(3.4)
K+(H) = {H ∈ Kh(H) H ≥ 0},
and K+(H) is a strict cone of Kh(H), that is, it is closed under addition and multiplication
with nonnegative numbers, and K+(H)T −K+(H) = {0}, where 0 denotes the null kernel.
More generally, one can define the signatures of K, denoted by κ±(K), as the posi-
tive/negative signatures of the inner product [·, ·]K. Then, if κ(K) = min{κ−(K), κ+(K)}
denotes the definiteness signature of K, the Hermtian H-kernel K is called quasi semi-
definite if κ(K) < ∞, that is, either κ−(K) or κ+(K) is finite.
A pairing (·, ·)F can be defined for arbitrary f, g ∈ F (H), provided at least one of f
and g has finite support, by
(3.5)
(f, g)F = Xx∈X
[f (x), g(x)]Hx.
When restricted to F0(H) this pairing becomes a nondegenerate inner product. To each
H-kernel K one associates the convolution operator, denoted also by K, and defined by
(3.6)
K : F0(H) → F (H),
K(x, y)f (y),
f ∈ F0(H).
(Kf )(x) =Xy∈X
Then
(3.7)
[f, g]K = (Kf, g)F ,
f, g ∈ F0(H).
Consequently, the kernel K is positive semidefinite (Hermitian) if and only if the corre-
sponding convolution operator K is positive semidefinite (Hermitian), that is, (Kf, g)F ≥
0 ((Kf, g)F = (f, Kg)F), for all f, g ∈ F0(H). Similar assertions can be made about the
signatures κ±(K), with appropriate definitions of signatures of Hermitian operators on
inner product spaces.
At this point, it is worth noting that there is no restriction of generality if one assumes
that all spaces Hx are Hilbert. To see this, fixing a fundamental symmetry Jx on each
Kreın space H, one can refer to the Hilbert spaces (Hx; h·, ·iJx) and, one considers the
kernel H on X defined by H(x, y) = JxK(x, y)Jy, for all x, y ∈ X. Taking into account
that A♯ = JxA∗Jy for any bounded linear operator A : Hx → Hy, it follows that all notions
like Hermitian, positive semidefinite, signatures, etc. defined for the kernel K, have word-
for-word transcriptions for the kernel H. However, since this survey is focusing on the
indefinite case, allowing right from the beginning Kreın spaces Hx brings more symmetry
and simpler formulas.
A reproducing kernel inner product space, with respect to the set X, the collection of
Kreın spaces H = {Hx}x∈X, and H-kernel K, is an inner product space (R, [·, ·]R) subject
to the following conditions:
(rk1) R ⊆ F (H).
(rk2) For all y ∈ X and all h ∈ Hy the map X ∋ x 7→ K(x, y)h ∈ Hx belongs to R.
(rk3) [f (x), h]Hx = [f, K(·, x)h]R, for all f ∈ R, x ∈ X, and h ∈ Hx.
The axiom (rk3) is usually called the reproducing property while the H-kernel K is called
the reproducing kernel of R. Note that K is necessarily a Hermitian H-kernel.
REPRODUCING KERNEL KREIN SPACES
5
According to the axiom (rk2), it is useful to consider the notation Ky = K(·, y) : X →
L(Hy, R), for y ∈ X. An immediate consequence of the axioms (rk1)–(rk3) is that the
inner product [·, ·]R is nondegenerate. Also, recalling that on any nondegenerate inner
product space the weak topology is separated, the following minimality property holds
(rk4) The span of {KxHx x ∈ X} is weakly dense in R.
Also, the reproducing kernel is uniquely determined by the reproducing kernel inner prod-
uct space.
Given a Hermitian H-kernel K on X, consider the subspace R0(H) of F (H) spanned
by Kxh, for all x ∈ X and all h ∈ Hx, on which one can define the inner product
(3.8)
[
mXi=1
Kxihi,
nXj=1
Kyj kj]R0 =
mXi=1
nXj=1
[K(yj, xi)hi, kj]Hyj
,
This definition can be proven to be correct: vectors in R0(H) may have different rep-
resentations as linear combinations of Kyk but the definition in (3.8) is independent of
these. The subspace R0(H) of F (H) is the range of the convolution operator K defined
at (3.6) and the inner product [·, ·]R0 is nondegenerate. In addition, (R0(H); [·, ·]R0) is a
reproducing kernel inner product space with reproducing kernel K.
In case the reproducing kernel inner product space (R; [·, ·]R) is a Kreın space, one talks
about a reproducing kernel Kreın space. The uniqueness of the reproducing kernel of a
Kreın space has a stronger characterisation.
Theorem 3.1. Let K be a Kreın space of H-valued vector fields on X, that is, K ⊆ F (H).
For each x ∈ X consider the linear operator E(x) : K → Hx of evaluation at x, that is,
E(x)f = f (x) for all f ∈ K. Then, K has a reproducing kernel if and only if E(x) is
bounded for all x ∈ X. In this case, the reproducing kernel of K is
(3.9)
K(x, y) = E(x)E(y)♯,
x, y ∈ X,
hence uniquely determined by K.
With the notation as in Theorem 3.1, it is useful to note that, if K is a reproducing
kernel Kreın space with reproducing kernel K, then Kx can be viewed as a linear operator
Hy → K and Kx = E(x)♯ for all x ∈ X. In particular,
(3.10)
K(x, y) = K ♯
xKy,
x, y ∈ X.
Conversely, given a Hermitian H-kernel K on X, the questions on existence and unique-
ness of a reproducing kernel Kreın space R ⊆ F (H) such that K is its reproducing kernel
are much more difficult. If (R; [·, ·]R) is a reproducing kernel Kreın space with reproducing
kernel K, in view of the axioms (rk1)–(rk3) and the minimality property (rk4), the inner
products [·, ·]R and [·, ·]R0 coincide on R0(H) and R0(H) is dense in R. Thus, existence of
a reproducing kernel Kreın space associated to a given H-kernel K on X depends heavily
on the possibility of "completing" the inner product space (R0(H); [·, ·]R0) to a Kreın
space R inside F (H), which is a core problem in the theory of indefinite inner product
spaces.
In order to tackle these questions, let us note that if R is a reproducing kernel Hilbert
space with reproducing kernel K, then K should be positive semidefinite. Conversely,
if K is a positive semidefinite H-kernel on X, then the inner product space (R0, [·, ·]R0)
6
A. GHEONDEA
defined at (3.8) is a pre-Hilbert space and the existence of a reproducing kernel Hilbert
space R with reproducing kernel K depends on whether (R0, [·, ·]R0) has a completion
inside of F (H). Actually, this is always the case and, moreover, uniqueness holds as
well, but these two facts are slightly more general, namely, they are true if K is quasi
semidefinite, in which case it is obtained a unique reproducing kernel Pontryagin space,
cf. Section 5.
In this survey, existence and uniqueness of reproducing kernel Kreın spaces associated
to Hermitian kernels are approached through the more abstract, but very useful, concept
of linearisation or Kolmogorov decomposition. By definition, a linearisation, sometimes
called Kolmogorov decomposition, of the H-kernel K is a pair (K; V ), subject to the
following conditions:
(kd1) K is a Kreın space and V (x) ∈ L(Hx, K) for all x ∈ X.
(kd2) K(x, y) = V (x)♯V (y) for all x, y ∈ X.
The linearisation (K; V ) is called minimal if the following condition holds as well:
(kd3) K = Wx∈X
V (x)Hx.
Two linearisations (Kj; Vj), j = 1, 2, of the same H-kernel K are called unitary equiv-
alent if there exists a Kreın space bounded unitary operator U : K1 → K2 such that
V2(x) = UV1(x) for all x ∈ X.
There is a close connection between the notion of reproducing kernel Kreın space with
reproducing H-kernel K and that of minimal linearisation of K.
Proposition 3.2. (1) Let R be a reproducing kernel Kreın space with reproducing H-
kernel K. Then, letting
(3.11)
V (x) = Kx = K(·, x) ∈ L(Hx, R),
the pair (R; V ) is a minimal linearisation of the H-kernel K.
Conversely, letting (K; V ) be a minimal linearisation of the H-kernel K, then,
(3.12)
R = {V (·)♯ k ∈ K},
is a vector subspace of F (H) which, with respect to the the inner product defined by
(3.13)
[V (·)♯h, V (·)♯k]R = [h, k]K,
h, k ∈ K,
is a Kreın space with reproducing kernel K.
(2) In the correspondence defined at (3.12) and (3.13), two unitary equivalent minimal
realisations of the same H-kernel K produce the same reproducing kernel Krein space and
hence, the correspondence between reproducing kernel Krein spaces and minimal linearisa-
tions is one-to-one, provided that unitary equivalent minimal linearisations are identified.
4. Some Examples
Example 4.1. Matrices. Let X = {1, 2, . . . , n} and Hx = C for each x ∈ X. A kernel on
X is simply a map K : X × X → C and hence, it can be identified with the n × n complex
matrix [K(i, j)]n
i,j=1 is
Hermitian. The vector space F = F0 = Cn hence, in view of the definition (3.6), the
convolution operator K : F → F is simply the linear operator K : Cn → Cn associated to
the matrix K.
i,j=1. The kernel K is Hermitian if and only the matrix [K(i, j)]n
REPRODUCING KERNEL KREIN SPACES
7
If K is a Hermitian kernel on X then the inner product [·, ·]K is the familiar inner
i,j=1. The signatures κ±(K) coincide,
i,j=1,
i,j=1
product associated to the Hermitian matrix [K(i, j)]n
respectively, with the number of positive/negative eigenvalues of the matrix [K(i, j)]n
counted with multiplicities. K is positive semidefinite if and only if the matrix [K(i, j)]n
is positive semidefinite.
Assuming that the kernel K on X is Hermitian, for each j = 1, . . . , n, the "map"
i=1 in Cn = F . The vector space R0, the
Kj : X → C is simply the column vector [K(i, j)]n
range of the convolution operator K, is the vector subspace of Cn generated by all the
column vectors [K(i, j)]n
i=1. In this particular case, R0 is the reproducing kernel Kreın
space with reproducing kernel K. The inner product [·, ·]R0 is defined at (3.8).
Example 4.2. Operator Block Matrices. Let X = {1, 2, . . . , n} but, this time, for each
x ∈ X one denotes by Hx a Hilbert space and let H = {Hx}x∈X. An H-kernel on X,
originally defined as a map K on X × X such that K(x, y) ∈ L(Hy, Hx), is naturally
identified with the operator block matrix [K(i, j)]n
i,j=1. The vector space F (H) = F0(H)
together with its natural inner product (·, ·)F defined at (3.5) is naturally identified with
i=1 Hi. Then, the convolution operator K associated to the H-
kernel K is identified with the bounded linear operator K : H → H naturally associated
to the operator block matrix [K(i, j)]n
i,j=1. For each j = 1, . . . , n, the map Kj = K(·, j)
is the operator block column matrix [K(i, j)]n
the Hilbert space H =Ln
i=1 : Hj → H.
The H-kernel K on X is Hermitian if and only if the corresponding operator block
matrix [K(i, j)]n
i,j=1 is Hermitian, if and only if the convolution operator K ∈ L(H) is
selfadjoint (Hermitian). In this case, the vector space R0(H) is identified with the range
of the convolution operator K as a subspace of H and is spanned by the ranges of the
operator block column matrices Kj = [K(i, j)]n
i=1 : Hj → H, for j = 1, . . . , n. The inner
product [·, ·]R0 is defined as in (3.8).
Example 4.3. The Hardy space H 2(D). Let D denote the open unit ball in the complex
field and consider the Szego kernel
(4.1)
S(z, w) =
1
1 − zw
=
∞Xn=0
wnzn,
z, w ∈ D,
the series converging absolutely and uniformly on any compact subset of D × D. Clearly,
S is a Hermitian scalar kernel on D and for every w ∈ D
Sw(z) = S(z, w) =
∞Xn=0
wnzn,
z, w ∈ D,
the series converging absolutely and uniformly on any compact subset of D, Sw is a
holomorphic function on D, and its Taylor coefficients are 1, w, w2, . . .. In order to describe
the reproducing kernel inner product space associated to the Szego kernel S, in view of
(3.8), one has to consider the vector space generated by the functions Sw, w ∈ D, and
complete it with respect to the inner product
(4.2)
hSu, Swi = Su(w) = S(w, u) =
∞Xn=0
unwn,
u, v ∈ D.
8
A. GHEONDEA
This completion is called the Hardy space, denoted by H 2(D), and in view of (4.2), is
(4.3)
H 2(D) = {f f (z) =
∞Xn=0
anzn,
2Xn=0
an2 < ∞}.
Since H 2(D) is a Hilbert space, it follows that the Szego kernel is positive semidefinite.
The Hardy space H 2(D) has some other special properties, among which, a distinguished
role is played by the boundary values of its functions. More precisely, for each f ∈ H 2(D),
(4.4)
there exists ef a function defined on the unit circle T = ∂D, such that
The function ef is uniquely determined by f , a.e. on T, and ef ∈ L2(T). Usually, there is
no distinction between a function f in H 2(D) and ef , that is, the tilde sign is not used at
all. In this way, an isometric embedding of H 2(D) into L2(T) is defined, in particular,
ef (eit) = lim
a.e. t ∈ [0, 2π).
f (reit),
r→1−
(4.5)
hf, giH 2(D) = hf, giL2(T) =
f (eit)g(eit)dt,
f, g ∈ H 2(D).
1
2πZ 2π
0
With respect to this embedding, H 2(D) is identified with the subspace of L2(T) of all
functions f whose Fourier coefficients of negative index vanish.
Example 4.4. The Drury-Arveson Space. Let Br(ξ) be the open ball of radius r and
n=1 ηnξn for
center ξ in the Hilbert space G = CN , with inner product hξ, ηiG = η · ξ =PN
n=1 in CN . We write Br instead of Br(0). The Szego kernel is
n=1 and η = (ηn)N
ξ = (ξn)N
(4.6)
S(ξ, η) =
1
1 − hξ, ηiG
,
ξ, η ∈ B1,
and note that S is a scalar Hermitian kernel on B1. We now describe a minimal lineari-
G⊗n be the Fock space associated to G = CN , where G⊗0 = C
sation of S. Let F (G) =
and G⊗n is the n-fold Hilbert space tensor product of G with itself, hence F (G) is a Hilbert
space. Let
∞Ln=0
(4.7)
Pn =
1
n! Xπ∈Sn
π
be the orthogonal projection of G⊗n onto its symmetric part, where π(ξ1 ⊗ . . . ⊗ ξn) =
ξπ−1(1) ⊗ . . . ⊗ ξπ−1(n) for any element π of the permutation group Sn on n symbols. Recall
Pn)F (G). For ξ ∈ B1 set ξ⊗0 = 1 and let
that the symmetric Fock space is F s(G) = (
∞Ln=0
G⊗n =Xn≥0
ξ⊗n denote the n-fold tensor product ξ ⊗ . . . ⊗ ξ, n ≥ 1. Note that
1
kMn≥0
ξ⊗nk2
F (G) =Xn≥0
kξ⊗nk2
kξk2n
G =
.
1 − kξk2
G
Hence Ln≥0
(4.8)
ξ⊗n ∈ F s(G) and one can define the mapping VS from B1 into F s(G),
VS(ξ) =Mn≥0
ξ⊗n,
ξ ∈ G.
REPRODUCING KERNEL KREIN SPACES
9
The pair (F s(G); VS) is a minimal linearisation of the kernel S. In order to see this,
VS(ξ) is also viewed as a bounded linear operator from C into F s(G) by VS(ξ)λ = λVS(ξ),
λ ∈ C, so that, for ξ, η ∈ B1,
VS(η)∗VS(ξ) = hVS(ξ), VS(η)iF s(G) =Xn≥0
hξ⊗n, η⊗niG⊗n =Xn≥0
(η · ξ)n =
1
1 − η · ξ
= S(ξ, η).
In particular, this shows that the Szego kernel is positive semidefinite. The set {VS(ξ)
ξ ∈ B1} is total in F s(G) since for n ≥ 1 and ξ ∈ G one has
dn
dtn V (tξ)t=0 = n!ξ⊗n.
The reproducing kernel Hilbert space associated to the Szego kernel S, called the Drury-
Arveson space and denoted by H 2(B1), is given by the completion of the linear space
generated by the functions Sη = S(·, η), η ∈ B1, with respect to the inner product
defined by hSη, SξiH 2(B1) = S(ξ, η), see (3.8). We use the multiindex notation: for n =
(n1, . . . , nN ) ∈ NN
N . Then,
anξn,
a function f holomorphic in B1, with Taylor series representation f (z) = Pn∈NN
0 , let n = n1 + · · · + nN , n! = n1! · · · nN !, and ξn = ξn1
belongs to H 2(B1) if and only if
1 · · · ξnN
0
(4.9)
kf k2
H 2(B1) = Xn∈NN
0
n!
n!
an2 < ∞.
Note that there exists a unitary operator Φ from the Drury-Arveson space H 2(B1) onto
F s(G) such that ΦSξ = VS(ξ), ξ ∈ B1.
If N = 1, the Drury-Arveson space coincides with the Hardy space H 2(D) described at
the previous example.
Example 4.5. Holomorphic Kernels. Given two Kreın spaces G and H let Ω be a sub-
region of the open unit disc D in the complex plane and Θ : Ω → L(G, H). One considers
the following kernels
(4.10)
(4.11)
(4.12)
KΘ(z, w) =
K eΘ(z, w) =
I − Θ(z)Θ(w)♯
1 − zw
I − eΘ(z)eΘ(w)♯
1 − zw
,
,
DΘ(z, w) =
KΘ(z, w)
eΘ(z) − eΘ(w)
z − w
Θ(z) − Θ(w)
z − w
K eΘ(z, w)
,
where eΘ(z) = Θ(z)♯. These are operator valued Hermitian holomorphic kernels that are
of interest in connection to Schur classes and their generalisations.
The classical Schur class corresponds to the case when G and H are Hilbert spaces and
the Schur kernel KΘ is positive semidefinite. In case G and H are Kreın or even genuine
Pontryagin spaces, in order to define the Schur class, positive semidefiniteness has to be
imposed on all KΘ, K eΘ and DΘ. Given κ ∈ N, the generalized Schur class corresponds
to the requirement that the negative signatures of each kernel KΘ, K eΘ and DΘ are κ.
10
A. GHEONDEA
When G = H and for appropriate regions in the complex field, the Carath´eodory kernel
CΘ
(4.13)
CΘ(z, ζ) =
1
2
Θ(z) − Θ(ζ)♯
1 − ζz
,
as well as the Nevanlinna kernel NΘ
(4.14)
NΘ(z, ζ) =
Θ(z) − Θ(ζ)♯
z − ζ
,
are of special interest. As in the case of the Schur classes, positive semidefiniteness of
the corresponding kernels define the Carath´eodory class and, respectively, the Nevanlinna
class of holomorphic functions. Generalized classes correspond to the case when the
appropriate kernels have fixed negative signatures.
Example 4.6. Toplitz Kernels. Let H be a Kreın space and X = Z, the set of integer
numbers. An H-kernel on Z, K : Z × Z → L(H), is called a Toplitz Kernel if K(i, j) =
T (i − j) for some function T : Z → L(H), called the symbol of K.
One considers the set T(H) of all Toplitz H-kernels. In the following it is considered
the subclass of Toplitz Hermitian H-kernels Th(H) and the subclass of Toplitz positive
semidefinite H-kernels T+(H). One notes that Th(H) is closed under addition, subtraction
and (left and right) multiplication with bounded operators on H. Also, T+(H) is a strict
cone of Th(H).
Considering the complex vector space F0(H) of all functions h : Z → H with finite
support and for an arbitrary Hermitian kernel H ∈ Kh(H) one associates the inner product
space (F0, [·, ·]H) as in the previous sections. On the vector space F0(H) one considers
two operators, the forward shift S+ defined by (S+h)(n) = h(n − 1), for all h ∈ F0(H)
and n ∈ Z, and the backward shift S− defined by (S−h)(n) = h(n + 1), for all h ∈ F0(H)
and all n ∈ Z. If H ∈ Kh(H) then it is a Toplitz kernel if and only if
[S+h, g]H = [h, S−g]H,
f, g ∈ F0(H).
If H is a Hermitian Toplitz H-kernel then both S+ and S− are isometric with respect to
the inner product [·, ·]H, that is, for all h, g ∈ F0(H) the equalities [S+h, S+g]H = [h, g]H
and [S−h, S−g]H = [h, g]H hold. The converse is also true, if either S+ or S− is isometric
with respect to the Hermitian H-kernel H then H is Toplitz.
Let H be a Toplitz Hermitian H-kernel. A Naımark dilation of H is, by definition, a
triple (U, Q; K) with the following properties:
(nd1) K is Kreın space, U ∈ L(K) is a unitary operator, and Q ∈ L(H, K).
(nd2) K = Wn∈Z
U nQH.
(nd3) H(i, j) = Q♯U i−jQ, i, j ∈ Z.
If the Toplitz Hermitian H-kernel H has a Naımark dilation (U, Q; K), letting V (n) =
U nQ ∈ L(H, K) it is readily verified that the pair (V ; K) is a linearisation of H.
REPRODUCING KERNEL KREIN SPACES
11
5. Quasi Semidefinite Kernels
In the following, for simplicity and because most of the applications presented in this
survey do not require the full generality as considered in Section 3, there is consider only
the case of H-kernels, that is, Hx = H for some fixed Kreın space H. A Hermitian kernel
H : X × X → L(H) is associated to an indefinite inner product space [·, ·]H on F (H),
the vector space of all complex valued functions f : X → H, as in (3.2). Recall that
κ±(H), the signatures of H, are defined as the positive/negative signatures of the inner
product [·, ·]H, while its rank of indefiniteness is defined by κ(H) = min{κ−(H), κ+(H)}.
A Hermitian H-kernel H is quasi semidefinite if κ(H) < ∞, that is, either κ−(H) or
κ+(H) is finite.
Theorem 5.1. Let H be quasi semidefinite H-kernel on X. Then:
(a) H admits a minimal linearisation to a Pontryagin space (K; V ) with κ(H) = κ(K),
unique up to a unitary equivalence.
(b) There exists a unique reproducing kernel Pontryagin space R on X and with re-
producing kernel H.
On the vector space F (H) of H-fields on X, the inner product [·, ·]H, see (3.2) is defined.
Since κ(H) is finite, the inner product space (F (H); [·, ·]H) is decomposable. To make a
choice, assume that κ−(H) < ∞, hence
(5.1)
F (H) = F−[+]F0[+]F+,
where F0 is the isotropic subspace, F± are positive/negative subspaces, and dim(F−) =
κ−(H) < ∞. Factoring out F0 one can assume, without loss of generality, that F0 = 0.
Then, let K+ denote the completion of (F+; [·, ·]H) to a Hilbert space and then K =
F−[+]K+ is a Pontryagin space with κ−(K) = κ−(H). For arbitrary x ∈ X, the linear
operator V (x) : H → K is defined by assigning to each vector h ∈ H the H-vector field
f : X → H given by f (x) = h and f (y) = 0 for all y ∈ X, y 6= x. Then (K; V ) is a
minimal lineariasation of H.
The uniqueness of minimal linearisations of H, modulo unitary equivalence, follows from
the fact that any dense linear subspace of a Pontryagin space with finite negative/positive
signature contains a maximal negative/positive subspace and the continuity of isometric
densely defined operators between Pontryagin spaces with the same negative/positive
signature.
From Proposition 3.2 (1), once one gets a minimal linearisation (K; V ) of H one imme-
diately obtains a reproducing kernel Pontryagin space R with reproducing kernel H as
defined in (3.12) and (3.13). The uniqueness of the reproducing kernel Pontryagin space
R follows from Proposition 3.2 (2) and the uniqueness, modulo unitary equivalence, of
the minimal linearisation of H.
There is a more direct but longer way of constructing the reproducing kernel Kreın
space associated to a quasi positive semidefinite H-kernel H, by considering the vector
space R0(H) linearly generated by Hxh, for x ∈ X and h ∈ H on which the inner product
[·, ·]R0 is defined as in (3.8) in such a way that (R0(H); [·, ·]R0) is a reproducing kernel inner
product space with H its reproducing kernel. The inner product space (R0(H); [·, ·]R0)
is nondegenerate and its negative signature is the same with κ−(H), hence finite. Then
R0(H) is decomposable, hence R0(H) = R−[+]R0,+, where dim(R−) = κ−(H) < ∞
12
A. GHEONDEA
is negative definite, while R0,+ is positive definite.
It can be proven that R0,+ has a
completion to a Hilbert space R+ inside F (H) and then R = R−[+]R+ is the reproducing
kernel Kreın space of H.
The next theorem points out a property of propagation to arbitrarily large domains of
holomorphy for quasi semidefinite holomorphic kernels.
Theorem 5.2. Let K be a Hermitian holomorphic H-kernel on some region Ω and let Ω0
be a subregion of Ω. If κ−(K0) < ∞, where K0 is the restriction of K to Ω0 × Ω0, then
κ−(K) = κ−(K0).
The idea of proof is to use Cauchy's Theorem in order to prove the propagation property
for positive semidefinite kernels, first for disks around the origin of the complex plane,
then for union of regions for which the kernel is positive definite, and finally to use the
decomposition of K = K+ − K−, where K± are positive semidefinite kernels and K− has
a reproducing Hilbert space of dimension κ−(K), that can be obtained from (5.1).
6. Induced Kreın Spaces
Let (H, h·, ·iH) be a Hilbert space and consider a bounded selfadjoint operator A on H.
A new inner product on H is defined by
(6.1)
[h, k]A = hAh, kiH,
h, k ∈ H.
In this section the properties of some Kreın spaces associated with this inner product are
described.
By definition, a Kreın space induced by A is a pair (K, Π), where
(ik1) K is a Kreın space and Π ∈ L(H, K) has dense range.
(ik2) [Πx, Πy]K = hAx, yiH, for all x, y ∈ H.
There exist at least two main constructions of Kreın spaces induced by selfadjoint
operators, and they are related by certain unitary equivalences. Two Kreın spaces (Ki, Πi),
i = 1, 2, induced by the same A, are unitary equivalent if there exists a unitary operator
U ∈ L(K1, K2) such that UΠ1 = Π2.
Example 6.1. The Induced Kreın Space (KA; ΠA). Let A denote the selfadjoint operator
with respect to the Hilbert space (H, h·, ·iH). Let H− and H+ be the spectral subspaces
corresponding to the semiaxis (−∞, 0) and, respectively, (0, +∞) and the operator A.
Then one gets the decomposition
H = H− ⊕ Ker A ⊕ H+.
Note that (H−, −[·, ·]A) and (H+, [·, ·]A) are positive definite inner product spaces and
hence they can be completed to Hilbert spaces K− and, respectively, K+. We can build
the Kreın space (KA, [·, ·]A) by letting
(6.2)
KA = K−[+]K+.
The operator ΠA ∈ L(H, KA) is, by definition, the composition of the orthogonal pro-
jection of H onto H ⊖ Ker A with the embedding of H ⊖ Ker A into KA. With these
definitions, it is readily verified that (KA, ΠA) is a Kreın space induced by A.
REPRODUCING KERNEL KREIN SPACES
13
In order to take a closer look at the strong topology of the Kreın space KA, consider the
seminorm H ∋ x 7→ kA1/2xk = hA1/2x, A1/2xi1/2. The kernel of this seminorm is ex-
actly Ker A and the completion of H⊖Ker A with respect to this norm is exactly the space
KA. Moreover, the strong topology of KA is induced by the extension of this seminorm.
The positive definite inner product on KA associated with the norm kA1/2 · k is hA·, ·i.
Hence, if A = SAA is the polar decomposition of A and SA denotes the corresponding
selfadjoint partial isometry, then it follows that SA can be extended by continuity to KA
and this extension is exactly the fundamental symmetry of KA corresponding to hA·, ·i.
Example 6.2. The Induced Kreın Space (BA; ΠBA). With notation as in Example 6.1,
consider the polar decomposition of A = SAA as in Example 6.1. Define the space
BA = Ran(A1/2) endowed with the positive definite inner product h·, ·iBA given by
(6.3)
hA1/2x, A1/2yiBA = hPH⊖Ker Ax, yiH,
x, y ∈ H.
This positive definite inner product is correctly defined and (BA, h·, ·iBA) is a Hilbert space.
To see this, just note that the operator A1/2 : H ⊖ Ker A → BA is a Hilbert space unitary
operator.
On BA one defines the inner product [·, ·]BA by
(6.4)
Since the operator A1/2 and SA commute it follows that
[A1/2x, A1/2y]BA = hSAx, yiH,
x, y ∈ H.
[a, b]BA = hSAa, biBA,
a, b ∈ BA.
The operator SABA is a symmetry. We define now a linear operator ΠBA : H → BA by
(6.5)
ΠBAh = Ah,
h ∈ H.
It follows that (BA, [·, ·]BA) is a Kreın space induced by A.
The Kreın space induced by a selfadjoint operator as in Example 6.2 is a genuine
operator range subspace. More precisely, to any Kreın space K continuously embedded in
the Hilbert space H one associates a selfadjoint operator A ∈ L(H) in the following way:
let ι : K → H be the inclusion operator which is supposed bounded and take A = ιι♯ ∈
L(H). Clearly A is selfadjoint and (K, ι♯) is a Kreın space induced by A. Conversely,
from Example 6.2 it is easy to see that BA is a Kreın space continuously embedded in H.
The connection between the induced Kreın spaces (KA, ΠA) and (BA, ΠBA) is explained
by the following
Proposition 6.3. The induced Kreın spaces (KA, ΠA) and (BA, ΠBA) are unitary equiv-
alent, more precisely, the mapping
extends uniquely to a unitary operator V ∈ L(KA, BA) such that V ΠA = ΠBA.
KA ∋ x 7→ Ax ∈ BA,
7. Uniqueness of Induced Kreın Spaces
The two examples of induced Kreın spaces described in Example 6.1 and Example 6.2
turned out to be unitary equivalent. However, in general, not all possible Kreın spaces
induced by a fixed selfadjoint operator are unitary equivalent. We denote by ρ(T ) the
resolvent set of the operator T .
14
A. GHEONDEA
Theorem 7.1. Let A be a bounded selfadjoint operator in the Hilbert space H. The
following statements are equivalent:
(i) The Kreın space induced by A is unique, modulo unitary equivalence.
(ii) There exists ǫ > 0 such that either (0, ǫ) ⊂ ρ(A) or (−ǫ, 0) ⊂ ρ(A).
(iii) For some (equivalently, for any) Kreın space {K, Π} induced by A, the range of Π
contains a maximal uniformly definite subspace of K.
The equivalence of (ii) and (iii) is a simple matter of spectral theory for bounded
selfadjoint operators. The implication (iii)⇒(i) comes from the fact that any densely
defined isometric operator whose domain contains a maximal uniformly definite subspace
is bounded. So only the idea of the implication (i)⇒(ii) is clarified.
Assuming that the statement (ii) does not hold, there exists a decreasing sequence of
numbers (µn)n≥1 with µn ∈ σ(A) and 0 < µn < 1 for all n ≥ 1, such that µn → 0
(n → ∞), and there exists a decreasing sequence of numbers (νn)n≥1 with −νn ∈ σ(−A)
and 0 < νn < 1 for all n ≥ 1, such that νn → 0 (n → ∞). Then, letting µ0 = ν0 = 1,
there exist sequences of orthonormal vectors {en}n≥1 and {fn}n≥1 such that
(7.1)
en ∈ E((µn, µn−1])H,
fn ∈ E([−νn−1, −νn))H, n ≥ 1.
As a consequence, one also gets
(7.2)
[Aei, fj] = 0,
i, j ≥ 1.
Define the sequence (λn)n≥1 by
λn = max{p1 − µ2
n,p1 − ν2
n}.
Then 0 < λn ≤ 1, λn ↑ 1 (n → ∞).
Consider the subspace Sn of the Kreın space KA, defined by
Sn = Cen +Cfn, n ≥ 1,
and then define the operators Un ∈ L(Sn),
Un =
1
p1 − λ2
n (cid:20) 1 −λn
λn −1 (cid:21) , n ≥ 1.
The operators Un are isometric in Sn. Further, one defines the linear manifold D0 in KA
by
and note that the closure of D0 = Wk≥1
By construction, the linear manifold
D0 = [k≥1
Sk
{ek, fk} is a regular subspace in KA.
D = D+ +D− = Ran(Π)
is dense in KA, where A = A+ −A− is the Jordan decomposition of A and D± = Ran(A±).
Also, D0 ⊆ D and the following decomposition holds
D = D0 +(D ∩ D⊥
0 ).
Then define a linear operator U in KA, with domain D0 and the same range, by USn = Un,
n ≥ 1, and U(D ∩ D⊥
0 ). The operator U is isometric, it has dense range
0 ) = I(D ∩ D⊥
REPRODUCING KERNEL KREIN SPACES
15
as well as dense domain, and it is unbounded since it maps uniformly definite subspaces
into subspaces that are not uniformly definite.
(KA, Π) is a Kreın space induced by A. Indeed, ΠH = UΠAH ⊇ D, and the latter is
dense in KA. Further,
[Πx, Πy] = [UΠAx, UΠAy] = [ΠAx, ΠAy] = [Ax, y],
x, y ∈ H.
Since ΠA is bounded it follows that Π is closed. By the Closed Graph Principle, it follows
that Π is bounded. Since U is unbounded it follows that (KA, ΠA) is not unitary equivalent
with (KA, Π). Thus, a contradiction with the assertion (i) is obtained.
Let Kk be two Kreın spaces continuously embedded into the Kreın space H and denote
by ιk : Kk → H the corresponding embedding operators, that is ιkh = h, h ∈ Kk, k = 1, 2.
We say that the Kreın spaces K1 and K2 correspond to the same selfadjoint operator A in
H if ι1ι♯
2 = A. If A is nonnegative, equivalently, the Kreın spaces Kk are actually
Hilbert spaces, this implies that K1 = K2. As a consequence of Theorem 7.1 the following
corollary is obtained.
1 = ι2ι♯
Corollary 7.2. Given a selfadjoint operator A ∈ L(H), the following statements are
mutually equivalent:
(a) There is a unique Kreın space K continuously embedded in H and associated to A.
(b) There exists ǫ > 0 such that either (−ǫ, 0) ⊂ ρ(A) or (0, ǫ) ⊂ ρ(A).
(c) There exists a Kreın space K continuously embedded in H, ι : K → H such that
ιι♯ = A and Ran(ι♯) contains a maximal uniformly definite subspace of K.
Indeed, let Ki be two Kreın spaces continuously embedded in H and let ιi : K → H
be the embedding operators, i = 1, 2. Assume that the induced Kreın spaces (Ki, ι♯
i)
are unitary equivalent, that is, there exists a unitary operator U ∈ L(K1, K2) such that
Uι2 = ι1. Then ι1 = ι2U ♯ and taking into account that ιi are embeddings, it follows that
U ♯x = ι2U ♯x = ι1x = x,
x ∈ K2,
hence K1 = K2 and ι1 = ι2. This shows that the two Kreın spaces coincide. We can now
apply Theorem 7.1 and get the equivalence of the statements (a), (b), and (c).
8. Existence of Reproducing Kernel Kreın Spaces
The next theorem clarifies the problem of existence of reproducing kernel Kreın spaces
associated to Hermitian H-kernels. Notation is as in Section 3. In addition, two positive
semidefinite H-kernels are called independent if for any P ∈ K+(H) such that P ≤ H, K,
it follows P = 0.
Theorem 8.1. Let H ∈ Kh(H). The following assertions are equivalent:
(1) There exists L ∈ K+(H) such that −L ≤ H ≤ L.
(1)′ There exists L ∈ K+(H) such that
[f, g]H ≤ [f, f ]1/2
L [g, g]1/2
L ,
f, g ∈ F0(H).
(2) H = H1 − H2 with H1, H2 ∈ K+(H).
(2)′ H = H+ − H− with H± ∈ K+(H) independent.
(3) There exists a Kolmogorov decomposition (V ; K) of H.
(4) There exists a Kreın space with reproducing kernel H.
16
A. GHEONDEA
Note that letting f = g in (1)′ one obtains (1). Conversely, let L ∈ K+(H) be such
that −L ≤ H ≤ L, that is [f, f ]H ≤ [f, f ]L, for all f ∈ F0(H). Let f, g ∈ F0(H).
Since H is Hermitian, one gets 4 Re[f, g]H = [f + g, f + g]H − [f − g, f − g]H and hence
4 Re[f, g]H ≤ [f + g, f + g]L + [f − g, f − g]L = 2[f, f ]L + 2[g, g]L. Let λ ∈ C be chosen
such that λ = 1 and Re[f, λg] = [f, λg]. Then
(8.1)
[f, λg]H ≤
1
2
[f, f ]L +
1
2
[g, g]L.
We distinguish two possible cases. First, assume that either [f, f ]L = 0 or [g, g]L = 0. To
make a choice assume [f, f ]L = 0. Consider the inequality (8.1) with g replaced by tg for
t > 0. Then [f, g]H ≤ t
2 [g, g]H. Letting t → 0 one gets [f, g]H = 0.
Second case, assuming that both [f, f ]L and [g, g]L are nontrivial, in (8.1) replace f by
L . Thus, the assertions (1)
g to get [f, g]H ≤ [f, f ]1/2
L [g, g]1/2
L
[f, f ]−1/2
and (1)′ are equivalent.
f and g by [g, g]−1/2
L
In order to describe the idea of the proof of the implication (1)′ ⇒(2)′, let KL be the
quotient-completion of (F0(H), [·, ·]L) to a Hilbert space. More precisely, letting NL =
{f ∈ F0 [f, f ]L = 0} denote the isotropic subspace of the positive semidefinite inner
product space (F0(H), [·, ·]L), one considers the quotient F0(H)/NL and complete it to a
Hilbert space KL. The inequality (1)′ implies that the isotropic subspace NL is contained
into the isotropic subspace NH of the inner product (F0, [·, ·]H). Therefore, [·, ·]H uniquely
induces an inner product on KL, also denoted by [·, ·]L such that the inequality in (1)′
still holds for all f, g ∈ KL. By the Riesz Representation Theorem one gets a selfadjoint
and contractive operator A ∈ L(KL), such that
(8.2)
[f, g]H = [Af, g]L,
f, g ∈ KL.
Let A = A+ −A− be the Jordan decomposition of A in KL. Then A± are also contractions
and hence
(8.3)
[A±f, f ]L ≤ [f, f ]L,
f ∈ KL.
It can be proven that the nonnegative inner products [A±·, ·] uniquely induce kernels
H± ∈ K+(H) such that [f, f ]H± ≤ [f, f ]L,
f ∈ F0(H), and H = H+ − H−.
Indeed, the inner product [A+·, ·] restricted to F0(H)/NH can be extended to an inner
product [·, ·]+ on F0(H) by letting it be null onto NL and hence
(8.4)
[f, f ]+ ≤ [f, f ]L,
f ∈ F0(H).
Let x, y ∈ X be arbitrary and x 6= y. Clearly, one can identify the Kreın space
Hx[+]Hy with the subspace of all H-fields f ∈ F0(H) such that supp f ⊆ {i, j}. With
this identification, one considers the restrictions of the inner products [·, ·]+ and [·, ·]L
to Hx[+]Hy. The inner product [·, ·]L is jointly continuous with respect to the strong
topology of Hx[+]Hy. By (8.4) and the equivalence of (1) and (1)′ one concludes that
the inner product [·, ·]+ is also jointly continuous with respect to the strong topology
of Hx[+]Hy and hence, by the Riesz Representation Theorem, there exists a selfadjoint
operator S ∈ L(Hx[+]Hy) such that
[f, g]+ = [Sf, g]Hx[+]Hy,
f, g ∈ Hx[+]Hy.
Define H+(x, y) = PHxSHy, H+(x, x) = PHxSHx, H+(j, j) = PHySHy and H+(j, i) =
PHy SHx = H+(x, y)♯.
REPRODUCING KERNEL KREIN SPACES
17
In this way one obtains a kernel H+ ∈ Kh(H) such that H+ ≤ L and
[f, g]H+ = [f, g]+,
f, g ∈ F0(H).
Since the inner product [·, ·]+ is nonnegative it follows that H+ ∈ K+(H).
Similarly one constructs the kernel H− ∈ K+(H) such that H− ≤ L and
[f, g]H− = [f, g]−,
f, g ∈ F0(H),
where the inner product [f, g]− is the extension of the restriction of the inner product
[A−f, g] to F0(H)/NL, by letting it be null onto NH.
From A = A+−A−, (8.2) and the constructions of the kernels H+ and H− one concludes
that H = H+ − H−.
Let P ∈ K+(H) be such that P ≤ H±. Then
(8.5)
[f, f ]P ≤ [f, f ]L,
f ∈ F0(H).
As before, [·, ·]P induces a nonnegative inner product [·, ·]P on KL such that (8.3) holds
for all f ∈ KL. From P ≤ H± one concludes that
[f, f ]P ≤ [A±f, f ]L,
f ∈ KL,
and, since A+A− = 0 this implies [f, f ]P = 0 for all f ∈ KL. Since by (8.5) one gets
NL ⊆ NP this implies that the inner product [·, ·]P is null onto the whole F0(H) and
hence P = 0.
The implications (2)′ ⇒(2) and (2)⇒(1) are clear.
The most interesting implication is (1)′ ⇒(3).
In a fashion similar to (1)′ ⇒(2)′,
one considers the quotient-completion Hilbert space KL, the representation (8.2) and the
Jordan decomposition A = A+ − A−. The latter yields in a canonical way a Kreın space
(K, [·, ·]H). We again consider NL and NH, the isotropic spaces of the inner product
spaces (F0(H), [·, ·]L) and, respectively, (F0(H), [·, ·]H). From the inequality (1)′ one gets
NL ⊆ NH.
For every x ∈ X and every vector h ∈ Hx one considers the function h ∈ F0(H) defined
by
(8.6)
h(y) =(cid:26) h,
0,
y = x,
y 6= x.
This identification of vectors with functions in F0(H) yields a natural embedding Hx ֒→
F0(H). With this embedding one defines linear operators V (x) : Hx → K by
V (x)h = h + NH ∈ F0(H)/NH ⊆ K,
h ∈ Hx.
It follows that the linear operators V (x) are bounded, for all x ∈ X, and that (K; V ) is a
minimal linearisation of the H-kernel H.
The idea of the proof of (3)⇒(1) deserves an explanation as well. Let (K, [·, ·]) be a
Kreın space and {V (x)}x∈X be a family of bounded linear operators V (x) ∈ L(Hx, K),
x ∈ X, such that
H(x, y) = V (y)♯V (x),
x, y ∈ X.
Fix on K a fundamental symmetry J and for each x ∈ X fix a fundamental symmetry Jx
on Hx. Then, defining the kernel L by
L(x, y) = JyV (y)∗V (x),
x, y ∈ X,
18
A. GHEONDEA
it can be proven that L ∈ K+(H) and that −L ≤ H ≤ L.
Example 8.2. (1) Let E be a reflexive real Banach space which is not a Hilbert space.
For example, one can take 1 < p < 2 and E = ℓp
R. Let E ′ denote its topological dual space.
On X = E × E ′ consider the Hermitian form
(8.7)
H((e, ϕ), (f, ψ)) = ϕ(f ) + ψ(e),
e, f ∈ E, ϕ, ψ ∈ E ′.
The Hermitian form H can be viewed as a Hermitian scalar kernel on X × X and it can
be proven that it cannot be written as a difference of two positive semidefinite scalar
kernels. Briefly, the idea is that X is a selfdual Banach space when given the norm
k(e, ψ)k2
E ′, and H is jointly continuous with respect to this norm, but the
topological inner product space (X; H; k · kX) is not decomposable.
X = kek2
E + kψk2
9. Uniqueness of Reproducing Kernel Kreın Spaces
Quasi semidefinite Hermitian kernels are associated with reproducing kernel Kreın
spaces which are unique, equivalently, they have linearisations having the uniqueness
property modulo unitary equivalence. In the general Hermitian case, the existence of a
reproducing kernel Kreın space does not imply that it is unique, equivalently, the exis-
tence of a linearisation does not imply its uniqueness, modulo unitary equivalence. Recall
that, two linearisations (K; V ) and (H, U) of the same H-kernel H are unitary equiva-
lent if there exists a unitary operator Φ ∈ L(K, H) such that for all x ∈ X one gets
U(x) = ΦV (x).
Let H be an H-kernel. If L ∈ K+(H) is such that −L ≤ H ≤ L then one denotes by
KL the quotient completion of (F0(H), [·, ·]L) to a Hilbert space and by A = AL ∈ L(KL)
the Gram operator of the inner product [·, ·]H with respect to the positive semidefinite
inner product [·, ·]L, that is, [h, k]H = [ALh, k]L for all h, k ∈ KL.
Theorem 9.1. Let H be an H-kernel on a set X which has a minimal linearisation,
equivalently, it is associated to a reproducing kernel of a reproducing kernel Kreın space
on X. The following assertions are equivalent:
(i) The H-kernel H has unique minimal linearisation, modulo unitary equivalence.
(ii) For any (equivalently, there exists a) positive semidefinite H-kernel L such that
−L ≤ H ≤ L there exists ǫ > 0 such that either (0, ǫ) ⊂ ρ(AL) or (−ǫ, 0) ⊂ ρ(AL).
(iii) H has a minimal linearisation (equivalently, any minimal linearisation) (K; V ) that
has fundamental decomposition K = K+[+]K− such that either K+ or K− is contained in
the linear manifold generated by V (x)Hx, x ∈ X.
(iv) The reproducing kernel Kreın space with reproducing kernel H is unique.
In order to explain the implication (i)⇒(ii), assume that there exists a positive semi-
definite H-kernel L such that −L ≤ H ≤ L and for any ǫ > 0 one gets (0, ǫ) ∩ σ(AL) 6= ∅
and (−ǫ, 0) ∩ σ(AL) 6= ∅. From Theorem 7.1 it follows that there exists two Kreın spaces
(K, Π) and (H, Φ) induced by the same selfadjoint operator AL, which are not unitary
equivalent. It is easy to see that the operator Ψ : Ran(Π) → Ran(Φ) defined by
ΨΠf = Φf,
f ∈ KL,
REPRODUCING KERNEL KREIN SPACES
19
is isometric, densely defined with dense range, and it is unbounded due to the non-
unitary equivalence of the two induced Kreın spaces. As a consequence, Ψ is closable and
its closure, denoted also by Ψ, shares the same properties.
Let (V ; K) be the minimal linearisation of H defined as in the proof of Theorem 8.1,
(1)′ ⇒(3). Define a new minimal linearisation (H; U) and prove that it is not uni-
tary equivalent with (V ; H). More precisely, let U(x) = ΨV (x) for all x ∈ X. Since
Ran(V (x)) ⊆ D(Ψ) and Ψ is closed it follows, via the Closed Graph Principle, that
U(x) ∈ L(Hx, K) for all x ∈ X.
Let x, y ∈ X be arbitrary and fix vectors h ∈ Hx and k ∈ Hy. Then
[U(y)k, U(x)h] = [ΨV (y)k, ΨV (x)h] = [V (y)k, V (x)h] = [H(x, y)k, h],
and hence U ♯
xUy = H(x, y). Also,
_j∈X
U(y)Hy = _j∈X
ΨV (y)Hy = Clos Ran(Φ) = H.
Thus, (H; U) is a minimal linearisation of the H-kernel H. On the other hand, since the
operator Ψ is unbounded it follows that the two minimal linearisations (K; V ) and (H; U)
are not unitary equivalent.
For the implication (ii)⇒(i), let (K; V ) and (H; U) be two minimal linearisations of H.
Let J and Jx be fundamental symmetries on K and, respectively, Hx, x ∈ X. We consider
the positive definite H-kernel LV defined by
LV (x, y) = JyV (y)∗V (x),
x, y ∈ X,
and as in the proof of Theorem 8.1 it follows that −LV ≤ H ≤ LV . We define a linear
operator ΠV : F0(H) → K by
(9.1)
V (x)hx,
h = (hx)x∈X ∈ F0(H).
ΠV (h) =Xx∈X
Taking into account of the axiom (kd2) in the definition of a minimal linearisation one
obtains
(9.2)
[ΠV h, ΠV k]H = [h, k]K,
h, k ∈ F0(H),
that is, the operator ΠV is isometric from (F0(H), [·, ·]H) into (K, [·, ·]K). In addition, ΠV
is also isometric when considered as a linear operator from (F0, [·, ·]LV ).
Similarly, considering the positive semidefinite H-kernel LU defined by
LU (x, y) = JyU(y)∗U(x),
x, y ∈ X,
one gets −LU ≤ H ≤ LU and, defining the linear operator ΠU : F0(H) → H by
(9.3)
one obtains
(9.4)
ΠU (h) =Xx∈X
U(x)hx,
h = (hx)x∈X ∈ F0(H),
[ΠU h, ΠU k]H = [h, k]K,
h, k ∈ F0(H),
that is, the operator ΠU is isometric from (F0, [·, ·]H) into (K, [·, ·]H) and ΠU is also
isometric when considered as a linear operator from (F0, [·, ·]LU ) into (K, h·, ·iJ).
20
A. GHEONDEA
Let L = LV + LU and clearly −L ≤ H ≤ L. Since LV ≤ L it follows that KL is
contractively embedded into KLV and hence ΠV induces a bounded operator ΠV : AL → K.
From (9.2) it follows
(9.5)
Π∗
V JΠV = AL.
By assumption, there exists ǫ > 0 such that either (−ǫ, 0) ⊂ ρ(AL) or (0, ǫ) ⊂ ρ(AL)
and hence from (9.5) and taking into account that by the minimality axiom (b) of the
minimal linearisation the operator ΠV has dense range, it follows that there exists a
uniquely determined unitary operator ΦV : KAL → K such that
(9.6)
ΦV h = ΠV h,
h ∈ AL,
where KAL is the Kreın space induced by the operator AL.
Similarly, performing the same operations with respect to the linearisation (H; V ) one
gets a uniquely determined unitary operator ΦU : KAL → H such that
(9.7)
ΦU h = ΠU h,
h ∈ AL.
Define the unitary operator Φ : K → H by
Φ = ΦU Φ−1
V .
Taking into account of (9.6) and (9.7), the definition of the operator ΠV as in (9.1), and
the definition of ΠU as in (9.3), it follows that
Φ(Xx∈X
V (x)hx) =Xx∈X
U(x)hx,
(hx)x∈X ∈ F0(H).
This implies readily that for all x ∈ X one has ΦV (x) = U(x) and hence the two Kol-
mogorov decompositions (K; V ) and (H; U) are unitary equivalent.
As a consequence of Theorem 9.1 one can obtain a rather general sufficient condition
of nonuniqueness. Let K and H be two positive semidefinite H-kernel. Then one consid-
ers the Hilbert spaces KK and KH, obtained by quotient completion of (F0, [·, ·]K) and,
respectively, of (F0, [·, ·]H). If H ≥ K then KH is contractively embedded into KK. The
kernel H is K-compact if the embedding of KH into KK is a compact operator.
Corollary 9.2. Let H+, H− ∈ K+(H) be two independent kernels, both of them of infinite
rank. If there exists a kernel K ∈ K+(H) such that H+ and H− are K-compact, then the
minimal linearisations of the kernel H+ − H− are not unique, modulo unitary equivalence.
Let H = H1 − H2. Clearly −K ≤ H ≤ K. Let A± ∈ L(KK) denote the Gram operator
of the kernel H±. Since H+ and H− are independent it follows that A = A+ − A− is
the Gram operator of H. Since H± are of infinite rank and K-compact it follows that
A± are compact operators of infinite rank in L(KK) and hence the spectra σ(A±) are
accumulating to 0. Then the spectrum σ(A) is accumulating to 0 from both sides. This
clearly contradicts the condition (ii) in Theorem 9.1 and hence the kernel H has non-
unique minimal linearisations.
(cid:3)
REPRODUCING KERNEL KREIN SPACES
21
10. Holomorphic Kernels: Single Variable Domains
Let Ω be a domain, a nonempty open subset, in the complex field C, and let H be
a Hilbert space. A kernel K : Ω × Ω → L(H) is called holomorphic if it is holomorphic
in the first variable, that is, for each w ∈ Ω, the map Ω ∋ z 7→ K(z, w) ∈ L(H)
and conjugate holomorphic in the second variable, that is, for each z ∈ Ω, the map
Ω ∋ w 7→ K(z, w) ∈ L(H) is holomorphic. Recall that, for Banach space valued functions
of complex variable, strong holomorphy is the same with weak holomorphy.
If K is
Hermitian, then K is a holomorphic kernel if and only if the map Ω ∋ z 7→ K(z, w) ∈ L(H)
is holomorphic for all w ∈ Ω.
Theorem 10.1. Let H be a Hilbert space and, for some r > 0, let K be a holomorphic
H-kernel on Dr = {z ∈ C z < r}. Then, there exists 0 < r′ ≤ r and a reproducing
kernel Kreın space on Dr′ with reproducing kernel KDr′ × Dr′.
The first step in the proof of this theorem is to observe that, without loss of generality,
one can assume r > 1. Indeed, if r ≤ 1 then, for some 0 < ρ < r small enough, the H-
kernel Kρ(z, w) = K(ρz, ρw) is holomorphic on Dr/ρ. If Kρ is the reproducing kernel Kreın
space with reproducing kernel Kρ restricted to Dr′′, for some 0 < r′′ ≤ r/ρ, let r′ = r′′ρ
and let K denote the vector space of functions f : Dr′ → H such that f (z) = F (z/ρ) for
some F ∈ Kρ and all z ∈ Dr′. On K there is defined the inner product [·, ·]K
f (z) = F (z/ρ), g(z) = G(z/ρ), F, G ∈ Kρ.
[f, g]K = [F, G]Kρ,
Then (K; [·, ·]K) is a Kreın space. For each w ∈ Dr′ the map z 7→ K(z, ρw) belongs to K
and, for each F ∈ Kρ, f (z) = F (z/ρ), and h ∈ H,
[f, K(·, ρw)h]K = [F, K(ρ·, ρw)h]Kρ = [F (ρw), h]H = [f (w), h]H,
hence K is a reproducing kernel Kreın space with reproducing kernel KDr′ × Dr′.
There are two main ideas of the proof. First, the Szego kernel S, see Example 4.3, plays
a distinguished role in holomorphy, and allows us to construct the convolution kernel of
K in the Hardy space H 2(D). The second idea is that, once the convolution operator
represented as a selfadjoint bounded operator on a Hilbert space of functions is defined,
the construction of the induced Kreın space as in Example 6.2 will provide the reproducing
kernel Kreın space with reproducing kernel K. Here are a few details.
Letting r > 1, there is considered the Hardy space H 2(D) ⊗ H, identified with a space
n=0 anzn, where (an)n≥0 is a sequence of vectors in H such
H < ∞. Also, the inner product on the Hilbert space H 2(D) ⊗ H
Thus, on H 2(D) ⊗ H one can define the analog of the convolution operator K
hf (eit), g(eit)iHdt,
f, g ∈ H 2(D) ⊗ H.
is
n≥0 kank2
of H-valued functions f (z) =P∞
that kf k2 =P∞
2π Z 2π
2π Z 2π
(Kf )(z) =
hf, gi =
(10.1)
0
1
1
0
K(z, eit)f (eit)dt,
f ∈ H 2(D) ⊗ H.
Letting M = supz,w≤1 kK(z, w)k < ∞, it follows that kKf k ≤ Mkf kH 2(D)⊗H, hence the
convolution operator K is a bounded linear operator in H 2(D) ⊗ H. On the other hand,
hKf, giH 2(D)⊗H =
hK(eit, eis)f (eis), g(eit)iHdtds,
1
4π2 Z 2π
0 Z 2π
0
22
A. GHEONDEA
hence K is selfadjoint. Since, for any w ∈ D1 and any h ∈ H, the function fw(z) =
h/(1 − wz) belongs to H 2(D) ⊗ H and by the Cauchy formula, (Kfw)(z) = K(z, w)h, the
range of the convolution operator K contains all the functions Kw(·)h, for w ∈ D and
h ∈ H.
Then, one can use the construction of the induced Kreın space (BK; ΠBK ) inside of
H 2(D) ⊗ H, as in Example 6.2 but applied to the bounded selfadjoint operator K in the
Hilbert space H 2(D) ⊗ H, in order to get the reproducing kernel Kreın space K with
reproducing kernel K.
11. Holomorphic Kernels: Several Variables Domains
In this section it is considered the analog of Theorem 10.1 in case the Hermitian kernels
are defined on domains in CN for N ≥ 2. We use the notation as in Example 4.4 where the
Drury-Arveson space was constructed as the reproducing kernel Hilbert space associated
to the Szego kernel. For simplicity, it is considered only scalar valued kernels.
Recall that a scalar-valued Hermitian kernel K, defined on a nonempty open subset O
of G = CN , is holomorphic on O if K(·, η) is holomorphic on O for each fixed η ∈ O.
Since K is Hermitian, it follows that K is conjugate holomorphic in the second variable.
Theorem 11.1. Let r > 0 and let K be a Hermitian holomorphic kernel on the open ball
Br in CN . Then, there exist 0 < r′ ≤ r and a reproducing kernel Kreın space K on Br′
with reproducing kernel KBr′ × Br′.
To a certain extent, the proof of this theorem follows a pattern similar to that of the
proof of Theorem 10.1, namely, first a scaling argument can be used in order to reduce the
proof to the case r > 1, then the convolution kernel of K can defined on the Drury-Arveson
space H 2(B1) associated to the Szego kernel S, see Example 4.4, and it can be proven
that this convolution operator is bounded and selfadjoint. Finally, the construction of
type (BA; ΠBA), see Example 6.2, can be used in order to produce a reproducing kernel
space K with reproducing kernel K. This proof shows, once again, a certain universality
property of the Szego kernel with respect to holomorphic Hermitian kernels.
Let K be a scalar Hermitian holomorphic kernel on Br, r > 0. Since K is Hermitian,
it follows that ζ 7→ K(ξ, ζ) is holomorphic on Br for each ξ ∈ Br, that is, letting {ej}N
denote the canonical orthonormal basis of G = CN , the conjugation can be defined by
j=1
ξ =
NXj=1
hξ, ejiej →
NXj=1
hξ, ejiej = ξ,
so that the function f (ξ, η) = K(ξ, η) is separately holomorphic on Br × Br. By Hartogs'
Theorem f is holomorphic on Br×Br, hence f is locally bounded. Similar to the argument
provided for Theorem 10.1, without loss of generality one can suppose that r > 1. Hence
there exist 1 < ρ < r and C > 0 such that:
(11.1)
and
(11.2)
K(ξ, η) ≤ C for all
ξ, η ∈ Bρ
K(ξ, η) = Xm≥0
pm(ξ, η)
REPRODUCING KERNEL KREIN SPACES
23
uniformly on Bρ, where each pm, m ≥ 0, is an m-homogeneous complex polynomial on
2N variables. There exists a continuous linear functional Am on Pm(G × G)⊗m, see (4.7),
such that
(11.3)
pm(ξ, η) = Am((ξ, η)⊗m),
for all ξ, η ∈ CN .
Using Cauchy Inequalities, for Bρ, one gets
(11.4)
hence
(11.5)
kAmk ≤ C
1
ρm ,
Xm≥0
kAmk2 ≤ CXm≥0
1
ρ2m = C
1
1 − 1/ρ2 = C ′ < ∞.
By the Riesz Representation Theorem, there exist am ∈ Pm(G × G)⊗m, m ≥ 0, such that
(11.6)
and
(11.7)
Am((ξ, η)⊗m) = h(ξ, η)⊗m, ami(G×G)⊗m,
kamk = kAmk,
(with a0 = A0 ∈ C). Since Pm(G × G)⊗m is isometrically isomorphic to (PmG⊗m)⊕(m+1),
it is deduced that there are ak
m ∈ PmG⊗m, k = 0, . . . , m, such that
(11.8)
and
(11.9)
h(ξ, η)⊗m, ami(G×G)⊗m =
miG⊗m,
mXk=0
hbk
m(ξ, η), ak
mXk=0
kak
mk2 = kamk2,
where b0
(11.6), and (11.8),
0 = 1 and bk
m(ξ, η) = ξ⊗(m−k) ⊗ η⊗k, m ≥ 1, k = 0, . . . , m. By (11.2), (11.3),
K(ξ, η) = Xm≥0
mXk=0
hbk
m(ξ, η), ak
mi =Xk≥0Xm≥k
hbk
m(ξ, η), ak
mi,
where the series converge absolutely on η by (11.4).
Using all these, it can be shown that Kη ∈ H 2(B1) for all η ∈ B1, where Kη(ξ) = K(ξ, η).
Then letting Kaη = Kη, η ∈ B1, one gets a bounded linear operator in H 2(B1) such that
K(ξ, η) = hKaξ, aηiH 2(B1).
This operator K is selfadjoint and it is the analog of the convolution operator for which,
applying the construction of type (BA; ΠBA) as in Example 6.2, one gets a reproducing
kernel Kreın space K with reproducing kernel KB1 × B1.
24
A. GHEONDEA
12. Comments
The theory of reproducing kernel Hilbert spaces and their positive semidefinite kernels
originates with the works of S. Zaremba [44], G. Szego [42], S. Bergman [7], and S. Bochner
[8]. E.H. Moore [33] also contributed significantly to this theory, but the first systemati-
sation and abstract presentation belongs to N. Aronszajn [4]. A different but equivalent
theory belongs to L. Schwartz [40], whose work remained almost unnoticed for a long time,
although it was the first to consider reproducing kernel Kreın spaces (Hermitian spaces,
as called there). So far, monographs on this subject have been written by T. Ando [3],
S. Saitoh [38], and the forthcoming title of S. Saitoh and Y. Sawano [39], which are good
sources for the large area of applications of the technique of reproducing kernel spaces in
complex functions theory, ordinary and partial differential equations, integral equations,
and approximation and numerical analysis.
So far, there are two monographs devoted to indefinite inner product spaces and their
linear operators, J. Bognar [9] and T.Ya. Azizov and I.S. Iokhvidov [6], where proofs,
examples, and counter-examples of the facts recalled in Sections:ks can be found.
The introductory material on Hermitian kernels in Section 3 follows T. Constantinescu
and A. Gheondea [14]. Theorem 3.1 is classical. The concept of linearisation originates
with J. Mercer [32], for the scalar case, and A.N. Kolmogorov [25], for the operator valued
case.
The Hardy Space H 2(D) originates with the G. Szego Kernel [42]. For the theory of
Hardy spaces there are monographs of P.L. Duren [19] and P. Koosis [26]. Bergman kernel
is also another important example of a positive semidefinite kernel but it falls out of this
chapter concern, see P.L. Duren and B. Schuster [20]. The Drury-Arveson space originates
with S.W. Drury [21] and W.B. Arveson [5]. The holomorphic kernels considered at
Example 4.5 make the main object of investigation of the monograph of D. Alpay et al.
[2]. The investigations of A.V. Potapov [36], L. de Branges [10]–[12], M.G. Kreın and
H. Langer [27]–[31], and H. Dym [22] highly motivate the interest for Hermitian kernels
with or without finite negative signatures. The study of Toplitz type kernels is related to
the investigations on operator dilations of M.A. Naımark [34] and B. Sz.-Nagy [43]. Our
short presentation as in Example 4.6 follows [14]. We only mention that there is a more
general and powerful theory of kernels invariant under actions of ∗-semigroups presented
in [15], motivated by problems in mathematical physics as in D.E. Evans and J.T. Lewis
[23] and K.R. Parthasaraty, K. Schmidt [35], and many others.
Theorem 5.1 essentially belongs to P. Sorjonen [41] but this result follows from the
more general theory of L. Schwartz [40] that has been obtained about ten years before.
Theorem 5.2 can be found in D. Alpay et al. [2].
Examples 6.1, 6.2, and Theorem 7.1 on induced Kreın spaces can be found in [14], while
Corollary 7.2 belongs to T. Hara [24]. Similar and, in a certain way, equivalent uniqueness
conditions, can be found in [13] and M.A. Dritschel [18]. B. ´Curgus and H. Langer [17]
proves that once non-equivalent induced Kreın spaces exist, there are infinitely many.
The characterisations of existence of reproducing Kreın spaces associated to given Her-
mitian kernels as in Theorem 8.1 belong essentially to L. Schwartz [40], but our presen-
tation follows [14]. Example 8.2 is from [40] as well, cf. [2].
The uniqueness Theorem 9.1 is from [14] while Corollary 9.2 is from [40].
REPRODUCING KERNEL KREIN SPACES
25
The result in Theorem 10.1 on single variable holomorphic kernels belongs to D. Alpay
[1], while its several variables generalisation in Theorem 11.1 is from [16]: for the basics
of several complex variables holomorphic functions facts used during the explanation of
the ideas of the proof, see R.M. Range [37].
References
[1] D. Alpay, Some remarks on reproducing kernel Kreın spaces, Rocky Mountain J., 21(1991), 1189–
1205.
[2] D. Alpay, A. Dijksma, J. Rovnyak, H. de Snoo, Schur Functions, Operator Colligations, and
Reproducing Kernel Pontryagin Spaces, Birkhauser Verlag, Basel 1997.
[3] T. Ando, Reproducing Kernel Spaces and Quadratic Inequalities, Lecture Notes, Hokkaido Univer-
sity, Research Institute of Applied Electricity, Division of Applied Mathematics, Sapporo, Japan,
1987.
[4] N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc. 68(2950), 337-404.
[5] W.B. Arveson, Subalgebras of C ∗-algebras. III: Multivariable operator theory, Acta Math.
181(1998), 159–228.
[6] T.Ya. Azizov, I.S. Iokhvidov, Linear Operators in Spaces with an Indefinite Metric, Pure and
Applied Mathematics, John Wiley & Sons, Ltd., Chichester 1989.
[7] S. Bergman, Uber die Entwicklung der harmonischen Funktionen der Ebene und des Raumes nach
Orthogonalfunktionen, Math. Ann. 86(1922), 238–271.
[8] S. Bochner, Uber orthogonale Systeme analytischer Funktionen, Math. Zeitschr. 14(1922), 180–
207.
[9] J. Bognar: Indefinite Inner Product Spaces, Springer-Verlag, Berlin 1976.
[10] L. de Branges: Complementation theory in Kreın spaces, Trans. Amer. Math. Soc., 305(1988),
277–291.
[11] L. de Branges, Krein spaces of analytic functions, J. Functional Analysis, 81(1988), 219–259.
[12] L. de Branges, A construction of Krein spaces of analytic functions, J. Functional Analysis,
98(1991), 1–41.
[13] T. Constantinescu, A. Gheondea, Elementary rotations of linear operators in Kreın spaces, J.
Operator Theory, 29(1993), 167–203.
[14] T. Constantinescu, A. Gheondea, Representations of hermitian kernels by means of Kreın
spaces, Publ. RIMS. Kyoto Univ., 33(1997), 917–951.
[15] T. Constantinescu, A. Gheondea, Representations of Hermitian kernels by means of Kreın
spaces II. Invariant kernels, Commun. Math. Phys. 216(2001), 409–430.
[16] T. Constantinescu, A. Gheondea, On L. Schwartz's boundedness condition for kernels, Posi-
tivity 10(2006), 65–86.
[17] B. ´Curgus, H. Langer, Continuous embeddings, completions and complementation in Krein
spaces, Rad. Mat. 12(2003), 37–79.
[18] M.A. Dritschel, The essential uniqueness property of linear operators in Kreın spaces, J. Funct.
Anal. 118(1993), 198–248.
[19] P.L. Duren, Theory of H
[20] P.L. Duren, A. Schuster, Bergman Spaces, Amer. Math. Soc., Providence R.I. 2004.
[21] S.W. Drury, A generalization of von Neumanns inequality to the complex ball, Proc. Amer. Math.
p Spaces, Academic Press, INC., New York, 1970.
Soc. 68(1978), 300–304.
[22] H. Dym, J-Contractive Matrix Functions, Reproducing Kernel Hilbert Spaces and Interpolation,
CBMS Regional Conference Series in Mathematics, Amer. Math. Soc., Providence R.I. 1989.
[23] D.E. Evans, J.T. Lewis: Dilations of Irreducible Evolutions in Algebraic Quantum Theory, Comm.
Dublin Inst. Adv. Studies Ser. A No. 24, Dublin Institute for Advanced Studies, Dublin, 1977.
[24] T. Hara, Operator inequalities and construction of Kreın spaces, Integral Equations and Operator
Theory, 15(1992), 551–567.
[25] A.N. Kolmogorov: Stationary sequences in Hilbert space, Bull. Math. Univ. Moscow, 2(1941),
1–40.
[26] P. Koosis, Introduction to H
p Spaces, Cambridge Mathematical Press, Cambridge, 1970.
26
A. GHEONDEA
[27] M.G. Krein, H. Langer, Uber einige Fortsetzungsprobleme, die eng mit der Theorie hermitescher
Operatoren im Raume Πκ zusammenhangen, I. Einige Funktionenklassen und ihre Darstellungen,
Math. Nachr. 77(1977), 187–236.
[28] M.G. Krein, H. Langer, Uber einige Fortsetzungsprobleme, die eng mit der Theorie hermitescher
Operatoren im Raume Πκ zusammenhangen, II. Verallgemeinerte Resolventen, u-Resolventen und
ganze Operatoren, J. Funct. Anal. 30(1978), 390–447.
[29] M.G. Krein, H. Langer, On some extension problems which are closely connected with the theory
of hermitian operators in a space Πκ. III: Indefinite analogues of the Hamburger and Stieltjes moment
problem, Part I, Beitr. Ana. 14(1979), 25-40; Part II, Beitr. Anal. 15(1981), 27–45.
[30] M.G. Krein, H. Langer, Some propositions on analytic matrix functions related to the theory of
operators in the space Πκ, Acta. Sci. Math. (Szeged) 1981, 181–205.
[31] M.G. Krein, H. Langer, On some extension problems which are closely connected with the theory
of hermitian operators in a space Πκ, IV. Continuous analogues of orthogonal polynomials on the
unit circle with respect to an indefinite weight and related continuation problems for some classes
of functions, J. Operator Theory 13(1985), 299-417.
[32] J. Mercer, Functions of positive and negative type and their connections with the theory of integral
equations, Philos. Trans. Roy. Soc. London Ser. A 209(1909), 415–446.
[33] E.H. Moore, General Analysis, Mem. Amer. Philos. Soc. I, 1935; II, 1939.
[34] M.A. Naımark: Positive-definite operator functions on a commutative group [Russian], Izvestya
Akad. Nauk SSSR, 1(1943), 234–244.
[35] K.R. Parthasaraty, K. Schmidt: Positive-Definite Kernels, Continous Tensor Products and
Central Limit Theorems of Probability Theory, Lecture Notes in Mathematics, Vol. 272, Springer-
Verlag, Berlin 1972.
[36] V.P. Potapov, The multiplicative structure of J-contractive matrix functions, Trudy Moskov. Mat.
Obshch. 4(1955), 125–236.
[37] R.M. Range, Holomorphic Functions and Integral Representations in Several Complex Variables,
Graduate Texts in Mathematics Vol. 108, Springer-Verlag, Berlin 1986.
[38] S. Saitoh, Theory of Reproducing Kernels and its Applications, Pitman Research Notes in Mathe-
matics Series, vol. 189, Longman Scientific and Technical, Harlow 1988
[39] S. Saitoh, Y. Sawano, The Theory of Reproducing Kernels – 60 Years Since N. Aronszajn, (to
appear)
[40] L. Schwartz, Sous espace Hilbertiens d'espaces vectoriel topologiques et noyaux associ´es (noyaux
reproduisants), J. Analyse Math., 13(1964), 115–256.
[41] P. Sorjonen, Pontrjagin Raume mit einem reproduzierenden Kern, Ann. Acad. Sci. Fenn. Ser. A
I Math. Dissertationes, 594(1973), 1–30.
[42] G. Szego, Uber die Randwerte einer analytischen Funktion, Math. Ann. 84(1921), 232–244.
[43] B. Sz.-Nagy, Prolongement des transformations de l'espace de Hilbert qui sortent de cet espace,
in Appendice au livre "Le¸cons d'analyse fonctionnelle" par F. Riesz et B. Sz.-Nagy, pp. 439-573
Akademiai Kiado, Budapest, 1955.
[44] S. Zaremba, L'´equation biharmonique et une classe remarquable de fonctions fondamentales har-
monique, Krak. Anz. (1907), 147–196.
Department of Mathematics, Bilkent University, 06800 Bilkent, Ankara, Turkey, and
Institutul de Matematica al Academiei Romane, C.P. 1-764, 014700 Bucures¸ti, Romania
E-mail address: [email protected] and [email protected]
|
1005.1585 | 2 | 1005 | 2010-12-29T12:56:59 | A really simple elementary proof of the uniform boundedness theorem | [
"math.FA"
] | I give a proof of the uniform boundedness theorem that is elementary (i.e. does not use any version of the Baire category theorem) and also extremely simple. | math.FA | math | A really simple elementary proof
of the uniform boundedness theorem
Alan D. Sokal∗
Department of Physics
New York University
4 Washington Place
New York, NY 10003 USA
[email protected]
May 7, 2010
revised October 20, 2010
to appear in the American Mathematical Monthly
Abstract
I give a proof of the uniform boundedness theorem that is elementary (i.e.,
does not use any version of the Baire category theorem) and also extremely
simple.
Key Words: Uniform boundedness; gliding hump; sliding hump; Baire category.
Mathematics Sub ject Classification (MSC 2000) codes: 46B99 (Primary);
46B20, 46B28 (Secondary).
0
1
0
2
c
e
D
9
2
]
A
F
.
h
t
a
m
[
2
v
5
8
5
1
.
5
0
0
1
:
v
i
X
r
a
∗Also at Department of Mathematics, University College London, London WC1E 6BT, England.
One of the pillars of functional analysis is the uniform boundedness theorem:
Uniform boundedness theorem. Let F be a family of bounded linear op-
erators from a Banach space X to a normed linear space Y .
If F is pointwise
bounded (i.e., supT ∈F kT xk < ∞ for all x ∈ X ), then F is norm-bounded (i.e.,
supT ∈F kT k < ∞).
The standard textbook proof (e.g., [17, p. 81]), which goes back to Stefan Banach,
Hugo Steinhaus, and Stanis law Saks in 1927 [3], employs the Baire category theorem
or some variant thereof.1 This proof is very simple, but its reliance on the Baire
category theorem makes it not completely elementary.
By contrast, the original proofs given by Hans Hahn [7] and Stefan Banach [2] in
1922 were quite different: they began from the assumption that supT ∈F kT k = ∞ and
used a “gliding hump” (also called “sliding hump”) technique to construct a sequence
(Tn ) in F and a point x ∈ X such that limn→∞ kTnxk = ∞.2 Variants of this proof
were later given by T. H. Hildebrandt [11] and Felix Hausdorff [9, 10].3 These proofs
are elementary, but the details are a bit fiddly.
Here is a real ly simple proof along similar lines:
Lemma. Let T be a bounded linear operator from a normed linear space X to a
normed linear space Y . Then for any x ∈ X and r > 0, we have
sup
x′∈B (x,r)
kT x′k ≥ kT kr ,
(1)
where B (x, r) = {x′ ∈ X : kx′ − xk < r}.
Proof. For ξ ∈ X we have
max(cid:8)kT (x + ξ )k, kT (x − ξ )k(cid:9) ≥ 1
2 (cid:2)kT (x + ξ )k + kT (x − ξ )k(cid:3) ≥ kT ξ k ,
(2)
where the second ≥ uses the triangle inequality in the form kα − β k ≤ kαk + kβ k.
Now take the supremum over ξ ∈ B (0, r). (cid:3)
Proof of the uniform boundedness theorem. Suppose that supT ∈F kT k =
∞, and choose (Tn )∞
n=1 in F such that kTnk ≥ 4n . Then set x0 = 0, and for n ≥ 1 use
the lemma to choose inductively xn ∈ X such that kxn − xn−1k ≤ 3−n and kTnxnk ≥
2
3 3−nkTnk. The sequence (xn ) is Cauchy, hence convergent to some x ∈ X ; and it is
easy to see that kx − xnk ≤ 1
2 3−n and hence kTnxk ≥ 1
6 3−nkTnk ≥ 1
6 (4/3)n → ∞. (cid:3)
1 See [4, p. 319, note 67] concerning credit to Saks.
2 Hahn’s proof is discussed in at least two modern textbooks: see [14, Exercise 1.76, p. 49]
and [13, Exercise 3.15, pp. 71–72].
3 See also [18, pp. 63–64] and [21, pp. 74–75] for an elementary proof that is closely related to
the standard “nested ball” proof of the Baire category theorem; and see [8, Problem 27, pp. 14–15
and 184] and [12] for elementary proofs in the special case of linear functionals on a Hilbert space.
2
Remarks. 1. As just seen, this proof is most conveniently expressed in terms of
a sequence (xn ) that converges to x. This contrasts with the earlier “gliding hump”
proofs, which used a series that sums to x. Of course, sequences and series are
equivalent, so each proof can be expressed in either language; it is a question of taste
which formulation one finds simpler.
2. This proof is extremely wasteful from a quantitative point of view. A quantita-
tively sharp version of the uniform boundedness theorem follows from Ball’s “plank
theorem” [1]: namely, if P∞
n=1 kTnk−1 < ∞, then there exists x ∈ X such that
limn→∞ kTnxk = ∞ (see also [15]).
3. A similar (but slightly more complicated) elementary proof of the uniform
boundedness theorem can be found in [6, p. 83].
4. “Gliding hump” proofs continue to be useful in functional analysis: see [20] for
a detailed survey.
5. The standard Baire category method yields a slightly stronger version of the
uniform boundedness theorem than the one stated here, namely: if supT ∈F kT xk < ∞
for a nonmeager (i.e., second category) set of x ∈ X , then F is norm-bounded.
6. The uniform boundedness theorem has generalizations to suitable classes of
non-normable and even non-metrizable topological vector spaces (see, e.g., [19, pp. 82–
87]). I leave it to others to determine whether any ideas from this proof can be carried
over to these more general settings.
7. More information on the history of the uniform boundedness theorem can be
found in [4, pp. 302, 319n67], [5, pp. 138–142], and [16, pp. 21–22, 40–43].
Acknowledgments. I wish to thank Keith Ball for reminding me of Hahn’s “gliding
hump” proof; it was my attempt to fill in the details of Keith’s sketch that led to
the proof reported here. Keith informs me that versions of this proof have been
independently devised by at least four or five people, though to his knowledge none
of them have bothered to publish it. I also wish to thank David Edmunds and Bob
Megginson for helpful correspondence, and three anonymous referees for valuable
suggestions concerning the exposition. Finally, I wish to thank Jurgen Voigt and
Markus Haase for drawing my attention to the elementary proofs in [6, 18, 21].
This research was supported in part by U.S. National Science Foundation grant
PHY–0424082.
References
[1] K. Ball, The plank problem for symmetric bodies, Invent. Math. 104 (1991)
535–543.
[2] S. Banach, Sur les op´erations dans les ensembles abstraits et leur application
aux ´equations int´egrales, Fund. Math. 3 (1922) 133–181; also in OEuvres avec
des Commentaires , vol. 2, ´Editions scientifiques de Pologne, Warsaw, 1979.
3
[3] S. Banach and H. Steinhaus, Sur le principe de la condensation des singularit´es,
Fund. Math. 9 (1927) 50–61.
[4] G. Birkhoff and E. Kreyszig, The establishment of functional analysis, Historia
Math. 11 (1984) 258–321.
[5] J. Dieudonn´e, History of Functional Analysis , North-Holland, Amsterdam,
1981.
[6] I. Gohberg, S. Goldberg, and M. A. Kaashoek, Basic Classes of Linear Opera-
tors , Birkhauser, Basel, 2003.
[7] H. Hahn, Uber Folgen linearer Operationen, Monatsh. Math. Phys. 32 (1922)
3–88.
[8] P. R. Halmos, A Hilbert Space Problem Book , 2nd ed., Springer-Verlag, New
York, 1982.
[9] F. Hausdorff, Zur Theorie der linearen metrischen Raume, J. Reine Angew.
Math. 167 (1932) 294–311.
[10] J. Hennefeld, A nontopological proof of the uniform boundedness theorem,
Amer. Math. Monthly 87 (1980) 217.
[11] T. H. Hildebrandt, On uniform limitedness of sets of functional operations, Bul l.
Amer. Math. Soc. 29 (1923) 309–315.
[12] S. S. Holland Jr., A Hilbert space proof of the Banach-Steinhaus theorem, Amer.
Math. Monthly 76 (1969) 40–41.
[13] B. D. MacCluer, Elementary Functional Analysis , Springer-Verlag, New York,
2009.
[14] R. E. Megginson, An Introduction to Banach Space Theory , Springer-Verlag,
New York, 1998.
[15] V. Muller and J. Vrsovsk´y, Orbits of linear operators tending to infinity, Rocky
Mountain J. Math. 39 (2009) 219–230.
[16] A. Pietsch, History of Banach Spaces and Linear Operators , Birkhauser,
Boston, 2007.
[17] M. Reed and B. Simon, Methods of Modern Mathematical Physics , vol. I, Func-
tional Analysis , Academic Press, New York, 1972.
[18] F. Riesz and B. Sz.-Nagy, Functional Analysis , Ungar, New York, 1955.
[19] H. H. Schaefer and M. P. Wolff, Topological Vector Spaces , 2nd ed., Springer-
Verlag, New York, 1999.
4
[20] C. Swartz, Infinite Matrices and the Gliding Hump , World Scientific, River
Edge, NJ, 1996.
[21] J. Weidmann, Linear Operators in Hilbert Spaces , Springer-Verlag, New York,
1980.
5
|
1010.0513 | 1 | 1010 | 2010-10-04T09:21:14 | The Range of Localization Operators and Lifting Theorems for Modulation and Bargmann-Fock Spaces | [
"math.FA"
] | We study the range of time-frequency localization operators acting on modulation spaces and prove a lifting theorem. As an application we also characterize the range of Gabor multipliers, and, in the realm of complex analysis, we characterize the range of certain Toeplitz operators on weighted Bargmann-Fock spaces. The main tools are the construction of canonical isomorphisms between modulation spaces of Hilbert-type and a refined version of the spectral invariance of pseudodifferential operators. On the technical level we prove a new class of inequalities for weighted gamma functions. | math.FA | math | THE RANGE OF LOCALIZATION OPERATORS AND
LIFTING THEOREMS FOR MODULATION AND
BARGMANN-FOCK SPACES
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
Abstract. We study the range of time-frequency localization operators acting
on modulation spaces and prove a lifting theorem. As an application we also
characterize the range of Gabor multipliers, and, in the realm of complex analysis,
we characterize the range of certain Toeplitz operators on weighted Bargmann-
Fock spaces. The main tools are the construction of canonical isomorphisms
between modulation spaces of Hilbert-type and a refined version of the spectral
invariance of pseudodifferential operators. On the technical level we prove a new
class of inequalities for weighted gamma functions.
1. Introduction
0
1
0
2
t
c
O
4
]
.
A
F
h
t
a
m
[
1
v
3
1
5
0
.
0
1
0
1
:
v
i
X
r
a
The precise description of the range of a linear operator is usually difficult, if not
impossible, because this amounts to a characterization of which operator equations
are solvable. In this paper we study the range of an important class of pseudodiffer-
ential operators, so-called time-frequency localization operators, and we prove an
isomorphism theorem between modulation spaces with respect to different weights.
The guiding example to develop an intuition for our results is the class of multi-
plication operators. Let m ≥ 0 be a weight function on Rd and define the weighted
m =(cid:16)RRd f (x)pm(x)p dx(cid:17)1/p
space Lp
= kf mkLp. Let Ma
be the multiplication operator defined by Maf = af . Then Lp
m is precisely the
range of the multiplication operator M1/m.
We will prove an similar result for time-frequency localization operators between
weighted modulation spaces. To set up terminology, let π(z)g(t) = e2πiξ·tg(t − x)
denote the time-frequency shift by z = (x, ξ) ∈ R2d acting on a function g on
Rd. The corresponding transform is the short-time Fourier transform of a function
defined by
m(Rd) by the norm kfkLp
Vgf (z) =ZRd
f (t)g(t − x) e2πiξ·t dt = hf, π(z)gi .
The standard function spaces of time-frequency analysis are the modulation spaces.
The modulation space norms measure smoothness in the time-frequency space
(phase space in the language of physics) by imposing a norm on the short-time
Key words and phrases. Localization operator, Toeplitz operator, Bargmann-Fock space, mod-
ulation space, Sjostrand class, spectral invariance, Hermite function.
K. G. was supported in part by the project P2276-N13 of the Austrian Science Foundation
(FWF).
1
2
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
Fourier transform of a function f . As a special case we mention the modulation
spaces M p
m(Rd) for 1 ≤ p ≤ ∞ and a non-negative weight function m. Let
h(t) = 2d/4e−πt2
= 2d/4e−πt·t,
t ∈ Rd
denote the (normalized) Gaussian. Then the modulation space M p
by the norm
m(Rd) is defined
The localization operator Ag
function, and the symbol or multiplier m is defined formally by the integral
m with respect to the "window" g, usually some test
kfkM p
m = kVhfkLp
m .
mf =ZR2d
Ag
m(z)Vgf (z)π(z)g dz .
Localization operators constitute an important class of pseudodifferential operators
and occur under different names such as Toeplitz operators or anti-Wick operators.
They were introduced by Berezin as a form of quantization [2], and are nowadays
applied in mathematical signal processing for time-frequency masking of signals and
for phase-space localization [11]. An equivalent form occurs in complex analysis
as Toeplitz operators on Bargmann-Fock space [3, 4, 7].
In hard analysis they
are used to approximate pseudodifferential operators and in some proofs of the
sharp Garding inequality and the Fefferman-Phong inequality [22, 23, 27]. For the
analysis of localization operators with time-frequency methods we refer to [9] and
the references given there, for a more analytic point of view we recommend [29,30].
A special case of our main result can be formulated as follows. By an isomorphism
between two Banach spaces X and Y we understand a bounded and invertible
operator from X onto Y .
Theorem 1.1. Let m be a non-negative continuous symbol on R2d satisfying m(w+
z) ≤ eawm(z) for w, z ∈ R2d and assume that m is radial in each time-frequency
coordinate. Then for suitable test functions g the localization operator Ag
m is an
isomorphism from the modulation space M p
µ/m(Rd) for all 1 ≤ p ≤ ∞
and moderate weights µ.
µ(Rd) onto M p
We see that the range of a localization operator exhibits the same behavior as the
multiplication operators. For the precise formulation with all assumptions stated
we refer to Section 4. The above isomorphism theorem can also be intepreted as
a lifting theorem, quite in analogy with the lifting property of Besov spaces [32].
However, whereas Besov spaces Bp,q
s with different smoothness s are isomorphic via
Fourier multipliers, the case of modulation spaces is more subtle because in this
case the time-frequency smoothness is parametrized by a weight function on R2d
rather than by a single numbers ∈ R. The lifting operators between modulation
spaces are precisely the localization operators with the weight m.
The isomorphism theorem for localization operators stated above is preceded
by many contributions, which were already listed in [19]. In particular, in [10] it
was shown that for moderate symbol functions with subexponential growth the
localization operator is a Fredholm operator, i.e., it differs from an invertible op-
erator only by a finite-rank operator. Here we show that such operators are even
THE RANGE OF LOCALIZATION OPERATORS
3
isomorphisms between the corresponding modulation spaces, even under weaker
conditions than needed for the Fredholm property. In the companion paper [19]
we proved the isomorphism property for weight functions of polynomial type, i.e.,
c(1 + z)−N ≤ m(z) ≤ C((1 + z)N .
(1)
The contribution of Theorem 1.1 is the extension of the class of possible weights.
In particular, the isomorphism property holds also for subexponential weight func-
tions of the form m(z) = eazb or m(z) = eaz/ log(e+z) for a > 0, 0 < b < 1, which
are often considered in time-frequency analysis. We remark that this generality
comes at the price of imposing the radial symmetry on the weight m. Therefore,
our results are not applicable to all situations covered by [19], where no radial
symmetry is required.
Although the extension to weights of ultra-rapid growth looks like a routine
generalization, it is not. The proof of Theorem 1.1 for weights of polynomial type is
based on a deep theorem of Bony and Chemin [6]. They construct a one-parameter
mt(Rd) for t ∈ R.
group of isomorphisms from L2(Rd) onto the modulation spaces M 2
Unfortunately, the pseudodifferential calculus developed in [6] requires polynomial
growth conditions, and, to our knowledge, an extention to symbols of faster growth
is not available.
On a technical level, our main contribution is the construction of canonical iso-
m(Rd) of Hilbert type. In
morphisms between L2(Rd) and the modulation spaces M 2
fact, such an isomorphism is given by a time-frequency localization operator with
Gaussian window h(t) = 2d/4e−πt2.
Theorem 1.2. Assume that m is a continuous moderate weight function of at
most exponential growth and radial in each time-frequency coordinate. Then the
localization operator Ah
m is an isomorphism from L2(Rd) onto M 2
1/m(Rd).
C−1 ≤Z ∞
0
θ(px/π)
xn
n!
e−xdx Z ∞
0
This theorem replaces the result of Bony and Chemin. The main point is that
Theorem 1.2 also covers symbols of ultra-rapid growth. Its proof requires most of
our efforts. We take a time-frequency approach rather than using classical methods
from pseudodifferential calculus. In the course of its proof we will establish new
inequalities for weighted gamma functions of the form
1
xn
n!
θ(px/π)
e−xdx ≤ C for all n ∈ N ∪ {0} .
The proof method for Theorem 1.1 may be of interest in itself. Once the canonical
isomorphisms are in place (Theorem 1.2), the proof of Theorem 1.1 proceeds as fol-
lows. It is easy to establish that Am is an isomorphism from M 2√m to M 2
1/√m, so that
the composition Ag
m is an isomorphism on M 2√m. Using the canonical isomor-
phisms of Theorem 1.2, one shows next that the operator V = Ah√mAg
1/√m
is an isomorphism on L2 and that the (Weyl) symbol of this operator belongs to
a generalized Sjostrand class. After these technicalities we apply the machinery of
spectral invariance of pseudodifferential operators from [16] to conclude that V is
invertible on all modulation spaces M p
µ (with µ compatible with the conditions on
1/mAg
1/mAg
mAh
4
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
m and the window g). Since V is a composition of three isomorphisms, we then
deduce that Ag
m is an isomorphism from M p
µ onto M p
µ/m.
As an application we prove (i) a new isomorphism theorem for so-called Ga-
bor multipliers, which are a discrete version of time-frequency localization oper-
ators, and (ii) an isomorphism theorem for Toeplitz operators between weighted
Bargmann-Fock spaces of entire functions. To formulate this result more explicitly,
for a non-negative weight function µ and 1 ≤ p ≤ ∞, let F p
µ(Cd) be the space of
entire functions of d complex variables defined by the norm
=ZCd F (z)pµ(z)pe−pπz2/2dz < ∞ ,
kFkp
F p
µ
µ/m(Cd) for every 1 ≤ p, q ≤ ∞ and every moderate weight µ.
and let P be the usual projection from L1
loc(Cd) to entire functions on Cd. The
Toeplitz operator with symbol m acting on a function F is defined to be TmF =
P (mF ). Then we show that
the Toeplitz operator Tm is an isomorphism from
µ (Cd) onto F p,q
F p,q
The paper is organized as follows: In Section 2 we provide the precise definition of
modulation spaces and localization operators, and we collect their basic properties.
In particular, we investigate the Weyl symbol of the composition of two localization
operators. In Section 3, which contains our main contribution, we construct the
canonical isomorphisms and prove Theorem 1.2. In Section 4 we derive a refinement
of the spectral invariance of pseudodifferential operators and prove the general
isomorphism theorem, of which Theorem 1.1 is a special case. Finally in Section 5
we give applications to Gabor multipliers and Toeplitz operators.
2. Time-Frequency Analysis and Localization Operators
We first set up the vocabulary of time-frequency analysis. For the notation we
follow the book [15]. For a point z = (x, ξ) ∈ R2d in phase space the time-frequency
shift of a function f is π(z)f (t) = e2πiξ·tf (t − x), t ∈ Rd.
loc(Rd)
The short-time Fourier transform: Fix a non-zero function g ∈ L1
which is usually taken in a suitable space of Schwartz functions. Then the short-
time Fourier transform of a function or distribution f on Rd is defined to be
Vgf (z) = hf, π(z)gi
(2)
provided the scalar product is well-defined for every z ∈ R2d. Here g is called a
"window function". If f ∈ Lp(Rd) and g ∈ Lp′(Rd) for the conjugate parameter
p′ = p/(p − 1), then the short-time Fourier transform can be written in integral
form as
z ∈ R2d ,
Vgf (z) =ZRd
f (t)g(t − x) e−2πiξ·t dt.
In general, the bracket h·,·i extends the inner product on L2(Rd) to any dual pairing
between a distribution space and its space of test functions, for instance g ∈ S(Rd)
and f ∈ S′(Rd), but time-frequency analysis often needs larger distribution spaces.
THE RANGE OF LOCALIZATION OPERATORS
5
Weight functions: We call a locally bounded, strictly positive weight function
m on R2d moderate, if
z∈R2d(cid:18)m(z + y)
m(z)
sup
m(z) (cid:19) := v(y) < ∞
m(z − y)
,
for all y ∈ R2d .
The resulting function v is a submultiplicative weight function, i.e., v is even and
satisfies v(z1 + z2) ≤ v(z1)v(z2) for all z1, z2 ∈ R2d, and then m satisfies
m(z1 + z2) ≤ v(z1)m(z2)
for all z1, z2 ∈ R2d .
(3)
Given a submultiplicative weight function v on R2d, any weight satisfying the
condition (3) is called v-moderate. For a fixed submultiplicative function v the set
Mv := {m ∈ L∞loc(R2d) : 0 < m(z1 + z2) ≤ v(z1)m(z2) ∀z1, z2 ∈ R2d}
contains all v-moderate weights.
We will use several times that every v-moderate weight m ∈ Mv satisfies the
following bounds:
1
v(z1 − z2) ≤
m(z1)
m(z2) ≤ v(z1 − z2)
for all z1, z2 ∈ R2d .
(4)
This follows from (3) by replacing z1 with z1 − z2.
Modulation spaces M p,q
m for arbitrary weights: For the general definition of
modulation spaces we choose the Gaussian function h(t) = 2d/4e−πt·t as the canoni-
cal window function. Then the short-time Fourier transform is defined for arbitrary
elements in the Gelfand-Shilov space (S1/2
1/2 )′(Rd) of generalized functions. The mod-
ulation space M p,q
1/2 )′(Rd) such
that the norm
m (Rd), 1 ≤ p, q < ∞ consists of all elements f ∈ (S1/2
m =(cid:16)ZRd(cid:16)ZRd Vhf (x, ξ)pm(x, ξ)p dx(cid:17)q/p
If p = ∞ or q = ∞, we make the usual modification and replace the
If m = 1, then we usually write M p,q
is finite.
integral by the supremum norm k · kL∞.
instead of M p,q
m = M p,p
m and M p = M p,p.
m . We also set M p
dξ(cid:17)1/q
= kVhfkLp,q
m
(5)
kfkM p,q
The reader who does not like general distribution spaces, may interprete M p,q
m (Rd)
as the completion of the finite linear combinations of time-frequency shifts H0 =
span{π(z)h : z ∈ R2d} with respect to the M p,q
m -norm for 1 ≤ p, q < ∞ and as
a weak∗-closure, when p = ∞ or q = ∞. These issues arise only for extremely
rapidly decaying weight functions. If m ≥ 1 and 1 ≤ p, q ≤ 2, then M p,q
m (Rd) is in
fact a subspace of L2(Rd). If m is of polynomial type (cf. (1)), then M p,q
m (Rd) is
a subspace of tempered distributions. This is the case that is usually considered,
although the theory of modulation spaces was developed from the beginning to
include arbitrary moderate weight functions [12, 15, 17].
Norm equivalence: Definition (5) uses the Gauss function as the canonical
window. The definition of modulation spaces, however, does not depend on the
6
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
particular choice of the window. More precisely, if g ∈ M 1
then there exist constants A, B > 0 such that
v , g 6= 0, and m ∈ Mv,
AkfkM p,q
m ≤ kVgfkLp,q
m ≤ BkfkM p,q
m = BkVhfkLp,q
m .
We will usually write
(6)
for the equivalent norms.
kVgfkLp,q
m ≍ kfkM p,q
m
Localization operators: Given a non-zero window function g ∈ M 1
symbol or multiplier m on R2d, the localization operator Ag
by
v (Rd) and a
m is defined informally
m(z)Vgf (z)π(z)g dz ,
(7)
mf =ZR2d
Ag
provided the integral exists. A useful alternative definition of Ag
definition
m is the weak
(8)
While in general the symbol m may be a distribution in a modulation space of the
form M∞1/v(R2d) [9,30], we will investigate only localization operators whose symbol
is a moderate weight function.
mf, kiL2(Rd) = hmVgf, VgkiL2(R2d) .
hAg
Taking the short-time Fourier transform of (7), we find that
Vg(Ag
mf )(w) =ZR2d
m(z)Vgf (z)hπ(z)g, π(w)gi dz =(cid:0)(mVgf ) ♮ Vgg(cid:1)(w) ,
with the usual twisted convolution ♮ defined by
Since
(F ♮ G)(w) =ZR2d
F 7→ZR2d
F (z)G(w − z)e2πiz1·(z2−w2) dz.
F (z)hπ(z)g, π(·)gi dz
is the projection from arbitrary tempered distributions on R2d onto functions of
the form Vgf for some distribution f , the localization operator can been seen as
the composition of a multiplication operator and the projection onto the space
of short-time Fourier transforms.
In this light, localization operators resemble
the classical Toeplitz operators, which are multiplication operators following by a
projection onto analytic functions. Therefore they are sometimes called Toeplitz
operators [19,28,30]. If the window g is chosen to be the Gaussian, then this formal
similarity can be made more precise. See Proposition 5.5.
2.1. Mapping Properties of Localization Operators. The mapping proper-
ties of localization operators on modulation spaces resemble closely the mapping
properties of multiplication operators between weighted Lp-spaces. The bound-
edness of localization operators has been investigated on many levels of general-
ity [9, 29, 33]. We will use the following boundedness result from [10, 30].
Lemma 2.1. Let m ∈ Mv and µ ∈ Mw. Fix g ∈ M 1
operator Ag
µm(Rd).
vw(Rd). Then the localization
1/m is bounded from M p,q
µ (Rd) to M p,q
THE RANGE OF LOCALIZATION OPERATORS
7
REMARK: The condition on the window g is required to make sense of Vgf for f
µ (Rd) and of Vgk for k in the dual M p′,q′
µm)′
1/(µm)(Rd) = (M p,q
in the domain space M p,q
of the target space M p,q
µm for the full range of parameters p, q ∈ [1,∞]. For fixed
p, q ∈ [1,∞] weaker conditions may suffice, because the norm equivalence (6) still
v for r ≤ min(p, p′, q, q′) [31].
holds after relaxing the condition g ∈ M 1
v into g ∈ M r
On a special pair of modulation spaces, Ag
1/m is even an isomorphism [19].
Likewise A1/m is an isomorphism from M 2
Lemma 2.2. Let g ∈ M 1
from M 2
θ (Rd) onto M 2
the composition Ag
morphism on M 2
1/mAg
1/θ(Rd).
1/θ(Rd).
v , m ∈ Mv, and set θ = m1/2. Then Ag
1/θ(Rd) onto M 2
m is an isomorphism on M 2
θ (Rd), and Ag
m is an isomorphism
θ (Rd). Consequently
1/m is an iso-
mAg
Lemma 2.2 is based on the equivalence
hAg
mf, fi = hm,Vgf2i = kVgf · θk2
2 ≍ kfk2
M 2
θ
,
(9)
1/mAg
and is proved in detail in [19, Lemma 3.4].
2.2. The Symbol of Ag
m. The composition of localization operators is no
longer a localization operator, but the product of two localization operators still has
a well behaved Weyl symbol. In the following we use the time-frequency calculus
of pseudodifferential operators as developed in [16, 18]. Compared to the standard
pseudodifferential operator calculus it is more restrictive because it is related to the
constant Euclidean geometry on phase space, on the other hand, it is more general
because it works for arbitrary moderate weight functions (excluding exponential
growth).
Given a symbol σ(x, ξ) on Rd × Rd ≃ R2d, the corresponding pseudodifferential
operator in the Weyl calculus Op(σ) is defined formally as
Op(σ)f (x) =ZZR2d
σ(cid:16)x + y
2
, ξ(cid:17)e2πi(x−y)·ξf (y) dydξ
with a suitable interpretation of the integral. If C is a class of symbols, we write
Op(C) = {Op(σ) : σ ∈ C} for the class of all pseudodifferential operators with
symbols in C. For the control of the symbol of composite operators we will use
the following characterization of the generalized Sjostrand class from [16]. For
the formulation associate to a submultiplicative weight v(x, ξ) on R2d the rotated
weight on R4d defined by
v(x, ξ, η, y) = v(−y, η).
(10)
For radial weights, to which we will restrict later, the distinction between v and v
is unnecessary.
Theorem 2.3. Fix a non-zero g ∈ M 1
in M∞,1
H ∈ L1
v . An operator T possesses a Weyl symbol
), if and only if there exists a (semi-continuous) function
, T ∈ Op(M∞,1
v
v(R2d) such that
v
hT π(z)g, π(y)gi ≤ H(y − z)
for all y, z ∈ R2d .
8
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
v
REMARK: This theorem says the symbol class M∞,1
is characterized by the off-
diagonal decay of its kernel with respect to time-frequency shifts. This kernel is
in fact dominated by a convolution kernel. The composition of operators can then
studied with the help of convolution relations. Clearly this is significantly easier
than the standard approaches that work with the Weyl symbol directly and the
twisted product between Weyl symbols. See [20] for results in this direction.
Theorem 2.4. Assume that g ∈ M 1
θ ∈ Mv1/2. Then Ag
1/θ ∈ Op(M∞,1
Proof. We distinguish the window g of the localization operator Ag
m from the win-
dow h used in the expression of the kernel hT π(z)h, π(y)hi. Choose h to be the
Gaussian, then h ∈ M 1
v for every submultiplicative weight v. Let us first write the
kernel hT π(z)h, π(y)hi informally and justify the convergence of the integrals later.
Recall that
vs(Rd), T ∈ Op(M∞,1
) for s ≥ 1/2, and
vs−1/2).
θT Ag
vs
T (Ag
1/θf ) = T(cid:16)ZR2d
θ(u)−1hf, π(u)giπ(u)g du(cid:17)
θT Ag
Then
hAg
=ZZR4d
Now set
1/θπ(z)h, π(y)hi = hT Ag
1/θπ(z)h, Ag
θπ(y)hi
1
θ(u)(cid:10)π(z)h, π(u)g(cid:11)(cid:10)T π(u)g, π(u′)g(cid:11) θ(u′)(cid:10)π(y)h, π(u′)g(cid:11) dudu′ .
G(z) = hg, π(z)hi = Vhg(z) and G∗(z) = G(−z)
(11)
and let H be a dominating function in L1
u). Since time-frequency shifts commute up to a phase factor, we have
v(R2d), so that hT π(u)g, π(u′)gi ≤ H(u′−
Before substituting all estimates into (11), we recall that θ is √v-moderate by
assumption and so (4) says that
hπ(z)h, π(u)gi = G(z − u) .
θ(u′)
θ(u) ≤ v(u′ − u)1/2
for all u, u′ ∈ R2d .
Now by (11) we get
θT Ag
G(z − u)H(u′ − u)G(y − u′) dudu′
θ(u′)
θ(u)
G(z − u)v(u′ − u)1/2H(u′ − u)G(y − u′) dudu′
1/θπ(z)h, π(y)hi
hAg
≤ ZR2dZR2d
≤ ZR2dZR2d
= (cid:16)G ∗ (v1/2H) ∗ G∗(cid:17)(z − y) .
vs(Rd) and thus G ∈ L1
θT Ag
Thus the kernel of Ag
assumption g ∈ M 1
1/θ is dominated by the function G ∗ (v1/2H) ∗ G∗. By
vs ) and thus
vs(R2d), and T ∈ Op(M∞,1
THE RANGE OF LOCALIZATION OPERATORS
9
H ∈ L1
vs(R2d). Then v1/2H ∈ L1
vs−1/2(R2d). Consequently
vs ⊆ L1
The characterization of Theorem 2.3 now implies that Ag
G ∗ (v1/2H) ∗ G∗ ∈ L1
vs−1/2 ∗ L1
vs ∗ L1
vs−1/2 .
θT Ag
(12)
1/θ ∈ Op(M∞,1
vs−1/2).
Corollary 2.5. Assume that g ∈ M 1
submultiplicative weight. Then Ag
1/mAg
v2w(Rd) and m ∈ Mv and w is an arbitrary
m ∈ Op(M∞,1
v w )).
Proof. In this case T is the identity operator and Id ∈ Op(M∞,1
v0 ) for every sub-
multiplicative weight v0(x, ξ, η, y) = v0(η, y). In particular Id ∈ Op(M∞,1
ev2 ). Now
replace the weight θ in Theorem 2.4 by m and the condition θ ∈ Mv1/2 by m ∈ Mv
and modify the convolution inequality (12) in the proof of Theorem 2.4.
3. Canonical Isomorphisms between Modulation Spaces of
Hilbert-Type
In [19] we have used a deep result of Bony and Chemin [6] about the existence of
isomorphisms between modulation spaces of Hilbert type and then extended those
isomorphisms to arbitrary modulation spaces. Unfortunately the result of Bony
and Chemin is restricted to weights of polynomial type and does not cover weights
moderated by superfast growing functions, such as v(z) = eazb for 0 < b < 1.
In this section we construct explicit isomorphisms between L2(Rd) and the modu-
lation spaces M 2
θ (Rd) for a general class of weights. We will assume that the weights
are radial in each time-frequency variable. Precisely, consider time-frequency vari-
ables
which we identify by
(x, ξ) ≃ z = x + iξ ∈ Cd ≃ R2d,
(x1, ξ1; x2, ξ2; . . . ; xd, ξd) = (z1, z2, . . . , zd) ∈ Cd ≃ R2d.
Then the weight function m should satisfy
(13)
(14)
m(z) = m0(z1, . . . ,zd)
for z ∈ R2d
d
+ = [0,∞)d. Without loss of generality, we may also
for some function m0 on R
assume that m is continuous on R2d. (Recall that only weights of polynomial type
occur in the lifting results in [19]. On the other hand, no radial symmetry is needed
in [19].)
For each multi-index α = (α1, . . . , αd) ∈ Nd
variate Hermite function by
0 we denote the corresponding multi-
hα(t) =
dYj=1
hαj (tj), where hn(x) =
21/4πn/2
n!1/2 eπx2 dn
dxn (e−2πx2
)
is the n-th Hermite function in one variable with the normalization khnk2 = 1.
10
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
Then the collection of all Hermite functions hα, α ≥ 0, is an orthonormal basis
of L2(Rd). By identifying R2d with Cd via (13), the short-time Fourier transform
of hα with respect to h(t) = 2d/4e−πt2 is simply
α! (cid:17)1/2
zα e−πz2/2 = e−πix·ξ eα(z) e−πz2
Vhhα(z) = e−πix·ξ(cid:16)πα
for z ∈ Cd . (15)
REMARK: We mention that a formal Hermite expansion f = Pα cαhα defines
a distribution in the Gelfand-Shilov space (S1/2
1/2)′, if and only if the coefficients
satisfy cα = O(eǫα) for every ǫ > 0. The Hermite expansion then converges
1/2)′ × S1/2
in the weak∗ topology. Here we have used the fact that the duality (S1/2
1/2
extends the L2-form h·,·i on S1/2
1/2 , and likewise the duality of the modulation spaces
1/θ(Rd). Consequently, the coefficient cα of a Hermite expansion is
θ (Rd) × M 2
M 2
uniquely determined by cα = hf, hαi for f ∈ (S1/2
In the following we take the existence and convergence of Hermite expansions
for functions and distributions in arbitrary modulation spaces for granted. By
dµ(z) = e−πz2 dz we denote the Gaussian measure on Cd.
Lemma 3.1. Assume that θ(z) = O(eaz) and that θ is radial in each coordinate.
θ (Rd).
(a) Then the monomials zα, α ≥ 0, are orthogonal in L2
(b) The finite linear combinations of the Hermite functions are dense in M 2
By using polar coordinates zj = rjeiϕj , where rj ≥ 0 and ϕj ∈ [0, 2π), we get
1/2 )′. See [21] for details.
θ(Cd, µ).
zα = rαeiα·ϕ
and dz = r1 · · · rd dϕdr
(16)
and the condition on θ in Lemma 3.1 can be recast as
(17)
for some appropriate function θ0 on [0,∞)d, and r = (r1, . . . , rd) and ϕ = (ϕ1, . . . , ϕd)
as usual.
θ(z) = θ0(r),
Proof. (a) This is well-known and is proved in [11,14]. In order to be self-contained,
we recall the arguments. By writing the integral over R2d in polar coordinates in
each time-frequency pair, (16) and (17) give
ZR2d
zαzβθ(z)2 e−πz2
dz =ZZRd
+×[0,2π)d
ei(α−β)·ϕrα+βe−πr2
θ0(r)2r1 · · · rd dϕdr .
The integral over the angles ϕj is zero, unless α = β, whence the orthogonality of
the monomials.
(b) Density: Assume on the contrary that the closed subspace in M 2
by the Hermite functions is a proper subspace of M 2
zero f ∈ (M 2
hα ∈ M 2
0 in (S1/2
θ (Rd) spanned
θ (Rd). Then there exists a non-
1/θ(Rd), such that hf, hαi = 0 for all Hermite functions
0. Consequently the Hermite expansion of f =Pαhf, hαihα =
θ (Rd), α ∈ Nd
1/2 )′, which contradicts the assumption that f 6= 0.
θ (Rd))′ = M 2
THE RANGE OF LOCALIZATION OPERATORS
11
Definition 1. The canonical localization operator Jm is the localization operator
Ah
m associated to the weight m and to the Gaussian window h = h0. Specifically,
Jmf =ZR2d
m(z)hf, π(z)hiπ(z)h dz .
(18)
(19)
For m = θ2 we obtain
hJmf, fi = hmVhf, VhfiR2d = kVhf θk2
2 = kfk2
M 2
θ
whenever f is in a suitable space of test functions.
Our main insight is that localization operators with respect to Gaussian windows
In view of the connection to
and radial symbols have rather special properties.
the localization operators on the Bargmann-Fock space (see below) this is to be
expected.
Theorem 3.2. If θ is a continuous, moderate function and radial in each time-
frequency coordinate, then each of the mappings
Jθ : M 2
J1/θ : M 2
Jθ : L2(Rd) → M 2
θ (Rd) → L2(Rd),
1/θ(Rd) → L2(Rd), J1/θ : L2(Rd) → M 2
1/θ(Rd)
θ (Rd)
is an isomorphism.
The proof is non-trivial and requires a number of preliminary results. In these
investigations we will play with different coefficients of the form
or, more generally,
τα(θ) := hJθhα, hαi,
τα,s(θ) := τα(θs) = hJθshα, hαi =ZR2d
θ(z)s πα
α! zα2e−πz2
dz ,
(20)
when θ is a weight function and s ∈ R. We note that the τα,s(θ) are strictly positive,
since θ is positive. If θ ≡ 1, then τα,s(θ) = 1, so we may consider the coefficients
τα,s(θ) as weighted gamma functions.
Proposition 3.3 (Characterization of M 2
moderate and radial function. Then
θ with Hermite functions). Let θ be a
kfk2
M 2
θ
=Xα≥0
hf, hαi2τα(θ2) .
(21)
Proof. Let f = Pα≥0 cαhα be a finite linear combination of Hermite functions.
Since the short-time Fourier transform of f with respect to the Gaussian h is given
by
Vhf (z) =X cαVhhα(z) = e−πix·ξXα≥0
cαeα(z)e−πz2/2
12
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
in view of (15), definition (18) gives
kfk2
M 2
θ
= ZR2d Vhf (z)2θ(z)2 dz
= Xα,β≥0
cαcβZR2d
eα(z)eβ(z)θ(z)2 e−πz2
dz =Xα≥0
cα2τα(θ2) .
In the latter equalities it is essential that the weight θ is radial in each time-
frequency coordinate so that the monomials eα are orthogonal in L2
θ(Cd, µ).
In the next proposition, which is due to Daubechies [11], we represent the canon-
ical localization operator by a Hermite expansion.
Proposition 3.4. Let θ be a moderate, continuous weight function on R2d that is
radial in each time-frequency coordinate.
Then the Hermite function hα is an eigenfunction of the localization operator Jθ
with eigenvalue τα(θ) for α ∈ Nd
0, and Jθ possesses the eigenfunction expansion
Proof. By Lemma 3.1(a) we find that, for α 6= β,
τα(θ)hf, hαihα
for all f ∈ L2(Rd) .
(22)
Jθf =Xα≥0
hJθhβ, hαi = ZCd
= ZCd
θ(z)Vhhβ(z) Vhhα(z) dz
θ(z)eβ(z)eα(z)e−πz2
dz = 0 .
This implies that Jθhα = chα and therefore c = chhα, hαi = hJθhα, hαi = τα(θ).
For a (finite) linear combination f =Pβ≥0 cβhβ, we obtain
Jθf =Xα≥0
hJθf, hαihα =Xα≥0Xβ≥0
cβhJθhβ, hαihα
=Xα≥0Xβ≥0
τβ(θ)δα,βcβhα =Xα≥0
τα(θ)cαhα .
1/θ(Rd) and because
1/θ(Rd) is one-to-one and possesses dense range in M 2
The proposition follows because the Hermite functions span M 2
the coefficients of a Hermite expansion are unique and given by cα = hf, hα).
Corollary 3.5. If θ is moderate and radial in each coordinate, then Jθ : L2(Rd) →
M 2
Proof. The coefficients in Jθf =Pα≥0 τα(θ)hf, hαihα are unique. If Jθf = 0, then
τα(θ)hf, hαi = 0, and since τα(θ) > 0 we obtain hf, hαi = 0 and thus f = 0. Clearly
the range of Jθ in M 2
1/θ(Rd) contains the finite linear combinations of Hermite
functions, and these are dense in M 2
1/θ(Rd) by Proposition 3.1.
1/θ(Rd).
THE RANGE OF LOCALIZATION OPERATORS
13
To show that Jθ maps L2(Rd) onto M 2
1/θ(Rd) is much more subtle. For this we
need a new type of inequalities valid for the weighted gamma functions in (20).
By Proposition 3.4 the number τα,s(θ) is exactly the eigenvalue of the localization
operator Jθs corresponding to the eigenfunction hα.
Proposition 3.6. If θ ∈ Mw is continuous and radial in each time-frequency
coordinate, then the mapping s 7→ τα,s(θ) is "almost multiplicative". This means
that for every s, t ∈ R there exists a constant C = C(s, t) such that
C−1 ≤ τα,s(θ)τα,t(θ)τα,−s−t(θ) ≤ C
for all multi-indices α .
(23)
Proof. The upper bound is easy. By Lemma 3.1 the Hermite function hα is a
common eigenfunction of Jθs, Jθt, and Jθ−s−t. Since the operator JθsJθtJθ−s−t is
bounded on L2(Rd) by repeated application of Lemma 2.1, we obtain that
τα,s(θ)τα,t(θ)τα,−s−t(θ) = kJθsJθtJθ−s−thαkL2 ≤ CkhαkL2 = C
(24)
for all α ≥ 0. The constant C is operator norm of JθsJθtJ−s−t on L2(Rd).
For the lower bound we rewrite the definition of τα,s(θ) and make it more explicit
by using polar coordinates zj = rjeiϕj , rj ≥ 0, ϕj ∈ [0, 2π), in each variable. Then
by assumption θ(z) = θ0(r) for some continuous moderate function θ0 on Rd
+, and
we obtain
τα,s(θ) =ZR2d
θ(z)s πα
dz
+
α! zα2e−πz2
= (2π)dZRd
. . .Z ∞
. . .Z ∞
0
0
0
= (2π)dZ ∞
=Z ∞
0
θ0(r)
πα
α!
r2αe−πr2
r1 · · · rd dr
θ0(r1, . . . , rd)
dYj=1
1
αj!
(πr2
j )αj e−πr2
j r1 · · · rd dr1 · · · drd
θ0(cid:0)pu1/π, . . . ,pud/π(cid:1) dYj=1
uαj
j
αj!
e−uj du1 · · · dud .
(25)
We focus on a single factor in the integral first. The function fn(x) = xne−x/n!
takes its maximum at x = n and
fn(n) =
1
n!
nne−n = (2πn)−1/2(cid:0)1 + O(n−1)(cid:1)
by Stirling's formula. Furthermore, fn is almost constant on the interval [n −
√n/2, n +√n/2] of length √n. On this interval the minimum of fn is taken at one
of the endpoints n ± √n/2, where the value is
fn(n ± √n/2) =
1
n!
(n ± √n/2)ne−(n±√n/2) =
1
n!
nn
en(cid:0)1 ±
1
2√n(cid:1)ne∓√n/2 .
14
Since
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
(2πn)1/2(cid:16) n
e(cid:17)n
n!
lim
n→∞
= 1
and
lim
n→∞(cid:0)1 ±
1
2√n(cid:1)n
e∓√n/2 = e−1/8
by Stirling's formula and straight-forward applications of Taylor's formula, we find
that
fn(x) ≥
c
√n
for x ∈ [n − √n/2, n + √n/2] and all n ≥ 1 .
(26)
Now consider the products of the fn's occuring in the integral above. For α =
For n = 0 we use the inequality f0(x) ≥ e−1/2 for x ∈ [0, 1/2].
(α1, . . . , αd) ∈ Nd
0 define the boxes
Cα =
dYj=1(cid:2)αj − 2−1√αj , αj + 2−1 max(√αj, 1)(cid:3) ⊆ Rd ,
with volume vol (Cα) = Qd
have
j=1pmax(2−1, αj). Consequently, on the box Cα we
dYj=1
uαj
j
αj!
e−uj ≥ C0
dYj=1
1
pmax(2−1, αj)
= C0(cid:0)vol Cα(cid:1)−1
(27)
for some constant C0 > 0 which is independent of α ∈ Nd
0.
Next, to take into account the coordinate change in (25), we define the box
(αj + 2−1 max(√αj, 1))1/2
√π
i ⊆ Rd ,
Furthermore the length of each edge of Dα is
Dα =
,
√π
dYj=1h (αj − 2−1√αj)1/2
π−1/2(cid:16)(cid:0)αj + √αj/2(cid:1)1/2
if z1, z2 ∈ Dα,
when αj ≥ 1 and likewise for αj = 0. Consequently,
−(cid:0)αj − √αj/2(cid:1)1/2(cid:17) ≤ π−1/2
then z1 − z2 ⊆ [−π−1/2, π−1/2]d .
(28)
After these preparations we start the lower estimate of τα,s(θ). Using (27) we
obtain
τα,s(θ) = Z ∞
0
0
. . .Z ∞
vol (Cα)ZCα
dYj=1
θ0(cid:0)pu1/π, . . . ,pud/π(cid:1)s
θ0(cid:0)pu1/π, . . . ,pud/π(cid:1)s
1
uαj
j
αj!
e−uj du1 · · · dud
du1 · · · dud .
≥ C0
Since θ is continuous, the mean value theorem asserts that there is a point z =
z(α, s) = (z1, z2, . . . , zd) ∈ Cα, such that
τα,s(θ) ≥ C0 θ0(cid:0)pz1/π, . . . ,pzd/π(cid:1)s
.
THE RANGE OF LOCALIZATION OPERATORS
15
Note that the point with coordinates ζ = ζ(α, s) =(cid:0)pz1/π, . . . ,pzd/π(cid:1) is in Dα,
consequently
Finally
τα,s(θ) ≥ C0 θ0(ζ(α, s))
for ζ(α, s) ∈ Dα .
τα,s(θ) τα,t(θ) τα,−s−t(θ) ≥ C 3
0 θ0(ζ)s θ0(ζ′)t θ0(ζ′′)−s−t
for points ζ, ζ′, ζ′′ ∈ Dα. Since the weight θ is a w-moderate, θ0 satisfies
θ0(z1)
θ0(z2) ≥
1
w(z1 − z2)
z1, z2 ∈ Rd .
(29)
Since ζ, ζ′, ζ′′ ∈ Dα, the differences ζ−ζ′′ and ζ′−ζ′′ are in the cube [−π−1/2, π−1/2]d
as observed in (28). We conclude the non-trivial part of this estimate by
w(z)(cid:17)−s−t
0(cid:16)
w(ζ′ − ζ′′)t ≥ C 3
w(ζ − ζ′′)s
τα,s(θ)τα,t(θ)τα,−s−t(θ) ≥ C 3
The proof is complete.
0
max
z∈[−π−1/2,π−1/2]d
1
1
= C .
The next result provides a sort of symbolic calculus for the canonical localization
operators Jθs. Although the mapping s → Jθs is not homomorphism from R to
operators, it is multiplicative modulo bounded operators.
Theorem 3.7. Let θ and µ be two moderate, continuous weight functions on R2d
that are radial in each time-frequency variable. For every r, s ∈ R there exists an
operator Vs,t that is invertible on every M 2
µ(Rd) such that
JθsJθtJθ−s−t = Vs,t .
Proof. For s, t ∈ R fixed, set γ(α) = τα,s(θ)τα,t(θ)τα,−s−t(θ) and
Vs,tf =Xα≥0
γ(α)hf, hαihα .
Clearly, Vs,t = JθsJθtJθ−s−t. Since C−1 ≤ γ(α) ≤ C for all α ≥ 0 by Proposition 3.6,
Proposition 3.3 implies that Vs,t is bounded on every modulation space M 2
µ. Like-
s,t f = Pα≥0 γ(α)−1hf, hαihα is bounded on
wise the formal inverse operator V −1
M 2
µ(Rd), consequently Vs,t is invertible on M 2
µ(Rd).
We can now finish the proof of Theorem 3.2.
Proof of Theorem 3.2. Choose s = 1 and t = −1, then JθJ1/θ = V1,−1 is invertible
on L2. Similarly, the choice s = −1, t = 1 yields that J1/θJθ = V−1,1 is invertible
θ . The factorization JθJ1/θ = V1,−1 implies that J1/θ is one-to-one from L2 to
on M 2
M 2
θ and that Jθ maps M 2
θ onto L2. The factorization J1/θJθ = V−1,1 implies that
θ to L2 and that J1/θ maps L2 onto M 2
Jθ is one-to-one from M 2
θ .
θ to L2 and that J1/θ is an
We have proved that Jθ is an isomorphism from M 2
isomorphism from L2 to M 2
θ . The other isomorphisms are proved similarly.
16
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
REMARK: In dimension d = 1 the invertibility of JθJθ−1 follows from the equiv-
alence τn,1(θ)τn,−1(θ) ≍ 1, which can be expressed as the following inequality for
weighted gamma functions:
C−1 ≤Z ∞
0
θ0(px/π)
xn
n!
e−xdx Z ∞
0
xn
n!
e−xdx ≤ C
(30)
1
θ0(px/π)
for all n ≥ 0. Here θ0 is the same as before. It is a curious and fascinating fact
that this inequality implies that the localization operator J1/θ is an isomorphism
between L2(R) and M 2
θ (R).
4. The General Isomorphism Theorems
In Theorem 4.3 we will state the general isomorphism theorems. The strategy
of the proof is similar to that of Theorem 3.2 in [19]. The main tools are the
theorems about the spectral invariance of the generalized Sjostrand classes [16] and
the existence of a canonical isomorphism between L2(Rd) and M 2
θ (Rd) established
in Theorem 3.2.
4.1. Variations on Spectral Invariance. We first introduce the tools concerning
the spectral invariance of pseudodifferential operators. Recall the following results
from [16].
Theorem 4.1. Let v be a submultiplicative weight on R2d such that
lim
n→∞
v(nz)1/n = 1
for all z ∈ R2d ,
(31)
) and T is invertible on L2(Rd),
and let v be the same as in (10). If T ∈ Op(M∞,1
then T −1 ∈ Op(M∞,1
for 1 ≤ p, q ≤ ∞ and all µ ∈ Mv.
).
v
v
Consequently, T is invertible simultaneously on all modulation spaces M p,q
µ (Rd)
Condition (31) is usually called the Gelfand-Raikov-Shilov (GRS) condition.
We prove a more general form of spectral invariance. Since we have formulated
all results about the canonical localization operators Jθ for radial weights only, we
will assume from now on that all weights are radial in each coordinate. In this case
v(x, ξ, η, y) = v(−η, y) = v(y, η),
v
) and T is invertible on M 2
If T ∈ Op(M∞,1
As a consequence T is invertible on every modulation space M p,q
and we do not need the somewhat ugly distinction between v and v.
Theorem 4.2. Assume that v satisfies the GRS-condition, θ2 ∈ Mv, and that both
v and θ are radial in each time-frequency coordinate.
θ (Rd), then T is invertible on L2(Rd).
µ (Rd) for 1 ≤
p, q ≤ ∞ and µ ∈ Mv.
Proof. Set eT = JθT J1/θ. By Theorem 3.2, J1/θ is an isomorphism from L2(Rd)
θ (Rd) onto L2(Rd), therefore eT is
onto M 2
an isomorphism on L2(Rd).
θ (Rd) and Jθ is an isomorphism from M 2
THE RANGE OF LOCALIZATION OPERATORS
17
(32)
Theorem 4.1 on the spectral invariance of the symbol class M∞,1
θ
M 2
θ
↑ J1/θ
L2(Rd)
T−→ M 2
↓ Jθ
eT−→ L2(Rd)
By Theorem 2.4 the operator eT is in Op(M∞,1
inverse operator eT also possesses a symbol in M∞,1
Now, since eT −1 = J−1
T −1 = J1/θeT −1Jθ .
1/θT −1J−1
, we find that
θ
v1/2 ). Since eT is invertible on L2(Rd),
v1/2 , i.e., eT −1 ∈ Op(M∞,1
v1/2 implies that the
v1/2 ).
Applying Theorem 2.4 once again, the symbol of T −1 must be in M∞,1. As a
consequence, T −1 is bounded on L2(Rd).
) and T is invertible on L2(Rd), it follows that T is also
Since T ∈ Op(M∞,1
invertible on M p,q
v
µ (Rd) for every weight µ ∈ Mv and 1 ≤ p, q ≤ ∞.
4.2. An isomorphism theorem for localization operators. We now combine
all steps and formulate and prove our main result, the isomorphism theorem for
time-frequency localization operators with symbols of superfast growth.
v2w(Rd), µ ∈ Mw and m ∈ Mv be such that m is
Theorem 4.3. Let g ∈ M 1
radial in each time-frequency coordinate and v satisfies (31). Then the localization
µ/m(Rd) for 1 ≤ p, q ≤ ∞.
operator Ag
m is an isomorphism from M p,q
µ (Rd) onto M p,q
Proof. Set T = Ag
(1) T = Ag
(2) T is invertible on M 2
1/mAg
1/mAg
m possesses a symbol in M∞,1
θ (Rd) by Lemma 2.2.
m. We have already established that
v w by Corollary 2.5.
These are the assumptions of Theorem 4.2, and therefore T is invertible on M p,q
for every µ ∈ Mw ⊆ Mvw and 1 ≤ p, q ≤ ∞. The factorization T = Ag
implies that Ag
M p,q
µ/m(Rd) onto M p,q
Now we change the order of the factors and consider the operator T ′ = Ag
m is one-to-one from M p,q
µ/m(Rd) and that Ag
µ (Rd) to M p,q
µ (Rd).
µ (Rd)
1/mAg
m
1/m maps
µ/m(Rd) to M p,q
µ/m(Rd) and factoring through M p,q
v w and is invertible on M 2
from M p,q
a symbol in M∞,1
that T ′ is invertible on all modulation spaces M p,q
T ′ = Ag
that Ag
1/m now yields that Ag
µ (Rd) onto M p,q
mAg
m maps M p,q
1/m is one-to-one from M p,q
µ/m(Rd).
1/m
µ (Rd). Again T ′ possesses
1/θ(Rd). With Theorem 4.2 we conclude
µ/m(Rd). The factorization of
µ (Rd) and
µ/m(Rd) to M p,q
mAg
As a consequence Ag
m is bijective from M p,q
µ (Rd) onto M p,q
µ/m(Rd), and Ag
1/m is
bijective from M p,q
µ/m(Rd) onto M p,q
µ (Rd).
18
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
5. Consequences for Gabor Multipliers and Toeplitz Operators on
Bargmann Fock Space
5.1. Gabor Multipliers. Gabor multipliers are time-frequency localization oper-
ators whose symbols are discrete measures. Their basic properties are the same,
but the discrete definition makes them more accessible for numerical computations.
Let Λ = AZ2d for some A ∈ GL(2d, R) be a lattice in R2d, g, γ suitable window
functions and m a weight sequence defined on Λ. Then the Gabor multiplier Gg,γ,Λ
is defined to be
m
m(λ)hf, π(λ)giπ(λ)γ .
(33)
Gg,γ,Λ
m f =Xλ∈Λ
m
The boundedness of Gabor multipliers between modulation spaces is formulated
and proved exactly as for time-frequency localization operators. See [13] for a
detailed exposition of Gabor multipliers and [9] for general boundedness results
that include distributional symbols.
Proposition 5.1. Let g, γ ∈ M 1
m ∈ Mv. Then Gg,γ,Λ
vw(Rd), m be a continuous moderate weight function
is bounded from M p,q
µ (R2d) to M p,q
µ/m(Rd).
Gabor frame for L2(Rd). This means that the frame operator Sg,Λf = M g,g,Λ
In the following we will assume that the set G(g, Λ) = {π(λ)g : λ ∈ Λ} is a
=
Pλ∈Λhf, π(λ)giπ(λ)g is invertible on L2(Rd). From the rich theory of Gabor frames
we quote only the following result: If g ∈ M 1
v and G(g, Λ) is a frame for L2(Rd),
1 ≤ p ≤ ∞, and m ∈ Mv, then
hf, π(λ)gipm(λ)p(cid:17)1/p
for all f ∈ M p
kfkM p
m(Rd).
(34)
1
m ≍(cid:16)Xλ∈Λ
We refer to [15] for a detailed discussion, the proof, and for further references.
Our main result on Gabor multipliers is the isomorphism theorem.
Theorem 5.2. Assume that g ∈ M 1
If m ∈ Mv is radial in each time-frequency coordinate, then Gg,g,Λ
phism from M p,q
v2w(Rd) and that G(g, Λ) is a frame for L2(Rd).
is an isomor-
µ/m(Rd) for every µ ∈ Mw and 1 ≤ p, q ≤ ∞.
µ (R2d) onto M p,q
The proof is similar to the proof of Theorem 4.3, and we only sketch the necessary
modifications. In the following we fix the lattice Λ and choose γ = g and drop the
reference to these additional parameter by writing Gg,γ,Λ
m in analogy to the
localization operator Ag
Proposition 5.3. If g ∈ M 1
isomorphism from M 2
v (Rd) and m ∈ Mv and θ = m1/2, then Gg
m is an
as Gg
θ (Rd) onto M 2
1/θ(Rd).
m.
m
m
Proof. By Proposition 5.1 Gg
mfkM 2
Next we use the characterization of M 2
multipliers:
θ to M 2
θ/m = M 2
1/θ and thus
m is bounded from M 2
kGg
1/θ ≤ CkfkM 2
θ by Gabor frames and relate it to Gabor
mf, fi =Xλ∈Λ
hGg
m(λ)hf, π(λ)gi2 ≍ kfk2
(35)
M 2
θ
.
.
θ
THE RANGE OF LOCALIZATION OPERATORS
19
This identity implies that Gg
equivalent norm on M 2
whence Gg
m is onto as well.
m is one-to-one on M 2
θ and that kGg
θ . Since Gg
m is self-adjoint, it has dense range in M 2
mfkM1/θ is an
1/θ,
Using the weight 1/m instead of m and 1/θ instead of θ, we see that Gg
1/m is an
isomorphism from M 2
1/θ(Rd) onto M 2
θ (Rd).
Proof of Theorem 5.2. We proceed as in the proof of Theorem 4.3. Define the
operators T = Gg
θ (Rd) to
M 2
θ (Rd), and since T is a composition of two isomorphisms, T is an isomorphism
on M 2
1/m. By Proposition 5.1 T maps M 2
θ (Rd). Likewise T ′ is an isomorphism on M 2
m and T ′ = Gg
1/mGg
1/θ(Rd).
mGg
As in Corollary 2.5 we verify that the symbol of T is in M∞,1
v w (R2d). Theorem 4.2
then asserts that T is invertible on L2(Rd) and the general spectral invariance
implies that T is an isomorphism on M p,q
µ (Rd). Likewise T ′ is an isomorphism on
M p,q
µ/m(Rd). This means that each of the factors of T must be an isomorphism on
the correct space.
5.2. Toeplitz Operators. The Bargmann-Fock space F = F 2(Cd) is the Hilbert
space of all entire functions of d variables such that
=Z d
C F (z)2e−πz2
kFk2
F
dz < ∞ .
(36)
Here dz is the Lebesgue measure on Cd = R2d.
Related to the Bargmann-Fock space F 2 are spaces of entire functions satisfying
weighted integrability conditions. Let m be a moderate weight on Cd and 1 ≤
m (Cd) is the Banach space of all entire functions of d
p, q ≤ ∞. Then the space F p,q
complex variables, such that
=ZRd(cid:16)ZRd F (x + iy)pm(x + iy)pe−pπx+iy2/2dx(cid:17)q/p
kFkq
Let P be the orthogonal projection from L2(Cd, µ) with Gaussian measure dµ(z) =
dy < ∞ .
F p,q
(37)
m
e−πz2 dz onto F 2(Cd). Then P is given by the formula
F (z)eπzwe−πz2
dz .
P F (w) =ZCd
(38)
We remark that a function on R2d with a certain growth is entire if and only if it
satisfies F = P F .
The classical Toeplitz operator on Bargmann-Fock space with a symbol m is
defined by TmF = P (mF ) for F ∈ F 2(Cd). Explicitly Tm is given by the formula
(39)
m(z)F (z)eπzw e−πz2
dz .
TmF (w) =ZCd
See [1–4, 11, 14] for a sample of references.
20
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
In the following we assume that the symbol of a Toeplitz operator is continu-
ous and radial in each coordinate, i.e., m(z1, . . . , zd) = m0(z1, . . .zd) for some
continuous function m0 from Rd
Theorem 5.4. Let µ ∈ Mw, and let m ∈ Mv be a continuous moderate weight
function such that one of the following conditions is fulfilled:
+ to R+.
(1) Either m is radial in each coordinate,
(2) or m is of polynomial type.
Then the Toeplitz operator Tm is an isomorphism from F p,q
1 ≤ p, q ≤ ∞.
µ (Cd) onto F p,q
µ/m(Cd) for
The formulation of Theorem 5.4 looks similar to the main theorem about time-
frequency localization operators. In fact, after a suitable translation of concepts,
it is a special case of Theorem 4.3.
To explain the connection, we recall the Bargmann transform that maps distri-
butions on Rd to entire functions on Cd.
Bf (z) = F (z) = 2d/4e−πz2/2ZRd
f (t)e−πt2
e2πt·zdt
z ∈ Cd .
(40)
If f is a distribution, then we interpret the integral as the action of f on the
function e−πt2e2πt·z.
The connection to time-frequency analysis comes from the fact that the Bargmann
transform is just a short-time Fourier transform in disguise [15, Ch. 3]. As before,
we use the normalized Gaussian h(t) = 2d/4e−πt2 as a window. Identifying the pair
(x, ξ) ∈ Rd × Rd with the complex vector z = x + iξ ∈ Cd, the short-time Fourier
transform of a function f on Rd with respect to h is
Vhf (z) = eπix·ξBf (z)e−πz2/2 .
(41)
In particular, the Bargmann transform of the time-frequency shift π(w)h is given
by
B(π(w)h)(z) = e−πiu·ηeπwze−πw2/2 w, z ∈ Cd, w = u + iη .
(42)
It is a basic fact that the Bargmann transform is a unitary mapping from L2(R) onto
the Bargmann-Fock space F 2(Cd). Furthermore, the Hermite functions hα, α ≥ 0
are mapped to the normalized monomials eα(z) = πα/2(α!)−1/2zα.
Let m′(z) = m(z). Then (41) implies that the Bargmann transform B maps
m (Rd) isometrically to F p,q
M p,q
m′ (Cd). By straight-forward arguments it follows that
m (Rd) onto the Fock space
the Bargmann transform maps the modulation space M p,q
F p,q
m′ (Cd) [15, 25].
The connection between time-frequency localization operators and Toeplitz op-
erators on the Bargmann-Fock space is given by the following statement.
Proposition 5.5. Let m be a moderate weight function on R2d ≃ Cd and set
m′(z) = m(z). The Bargmann transform intertwines Jm and Tm′, i.e., for all
f ∈ L2(Rd) (or f ∈ (S1/2
1/2)∗) we have
B(Jmf ) = Tm′(Bf ) .
(43)
THE RANGE OF LOCALIZATION OPERATORS
21
Proof. This fact is well-known [7,11]. For completeness and consistency of notation,
we provide the formal calculation.
We take the short-time Fourier transform of Jm = Ah
m with respect to h. On the
one hand we obtain that
hJmf, π(w)hi = eπiu·ηB(Jmf )(w) e−πw2/2 ,
(44)
and on the other hand, after substituting (41) and (42), we obtain that
hJmf, π(w)hi = ZR2d
= ZCd
= eπiu·ηZCd
m(z)Vhf (z)Vh(π(w)h)(z) dz
m(z)Bf (z) B(π(w)h)(z)e−πz2
dz
(45)
m(z) Bf (z)eπzw e−πz2
dz e−πw2/2 .
Comparing (44) and (45) we obtain that
B(Jmf )(w) =ZCd
m(z)Bf (z)eπzw e−πz2
dz = Tm′Bf (w) .
Proof of Theorem 5.4. First assume that (1) holds, i.e., m is radial in each vari-
able. The Bargmann transform is an isomorphism between the modulation space
m (Rd) and the Bargmann-Fock space F p,q
M p,q
m′ (Cd) for arbitrary moderate weight
function m and 1 ≤ p, q ≤ ∞. By Theorem 4.3 the canonical localization operator
µ (Rd) onto M p,q
Jm is an isomorphism from M p,q
µ/m(Rd). Since Tm′ is a composition
of three isomorphism (Proposition 5.5 and (46)), the Toeplitz operator Tm′ is an
isomorphism from F p,q
µ′/m′(Cd).
µ′ (Cd) onto F p,q
F p,q
↑ B
M p,q
µ
µ′
µ′/m′
Tm′
−→ F p,q
↑ B
Jm−→ M p,q
µ/m
.
(46)
Finally replace m′ and µ′ by m and µ.
By using Theorem 3.2 in [19] instead of Theorem 4.3, the same arguments show
that the result follows when (2) is fulfilled.
References
[1] V. Bargmann. On a Hilbert space of analytic functions and an associated integral transform.
Comm. Pure Appl. Math., 14:187–214, 1961.
[2] F. A. Berezin. Wick and anti-Wick symbols of operators. Mat. Sb. (N.S.), 86(128):578–610,
1971.
[3] C. A. Berger and L. A. Coburn. Toeplitz operators on the Segal-Bargmann space. Trans.
Amer. Math. Soc., 301(2):813–829, 1987.
[4] C. A. Berger and L. A. Coburn. Heat flow and Berezin-Toeplitz estimates. Amer. J. Math.,
116(3):563–590, 1994.
22
KARLHEINZ GR OCHENIG AND JOACHIM TOFT
[5] P. Boggiatto, E. Cordero, and K. Grochenig. Generalized anti-Wick operators with symbols
in distributional Sobolev spaces. Integral Equations Operator Theory, 48(4):427–442, 2004.
[6] J.-M. Bony and J.-Y. Chemin. Espaces fonctionnels associ´es au calcul de Weyl-Hormander.
Bull. Soc. Math. France, 122(1):77–118, 1994.
[7] L. A. Coburn. The Bargmann isometry and Gabor-Daubechies wavelet localization operators.
In Systems, approximation, singular integral operators, and related topics (Bordeaux, 2000),
volume 129 of Oper. Theory Adv. Appl., pages 169–178. Birkhauser, Basel, 2001.
[8] J. B. Conway. A course in functional analysis. Springer-Verlag, New York, second edition,
1990.
[9] E. Cordero and K. Grochenig. Time-frequency analysis of localization operators. J. Funct.
Anal., 205(1):107–131, 2003.
[10] E. Cordero and K. Grochenig. Symbolic calculus and fredholm property for localization
operators. J. Fourier Anal. Appl., 12(4):345–370, 2006.
[11] I. Daubechies. Time-frequency localization operators: a geometric phase space approach.
IEEE Trans. Inform. Theory, 34(4):605–612, 1988.
[12] H. G. Feichtinger and K. Grochenig. Banach spaces related to integrable group representa-
tions and their atomic decompositions. I. J. Functional Anal., 86(2):307–340, 1989.
[13] H. G. Feichtinger and K. Nowak. A first survey of Gabor multipliers. In Advances in Gabor
analysis, Appl. Numer. Harmon. Anal., pages 99–128. Birkhauser Boston, Boston, MA, 2003.
[14] G. B. Folland. Harmonic Analysis in Phase Space. Princeton Univ. Press, Princeton, NJ,
1989.
[15] K. Grochenig. Foundations of time-frequency analysis. Birkhauser Boston Inc., Boston, MA,
2001.
[16] K. Grochenig. Time-frequency analysis of Sjostrand's class. Revista Mat. Iberoam., 22(2):703–
724, 2006.
[17] K. Grochenig. Weight functions in time-frequency analysis. In e. a. L. Rodino, M.-W. Wong,
editor, Pseudodifferential Operators: Partial Differential Equations and Time-Frequency
Analysis, volume 52, pages 343 – 366. Fields Institute Comm., 2007.
[18] K. Grochenig and Z. Rzeszotnik. Banach algebras of pseudodifferential operators and their
almost diagonalization. Ann. Inst. Fourier (Grenoble), 58(7):2279–2314, 2008.
[19] K. Grochenig and J. Toft. Isomorphism properties of Toeplitz operators and pseudo-
differential operators between modulation spaces. J. Anal. Math., To appear.
[20] A. Holst, J. Toft and P. Wahlberg. Weyl product algebras and modulation spaces. J. Funct.
Anal., 251:463-491, 2007.
[21] A. J. E. M. Janssen. Bargmann transform, Zak transform, and coherent states. J. Math.
Phys., 23(5):720–731, 1982.
[22] N. Lerner. The Wick calculus of pseudo-differential operators and some of its applications.
Cubo Mat. Educ., 5(1):213–236, 2003.
[23] N. Lerner and Y. Morimoto. A Wiener algebra for the Fefferman-Phong inequality. In Sem-
inaire: Equations aux D´eriv´ees Partielles. 2005–2006, S´emin. ´Equ. D´eriv. Partielles, pages
Exp. No. XVII, 12. ´Ecole Polytech., Palaiseau, 2006.
[24] M. A. Shubin. Pseudodifferential Operators and Spectral Theory. Springer-Verlag, Berlin,
second edition, 2001. Translated from the 1978 Russian original by Stig I. Andersson.
[25] M. Signahl, J. Toft Mapping properties for the Bargmann transform on modulation spaces,
Preprint, arXiv:1002.3061.
[26] S. Thangavelu. Lectures on Hermite and Laguerre expansions, volume 42 of Mathemati-
cal Notes. Princeton University Press, Princeton, NJ, 1993. With a preface by Robert S.
Strichartz.
[27] J. Toft. Regularizations, Decompositions and Lower Bound Problems in the Weyl Calculus.
Comm. Partial Differential Equations, 25 (7& 8): 1201–1234, 2000.
[28] J. Toft. Subalgebras to a Wiener type algebra of pseudo-differential operators. Ann. Inst.
Fourier (Grenoble), 51(5):1347–1383, 2001.
THE RANGE OF LOCALIZATION OPERATORS
23
[29] J. Toft. Continuity properties for modulation spaces, with applications to pseudo-differential
calculus. I. J. Funct. Anal., 207(2):399–429, 2004.
[30] J. Toft. Continuity and Schatten-von Neumann Properties for Toeplitz Operators on Modu-
lation Spaces in: J. Toft, M. W. Wong, H. Zhu (Eds), Modern Trends in Pseudo-Differential
Operators, Operator Theory Advances and Applications Vol 172, Birkhauser Verlag, Basel:
pp. 313–328, 2007.
[31] J. Toft. Multiplication properties in pseudo-differential calculus with small regularity on the
symbols. J. Pseudo-Differ. Oper. Appl. 1(1): 101–138, 2010.
[32] H. Triebel. Theory of function spaces. Birkhauser Verlag, Basel, 1983.
[33] M. W. Wong. Wavelets Transforms and Localization Operators, volume 136 of Operator
Theory Advances and Applications. Birkhauser, 2002.
Faculty of Mathematics, University of Vienna, Nordbergstrasse 15, A-1090 Vi-
enna, Austria
E-mail address: [email protected]
Department of Computer science, Physics and Mathematics, Linnaeus University,
Vaxjo, Sweden
E-mail address: [email protected]
|
1707.01282 | 5 | 1707 | 2017-10-31T01:09:29 | Addition Formulas of Leaf Functions According to Integral Root of Polynomial Based on Analogies of Inverse Trigonometric Functions and Inverse Lemniscate Functions | [
"math.FA"
] | The second derivative of a function r(t) with respect to a variable t is equal to -n times the function raised to the 2n-1 power of r(t); using this definition, an ordinary differential equation is constructed. Graphs with the horizontal axis as the variable t and the vertical axis as the variable r(t) are created by numerically solving the ordinary differential equation. These graphs show several regular waves with a specific periodicity and waveform depending on the natural number n. In this study, the functions that satisfy the ordinary differential equation are presented as the leaf functions. For n = 1, the leaf function become a trigonometric function. For n = 2, the leaf function becomes a lemniscatic elliptic function. These functions involve an additional theorem. In this paper, based on the additional theorem for n = 1 and n = 2, the double angle and additional theorem for n = 3 are presented. | math.FA | math | Addition Formulas of Leaf Functions According to
Integral Root of Polynomial Based on Analogies of
Inverse Trigonometric Functions and Inverse
Lemniscate Functions
Kazunori Shinohara
Department of Mechanical System Engineering, Daido University
10-3 Takiharu-cho, Minami-ku, Nagoya 457-8530, Japan
Abstract
Inverse trigonometric functions arcsin(r) and arccos(r) are obtained by integrating
the reciprocals of the root of the polynomial 1 − t2. Based on this analogy, the
inverse lemniscate functions arcsl(r) and arccl(r) are created by integrating the
reciprocal of the root of the polynomial 1 − t4. The lemniscate functions are then
extended to the Jacobi elliptic functions, which in turn are improved using the theta
functions. However, to the best of our knowledge, the integration of the reciprocals
of the root of the polynomial 1 − t6 has never been published. The inverse functions
based on 1 − t6 are defined as arcsleaf3(r) and arccleaf3(r), and they were
determined to produce waves different from those of the trigonometric and
lemniscate functions. This paper presents the addition formulas of the artificially
created functions sleaf3(l) and cleaf3(l). These formulas were numerically verified
through examples.
Keywords: Leaf functions, Trigonometric functions, Lemniscatic elliptic functions,
Inverse functions, Addition formulas
1 Introduction
The second derivative of a function r(l) with respect to a variable l is equal to -n
times the function raised to the 2n-1 power of r(l) using this definition, an ordinary
differential equation is constructed. Graphs with the horizontal axis as the variable
l and the vertical axis as the variable r(l) are created by numerically solving the
ordinary differential equation. These graphs show several regular waves with a
specific periodicity and waveform depending on the natural number n [1] [2].
lrd
2
dl
2
The definition of the arc leaf function arcsleafn(r) is obtained by the above equation.
arcsleaf
(1.1)
(1.2)
lrn
1
l
n
2
r
r
n
0
dt
t
1
2
n
In case of n=3, the above ordinary differential equations are as follows:
lrd
2
dl
2
3
5
lr
(1.3)
These equations are related in the nonlinear equations (Quintic Duffing equation)
[3]. To obtain the solution accuracy, or to grasp behavior of the solution, the
addition formula often be demanded in the solution process [4] [5] [6]. Historically,
the trigonomic function (in n=1 of the Eq. (1.2)) and the lemniscatic elliptic
function (in n=2 of the Eq. (1.2)) have been discussed [7] [8] [9] [10] [11]. However,
in case of n=3, the integration of the reciprocals of the root of the polynomial 1− t6
has never been published.
The variable n represents a natural number. For n = 1, the above equation is written
as follows:
arcsleaf
(1.4)
arcsin
l
r
r
1
r
10
dt
2
t
1
2
l
2
l
2
cos
l
l
or
sin2
sleaf
sleaf
The function arcsin(r) represents inverse trigonometric functions. The double angle
and the addition formulas of the function sin(l) are well known, as follows:
sin
sin(
or
For n = 2, Eq.(1.2) is rewritten as follows:
l
cos
1
l
l
2
sleaf
l
2
cleaf
sin
l
1
sin
l
1
)
2
sleaf
(1.5)
(1.6)
l
1
cos
l
cleaf
1
l
sleaf
cleaf
l
l
l
l
1
1
2
1
2
l
1
1
2
1
1
1
dt
1
t
arcsleaf
r
2
r
0
arcsl
r
l
4
(1.7)
Function arcsl(r) represents the inverse lemniscatic elliptic function [12] [13] [14].
The double angle and the addition formulas are as follows:
sl
l
2
2
lsl
1
4
lsl
1
4
lsl
lsl
1
l
2
lsl
1
1
or
sleaf
l
1
2
l
2
or
sleaf
2
l
2
2
sleaf
l
1
2
sleaf
1
2
sleaf
4
l
4
l
2
(1.8)
lsl
1
4
4
lsl
lsl
2
2
2
lsl
lsl
1
1
sleaf
l
1
1
2
2
1
2
sleaf
1
2
l
2
sleaf
4
2
l
1
2
sleaf
2
sleaf
l
2
l
2
2
1
2
sleaf
2
4
l
1
(1.9)
In this study, the double angles with respect to the leaf functions sleaf3(l) and
cleaf3(l) were derived by using the analogous of the trigonometric functions and
lemniscatic elliptic function. Next, the addition formulas of these leaf functions
were derived.
2 Double angle
2.1 Leaf function: sleaf3(l)
For n = 3, the double angle is derived. Variables l1 and l2 are defined as follows:
l
1
l
2
r
1
0
r
2
0
dt
1
6
t
dt
1
t
6
3
1
l
sleaf
Variable l1 is related to variable r1 as follows:
r
1
Variable l2 is related to variable r2 as follows:
(2.1)
(2.2)
(2.3)
l
2
3
2
2l
sleaf
(2.4)
r
2
We try to find the relationship between r1 and r2 when the following relationship
exists.
l
1
The following equation is obtained by using Eqs. (2.1), (2.2), and (2.5).
The relationship between r1 and r2 is obtained as the double angle formula of the
leaf function:
(2.5)
(2.6)
dt
1
dt
1
2
r
2
t
t
r
1
6
0
6
0
(2.7)
2
r
1
r
r
1
6
2
2
r
81
6
2
By differentiating it with respect to variable r2, the following equation is obtained.
dr
1
dr
2
The following equation is transformed by using the variable r2.
r
2
40
6
2
r
81
6
2
r
12
2
r
6
2
16
1
(2.8)
3
2
1
1
r
6
1
r
81
6
2
r
20
6
2
3
2
r
8
12
2
1
The following equation is obtained by using Eqs. (2.8) and (2.9).
1
1
r
6
1
dr
1
dr
2
3
r
81
6
2
2
r
r
20
8
6
12
2
2
1
r
2
40
6
2
r
81
6
2
3
2
r
16
12
2
r
1
6
2
2
1
r
6
2
The following equation is derived from the above equation.
dr
1
1
2
r
6
1
dr
2
1
r
6
2
(2.9)
(2.10)
(2.11)
Eq. (2.6) satisfies the above equation. The double angle of the leaf function is
obtained as follows:
(1) In case of an inequality
l
2
m
1
2
3
1
2
3
m
(See [1] [2] for the constant π3.)
2
sleaf
3
l
2
2
sleaf
l
1
3
sleaf
81
sleaf
6
l
3
3
6
l
(2) In case of an inequality
2
m
1
2
3
l
2
m
3
2
3
sleaf
l
2
3
2
sleaf
l
sleaf
1
3
6
sleaf
l
81
3
3
6
l
(2.12)
(2.13)
The differentiation of the leaf function sleaf3(l) is defined by the domain of the
variable l. The sign is defined as follows [1] [2].
(1) In case of the inequality
l
2
m
2
m
1
2
3
1
2
3
(2.14)
(2.15)
3
1
l
sleaf
sleaf
d
dl
(2) In case of the inequality
3
6
l
sleaf
l
3
1
sleaf
6
l
3
d
dl
2
m
1
2
3
l
2
m
3
2
3
6
l
3
1
sleaf
depends on the domain of variable l.
The sign of the term
Because of change in sign, it is necessary to classify Eqs. (2.12) and (2.13).
2.2 Leaf function: cleaf3(l)
The leaf function sleaf3(l) is related to the leaf function cleaf3(l) as follows[1] [2]:
2
l
cleaf
3
2
l
2
sleaf
3
2
l
cleaf
3
2
l
1
(2.16)
sleaf
3
(2.17)
Variable l in the above equation is replaced by variable 2l.
cleaf
3
cleaf
3
sleaf
3
sleaf
3
2
l
2
l
2
l
2
l
2
1
2
2
2
2
By substituting Eq. (2.12) or Eq. (2.13) into Eq. (2.17), we get the following
equation.
cleaf
3
l
2
l
cleaf
2
2
cleaf
l
81
3
2
cleaf
2
cleaf
l
8
3
4
l
1
cleaf
8
6
3
3
8
l
3
(2.18)
When the function cleaf3(2l) is derived from Eq. (2.17), both plus and minus signs
occur in the function cleaf3(2l). We have to choose the plus sign so that the initial
condition, namely, cleaf3(0) =1 [1] [2] is satisfied.
3 Additional formulas
3.1 Leaf function: sleaf3(l1+l2)
With respect to arbitrary variables l1 and l2, the following equation is satisfied.
4
l
1
(3.1)
(3.2)
(3.3)
(3.4)
(3.5)
(3.6)
l
3
3
sleaf
sleaf
l
1
l
2
sleaf
sleaf
l
41
2
3
3
sleaf
sleaf
l
l
3
2
3
1
4
2
sleaf
sleaf
l
l
4
1
3
4
3
3
3
l
3
2
2
2
sleaf
sleaf
l
4
1
2
sleaf
l
1
3
2
sleaf
2
3
4
l
1
2
3
sleaf
l
1
2
l
1
sleaf
2
l
3
3
sleaf
3
l
l
1
sleaf
sleaf
41
2
2
l
2
l
2
3
3
2
The above equation is set as follows:
llg
,
2
1
lg
1,
llp
,
2
12
2
llp
,
13
llp
,
2
11
llp
,
2
13
sleaf
l
13
2
l
l
2
2
2
llp
,
2
11
sleaf
l
13
l
2
sleaf
3
l
2
sleaf
3
l
2
l
13
sleaf
l
1
lp
12
,
l
2
sleaf
l
13
3
sleaf
3
l
2
sleaf
l
13
sleaf
3
l
2
3
llp
,
2
13
41
sleaf
4
l
13
sleaf
3
l
2
2
4
sleaf
2
l
13
sleaf
3
l
2
4
llp
,
2
12
l
1
sleaf
3
3
3
l
2
lp
13
l
1
,
l
2
For Eqs. (3.4)–(3.6), the following equations are obtained by differentiating with
respect to variable l1.
l
1
l
l
,
2
3
3
2
3
sleaf
l
sleaf
2
3
5
l
1
3
(3.7)
lp
1
1
l
1
sleaf
l
1
sleaf
l
2
sleaf
2
l
13
sleaf
3
l
2
sleaf
2
l
13
sleaf
3
l
13
sleaf
l
1
sleaf
l
1
2
l
2
sleaf
3
l
13
l
2
3
l
13
sleaf
l
1
(3.8)
sleaf
l
3
13
l
13
sleaf
8
sleaf
l
1
2
sleaf
3
l
2
2
l
13
sleaf
l
1
sleaf
16
3
l
13
2
2
2
sleaf
8
l
sleaf
3
13
l
sleaf
l
13
2
sleaf
3
l
2
The following equation is substituted into Eq. (3.7).
2
l
3
sleaf
5
l
3
sleaf
l
2
3
l
2
4
l
13
sleaf
l
1
(3.9)
(3.10)
For Eqs. (3.4)–(3.6), the following equations are obtained by differentiating with
respect to variable l2.
lp
11
l
sleaf
3
l
(3.11)
sleaf
3
sleaf
3
l
13
l
13
sleaf
l
1
5
l
l
l
,
2
2
2
2
2
lp
1
2
l
2
,
l
2
l
1
sleaf
3
3
l
1
l
2
sleaf
3
l
2
sleaf
3
sleaf
3
2
l
1
3
sleaf
3
l
2
2
l
2
2
sleaf
3
l
sleaf
3
3
sleaf
3
l
2
2
lp
1
3
l
2
,
l
2
16
sleaf
l
2
3
3
sleaf
2
l
1
3
l
2
3
sleaf
l
2
2
l
1
sleaf
2
l
1
3
2
sleaf
l
2
3
2
l
2
l
1
sleaf
3
l
2
(3.12)
2
l
2
4
l
1
3
sleaf
l
3
sleaf
8
sleaf
l
3
2
l
2
3
sleaf
l
2
(3.13)
8
sleaf
l
2
3
sleaf
3
For Eq. (3.2), the following equation is obtained by differentiating with respect to
variable l1.
2
l
2
lp
11
,
l
2
,
lg
1
l
1
lp
11
,
l
2
2
2
lg
1
,
l
2
lp
1
2
l
,
lp
3
1
2
2
l
,
2
l
l
,
lp
1
3
lp
1
2
2
2
,
l
lp
11
l
1
lg
,
2
1
lp
1
3
l
1
2
l
,
2
2
,
l
2
2
lp
1
2
l
1
,
l
2
2
lp
1
3
,
l
2
(3.14)
Using Eqs. (3.11)–(3.13), the numerator of Eq. (3.16) is expanded as follows:
2
2
2
2
2
2
2
2
,
l
,
l
,
l
,
l
,
l
,
l
lp
11
lp
1
3
lp
1
2
l
1
lp
11
l
1
Using Eqs. (3.7)–(3.9), the numerator of Eq. (3.14) is expanded as follows (see
Appendix B):
lp
2
11
2
sleaf
8
3
2
8
l
sleaf
l
sleaf
3
1
3
sleaf
l
l
1
3
l
1
sleaf
sleaf
l
3
1
3
sleaf
l
l
3
l
lp
1
2
lp
1
3
l
1
sleaf
l
3
1
16
l
sleaf
3
1
16
sleaf
8
3
sleaf
l
3
8
sleaf
3
sleaf
3
l
24
sleaf
3
24
sleaf
3
l
l
1
l
l
1
l
l
1
l
1
l
1
(3.15)
sleaf
3
sleaf
3
sleaf
3
sleaf
3
sleaf
3
lp
1
2
l
,
l
,
l
5
8
2
5
2
5
2
2
6
2
6
3
2
5
2
8
2
3
4
2
2
4
2
2
2
2
2
On the other hand, for Eq. (3.2), the following equation is obtained by
differentiating with respect to variable l2.
2
2
,
l
2
lp
11
,
l
2
,
l
2
lp
11
l
lg
1
l
2
lp
11
,
l
2
2
2
lg
1
,
l
2
lp
1
2
l
,
lp
13
2
2
,
l
2
2
lp
,
1
2
lp
13
l
,
2
l
2
2
lg
,
2
1
lp
13
l
2
2
,
l
2
l
2
lp
1
2
l
2
,
l
2
2
lp
13
,
l
2
(3.16)
2
,
lp
11
l
2
l
8
1
l
sleaf
1
3
l
2
2
lp
1
2
,
l
2
3
sleaf
l
2
3
5
l
sleaf
1
16
sleaf
6
,
l
2
2
lp
1
2
l
l
l
1
3
2
3
2
5
2
l
l
1
6
8
sleaf
sleaf
l
3
3
2
3
3
2
l
sleaf
1
16
sleaf
3
3
5
8
2
l
l
1
l
sleaf
,
lp
2
11
2
sleaf
8
2
8
sleaf
sleaf
3
4
2
l
sleaf
l
1
3
3
2
3
3
,
,
l
l
lp
1
1
lp
1
3
24
sleaf
l
sleaf
1
sleaf
l
l
1
sleaf
sleaf
l
24
1
sleaf
l
sleaf
l
l
3
5
2
3
4
3
8
2
3
2
3
l
2
Using the above equation, the following equation is derived.
2
lp
1
2
,
l
2
2
,
l
2
lp
1
3
l
2
2
3
2
(3.17)
(3.18)
l
2
,
lg
1
l
1
,
l
2
lg
1
l
2
The following equation that satisfies Eq. (3.18) (see Appendix A) is derived.
2
2
,
l
l
0,
lg
1
lg
1
Using the initial condition sleaf3(0)=0 and ∂sleaf3(0)/∂l=1, the function g(l1+l2, 0)
is written as follows:
(3.19)
0
3
2
3
l
l
1
sleaf
sleaf
l
41
3
1
3
l
sleaf
sleaf
sleaf
sleaf
l
2
sleaf
l
4
0
3
2
l
sleaf
sleaf
l
0
3
2
1
2
sleaf
l
l
4
0
1
4
2
3
3
4
2
2
3
2
3
2
0
3
sleaf
l
1
3
l
1
l
2
2
2
l
sleaf
2
sleaf
l
3
1
23
0
sleaf
3
4
0
lg
1
2
2
0,
l
l
sleaf
3
1
l
sleaf
l
41
3
1
2
l
l
sleaf
1
3
2
Using Eqs. (3.19) and (3.20), the following equation is obtained.
4
0
3
(3.20)
lg
1
,
l
2
2
l
2
sleaf
l
13
sleaf
sleaf
41
sleaf
sleaf
41
l
13
l
13
l
13
l
13
4
lg
1
sleaf
3
l
2
sleaf
3
sleaf
3
sleaf
3
3
4
l
0,
2
l
2
2
l
2
l
2
2
l
2
l
2
sleaf
3
sleaf
4
sleaf
l
13
sleaf
4
l
13
2
sleaf
l
1
2
sleaf
l
3
2
23
l
2
4
sleaf
l
2
3
l
13
sleaf
3
2
l
13
4
(3.21)
The differentiation ∂sleaf3(l1)/dl1 and ∂sleaf3(l2)/dl2 depends on the domains of
variables l1 and l2. It is necessary to classify domains. The additional formulas can
be summarized as follows.
(i)
In case both (4m-1)π3/2≦l1≦(4m+1)π3/2 and (4k-1)π3/2≦l2≦(4k+1)π3/2
or both (4m+1)π3/2≦l1≦(4m+3)π3/2 and (4k+1)π3/2≦l2≦(4k+3)π3/2 (Variables
m and k represent integers; for the constant π3, see [1][2].)
sleaf
3
2
l
l
1
sleaf
sleaf
41
2
3
3
l
1
l
1
4
3
2
6
l
sleaf
3
1
sleaf
41
3
sleaf
l
2
3
2
sleaf
l
sleaf
l
l
sleaf
1
3
2
3
4
2
sleaf
l
l
sleaf
4
3
2
1
23
sleaf
sleaf
l
l
2
1
3
3
2
sleaf
l
sleaf
l
4
1
4
3
2
3
2
3
3
sleaf
1
3
sleaf
l
1
2
2
6
l
1
l
4
2
3
(3.22)
(ii)
In case both (4m+1)π3/2≦l1≦(4m+3)π3/2 and (4k-1)π3/2≦l2≦(4k+1)π3/2
or both (4m-1)π3/2≦l1≦(4m+1)π3/2 and (4k+1)π3/2≦l2≦(4k+3)π3/2 (Variables
m and k represent integers)
sleaf
1
3
sleaf
l
1
2
2
6
l
1
l
4
2
3
(3.23)
sleaf
3
2
l
l
1
sleaf
sleaf
41
2
3
3
l
1
l
1
4
3
2
6
l
sleaf
3
1
sleaf
41
3
sleaf
l
2
3
2
sleaf
l
sleaf
l
l
sleaf
1
3
2
3
4
2
sleaf
l
l
sleaf
4
3
2
1
23
sleaf
sleaf
l
l
1
2
3
3
2
sleaf
sleaf
l
l
4
1
4
3
2
3
2
3
3
3.2 Leaf function: cleaf3(l1+l2)
With respect to arbitrary variables l1 and l2, the following equation is satisfied.
cleaf
3
2
l
l
1
sleaf
3
sleaf
41
3
2
l
1
l
2
4
l
2
sleaf
cleaf
l
3
3
1
l
2
cleaf
sleaf
l
l
41
1
2
3
3
3
cleaf
sleaf
cleaf
l
l
1
3
2
3
3
2
2
sleaf
l
l
cleaf
4
3
1
3
2
4
sleaf
3
l
2
2
sleaf
4
3
23
l
2
cleaf
3
4
l
1
2
l
1
cleaf
3
l
1
2
cleaf
3
l
2
4
l
1
(3.24)
The additional formulas of the leaf function cleaf3(l1 + l2) can be similarly proved
as described in section 3.1. The differentials ∂cleaf3(l1)/dl1 and ∂cleaf3(l2)/dl2
depend on the domains of variables l1 and l2. It is necessary to classify domains.
The additional formulas can be summarized as follows.
(i) In case both 2kπ3≦l1≦(2k+1)π3 and (4m-1)π3/2≦l2≦(4m+1)π3/2 or
(2k+1)π3≦l1≦(2k+2)π3 and (4m+1)π3/2≦l2≦(4m+3)π3/2 (Variables m and k
represent integers)
cleaf
3
2
l
l
1
sleaf
3
sleaf
41
3
2
3
l
1
l
2
4
cleaf
l
3
1
sleaf
41
3
cleaf
l
3
2
2
cleaf
l
3
1
6
l
sleaf
sleaf
l
1
3
2
2
3
4
2
l
cleaf
l
sleaf
4
3
1
2
3
23
sleaf
cleaf
l
l
1
2
3
3
2
l
sleaf
cleaf
4
3
3
4
l
1
2
1
l
2
cleaf
3
2
cleaf
3
2
6
l
1
l
1
4
(3.25)
(ii) In case both (2k+1)π3≦l1≦(2k+2)π3 and (4m-1)π3/2≦l2≦(4m+1)π3/2
or 2kπ3≦l1≦(2k+1)π3 and (4m+1)π3/2≦l2≦(4m+3)π3/2 (Variables m and k
represent integers)
cleaf
3
2
l
l
1
sleaf
sleaf
41
3
2
3
3
l
1
l
2
4
cleaf
l
3
1
sleaf
41
3
cleaf
l
3
2
2
cleaf
l
3
1
3
6
l
sleaf
sleaf
l
1
3
2
2
4
2
l
cleaf
l
sleaf
4
3
1
2
23
cleaf
sleaf
l
l
1
2
3
3
2
sleaf
l
cleaf
4
3
4
l
1
3
2
3
1
l
2
cleaf
3
2
cleaf
3
2
6
l
1
l
1
4
(3.26)
4 Numerical example
The numerical data of the leaf functions sleaf3(l) and cleaf3(l) are presented in
Table 1. Using Eqs. (3.22), (3.23), (3.25), and (3.26), the values of sleaf3(l1+ l2) and
cleaf3(l1+l2) can be calculated by using the values of sleaf3(l1), sleaf3(l2), cleaf3(l1),
and cleaf3(l2). As an example, consider l1 = 0.2 and l2 = 0.3.
6
3
4
3
2
6
3
4
2
sleaf
Then, sleaf3(0.2+0.3) can be obtained. We have to choose Eq. (3.22) for the
inequalities -π3/2≦0.2≦π3/2 and -π3/2≦0.3≦π3/2. By substituting l1= 0.2 and l2
= 0.3 into Eq. (3.22), we get the following equation:
sleaf
sleaf
sleaf
sleaf
2.0
13.0
3.0
12.0
3
3
3
3
2
2
4
sleaf
sleaf
sleaf
sleaf
3.0
2.0
41
4
2.0
3.0
3
3
23
sleaf
sleaf
sleaf
2.0
3.0
3.0
2.0
3
3
3
2
2
sleaf
sleaf
sleaf
3.0
4
3.0
2.0
2.0
199999
.0
299984
.0
299984
.0
1
4
2
2
299984
199999
.0
.04
.0
299984
199999
23
3
199999
299984
.0
299984
.0
.0
2
2
299984
.0
299984
.04
.0
199999
3.02.0
sleaf
sleaf
41
3
199999
.0
.041
199999
.0
199999
.041
(4.1)
2494431
.0
Therefore, the value of sleaf3(0.5) is obtained as follows:
sleaf
3
3.02.0
49944
2494431
sleaf
3
5.0
.0
.0
(4.2)
3
4
4
2
6
4
3
1
3
3
6
3
4
Table 1 Numerical data of leaf functions sleaf3(l) and cleaf3(l)
l
sleaf3(l)
0.000000
0.100000
0.199999
0.299984
0.399883
0.499443
0.598009
0.694183
0.785387
0.867486
0.934768
0.980708
0.999692
0.989090
0.950393
0.888560
0.810064
0.720972
0.625896
0.527828
0.428461
cleaf3(l)
1.000000
0.985184
0.942810
0.878184
0.797825
0.707632
0.611979
0.513647
0.414176
0.314304
0.214324
0.114325
0.014325
-0.085670
-0.185670
-0.285660
-0.385580
-0.485220
-0.583990
-0.680640
-0.772770
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
l
sleaf3(l)
0.328621
0.228649
0.128651
0.028651
-0.071350
-0.171350
-0.271340
-0.371280
-0.470980
-0.569930
-0.667000
-0.759970
-0.845200
-0.917390
-0.970090
-0.997240
-0.995140
-0.964110
-0.908270
-0.833880
-0.747280
cleaf3(l)
-0.856490
-0.926290
-0.975670
-0.998770
-0.992410
-0.957500
-0.898590
-0.822090
-0.734190
-0.639750
-0.541980
-0.442740
-0.342940
-0.242970
-0.142980
-0.042980
0.057024
0.157024
0.257019
0.356971
0.456727
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3.0
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
4.0
4.1
5 Conclusions
By using the analogy of the inverse trigonometric and lemniscate functions, higher
order of functions were artificially created as leaf functions. The following
conclusions can be drawn from this study:
・The artificially created leaf function r = sleaf3(l) corresponds to r = sin(l) and r
= sl(l), whereas the artificially created leaf function r = cleaf3(l) corresponds to r =
cos(l) and r = cl (l). The waves obtained from the trigonometric functions sin(l) and
cos(l) have a periodicity of 6.28… and an amplitude of 1, while the waves obtained
from the lemniscate functions have a periodicity of 5.24… and an amplitude of 1.
The artificially created functions sleaf3(l) and cleaf3(l) also produce regular waves
with 4.85 … periodicity and amplitude of 1. These functions produce waves
different from those of both the trigonometric and lemniscate functions. The prefix
"s" and "c" of these functions are classified by the following initial conditions: r(0)
= 0, dr(0)/dl = 1 and r(0) = 1, dr(0)/dl = 0.
・Additional formulas of leaf functions sleaf3(l1 +l2) and cleaf3(l1 +l2) were also
derived.
・The values of the leaf functions sleaf3(l1 +l2) can be numerically obtained using
the values of sleaf3(l1), sleaf3(l2), while those of cleaf3(l1 +l2) can be obtained using
values of sleaf3(l1), sleaf3(l2), cleaf3(l1), and cleaf3(l2) through the additional
formulas.
This study will be the first step in clarifying solution behaviors of nonlinear
equations consisting of both the second-order differentials and the higher order
polynomials.
References
[1] K. Shinohara, Special function: leaf function r=sleafn(l) (First report), Bulletin
of Daido University, 51 (2015), 23–38. http://id.nii.ac.jp/1277/00000049/
[2] K. Shinohara, Special function: leaf function r=cleafn(l) (Second report),
Bulletin of Daido University, 51 (2015), 39–68. http://id.nii.ac.jp/1277/00000050/
[3] Ivana Kovacic, Michael J. Brennan, The Duffing Equation: Nonlinear
Oscillators and their Behaviour, John Wiley & Sons, 2011.
https://doi.org/10.1002/9780470977859
[4] A. Beléndez, G. Bernabeu, J. Francés, D.I. Méndez, S. Marini, An accurate
closed-form approximate solution for the quintic Duffing oscillator equation,
Mathematical and Computer Modelling, 52 (2010), 637-641.
https://doi.org/10.1016/j.mcm.2010.04.010
[5] A. Kargar, M. Akbarzade, Analytical solution of nonlinear cubic-quintic
Duffing oscillator using global error minimization method, Advanced Studies in
Theoretical Physics, 6 (2012), 467–471.
[6] A. E. Zúñiga, Exact solution of the cubic-quintic Duffing oscillator, Applied
Mathematical Modelling, 37 (2013), 2574-2579.
https://doi.org/10.1016/j.apm.2012.04.005
[7] N. I. Akhiezer, Elements of the Theory of Elliptic Functions, (translations of
mathematical monographs), Vol. 79, American Mathematical Society, 1990.
[8] D. F. Lawden, Elliptic Functions and Applications, Applied Mathematical
Sciences, Vol. 80, Springer, 1989. https://doi.org/10.1007/978-1-4757-3980-0
[9] C. H. Liang, Notes on the ellipsoidal function, Springer, 2016.
[10] H. McKean and V. Moll, Elliptic Curves: Function Theory, Geometry,
Arithmetic, Cambridge University Press, 1999.
[11] R. Ranjan, Elliptic and Modular Functions from Gauss to Dedekind to Hecke,
Cambridge University Press, 2017. https://doi.org/10.1017/9781316671504
[12] G.C. Fagnano, Produzioni Matematiche Del Marchese Giulio Carlo De'
Toschi Di Fagnano, Pesaro Gavelli, 1750.
[13] L. Euler, Leonhardi Euleri Opera Omnia Series (cid:31)-(cid:31), Birkhaeuser, 1911.
[14] C. F. Gauss, Disquisitiones Arithmeticae, Yale University Press, 1965.
Appendix A
Let g(l1, l2) be a differentiable function. The necessary and sufficient condition for
satisfying g(l1, l2) = g(l1+l2,0) is that
holds. Function h(x,y) is
l
l
2
2
,
lg
1
l
1
defined as follows:
yxh
,
xg
xy
,
lg
1
l
,
2
y
(A.1)
x
y
x
y
2
(A.2)
l
1
l
By differentiating the above equation with respect to y, we obtain the following
equation.
yxh
,
y
xg
xy
,
l
1
(A.3)
xy
,
l
(A.4)
xg
xg
xg
l
2
y
y
y
y
y
2
l
1
y
xy
,
l
2
xy
,
l
1
Therefore, if the equation ∂g/∂l1=∂g/∂l2 holds, the following equation holds.
(A.5)
yxh
,
y
0
,
l
b
l
1
bxh
(A.6)
From the above equation, we find that h(x,y) is a function of x, not a function of y.
Therefore, the following equation holds for any constants a and b.
axh
,
Here, we set the following equation:
x
a
From the above two equations, the following equation is obtained.
axh
,
(A.8)
(A.9)
(A.7)
2
2
l
1
l
1
l
1
l
1
l
1
l
l
l
l
l
l
l
l
,
,
2
2
2
2
2
lg
1
,
2
2
2
2
1
lh
2
2
1
lg
2
2
2
2
bxh
,
l
2
l
1
,
1
lh
2
l
2
2
1
lg
2
l
2
l
1
l
2
l
1
,
l
2
l
1
2
2
l
2
2
lg
1
l
2
0,
(A.10)
The following equation is obtained by using Eqs. (A.6), (A.9), and (A.10).
,
l
2
lg
1
l
2
0,
(A.11)
xy
,
xg
Conversely, if the above equation holds, the following relational expression can
be obtained by using Eqs. (A.1) and (A.11).
yxh
,
We differentiate the above equation with y. Because g(2x, 0) is a function of x and
not a function of y, the following equation is obtained.
(A.12)
0,2
xg
y
yxh
,
xg
0,2
0
y
lg
1
y
yxh
,
y
xg
xg
y
xy
,
l
1
y
xy
,
l
1
xg
l
1
y
xy
,
l
2
xg
y
xy
,
l
2
l
2
y
y
0
Therefore, the following equation is obtained.
xg
xg
y
y
xy
,
l
1
xy
,
l
2
2
l
l
1
2
Here, we use the following equation:
x
y
Using Eq. (A.15), we obtain the following equation:
lg
1
l
1
lg
1
l
2
l
1
l
l
l
,
,
2
2
2
2
(A.13)
(A.14)
(A.15)
(A.16)
(A.17)
(A.18)
Appendix B
The first term of the Eq. (3.15) is transformed as follows:
lp
1
1
lp
2
1
l
l
,
,
2
2
sleaf
3
sleaf
4
l
2
l
1
3
l
1
sleaf
l
1
3
3
sleaf
l
2
3
2
4
l
2
l
1
sleaf
l
2
sleaf
l
1
3
3
sleaf
l
2
6
3
sleaf
l
2
3
1
sleaf
3
l
1
3
sleaf
6
l
1
3
sleaf
l
2
3
l
2
sleaf
4
l
2
4
sleaf
3
l
1
3
3
l
2
3
6
sleaf
l
l
1
sleaf
l
1
2
3
2
,
l
lp
1
2
l
1
l
l
2
2
3
2
2
l
1
2
,
l
lp
1
1
l
1
sleaf
l
1
l
1
sleaf
3
3
sleaf
l
2
sleaf
l
sleaf
l
1
3
l
1
3
l
2
6
2
l
l
1
3
1
3
sleaf
6
l
1
3
sleaf
3
sleaf
l
1
3
sleaf
3
sleaf
l
1
3
sleaf
l
1
3
2
sleaf
l
sleaf
l
1
sleaf
3
3
l
1
l
1
sleaf
l
2
6
3
3
sleaf
sleaf
sleaf
l
1
3
sleaf
4
3
l
1
l
3
2
l
2
4
sleaf
6
l
1
3
sleaf
l
2
3
3
4
sleaf
l
1
sleaf
l
l
3
2
3
2
The following equation is applied [1] [2].
l
1
(B.1)
sleaf
3
l
1
sleaf
3
6
l
(B.2)
d
dl
|
1910.13285 | 1 | 1910 | 2019-10-29T14:24:31 | Boundedness of operators on certain power-weighted Morrey spaces beyond the Muckenhoupt weights | [
"math.FA"
] | We prove that for operators satistying weighted inequalities with $A_p$ weights the boundedness on a certain class of Morrey spaces holds with weights of the form $|x|^\alpha w(x)$ for $w\in A_p$. In the case of power weights the shift with respect to the range of Muckenhoupt weights was observed by N.~Samko for the Hilbert transform, by H.~Tanaka for the Hardy-Littlewood maximal operator, and by S.~Nakamura and Y.~Sawano for Calder\'on-Zygmund operators and others. We extend the class of weights and establish the results in a very general setting, with applications to many operators. For weak type Morrey spaces, we obtain new estimates even for the Hardy-Littlewood maximal operator. Moreover, we prove the necessity of certain $A_q$ condition. | math.FA | math |
BOUNDEDNESS OF OPERATORS ON CERTAIN
POWER-WEIGHTED MORREY SPACES BEYOND THE
MUCKENHOUPT WEIGHTS
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
Abstract. We prove that for operators satistying weighted in-
p weights the boundedness on a certain class of
equalities with A
Morrey spaces holds with weights of the form xαw(x) for w ∈ A
p.
In the case of power weights the shift with respect to the range of
Muckenhoupt weights was observed by N. Samko for the Hilbert
transform, by H. Tanaka for the Hardy-Littlewood maximal oper-
ator, and by S. Nakamura and Y. Sawano for Calder´on-Zygmund
operators and others. We extend the class of weights and establish
the results in a very general setting, with applications to many
operators. For weak type Morrey spaces, we obtain new estimates
even for the Hardy-Littlewood maximal operator. Moreover, we
prove the necessity of certain A
q condition.
1. Introduction
For 1 ≤ p < ∞ and 0 ≤ λ < n, let the Morrey space Lp,λ(w) be the
collection of all measurable functions f such that
(1.1)
kf kLp,λ(w) := sup
x∈Rn,r>0(cid:18) 1
rλZB(x,r)
f pw(cid:19)1/p
< ∞.
We also consider the weak Morrey space W Lp,λ(w), for which
kf kW Lp,λ(w) :=
x∈Rn,r>0,t>0(cid:18)tp w({y ∈ B(x, r) : f (y) > t})
sup
rλ
(cid:19)1/p
< ∞.
(Here and in what follows w(A) stands for the integral of w over A.)
Clearly, Lp,λ(w) ⊂ W Lp,λ(w).
N. Samko proved in [7] that the Hilbert transform is a bounded
operator on Lp,λ(xα) for 0 < λ < 1 and λ − 1 < α < λ + p − 1. This
2010 Mathematics Subject Classification. Primary .
Key words and phrases. Morrey spaces, Muckenhoupt weights, Hardy-
Littlewood maximal operator, Calder´on-Zygmund operators.
The first author is supported by the grants MTM2014-53850-P of the Minis-
terio de Econom´ıa y Competitividad (Spain) and grant IT-641-13 of the Basque
Gouvernment.
1
2
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
range of values of α shows a shift with respect to the corresponding
range in the Ap class, which is −1 < α < p − 1. In [10], H. Tanaka
explored the boundedness on Lp,λ(w) of the Hardy-Littlewood maximal
operator and was able to describe necessary conditions and sufficient
conditions, but not a characterization. Nevertheless, for power weights
w(x) = xα he obtained the sharp range λ − n ≤ α < λ + n(p − 1),
which in the one-dimensional case coincides with the range obtained
by Samko for the Hilbert transform except at the endpoint α = λ − n.
Later on, S. Nakamura and Y. Sawano in [6] studied the boundedness of
the Riesz transforms and other singular integrals and obtained similar
shifted ranges for the case of Lp,λ(xα) (with open left endpoint).
In [3] the authors of this paper proved a general result involving
Muckenhoupt weights, under the assumptions of the extrapolation the-
orem for Ap weights. When particularized for the Hardy-Littlewood
maximal operator or for Calder´on-Zygmund operators, the bounded-
ness on Lp,λ(w) was obtained for w ∈ Ap ∩ RHσ in the range 0 ≤ λ ≤
n/σ′, which for w(x) = xα gives the range λ − n ≤ α < n(p − 1).
In this paper the results in [3] for Lp,λ(w) spaces are extended to the
weights xαw(x) for w ∈ Ap ∩ RHσ and 0 < α < λ. (We recall the
definitions of the weight classes Ap and the reverse Holder classes RHσ
in Section 2.) In particular, we prove the following theorem.
Theorem 1.1. Let 1 ≤ p0 < ∞ and let F be a collection of nonnegative
measurable pairs of functions. Assume that for every (f, g) ∈ F and
every w ∈ Ap0 we have
(1.2)
kgkLp0 (w) ≤ Ckf kLp0 (w),
where C does not depend on the pair (f, g) and it depends on w only
(defined at the beginning of Section 2). Then for
in terms of [w]Ap0
1 < p < ∞ and w ∈ Ap ∩ RHσ it holds
(1.3)
kgkLp,λ(xαw) ≤ Ckf kLp,λ(xαw),
for 0 ≤ λ < n/σ′ and 0 ≤ α < λ. In particular, for power weights of
the form xβ = xαw(x), the estimate (1.3) holds for λ − n < β <
λ + n(p − 1). If the hypothesis holds for p0 = 1, then (1.3) also holds
for p = 1, and in the case of power weights for λ − n < β < λ.
When we say that (1.2) holds for every w ∈ Ap0 we mean that if the
right-hand side is finite for a fixed w, then also the left-hand side is
finite for the same w and the inequality holds. The conclusion of the
if f is in Lp,λ(xαw),
theorem is to be understood in the same way:
then g is in the same space and the inequality holds.
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
3
To make things clear let us say that the weights appearing in the the-
orem are always in some Muckenhoupt class. Indeed, xαw ∈ Ap+λ/n
for 0 ≤ α < λ and w ∈ Ap. Being in Ap+λ/n is not a particular restric-
tion for the weights in our theorem, because we show in Proposition
5.2 the necessity of u ∈ Ap+λ/n for the boundedness of the Hardy-
Littlewood maximal operator in the Morrey space Lp,λ(u) (and even
for the weak-type boundedness). We are thus forced to deal with Muck-
enhoupt weights. When we say that the results go beyond the Muck-
enhoupt range, we mean that for a fixed value of p the boundedness
of the involved operators holds for weights which are not necessarily in
Ap.
This theorem has a number of applications because we know that
for many operators T , the pairs (f , T f ) satisfy its assumptions. In
particular, we recover the results for the Hardy-Littlewood maximal
operator (except the left endpoint, that is, β = λ − n), the Hilbert
transform and the Calder´on-Zygmund operators mentioned above. But
it extends also to Littlewood-Paley operators, rough singular integrals
and others. Moreover, in those examples the case p = 1 of the theorem
provides a weak-type result, from L1,λ(xβ) to W L1,λ(xβ) for λ − n <
β < λ, which for β > 0 is new even for the Hardy-Littewood maximal
operator.
We present some preliminary results in Section 2. The proof of The-
orem 1.1 is in Section 3, where we also prove another theorem suited
to operators satisfying the assumptions of the so-called limited range
extrapolation. In Section 4 we establish embeddings which allow to de-
fine the operators in the Morrey spaces by restriction. In Section 5 we
prove the necessity of the Ap+λ/n condition for the Hardy-Littlewood
maximal operator and of Ap+λ for the Hilbert transform in the case of
weak-type estimates. This implies the necessity of the range of power
weights for positive exponents. We also give an easy proof of the ne-
cessity for negative exponents. In the case of the strong estimates and
power weights this was proved by Tanaka in [10] checking his more
general necessary condition. We extend the necessity to the weak-type
estimates. In the same section we prove the estimate for the left end-
point for power weights (that is, for the weight xλ−n).
2. Preliminary results
Let w ∈ Lloc
1 (Rn) with w > 0 almost everywhere. We say that w is
a Muckenhoupt weight belonging to Ap for 1 < p < ∞ if
[w]Ap ≡ sup
B
w(B)
B (cid:18)w1−p′(B)
B (cid:19)p−1
< ∞,
4
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
where the supremum is taken over all Euclidean balls B in Rn. The
quantity [w]Ap is the Ap constant of w. We say that w belongs to A1
if, for any Euclidean ball B,
w(B)
B
≤ cw(x) for almost all x ∈ B.
The A1 constant of w, denoted by [w]A1, is the smallest constant c for
which the inequality holds.
We say that a nonnegative locally integrable function w on Rn be-
longs to the reverse Holder class RHσ for 1 < σ < ∞ if it satisfies the
reverse Holder inequality with exponent σ, i.e.,
(cid:18) 1
BZB
σ
w(x)σdx(cid:19) 1
≤
c
BZB
w(x)dx,
where the constant c is universal for all Euclidean balls B ⊂ Rn.
Remark 2.1. Some results for weights are the following. We use repeat-
edly the first two throughout the paper.
(i) If Mh < ∞ a.e., then (Mh)1/s ∈ A1 and its A1 constant de-
pends on s, but not on h. Moreover, (Mh)1/s ∈ A1 ∩ RHσ if
s > σ.
(ii) Let w ∈ RHσ. For any ball B and any measurable E ⊂ B it
holds that
(2.1)
w(E)
w(B)
≤ c(cid:18)E
B(cid:19)1/σ′
.
Since w ∈ Ap implies that w ∈ RHσ for some σ, the inequality
holds for each Ap weight for the appropriate σ.
(iii) Weights simultaneously in Ap and RHσ can be described ([4])
as
(2.2)
Ap ∩ RHσ = {w : wσ ∈ Aσ(p−1)+1}.
Remark 2.2. When dealing with the definition of the norm (1.1), we
only need to take into account two types of balls: balls centered at the
origin and balls of the form B(x, r) with r < x/4. Indeed, if we have a
ball B(x, r) with r ≥ x/4, it holds that B(x, r) ⊂ B(0, 5r), and since
the radii are comparable we can replace the smaller ball by the larger
one.
The following lemma provides an estimate which is used in the proofs
of the theorems.
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
5
Lemma 2.3. Let 1 ≤ p < ∞ and 0 ≤ λ < n. Let f ≥ 0 in Lp,λ(xαw)
and r > 0. If α < λ, then
(2.3)
(cid:18)ZB(0,r)
p
f pw(cid:19) 1
≤ Cr
λ−α
p kf kLp,λ(xαw).
The constant depends only on α, λ and p.
Proof. Let Aj = B(0, 2−j+1r) \ B(0, 2−jr), j ∈ N. Then
ZB(0,r)
f pw ≤ C
≤ C
∞Xj=1ZAj
∞Xj=1
f (y)pw(y)(cid:18) y
2−jr(cid:19)α
dy
(2−jr)λ−αkf kp
Lp,λ(xαw).
If α < λ the series is convergent and (2.3) follows.
(cid:3)
Remark 2.4. We deal with integrals of the typeRB f pw where w is a
certain Ap weight and B is a ball. This can be written asR f pwχB.
But wχB cannot be an Ap weight for any p because Ap-weights cannot
vanish in a set of positive measure. Hence, the second proof of Theorem
6.1 of [1] is not correct, because it claims that rλ−nχB(x,r) is an A1
weight.
3. Main theorems
In this section we first prove Theorem 1.1, starting with the assump-
tion for p0 = 1. This case is important because the proof of the theorem
is simpler and the general case p0 ∈ (1, ∞) can be reduced to this one by
a scaling argument. In the applications it is also significant because for
a number of operators weighted weak-type (1, 1) estimates are known
and our result provides weak-type Morrey estimates for them.
Proof of Theorem 1.1. Case p0 = 1. We assume first that (1.2) holds
with p0 = 1.
Let Br := B(x, r) be one of the balls considered in Remark 2.2. Let
w ∈ Ap. We have
(3.1)
ZBr
g(y)pyαw(y)dy ≤
gpw,
rαZBr
CxαZBr
if x = 0;
gpw,
if 0 < r ≤ x/4.
6
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
Proof for 1 < p < ∞. In both cases we are left with the integral of
gpw on the ball Br, which we handle as in [3]. Using duality we have
p
gpw(cid:19) 1
(cid:18)ZBr
ghw ≤ZRn
ZBr
=
sup
Lp′
(w,Br )
h : khk
ghw.
=1ZBr
s ≤ cZRn
1
Fix such a function h and we have for s > 1 that
(3.2)
gM(hswsχBr )
f M(hswsχBr )
1
s ,
1
provided that M(hswsχBr )
s ∈ A1. According to Remark 2.1 (i) we
need M(hswsχBr )(x) < ∞ a.e. We check that hswsχBr ∈ L1 for ap-
propriate s > 1, and get a bound for its integral for future use. To this
end, we choose s > 1 such that w1−p′ ∈ Ap′/s, which is possible because
w1−p′ ∈ Ap′. We have
(cid:18)ZBr
s
hsws−1w(cid:19) 1
≤(cid:18)ZBr
(3.3)
hp′
w(cid:19) 1
p′(cid:18)ZBr
(Br)− 1
w(s−1) p′
p′−s
s − 1
p′
+1(cid:19) 1
≤ c Br
1
s w1−p′
p′ ≤ c w(Br)
1
p r− n
s′ ,
where the second inequality holds because w1−p′ ∈ Ap′/s (the exponent
of w in the integral is the same as (1 − p′)(1 − (p′/s)′)) and in the last
one we use
cnrn = Br ≤ w(Br)
1
p w1−p′
(Br)
1
p′ .
We split the last integral of (3.2) into the integral over B2r and over
its complement. On the one side we have
ZB2r
f M(hswsχBr )
1
s ≤(cid:18)ZB2r
f pw(cid:19) 1
p(cid:18)ZB2r
M(hswsχBr )
p′
p′
s w1−p′(cid:19) 1
.
The last term is bounded by a constant because M is bounded on
Lp′/s(w1−p′) and we get a constant times the norm of h in Lp′(w, Br),
which is 1. Now we have
f pw ≤
(3.4)
ZB2r
C(2r)λ−αkf kp
Lp,λ(xαw),
Cx−α(2r)λkf kp
Lp,λ(xαw),
if x = 0;
if 0 < r ≤ x/4;
where in the first case we use (2.3), and in the second case we use
x ∼ y for y ∈ B2r.
To deal with the integral on Rn \ B2r we decompose it into annuli
and use that on B2j+1r \ B2j r the maximal operator is comparable to
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
7
(2jr)−nRBr
hsws. Using (3.3) we get
∞Xj=1ZB2j+1 r\B2j r
1
p
w(Br)
2jn/srn
f
(3.5)
≤
≤
f pw! 1
∞Xj=1 ZB2j+1 r\B2j r
f pw! 1
∞Xj=1 ZB2j+1 r
p
jn
s′ 2− jn
pσ′ ,
2
p w1−p′ (B2j+1r)
2jn/srn
1
p′ w(Br)
1
p
where in the last step we use
w (B2j+1r)
1
p w1−p′
(B2j+1r)
1
p′ ≤ CB2j+1r = C ′2jnrn,
and (2.1) for Br and B2j+1r.
If Br is centered at the origin, we use (2.3) in the last term of (3.5)
to get
(3.6)
C
∞Xj=1
2j[(λ−α− n
σ′ ) 1
p + n
s′ ]r(λ−α) 1
p kf kLp,λ(xαw).
The series is convergent if we choose s close enough to 1, because we
assume λ < n/σ′.
If the ball is centered at x and r ≤ x/4, we can assume x = 2N r
for some N ≥ 2, increasing slightly r if necessary. For the integral in
(3.5) we distinguish the cases j ≤ N −2 and j ≥ N −1. In the first case,
if y ∈ B2j+1r, then y ∼ x. In the second case, B2j+1r ⊂ B(0, 2j+2r).
As a consequence, for j ≤ N − 2,
ZB2j+1 r
f pw ≤ Cx−αZB2j+1 r
f (y)pw(y)yαdy
≤ Cx−α(2jr)λkf kp
Lp,λ(xαw).
For j ≥ N − 1 we use (2.3) and we obtain
f pw ≤ C(2jr)λ−αkf kp
Lp,λ(xαw).
f pw ≤ZB(0,2j+2r)
ZB2j+1 r
kf kLp,λ(xαw) N −2Xj=1
Inserting this into (3.5) we obtain a constant times
x− α
p r
λ
p 2j[(λ− n
σ′ ) 1
p + n
s′ ] +
r(λ−α) 1
p 2j[(λ−α− n
σ′ ) 1
p + n
s′ ]! .
∞XN −1
8
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
The first sum is bounded independently of N if we take s close enough
to 1, because λ < n/σ′. For this same reason the series is convergent
and the value of its sum is
s′ ]r(λ−α) 1
s′ ] ≤ C ′x− α
p r
p = Cx− α
p r
C2N [(λ−α− n
σ′ ) 1
p + n
λ
p 2N [(λ− n
σ′ ) 1
p + n
λ
p ,
where C ′ is independent of N because the exponent of 2N in the middle
term is negative. Taking into account (3.4), (3.6) and the recent bound,
the right-hand side of (3.1) is bounded as desired.
Proof for p = 1. The proof is similar but easier because we do not
need to use a duality argument and there is no h as in the previous
situation.
We have w ∈ A1 ∩ RHσ (which is the same as wσ ∈ A1). In the
construction of the A1 weight M(wsχBr )1/s we choose 1 < s < σ.
Then M(wsχBr )1/s ≤ M(ws)1/s ≤ w a.e. and when we integrate on
B2r we obtain (3.4) with p = 1.
When we are in B2j+1r \ B2j r we have
M(wsχBr )1/s ≤ C(cid:18)ws(Br)
(2jr)n(cid:19)1/s
.
Using that w ∈ RHs (because s < σ), ws(Br)1/s ≤ CBr−1/s′w(Br).
Instead of (3.5) we have now
∞Xj=1ZB2j+1 r\B2j r
f w 2jn/s′ w(Br)
w(B2j+1r)
≤
∞Xj=1ZB2j+1 r\B2j r
The proof continues as before.
f w 2j( n
s′ − n
σ′ ).
Case p0 > 1. By the usual extrapolation theorem the assumption
(1.2) is valid for any p0 ∈ (1, ∞). Given p > 1 and w ∈ Ap ∩ RHσ, we
choose p0 > 1 for which w ∈ Ap/p0. The assumption holds in the form
kgp0kL1(v) ≤ Ckf p0kL1(v),
for v ∈ A1 ⊂ Ap0 and we can apply the previous part of the proof to
the pair (f p0, gp0) to get the Morrey estimate with exponent p/p0. (cid:3)
We can generalize Theorem 1.1 to a setting in which weighted in-
equalities is a restricted range are assumed.
Theorem 3.1. Let 1 < b < ∞ and 1 ≤ p0 < b. Let F be a collection
of nonnegative measurable pairs of functions. Assume that for every
(f, g) ∈ F and every w ∈ Ap0 ∩ RH(b/p0)′ we have
(3.7)
kgkLp0 (w) ≤ Ckf kLp0 (w),
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
9
where C does not depend on the pair (f, g) and it depends on w only
in terms of the Ap0 and RH(b/p0)′ constants of w. Then if 1 < p < b
and w ∈ Ap ∩ RHσ with σ > (b/p)′, it holds that
kgkLp,λ(xαw) ≤ Ckf kLp,λ(xαw),
for 0 ≤ λ < n( 1
of the form xβ = xαw(x) and for
σ′ − p
b ) and 0 ≤ α < λ. In particular, for power weights
it holds that
λ − n(cid:16)1 −
p
b(cid:17) < β < λ + n(p − 1),
kgkLp,λ(xβ) ≤ Ckf kLp,λ(xβ).
Moreover, if the hypothesis holds for p0 = 1, then the results are valid
for p = 1.
Proof. The proof is similar to that of Theorem 1.1. Starting with p0 =
1, to apply the hypothesis in (3.2) we need to assume s > b′ because the
weight has to be in A1 ∩RHb′. The estimate (3.3) and the boundedness
of M in Lp′/s(w1−p′) need w1−p′ ∈ Ap′/s. We are assuming that w ∈
Ap∩RHσ with σ > (b/p)′, in particular, w ∈ Ap∩RH(b/p)′. According to
(2.2) this is the same as saying w(b/p)′ ∈ A(b/p)′(p−1)+1, which by duality
yields w(b/p)′(1−q′) ∈ Aq′ with q = (b/p)′(p−1)+1, that is, w1−p′ ∈ Ap′/b′.
Then there exists s > b′ for which w1−p′ ∈ Ap′/s as needed. The proof
continues as before, and we only need to add the condition that makes
the series convergent. This condition is λ < n( 1
If we assume p0 > 1 in (3.7), by the usual extrapolation theorem we
can consider any p0 ∈ (1, b). Given p and w we proceed again as before
by choosing p0 close enough to 1 such that w ∈ Ap/p0 and working with
the pairs (f p0, gp0).
(cid:3)
σ′ − p
b ).
The formulation of the extrapolation theorem in terms of pairs of
functions provides several extensions as corollaries (see [2, p. 21 -- 22]).
In a similar way, we can get similar extensions in the Morrey set-
ting. We state the scaling and weak-type extensions in the following
two corollaries, and leave to the interested reader the extension to the
vector-valued setting.
Corollary 3.2. Let 0 < p− ≤ p0 < p+ ≤ ∞. Let F be a collection
of nonnegative measurable pairs of functions. Assume that for every
(f, g) ∈ F and every w ∈ A p0
p−
∩ RH(cid:16) p+
p0 (cid:17)′ we have
(3.8)
kgkLp0 (w) ≤ Ckf kLp0 (w),
10
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
where C does not depend on the pair (f, g) and it depends on w only in
terms of the A p0
p0 (cid:17)′ constants of w. Then if p− < p < p+
p−
and RH(cid:16) p+
and w ∈ A p
p−
∩ RHσ with σ >(cid:16) p+
p(cid:17)′
, it holds that
kgkLp,λ(xαw) ≤ Ckf kLp,λ(xαw),
for 0 ≤ λ < n( 1
of the form xβ = xαw(x) and for
σ′ − p
p+
) and 0 ≤ α < λ. In particular, for power weights
λ − n(cid:18)1 −
p
p+(cid:19) < β < λ + n(
p
p−
− 1),
kgkLp,λ(xβ) ≤ Ckf kLp,λ(xβ).
it holds that
Moreover, if the hypothesis holds for p0 = p−, then the conclusion is
valid for p = p−.
If p+ = ∞ this is a corollary to Theorem 1.1 and if p+ < ∞ to
Theorem 3.1. The proof is immediate from the theorems if we read
(3.8) as
kgp−kLfp0 (w) ≤ Ckf p−kLfp0 (w),
and b = p+
p−
.
for every w ∈ Aep0 ∩ RH(b/ ep0)′ with ep0 = p0
Corollary 3.3. If in Theorem 1.1, Theorem 3.1 or Corollary 3.2 the
assumptions hold as weak-type inequalities, that is, with kgkLp0,∞(w) in-
stead
kgkLp0(w), then the conclusions also hold in the weak sense, that is,
with kgkW Lp,λ(xαw) instead of kgkLp,λ(xαw).
of
p−
To prove this case, the weak-type hypothesis can be read as a strong
type inequality for the pair (f, tχ{g>t}) with constants uniform in t.
4. Embeddings and applications
The proof of Theorem 1.1 shows that
ZRn
f M(hswsχBr )
1
s ≤ Ckf kLp,λ(xαw),
for 0 ≤ λ < n, w ∈ Ap and appropriate s > 1. This implies the
continuous embedding
Lp,λ(xαw) ֒→ L1(M(hswsχBr )
1
s ).
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
11
In particular, choosing the ball B(0, 1) and h = cw−1, we have
Lp,λ(xαw) ֒→ L1((1 + x)−β) for some β < n. Since M(hswsχBr )
s ∈
A1, we have
1
Lp,λ(xαw) ⊂ [u∈A1
L1(u).
By the scaling argument at the end of the proof of the same theorem,
if p > 1,
(4.1)
Lq(u),
q > 1.
Lp,λ(xαw) ⊂ [u∈Aq
(The right-hand side is independent of q; see [5] and [3].) Suitable em-
beddings can be written for the weighted spaces appearing in Theorem
3.1, hence in Corollary 3.2.
The applications to boundedness of operators in Morrey spaces are
corollaries of the general theorems and of the embeddings just men-
tioned. For instance, the basic one is the following.
Corollary 4.1. Assume that for some p0 ∈ [1, ∞), T is an operator
Lp0(w) into the space of measurable functions and
acting fromSw∈Ap0
satisfying
kT f kLp0 (w) ≤ Ckf kLp0 (w)
(4.2)
for all f ∈ Lp0(w) and w ∈ Ap0, with a constant depending on [w]Ap0
.
Then for every 1 < p < ∞ (and also p = 1 if p0 = 1), w ∈ Ap ∩ RHσ,
0 ≤ λ < n/σ′ and 0 < α < λ, we have that T is well defined on
Lp,λ(xαw) by restriction and, moreover,
(4.3)
kT f kLp,λ(xαw) ≤ Ckf kLp,λ(xαw).
For power weights xβ, T is well defined and bounded on Lp,λ(xβ) if
λ − n < β < λ + n(p − 1).
If (4.2) is replaced by the weak estimate from L1(w) to L1,∞(w) for
w ∈ A1, then (4.3) holds from L1,λ(xαw) to W L1,λ(xαw).
The definition by embedding is guaranteed by (4.1) and the size
estimate by Theorem 1.1.
There are many operators satisfying the assumptions of the theo-
rem: the Hardy-Littlewood maximal operator, Calder´on-Zygmund op-
erators, rough operators with kernel x−nΩ(x/x) with Ω ∈ L∞(Sn−1)
and integral zero, commutators (in this case the weighted weak-type
(1, 1) does not hold), square functions (including some Littlewood-
Paley type operators, Lusin area integral, gλ functions, Marcinkiewicz
integral), etc.
12
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
Similar corollaries can be written for our other general results. All
the applications mentioned in [3] for the spaces Lp,λ(w) are now ex-
tended to Lp,λ(xαw) with 0 ≤ α < λ by the theorems in this paper.
Note that in [3] the space Lp,ν(w) was denoted as Lr
p(λ, w, Rn) with
n + rp = ν.
5. Necessary conditions for M and H, and the endpoint
for M
Tanaka proved in [10] that M is bounded on Lp(xβ) if and only
if λ − n ≤ β < λ + n(p − 1). The necessity of the upper bound
means that 0 ≤ α < λ in Theorem 1.1 is optimal. Tanaka's proof
uses a general necessary condition involving duality in Morrey spaces.
Avoiding duality, we prove first a necessary condition in terms of the
Aq scale from which the necessity of the upper bound follows, and next
the necessity of the lower bound is proved in a direct way. In all cases
our necessary conditions are valid for the weak-type (p, p) estimates.
The sufficiency for λ − n < β < λ + n(p − 1) comes from Theorem 1.1.
We give in Proposition 5.7 below a direct proof of the boundedness at
the endpoint Lp,λ(xλ−n) for 1 < p < ∞ and the corresponding weak
estimate for p = 1. The weak estimates are not in [10].
pn
Lemma 5.1. Let 1 ≤ p < ∞ and 0 ≤ λ < n. The embedding
n−λ ) ֒→ Lp,λ(w) holds with constant depending only on n, λ
L
and p, not on w.
n−λ (w
n
Proof. Let B be a ball of radius r. Then
1
rλZB
f pw ≤
f
pn
n−λ w
1
rλ(cid:18)ZB
n
n
n−λ(cid:19)1− λ
B
λ
n ≤ cn,λkf kp
L
pn
n−λ (w
n
n−λ )
. (cid:3)
Proposition 5.2. Let 1 ≤ p < ∞ and 0 ≤ λ < n. If M is bounded
from Lp,λ(w) to W Lp,λ(w), then w ∈ Ap+λ/n.
Proof. Let B be a ball of radius r. Define f = σχB with σ nonnegative
to be chosen later. For x ∈ B, we have Mf (x) ≥ σ(B)/B. If t <
σ(B)/B, then B = {x ∈ B : Mf (x) > t}. Assuming that M is
bounded from Lp,λ(w) to W Lp,λ(w) we have
tw(B)1/p
rλ/p ≤ CkσχBkLp,λ(w) ≤ C ′kσχBk
pn
n−λ (w)
L
= C ′(cid:18)ZB
pn
n−λ w
σ
n
pn
n−λ(cid:19) n−λ
,
where we used Lemma 5.1 in the second inequality. Let t tend to
σ(B)/B and choose σ such that σ = σ
n = w.
n−λ , that is, σ1−p− λ
n−λ w
pn
n
ZB(0,δ)
xβdx ∼ δβ+n,
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
13
We get
w(B)σ(B)p+ λ
n −1
Bp+ λ
n
≤ C,
with a constant independent of B. Therefore, w ∈ Ap+λ/n.
To be precise, we do not know a priori that σ(B) is finite for the
n = w. As usual, to overcome this problem, we define σǫ
(cid:3)
= w + ǫ for ǫ > 0 and let ǫ tend to 0.
choice σ1−p− λ
1−p− λ
by σ
n
ǫ
Proposition 5.3. Let 1 ≤ p < ∞ and 0 ≤ λ < n. If M is bounded
from Lp,λ(xβ) to W Lp,λ(xβ), then λ − n ≤ β < λ + n(p − 1).
Proof. According to the previous proposition, xβ ∈ Ap+λ/n and this
requires β < λ + n(p − 1).
For the lower bound, first we observe that if the characteristic func-
tion of a ball centered at the origin is in W Lp,λ(xβ), then λ − n ≤ β.
Indeed, let δ be small and t < 1. Then for every x ∈ B(0, δ), the
function is bigger than t at x. Since
we want δβ+n ≤ Kδλ for small δ. Therefore, β + n ≥ λ is necessary.
Let f be the characteristic function of the ball centered at 1 with
radius 1/2. This function is clearly in Lp,λ(xβ) for any β and for
0 ≤ λ < n. The maximal operator acting on f satisfies Mf (x) ≥ c
for some c > 0 and x ∈ B(0, 1). Consequently, Mf /∈ W Lp,λ(xβ) for
β < λ − n.
(cid:3)
A direct proof of the necessity of β < λ + n(p − 1) is obtained as
follows. For β ≥ λ + n(p − 1), the function x−nχB(0,1) is in Lp,λ(xβ)
and is not locally integrable. Therefore, β < λ + n(p − 1) is necessary.
Proposition 5.4. Let 1 ≤ p < ∞ and 0 ≤ λ < 1.
transform H is bounded from Lp,λ(w) to W Lp,λ(w), then w ∈ Ap+λ.
If the Hilbert
Proof. First we observe that for characteristic functions of sets the
norms in Lp,λ(w) and W Lp,λ(w) are the same.
Given an interval I, let I ′ be the adjacent interval of the same length,
placed at the right of I. Note that for x ∈ I ′, H(χI)(x) > 1/(2π).
Assuming the weak boundedness of H we have
kχI ′kW Lp,λ(w) ≤ k2πH(χI)χI ′kW Lp,λ(w)
≤ 2πkH(χI)kW Lp,λ(w) ≤ CkχIkLp,λ(w).
Interchanging the role of I and I ′ we deduce that kχIkW Lp,λ(w) and
kχI ′kW Lp,λ(w) are comparable.
≤ Ck2πH(σχI)χI ′kW Lp,λ(w)
≤ 2πCkH(σχI)kW Lp,λ(w) ≤ C ′kσχIkLp,λ(w).
σ(I)
I
w(I)1/p
Iλ/p ≤ C(cid:13)(cid:13)(cid:13)(cid:13)
Using Lemma 5.1 we get
σ(I)
I
χI ′(cid:13)(cid:13)(cid:13)(cid:13)W Lp,λ(w)
Iλ/p ≤ C ′′(cid:18)ZI
w(I)1/p
σ(I)
I
14
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
For σ to be chosen later, we observe that H(σχI)(x) > σ(I)(2πI)−1
for x ∈ I ′. Also
w(I)
Iλ ≤ kχIkp
W Lp,λ(w) ≤ CkχI ′kp
W Lp,λ(w),
where the first inequality holds by the definition of the norm. Then
p
1−λ w
σ
p
1
1−λ(cid:19) 1−λ
p
,
1
1−λ .
(cid:3)
from which w ∈ Ap+λ follows if we choose σ = σ
1−λ w
Remark 5.5. In the case of the Hilbert transform one could prefer to
assume that it is defined a priori only for Schwartz functions through
the principal value formula. The proof given here can be adapted to
such assumption by approximating the involved functions by smooth
ones. Without affecting the proof one can take the intervals I and I ′
separated by a distance equal to their length, instead of taking them
adjacent, so that there is some room for the approximation.
The proof of Proposition 5.4 can be adapted to higher dimensions to
obtain the necessity of the condition Ap+λ/n for the Riesz transforms
and other singular integral operators satisfying a nondegeneracy con-
dition (see [9, Chapter V, §4.6] for the similar result in the Lebesgue
setting).
Proposition 5.6. Let 1 ≤ p < ∞ and 0 ≤ λ < 1. If H is bounded
from Lp,λ(xβ) to W Lp,λ(xβ), then λ − 1 < β < λ + p − 1.
Proof. The condition β < λ + p − 1 is a consequence of xβ ∈ Ap+λ as
required by the previous proposition.
As in the proof for the maximal operator in Proposition 5.3, the
estimate H(χ1,2)(x) ≥ 1/(2π) for x ∈ (0, 1) is enough to get β ≥ λ − 1.
To rule out the Morrey estimate for β ≥ λ − 1 we consider χ(0,1),
which is in Lp,λ(xλ−1). A direct computation shows that Hχ(0,1)(x) =
π−1 log(x/x − 1). Then Hχ(0,1)(x) ≥ c(− log x)χ(0,1/4)(x), and the
last function is not in W Lp,λ(xλ−1).
(cid:3)
The result in this proposition was proved for the strong estimates
(hence, 1 < p < ∞) by N. Samko in [7, Theorem 4.7], and in [8] she
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
15
discussed the necessity of a more general condition. Our result gives
also the necessity for the weak estimates (1 ≤ p < ∞).
In the next proposition we give a direct proof of the boundedness
of M for the Morrey spaces with weight xλ−n (the endpoint of the
allowed range).
In [10] this is a consequence of a certain sufficient
condition involving duality.
Proposition 5.7. Let 0 ≤ λ < n. M is bounded on Lp,λ(xλ−n) for
1 < p < ∞ and from L1,λ(xλ−n) to W L1,λ(xλ−n).
Proof. Let f ∈ Lp,λ(xλ−n) with λ > 0. Assume that it is nonnegative.
Consider the ball Br := B(x, r). Decompose f as f1 + f2, where f1 =
f χB2r . Using the subadditivity of M we have
Mf (y) ≤ Mf1(y) + Mf2(y).
Using the boundedness of M on Lp(xλ−n) we have
ZBr
(Mf1)(y)pyλ−ndy ≤ C1ZB2r
f (y)pyλ−ndy ≤ C2rλkf kp
Lp,λ(xλ−n).
On the other hand, Mf2(y) is almost constant on Br in the sense that
Mf2(y1)
∼ Mf2(y2) ∼ Mf2(x) for y1, y2 ∈ Br. Then
(5.1)
Moreover,
ZBr
(Mf2)(y)pyλ−ndy ≤ C1(Mf2)(x)pZBr
yλ−ndy.
Mf2(x) ∼ sup
R>2r
1
RnZBR\B2r
f.
We distinguish two types of balls as in Remark 2.2.
In the case of a ball centered at 0, we have
1
RnZBR\B2r
f ≤
1
Rn(cid:18)ZBR
f p(cid:19)1/p
BR1/p′
≤ Ckf kLp,λ(xλ−n),
using (2.3) with w ≡ 1 and α = λ − n. Since the last integral in (5.1)
is Crλ we get the desired estimate.
For a ball centered at x 6= 0 with radius r ≤ x/4 we consider first
R ≥ x/2. In such case,
1
RnZBR\B2r
f ≤
1
RnZB(0,3R)
f ≤ Ckf kLp,λ(xλ−n).
16
JAVIER DUOANDIKOETXEA AND MARCEL ROSENTHAL
The last integral in (5.1) is Cxλ−nrn and this is bounded by Crλ
because r < x and λ − n is negative. Let now 2r < R < x/2. Then
1
RnZBR\B2r
f ≤
1
Rn(cid:18)ZBR
f p(cid:19)1/p
Rn/p′
≤ Cx
n−λ
p R
λ−n
p kf kLp,λ(xλ−n).
Replacing the last integral in (5.1) by Cxλ−nrn the needed estimate
holds because Rλ−nrn ≤ rλ due to R > r.
The proof of the weak type for p = 1 is similar.
(cid:3)
The range λ − n < β < λ + n(p − 1) for weights of type xβ corre-
sponds to all the power weights in Ap+λ/n ∩ RHn/(n−λ). On the other
hand, the result in Theorem 1.1 is valid for λ < n/σ′, that is, we need
w ∈ RHσ for some σ > n/(n − λ) to get the estimate. Such a σ exists
for any weight in RHn/(n−λ) by the self-improvement property of the
reverse Holder inequalities (Gehring's lemma). The endpoint weight
xλ−n, for which the estimates for M hold, is not in RHn/(n−λ), but it
is in RHσ for every σ < n/(n − λ). One could guess the necessity of
a reverse Holder condition of this type in Proposition 5.2, but we have
not been able to get it.
The weights in Ap+λ/n ∩ RHn/(n−λ) are characterized by the factor-
ization u−λ/nw with u ∈ A1 and w ∈ Ap ∩ RHn/(n−λ). The result in
Theorem 1.1 covers all the weights of this type for which u is a power
weight in A1. By translation invariance power weights can be taken
to be centered at a point different from the origin. The sufficiency of
Ap+λ/n ∩ RHn/(n−λ) in Theorem 1.1 remains an open question for us.
References
[1] Adams, D. R.: Morrey spaces, Lecture Notes in Applied and Numerical Har-
monic Analysis. Birkhauser/Springer, Cham (2015).
[2] Cruz-Uribe, D. V., Martell, J. M., P´erez, C.: Weights, extrapolation and the
theory of Rubio de Francia. Operator Theory: Advances and Applications 215,
Birkhauser, Basel (2011).
[3] Duoandikoetxea, J., Rosenthal, M.: Extension and boundedness of operators
on Morrey spaces from extrapolation techniques and embeddings. J. Geom.
Anal. 28, 3081 -- 3108 (2018).
[4] Johnson, R., Neugebauer, C. J.: Change of variable results for A
p and reverse
Holder RH
r-classes. Trans. Amer. Math. Soc. 328, 639 -- 666 (1991).
[5] Knese, G., McCarthy, J. E., Moen, K.: Unions of Lebesgue spaces and A1
majorants. Pacific J. Math. 280, 411 -- 432 (2016).
[6] Nakamura, S., Sawano, Y.: The singular integral operator and its commutator
on weighted Morrey spaces. Collect. Math. 68, 145 -- 174 (2017).
[7] Samko, N.: Weighted Hardy and singular operators in Morrey spaces. J. Math.
Anal. Appl. 350, 56 -- 72 (2009).
BOUNDEDNESS OF OPERATORS ON WEIGHTED MORREY SPACES
17
[8] Samko, N.: On a Muckenhoupt-type condition for Morrey spaces. Mediterr. J.
Math. 10, 941 -- 951 (2013).
[9] Stein, E. M.: Harmonic analysis: real-variable methods, orthogonality, and
oscillatory integrals. Princeton University Press, Princeton, NJ (1993).
[10] Tanaka, H.: Two-weight norm inequalities on Morrey spaces. Ann. Acad. Sci.
Fenn. Math. 40, 773 -- 791 (2015).
(J. D.) Universidad del Pa´ıs Vasco/Euskal Herriko Unibertsitatea,
Departamento de Matem´aticas/Matematika saila, Apdo. 644, 48080
Bilbao, Spain
E-mail address: [email protected], [email protected]
|
1202.6390 | 1 | 1202 | 2012-02-28T21:52:17 | Higher Order Spreading Models | [
"math.FA"
] | We introduce the higher order spreading models associated to a Banach space $X$. Their definition is based on $\ff$-sequences $(x_s)_{s\in\ff}$ with $\ff$ a regular thin family and the plegma families. We show that the higher order spreading models of a Banach space $X$ form an increasing transfinite hierarchy $(\mathcal{SM}_\xi(X))_{\xi<\omega_1}$. Each $\mathcal{SM}_\xi (X)$ contains all spreading models generated by $\ff$-sequences $(x_s)_{s\in\ff}$ with order of $\ff$ equal to $\xi$. We also provide a study of the fundamental properties of the hierarchy. | math.FA | math |
HIGHER ORDER SPREADING MODELS
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Abstract. We introduce the higher order spreading models associated to a
Banach space X. Their definition is based on F -sequences (xs)s∈F with F a
regular thin family and the plegma families. We show that the higher order
spreading models of a Banach space X form an increasing transfinite hierarchy
(SMξ(X))ξ<ω1 . Each SMξ(X) contains all spreading models generated by
F -sequences (xs)s∈F with order of F equal to ξ. We also provide a study of
the fundamental properties of the hierarchy.
1. Introduction
Spreading models have been invented by A. Brunel and L. Sucheston [7] in the
middle of 70's and since then have a constant presence in the evolution of the Banach
space theory. Recall that a sequence (en)n in a seminormed space (E, k·k∗) is called
a spreading model of the space X if there exists a sequence (xn)n in X which is
Schreier almost isometric to (en)n, that is for some null sequence (δn)n of positive
reals we have
(1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
kXi=1
aixni(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXi=1
< δk,
aiei(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i=1 ∈ [−1, 1]k. We also say that the sequence
for every k ≤ n1 < . . . < nk and (ai)k
(xn)n which satisfies (1) generates (en)n as a spreading model. By an iterated
use of Ramsey's theorem [17], Brunel and Sucheston proved that every bounded
sequence in a Banach space X has a subsequence generating a spreading model.
It is easy to see that any sequence (en)n satisfying (1) is spreading1. The impor-
tance of the spreading models arise from the fact that they connect in an asymptotic
manner the structure of an arbitrary Banach space X to the corresponding one of
spaces generated by spreading sequences. The definition of the spreading model
resembles the finite representability2 of the space generated by the sequence (en)n
into the space (X, k · k). However there exists a significant difference between the
two concepts. Indeed in the frame of the finite representability there are two classi-
cal achievements: Dvoretsky's theorem [8] asserting that ℓ2 is finitely representable
in every Banach space X and also Krivine's theorem [12] asserting that for every
linearly independent sequence (xn)n in X there exists a 1 ≤ p ≤ ∞ such that ℓp is
block finitely representable in the subspace generated by (xn)n. On the other hand
12010 Mathematics Subject Classification: Primary 46B03, 46B06, 46B25, 46B45, Secondary
05D10
2Keywords: Spreading models, Ramsey theory, thin families, plegma families
1A sequence (en)n in a seminormed space (E, k · k∗) is called spreading if for every n ∈ N,
k1 < . . . < kn in N and a1, . . . , an ∈ R we have that k Pn
j=1 aj ejk∗ = k Pn
j=1 aj ekj k∗
2A Banach space Y is finitely representable in X if for every finite dimensional subspace F of
Y and every ε > 0 there exists T : F → Y bounded linear injection such that kT k · kT −1k < 1 + ε.
1
2
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
E. Odell and Th. Schlumprecht [15] have shown that there exists a reflexive space
X admitting no ℓp as a spreading model. Thus the spreading models of a space X
lie strictly between the finitely representable spaces in X and the spaces that are
isomorphic to a subspace of X.
The spreading models associated to a Banach space X can be considered as a
cloud of Banach spaces, including many members with regular structure, surround-
ing the space X and offering information concerning the local structure of X in an
asymptotic manner. Our aim is to enlarge that cloud and to fill in the gap between
spreading models and the spaces which are finitely representable in X. More pre-
cisely we extend the Brunel -- Sucheston concept of a spreading model and we show
that under the new definition the spreading models associated to a Banach space
X form a whole hierarchy of classes of spaces indexed by the countable ordinals.
The first class of this hierarchy is the classical spreading models. The initial step of
this extension has already been done in [4] where the class of k-spreading models
was defined for every positive integer k. The transfinite extension introduced in
the present paper requires analogue ingredients that we are about to describe.
The first one is the F -sequences; that is sequences of the form (xs)s∈F where
the index set F is a regular thin family of finite subsets of N (see Definition 2.7).
Typical examples of such families are the k-element subsets of N and also the max-
imal elements of the ξth-Schreier family Sξ (see [2]). A subsequence of (xs)s∈F is
a restriction of the F -sequence on an infinite subset of N, i.e.
it is of the form
(xs)s∈F ↾M where F ↾ M = F ∩ [M ]<∞. Among others we study the convergence
of F -sequences in a topological space (X, T ). In this setting we show that when
the closure of (xs)s∈F in (X, T ) is a compact metrizable space then we can always
restrict to an infinite subset M of N where the subsequence (xs)s∈F ↾M is subordi-
s ∈ F ↾ M (see Definition 5.8 and Theorem 5.10).
nated ; that is if bF = {t ∈ [N]<∞ : ∃s ∈ F such that t is an initial segment of s}
then there exists a continuous map ϕ : bF ↾ M → X with ϕ(s) = xs for every
The second ingredient is the plegma families which extend the corresponding one
in [4]. Roughly speaking a plegma family is a sequence (s1, . . . , sl) of non empty
finite subsets of N where the first elements of s1, ..., sl are in increasing order and
they lie before their second elements which are also in increasing order and so on
(see Definition 3.1). Here the plegma families do not necessarily include sets of
equal size.
The F -sequences and the plegma families are the key components for the def-
inition of the higher order spreading models which goes as follows. Given an F -
sequence (xs)s∈F in a Banach space X and a sequence (en) in a seminormed space
(E, k · k∗) we will say that (xs)s∈F generates (en)n as an F -spreading model if for
some null sequence (δn)n of positive reals we have
(2)
kXi=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
aixsi(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXi=1
< δk,
aiei(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i=1 ∈ [−1, 1]k and every plegma family (si)k
for every (ai)k
i=1 in F with k ≤ min s1.
Note that the F -sequences (xs)s∈F generate a higher order spreading model just
as the ordinary sequences (xn)n do in the classical definition. Moreover since the
family of all k-element subsets of N is a regular thin family, the above definition
extends the classical definition of the spreading model as well as the one of k-
spreading models given in [4].
HIGHER ORDER SPREADING MODELS
3
Brunel -- Sucheston's theorem [6] is extended in the frame of the bounded F -
sequences (xs)s∈F in a Banach space X. Namely, every bounded F -sequence in
X contains a subsequence (xs)s∈F ↾M generating an F -spreading model. The proof
is based on the fact that plegma families with elements in a regular thin family
satisfy strong Ramsey properties. It is notable that the concept of the F -spreading
model is independent of the particular family F and actually depends only on
the order3 of the family F . Namely, if (en)n is an F -spreading model then it is
also a G-spreading model for every regular thin family G with o(G) ≥ o(F ). This
fact allow us to classify all the F -spreading models of a Banach space X as an
increasing transfinite hierarchy of the form (SMξ(X))ξ<ω1 . Let's point out that
the ξ-spreading models of X have a weaker asymptotic relation to the space X as
ξ increases to ω1.
The infinite graphs with vertices from a regular thin family and edges the plegma
pairs is the key for the proof of the above results. Specifically, it is shown that if G
and F are two regular thin families with o(G) ≥ o(F ) then there exist an infinite
subset M of N and a plegma preserving map ϕ : G ↾ M → F (that is (ϕ(s1), ϕ(s2))
is a plegma pair in F whenever (s1, s2) is plegma pair in G ↾ M ). Moreover, it
is also shown that such an embedding is forbidden if we wish to go from families
of lower to families of higher order. More precisely, if o(F ) < o(G) then for every
M ∈ [N]∞ and ϕ : F ↾ M → G there exists L ∈ [M ]∞ such that for every plegma
pair (s1, s2) in F ↾ L neither (φ(s1), φ(s2)) nor (φ(s2), φ(s1)) is a plegma pair in G
(see Theorems 3.16 and 3.18).
The paper is organized as follows.
In Section 2 we review some basic facts
concerning families of finite subsets of N and we define the regular thin families.
In Section 3 we study the plegma families and their properties. In Section 4 we
introduce the definition of the higher order spreading models.
In Section 5 we
deal with F -sequences (xs)s∈F in a general topological space. Finally, in Section
6 we study the F -sequences which generate several classes of spreading sequences
as spreading models. In this last section we show that several well known results
concerning the classical spreading models remain valid in the higher order setting.
For instance, we show that a subordinated, seminormalized and weakly null F -
sequence generates an unconditional spreading model.
The present paper is an updated version of the first part of [3]. The second part
which deals with certain examples will be presented elsewhere.
1.1. Preliminary notation and definitions. By N = {1, 2, . . .} we denote the
set of all positive integers. Throughout the paper we shall identify strictly increasing
sequences in N with their corresponding range i.e. we view every strictly increasing
sequence in N as a subset of N and conversely every subset of N as the sequence
resulting from the increasing ordering of its elements. We will use capital letters as
L, M, N, ... to denote infinite subsets and lower case letters as s, t, u, ... to denote
finite subsets of N. For every infinite subset L of N, [L]<∞ (resp. [L]∞) stands for
the set of all finite (resp. infinite) subsets of L. For an L = {l1 < l2 < ...} ∈ [N]∞
and a positive integer k ∈ N, we set L(k) = lk. Similarly, for a finite subset
s = {n1 < .. < nm} of N and for 1 ≤ k ≤ m we set s(k) = nk. For an L = {l1 <
l2 < ...} ∈ [N ]∞ and a finite subset s = {n1 < .. < nm} (resp. for an infinite subset
3The order of F , denoted by o(F ) is a countable ordinal which measures the complexity F
(see Section 2 for the precise definition). For example the family of k-element subsets of N has
order k, while the ξth-Schreier has order o(Sξ) = ωξ
4
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
N = {n1 < n2 < ...} of N), we set L(s) = {ln1, ..., lnm} = {L(s(1)), ..., L(s(m))}
(resp. L(N ) = {ln1, ln2, ...} = {L(N (1)), L(N (2)), ...}).
For s ∈ [N]<∞ by s we denote the cardinality of s. For L ∈ [N]∞ and m ∈ N,
we denote by [L]m the set of all s ∈ [L]<∞ with s = m. Also for every nonempty
s ∈ [N]<∞ and 1 ≤ k ≤ s we set sk = {s(1), . . . , s(k)} and s0 = ∅. Moreover,
for s, t ∈ [N]<∞, we write t ⊑ s (resp. t ⊏ s) to denote that t is an initial (resp.
proper initial) segment of s. Also, for every s, t ∈ [N]<∞ we write t < s if either at
least one of them is the empty set, or max t < min s.
Concerning Banach space theory, although the notation that we follow is the
standard one, we present for the sake of completeness some basic concepts that
we will need. Let X be a Banach space. We say that a sequence (xn)n in X is
bounded (resp. seminormalized) if there exists M > 0 (resp. M1, M2 > 0) such
that kxnk ≤ M (resp. M1 ≤ kxnk ≤ M2) for all n ∈ N. The sequence (xn)n is
called Schauder basic if there exists a constant C ≥ 1 such that
(3)
anxn(cid:13)(cid:13)(cid:13),
kXn=1
(cid:13)(cid:13)(cid:13)
anxn(cid:13)(cid:13)(cid:13) ≤ C(cid:13)(cid:13)(cid:13)
mXn=1
for every k ≤ m in N and a1, ..., am ∈ R. Finally, we say that (xn)n is C-
unconditional, if for every m ∈ N, F ⊆ {1, ..., m} and a1, ..., am ∈ R, it holds
that
(4)
(cid:13)(cid:13)(cid:13)Xn∈F
anxn(cid:13)(cid:13)(cid:13) ≤ C(cid:13)(cid:13)(cid:13)
mXn=1
anxn(cid:13)(cid:13)(cid:13).
2. Regular thin families
In this section we define the regular thin families of finite subsets of N and we
study their basic properties. The definition is based on two well known concepts,
namely that of regular families traced back to [2] and that of thin families defined
in [14] and extensively studied in [16] and [13].
2.1. On families of finite subsets of N. We start with a review of the basic
concepts concerning families of finite subsets of N. For a more detailed exposition
the reader can refer to [5].
2.1.1. Ramsey properties of families of finite subsets of N. For a family F ⊆ [N]<∞
and L ∈ [N]∞, we set
(5)
F ↾ L = {s ∈ F : s ⊆ L} = F ∩ [L]<∞.
Recall the following terminology from [11]. Let F ⊆ [N]<∞ and M ∈ [N]∞. We
say that F is large in M if for every L ∈ [M ]∞, F ↾ L is nonempty. We say that
F is very large in M if for every L ∈ [M ]∞ there exists s ∈ F such that s ⊑ L.
The following is a restatement ( see [11]) of a well known theorem of C. St. J. A.
Nash-Williams [14] and F. Galvin and K. Prikry [9].
Theorem 2.1. Let F ⊆ [N]<∞ and M ∈ [N]∞. If F is large in M then there exists
L ∈ [M ]∞ such that F is very large in L.
HIGHER ORDER SPREADING MODELS
5
2.1.2. The order of a family of finite subsets of N. Let F ⊆ [N]<∞ be a nonempty
family of finite subsets of N. The order of F ⊆ [N]<∞ is defined as follows (see also
[16]). First, we assign to F its (⊑ −)closure, i.e. the set
(6)
bF = {t ∈ [N]<∞ : ∃s ∈ F with t ⊑ s},
which is a tree under the initial segment ordering. If bF is ill-founded (i.e. there
exists an infinite sequence (sn)n in bF such that sn ⊏ sn+1) then we set o(F ) = ω1.
Otherwise for every maximal element s of bF we set o bF (s) = 0 and recursively for
every s in bF we define
(7)
o bF (s) = sup{o bF (t) + 1 : t ∈ bF and s ⊏ t}.
The order of F denoted by o(F ) is defined to be the ordinal o bF (∅). For instance
o({∅}) = 0 and o([N]k) = k, for every k ∈ N.
For every n ∈ N, we define
(8)
F(n) = {s ∈ [N]<∞ : n < s and {n} ∪ s ∈ F },
where n < s means that either s = ∅ or n < min s. It is easy to see that for every
nonempty family F ⊆ [N]<∞ we have that
(9)
o(F ) = sup{o(F(n)) + 1 : n ∈ N}.
2.1.3. Regular families. A family R ⊆ [N]<∞ is said to be hereditary if for every
s ∈ F and t ⊆ s we have that t ∈ F and spreading if for every n1 < ... < nk and
m1 < ... < mk with n1 ≤ m1, ..., nk ≤ mk, we have that {m1, ..., mk} ∈ R whenever
{n1, ..., nk} ∈ R. Also R is called compact if the set {χs ∈ {0, 1}N : s ∈ R} of
characteristic functions of the members of R, is a closed subset of {0, 1}N under the
product topology.
A family R of finite subsets of N will be called regular if it is compact, hereditary
and spreading. Notice that for every regular family R, bR = R and R(n) is also
regular for every n ∈ N. Moreover, using equation (9) and by induction on the
order of R we easily get the following.
Proposition 2.2. Let R be a regular family. Then o(R ↾ L) = o(R), for every
L ∈ [N]∞.
Exploiting the method of [16] we have the next result.
Proposition 2.3. For every ξ < ω1 there exists a regular family Rξ with o(Rξ) = ξ.
Proof. For ξ = 0 we set R0 = {∅}. We proceed by induction on ξ < ω1. Assume
that for some ξ < ω1 and for each ζ < ξ we have defined a regular family Rζ with
o(Rζ ) = ζ. If ξ is a successor ordinal, i.e. ξ = ζ + 1, then we set
If ξ is a limit ordinal, then we choose a strictly increasing sequence (ζn)n such that
ζn → ξ and we set
Rξ =n{n} ∪ s : n ∈ N, s ∈ Rζ and n < so.
Rξ =[n ns ∈ Rζn : min s ≥ no =[n
Rζn ↾ [n, +∞).
It is easy to check that Rξ is a regular family with o(Rξ) = ξ for all ξ < ω1.
(cid:3)
6
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
We will need some combinatorial properties of regular families. To this end we
give the following definition.
For every R ⊆ [N]<∞ and L ∈ [N]∞, let
(10)
L(R) = {L(s) : s ∈ R}.
Notice that o(R) = o(L(R)) and if R is compact (resp. hereditary) then L(R) is
also compact (resp. hereditary). It is also easily verified that if R is spreading then
L1(R) ⊆ L2(R), for every L1 ⊆ L2 in [N]∞ and more generally,
(11)
for every k ∈ N and L1, L2 ∈ [N]∞ satisfying {L1(j) : j > k} ⊆ {L2(j) : j > k}
(where L(R(k)) = {L(s) : s ∈ R(k)}).
L1(R(k)) ⊆ L2(R(k)),
Proposition 2.4. Let R, S be regular families of finite subsets of N with o(R) ≤
o(S). Then for every M ∈ [N]∞ there exists L ∈ [M ]∞ such that L(R) ⊆ S.
Proof. If o(R) = 0, i.e. R = {∅}, then the conclusion trivially holds. Suppose that
for some ξ < ω1 the proposition is true for every regular families R′, S′ ⊆ [N]<∞
such that o(R′) < ξ and o(R′) ≤ o(S′). Let R, S be regular with o(R) = ξ and let
M ∈ [N]∞. By equation (9) we have that o(R(1)) < o(R). Hence o(R(1)) < o(S)
and so there is some l1 ∈ N such that o(R(1)) ≤ o(S(l1)). Since S is spreading
we have that o(S(l1)) ≤ o(S(n)) for all n ≥ l1 and therefore we may suppose that
l1 ∈ M . Since R(1) and S(l1) are regular families, by our inductive hypothesis there
is L1 ∈ [M ]∞ such that L1(R(1)) ⊆ S(l1).
Proceeding in the same way we construct a strictly increasing sequence (lj)j in
M and a decreasing sequence M = L0 ⊃ L1 ⊃ ... of infinite subsets of M such that
(i) lj+1 ∈ Lj, (ii) lj+1 > Lj(j) and (iii) Lj(R(j)) ⊆ S(lj ), for all j ≥ 1.
We set L = {lj}j and we claim that L(R) ⊆ S. Indeed, by the above construction
we have that for every k ∈ N, {L(j)}j>k ⊆ {Lk(j)}j>k. Therefore by (11) and (iii)
above, we get that
(12)
L(R(k)) ⊆ Lk(R(k)) ⊆ S(lk).
It is easy to see that L(R(k)) = L(R)(lk) and so by (12) we have that L(R)(lk) ⊆
S(lk). Since this holds for every k ∈ N, we conclude that L(R) ⊆ S.
(cid:3)
The next corollary is an immediate consequence.
Corollary 2.5. Let R, S be regular families of finite subsets of N with o(R) =
o(S). Then for every M ∈ [N]∞ there exists L ∈ [M ]∞ such that L(R) ⊆ S and
L(S) ⊆ R.
2.1.4. Thin families. A family F of finite subsets of N is called thin if there do not
exist s, t in F such that s is a proper initial segment of t. The following result is
contained in [14] and [16]. Since it plays a crucial role in the sequel for the sake of
completeness we present its proof.
Proposition 2.6. Let F ⊆ [N]<∞ be a thin family. Then for every finite partition
i=1Fi, (k ≥ 2) of F and every M ∈ [N]∞ there exist L ∈ [M ]∞ and
F = ∪k
1 ≤ i0 ≤ k such that F ↾ L ⊆ Fi0 .
Proof. It suffices to show the result only for k = 2 since the general case follows
easily by induction. So let F = F1 ∪ F2 and M ∈ [N]∞. Then either there is
L ∈ [M ]∞ such that F1 ↾ L = ∅ or F1 is large in M . In the first case it is clear
HIGHER ORDER SPREADING MODELS
7
In the second case by Theorem 2.1 there is L ∈ [M ]∞ such
that F ↾ L ⊆ F2.
that F1 is very large in L. We claim that F ↾ L ⊆ F1. Indeed, let s ∈ F ↾ L.
We choose N ∈ [L]∞ such that s ⊑ N and let t ⊑ N such that t ∈ F1. Then
s, t are ⊑-comparable members of F and since F is thin s = t ∈ F1. Therefore
F ↾ L ⊆ F1.
(cid:3)
2.2. Regular thin families. We are now ready to introduce the main concept of
this section.
Definition 2.7. A family F of finite subsets of N will be called regular thin if (a)
F is thin and (b) the ⊑-closure bF of F is a regular family.
We will use the following notation. For a family R ⊆ [N]<∞ we set
The next lemma allow us to construct regular thin families from regular ones.
(13)
M(R) = {s ∈ R : s is ⊑ -maximal in R}.
Notice that a family F ⊆ [N]<∞ is thin if and only if F = M(bF ).
Lemma 2.8. Let R be a regular family. Then the family M(R) is thin and satisfies
\M(R) = R. Therefore M(R) is regular thin with o(M(R)) = o(R).
Proof. Since M(R) ⊆ R and R is hereditary, we have that \M(R) ⊆ R. To show
that R ⊆ \M(R) notice that for every s ∈ R there exists a t ∈ M(R) such that
s ⊑ t, otherwise R would not be compact. Hence \M(R) = R and clearly M(R) is
thin. Thus M(R) is regular thin. Finally, by the definition of the order, we have
o(M(R)) = o( \M(R)) and hence o(M(R)) = o(R).
(cid:3)
Corollary 2.9. For every ξ < ω1 there is a regular thin family Fξ with o(Fξ) = ξ.
Proof. Let ξ < ω1 and Rξ be a regular family with o(Rξ) = ξ. Then Fξ = M(Rξ)
is as desired.
(cid:3)
all regular thin families and the set of all regular ones. Moreover, the inverse map
sends each regular family R to M(R).
Corollary 2.10. The map which sends F to bF is a bijection between the set of
Proof. By the definition of regular thin families, the map F → bF sends each regular
thin family to a regular one. By Lemma 2.8 we get that the map is 1-1, onto and
the inverse map sends each regular family R to M(R).
(cid:3)
Remark 1. If F is a regular thin family with o(F ) = k < ω, then it is easy
to see that there exists n0 such that F ↾ [n0, ∞) = {s ∈ [N]k : min s ≥ n0}.
Therefore, for each k < ω, the family [N]k is essentially the unique regular thin
family of order k. However this does not remain valid for regular thin families of
order ξ ≥ ω. For instance for every unbounded increasing map f : N → N the
family F = {s ∈ [N]<∞ : s = f (min s)}, is a regular thin family of order ω.
Lemma 2.11. Let R be a regular family and L ∈ [N]∞. Then M(R) ↾ L = M(R ↾
L) and setting M = M(R), \M ↾ L = R ↾ L and o(M ↾ L) = o(R).
Proof. It is easy to see that M(R) ↾ L ⊆ M(R ↾ L). To show the converse
inclusion let s ∈ M(R ↾ L) and assume that s /∈ M(R). Since R ↾ L ⊆ R, s ∈ R
and therefore there exists some t ∈ M(R) with s ⊏ t. Since R is spreading this
8
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
yields that exists t′ ∈ R ↾ L with s ⊏ t′. Thus s /∈ M(R ↾ L), a contradiction.
Therefore s ∈ M(R). Since s ∈ [L]<∞, we have that s ∈ M(R) ↾ L. Therefore
M(R ↾ L) = M(R) ↾ L.
Since M ↾ L = M(R) ↾ L ⊆ R ↾ L and R ↾ L is hereditary, we have that
\M ↾ L ⊆ R ↾ L. Conversely, let s ∈ R ↾ L. Since R ↾ L is compact there is
t ∈ M(R ↾ L) = M(R) ↾ L with s ⊑ t. Hence s ∈ \M ↾ L and \M ↾ L = R ↾ L.
Finally, o(M ↾ L) = o( \M ↾ L) = o(R ↾ L) = o(R), where the last equality
(cid:3)
follows by Proposition 2.2.
Corollary 2.12. Let F be a regular thin family and L ∈ [N]∞. Then F ↾ L =
M(bF ↾ L), \F ↾ L = bF ↾ L and o(F ↾ L) = o(F ).
Proof. Since F is thin we have that F = M(bF ). Setting R = bF in Lemma 2.11
Corollary 2.13. Let F be a regular thin family. Then for every M ∈ [N]∞ there
exists L ∈ [M ]∞ such that F ↾ L is very large in L.
the result follows.
(cid:3)
Proof. If F is regular thin then since bF is spreading, bF ↾ N is nonempty for every
N ∈ [N]∞. Since bF ↾ N = \F ↾ N , we get that F ↾ N is nonempty too, i.e. F is
large in N. Therefore, by Theorem 2.1, for every M ∈ [N]∞ there exists L ∈ [M ]∞
such that F ↾ L is very large in L.
(cid:3)
Definition 2.14. For two families F , G of finite subsets of N, we write F ⊑ G
(resp. F ⊏ G) if every element in F has an extension (resp. proper extension) in
G and every element in G has an initial (resp. proper initial) segment in F .
The following proposition is a consequence of a more general result from [10].
Proposition 2.15. Let F , G ⊆ [N]<∞ be regular thin families with o(F ) < o(G).
Then for every M ∈ [N]∞ there exists L ∈ [M ]∞ such that F ↾ L ⊏ G ↾ L.
Proof. By Corollary 2.13 we have that there exists L1 ∈ [M ]∞ such that both F , G
are very large in L1. So for every L ∈ [L1]∞ and every t ∈ G ↾ L there exists
s ∈ F ↾ L such that s, t are comparable.
Let G1 be the set of all elements of G which have a proper initial segment in F
and G2 = G \ G1. By Proposition 2.6 there exist i0 ∈ {1, 2} and L ∈ [L1]∞ such
that G ↾ L ⊆ Gi0 . It suffices to show that i0 = 1. Indeed, if i0 = 2 then for every
t ∈ G ↾ L there is s ∈ F such that t ⊑ s. This in conjunction with Corollary 2.12
yields that o(G) = o(G ↾ L) ≤ o(F ) which is a contradiction.
(cid:3)
A similar but weaker result holds when o(F ) = o(G).
Proposition 2.16. Let F , G ⊆ [N]<∞ be regular thin families with o(F ) = o(G).
Then there exists L0 ∈ [N]∞ such that for every M ∈ [N]∞ there exists L ∈
[L0(M )]∞ such that L0(F ) ↾ L ⊑ G ↾ L.
M ∈ [N]∞. Notice that L0(F ) and G are large in L0(M ). Hence by Theorem
2.1 there exists L ∈ [L0(M )]∞ such that L0(F ) and G are very large in N . Since
(cid:3)
Proof. By Proposition 2.4 there exists L0 ∈ [N]∞ such that L0(bF ) ⊆ bG. Let
L0(bF ) ⊆ bG, we conclude that L0(F ) ↾ L ⊑ G ↾ L.
Technically the above two propositions are incorporated in one as follows.
HIGHER ORDER SPREADING MODELS
9
Corollary 2.17. Let F , G ⊆ [N]<∞ be regular thin families with o(F ) ≤ o(G). Then
there exists L0 ∈ [N]∞ such that for every M ∈ [N]∞ there exists L ∈ [L0(M )]∞
such that L0(F ) ↾ L ⊑ G ↾ L.
Proof. If o(F ) < o(G), we set L0 = N. Then L0(F ) = F and L0(M ) = M and the
conclusion follows by Proposition 2.15. If o(F ) = 0(G) the result is immediate by
Proposition 2.16.
(cid:3)
3. Plegma families
In this section we introduce the notion of plegma families initially appeared in
[4] for k-subsets of N. Here we do not assume that the members of a plegma family
are necessarily of the same cardinality.
3.1. Definition and basic properties. We begin by stating the definition of a
plegma family.
Definition 3.1. Let l ∈ N and s1, ..., sl be nonempty finite subsets of N. The
l−tuple (sj)l
j=1 will be called a plegma family if the following are satisfied.
(i) For every i, j ∈ {1, ..., l} and k ∈ N with i < j and k ≤ min(si, sj), we
have that si(k) < sj(k).
(ii) For every i, j ∈ {1, ..., l} and k ∈ N with k ≤ min(si, sj − 1), we have
that si(k) < sj(k + 1).
For instance a pair ({n1}, {n2}) of singletons is plegma iff n1 < n2 and a pair of
doubletons ({n1, m1}, {n2, m2}) is plegma iff n1 < n2 < m1 < m2. More generally
for two non empty s, t ∈ [N]<∞ with s ≤ t the pair (s, t) is a plegma pair iff
s(1) < t(1) < s(2) < t(2) < ... < s(s) < t(s). Of course the situation is more
involved when the size of a plegma family is large.
Below we gather together some stability properties of plegma families. We omit
the proof as it is a direct application of the definition.
Proposition 3.2. Let (sj)l
are satisfied.
j=1 be a family of finite subsets of N. Then the following
(i) If (sj)l
j=1 is a plegma family then (sjm )k
every 1 ≤ k ≤ l and 1 ≤ j1 < . . . < jk ≤ l.
m=1 is also a plegma family, for
(ii) The family (sj)l
j=1 is a plegma family iff (sj1 , sj2 ) is a plegma pair, for
every 1 ≤ j1 < j2 ≤ l.
(iii) If (sj)l
j=1 is a plegma family then (tj)l
j=1 is also a plegma family, whenever
∅ 6= tj ⊑ sj for 1 ≤ j ≤ l.
(iv) If (sj)l
j=1 is a plegma family then (L(sj))l
j=1 is also a plegma family, for
every L ∈ [N]∞.
such that s1, ..., sl ∈ F and (sj)l
For every family F ⊆ [N]<∞ and l ∈ N we denote by Plml(F ) the set of all (sj)l
j=1
j=1 is a plegma family. We also set Plm(F ) =
l=1 Plml(F ). Our mai aim is to show that for every l ∈ N, Plml(F ) is a Ramsey
family. To this end we need some preparatory lemmas.
S∞
Lemma 3.3. Let F be a regular thin family and l ∈ N. Then for every (sj)l
Plml(F ) we have that s1 ≤ . . . ≤ sl.
j=1 ∈
10
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Proof. By (ii) of Proposition 3.2 it suffices to show the conclusion for l = 2. Assume
on the contrary that there exists a plegma pair (s1, s2) in F with s1 > s2. We pick
s ∈ [N]<∞ such that s = s1, s2 ⊏ s and s(s2 + 1) > max s1. By the definition
of the plegma family, we have that for every 1 ≤ k ≤ s2, s1(k) < s2(k) = s(k).
Hence, for every 1 ≤ k ≤ s1, we have that s1(k) ≤ s(k). By the spreading property
s2 6∈ F , which is a contradiction.
of bF we get that s ∈ bF . But since s2 is a proper initial segment of s we get that
Lemma 3.4. Let F be a thin family of finite subsets of N and l ∈ N. Let
(sj)l
j=1sj ⊑
∪l
j=1 in Plml(F ) with s1 ≤ . . . ≤ sl, t1 ≤ . . . ≤ tl and ∪l
j=1tj. Then (sj)l
j=1 = (tj)l
j=1, (tj)l
(cid:3)
j=1 and consequently ∪l
j=1sj = ∪l
j=1tj.
j=msj and t = ∪l
Proof. Suppose that for some 1 ≤ m ≤ l we have that (si)i<m = (ti)i<m. We will
show that sm = tm. Let s = ∪l
j=mtj. Then by our assumptions
s ⊑ t. Moreover since sm ≤ . . . ≤ sl and tm ≤ . . . ≤ tl, we easily conclude
that sm(j) = s(cid:0)(j − 1)(l − m + 1) + 1(cid:1), for all 1 ≤ j ≤ sm and similarly tm(j) =
t(cid:0)(j − 1)(l − m + 1) + 1(cid:1), for all 1 ≤ j ≤ tm. Hence, as s ⊑ t, we get that for all
1 ≤ j ≤ min{tm, sm}, sm(j) = tm(j). Therefore sm and tm are ⊑-comparable.
Since F is thin we have sm = tm. By induction on 1 ≤ m ≤ l, we obtain that
sj = tj for every 1 ≤ j ≤ l.
(cid:3)
Theorem 3.5. Let M be an infinite subset of N, l ∈ N and F be a regular thin
family. Then for every finite partition Plml(F ↾ M ) = ∪p
i=1Pi, there exist L ∈
[M ]∞ and 1 ≤ i0 ≤ p such that Plml(F ↾ L) ⊆ Pi0.
Proof. Let U = {∪l
j=1 ∈ Plml(F ↾ M )}. By Lemmas 3.3 and 3.4 we
get that U is thin and the map Φ : Plml(F ↾ M ) → U sending each plegma l-tuple
(sj)l
j=1sj, is an onto bijection. We
set Ui = Φ(Pi), for 1 ≤ j ≤ p. Then U = ∪p
i=1Ui and by Proposition 2.6 there exist
j0 and L ∈ [M ]∞ such that U ↾ L ⊆i0 Ui0 or equivalently Plml(F ↾ L) ⊆ Pj0 .
(cid:3)
j=1 with si ∈ F ↾ M for 1 ≤ i ≤ l, to its union ∪l
j=1sj : (sj)l
3.2. Plegma paths. In this subsection we introduce the definition of the plegma
paths in finite subsets of N and we present some of their properties. Such paths
will be used in the next subsection for the study of plegma preserving maps.
Definition 3.6. Let k ∈ N and s0, ..., sk be nonempty finite subsets of N. We will
say that (sj)k
j=0 is a plegma path of length k from s0 to sk, if for every 0 ≤ j ≤ k−1,
the pair (sj, sj+1) is plegma. Similarly, a sequence (sj)j of nonempty finite subsets
of N will be called an infinite plegma path if for every j ∈ N the pair (sj, sj+1) is
plegma.
The next simple lemma is very useful for the following.
Lemma 3.7. Let (s0, . . . , sk−1, s) be a plegma path of length k from s0 to s such
that s0 < s. Then
k ≥ min{si : 0 ≤ i ≤ k − 1}.
Proof. Suppose that k < min{si : 0 ≤ i ≤ k − 1}. Then s(1) < sk−1(2) <
sk−2(3) < . . . < s1(k) < s0(k + 1), which contradicts that s0 < s.
(cid:3)
For a family F ⊆ [N]<∞ a plegma path in F is a (finite or infinite) plegma path
which consists of elements of F . It is easy to verify the existence of infinite plegma
paths in F whenever F is very large in an infinite subset L of N. In particular, let
HIGHER ORDER SPREADING MODELS
11
s ∈ F ↾ L satisfying the next property: for every j = 1, ..., s − 1 there exists l ∈ L
such that s(j) < l < s(j + 1). Then it is straightforward that there exists s′ ∈ F ↾ L
such that the pair (s, s′) is plegma and moreover s′ shares the same property with
s. Based on this one can built an infinite plegma path in F of elements having the
above property.
The above remarks motivate the following definition. For every F ⊆ [N]<∞ and
L ∈ [N]∞, we set
The proof of the next lemma follows the same lines with the one of Lemma 2.11.
(14) F ↾↾ L =ns ∈ F ↾ L : ∀1 ≤ j ≤ s − 1 ∃ l ∈ L with s(j) < l < s(j + 1)o.
Lemma 3.8. Let F be a regular thin family and L ∈ [N]∞. Then \F ↾↾ L = bF ↾↾ L,
i.e. s ∈ F ↾↾ L iff s is ⊑-maximal in bF ↾↾ L.
We are now ready to present the main result of this subsection. In terms of graph
theory it states that in the (directed) graph with vertices the elements of F ↾↾ L
and edges the plegma pairs (s, t) in F ↾↾ L, the distance between two vertices s0
and s with s0 < s is equal to the cardinality of s0.
Theorem 3.9. Let F be a regular thin family and L ∈ [N]∞. Assume that F is
very large in L. Then for every s0, s ∈ F ↾↾ L with s0 < s there exists a plegma
path (s0, . . . , sk−1, s) in F ↾↾ L of length k = s0 from s0 to s. Moreover k = s0
is the minimal length of a plegma path in F ↾↾ L from s0 to s.
Proof. By Lemmas 3.7 and 3.3 every plegma path in F from s0 to s is of length
at least s0. Therefore for s0 < s a plegma path of the form (s0, . . . , sk−1, s) with
s0, ..., sk−1, s ∈ F and k = s0 certainly is of minimal length.
We will actually prove a slightly more general result. Namely we will show that
t from t to s such that all its elements except perhaps t belong to F ↾↾ L.
for every t in bF ↾↾ L and s ∈ F ↾↾ L with t < s there exists a plegma path of length
For the proof we will use induction on the length of t. The case t = 1 is trivial,
since for every s ∈ [N]<∞ with t < s the pair (t, s) is already a plegma path of
length 1 from t to s. Suppose that for some k ∈ N the above holds for all t in
exist n1 < n2 < . . . < nk+1 in N such that nj − nj−1 > 1, for 2 ≤ j ≤ k and
t = {L(nj) : 1 ≤ j ≤ k + 1}. We set t0 = {L(nj − 1) : 2 ≤ j ≤ k + 1}.
Since nj − 1 > nj−1 we have that t0 is of equal cardinality and pointwise strictly
bF ↾↾ L with t = k.
Let t ∈ bF ↾↾ L with t = k + 1 and s ∈ F ↾↾ L with t < s. Then there
greater than t \ {max t}. Hence, since bF is spreading, we have t0 ∈ bF ↾↾ L and
moreover t0 cannot be a ⊑-maximal element of bF ↾↾ L. By Lemma 3.8 we have that
F ↾↾ L is the set of all ⊑-maximal elements of bF ↾↾ L. Therefore, we conclude that
t0 ∈ bF \ F . Thus, since t0 = k, by the inductive hypothesis, there exists a plegma
Let l = s1. Since (t0, s1) is a plegma pair with t0 ∈ bF \ F and s1 ∈ F , arguing
as in Lemma 3.3, we see that l ≥ k + 1. Moreover since s1 ∈ F ↾↾ L, there exist
m1 < . . . < ml in N such that mj − mj−1 > 1 and s1 = {L(mj) : 1 ≤ j ≤ l}. Notice
that n2 ≤ m1 < n3 ≤ m2 < ... < nk ≤ mk−1 < mk+1 − 1.
path (t0, s1, . . . , sk−1, s) of length k = t0 from t0 to s with all s1, . . . , sk−1, s in
F ↾↾ L.
We set w = t0 ∪ {L(mj − 1) : k + 1 ≤ j ≤ l} and let L′ ∈ [L]∞ such that w is an
initial segment of L′. Notice that w = l. Since F is very large in L there exists
12
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
that t0 < s0 ≤ l and therefore t0 ⊏ s0 ⊑ w.
s0 ∈ F with s0 initial segment of L′. Using again that bF is spreading it is shown
It is easy to check that (t, s0) and (s0, s1) are plegma pairs. Hence the sequence
(t, s0, . . . , sk−1, s) is a plegma path of length k + 1 from t to s with s0, . . . , sk−1, s ∈
F ↾↾ L. The proof of the inductive step as well as of the theorem is complete. (cid:3)
We close this section by presenting an application of the above theorem. We
start with the following definition.
Let X be a set, M ∈ [N]∞, F ⊆ [N]<∞ and ϕ : F → X. We will say that ϕ is
hereditarily nonconstant in M if for every L ∈ [M ]∞ the restriction of ϕ on F ↾ L is
nonconstant. In particular if M = N then we will simply say that ϕ is hereditarily
nonconstant.
Lemma 3.10. Let F be a regular thin family, X be a set and ϕ : F → X be
hereditarily nonconstant. Then for every N ∈ [N]∞ there exists L ∈ [N ]∞ such
that for every plegma pair (s1, s2) in F ↾ L, ϕ(s1) 6= ϕ(s2).
Proof. By Theorem 3.5 there exists an L ∈ [N ]∞ such that either ϕ(s1) 6= ϕ(s2),
for all plegma pairs (s1, s2) in F ↾ L, or ϕ(s1) = ϕ(s2) for all plegma pairs (s1, s2) in
F ↾ L. The second alternative is excluded. Indeed, suppose that ϕ(s1) = ϕ(s2), for
every plegma pair (s1, s2) in F ↾ L. By Corollary 2.13 we may also assume that F ↾
L is very large in L. Let s0 be the unique initial segment of L0 =(cid:8)L(2ρ) : ρ ∈ N(cid:9) in
0 =(cid:8)L(2ρ) : ρ ∈ N and ρ > k(cid:9). By Theorem 3.9 for
F ↾ L and let k = s0. We set L′
every s ∈ F ↾ L′
Therefore for every s ∈ F ↾ L′
which contradicts that ϕ is hereditarily nonconstant.
0 there exist a plegma path (s0, s1, . . . , sk−1, s) of length k in F ↾ L.
0 we have that ϕ(s) = ϕ(sk−1) = . . . = ϕ(s1) = ϕ(s0),
(cid:3)
Proposition 3.11. Let F be a regular thin family, M ∈ [N]∞ and ϕ : F → N be
hereditarily non constant in M . Let also g : N → N. Then there exists N ∈ [M ]∞
such that for every plegma pair (s1, s2) in F ↾ N , ϕ(s2) − ϕ(s1) > g(n), where
min s2 = N (n).
Proof. By Theorem 3.5 there exists L ∈ [M ]∞ such that one of following holds.
(i) For every plegma pair (s1, s2) in F ↾ L, we have ϕ(s1) = ϕ(s2).
(ii) For every plegma pair (s1, s2) in F ↾ L, we have ϕ(s1) > ϕ(s2).
(iii) For every plegma pair (s1, s2) in F ↾ L, we have ϕ(s1) < ϕ(s2).
Since ϕ is hereditarily non constant in M , by Lemma 3.10 case (i) is excluded.
Similarly case (ii) cannot occur since otherwise, (ϕ(sn))n would form a strictly
decreasing sequence in N whenever (sn)n is an infinite plegma path in F ↾ L.
Therefore, case (iii) holds. We choose N ∈ [L]∞ such that for every n ≥ 2, we have
that
Let (s1, s2) be a plegma pair in F ↾ N and let n ∈ N such that min s2 = N (n).
Notice for every 1 ≤ k ≤ s1, we have
g(j).
j≤n
(cid:12)(cid:12)(cid:12)nl ∈ L : N (n − 1) < l < N (n)o(cid:12)(cid:12)(cid:12) ≥ max
(cid:12)(cid:12)(cid:12)nl ∈ L : s1(k) < l < s2(k)o(cid:12)(cid:12)(cid:12) ≥ g(n).
(cid:12)(cid:12)(cid:12)nl ∈ L : s2(k − 1) < l < s2(k)o(cid:12)(cid:12)(cid:12) ≥ g(n).
Similarly for every s1 < k ≤ s2, we have
HIGHER ORDER SPREADING MODELS
13
The above yield that there exist t1, . . . , tg(n) ∈ F ↾ L such that the (g(n) + 2)-tuple
(s1, t1, . . . , tg(n), s2) is plegma. Hence ϕ(s2) − ϕ(s1) > g(n).
(cid:3)
Corollary 3.12. Let F be a regular thin family, M ∈ [N]∞ and ϕ : F → N be
hereditarily non constant in M . Then there exists N ∈ [M ]∞ such that for every
plegma pair (s1, s2) in F ↾ N we have ϕ(s2) − ϕ(s1) > 1.
3.3. Plegma preserving maps. Let F ⊆ [N]<∞ and ϕ : F → [N]<∞. We will
say that the map ϕ is plegma preserving if (ϕ(s1), ϕ(s2)) is a plegma pair whenever
(s1, s2) is a plegma pair in F .
Lemma 3.13. Let F ⊆ [N]<∞ and ϕ : F → [N]<∞. If ϕ is plegma preserving then
for every l ∈ N and (sj)l
j=1 is a plegma l-tuple.
j=1 ∈ Plm(F ) we have that (ϕ(sj ))l
Proof. Let l ∈ N and (sj)l
j=1 be a plegma l-tuple in F . Then for every 1 ≤ j1 <
j2 ≤ l we have that (sj1 , sj2 ) is plegma and thus (ϕ(sj1 ), ϕ(sj2 )) is plegma. Hence,
by (ii) of Proposition 3.2, (ϕ(sj))l
(cid:3)
j=1 is a plegma l-tuple.
Proposition 3.14. Let F be a regular thin family and ϕ : F → [N]∞. Then for
every M ∈ [N]∞ there is L ∈ [M ]∞ such that exactly one of the following holds.
(i) The restriction of ϕ on F ↾ L is plegma preserving.
(ii) For every (s1, s2) ∈ Plm2(F ↾ L) neither (ϕ(s1), ϕ(s2)) nor (ϕ(s2), ϕ(s1))
is a plegma pair.
Proof. Assume that there is M ∈ [N]∞ such that for every L ∈ [M ]∞ neither (i)
nor (ii) holds true. Then by Theorem 3.5 there exists L ∈ [M ]∞ such that for
every (s1, s2) ∈ Plm2(F ↾ N ) we have that (ϕ(s2), ϕ(s1)) is plegma. But this
is impossible. Indeed, otherwise for an infinite plegma path (sn)n in F ↾ N the
sequence (min sn)n would form a strictly decreasing infinite sequence in N.
(cid:3)
For a family F ⊆ [N]<∞ and a plegma preserving map ϕ : F → [N]<∞ we will
say that ϕ is normal provided that ϕ(s1) ≤ ϕ(s2) for every plegma pair (s1, s2)
in F and ϕ(s) ≤ s for every s ∈ F .
Theorem 3.15. Let F be a regular thin family, M ∈ [N]∞ and ϕ : F ↾ M → [N]<∞
be a plegma preserving map. Then there exists L ∈ [M ]∞ such that the restriction
of ϕ on F ↾ L is a normal plegma preserving map.
Proof. By Theorem 3.5 there exists N ∈ [M ]∞ such that either (a) ϕ(s1) ≤
ϕ(s2), for every plegma pair (s1, s2) in F ↾ N , or (b) ϕ(s1) > ϕ(s2), for every
plegma pair (s1, s2) in F ↾ N . Alternative (b) cannot occur since otherwise for an
infinite plegma path (sn)n in F ↾ N the sequence (ϕ(sn))n would form a strictly
decreasing sequence in N. By Theorem 2.6 there exists L ∈ [N ]∞ such that either
(c) ϕ(s) ≤ s, for every s ∈ F ↾ L, or (d) ϕ(s) > s, for every s ∈ F ↾ L. We
claim that (d) cannot hold true. Indeed, since ϕ on F ↾ L is plegma preserving,
using a plegma path of enough large length, we may choose s0, s in F ↾↾ L such
that min s0 < min s and min ϕ(s0) < min ϕ(s). Let k0 = s0. Then by Proposition
3.9 there exists a plegma path (si)k0
i=0 in F ↾↾ L from s0 to s = sk0 of length k0.
By Lemma 3.13 (ϕ(si))k0
i=0 is also a plegma path of length k0 from ϕ(s0) to ϕ(sk0 )
and by Lemma 3.7 we have that
(15)
min{ϕ(si) : 0 ≤ i ≤ k0 − 1} ≤ k0.
14
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Moreover by Lemma 3.4, we have s0 ≤ s1 ≤ ... ≤ sk0 . Hence if (d) holds true
then
(16)
min{ϕ(si) : 0 ≤ i ≤ k0 − 1} > min{si : 0 ≤ i ≤ k0 − 1} ≥ k0,
which contradicts (15).
Therefore we conclude that (a) and (c) hold true i.e. the restriction of ϕ on
(cid:3)
F ↾ L is a normal plegma preserving map.
3.4. Plegma preserving maps between thin families. In this subsection we
are concerned with the question of the existence of a plegma preserving map ϕ :
G → F , where G and F are regular thin families. We shall show that such maps
exist only when o(G) ≥ o(F ). We start with the positive result.
Theorem 3.16. Let F , G be regular thin families with o(F ) ≤ o(G). Then for every
M ∈ [N]∞ there is N ∈ [N]∞ and a plegma preserving map ϕ : G ↾ N → F ↾ M .
Moreover, for every l ∈ N and t ∈ G ↾ N , if min t ≥ N (l) then min ϕ(t) ≥ M (l).
Proof. Let M ∈ [N]∞. By Corollary 2.17 there exists L0 ∈ [N]∞ and N ∈ [L0(M )]∞
such that L0(F ) ↾ N ⊑ G ↾ N . Thus for every t ∈ G ↾ N there exists a unique
st ∈ F such that L0(st) ⊑ t. Moreover, L0(st) ⊑ t ⊆ N ⊆ L0(M ) and therefore
st ⊆ M . We define ϕ : G ↾ N → F ↾ M , by setting ϕ(t) = st. To see that ϕ
is plegma preserving, let (t1, t2) be a plegma pair in G ↾ N . Then L0(ϕ(ti)) ⊑ ti,
for i ∈ {1, 2} and therefore by (iv) of Proposition 3.2 (L0(ϕ(t1), (L0(ϕ(t2)) and
(ϕ(t1), ϕ(t2)) are also plegma pairs. Hence ϕ is plegma preserving.
Finally, let l ∈ N and t ∈ G ↾ N with min t ≥ N (l). Since L0(ϕ(t)) ⊑ t, we have
that min L0(ϕ(t)) = min t and therefore L0(min ϕ(t)) = min L0(ϕ(t)) = min t ≥
N (l) ≥ L0(M )(l) = L0(M (l)). Hence min ϕ(t) ≥ M (l).
(cid:3)
For the following we shall need the next definition. Let F ⊆ [N]<∞ and L ∈ [N]∞.
We define
(17)
L−1(F ) =nt ∈ [N]<∞ : L(t) ∈ Fo.
It is easy to see that for every family F ⊆ [N]<∞ and L ∈ [N]∞ the following are
satisfied.
(a) If F is very large in L then the family L−1(F ) is very large in N.
(b) If F is regular thin then so does the family L−1(F ).
(c) o(L−1(F )) = o(F ↾ L). In particular if F is regular thin then o(L−1(F )) =
o(F ).
Lemma 3.17. Let F be a regular thin family, L ∈ [N]∞ such that F is very
large in L. Let ϕ : F ↾ L → [N]<∞ be a normal plegma preserving map. Let
ψ : L−1(F ) → [N]<∞ defined by ψ(u) = ϕ(L(u)) for every u ∈ L−1(F ). Then ψ is
a normal plegma preserving map which in addition satisfies the following property.
If u ∈ L−1(F ) ↾↾ N and w = ψ(u) then u(i) ≤ w(i) for every 1 ≤ i ≤ w.
Proof. It is easy to check that ψ is a normal plegma preserving map. Therefore we
pass to the proof of the property of ψ. First, by induction on k = u(1), we shall show
that u(1) ≤ ψ(u)(1), for all u ∈ L−1(F ) ↾↾ N. Indeed, if u(1) = 1 then obviously
ψ(u)(1) ≥ 1 = u(1). Suppose that for some k ∈ N and every u ∈ L−1(F ) ↾↾ N with
u(1) = k we have that ψ1(u)(1) ≥ k. Let u ∈ L−1(F ) ↾↾ N with u(1) = k + 1. Since
L−1(F ) is regular thin and very large in N, we easily see that there exists a unique
u′ ∈ L−1(F ) with u′ ⊑ {u(ρ) − 1 : 1 ≤ ρ ≤ u}. Notice that (u′, u) is a plegma pair
HIGHER ORDER SPREADING MODELS
15
in L−1(F ) and u′(1) = k. Since ψ is a normal plegma preserving map we have that
(ψ(u′), ψ(u)) is also a plegma pair. Hence ψ(u)(1) > ψ(u′)(1) ≥ u′(1) = k, that is
ψ(u)(1) ≥ k + 1 = u(1).
Suppose now that for some i ∈ N and every u ∈ L−1(F ) ↾↾ N with i ≤ ψ(u),
u(i) ≤ ψ(u)(i). Let u ∈ L−1(F ) ↾↾ N with i+1 ≤ ψ(u). Since L−1(F ) is very large
in N, there exists u′ ∈ L−1(F ) ↾↾ N such that {u(ρ)−1 : 2 ≤ ρ ≤ u} ⊑ u′. Observe
that (u, u′) is plegma pair in L−1(F ), u ≤ u′ and u(i + 1) = u′(i) + 1. Since ψ is
normal plegma preserving, we have that (ψ(u), ψ(u′)) is also a plegma pair and in
addition i+1 ≤ ψ(u) ≤ ψ(u′). Hence, ψ(u)(i+1) > ψ(u′)(i) ≥ u′(i) = u(i+1)−1,
that is ψ(u)(i + 1) ≥ u(i + 1). By induction on i ∈ N the proof is complete.
(cid:3)
Theorem 3.18. Let F , G be regular thin families with o(F ) < o(G) and M ∈ [N]∞.
Then there is no plegma preserving map from F ↾ M to G. More precisely for every
M ∈ [N]∞ and ϕ : F ↾ M → G and there exists L ∈ [M ]∞ such that for every
plegma pair (s1, s2) in F ↾ L neither (φ(s1), φ(s2)) nor (φ(s2), φ(s1)) is a plegma
pair.
Proof. Assume that there exist M ∈ [N]∞ and ϕ : F ↾ M → G such that ϕ is plegma
preserving. By Theorem 3.15 there exists L ∈ [M ]∞ such that the restriction of
ϕ on F ↾ L is a normal plegma preserving map. By Corollary 2.13 we may also
assume that F is very large in L. Since o(L−1(F )) = o(F ) < o(G) by Proposition
2.15 we have that there exists N ∈ [N]∞ such that L−1(F ) ↾ N ⊏ G ↾ N . We may
assume that N (i + 1) − N (i) > 1 and therefore L−1(F ) ↾ N ⊆ L−1(F ) ↾↾ N. Let
ψ : L−1(F ) → [N]<∞ defined by ψ(u) = ϕ(L(u)) for every u ∈ L−1(F ).
Pick u0 ∈ L−1(F ) ↾ N and set w0 = ψ(u0). Since L−1(F ) ↾ N ⊏ G ↾ N , we
have that u0 ∈ bG \ G and since ϕ takes values in G, we have that w0 ∈ G. We are
now ready to derive a contradiction. Indeed, by Lemma 3.17 we have that ψ is
a normal plegma preserving map which implies that w0 ≤ u0. Moreover, since
L−1(F ) ↾ N ⊆ L−1(F ) ↾↾ N, we have that u0 ∈ L−1(F ) ↾↾ N. Hence, again by
Lemma 3.17, we get that u0(i) ≤ w0(i), for every 1 ≤ i ≤ w0. Summarizing we
have that u0 ∈ bG \ G, w0 ≤ u0 and u0(i) ≤ w0(i), for every 1 ≤ i ≤ w0. Since bG
is spreding we conclude that w0 ∈ bG \ G, which is impossible. Therefore there is no
M ∈ [N]∞ and ϕ : F ↾ M → G such that ϕ is plegma preserving. By Proposition
3.14, we get that for every M ∈ [N]∞ and ϕ : F ↾ M → G there is L ∈ [M ]∞ such
that for every plegma pair (s1, s2) in F ↾ L neither (φ(s1), φ(s2)) nor (φ(s2), φ(s1))
is plegma.
(cid:3)
4. The hierarchy of spreading models
In this section we define the class of the ξ-spreading models of a Banach space
X for every countable ordinal 1 ≤ ξ < ω1. The definition is a transfinite extension
of the corresponding one of the finite order spreading models given in [4]. The
basic ingredients of this extension are the concepts of the F -sequences in X, i.e.
sequences of the form (xs)s∈F with xs ∈ X for every s ∈ F and the plegma families
with members in F where F is a regular thin family.
4.1. The F -spreading models of a Banach space X. Let X be a Banach space
and F ⊆ [N]<∞ be a regular thin family. By the term F -sequence in X we will
mean a map ϕ : F → X. An F -sequence in X will be usually denoted by (xs)s∈F ,
where xs = ϕ(s) for all s ∈ F . Also, for every M ∈ [N]∞, the map ϕ : F ↾ M → X
16
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
will be called an F -subsequence of (xs)s∈F and will be denoted by (xs)s∈F ↾M . An
F -sequence (xs)s∈F in X will be called bounded (resp. seminormalized) if there
exists C > 0 (resp. 0 < c < C) such that kxsk ≤ C (resp. c ≤ kxsk ≤ C) for every
s ∈ F .
Lemma 4.1. Let (xs)s∈F be a bounded F -sequence in X. Let k ∈ N, N ∈ [N]∞
and δ > 0. Then there exists M ∈ [N ]∞ such that
kXj=1
ajxtj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
≤ δ
ajxsj(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
kXj=1
an
j xtj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
≤
δ
3
,
an
j xsj(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for every (tj)k
j=1, (sj)k
j=1 ∈ Plmk(F ↾ M ) and a1, ..., ak ∈ [−1, 1].
Proof. Let (an)n0
(an
Nn0 satisfying
1 , ..., an
n=1 be a δ
3l −net of the unit ball of (Rk, k · k∞). Setting an =
k ) for every 1 ≤ n ≤ n0, we inductively construct N = N0 ⊇ N1 ⊇ . . . ⊇
for every 1 ≤ n ≤ n0 and every (sj)k
j=1, (tj)k
j=1 ∈ Plmk(F ↾ Nn).
have been constructed. Define gn : Plmk(F ↾ Nn−1) → [0, lC] by gn((sj)k
j xsj k. By dividing the interval [0, lC] into disjoint intervals of length δ
The inductive step of the construction has as follows. Suppose that N0, . . . , Nn−1
j=1) =
3 and
j=1) ≤ δ
3 ,
applying Theorem 3.5, there is Nn ⊆ Nn−1 such that gn((tj )k
for every (tj)k
kPk
j=1)−gn((sj )k
j=1 ∈ Plmk(F ↾ Nn).
j=1, (sj)k
j=1 an
We set M = M (Nn0). By (19) and taking into account that (an)n0
n=1 is a δ
of the unit ball of (Rk, k · k∞) it is easy to see that L is as desired.
Lemma 4.2. Let (xs)s∈F be a bounded F -sequence in X. Let l ∈ N, N ∈ [N]∞
and δ > 0. Then there exists M ∈ [N ]∞ such that
3 −net
(cid:3)
for every 1 ≤ k ≤ l, (tj)k
j=1, (sj)k
j=1 ∈ Plmk(F ↾ M ) and a1, ..., ak ∈ [−1, 1].
Proof. It follows easily by an iterated use of Lemma 4.1.
(cid:3)
Lemma 4.3. Let (xs)s∈F be a bounded F -sequence in X. Then for every sequence
(δn)n of positive real numbers and N ∈ [N]∞ there exists M ∈ [N ]∞ satisfying
kXj=1
ajxtj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
≤ δ
ajxsj(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
kXj=1
ajxtj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
ajxsj(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ δl,
for every 1 ≤ k ≤ l, a1, ..., ak ∈ [−1, 1] and (tj)k
that s1(1), t1(1) ≥ M (l).
j=1, (sj )k
j=1 ∈ Plmk(F ↾ M ) such
Proof. It is straightforward by Lemma 4.2 and a standard diagonalization.
(cid:3)
Hence, assuming in the above lemma that (δn)n is a null sequence, we get that
for every l ∈ N and every sequence(cid:0)(sn
1 (1) → ∞ the sequence (kPl
j=1 ajxn
sn
j )l
j=1(cid:1)n of plegma l−tuples in F ↾ M with
sj k)n is convergent and its limit is independent
(18)
(19)
(20)
(21)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
HIGHER ORDER SPREADING MODELS
17
from the choice of ((sn
under which the natural Hamel basis (en)n satisfies
j )l
j=1)n. Actually, we may define a seminorm k · k∗ on c00(N)
(22)
≤ δl,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
kXj=1
ajxsj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
ajej(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for all 1 ≤ k ≤ l, (ai)k
i=1 in [−1, 1] and (sj)k
j=1 ∈ Plmk(F ↾ M ) with s1(1) ≥ M (l).
Let's notice here that there exist bounded F -sequences in Banach spaces such
that no seminorm resulting from Lemma 4.3 is a norm. For example this happens
in the case where (xs)s∈F is constant. Moreover, even if the k · k∗ is a norm on
c00(N), the sequence (en)n is not necessarily Schauder basic.
We are now ready to give the definition of the F -spreading models of a Banach
space X.
Definition 4.4. Let X be a Banach space, F be a regular thin family, (xs)s∈F be
an F -sequence in X. Let (E, k · k∗) be an infinite dimensional seminormed linear
space with Hamel basis (en)n. Also let and M ∈ [N]∞ and (δn)n be a null sequence
of positive real numbers.
We will say that the F -subsequence (xs)s∈F ↾M generates (en)n as an F -spreading
i=1 in [−1, 1] and
model (cid:0)with respect to (δn)n(cid:1) if for every l ∈ N, 1 ≤ k ≤ l, (ai)k
j=1 ∈ Plmk(F ↾ M ) with s1(1) ≥ M (l), we have
(sj)k
(23)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
kXj=1
ajxsj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
ajej(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ δl.
We will also say that (xs)s∈F admits (en)n as an F -spreading model if there exists
M ∈ [N]∞ such that (xs)s∈F ↾M generates (en)n as an F -spreading model.
Finally, for a subset A of X, we will say that (en)n is an F -spreading model of
A if there exists an F -sequence (xs)s∈F in A which admits (en)n as an F -spreading
model.
The next remark is straightforward.
Remark 2. Let M ∈ [N]∞ such that (xs)s∈F ↾M generates (en)n as an F -spreading
model. Then the following are satisfied.
(i) The sequence (en)n is spreading, i.e. for every n ∈ N, k1 < . . . < kn in N
and a1, . . . , an ∈ R we have that kPn
j=1 ajejk∗ = kPn
(ii) For every M ′ ∈ [M ]∞ we have that (xs)s∈F ↾M ′ generates (en)n as an F -
j=1 ajekj k∗.
spreading model.
(iii) For every null sequence (δ′
n)n of positive reals there exists M ′ ∈ [M ]∞ such
that (xs)s∈F ↾M ′ generates (en)n as an F -spreading model with respect to
(δ′
n)n.
By Lemma 4.3 we get the following.
Theorem 4.5. Let F be a regular thin family and X a Banach space. Then every
bounded F -sequence in X admits an F -spreading model. In particular for every
bounded F -sequence (xs)s∈F in X and every N ∈ [N]∞ there exists M ∈ [N ]∞
such that (xs)s∈F ↾M generates F -spreading model.
18
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
4.2. Spreading models of order ξ. In this subsection we show that Definition
4.4 is independent of the particular regular thin family F and actually depends on
the order of F . More precisely we have the following.
Lemma 4.6. Let F , G be regular thin families with o(F ) ≤ o(G). Let X be a Banach
space and (xs)s∈F be an F -sequence in X which admits an F -spreading model (en)n.
Then there exists a G-sequence (wt)t∈G with {wt : t ∈ G} ⊆ {xs : s ∈ F } and which
admits (en)n as a G-spreading model.
Proof. Let M ∈ [N]∞ and (δn) ց 0 be such that (xs)s∈F ↾M generates (en)n as an
F -spreading model with respect to (δn)n. By Theorem 3.16 there exist N ∈ [N]∞
and a plegma preserving map ϕ : G ↾ N → F ↾ M such that min ϕ(t) ≥ M (l), for
every l ∈ N and t ∈ G ↾ N with min t ≥ N (l). For every t ∈ G ↾ N let wt = xϕ(t)
and for every t ∈ G \ (G ↾ N ) let wt = xs0 where s0 is an arbitrary element of F .
We claim that (wt)t∈G↾N generates (en)n as a G-spreading model with respect to
(δn)n.
Indeed, fix l ∈ N, 1 ≤ k ≤ l, (aj )k
j=1 in [0, 1] and (tj )k
j=1 ∈ G ↾ N with
t1(1) ≥ N (l). Let sj = ϕ(tj ), for all 1 ≤ j ≤ k. Then (sj)k
j=1 ∈ Plml(F ↾ M )
and s1(1) ≥ M (l). Therefore, since (xs)s∈F ↾M generates (en)n as an F -spreading
model with respect to (δn)n, we have
kXj=1
ajwtj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
and the proof is complete.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
kXj=1
ajej(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
kXj=1
ajxsj(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
kXj=1
ajej(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ δl
(cid:3)
Corollary 4.7. Let X be a Banach space, A ⊆ X and F , G be regular thin families
with o(F ) = o(G). Then (en)n is an F -spreading model of A iff (en)n is a G-
spreading model of A.
The above permits us to give the following definition.
Definition 4.8. Let A be a subset of a Banach space X and ξ ≥ 1 be a countable
ordinal. We will say that (en)n is a ξ-spreading model of A if there exists a regular
thin family F with o(F ) = ξ such that (en)n is an F -spreading model of A. The
set of all ξ-spreading models of A will be denoted by SMξ(A).
Notice that by Lemma 4.6 we have
(24)
SMζ(A) ⊆ SMξ(A),
for every 1 ≤ ζ < ξ < ω1.
The following is an extension of Example 1 in [4].
It shows that for a given
ξ < ω1 and a regular thin family G there exists a norm on c00(G) such that setting
A = {es : s ∈ G} (where (es)s∈F is the natural Hamel basis of c00(G)), we have
SMζ(A) ( SMξ(A), for every ζ < ξ.
Example 1. Let 1 ≤ ξ < ω1, G be a regular thin family of order ξ and (es)s∈G be
the natural Hamel basis of c00(G). Let (E, k·k) be a Banach space with a normalized
spreading and 1-unconditional basis (en)n which in addition is not equivalent to the
usual basis of c0. Let XG be the completion of c00(G) under the norm k · kF defined
HIGHER ORDER SPREADING MODELS
19
by
(25)
kxkG = supn(cid:13)(cid:13)(cid:13)
lXi=1
ati ei(cid:13)(cid:13)(cid:13) : l ∈ N, (ti)l
for every x =Pt∈G atet ∈ c00(G).
i=1 ∈ Plml(G) and l ≤ t1(1)o,
Let A = {et : t ∈ G}. It is easy to see that (et)t∈G generates (en)n as an G-
spreading model, i.e. (en)n belongs to SMξ(A). Let ζ < ξ. We claim that for
every (e′
n)n is generated by a constant F -sequence with
o(F ) = ζ or it is isometric to the usual basis of c0. Thus (en)n /∈ SMζ(A).
n)n ∈ SMζ(A) either (e′
Indeed, let (e′
n)n ∈ SMζ (A). Then there exists a regular thin family F of order
ζ, an F -sequence (xs)s∈F in A and M ∈ [N]∞ such that (xs)s∈F ↾M generates (e′
n)n
as an F -spreading model. Since {xs}s∈F ⊆ A, we may define ϕ : F ↾ M → G by
choosing for each s ∈ F ↾ M an element ϕ(s) ∈ G satisfying eϕ(s) = xs.
Assume for the following that (e′
n)n ∈ SMζ(A) is not generated by a constant
F -sequence with o(F ) = ζ. By part (ii) of Remark 2 we have that ϕ is hereditarily
nonconstant and by Lemma 3.10 there exists N ∈ [M ]∞ such that for every plegma
pair (s1, s2) in F ↾ N , ϕ(s1) 6= ϕ(s2). Moreover since o(F ) < o(G) by Theorem
3.18 we have that there exists L ∈ [N ]∞ such that for every plegma pair (s1, s2) in
G ↾ L neither (ϕ(s1), ϕ(s2)), nor (ϕ(s2), ϕ(s1)) is a plegma pair. Therefore, by part
(i) of Proposition 3.2, we conclude that for every 1 ≤ k ≤ l, (sj)k
j=1 ∈ Plmk(F ↾ L)
and (ti)l
i=1 ∈ Plml(G), we must have
Hence, for every k ∈ N, a1, . . . , ak ∈ R and (sj)k
j=1 ∈ Plmk(F ↾ L), we have
(26)
(27)
(cid:12)(cid:12)(cid:8)j ∈ {1, . . . , k} : ϕ(sj) ∈ {ti : 1 ≤ i ≤ l}(cid:9)(cid:12)(cid:12) ≤ 1.
ajxsj(cid:13)(cid:13)(cid:13)G
kXj=1
ajeϕ(sj)(cid:13)(cid:13)(cid:13)G
=(cid:13)(cid:13)(cid:13)
kXj=1
max
1≤j≤k
(25),(26)
=
(cid:13)(cid:13)(cid:13)
aj,
i.e. (e′
n)n is isometric to the usual basis of c0.
A natural question arising from the above is the following.
Question. Let X be a separable Banach space.
countable ordinal ξ such that SMζ(X) = SMξ(X) for every ζ > ξ?
Is it true that there exists a
The above question can be also stated in an isomorphic version, i.e. whether
every sequence in SMζ(X) is equivalent to some sequence in SMξ(X) and vice
versa.
Remark 3. In a forthcoming paper we will provide examples establishing the
hierarchy of the higher order spreading models and also illustrating the boundaries
of the theory. Specifically we will show the following.
(1) For every countable limit ordinal ξ there exist a Banach space X such that
SMξ(X) properly includes up to equivalence ∪ζ<ξSMζ(X).
(2) There exist a Banach space X such that for every ξ < ω1 and every (en)n ∈
SMξ(X), the space E generated by (en)n does not contain any isomorphic
copy of c0 or ℓp for all 1 ≤ p < ∞.
The above results require a deeper study of the structure of F -sequences generating
ℓ1-spreading models (see also [4] for the finite order case).
20
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
5. F -sequences in topological spaces
Let (X, T ) be a topological space and F be a regular thin family. As we have
already defined in the previous section, an F -sequence in X is any map of the
form ϕ : F → X and generally an F -subsequence in X is any map of the form
ϕ : F ↾ M → X. In this section we will study the topological properties of F -
sequences. The particular case where F = [N]k, k ∈ N had been studied in [4].
5.1. Convergence of F -sequences. We introduce the following natural definition
of convergence of F -sequences.
Definition 5.1. Let (X, T ) be a topological space, F a regular thin family, M ∈
[N]∞, x0 ∈ X and (xs)s∈F an F -sequence in X. We will say that the F -subsequence
(xs)s∈F ↾M converges to x0 if for every U ∈ T with x0 ∈ U there exists m ∈ N such
that for every s ∈ F ↾ M with min s ≥ M (m) we have that xs ∈ U .
It is immediate that if an F -subsequence (xs)s∈F ↾M in a topological space X is
convergent to some x0, then every further F -subsequence is also convergent to x0.
Also notice that if o(F ) ≥ 2 then the convergence of (xs)s∈F ↾M does not in general
imply that {xs : s ∈ F ↾ M } is a relatively compact subset of X. For instance, let
i=s(1) ei, where (ei)i is the
usual basis of c0. According to Definition 5.1 the [N]2-sequence (xs)s∈[N]2 weakly
= {xs : s ∈ [N]2} ∪ {0} which is not a weakly
converges to zero but {xs : s ∈ [N]2}
compact subset of c0.
(xs)s∈[N]2 be the [N]2-sequence in c0 defined by xs =Ps(2)
w
Proposition 5.2. Let (X, TX ), (Y, TY ) be two topological spaces and f : Y → X
be a continuous map. Let F be a regular thin family, M ∈ [N]∞ and (ys)s∈F an
F -sequence in Y . Suppose that the F -subsequence (ys)s∈F ↾M is convergent to some
y ∈ Y . Then the F -subsequence (f (ys))s∈F ↾M is convergent to f (y).
Proof. Let UX ∈ TX , with f (y) ∈ UX . By the continuity of f there exists UY ∈ TY
such that y ∈ UY and f [UY ] ⊆ UX . Since (ys)s∈F ↾M is convergent to y, there exists
m ∈ N such that for every s ∈ F ↾ M with min s ≥ M (m) we have that ys ∈ UY
and therefore f (ys) ∈ f [UY ] ⊆ UX .
(cid:3)
For the rest of this section we shall restrict to F -sequences in metric spaces.
Definition 5.3. Let (X, ρ) be a metric space, F a regular thin family, M ∈
[N]∞ and (xs)s∈F an F -sequence in (X, ρ). We will say that the F -subsequence
(xs)s∈F ↾M is Cauchy if for every ε > 0 there exists m ∈ N such that for every
s1, s2 ∈ F ↾ M with min s1, min s2 ≥ M (m), we have that ρ(xs1 , xs2 ) < ε.
Proposition 5.4. Let M ∈ [N]∞, F be a regular thin family and (xs)s∈F be an
F -sequence in a complete metric space (X, ρ). Then the F -subsequence (xs)s∈F ↾M
is Cauchy if and only if (xs)s∈F ↾M is convergent.
Proof. If the F -subsequence (xs)s∈F ↾M is convergent, then it is straightforward
that (xs)s∈F ↾M is Cauchy. Concerning the converse we have the following. Suppose
that the F -subsequence (xs)s∈F ↾M is Cauchy. Let (sn)n be a sequence in F ↾ M
such that min sn → ∞. It is immediate that (xsn )n forms a Cauchy sequence in
X. Since (X, ρ) is complete, there exists x ∈ X such that the sequence (xsn )n
converges to x. We will show that the F -subsequence (xs)s∈F ↾M converges to x.
Indeed, let ε > 0. Since (xs)s∈F ↾M is Cauchy, there exists k0 ∈ N such that for
HIGHER ORDER SPREADING MODELS
21
every t1, t2 ∈ F ↾ M with min t1, min t2 ≥ M (k0) we have that ρ(xt1 , xt2 ) < ε
2 .
Since the sequence (xsn )n converges to x and min sn → ∞, there exists n0 ∈ N
) < ε/2. Hence for every s ∈ F ↾ M such
such that min sn0 ≥ M (k0) and ρ(x, xsn0
, xs) < ε and the
that min s ≥ M (k0), we have that ρ(x, xs) ≤ ρ(x, xsn0
proof is completed.
(cid:3)
) + ρ(xsn0
Lemma 5.5. Let M ∈ [N]∞, F be a regular thin family and (xs)s∈F an F -sequence
in a metric space (X, ρ). Suppose that for every ε > 0 and L ∈ [M ]∞ there exists
a plegma pair (s1, s2) in F ↾ L such that ρ(xs1 , xs2 ) < ε. Then the F -subsequence
(xs)s∈F ↾M has a further Cauchy subsequence.
Proof. Let (εn)n be a sequence of positive reals such that P∞
n=1 εn < ∞. Using
Theorem 3.5, we inductively construct a decreasing sequence (Ln)n in [M ]∞, such
that for every n ∈ N and for every plegma pair (s1, s2) in F ↾ Ln we have that
ρ(xs1 , xs2 ) < εn. Let L′ be a diagonalization of (Ln)n, i.e. L′(n) ∈ Ln for all n ∈ N,
and L = {L′(2n) : n ∈ N}.
exists n0 ∈ N such that P∞
We claim that the F -subsequence (xs)s∈F ↾L is Cauchy. Indeed let ε > 0. There
2 . Let s0 be the unique initial segment of
{L(n) : n ≥ n0} in F . If max s0 = L(k) then we set k0 = k + 1. Then for every
s1, s2 ∈ F ↾ L with min s1, min s2 ≥ L(k0), by Theorem 3.9 there exist plegma
j=1 in F ↾ L′ from s0 to s1, s2 respectively. Then for i = 1, 2 we
paths (s1
have that
j=1, (s2
εn < ε
n=n0
j )s0
j )s0
ρ(xs0 , xsi ) ≤
ρ(xsi
j
, xsi
j+1
) <
s0−1Xj=0
s0−1Xj=0
εn0+j <
ε
2
which implies that ρ(xs1 , xs2 ) < ε.
Definition 5.6. Let ε > 0, L ∈ [N]∞, F be a regular thin family and (xs)s∈F an
F -sequence in a metric space X. We will say that the subsequence (xs)s∈F ↾L is
plegma ε-separated if for every plegma pair (s1, s2) in F ↾ L, ρ(xs1 , xs2 ) > ε.
(cid:3)
The following proposition is actually a restatement of Lemma 5.5.
Proposition 5.7. Let M ∈ [N]∞, F be a regular thin family and (xs)s∈F an
F -sequence in a metric space X. Then the following are equivalent.
(i) The F -subsequence (xs)s∈F ↾M has no further Cauchy subsequence.
(ii) For every N ∈ [M ]∞ there exist ε > 0 and L ∈ [N ]∞ such that the subse-
quence (xs)s∈F ↾L is plegma ε-separated.
Proof. (i)⇒(ii): Assume that (ii) is not true. Then there is N ∈ [M ]∞ such that
for every ε > 0 and L ∈ [N ]∞ there exists a plegma pair (s1, s2) in F ↾ L such that
ρ(xs1 , xs2 ) < ε. By Lemma 5.5 the F -subsequence (xs)s∈F ↾N has a further Cauchy
subsequence. Since N ⊆ M this means that (xs)s∈F ↾M has a further Cauchy
subsequence which is a contradiction.
(ii)⇒(i): Suppose that (i) does not hold. Then there exists N ∈ [M ]∞ such that
(xs)s∈F ↾N is Cauchy. Let ε > 0 and L ∈ [N ]∞. Then (xs)s∈F ↾L is also Cauchy and
therefore (xs)s∈F ↾L is not plegma ε-separated, a contradiction.
(cid:3)
5.2. Subordinated F -sequences. By identifying every subset of N with its char-
acteristic function, a thin family F becomes a discrete subspace of {0, 1}N (under
the usual product topology) with bF being its closure. This in particular yields that
every φ : F → (X, T ) is automatically continuous.
In this subsection we show
22
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
that for every regular thin family F , M ∈ [N]∞ and ϕ : F ↾ M → (X, T ) such
that the closure of ϕ(F ↾ M ) is a compact metrizable subspace of X there exist
following definition.
Definition 5.8. Let (X, T ) be a topological space, F be a regular thin family, M ∈
[N]∞ and (xs)s∈F be an F -sequence in X. We say that (xs)s∈F ↾M is subordinated
L ∈ [M ]∞ and a continuous extension bϕ : bF ↾ L → (X, T ). We start with the
(with respect to (X, T )) if there exists a continuous map bϕ : bF ↾ M → (X, T ) with
bϕ(s) = xs for every s ∈ F ↾ M .
Assume that (xs)s∈F ↾M is subordinated. Then since F is dense in bF , there
exists a unique continuous map bϕ : bF ↾ M → (X, T ) witnessing this. Moreover, for
the same reason we have {xs : s ∈ F ↾ M } = bϕ(cid:0)bF ↾ M(cid:1), where {xs : s ∈ F ↾ M }
is the T -closure of {xs : s ∈ F ↾ M } in X. Therefore {xs : s ∈ F ↾ M } is a
countable compact metrizable subspace of (X, T ) with Cantor-Bendixson index at
most o(F ) + 1. Another property of subordinated F -sequences is stated in the next
proposition.
Proposition 5.9. Let (X, T ) be a topological space, F be a regular thin family and
(xs)s∈F be an F -sequence in X. Let M ∈ [N]∞ such that (xs)s∈F ↾M is subordinated.
(X, T ) is the continuous map witnessing the fact that (xs)s∈F ↾M is subordinated
Proof. Via the identity map we may consider the family F as an F -sequence in
let (ys)s∈F be the F -sequence in Y , with ys = s
for every s ∈ F . As we have already noticed (ys)s∈F converges to the empty set.
Then (xs)s∈F ↾M is a convergent F -subsequence in X. In particular, if bϕ : bF ↾ M →
then (xs)s∈F ↾M is convergent to bϕ(∅).
the metric space Y = bF i.e.
Hence, since bϕ : bF ↾ M → (X, T ) is continuous, by Proposition 5.2 we get that
(cid:0)bϕ(ys)(cid:1)s∈F ↾M converges to bϕ(∅). Since bϕ(ys) = bϕ(s) = xs for every s ∈ F ↾ M ,
this means that (xs)s∈F ↾M is convergent to bϕ(∅).
Theorem 5.10. Let F be a regular thin family and (xs)s∈F be an F -sequence in a
topological space (X, T ). Then for every M ∈ [N]∞ such that {xs : s ∈ F ↾ M } is a
compact metrizable subspace of (X, T ) there exists L ∈ [M ]∞ such that (xs)s∈F ↾L
is subordinated.
(cid:3)
Proof. We will use induction on the order of the regular thin family F . If o(F ) = 0
(i.e. the family F is the singleton F = {∅}) the result trivially holds. Let ξ < ω1
and assume that the theorem is true when o(F ) < ξ.
We fix a regular thin family F with o(F ) = ξ, an F -sequence (xs)s∈F in a
topological space (X, T ) and M ∈ [N]∞ such that {xs : s ∈ F ↾ M } is a compact
metrizable subspace of (X, T ). By passing to an infinite subset of M if it is necessary
we may also suppose that F is very large in M . Let ρ be a compatible metric for
the subspace X0 = {xs : s ∈ F ↾ M }. We shall construct (a) a strictly increasing
sequence (mn)n in M , (b) a decreasing sequence M = M0 ⊇ M1 ⊇ ... of infinite
subsets of M (c) a sequence bϕn of maps with bϕn : \F(mn) ↾ Mn → X and (d)
a decreasing sequence of closed balls (Bn)n in X0 such that for every n ∈ N the
following are satisfied:
(i) mn = min Mn−1 and Mn ⊆ Mn−1 \ {mn},
(ii) diam Bn < 1/n,
HIGHER ORDER SPREADING MODELS
23
(iii) the map bϕn is continuous,
(iv) bϕn(u) = x{mn}∪u, for every u ∈ F(mn) ↾ Mn and
(v) {bϕn(u) : u ∈ \F(mn) ↾ Mn} ⊆ Bn.
We shall present the general inductive step of the above construction so let us
assume that the construction has been carried out up to some n ∈ N. We set
mn+1 = min Mn. Since F is very large in M we have that G = F(mn+1) = {u ∈
[N]<∞ : mn+1 < u and {mn+1} ∪u ∈ F } is a regular thin family. For each u ∈ G we
set yu = x{mn+1}∪u and we form the G-sequence (yu)u∈G. Let M ′
n = Mn \ {mn+1}.
Since Y = {yu : u ∈ G ↾ M ′
n} ⊆ {xs : s ∈ F ↾ M }, we have that the closure of
Y in (X, T ) is also a compact metrizable subspace of (X, T ). Thus Y is a totally
bounded metric space and therefore by Theorem 2.1 and passing to an infinite
subset of M ′
n if it is necessary, we may also suppose that there exists a ball Bn+1
of X0 with diam Bn+1 < (n + 1)−1 and such that
(28)
{yu : u ∈ G ↾ M ′
n} ⊆ Bn+1.
n = Mn \ {mn+1} such that the
sis, there exists an infinite subset Mn+1 of M ′
Moreover, o(G) = o(cid:0)F(mn+1)(cid:1) < o(F ) = ξ. Hence, by our inductive hypothe-
G-subsequence (yu)u∈G↾Mn+1 is subordinated. Let bϕn+1 : bG ↾ Mn+1 → X be the
continuous map witnessing this fact. Then bϕn+1(u) = yu = x{mn+1}∪u, for every
u ∈ F(mn+1) ↾ Mn+1 and by the continuity of bϕn+1 we have
(29) (cid:8)bϕn+1(u) : u ∈ \F(mn+1) ↾ Mn+1(cid:9) ⊆ {yu : u ∈ F(mn+1) ↾ Mn+1}
which completes the proof of the inductive step.
We set M ′ = {mn : n ∈ N}. Since lim diam Bn = 0 and X0 is a compact metric
(28)
⊆ Bn+1,
space there exists a strictly increasing sequence (kn)n and x0 ∈ X0 such that
(30)
lim dist(cid:0)x0, Bkn(cid:1) = 0.
We set L = {mkn : n ∈ N} and we define bϕ : bF ↾ L → X as follows. For s = ∅, we
set bϕ(∅) = x0. Otherwise, if n is the unique positive integer such that mkn = min s
we set bϕ(s) = bϕkn (s \ {min s}) = xt. It is easy to check that bϕ is well defined. To
see that bϕ is continuous let (sn)n be a sequence in bF ↾ L and s ∈ bF ↾ L such that
(30) we obtain that bϕ(sn) → x0 = bϕ(∅). Otherwise, let mkn0
for all but finitely many n and since bϕkn0
sn → s. If s = ∅ we have that min sn → +∞, thus using condition (v) and equation
= min s. Then
(sn),
(cid:3)
, for all but finitely many n. Therefore, bϕ(sn) = bϕkn0
is continuous, bϕ(sn) → bϕ(s).
min sn = min s = mkn0
6. F -sequences generating spreading models
Let (xn)n be a sequence in a Banach space X generating a spreading model
(en)n.
It is well known (see [6], [7]) that if (xn)n is norm convergent then the
seminorm in the space E generated by (en)n is not a norm. Furthermore if (xn)n
is weakly null and seminormalized then (en)n is 1-unconditional. In this section
we show that analogues of these results remain true in the higher order setting
of ξ-spreading models. We begin with a short review of the basic properties of
spreading sequences.
(31)
Hence,
(32)
(cid:13)(cid:13)(cid:13)
k−1Xj=1
(cid:13)(cid:13)(cid:13)ek − ek+1(cid:13)(cid:13)(cid:13)∗
ajej + akek(cid:13)(cid:13)(cid:13)∗
ak (cid:13)(cid:13)(cid:13)
k−1Xj=1
=(cid:13)(cid:13)(cid:13)
k−1Xj=1
ajej + akek(cid:13)(cid:13)(cid:13)∗
ajej + akek+1(cid:13)(cid:13)(cid:13)∗
+(cid:13)(cid:13)(cid:13)
k−1Xj=1
1
≤
= 0.
ajej + anek+1(cid:13)(cid:13)(cid:13)∗! = 0.
24
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
quence (en)n in E is called spreading if kPn
6.1. Spreading sequences. Let (E, k · k∗) be a seminormed linear space. A se-
j=1 ajekj k∗, for every
n ∈ N, a1, . . . , an ∈ R and k1 < . . . < kn in N. As we have already mentioned every
spreading model of any order of a Banach space is a spreading sequence. In this
subsection we shall briefly recall some well known results on spreading sequences
that we shall later use (for a more detailed exposition see [1], [5], [6], [7]). We start
with the following elementary lemma.
j=1 ajejk∗ = kPn
Lemma 6.1. Let (E, k · k∗) be a seminormed linear space and (en)n be a spreading
sequence in E. Then the following are equivalent.
(i) There exist k ∈ N and a1, . . . , ak ∈ R not all zero with kPk
(ii) For every n, m ∈ N, ken − emk∗ = 0.
j=1 aieik∗ = 0.
Proof. The implication (ii)⇒(i) is straightforward. To show the converse let k ∈ N
j=1 ajejk∗ = 0. Since (en)n is
and a1, . . . , an ∈ R not all zero, such that kPk
spreading we may suppose that aj 6= 0 for all 1 ≤ j ≤ n. Moreover notice that
Since (en)n is spreading we get that ken − emk∗ = 0 for every n, m ∈ N.
(cid:3)
Spreading sequences in seminormed linear spaces satisfying (i) or (ii) of the above
lemma will be called trivial. As a consequence we have that if (en)n is non trivial
then the restriction of the seminorm k · k∗ to the linear subspace generated by (en)n
is actually norm.
As in [4], we classify the non trivial spreading sequences into the following three
categories: (1) the singular, i.e.
the non trivial spreading sequences which are
not Schauder basic, (2) the unconditional and (3) the conditional Schauder basic
spreading sequences, i.e. the spreading sequences which are Schauder basic but not
unconditional.
Proposition 6.2. Let (en)n be a non trivial spreading sequence.
(i) If (en)n is weakly null then it is 1-unconditional.
(ii) If (en)n is unconditional then either it is equivalent to the usual basis of ℓ1
Proof. (i) See [1]. (ii) Since (en)n is unconditional there exist C > 0 such that
i=1 ei(cid:13)(cid:13)(cid:13) = 0).
or it is norm Ces`aro summable to 0 (i.e. limn→∞(cid:13)(cid:13)(cid:13) 1
nPn
(cid:13)(cid:13)(cid:13)Pn
i=1 εiaiei(cid:13)(cid:13)(cid:13) ≤ C(cid:13)(cid:13)(cid:13)Pn
i=1 aiei(cid:13)(cid:13)(cid:13), for every n ∈ N, a1, . . . , an ∈ R and ε1, . . . , εn ∈
{−1, 1}. Also since it is spreading and nontrivial there exists M > 0 such that
kenk = M for all n ∈ N. Suppose that (en)n is not Ces`aro summable to zero.
Then there exist θ > 0 and a strictly increasing sequence of natural numbers (pn)n
such that k 1
i=1 eik > θ, for all i ∈ {1, . . . , pn}. Hence for every n ∈ N there
nk = 1 such that x∗
exist x∗
i=1 ei) > θ. For every n ∈ N, we set
n with kx∗
n( 1
pnPpn
pnPpn
HIGHER ORDER SPREADING MODELS
25
In = {1, . . . , pn} and let An = {i ∈ In : x∗
n(ei) > θ
θ < x∗
n(cid:16) 1
pn Xi∈In
ei(cid:17) =
1
pn
x∗
n(cid:16) Xi∈An
ei(cid:17) +
1
pn
2 }. Then we have
1
pn
n(cid:16) Xi∈In\An
ei(cid:17) ≤
x∗
AnC +
θ
2
Hence An ≥ θ
to show that (en)n is equivalent to the usual basis of ℓ1.
a1, . . . , an ∈ R and choose n0 ∈ N such that An0 ≥ n. Then
2C pn which gives that limn→∞ An = +∞. We are now ready
Indeed, let n ∈ N,
M
nXi=1
ai ≥(cid:13)(cid:13)(cid:13)
nXi=1
aiei(cid:13)(cid:13)(cid:13) ≥
≥
1
C(cid:13)(cid:13)(cid:13)
1
C
1
aiei(cid:13)(cid:13)(cid:13) =
C(cid:13)(cid:13)(cid:13)
nXi=1
nXi=1
aieAn0 (i)(cid:17) ≥
n(cid:16) nXi=1
aieAn0 (i)(cid:13)(cid:13)(cid:13)
nXi=1
θ
2C
· x∗
ai.
(cid:3)
Proposition 6.3. Let (en)n be a singular sequence and let E be the space gener-
ated by (en)n. Then there is e ∈ E \ {0} such that (en)n is weakly convergent to
e. Moreover if e′
n)n is spreading, 1-unconditional and Ces`aro
summable to zero.
n = en − e then (e′
Proof. Since (en)n is equivalent to all its subsequences and it is not Schauder basic,
every subsequence of (en)n is not Schauder basic. In particular, every subsequence
of (en)n cannot be non trivial weak-Cauchy or weakly null. Hence, by Rosenthal's
ℓ1-theorem [18] (en)n is weakly convergent to a non zero element e ∈ E.
Let e′
n = en − e. To show that (e′
n)n is spreading, let n ∈ N, λ1, ..., λn ∈ R and
block subsequence (wm)m of (e′
such that kn < supp(wm) for all m ≥ m0. Then by (33) we have
n)n is weakly null we may choose a convex
n)n which norm converges to zero. Let m0 ∈ N be
i=1 λi = λ. Since (e′
(34)
λie′
(cid:13)(cid:13)(cid:13)
nXi=1
ki − λwm(cid:13)(cid:13)(cid:13),
for all m ≥ m0. Hence, by taking limits, we get that(cid:13)(cid:13)(cid:13)Pn
i − λwm(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
n)n is non trivial. Finally, since (e′
that is the sequence (e′
mk, by
Lemma 6.1 we have that (e′
n)n is weakly null, by
Proposition 6.2 it is also 1-unconditional and norm Ces´aro summable to zero. (cid:3)
n)n is spreading. Moreover, since ken − emk = ke′
i(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)Pn
i=1 λie′
n − e′
ki(cid:13)(cid:13)(cid:13),
nXi=1
i=1 λie′
λie′
In the following the above decomposition en = e′
n + e of a singular spreading
sequence (en)n will be called the natural decomposition of (en)n.
6.2. F -sequences generating non singular spreading models. We start with
a characterization of the F -sequences in a Banach space X which generate a trivial
spreading model.
(33)
k1 < ... < kn in N. IfPn
i(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
nXi=1
Generally let Pn
λie′
i=1 λi = 0, then notice that
nXi=1
λiei(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
nXi=1
λieki(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
nXi=1
λie′
ki(cid:13)(cid:13)(cid:13)
26
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Theorem 6.4. Let X be a Banach space, F be a regular thin family and (xs)s∈F
be an F -sequence in X and M ∈ [N]∞. Let (E, k · k∗) be an infinite dimensional
seminormed linear space with Hamel basis (en)n such that (xs)s∈F ↾M generates
(en)n as an F -spreading model. Then the following are equivalent:
(i) The sequence (en)n is trivial.
(ii) For every ε > 0 and every L ∈ [M ]∞, there exists a plegma pair (s1, s2) in
F ↾ L such that kxs1 − xs2 k < ε.
(iii) The F -subsequence (xs)s∈F ↾M contains a further norm Cauchy subsequence.
(iv) There exists x ∈ X such that every subsequence of (xs)s∈F ↾M contains a
further subsequence convergent to x.
Proof. (i)⇒(ii): Let ε > 0 and L ∈ [M ]∞. Since the F -subsequence (xs)s∈F ↾L also
generates (en)n as an F -spreading model (see Remark 2), there exists n0 ∈ N such
that for every plegma pair (s1, s2) in F ↾ L with min s1 ≥ L(n0), we have
Let (s1, s2) be such a plegma pair. Since (en)n is trivial we have ke1 − e2k∗ = 0
(35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)xs1 − xs2(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)e1 − e2(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
and therefore by (35) we obtain that(cid:13)(cid:13)(cid:13)xs1 − xs2(cid:13)(cid:13)(cid:13) < ε.
< ε.
(ii)⇒(iii): This follows by Lemma 5.5.
(iii)⇒(i): Using that (xs)s∈F ↾M contains a further norm Cauchy subsequence,
2 ))n of plegma pairs in F ↾ M such that
1 , sn
we easily construct a sequence ((sn
sn
1 (1) → ∞ and kxsn
1 − xsn
2 k < 1/n. Then we have
(cid:13)(cid:13)e1 − e2k∗ = lim
n→∞
(36)
kxsn
1 − xsn
2 k = 0.
Thus, by Lemma 6.1, we conclude that (en)n is trivial.
(iv)⇒(iii): It is straightforward.
(i)⇒(iv) Since every subsequence of (xs)s∈F ↾M generates (en)n as an F -spreading
model we have that for every L ∈ [M ]∞ (xs)s∈F ↾M generates a trivial spreading
model. By the implication (i)⇒(iii) and Proposition 5.4, we have that every sub-
sequence of (xs)s∈F ↾M contains a further convergent subsequence. It remains to
show that all the convergent subsequences of (xs)s∈F ↾M have a common limit.
To this end, let L1, L2 ∈ [M ]∞, x1, x2 ∈ X such that (xs)s∈F ↾Li converges to
xi for i ∈ {1, 2} and let ε > 0. Hence there exists n0 ∈ N such that for every
s ∈ F ↾ L1 and t ∈ F ↾ L2 with min s ≥ L1(n0) and min t ≥ L2(n0) we have
(37)
kx1 − xsk, kx2 − xtk <
ε
3
Since (xs)s∈F ↾M generates the trivial sequence (en)n as an F -spreading model, we
may also assume that
(38)
(39)
for every plegma pair (s1, s2) in F ↾ M with min s1 ≥ M (n0).
It is easy to see that we can choose s1 ∈ F ↾ L1 with min s1 ≥ L1(n0) and
s2 ∈ F ↾ L2 with min s2 ≥ L2(n0) such that (s1, s2) is a plegma pair. Then by (37)
and (38) we have
,
<
ε
3
(cid:13)(cid:13)xs1 − xs2(cid:13)(cid:13) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)xs1 − xs2(cid:13)(cid:13) −(cid:13)(cid:13)e1 − e2(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:13)(cid:13)(cid:13)x1 − x2(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)x1 − xs1(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)xs1 − xs2(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)x2 − xs2(cid:13)(cid:13)(cid:13) < ε.
HIGHER ORDER SPREADING MODELS
Since (39) holds for every ε > 0 we get that x1 = x2. The proof is complete.
27
(cid:3)
We proceed to present a sufficient condition for an F -sequence to generate a
Schauder basic spreading model. We need the next definition.
Definition 6.5. Let A be a countable seminormalized subset of a Banach space
X. We say that A admits a Skipped Schauder Decomposition (SSD) if there exist
C ≥ 1 and a pairwise disjoint sequence (Ak)k of finite subsets of A such that the
following are satisfied.
k=1Ak = A.
(i) ∪∞
(ii) For every N ∈ [N]∞ not containing two successive integers and every se-
quence (xk)k∈N with xk ∈ Ak for all n ∈ N , (xk)k∈N is a Schauder basic
sequence of constant C.
The following proposition is well known but for the sake of completeness we
outline its proof.
Proposition 6.6. Let (xn)n be a seminormalized weakly null sequence in a Banach
space X. Then for every ε > 0 the set A = {xn : n ∈ N} admits a SSD with constant
C = 1 + ε.
Proof. We may assume that X has a Schauder basis (en)n with basis constant K=1
(for example we may assume that X = C[0, 1]). By induction and using the sliding
hump argument, we define (1) a partition (Fn)n of N into finite pairwise disjoint sets
and (2) a sequence (yn)n of finitely supported vectors in X such that the following
are fulfilled:
(i) For every k ∈ N and n ∈ Fk, we have kxn − ynk < ε/2k and
(ii) for every k2 > k1 with k2 − k1 > 1, n1 ∈ Fk1 and n2 ∈ Fk2 , we have
max supp(yn1) < min supp(yn2).
Setting Ak = {xn : n ∈ Fk}, k ∈ N, it is easy to check that (Ak)k satisfies conditions
(i) and (ii) of Definition 6.5.
(cid:3)
Theorem 6.7. Let A be a subset of a Banach space X. If A admits a SSD with
constant C, then every non trivial spreading model of any order of A is Schauder
basic of constant C.
Proof. Let 1 ≤ ξ be a countable ordinal and (en)n be a non trivial spreading model
of A of order ξ. Let F be a regular thin family with o(F ) = ξ, (xs)s∈F be an
F -sequence in A and M ∈ [N]∞ such that (xs)s∈F ↾M generates (en)n as an F -
spreading model. Let (Ak)k be a partition of A satisfying condition (ii) of Def. 6.5.
Finally, let ϕ : F ↾ M → N, defined by ϕ(s) = k if xs ∈ Fk.
Observe that ϕ is hereditarily nonconstant in M . Indeed, otherwise there exists
L ∈ [M ]∞ and k0 ∈ N such that xs ∈ Fk0 for every s ∈ F ↾ L and s ∈ F ↾ L.
Since Fk0 is finite, by Proposition 2.6 there exists N ∈ [L]∞ such that (xs)s∈F ↾N is
constant. By part (ii) of Remark 2, the F -sequence (xs)s∈F ↾N also generates (en)n
as an F -spreading model. But then, since (xs)s∈F ↾N is constant, the sequence (en)n
should be trivial which is a contradiction. Hence ϕ is hereditarily nonconstant in M
and therefore by Corollary 3.12 there exists N ∈ [M ]∞ such that ϕ(s2) − ϕ(s1) > 1
for every plegma pair (s1, s2) in F ↾ N . By the SSD property of A we have that
28
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
for every 1 ≤ m < l ∈ N and for every
(40)
k
mXj=1
ajxsj k ≤ Ck
ajxsj k,
lXj=1
for every plegma l-tuple (sj)l
(en)n is a Schauder basic sequence with constant C.
j=1 in F ↾ N and a1, ..., al ∈ R. This easily yields that
(cid:3)
6.3. Spreading models generated by subordinated F -sequences. Let (xn)n
be a weakly convergent sequence in a Banach space X which is not norm Cauchy
and assume that it generates a spreading model (en)n. It is well known (see [6],
[7]) that (en)n is either an unconditional or a singular spreading sequence. In [4]
we extended this fact for subordinated k-sequences. Here we will show that similar
results also hold true for F -sequences where F is a regular thin family.
6.3.1. Unconditional spreading models. Let n ∈ N and for every 1 ≤ i ≤ n let
Fi ⊆ [N]<∞. We will say that (Fi)n
i=1 is completely plegma connected if for every
choice of si ∈ Fi, the n-tuple (si)n
i=1 is a plegma family. Also for a subset A of a
Banach space X, convA denotes the convex hull of A.
Lemma 6.8. Let X be a Banach space, n ∈ N, F1, . . . , Fn be regular thin families,
family (Fi)n
every i = 1, ..., n.
and L ∈ [N]∞. Assume that for every i = 1, ..., n, there exists a continuous map bϕi :
cFi ↾ L → (X, w). Then for every ε > 0 there exists a completely plegma connected
i=1 such that Fi ⊆ [Fi ↾ L]<∞ and dist(cid:16)bϕi(∅), conv bϕi(Fi)(cid:17) < ε, for
i=1(cid:1) := max{o(Fi) : 1 ≤ i ≤ n}.
Proof. We will use induction on o(cid:0)(Fi)l
i=1(cid:1) = 0, i.e. Fi = {∅} for all 1 ≤ i ≤ n the result follows trivially. Let
o(cid:0)(Fi)l
i=1(cid:1) < ξ. Let n ∈ N,
1 ≤ ξ < ω1 and suppose that the lemma holds true if o(cid:0)(Fi)n
i=1(cid:1) = ξ and assume
L ∈ [N]∞ and F1, . . . , Fn be regular thin families with o(cid:0)(Fi)l
that for every 1 ≤ i ≤ n, there exists a continuous map bϕi :cFi ↾ L → (X, w).
every singleton {l} with l ∈ L belongs to cFi. By the continuity of bϕi, we get that
w- liml∈Lbϕi({l}) = bϕi(∅). By Mazur's theorem, we may choose a finite subset
Fix i ∈ {1, ..., n}. We may suppose that Fi is very large in L and therefore
mi} of L such that
1 < ... < li
Λi = {li
If
(41)
dist(cid:16)bϕi(∅), conv bϕi(Λi)(cid:17) < ε/2,
for every 1 ≤ i ≤ n. We may also assume that
(42)
Λ1 < ... < Λn.
j < s and {li
Let Λ = ∪n
We set Gi
i=1Λi and let M = {l ∈ L : l > max Λ}. Fix 1 ≤ i ≤ n and 1 ≤ j ≤ mi.
j = (Fi)(li
j ↾
j is a continuous
j ) = {s ∈ [N]<∞ : li
map and since o(Gi
our inductive assumption we may choose a completely plegma connected family
M → (X, w), defined by bϕi
j} ∪ s). Notice that bϕi
j (s) = bϕi({li
i=1(cid:17) < o(cid:0)(Fi)n
j ) < o(Fi), o(cid:16)(cid:0)(cid:0)Gi
j=1(cid:1)n
j(cid:1)mi
(cid:0)(cid:0)Gi
j(cid:1)mi
j=1(cid:1)n
j )(cid:17) < ε/2,
dist(cid:16)bϕi
j(∅), conv bϕi
j} ∪ s ∈ Fi} and let bϕi
j : cGi
i=1(cid:1) = ξ. Therefore using
i=1 such that Gi
j ↾ M ]<∞ and
j ⊆ [Gi
j(Gi
(43)
HIGHER ORDER SPREADING MODELS
29
(44)
for every 1 ≤ i ≤ n and 1 ≤ j ≤ mi.
j}∪s : s ∈ Gi
Hence, equation (43) translates to
for every 1 ≤ i ≤ n and 1 ≤ j ≤ mi.
i=1 is completely
j (Gi
j ).
For every 1 ≤ i ≤ n and 1 ≤ j ≤ mi we set F i
j(cid:9). By equation
j) = bϕi(F i
j =(cid:8){li
j=1(cid:1)n
j(cid:1)mi
i=1, we easily see that(cid:0)(cid:0)F i
j(∅) = bϕi({li
j}) and bϕi
j )(cid:17) < ε/2,
j=1(cid:1)n
j(cid:1)mi
(42) and the choice of(cid:0)(cid:0)Gi
plegma connected. Moreover, observe that bϕi
dist(cid:16)bϕi({li
j}), conv bϕi(F i
For every 1 ≤ i ≤ n, let Fi =Smi
By (44) we have dist(cid:0)x, conv bϕi(Fi)(cid:1) < ε/2, for every x ∈ conv bϕi(Λi) and therefore
by (41) we conclude that dist(cid:16)bϕi(∅), conv bϕi(Fi)(cid:17) < ε.
i=1 is a completely plegma
connected family with Fi ⊆ [Fi ↾ L]<∞, for every i = 1, ..., n. Finally, fix 1 ≤ i ≤ n.
Theorem 6.9. Let X be a Banach space, F be a regular thin family and L ∈ [N]∞.
Let (xs)s∈F ↾L be an F -subsequence in X generating an F -spreading model (en)n.
Also assume that (xs)s∈F ↾L is seminormalized, subordinated (with respect to the
weak topology of X) and weakly null. Then (en)n is an 1-unconditional spreading
sequence.
j . Clearly (Fi)n
j=1 F i
(cid:3)
Proof. We first show that (en)n is non trivial. Indeed, otherwise by Theorem 6.4
there exists M ∈ [L]∞ and x0 ∈ X such that the F -subsequence (xs)s∈F ↾M is norm
convergent to x0. Since M ⊆ L, (xs)s∈F ↾M is also weakly null and therefore x0 = 0.
But this is a contradiction since (xs)s∈F ↾L is seminormalized.
We proceed to show that (en)n is 1-unconditional. Fix n ∈ N, 1 ≤ p ≤ n and
a1, . . . , an ∈ [−1, 1]. It suffices to show that for every ε > 0 we have
Indeed, fix ε > 0. Since (xs)s∈F ↾L generates (en)n as an F -spreading model, by
passing to a final segment of L if it is necessary we may assume that
+ ε.
i6=p
nXi=1
aiei(cid:13)(cid:13)(cid:13)∗
aiei(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
<
nXi=1
aiei(cid:13)(cid:13)(cid:13)∗
<(cid:13)(cid:13)(cid:13)
and (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
nXi=1
ε
3
aixsi(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
nXi=1
aiei(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
<
ε
3
,
(45)
(cid:13)(cid:13)(cid:13)
(46)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
nXi=1
i6=p
aixsi(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
nXi=1
i6=p
for every plegma n-tuple (si)n
i=1 in F ↾ L. Since (xs)s∈F ↾L is subordinated with
for all i = 1, ..., n), there exist a completely plegma connected family (Fi)n
sequence (xi)n
respect to the weak topology, there exists a continuous map bϕ : F ↾ L → (X, w)
such that bϕ(s) = xs for every s ∈ F ↾ L. Since (xs)s∈F ↾L is weakly convergent to
bϕ(∅) we have that bϕ(∅) = 0. Therefore by Lemma 6.8 (for Fi = F and bϕi = bϕ,
i=1 in X such that Fi ⊆ [F ↾ L]<∞, xi ∈ conv bϕ(Fi) and kxik < ε/3,
for every 1 ≤ i ≤ n. Let (µs)s∈Fp be a sequence in [0, 1] such thatPs∈Fp
and xp =Ps∈Fp
µsbϕ(t) and for each i 6= p choose si ∈ Fi. By the above we have
and kxpk = kPs∈Fp
that for every s ∈ Fp the n-tuple (s1, . . . , sp−1, s, sp+1, . . . , sn) is a plegma family
µsxsk < ε
3 .
i=1 and a
µs = 1
30
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Therefore by (46) we have
nXi=1
i6=p
(cid:13)(cid:13)(cid:13)
aiei(cid:13)(cid:13)(cid:13)∗
i6=p
aixsi(cid:13)(cid:13)(cid:13) +
≤(cid:13)(cid:13)(cid:13)
nXi=1
µs(cid:13)(cid:13)(cid:13)
nXi=1
≤ Xs∈Fp
aiei(cid:13)(cid:13)(cid:13)∗
=(cid:13)(cid:13)(cid:13)
nXi=1
i6=p
+ ε.
The proof is complete.
i6=p
ε
3
≤(cid:13)(cid:13)(cid:13)
nXi=1
aixsi + apxs(cid:13)(cid:13)(cid:13) +
aixsi + ap Xs∈Fp
≤ Xs∈Fp
2ε
3
ε
3
µsxs(cid:13)(cid:13)(cid:13) + ap
aiei(cid:13)(cid:13)(cid:13)∗
µs(cid:16)(cid:13)(cid:13)(cid:13)
nXi=1
+
ε
3
+
ε
3(cid:17) +
2ε
3
(cid:3)
6.3.2. Singular or isomorphic to ℓ1 spreading models. We proceed to show an ana-
logue of Theorem 6.9 for subordinated F -sequences which are not weakly null. We
will need the following lemma.
are both Ces´aro summable to zero. Suppose that for every n ∈ N and λ1, . . . , λn
Lemma 6.10. Let (en)n and (een)n be two non trivial spreading sequences which
withPn
i=1 λi = 0, we have that
(47)
to zero, we have limm→∞
Then the map en →een extends to a linear isometry from < (en)n > onto < (een)n >.
Proof. Let n ∈ N and λ1, . . . , λn ∈ R. Since (en)n (resp. (een)n) is Ces´aro summable
j=1een+j = 0). Let
λ =Pn
j=1 en+j = 0 (resp. limm→∞
λ
m = 0 and therefore,
mPm
1
1
nXi=1
=(cid:13)(cid:13)(cid:13)
.
λieei(cid:13)(cid:13)(cid:13)∗∗
nXi=1
λiei(cid:13)(cid:13)(cid:13)∗
(cid:13)(cid:13)(cid:13)
mPm
i=1 λi −Pm
m→∞(cid:13)(cid:13)(cid:13)
nXi=1
m→∞(cid:13)(cid:13)(cid:13)
nXi=1
j=1
(47)
= lim
= lim
i=1 λi. ThenPn
λiei(cid:13)(cid:13)(cid:13)∗
(cid:13)(cid:13)(cid:13)
nXi=1
λiei −
λ
m
λieei −
en+j(cid:13)(cid:13)(cid:13)∗
mXj=1
mXj=1een+j(cid:13)(cid:13)(cid:13)∗∗
λ
m
nXi=1
=(cid:13)(cid:13)(cid:13)
λieei(cid:13)(cid:13)(cid:13)∗∗
(cid:3)
The next lemma is from [4]. We reproduce it for the sake of completeness.
i=1 ai = 0.
i=1 aieik = kPn
Lemma 6.11. Let X be a Banach space, F be a regular thin family and (xs)s∈F ↾L
be an F -subsequence in X. Let x0 ∈ X and set x′
s = xs − x0, for all s ∈ F ↾ L.
Assume that (xs)s∈F ↾L and (x′
respectively. Then the following hold.
s)s∈F ↾L generate F -spreading models (en)n and (een)n
(i) kPn
i=1 aieeik, for every n ∈ N and a1, . . . , an ∈ R with
Pn
(ii) The sequence (en)n is trivial if and only if (een)n is trivial.
(iii) The sequence (en)n is equivalent to the usual basis of ℓ1 if and only if (een)n
i=1 ai = 0, we have Pn
with Pn
si . Since (en)n and (een)n are
Proof. (i) Notice that for every n ∈ N, s1, ..., sn in F ↾ L and a1, . . . , an ∈ R
generated by (xs)s∈F ↾L and (xs)s∈F ↾L the result follows.
i=1 aixsi = Pn
is equivalent to the usual basis of ℓ1.
i=1 aix′
HIGHER ORDER SPREADING MODELS
31
(ii) It follows by part (i) and Lemma 6.1.
i=1 a′
i=1 a′
1, . . . , a′
i/2 and an+i = −a′
to the usual basis of ℓ1.
n ∈ N and a′
ai = a′
(iii) We fix ε > 0. If (een)n is not equivalent to the usual basis of ℓ1 then there exist
ieeik < ε. Setting
i=1 aieik = kP2n
kP2n
i = 1 and kPn
i/2, for all 1 ≤ i ≤ n, we haveP2n
i=1 ai = 0 and therefore,
i=1 ai = 1, (en)n is also not equivalent
(cid:3)
n ∈ R such that Pn
i=1 aieeik < ε. SinceP2n
Theorem 6.12. Let X be a Banach space, F be a regular thin family and L ∈ [N]∞.
Let (xs)s∈F ↾L be an F -subsequence in X generating a non trivial F -spreading model
(en)n. Also assume that (xs)s∈F ↾L is subordinated and let x0 be the weak-limit of
(xs)s∈F ↾L. Finally, let x′
s = xs − x0, for every s ∈ F ↾ L. If x0 6= 0 then exactly
one of the following holds.
(i) The sequence (en)n as well as every spreading model of (x′
s)s∈F ↾L is equiv-
alent to the usual basis of ℓ1.
(ii) The sequence (en)n is singular and if en = e′
n+e is its natural decomposition
n)n is the unique (up to isometry) F -spreading model
then kek = kx0k and (e′
of (x′
s)s∈F ↾L.
s)s∈F ↾L. By passing to an infinite
If (en)n is equivalent to the usual basis of ℓ1 then by Lemma 6.11, we have that
subset of L if it is necessary we may assume that (x′
F -spreading model.
Proof. Let (een)n be an F -spreading model of (x′
s)s∈F ↾L generates (een)n as an
the same holds for (een)n and hence (i) is satisfied. Otherwise, again by Lemma
6.11, (een)n is also non trivial and not equivalent to the ℓ1-basis. Let us denote by
k · k∗ (resp. k · k∗∗) the norm of the space generated by (en)n (resp. (een)n). Since
(een)n is non trivial, we have that keenk∗∗ > 0 and therefore (by passing to a final
Theorem 6.9, (een)n is 1-unconditional. Moreover, since (een)n is not equivalent to
the usual basis of ℓ1, by Proposition 6.2 (ii), we conclude that (een)n is norm Ces`aro
segment of L if it is necessary) we may assume that (x′
It is also easy to see that (x′
s)s∈F ↾L is seminormalized.
s)s∈F ↾M is subordinated and weakly null. Therefore by
summable to zero. Hence, by part (i) of Lemma 6.11, we have
For every n ∈ N choose a sequence (sn
Since xs − x0 = x′
s, for every s ∈ F ↾ L, we have
j=1 ∈ Plmn(F ↾ L) such that min sn
1 ≥ L(n).
n + e be the natural decomposition of (en)n. By (50) and the fact that (e′
it is singular. Let
By (48) and (50), we get that (en)n is not Schauder basic, i.e.
en = e′
n)n
is Ces`aro summable to zero, we have that kek = kx0k. To complete the proof it
n)n are isometrically equivalent. By Lemma 6.10
remains to show that (een)n and (e′
(48)
(49)
(50)
Therefore
ej −
1
n
lim
1
n
nXj=1
n→∞(cid:13)(cid:13)(cid:13)
n→∞(cid:13)(cid:13)(cid:13)x0 −
n→∞(cid:13)(cid:13)(cid:13)
1
n
lim
lim
j )n
= lim
2nXj=n+1
ej(cid:13)(cid:13)(cid:13)∗
j(cid:13)(cid:13)(cid:13) = lim
n→∞(cid:13)(cid:13)(cid:13)
n→∞(cid:13)(cid:13)(cid:13)
ej(cid:13)(cid:13)(cid:13)∗
n→∞(cid:13)(cid:13)(cid:13)
nXj=1
nXj=1
= lim
1
n
1
n
1
n
1
n
2nXj=n+1eej(cid:13)(cid:13)(cid:13)∗∗
nXj=1een(cid:13)(cid:13)(cid:13)∗∗
nXj=1eej −
j(cid:13)(cid:13)(cid:13) = lim
n→∞(cid:13)(cid:13)(cid:13)
j(cid:13)(cid:13)(cid:13) = kx0k > 0.
1
n
xsn
x′
sn
xsn
nXj=1
nXj=1
1
n
= 0.
= 0.
32
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
it suffices to show that
,
λieei(cid:13)(cid:13)(cid:13)∗∗
i=1 λ1 = 0.
(51)
λie′
(cid:13)(cid:13)(cid:13)
nXi=1
i(cid:13)(cid:13)(cid:13)∗
=(cid:13)(cid:13)(cid:13)
nXi=1
for every n ∈ N and λ1, . . . λn ∈ R with Pn
λ1, . . . λn ∈ R with Pn
λiei(cid:13)(cid:13)(cid:13)∗
(cid:13)(cid:13)(cid:13)
nXi=1
nXi=1
k→∞(cid:13)(cid:13)(cid:13)
such that limk→∞ min sk
and the proof is complete.
=(cid:13)(cid:13)(cid:13)
i(cid:13)(cid:13)(cid:13)∗
nXi=1
= lim
λixsk
λie′
i=1 λi = 0. For each k ∈ N choose (sk
1 = +∞. Then
Indeed, fix n ∈ N and
j=1 ∈ Plmn(F ↾ L)
j )n
i(cid:13)(cid:13)(cid:13) = lim
k→∞(cid:13)(cid:13)(cid:13)
nXi=1
λix′
sk
i(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
nXi=1
λieei(cid:13)(cid:13)(cid:13)∗∗
(cid:3)
6.3.3. Weakly relatively compact F -sequences. Let X be a Banach space and ξ <
ω1. By SMwrc
(X) we will denote the set of all spreading sequences (en)n such
that there exists a weakly relatively compact subset W of X which admits (en)n
as a ξ-spreading model. We also set
ξ
SMwrc(X) = [ξ<ω1
SMwrc
ξ
(X).
w
Hence (en)n ∈ SMwrc
(X) if and only if there exists an F -sequence (xs)s∈F such
ξ
is a weakly compact subset of X and for some L ∈ [N]∞,
that {xs : s ∈ F }
(xs)s∈F ↾L generates (en)n as an F -spreading model. The F -sequences with weakly
relatively compact range will be called weakly relatively compact ("wrc" in short).
The following proposition says that every wrc F -sequence always contains a subor-
dinated subsequence.
Proposition 6.13. Let X be a Banach space, F be a regular thin family and
(xs)s∈F be a weakly relatively compact F -sequence in X. Then for every M ∈ [N]∞
there exists L ∈ [M ]∞ such that the F -subsequence (xs)s∈F ↾L is subordinated with
respect to the weak topology.
ξ
ζ
w
(X) ⊆ SMwrc
Proof. Let M ∈ [N]∞. Since the weak topology on every separable weakly compact
subset of a Banach space is metrizable, we have that {xs : s ∈ F }
is compact
metrizable. By Theorem 5.10 the result follows.
(cid:3)
Proposition 6.14. Let X be a Banach space, ξ < ω1 and (en)n ∈ SMwrc
(X).
Then for every regular thin family G with o(G) ≥ ξ there exist a weakly relatively
compact G-sequence (wt)t∈G in X and L ∈ [N]∞ such that (wt)t∈G↾L is subordi-
nated with respect to the weak topology and generates (en)n as a G-spreading model.
Consequently SMwrc
Proof. Since (en)n ∈ SMwrc
(X) there exists a weakly relatively compact subset
A of X such that A admits (en)n as a ξ-spreading model. Hence there exists a
regular thin family F of order ξ, an F -sequence (xs)s∈F in A and M ∈ [N]∞ such
that (xs)s∈F ↾M generates (en)n as an F -spreading model. By Lemma 4.6 there
exist a G-sequence (wt)t∈G and N ∈ [N]∞ such that (wt)t∈G↾N generates (en)n
as a G-spreading model and moreover {wt : t ∈ G} ⊆ {xs : s ∈ F } ⊆ A. Hence
(wt)t∈G is a weakly relatively compact G-sequence. By Proposition 6.13 there exists
L ∈ [N ]∞ such that (wt)t∈G↾L is subordinated with respect to the weak topology.
Clearly (wt)t∈G↾L also generates (en)n as a G-spreading model and the proof is
complete.
(cid:3)
(X), for every 1 ≤ ζ < ξ < ω1.
ξ
ξ
HIGHER ORDER SPREADING MODELS
33
Proposition 6.14 implies that every (en)n in SMwrc(X) is generated by a subor-
dinated F -subsequence. Hence, by Theorems 6.9 and 6.12 we obtain the following.
Corollary 6.15. Let X be a Banach space, F be a regular thin family and (xs)s∈F
be a weakly relatively compact F -sequence. Let (en)n be a spreading sequence and
assume that (xs)s∈F admits (en)n as an F -spreading model. Then exactly one of
the following holds:
(i) The sequence (en)n is trivial.
(ii) The sequence (en)n is singular.
x0 ∈ X such that if en = e′
the F -subsequence (x′
generates the sequence (e′
In this case there exist L ∈ [N]∞ and
n + e is the natural decomposition of (en)n then
s = xs − x0 for all s ∈ F ↾ L,
s)s∈F ↾L, defined by x′
n)n as an F -spreading model and kx0k = kek.
(iii) The sequence (en)n is Schauder basic. In this case (en)n is unconditional.
6.4. F -sequences generating singular spreading models. Let X be a Banach
space and (xn)n be a sequence in X which generates a singular spreading model
(en)n and let en = e′
n + e be the natural decomposition of (en)n. It can be shown
that there exists x ∈ X \ {0} such that kxk = kek and setting x′
n)n is
the unique spreading model of (x′
n)n. In the following we will present an extension
of this fact for F -sequences in a Banach space X. We start with the next lemma.
n = xn − x, (e′
Lemma 6.16. Let F be a regular thin family, M ∈ [N]∞ and (xs)s∈F be an F -
sequence in a Banach space X such that (xs)s∈F ↾M generates a singular F -spreading
model (en)n. Then there exists L ∈ [M ]∞ satisfying the next property.
For every ε > 0 there exists m0 ∈ N such that
(52)
1
n
nXj=1
xsj −
1
m
mXj=1
(cid:13)(cid:13)(cid:13)
xtj(cid:13)(cid:13)(cid:13) < ε,
for every n, m ≥ m0 and (sj)n
t1(1) ≥ L(m).
j=1, (tj)m
j=1 ∈ Plm(F ↾ L) with s1(1) ≥ L(n) and
Proof. Initially we notice that a weaker version of the lemma holds true, that is
for every ε > 0 there exists k0 ∈ N such that for every n, m ≥ k0 and every
(sj)n+m
j=1 ∈ Plm(F ↾ M ) with s1(1) ≥ M (n + m), we have
Indeed, let ε > 0. Since (en)n is singular, it is weakly convergent to some e and
moreover setting e′
n)n is Ces´aro summable to zero. Hence
< ε/4, for all n ≥ n0. Hence for
n = en −e, the sequence (e′
i=1 e′
(53)
1
n
(cid:13)(cid:13)(cid:13)
1
m
xsj −
nXj=1
we may choose n0 ∈ N such that (cid:13)(cid:13)(cid:13) 1
nPn
=(cid:13)(cid:13)(cid:13)
en+i(cid:13)(cid:13)(cid:13)∗
every n, m ≥ n0, we have
mXi=1
nXi=1
(cid:13)(cid:13)(cid:13)
1
m
ei −
(54)
1
n
xsn+j(cid:13)(cid:13)(cid:13) < ε.
mXj=1
i(cid:13)(cid:13)(cid:13)∗
nXi=1
e′
i −
1
m
mXi=1
1
n
Since (xs)s∈F ↾M generates (en)n as an F -spreading model we can find k0 ≥ n0 such
that for every n, m ≥ k0 and every (sj )n+m
j=1 ∈ Plm(F ↾ M ) with s1(1) ≥ M (n + m)
equation (53) is satisfied.
< ε/2
e′
n+i(cid:13)(cid:13)(cid:13)∗
34
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Let (εk)k be a sequence of positive real numbers such thatPk εk < +∞. By the
above we can choose an increasing sequence (nk)k in N such that for every k ∈ N,
n, m ≥ nk and (sj)n+m
j=1 ∈ Plm(F ↾ M ) with s1(1) ≥ M (n + m), it holds that
(55)
xsj −
1
m
1
n
nXj=1
(cid:13)(cid:13)(cid:13)
mXj=1
xsn+j(cid:13)(cid:13)(cid:13) < εk.
We may also assume that F is very large in M and 2nk < nk+1, for every k ∈ N.
We set L = {M (2nk + nk+1) : k ∈ N} and we shall show that L satisfies the
conclusion of the lemma. To this end we shall use an appropriate map sending each
s ∈ F ↾ L to a plegma family in F ↾ M . First, for every s ∈ F ↾ L and p = 1, ..., s,
let k(s(p)) be the unique positive integer k satisfying s(p) = L(k) = M (2nk +nk+1).
We define Φ : F ↾ L → Plm(F ↾ M ) as follows. For every s ∈ F ↾ L, we assign
j is the unique element of F ↾ M ,
where vs
the nk(s(1))-tuple Φ(s) = (cid:0)vs
satisfying
j(cid:1)nk(s(1))
j=1
vs
j ⊑(cid:8)M(cid:0)2nk(s(p)) + nk(s(p))+1 − nk(s(1)) + j(cid:1) : p = 1, . . . , s(cid:9).
j , j = 1, ..., nk(s(1)), follows easily from the fact that F is regular
(56)
The existence of vs
thin and very large in M .
Below we state some useful properties of Φ. Their verification is straightforward.
(P1) For every s ∈ F ↾ L, Φ(s) ∈ Plmnk (F ↾ M ), vs
1(1) > M (nk + nk+1) and
vs
nk = s, where k = k(s(1)),
(P2) for every (s1, s2) ∈ Plm2(F ↾ L), the concatenation Φ(s1)⌢Φ(s2) belongs
to Plm(F ↾ M ).
We are now ready to prove that L is actually the desired set. Fix a positive integer
k and let us denote by s the unique element of F ↾ L such that s ⊑ {L(i) : i ≥ k}.
Notice that s(1) = L(k) = M (2nk + nk+1) and therefore k(s(1)) = k. Also let
mk = max{nk, k + s + 1}. We claim that
(57)
1
nk
nkXj=1
xvs
j
−
1
m
mXj=1
(cid:13)(cid:13)(cid:13)
xtj(cid:13)(cid:13)(cid:13) <
k+sXl=k
εl,
for every m ≥ mk and (tj )m
j=1 ∈ F ↾ L with t1(1) ≥ L(m).
1
, . . . , vwl+1
1 , . . . , vwl
Indeed, let m ≥ mk and (tj )m
j=1 ∈ F ↾ L with t1(1) ≥ L(m). Notice that
max s = L(k + s − 1) < L(m) ≤ t1(1) = min t1. Hence, by Theorem 3.9 there
exists a plegma path (wl)l0
l=0 in F ↾ L from w0 = s to wl0 = t1 of length l0 = s.
Notice that k(wl(1)) ≥ k + l which implies that nk(wl(1)) ≥ nk+l and therefore
length nk+l + nk+l+1 with vwl
Hence by (55) we get
1 , . . . , vwl
(cid:0)vwl
we have that (cid:0)vwl
(cid:13)(cid:13)(cid:13)
nk+l , vwl+1
1 (1) > M (nk(wl(1)) + nk(wl(1)+1)) > M (nk+l + nk+l+1).
nk+l(cid:1) is a subfamily of Φ(wl). Thus, by properties (P1) and (P2) above,
nk+l+1(cid:1) is a plegma family in F ↾ L of
nk+l+1Xj=1
nk+sXj=1
j (cid:13)(cid:13)(cid:13) <
(cid:13)(cid:13)(cid:13) < εk+l,
k+s−1Xl=k
nk+lXj=1
nkXj=1
for every l = 0, ..., l0 − 1. Thus,
nk+l+1
nk+s
xvs
j
−
xv
wl+1
j
(cid:13)(cid:13)(cid:13)
1
nk
xv
wl
j
−
xv
t1
1
nk+l
(59)
(58)
εl.
1
1
HIGHER ORDER SPREADING MODELS
35
1
−
(60)
xv
t1
j
1
m
Now (57) follows by (59) and (60).
Similarly, since m > k + s = k + l0 we have that nm > nk+s. Also since
t1(1) ≥ L(m) = M (2nm + nm+1) we have that k(t1(1)) ≥ m. Hence nk(t1(1)) ≥
Finally, by (57) and a triangle inequality we obtain that
M (2nm + nm+1) ≥ M (nk+s + m) and so again by (55), we have
1 , ..., vt1
nm > nk+s which implies that (cid:0)vt1
Therefore,(cid:0)vt1
nk+sXj=1
nk+s
1 , ..., vt1
nk+s(cid:1) is a proper subfamily of Φ(t1).
nk+s, t1, ..., tm(cid:1) is a plegma family in F ↾ L. Moreover t1(1) ≥
xtj(cid:13)(cid:13)(cid:13) < εk+s.
(cid:13)(cid:13)(cid:13)
mXj=1
xtj(cid:13)(cid:13)(cid:13) < 2
k+sXl=k
mXj=1
nXj=1
(cid:13)(cid:13)(cid:13)
j=1, (tj )m
for every k ∈ N, n, m ≥ mk and (sj)n
j=1 ∈ Plm(F ↾ L) with s1(1) ≥ L(n)
(cid:3)
and t1(1) ≥ L(m). SincePk εk < +∞ the proof is complete.
Theorem 6.17. Let F be a regular thin family, M ∈ [N]∞ and (xs)s∈F be an F -
sequence in a Banach space X such that (xs)s∈F ↾M generates a singular F -spreading
model (en)n. Let en = e′
n + e be the natural decomposition of (en)n. Then there
exist x ∈ X with kxk = kek∗ and L ∈ [M ]∞ such that setting x′
s = xs − x the
F -subsequence (x′
n)n as a unique (up to isometry) F -spreading
model.
s)s∈F ↾L admits (e′
xsj −
1
m
(61)
1
n
εl,
Proof. We start by determining the element x ∈ X. Let L ∈ [M ]∞ satisfying
Lemma 6.16. For every k ∈ N we set
Ak =n 1
n
nXi=1
xsi : (si)n
i=1 ∈ Plm(F ↾ L) and s1(1) ≥ n ≥ ko.
Clearly the sequence (Ak)k is decreasing and by Lemma 6.16, diam(Ak) → 0.
Therefore there exists a unique x ∈ X such that ∩∞
k=1Ak = {x}.
We continue to show that kek = kxk. Notice that by the choice of x, we have
that for every ε > 0 there exists n0 ∈ N such that for all n ≥ n0
For each n ∈ N we pick (sn
have
i )n
i=1 ∈ Plm(F ↾ L), with sn
i (1) ≥ L(n). By (62), we
Also, since (xs)s∈F ↾L generates (en)n as an F -spreading model, we get
Moreover, since (e′
n)n is Ces`aro summable to zero, we have
= 0.
(62)
(63)
(64)
(65)
lim
i
1
n
1
n
xsn
nXj=1
xsj − x(cid:13)(cid:13)(cid:13) < ε.
(cid:13)(cid:13)(cid:13)
n (cid:13)(cid:13)(cid:13)
− x(cid:13)(cid:13)(cid:13) = 0.
nXi=1
ei(cid:13)(cid:13)(cid:13)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)
nXi=1
nXi=1
n (cid:13)(cid:13)(cid:13)Pn
− e(cid:13)(cid:13)(cid:13)∗
i=1 ei
n
= 0.
xsn
1
n
lim
1
n
lim
n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)
36
S. A. ARGYROS, V. KANELLOPOULOS, AND K. TYROS
Hence,
kek∗
(65)
= lim
1
n
nXi=1
n (cid:13)(cid:13)(cid:13)
ei(cid:13)(cid:13)(cid:13)∗
(64)
= lim
xsn
(63)
= kxk.
1
n
nXi=1
n (cid:13)(cid:13)(cid:13)
i(cid:13)(cid:13)(cid:13)
We proceed now to show that (e′
n)n is the unique F -spreading model of (x′
where x′
s = xs − x, s ∈ F ↾ L. Let N ∈ [L]∞ such that (x′
s)s∈F ↾L,
s)s∈F ↾N generates an
(66)
F -spreading model (een)n. We will show that (een)n is isometric to (e′
for every n ∈ N, by (62) we conclude that (een)n is Ces´aro summable to zero. Hence
by Lemma 6.10 it suffices to show that
n)n. Since,
nXj=1
nXj=1
xsj − x,
x′
sj =
1
n
1
n
,
λieei(cid:13)(cid:13)(cid:13)∗∗
(67)
λie′
(cid:13)(cid:13)(cid:13)
nXi=1
i(cid:13)(cid:13)(cid:13)∗
=(cid:13)(cid:13)(cid:13)
nXi=1
for every n ∈ N and λ1, . . . λn ∈ R with Pn
λ1, . . . λn ∈ R withPn
λiei(cid:13)(cid:13)(cid:13)∗
(cid:13)(cid:13)(cid:13)
nXi=1
nXi=1
i=1 λi = 0. Also let ((sk
k→∞(cid:13)(cid:13)(cid:13)
and the proof is complete.
such that limk→∞ sk
1(1) = +∞. Then
=(cid:13)(cid:13)(cid:13)
i(cid:13)(cid:13)(cid:13)∗
nXi=1
= lim
λixsk
λie′
Indeed, let n ∈ N and
i=1 λ1 = 0.
j )n
j=1)k be a sequence in Plmn(F ↾ L)
i(cid:13)(cid:13)(cid:13) = lim
k→∞(cid:13)(cid:13)(cid:13)
nXi=1
λix′
sk
i(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
nXi=1
λieei(cid:13)(cid:13)(cid:13)∗∗
(cid:3)
We close by a strengthening of Theorem 6.17 for Banach spaces with separable
dual. We will need the following lemma.
Lemma 6.18. Let F be a regular thin family and (ys)s∈F be an F -sequence in a
Banach space X. Let L ∈ [N]∞ and suppose that for every ε > 0 and N ∈ [L]∞
there exist k ∈ N, λ1, . . . , λk > 0 and (sj)k
j=1 λj = 1
j=1 λj ysj k < ε. Then for every x∗ ∈ X ∗, ε > 0 and N ∈ [L]∞ there exists
j=1 ∈ Plm(F ↾ N ) such thatPk
M ∈ [N ]∞ such that x∗(ys) < ε for every s ∈ F ↾ M .
and kPk
Proof. Let x∗ ∈ X ∗, ε > 0 and N ∈ [L]∞. By Proposition 2.6 there exists M ∈
[N ]∞ such that exactly one of the following holds.
(a) x∗(ys) < ε, for every
s ∈ F ↾ M , or (b) x∗(ys) ≥ ε, for every s ∈ F ↾ M , or (c) x∗(ys) ≤ −ε, for
every s ∈ F ↾ M . It suffices to show that cases (b) and (c) cannot occur. Indeed,
suppose (b) holds true (the proof for case (c) is similar). By our assumption there
exist k ∈ N, λ1, . . . , λk > 0 and (sj)k
j=1 λj = 1 and
j=1 ∈ Plm(F ↾ M ) such thatPk
j=1 λj ysj k < ε. But then
ε >(cid:13)(cid:13)(cid:13)
kXj=1
λj ysj(cid:13)(cid:13)(cid:13) ≥ x∗(cid:16) kXj=1
λjysj(cid:17) =
kXj=1
λj x∗(ysj ) ≥ ε,
(cid:3)
which is a contradiction.
kPk
(68)
Corollary 6.19. Let X be a Banach space with separable dual. Let F be a regular
thin family, (xs)s∈F be an F -sequence in X and M ∈ [N]∞ such that (xs)s∈F ↾M
generates a singular F -spreading model (en)n. Let en = e′
n + e be the natural
decomposition of (en)n. Then there exist x ∈ X with kxk = kek∗ and N ∈ [M ]∞
HIGHER ORDER SPREADING MODELS
37
such that setting x′
(up to isometry) F -spreading model.
s = xs −x, (x′
s)s∈F ↾N is weakly null and admits (e′
n)n as a unique
s)s∈F ↾L admits (e′
Proof. By Theorem 6.17, there exist x ∈ X with kxk = kek∗ and L ∈ [M ]∞ such
that (x′
n)n as a unique F -spreading model. By applying Lemma
6.18 (for xs in place of ys) and a standard diagonalization for a countable dense
subset of X ∗ we may choose N ∈ [L]∞ such that the F -subsequence (x′
s)s∈F ↾N is
in addition weakly null.
(cid:3)
References
[1] F. Albiac, N. J. Kalton, Topics in Banach Space Theory, Springer, 2006.
[2] D. E. Alspach, S. A. Argyros, Complexity of weakly null sequences, Dissertationes Math.,
(1992).
[3] S.A. Argyros, V. Kanellopoulos and K. Tyros, Higher order Spreading Models in Banach space
theory, preprint, available at http://arxiv.org/abs/1006.0957.
[4] S. A. Argyros, V. Kanellopoulos and K. Tyros, Finite order Spreading Models, preprint, avail-
able at http://arxiv.org/abs/1105.2732.
[5] S.A. Argyros, S. Todorcevic, Ramsey methods in analysis, Birkhauser Verlag, Basel, 2005.
[6] B. Beauzamy, J.-T. Laprest´e, Mod`eles ´etal´es des espaces de Banach, Travaux en Cours, Her-
mann, Paris, 1984, iv+210 pp.
[7] A. Brunel, L. Sucheston On B-convex Banach spaces, Math. Systems Theory 7 (1974), no. 4,
294 -- 299.
[8] A. Dvoretzky, Some results on convex bodies and Banach spaces, 1961 Proc. Internat. Sympos.
Linear Spaces (Jerusalem, 1960) pp. 123 -- 160 Jerusalem Academic Press, Jerusalem;
[9] F. Galvin, K. Prikry, Borel sets and Ramsey's theorem, J. Symbolic Logic, 38, (1973), 193 --
198.
[10] I. Gasparis, A dichotomy theorem for subsets of the power set of the natural numbers, Proc.
Amer. Math. Soc., 129, (2001), no. 3, 759 -- 764.
[11] W.T Gowers, Ramsey Methods in Banach Spaces, in Handbook of the geometry of Banach
Spaces, vol. 2, (2003), 1072-1097.
[12] J. L. Krivine, Sous-espaces de dimension finie des espaces de Banach r´eticul´es, Ann. of Math.
(2) 104 (1976), no. 1, 1 -- 29.
[13] J. Lopez-Abad, S. Todorcevic, Pre-compact families of finite sets of integers and weakly null
sequences in Banach spaces, Topology Appl. 156 (2009), no. 7, 1396 -- 1411.
[14] C.St.J.A. Nash-Williams, On well quasi-ordering transfinite sequences, Proc. Cambr. Phil.
Soc., 61, (1965), 33-39.
[15] E. Odell, Th. Schlumprecht, On the richness of the set of p's in Krivine's theorem, Geometric
aspects of functional analysis (Israel, 1992 -- 1994), 177 -- 198, Oper. Theory Adv. Appl., 77,
Birkhauser, Basel, 1995.
[16] P. Pudlak, V. Rodl, Partition theorems for systems of finite subsets of integers, Discrete
Math. 39, (1982), no. 1, 67 -- 73.
[17] F. P. Ramsey, On a problem of formal logic, Proc. London Math. Soc. 30 (2), (1929), 264286.
(2001)
[18] H. Rosenthal, A characterization of Banach spaces containing l1, Proc. Nat. Acad. Sci. U.S.A.
71 (1974), 2411 -- 2413.
National Technical University of Athens, Faculty of Applied Sciences, Department
of Mathematics, Zografou Campus, 157 80, Athens, Greece
E-mail address: [email protected]
National Technical University of Athens, Faculty of Applied Sciences, Department
of Mathematics, Zografou Campus, 157 80, Athens, Greece
E-mail address: [email protected]
Department of Mathematics, University of Toronto, Toronto, Canada, M5S 2E4
E-mail address: [email protected]
|
1310.2187 | 3 | 1310 | 2015-03-11T17:23:43 | Schur--Agler and Herglotz--Agler classes of functions: positive-kernel decompositions and transfer-function realizations | [
"math.FA"
] | We discuss transfer-function realization for multivariable holomorphic functions mapping the unit polydisk or the right polyhalfplane into the operator analogue of either the unit disk or the right halfplane (Schur/Herglotz functions over either the unit polydisk or the right polyhalfplane) which satisfy the appropriate stronger contractive/positive real part condition for the values of these functions on commutative tuples of strict contractions/strictly accretive operators (Schur--Agler/Herglotz--Agler functions over either the unit polydisk or the right polyhalfplane). As originally shown by Agler, the first case (polydisk to disk) can be solved via unitary extensions of a partially defined isometry constructed in a canonical way from a kernel decomposition for the function (the {\em lurking-isometry method}). We show how a geometric reformulation of the lurking-isometry method (embedding of a given isotropic subspace of a Kre\u{\i}n space into a Lagrangian subspace---the {\em lurking-isotropic-subspace method}) can be used to handle the second two cases (polydisk to halfplane and polyhalfplane to disk), as well as the last case (polyhalfplane to halfplane) if an additional growth condition at $\infty$ is imposed. For the general fourth case, we show how a linear-fractional-transformation change of variable can be used to arrive at the appropriate symmetrized nonhomogeneous Bessmertny\u{\i} long-resolvent realization. We also indicate how this last result recovers the classical integral representation formula for scalar-valued holomorphic functions mapping the right halfplane into itself. | math.FA | math |
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES OF
FUNCTIONS: POSITIVE-KERNEL DECOMPOSITIONS AND
TRANSFER-FUNCTION REALIZATIONS
JOSEPH A. BALL AND DMITRY S. KALIUZHNYI-VERBOVETSKYI
Abstract. We discuss transfer-function realization for multivariable holomor-
phic functions mapping the unit polydisk or the right polyhalfplane into the
operator analogue of either the unit disk or the right halfplane (Schur/Herglotz
functions over either the unit polydisk or the right polyhalfplane) which sat-
isfy the appropriate stronger contractive/positive real part condition for the
values of these functions on commutative tuples of strict contractions/strictly
accretive operators (Schur -- Agler/Herglotz -- Agler functions over either the unit
polydisk or the right polyhalfplane). As originally shown by Agler, the first
case (polydisk to disk) can be solved via unitary extensions of a partially de-
fined isometry constructed in a canonical way from a kernel decomposition
for the function (the lurking-isometry method). We show how a geometric
reformulation of the lurking-isometry method (embedding of a given isotropic
subspace of a Kreın space into a Lagrangian subspace -- the lurking-isotropic-
subspace method) can be used to handle the second two cases (polydisk to
halfplane and polyhalfplane to disk), as well as the last case (polyhalfplane
to halfplane) if an additional growth condition at ∞ is imposed. For the
general fourth case, we show how a linear-fractional-transformation change of
variable can be used to arrive at the appropriate symmetrized nonhomoge-
neous Bessmertnyı long-resolvent realization. We also indicate how this last
result recovers the classical integral representation formula for scalar-valued
holomorphic functions mapping the right halfplane into itself.
1. Introduction
For U, Y coefficient separable Hilbert spaces, we define the operator-valued Schur
class S(U ,Y) (over the unit disk D) to consist of all holomorphic functions S on
the unit disk D with values in the closed unit ball of the space L(U ,Y) of bounded
linear operators from U to Y, i.e., subject to kS(ζ)k ≤ 1 for all ζ ∈ D. The fol-
lowing result linking the theories of holomorphic functions, linear operators, and
input/state/output linear systems is now well known (see e.g. [10] for a full discus-
sion where multivariable extensions are also treated).
Theorem 1.1. Given a function S : D → L(U ,Y), the following are equivalent.
(1) S ∈ S(U ,Y).
2010 Mathematics Subject Classification. 32A10; 47A48; 47A56.
Key words and phrases. Schur -- Agler class; Herglotz -- Agler class; Bessmertnyı long resol-
vent representation; positive-kernel decomposition; transfer-function realization; lurking-isometry
method; lurking-isotropic-subspace method.
The authors were partially supported by US -- Israel BSF grant 2010432. The second author
was also partially supported by NSF grant DMS-0901628, and wishes to thank the Department
of Mathematics at Virginia Tech for hospitality during his sabbatical visit in January -- May 2013,
when a significant part of work on the paper was done.
1
2
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
(2) The de Branges -- Rovnyak kernel
KS(ω, ζ) =
I − S(ω)∗S(ζ)
1 − ωζ
is a positive kernel on D, i.e., there is an auxiliary Hilbert space X and a
holomorphic function H : D → L(U ,X ) which gives rise to a Kolmogorov
decomposition for KS:
(3) S has a unitary transfer-function realization, i.e., there is an auxiliary
KS(ω, ζ) = H(ω)∗H(ζ).
Hilbert state space X and a unitary colligation matrix
U =(cid:20)A B
C D(cid:21) : (cid:20)X
U(cid:21) →(cid:20)X
Y(cid:21)
so that
(3′) Condition (3) above holds where the colligation matrix U is taken to be any
S(ζ) = D + ζC(I − ζA)−1B for ζ ∈ D.
of (i) coisometric, (ii) isometric, or (iii) contractive.
It is natural to seek extensions of the Schur class to the multivariable setting
where the disk D is replaced by the polydisk
Dd = {ζ = (ζ1, . . . , ζd) ∈ Cd : ζk < 1 for k = 1, . . . , d}.
We therefore define the d-variable Schur class Sd(U ,Y) to consist of holomorphic
functions S : Dd → L(U ,Y) subject to kS(ζ)k ≤ 1 for all ζ ∈ Dd. It was the pro-
found observation of Agler [1] that, unless d ≤ 2, a characterization of Sd(U ,Y) of
the same form as Theorem 1.1 is not possible. Instead, we define what is now called
the Schur -- Agler class, denoted as SAd(U ,Y), to consist of holomorphic functions
S : Dd → L(U ,Y) such that kS(T1, . . . , Td)k ≤ 1 whenever T = (T1, . . . , Td) is a
commutative d-tuple of strict contraction operators on a fixed separable infinite-
dimensional Hilbert space K. Here the functional calculus defining S(T1, . . . , Td)
can be given by
S(T1, . . . , Td) = Xn∈Zd
+
Sn ⊗ T n
Snζn is the
multivariable Taylor expansion for S centered at the origin 0 ∈ Dd and where we
use standard multivariable notation:
(convergence in the strong operator topology) where S(ζ) = Pn∈Zd
if n = (n1, . . . , nd) ∈ Zd
+.
1 ··· ζnd
d ,
T n = T n1
1
··· T nd
ζn = ζn1
d
+
The following result due to Agler [1] (see also [2, 13]) has had a profound impact
on the subject over the years.
Theorem 1.2. Given a function S : Dd → L(U ,Y), the following are equivalent.
(1) S ∈ SAd(U ,Y).
(2) S has an Agler decomposition in the sense that there exists d L(Y)-valued
positive kernels K1, . . . , Kd on Dd such that
I − S(ω)∗S(ζ) =
dXk=1
(1 − ωkζk)Kk(ω, ζ).
(1.1)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
3
(3) S has a unitary Givone -- Roesser d-dimensional transfer-function realiza-
tion, i.e., there is an auxiliary Hilbert state space X with a d-fold orthog-
onal direct-sum decomposition X = X1 ⊕ ··· ⊕ Xd together with a unitary
colligation matrix
U =(cid:20)A B
C D(cid:21) : (cid:20)X
U(cid:21) →(cid:20)X
Y(cid:21)
so that
S(ζ) = D + C(I − P (ζ)A)−1P (ζ)B for ζ ∈ Dd.
(1.2)
where we have set
P (ζ) = ζ1P1 + ··· + ζdPd
where Pk is the orthogonal projection of X onto Xk for each k = 1, . . . , d.
(3′) Condition (3) above holds where the colligation matrix U is taken to be any
of (i) coisometric, (ii) isometric, or (iii) contractive.
The goal of this paper is to study parallel results for an assortment of linear-
fractional transformed versions of the Schur -- Agler class. Specifically, we seek anal-
ogous characterizations of the following classes of holomorphic functions:
(1) The Herglotz -- Agler class over the unit polydisk Dd, denoted HA(Dd,L(U))
defined as the class of all holomorphic functions F : Dd → L(U) such that
F (T ) has positive real part
F (T ) + F (T )∗ ≥ 0
for all commutative d-tuples T = (T1, . . . , Td) of strict contractions on the
Hilbert space K.
(2) The Schur -- Agler class over the right polyhalfplane
Πd = {z = (z1, . . . , zd) ∈ Cd : zk + zk > 0 for k = 1, . . . , d},
denoted by SA(Πd,L(U ,Y)), consisting of all holomorphic L(U ,Y)-valued
functions s on Πd such that ks(A)k ≤ 1 for all strictly accretive com-
mutative d-tuples A = (A1, . . . , Ad) of operators on K (i.e., such that
Ak + A∗k ≥ cI for some constant c > 0, k = 1, . . . , d).
(3) The Herglotz -- Agler class over Πd, denoted by HA(Πd,L(U)), consisting
of all holomorphic functions f on Πd such that f (A) + f (A)∗ ≥ 0 for all
strictly accretive commutative d-tuples A of operators on K.
(4) The Nevanlinna -- Agler class over the upper polyhalfplane
zk − zk
2i
> 0 for k = 1, . . . , d},
denoted by NA((iΠ)d,L(U)), consisting of all holomorphic L(U)-valued
(iΠ)d = {z = (z1, . . . , zd) ∈ Cd :
2i (ef (eA) − ef (eA)∗) ≥ 0 whenever eA =
functions ef on (iΠ)d such that 1
(eA1, . . . , eAd) is a commutative d-tuple of operators on K, each with strictly
positive-definite imaginary part (i.e., such that 1
constant c > 0, k = 1, . . . , d).
2i (Ak − A∗k) ≥ cI for some
To be consistent with the more detailed notation used for these variants of the
Schur -- Agler class, we will also use the notation SA(Dd,L(U ,Y)) for the Schur --
Agler class SAd(U ,Y) over the polydisk Dd discussed above.
4
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
We note that our convention is to use the term Herglotz for functions with values
having positive real part, and Nevanlinna for functions with values having positive
imaginary part; we recognize that these conventions are by no means universal (see
e.g. [6]).
For the single-variable case such realization results have been explored in a sys-
tematic way in [35] and [12]. For the multivariable setting, apart from the now clas-
sical Schur -- Agler class over the polydisk SAd(U ,Y), the only results along these
lines which we are aware of are those in the recent paper of Agler -- McCarthy -- Young
[3] and of Agler -- Tully-Doyle -- Young [4, 5].
The approach in [35] (in the single-variable setting) is to use a linear-fractional-
transformation (LFT) change of variables (on the domain and/or the range side)
to reduce the desired result to the corresponding result for the Schur class over the
unit disk. This is also the main tool in [3, 4, 5]: use an LFT Cayley-transform
change of variables to reduce results for the Nevanlinna -- Agler class to the corre-
sponding known results for the Schur -- Agler class. However the procedure is rather
intricate due to the added subtleties involved in handling points at infinity in the
multivariable case.
In contrast, the approach in [12] is to apply a projective version of the lurk-
ing isometry argument (roughly, a lurking-isotropic-subspace argument in a Kreın-
space setting) to arrive at the desired realization result via a direct but unified
Kreın-space geometric argument. One of the main contributions of the present
work is to extend this approach to the multivariable setting. The main difficulty is
to guarantee that a naturally defined isotropic subspace is actually a graph space
with respect to a system of coordinates not coming from a fundamental decompo-
sition of the ambient Kreın space. We show how this difficulty can be overcome
for the case of the Dd-Herglotz -- Agler class and Πd-Schur -- Agler class. For the Πd-
Herglotz -- Agler class, we are able to overcome the difficulty only in a special case
(associated with the imposition of a growth condition at infinity), thereby recov-
ering parallel results from [3]. For the most general Πd-Herglotz -- Agler function
f , we follow the LFT change-of-variable approach of [4] combined with the more
general realization formalism (Schur complement of an operator pencil) suggested
by the work of Bessmertnyı (see [18, 19, 20, 21, 28]) to arrive at a realization for-
mula for the most general Πd-Herglotz -- Agler function. We note that the original
Bessmertnyı class involved additional symmetries leading to strong rigidity results.
It was conjectured in [9] that an appropriate weakening of the metric conditions
for the Bessmertnyı operator pencil should lead to a representation for the most
general Πd-Herglotz -- Agler function. Here we show that this conjecture is correct
once one identifies the appropriate modification: one must allow the nonhomoge-
neous skew-adjoint term in the nonhomogeneous Bessmertnyı operator pencil to be
unbounded (more precisely, a certain flip Π-impedance-conservative system node
in the sense of [35]).
There has been a lot of work on transfer-function realization for the single-
variable Schur and Herglotz classes over the right half plane. The most influential
for our point of view toward multivariable generalizations is the work of Arov-
Nudelman [7] and of Staffans and collaborators (see [34, 35, 12, 29, 37] as well as
the treatise [36] and the references there). There is also a complementary approach
to such realization theory (upper halfplane rather than right halfplane version)
with emphasis on the theory of selfadjoint extensions of densely defined symmetric
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
5
operators on a Hilbert space (see [14, 15, 16] as well as the recent book [6] and the
references there).
The paper is organized as follows. Section 2 highlights the main ideas from
Kreın-space geometry and from infinite-dimensional systems theory, in particular,
the idea of a system node, which will be used in the later sections. Section 3 presents
our results for the Herglotz -- Agler class over the polydisk while Section 4 does the
same for the Schur -- Agler class over the right polyhalfplane. Section 5 presents our
results for the restricted Herglotz -- Agler class over the right polyhalfplane where a
growth condition at infinity is imposed on the functions to be realized. With this
added restriction, the lurking-isotropic-subspace method from [12] adapts well to
lead to a classical type realization (but with an in general unbounded Π-impedance-
conservative system node) for the Πd-Herglotz -- Agler function. Section 6 identifies
the nonhomogeneous unbounded Bessmertnyı operator pencils which then lead to a
realization for the most general Herglotz -- Agler function over Πd. We also mention
that the results parallel to the results of this paper for the four classes under
discussion (Dd-Schur -- Agler, Dd-Herglotz -- Agler, Πd-Schur -- Agler and Πd-Herglotz --
Agler) for the rational matrix-valued (Cayley) inner case, where the emphasis is
on obtaining realizations with finite-dimensional state space, are obtained in our
companion paper [11].
We shall have occasion to need a Cayley transform (with both scalar and operator
argument) acting between the right halfplane and the unit disk. Following the
conventions in [28] and [11], we shall make use of the following version:
ζ ∈ D 7→ w =
w ∈ Π 7→ ζ =
1 + ζ
1 − ζ ∈ Π, with inverse given by
w − 1
w + 1 ∈ D.
(1.3)
For ζ = (ζ1, . . . , ζd) a point in the unit polydisk Dd, we continue to use the notation
for the corresponding point in the right polyhalfplane Πd. Similarly, given a point
w = (w1, . . . , wd) in the right polyhalfplane Πd, we use the notation
1 + ζ
1 − ζ
:=(cid:18) 1 + ζ1
1 − ζ1
, . . . ,
1 + ζd
1 − ζd(cid:19)
w − 1
w + 1
:=(cid:18) w1 − 1
w1 + 1
, . . . ,
wd − 1
wd + 1(cid:19)
(1.4)
(1.5)
for the associated point in the polydisk.
Acknowledgement. The authors are thankful to the anonymous referee for his/her
careful reading and constructive remarks.
2. Preliminaries
2.1. Decompositions of the identity. Given a Hilbert space X , we shall say
that a collection of d operators (Y1, . . . , Yd) on X forms a d-fold positive decom-
position of the identity IX if each Yk is a selfadjoint contraction (0 ≤ Yk ≤ IX
k=1 Yk = IX ). In case
(Y1, . . . , Yd) = (P1, . . . , Pd) consists of orthogonal projection operators (necessar-
ily with pairwise orthogonal ranges) we shall say that (P1, . . . , Pd) forms a d-fold
spectral decomposition of IX . Note that d-fold spectral decompositions (P1, . . . , Pd)
for 1 ≤ k ≤ d) which together sum up to the identity (Pd
6
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
arise in the realization formula for the Schur -- Agler class in Theorem 1.2. We shall
see that the more general positive decompositions are needed for the realization
formulas for functions in the Πd-Schur -- Agler class and in the Πd-Herglotz -- Agler
class, as already discovered in [3, 4, 5].
From the definitions we see that any spectral decomposition (P1, . . . , Pd) is
also a positive decomposition. There is also a result in the converse direction:
if (Y1, . . . , Yd) is a positive decomposition of IX , then there exist a Hilbert space eX ,
, and an isometric embedding ι : X → eX
a spectral decomposition (P1, . . . , Pd) of I eX
such that
Yk = ι∗Pkι for k = 1, . . . , d.
This can be seen as a consequence of the Naimark dilation theorem (apply [30,
Theorem 4.6] with the measurable space X taken to be the finite set {k ∈ N : 1 ≤
k ≤ d}). To prove the result for this simple case of the Naimark dilation theorem,
i=1 X by
simply define an isometric embedding of X intoLd
ι =
onto the k-th block in the direct-sum spaceLd
Q1
...
Qd
i=1 X .
where Qk provides a factorization Yk = Q∗kQk and take Pk equal to the projection
2.2. Basics on the geometry of Kreın spaces. In this Section we review some
basics about the geometry of Kreın spaces and Kreın-space operator theory which
we shall need in the sequel. Other resources on this topic is a similar survey section
in the paper [12] as well as the more complete treatises [22, 8].
A Kreın space by definition is a linear space K endowed with an indefinite inner
product [·,·] which is complete in the following sense: there are two subspaces K+
and K− of K such that the restriction of [·,·] to K+ × K+ makes K+ a Hilbert
space while the restriction of −[·,·] to K− × K− makes K− a Hilbert space, and
K = K+[ +]K− is a [·,·]-orthogonal direct sum decomposition of K. In this case the
decomposition K = K+[ +]K− is said to form a fundamental decomposition for K.
Fundamental decompositions are never unique except in the trivial case where one
of K+ or K− is equal to the zero space.
Unlike the case of Hilbert spaces where closed subspaces all look the same, there
is a rich geometry for subspaces of a Kreın space. A subspace M of a Kreın space
K is said to be positive, isotropic, or negative depending on whether [u, u] ≥ 0 for
all u ∈ M, [u, u] = 0 for all u ∈ M (in which case it follows that [u, v] = 0 for all
u, v ∈ M as a consequence of the Cauchy-Schwarz inequality), or [u, u] ≤ 0 for all
u ∈ M. Given any subspace M, we define the Kreın-space orthogonal complement
M[⊥] to consist of all v ∈ K such that [u, v] = 0 for all u ∈ K. Note that the
statement that M is isotropic is just the statement that M ⊂ M[⊥]. If it happens
that M = M[⊥], we say that M is a Lagrangian subspace of K.
Examples of such subspaces arise from placing appropriate Kreın-space inner
products on the direct sum H1 ⊕ H2 of two Hilbert spaces and looking at graphs
of operators of an appropriate class.
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
7
Example 2.1. Suppose that H′ and H are two Hilbert spaces and we take K to
be the external direct sum H′ ⊕ H with inner product
(cid:20)(cid:20)x
y(cid:21) ,(cid:20)x′
y′(cid:21)(cid:21) =(cid:28)(cid:20)IH′
0 −IH(cid:21)(cid:20)x
0
y(cid:21) ,(cid:20)x′
y′(cid:21)(cid:29)H′⊕H
where h·,·iH′⊕H is the standard Hilbert-space inner product on the direct-sum
Hilbert space H′ ⊕ H. In this case it is easy to find a fundamental decomposition:
take K+ =(cid:2) H
{0}(cid:3) and K− =h {0}
H to H′ and let M be the graph of T :
H′i. Now let T be a bounded linear operator from
M = GT =(cid:26)(cid:20)T x
x(cid:21) : x ∈ H(cid:27) ⊂ K.
Then a good exercise is to work out the following facts:
• GT is negative if and only if kTk ≤ 1.
• GT is isotropic if and only if T is isometric (T ∗T = IH).
• GT is Lagrangian if and only if T is unitary: T ∗T = IH and T T ∗ = IH′ .
Example 2.2. Let H be a Hilbert space and set K equal to the direct-sum space
K = H ⊕ H with indefinite inner product given by
In this case a choice of fundamental decomposition is not so obvious; one such
choice is
IH
IH
0(cid:21)(cid:20)x
(cid:20)(cid:20)x
y(cid:21) ,(cid:20)x′
K+ =(cid:26)(cid:20)x
y′(cid:21)(cid:21) =(cid:28)(cid:20) 0
y(cid:21) ,(cid:20)x′
x(cid:21) : x ∈ H(cid:27) , K− =(cid:26)(cid:20)−y
y′(cid:21)(cid:29)H⊕H
y (cid:21) : y ∈ H(cid:27) .
.
Then the exercise parallel to that suggested in Example 2.1 is to work out the
following: given a closed operator T ∈ L(H) with dense domain D(T ) ⊂ H, let
GT =(cid:26)(cid:20)T x
x(cid:21) : x ∈ D(T )(cid:27) ⊂ K
be its graph space. Then:
• GT is negative if and only if T is dissipative: hT x, xi + hx, T xi ≤ 0 for all
x ∈ D(T ).
• GT is maximal negative, i.e., GT is negative and is not contained in any
properly larger negative subspace, if and only if T is maximal dissipative,
i.e., T is dissipative and has no proper dissipative extension. An equivalent
condition is T is dissipative and the operator I + T is onto (or equivalently
wI + T is onto for all w in the right halfplane Π+) (see [31]).
• GT is isotropic if and only if T is skew-symmetric or T ⊂ −T ∗, i.e. hT x, xiH+
hx, T xiH = 0 for all x ∈ D(T ).
• GT is Lagrangian if and only if T is skew-adjoint or T = −T ∗ (i.e., T ⊂ −T ∗
and T and T ∗ have the same domain: y, z ∈ H such that hT x, yiH = hx, ziH
implies that y ∈ D(T ) and z = −T y. A closely related result is proved in
Corollary 2.7 below.
We shall have use for the following connection between Examples 2.1 and 2.2.
Consider Example 2.1 for the case where H′ = H and call this Kreın space K1. Let
8
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
K2 be the Kreın space H ⊕ H with Kreın-space inner product the negative of the
inner product as in Example 2.2:
(cid:20)(cid:20)x
y(cid:21) ,(cid:20)x′
y′(cid:21)(cid:21)K2
=(cid:28)(cid:20) 0
−IH
−IH
0 (cid:21)(cid:20)x
y(cid:21) ,(cid:20)x′
y′(cid:21)(cid:29)H⊕H
.
Note that this modification of the inner product just interchanges positive and
negative subspaces and preserves isotropic and Lagrangian subspaces. Then the
operator Γ defined by
Γ :=
1
√2(cid:20) IH
−IH IH(cid:21) : (cid:20)H
H(cid:21) →(cid:20)H
H(cid:21)
IH
defines a Kreın-space isomorphism between the Kreın space K1 as in Example 2.2
and the Kreın space K2 as in Example 2.1; this is a consequence of the bijectivity
of Γ along with the identity
Γ∗(cid:20) 0 −I
0 (cid:21) Γ =(cid:20)I
−I
0 −I(cid:21) .
0
It follows that Γ maps K1-Lagrangian subspaces to K2-Lagrangian subspaces. In
particular, if U is unitary and does not have 1 as an eigenvalue, we define the Cayley
transform Y of T according to the formula
D(Y ) = Ran(I − U ) and Y = (I + U )(I − U )−1.
(2.1)
Note that
ΓGU =
I
1
−I
x (cid:21) : x ∈ H(cid:27)
√2(cid:20) I
I(cid:21)(cid:26)(cid:20)U x
=(cid:26)(cid:20)(I + U )x
(I − U )x(cid:21) : x ∈ H(cid:27)
=(cid:26)(cid:20)(I + U )(I − U )−1
=(cid:26)(cid:20) Y
IH(cid:21) y : y ∈ D(Y )(cid:27) = GY .
I
(cid:21) (I − U )x : x ∈ H(cid:27)
There results the following fact concerning the Cayley transform map of this type.
Proposition 2.3. Suppose that U is a linear operator on a Hilbert space H which
does not have 1 as an eigenvalue. Define the Cayley transform Y of U as in (2.1).
Then U is unitary if and only if Y is skew-adjoint.
2.3. Well-posed linear systems and system nodes. A continuous-time in-
put/state/output (i/s/o) linear system is a system of equations of the form
Σ : (cid:26) x(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)
(2.2)
where x(t) takes values in the state space X , u(t) takes values in the input space U,
and y(t) takes values in the output space Y. Under the assumption that the system
matrix (sometimes also called the colligation matrix)
Σ :=(cid:20)A B
C D(cid:21) : (cid:20)X
U(cid:21) →(cid:20)X
Y(cid:21)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
9
consists of bounded operators, imposition of the initial condition x(0) = 0 and
application of the Laplace transform
bx(w) =Z ∞
0
e−wtx(t) dt,
leads to the input-output relation in the frequency domain
by(w) =Z ∞
0
e−wty(t) dt
e−wtu(t) dt,
0
bu(w) =Z ∞
by(w) = TΣ(w)bu(w)
where
is the transfer function of the linear system Σ (2.2). The converse question of when
an L(U ,Y)-valued function f on the right halfplane Π can be realized as f = TΣ
TΣ(w) = D + C(wI − A)−1B
C D(cid:21) has generated much interest over the years. The
for some system Σ = (cid:20)A B
obvious necessary condition is that w 7→ f (w) be analytic on some right halfplane
but this is not sufficient if we limit our attention to systems Σ consisting only of
bounded operators A, B, C, D. While it is clear that the right generalization of
the state operator A is that it should be the generator of a C0-semigroup, exactly
how to handle the remaining operators B, C, D so as to get a meaningful theory
containing compelling examples of interest was not so clear, but some progress was
made already in the 1970s (see [26, 27]). It is now understood that a useful notion
of generalized system matrix Σ is that associated with so-called well-posed system.
Roughly, a well-posed linear system is an i/s/o linear system for which the integral
form of the system operators A, B, C, D satisfy natural compatibility conditions and
the integral form of the system matrix
Ct
0
0
Dt
0(cid:21) : (cid:20) x(0)
(cid:20)At Bt
of the system matrix Σ =(cid:20)A B
C D(cid:21) appearing in (2.2)) is the notion of system node
makes sense as a bounded operator from X ⊕ L2
([0, t)) for each
U
t > 0 (see [36] for complete details). The "right" infinitesimal object (the analogue
u[0,t)(cid:21) 7→(cid:20) x(t)
y[0,t)(cid:21)
defined as follows; this notion is well laid out in the work of Staffans [36, 35] where
it is acknowledged that much of the idea was already anticipated in the earlier work
of Salamon [32] and Smuljan [33].
([0, t)) to X ⊕ L2
Y
(2.3)
We first make some preliminary observations. A system node Σ is still an oper-
ator from X ⊕ U to X ⊕ Y but now allowed to be unbounded with some domain
D(Σ) ⊂ X ⊕ U. We may then split Σ in the form
Σ2(cid:21)
Σ =(cid:20)Σ1
where Σ1 : D(Σ) → X and Σ2 : D(Σ) → Y. However we allow the possibility that
D(Σ) does not split D(Σ) =(cid:20)D(Σ)1
D(Σ)2(cid:21) as the direct sum of a linear manifold D(Σ)1
in X with a linear manifold D(Σ)2 in U. To keep the parallel with the classical
case, we therefore write
Σ =(cid:20)A&B
C&D(cid:21)
10
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
splitting (cid:2)AX B(cid:3) where AX and B are operators mapping the spaces X and U
with the notation A&B (and similarly C&D) suggesting that the common domain
of A&B and C&D in X ⊕ U may not have a splitting. However A&B will have a
into a larger "rigged" space X−1 which is part of a so-called Gelfand triple defined
as follows.
We assume that A is an in general unbounded closed operator with dense domain
D(A) := X1 in X and with nonempty resolvent set. Then X1 is a Hilbert space in
its own right with respect to the X1-norm given by
kxk1 := k(αI − A)xkX
where α is any fixed number in the resolvent set of A. While the norm depends
on the choice of α, any other choice α′ of α leads to the same space but with an
equivalent norm, as long as α and α′ are in the same connected component of the
resolvent set of A. If A is the generator of a C0-semigroup (the most interesting
case for us), one can take the connected component of the resolvent set to be a
right half plane. Note that (αI − A)X1 can be viewed as an isometry from X1 onto
X .
We next construct another Hilbert space X−1 as the completion of X in the norm
kxk−1 := k(αI − A)−1xkX .
If {xn}n∈Z+ is a sequence in X1 converging to x ∈ X in X -norm, then the sequence
{(αI − A)xn}n∈Z+ is Cauchy in X−1-norm and hence converges to an element
x−1 ∈ X−1. One can check that this element x−1 ∈ X−1 is independent of the
choice of sequence {xn} ⊂ X1 converging in X -norm to x; we denote this element
y ∈ X−1 by y = (αI − A)x and then define an extension AX : X → X−1 by
AX : x 7→ αx − (αI − A)x ∈ X−1 if x ∈ X .
We then have that the extended operator (αI − A)X is an isometry from X onto
X−1 and we have the nested inclusions X1 ⊂ X ⊂ X−1 with continuous and dense
injections. We will on occasion simplify the notation AX to simply A when the
meaning is clear; thus for x ∈ X and x not necessarily in X1 = D(A), the element
Ax is still defined but as an element of X−1.
It is also useful to note the role of these spaces in duality pairings. First, we
note that the constructions in the previous paragraph can be carried out using the
1 = (αI − A∗)−1X
operator A∗ in place of A. When this is done we get spaces X ⋆
(with the norm kxk1,⋆ := k(αI − A∗)xkX and X ⋆
−1 equal to the completion of X in
−1-norm kxk−1,⋆ = k(αI − A∗)−1xkX with the properties that (αI − A∗)−1
the X ⋆
is an isometry from X onto X ⋆
1 and (αI − A∗) extends to an isometry from X onto
X ⋆
−1 with the nesting X ⋆
−1. Given any x ∈ X , we can view x as a linear
functional on X ⋆
1 using the X -pairing:
1 ⊂ X ⊂ X ⋆
ℓx(x⋆
1) := hx⋆
1, xiX for x⋆
1 ∈ X ⋆
1 .
(2.4)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
11
If we write x⋆
1 = (αI − A∗)−1y with y ∈ X , then
ℓx(x⋆
1) = hx⋆
1, xiX
= h(αI − A∗)−1y, xiX
= hy, (αI − A)−1xiX
≤ kykXk(αI − A)−1xkX
= kx⋆
1k1,⋆kxk−1
with equality if we take x⋆
conclude that the linear-functional norm of ℓx is equal to the X−1-norm of X :
1 = (αI − A∗)−1(αI − A)−1x (or y = (αI − A)−1x). We
kℓxk(X ⋆
1 )∗ = kxk−1
1 )∗. It is not difficult to see that this
and X can be identified with a subspace of (X ⋆
subspace is dense and hence, after taking completions, we have that X−1 is naturally
isomorphic to the dual space of X ⋆
1 via the X -pairing (2.4). For our application to
infinite-dimensional linear systems, in practice the unbounded closed operator A in
this construction will also be taken to be the generator of a C0-semigroup, so as to
make sense of a differential equation of the form dx
dt (t) = Ax(t).
We are now ready to introduce the notion of system node.
Definition 2.4. By a system node Σ on the collection of three Hilbert spaces
(U ,X ,Y), we mean a linear operator
such that:
Σ :=(cid:20)A&B
C&D(cid:21) : (cid:20)X
U(cid:21) ⊃ D(Σ) →(cid:20)X
Y(cid:21)
(1) Σ is a closed operator with domain D(Σ) dense in(cid:20)X
U(cid:21),
(2) If we define an operator A with domain D(A) = {x ∈ X : [ x
Ax = Σ [ x
0 ] for x ∈ D(A),
0 ] ∈ D(Σ)} by
(2.5)
then A is the generator of a C0-semigroup on X.
(3) Let AX : X → X−1 be the extension of the operator A : X1 = D(A) →
X (as in (2.5)) as described in the previous paragraph. Then there is a
bounded linear operator B : U → X−1 so that we recover the operator
] → X−1 to D(Σ):
A&B as the restriction of the operator(cid:2)AX B(cid:3) : [ X
U
Σ =(cid:2)AX B(cid:3) D(Σ).
(4) C&D is a bounded operator from D(Σ) to Y
C&D ∈ L(D(Σ),Y)
where D(Σ) carries the graph norm.
(5) The domain D(Σ) of Σ is characterized as
D(Σ) =(cid:26)(cid:20)x
U(cid:21) : AX x + Bu ∈ X(cid:27) .
Given u ∈ U , there exists xu ∈ D(Σ) so that [ xu
A consequence of this definition of system node is the following fact:
u ] ∈ D(Σ).
u(cid:21) ∈(cid:20)X
(2.6)
(2.7)
12
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
in part (5) of Definition 2.4 to check that this [ xu
Indeed, it suffices to take xu =(cid:0)(αI − A)X(cid:1)−1
u ] = AX(cid:0)(αI − A)X(cid:1)−1
(cid:2)AX B(cid:3) [ xu
= (A − αI)X(cid:0)(αI − A)X(cid:1)−1
= α ·(cid:0)(αI − A)X(cid:1)−1
Bu ∈ X .
Bu + Bu
Bu. To see this, we use the criterion
u ] is in D(Σ):
Bu + α ·(cid:0)(αI − A)X(cid:1)−1
Bu + Bu
In the definition of system node, we took pains to write the top component in
] does not split; yet in the end
the form A&B to indicate that its domain in [ X
U
we found a splitting by extending to a larger space(cid:2) X−1
restriction of an extended operator (cid:2)AX B(cid:3) : (cid:2) X−1
split. Similarly, there is at least a partial splitting for the operator C&D : D(Σ) →
Y. Indeed, we have seen that X1 = D(A) = {x ∈ X : [ x
0 ] ∈ D(Σ} is a dense subset
of X . We may therefore define an operator C : X1 → Y by
U (cid:3) and writing A&B as the
U (cid:3) → Y whose domain does
Cx = C&D [ x
0 ] for x ∈ X1.
(2.8)
Since C&D is a bounded operator from D(Σ) (graph norm) to Y, it follows that
C so defined is a bounded operator from X1 (graph norm induced by the operator
A) to Y. In practice we assign no independent meaning to the D in C&D except
under some additional hypotheses (e.g., for the case of a regular system -- see the
paper of Weiss [39] or [36, Section 5.6]). To this point we have at least versions of
all the usual constituents for a linear input/state/output linear system:
A : X1 → X main operator or state dynamics,
B : U → X−1 input or control operator,
C : X1 → Y output or observation operator,
C&D : D(Σ) → Y combined observation/feedthrough operator,
(2.9)
and we lack in general an independent well-defined feedthrough operator D.
The formula for xu in (2.7) can use any point w in the connected component (e.g.,
an appropriate right half plane) of the resolvent set of A containing α. For clarity,
Bu to indicate the dependence of xu on the
follows that the expression
let us write xu(w) =(cid:0)(wI − A)X(cid:1)−1
point w in the right half plane. From the fact that(cid:2) xu(w)
TΣ(w) : u 7→ C&D(cid:20)xu(w)
u (cid:21) = C&D(cid:20)(cid:0)(wI − A)X(cid:1)−1
(cid:3) ∈ D(Σ) = D(C&D) it
(cid:21)
Bu
u
u
is well defined and defines the transfer function of the system node. Let α be any
fixed point in the resolvent set of A. Then we can recover the value of the transfer
function at any point w in the same connected component of the resolvent set of A
from its value at the fixed point α according to the recipe
TΣ(w)u = (TΣ(w) − TΣ(α)) u + TΣ(α)u
= C(xu(w) − xu(α)) + TΣ(α)u
=(cid:16)(α − w)C(wI − A)−1(cid:0)(αI − A)X(cid:1)−1
B + TΣ(α)(cid:17) u
Conversely, start with any semigroup generator A on X with domain D(A) = X1
with induced Gelfand rigging X1 ⊂ X ⊂ X−1, any input operator B : U → X−1,
(2.10)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
13
and an output operator C : X1 → Y along with a value TΣ(α) ∈ L(U ,Y) for the
transfer function at the point α. Define
D(Σ) =(cid:26)(cid:20)x
u(cid:21) : AX x + Bu ∈ X(cid:27) with
A&B =(cid:2)AX B(cid:3) D(Σ), C&D [ x
u ] = C(x − xu(α)) + TΣ(α)u.
(2.11)
Then Σ so defined is a system node with value of its transfer function at α equal to
the prescribed value TΣ(α). Note here that x− xu(α) is in X1 since [ x
are in D(Σ) and hence so also is
u ] and(cid:2) xu(α)
(cid:3)
u
(cid:20)x − xu(α)
0
(cid:21) =(cid:20)x
u(cid:21) −(cid:20)xu(α)
u (cid:21)
(2.12)
resulting in x − xu(α) ∈ X1. Moreover, we recover the transfer function TΣ at a
general point from A, B, C and TΣ(α) via the formula (2.10). (see [34, Lemma 2.2]
for more complete details).
Given a system node Σ, it is possible to make sense of the associated system of
differential equations
(cid:20) x(t)
y(t)(cid:21) = Σ(cid:20)x(t)
u(t)(cid:21) ,
as long as u ∈ C2([0,∞);U) and(cid:20)x(0)
u(0)(cid:21) ∈ D(Σ) (see [29, Proposition 2.6]). Appli-
x(0) = x0
(2.13)
cation of the Laplace transform
0
e−wtx(t) dt
x(t) 7→bx(w) :=Z ∞
by(w) = C(wI − A)−1x(0) + TΣ(w)bu(w)
to the system equations (2.13) leads us to the input-output property of the transfer
function (2.10):
for w with sufficiently large real part.
We mention that any well-posed linear system (2.3) is the integral form of the
dynamical system associated with a system node (see e.g. [36]); however there are
system nodes for which the associated dynamical system (2.13) is not well-posed
(i.e., one or more of the block operators Bt
0 appearing in (2.3) fail to exist
as bounded operators between the appropriate spaces), despite the fact that the
infinitesimal form of the system equations (2.13) does make sense.
0, Dt
0, Ct
The following examples of system nodes will be useful in the sequel.
In this
discussion we make use of Kreın-space geometry notions discussion in Section 2.2.
Example 2.5. Suppose that Σ = (cid:20)A&B
operator such that its graph space
Y(cid:21) is a closed
C&D(cid:21) : (cid:20)X
U(cid:21) ⊃ D(Σ) → (cid:20)X
D(Σ) ⊂
X
Y
X
U
A&B
C&D
0
I
0 I
G(Σ) :=
14
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
is a Lagrangian subspace of K := X ⊕ Y ⊕ X ⊕ U, where K is given a Kreın space
structure using the signature operator
J =
0
0
IX
0
0
IY
0
0
0
IX
0
0
0
0
0 −IU
.
Ct
0
0
Dt
Then Σ is a system node (see [12, Proposition 4.9]). In fact, Schur-class functions
over the right halfplane Π are characterized as those functions s having a realization
s(w) = TΣ(w) as in (2.10) with a system node Σ of this form (see [12, Theorem
4.10]). These systems are also characterized by the energy-balance property that
the block matrixh At Bt
0i associated with the integral form (2.3) of the system equa-
tions is unitary for each t (see [35]). For this reason any such system node is said to
be a Π-scattering conservative system node. In particular, any Π-scattering conser-
vative system is well-posed. For more recent information concerning Π-scattering
conservative system nodes and the related notion of Π-scattering dissipative system
we refer to the recent work of Malinen, Staffans, and Weiss ([29], [37], [38]). In par-
nodes (where the block matrixh At Bt
ticular, the result of [37] is that a linear operator S : (cid:20)X
0i in (2.3) is assumed only to be contractive),
Y(cid:21) is a Π-scattering
passive system node if and only if it is closed with its graph equal to a maximal
J -negative subspace of X ⊕ Y ⊕ X ⊕ U. Much of this work also has a focus of
fitting physical examples into this framework; a recent accomplishment was to fit
Maxwell's equations into this framework (see [40]).
U(cid:21) →(cid:20)X
0
Dt
Ct
0
Example 2.6. Suppose that Σ = (cid:20)A&B
C&D(cid:21) : (cid:20)X
U(cid:21) ⊃ D(Σ) → (cid:20)X
U(cid:21) is a closed
operator with output space Y taken to be the same as the input space U such that
(1) its graph G(Σ) is a Lagrangian subspace of K, but where now K is given
the Kreın-space inner product induced by the signature operator J ′ given
by
,
u(cid:21) is in D(Σ).
(2) for each u ∈ U there is an xu ∈ X so that(cid:20)xu
J ′ =
0
0
0
0
0
IX
0 −IU
0
IX
0 −IU
0
0
0
0
and
(2.14)
Then Σ is a system node (see [12, Proposition 4.11]). In fact, Herglotz functions
over the right halfplane which also satisfy the growth condition at infinity
lim
t→+∞
t−1f (t)u = 0 for each u ∈ U
are characterized as those functions f having a realization f (w) = TΣ(w) as in
(2.10) with a system node Σ of this form (see [12, Theorem 4.12]). The trajectories
(u(t), x(t), y(t)) satisfy the alternative energy-conservation law
kx(t)k2
X − kx0k2
X = 2Z t
0
Rehy(t), u(t)iU dt
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
15
and Σ is called a Π-impedance-conservative system node (see [35]). As the transfer
function for a Π-impedance-conservative system node need not be bounded in the
right halfplane, it follows that Π-impedance-conservative system nodes need not be
well-posed in general (see [34]).
In connection with Example 2.6 we shall have use for the following additional
fact.
0
operator such that
system node as in Example 2.6.
(1) Y is skew-adjoint: Y = −Y∗, and
U(cid:21) is a closed densely defined
u(cid:21) ∈ D(Y).
Corollary 2.7. Suppose that Y : D(Y) ⊂(cid:20)X
U(cid:21) →(cid:20)X
(2) for each u ∈ U there is an xu ∈ X such that(cid:20)xu
Set J ′0 =(cid:20)IX
then Y := −ΣJ′0 : D(Y) ⊂(cid:20)X
D(Y) =n(cid:20)x
0 −IU(cid:21) and set Σ := −YJ ′0. Then Σ is a Π-impedance-conservative
U(cid:21) is a closed operator with the dense domain
U(cid:21) : J′0(cid:20)x
Conversely, if Σ is a Π-impedance-conservative system node as in Example 2.6,
−u(cid:21) ∈ D(Σ)o
Proof. Given any closed, densely defined operator Σ : D(Σ) ⊂ (cid:20)X
u(cid:21) =(cid:20) x
satisfying conditions (1) and (2).
U(cid:21) →(cid:20)X
u(cid:21) ∈(cid:20)X
U(cid:21) → (cid:20)X
U(cid:21), the
following translation of the definition of adjoint operator to graph spaces is well
known:
where here the orthogonal complement is with respect to the standard Hilbert space
inner product. More generally, compute the J ′-orthogonal complement (where
(cid:18)(cid:20)Σ
I(cid:21)D(Σ)(cid:19)⊥
=(cid:20) I
−Σ∗(cid:21)D(Σ∗)
0(cid:21) by the definition (2.14) of J ′) as follows:
J ′0y1(cid:21) ∈(cid:18)(cid:20)Σ
I(cid:21)D(Σ)(cid:19)⊥J ′
J ′0
J ′ =(cid:20) 0 J ′0
y2(cid:21) ∈(cid:18)(cid:20)Σ
(cid:20)y1
I(cid:21)D(Σ)(cid:19)⊥
=(cid:20) I
−Σ∗(cid:21)D(Σ∗)
I
y2(cid:21) =(cid:20)J ′0y2
⇔ J ′(cid:20)y1
y2(cid:21) ∈(cid:20)−J ′0Σ∗J ′0
⇔(cid:20)y1
(cid:20)Σ
I(cid:21)D(Σ) =(cid:20)−J ′0Σ∗J ′0
Σ = −J ′0Σ∗J ′0.
I
(cid:21)D(Σ∗J ′0).
(cid:21)D(Σ∗J ′0)
Thus condition (1) in Example 2.6 for Σ to be a Π-impedance-conservative system
node translates to
or simply
An equivalent condition is: the operator Y := ΣJ ′0 (or equivalently −Y = −ΣJ ′0)
is skew-adjoint: Y = −Y∗.
16
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
Conversely, if Y is any skew-adjoint operator with dense domain in (cid:20)X
U(cid:21), then
Σ = −YJ ′0 satisfies condition (1) in Example 2.6. Since(cid:20)x
u(cid:21) ∈ D(Σ) if and only if
(cid:20) x
−u(cid:21) ∈ D(Y), condition (2) in Example 2.6 is equivalent to condition (2) in the
statement of the corollary.
(cid:3)
The next result gives a model for Π-impedance-conservative system nodes as
described in Corollary 2.7.
Proposition 2.8. Let (T, V0, R) be a triple of operators such that:
(1) T is a densely defined skew-adjoint operator on the Hilbert space X ,
(2) V0 ∈ L(U ,X ) is a bounded linear operator from the input-output space U
(3) R is a bounded skew-adjoint operator on U.
into X , and
U(cid:21) →(cid:20)X
u(cid:21) ∈(cid:20)X
Define an operator Σ : D(Σ) ⊂(cid:20)X
D(Σ) =(cid:26)(cid:20)x
V ∗0 x + V ∗0 T (x − V0u) + Ru(cid:21) ∈(cid:20)X
U(cid:21) as follows. Set
U(cid:21) : x − V0u ∈ D(T )(cid:27)
U(cid:21) for (cid:20)x
Σ(cid:20)x
u(cid:21) =(cid:20)
T (x − V0u) + V0u
and then define
u(cid:21) ∈ D(Σ).
(2.15)
(2.16)
Equivalently, Σ can be defined as the system node constructed from the data
A = T ∈ L(X1,X ), B = (I − T )V0 ∈ L(U ,X−1),
C = V ∗0 (I − T )∗ ∈ L(X1,U), α = 1 and TΣ(1) = V ∗0 V0 + iR ∈ L(U)
(2.17)
according to the recipe (2.11). Then Σ is a Π-impedance-conservative system node.
Conversely, any Π-impedance-conservative system node arises in this way from
a triple of operators T, V0, R satisfying conditions (1), (2), and (3) above.
Remark 2.9. We note that, in case T is bounded, the operator Σ given by (2.15)
and (2.16) is simply
I − T
0
0
(2.18)
0
T
Σ =(cid:20)I
0 V ∗0(cid:21)(cid:20)
(I − T )∗ −T (cid:21)(cid:20)I
One way to make sense of the first term in this formula for the general case where T
0 V0(cid:21) +(cid:20)0
0 R(cid:21) .
is allowed to be unbounded is as follows. We may view eΣ =(cid:20)
(I − T )∗ −T (cid:21) =
X−1(cid:21), where X−1 is the rigged level-
(cid:20) T
x2(cid:21) : T (x1 − x2) ∈ X(cid:27) .
−T (cid:21) as an operator from(cid:20)X
D =(cid:26)(cid:20)x1
I + T
(−1) space associated with the skew-adjoint operator T as explained in discussion
at the beginning of this section. It is natural to introduce a domain
X(cid:21) to(cid:20)X−1
X(cid:21)(cid:27) =(cid:26)(cid:20)x1
x2(cid:21) ∈(cid:20)X
I − T
I − T
(2.19)
T
x2(cid:21) : eΣ(cid:20)x1
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
17
the operator Σ via the formula (2.18) with domain given by
and define an operator eΣ0 : D ⊂(cid:20)X
X(cid:21) →(cid:20)X
X(cid:21) by eΣ0 = eΣD. Then we may define
0 V0(cid:21)(cid:20)x
u(cid:21) ∈ D(eΣ0)(cid:27) .
D(Σ) =(cid:26)(cid:20)x
u(cid:21) : (cid:20)I
Proof of Proposition 2.8. We first show that Σ defined as in (2.15), (2.16) satisfies
conditions (1) and (2) in Corollary 2.7. As for condition (2), note that if we set
xu = V0u for each u ∈ U, then xu − V0u = 0 ∈ D(T ), so condition (2) is satisfied.
It remains to verify condition (1).
Toward this end, by the representation (2.18) for Σ as explained in Remark 2.9,
0
it suffices to show that the operator
Σ′ :=(cid:20)−I
I(cid:21)(cid:20) T
I + T −T + R(cid:21) =(cid:20) −T
with domain D as in (2.19) is skew-adjoint. Note that(cid:20)y1
T + I −T + R(cid:21)
y2(cid:21) ∈ D(cid:18)(cid:20) −T
T − I
I − T
means that the sesquilinear form
0
0
T + I −T + R(cid:21)∗(cid:19)
T − I
y2(cid:21)(cid:29)
x2(cid:21) ,(cid:20)y1
T − I
(cid:28)(cid:20) −T
T + I −T + R(cid:21)(cid:20)x1
x2 ] ∈ D(cid:0)(cid:2) −T
T−I
T − I
= hT (x1 − x2),−y1 + y2i − hx2, y1i + hx1, y2i + hRx2, y2i
T +I −T +R(cid:3)(cid:1) is continuous in the argument [ x1
defined for [ x1
x2 ] in the X ⊕X
norm. As the second, third, and fourth terms are automatically bounded in the
[ x1
x2 ]-argument and the element x1 − x2 is an arbitrary element of D(T ) (e.g., take
x2 = 0 and x1 ∈ D(T )), it follows that necessarily y2 − y1 ∈ D(T ∗) = D(T ) (i.e.,
[ y1
y2 ] ∈ D) and the calculation above continues as
y2(cid:21)(cid:29)
x2(cid:21) ,(cid:20)y1
(cid:28)(cid:20) −T
T + I −T + R(cid:21)(cid:20)x1
= hx1 − x2, T ∗(y2 − y1)i − hx2, y1i + hx1, y2i + hRx2, y2i
x2(cid:21) , −(cid:20) −T
x2(cid:21) ,(cid:20) −T y2 + T y1 + y2
=(cid:28)(cid:20)x1
T y2 − T y1 − y1 + R∗y2(cid:21)(cid:29) =(cid:28)(cid:20)x1
y2(cid:21)(cid:29)
T + I −T + R(cid:21)(cid:20)y1
where we use R is bounded with R = −R∗ in the last two steps. The skew-
adjointness of the operator Σ′ now follows as wanted. As a consequence of Corollary
2.7 it follows in particular that Σ is a system node. The equivalence of the original
definition (2.16) of Σ with the alternative formulation based on the data set (2.17)
is a simple consequence of the identities (2.11) along with computation of the value
of the transfer function at the point 1 ∈ Π: TΣ(1) = V ∗0 V0 + iR; this in turn is a
routine verification which we leave to the reader.
For the converse statement, we let Σ′ =(cid:20)A′&B′
C′&D′(cid:21) be any Π-impedance-conser-
vative system node. As (cid:20)−(A′&B′)
C′&D′ (cid:21) is skew-adjoint, from the duality theory of
system nodes (see e.g. [36]) one can see that necessarily A′ = −A′∗ and C′ = B′∗;
here B′ ∈ L(U ,X ′
−1), C′ ∈ L(X ′1,U) and computation of the adjoint B′∗ is with
respect to the duality pairing between X ′
−1 and X ′1 via the X ′ pairing (2.4) (note
that X ′1 = X ′⋆
1 since A′ = −A′∗). Set T = A′. As T is skew-adjoint, both 1 and −1
are in the resolvent set of T and we may define a bounded operator V0 ∈ L(U ,X ′)
T − I
18
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
by V0 = (I − T )−1B′. We then have B′ = (I − T )V0 and C′ = V ∗0 (I + T ). We now
have all the ingredients to define another Π-impedance-conservative system node
Σ0 =(cid:20)I
0 V ∗0(cid:21)(cid:20) T
0
I + T
zero). When we write Σ0 in the form
I − T
−T (cid:21)(cid:20)I
0
0 V0(cid:21) as in (2.15), (2.16) (with R taken equal to
C0&D0(cid:21) ,
Σ0 =(cid:20)A0&B0
we see that the construction gives that D(Σ′) = D(Σ0), i.e., D(A0&B0) = D(A′&B′),
with A0&B0 = A′&B′, so
A0 := A0&B0D(Σ0)∩[ X0 ] = A′&B′D(Σ0)∩[ X0 ] =: A′, B0 = B′ ∈ L(U ,X−1),
C0 := C0&D0D(Σ0)∩[ X0 ] = C′&D′D(Σ)∩[ X0 ] =: C′.
As observed in (2.12), for any fixed choice of α ∈ Π (e.g., α = 1) an element
[ x
u ] ∈ D(Σ′) can be decomposed as
(cid:20)x
u(cid:21) =(cid:20)x − (αI − A)−1Bu
0
(cid:21) +(cid:20)(αI − A)−1Bu
(cid:21)
u
where each summand is again in D(Σ′) = D(Σ0). Then making use of the formulas
(2.11) we compute
and similarly
0
Ax + Bu
C(x − (αI − A)−Bu)(cid:21) +(cid:20)
C(x − (αI − A)−1Bu)(cid:21) +(cid:20)
TΣ(α)u(cid:21)
TΣ0(α)u(cid:21)
(TΣ(α) − TΣ0 (α)) u(cid:21) .
Ax + Bu
0
Σ(cid:20)x
u(cid:21) =(cid:20)
Σ0(cid:20)x
u(cid:21) =(cid:20)
u(cid:21) = Σ0(cid:20)x
Σ(cid:20)x
0 −I(cid:3) and Σ0(cid:2) I 0
0
0 R ] with R := TΣ(α) − TΣ0(α) equal to a bounded operator on
u(cid:21) +(cid:20)
0 −I(cid:3) are skew-adjoint, it is necessarily the case that
(cid:3)
from which we read off
Thus Σ = Σ0 + [ 0 0
U. As both Σ(cid:2) I 0
R = −R∗ as well.
We need a generalization of the formula for the transfer function (2.10) to the
setting where the resolvent (wI − A)−1 is replaced by the structured resolvent
(Y (w) − A)−1. For the statement of this result it is convenient to assume that A is
maximal dissipative. Recall from the discussion in Example 2.2 above that a densely
defined closed Hilbert-space operator A is maximal dissipative if it is dissipative
(RehAx, xi ≤ 0 for all x ∈ D(A)) and I + A is onto; it then follows that wI + A
is onto for all w ∈ Π and that A is the generator of a contractive semigroup (see
[31]).
Proposition 2.10. Let A be a maximal dissipative operator on the Hilbert space
X (and hence A is the generator of a contractive C0-semigroup with resolvent set
containing the right halfplane Π), and suppose that (Y1, . . . , Yd) is a d-fold positive
decomposition of IX . For w = (w1, . . . , wd) ∈ Cd, we set Y (w) = w1Y1 +···+wdYd.
Then the following observations hold true:
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
19
(1) For w ∈ Πd, Y (w) − A is invertible with
k(Y (w) − A)−1k ≤
minj Re wj
1
.
(2.20)
(2) For w ∈ Πd, the operator (Y (w) − A)−1 is a bicontinuous bijection from X
(3) For w ∈ Πd, the operator Y (w) − A : X1 → X has an extension
onto X1.
(Y (w) − A)X : X → X−1
which is a bicontinuous bijection from X onto X−1.
(4) For any system node Σ =(cid:20)A&B
C&D(cid:21) containing A as its state operator/semi-
group generator, it holds that
(cid:20)((Y (w) − A)X )−1Bu
u
(cid:21) ∈ D(Σ).
Proof. (1) For w = (w1, . . . , wd) ∈ Πd and x ∈ D(A), we have
k(Y (w) − A)xk · kxk ≥ h(Y (w) − A)x, xi
≥ Reh(Y (w) − A)x, xi
≥ RehY (w)x, xi
=
dXk=1
h(Re wk)Ykx, xi
dXk=1
≥ minj(Re wj)
= minj(Re wj)kxk2.
hYkx, xi
(2.21)
Thus Y (w) − A is bounded below and has a left inverse. A similar argument with
Y (w) − A replaced by (Y (w) − A)∗ shows that Y (w) − A also has a right inverse.
The computation (2.21) gives the estimate (2.20).
(2) Next note that for w ∈ Πd and x ∈ X ,
A(Y (w) − A)−1x = −x + Y (w)(Y (w) − A)−1x ∈ X
from which we conclude that (Y (w) − A)−1 maps X into X1. Conversely if x ∈ X1,
then y = (Y (w) − A)x ∈ X and we recover x as x = (Y (w) − A)−1y. We conclude
that (Y (w) − A)−1 maps X bijectively to X1. The fact that (Y (w) − A)−1 is
bicontinuous is then a consequence of the open mapping theorem.
(3) We use the dual version of a result from part (2): for w ∈ Πd,
Y (w)∗ − A∗ : X ⋆
1 → X
is a bicontinuous bijection from X ⋆
adjoints with respect to the X -pairing, we get an operator
1 (with the (1, ⋆)-norm) onto X . If we then take
(Y (w)∗ − A∗)∗ : X → X−1
which must also be a bicontinuous bijection, but now from X to X−1. Since the
duality is with respect to the X -pairing (2.4), it is clear that this map provides an
extension of the operator Y (w) − A : X1 → X :
(Y (w) − A)X := (Y (w)∗ − A∗)∗ : X → X−1.
20
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
(4) To show that(cid:20)((Y (w) − A)X )−1Bu
D(Σ) in Definition 2.4 we need only show that
u
(cid:21) is in D(Σ), by the characterization of
(cid:2)AX B(cid:3)(cid:20)((Y (w) − A)X )−1Bu
u
(cid:21) ∈ X .
(2.22)
As Bu ∈ X−1, (Y (w) − A)−1 maps X−1 to X by part (3) and Y (w) is a bounded
operator on X , we conclude that
Y (w)((Y (w) − A)X )−1Bu ∈ X .
We now rewrite the expression in (2.22) as
(cid:2)AX B(cid:3)(cid:20)((Y (w) − A)X )−1Bu
u
(cid:21)
= [(A − Y (w))X + Y (w)]((Y (w) − A)X )−1Bu + Bu
= Y (w)((Y (w) − A)X )−1Bu
to conclude that (2.22) holds as desired.
(cid:3)
Remark 2.11. We note that as a consequence of property (4) in Proposition 2.10
it is possible to define the transfer function associated with the structured resolvent
(Y (w) − A)−1 via
TΣ,{Yk}(w) = [C&D] (cid:20)(cid:0)(Y (w) − A)X(cid:1)−1
u
Bu
(cid:21) .
It is tempting to view this as the transfer function for a multidimensional linear
system
Pd
k=1 Yk
∂x
∂tk
(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t).
(2.23)
where t = (t1, . . . , td) (a continuous-time version of a multidimensional linear sys-
tem of Fornasini -- Marchesini type -- see [25]). However, in general it is not clear how
to extend the operators Y1, . . . , Yd ∈ L(X ) to operators YkX−1 in a sensible way.
Special cases where this is possible are: (1) the case where B and C are bounded,
i.e., B ∈ L(U ,X ) and C ∈ L(X ,Y): in this situation the system (2.23) makes sense
with state space taken simply to be X , and (2) in case the operators Yk commute
with A: in this case one can extend Yk to X−1 via the formula
YkX−1 = (αI − A)Yk(αI − A)−1 : X−1 → X−1.
We note that continuous-time counterparts of Fornasini -- Marchesini models of a
somewhat different form have been considered in the literature (see [23] and the
references there).
3. The Herglotz -- Agler class over the polydisk
In this section we present our results for the Herglotz -- Agler class over the poly-
disk Dd.
Theorem 3.1. Given a function F : Dd → L(U), the following are equivalent.
(1) F is in the Herglotz -- Agler class HA(Dd,L(U)).
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
21
(2) F has a Dd-Herglotz -- Agler decomposition, i.e., there exist L(U)-valued pos-
itive kernels K1, . . . , Kd on Dd such that
F (ω)∗ + F (ζ) =
(1 − ωkζk)Kk(ω, ζ).
(3.1)
dXk=1
(3) There exist a Hilbert space X , a d-fold spectral decomposition (P1, . . . , Pd)
of IX with associated operator pencil P (ζ) = ζ1P1 + ··· + ζdPd (see Section
2.1), and a bounded colligation matrix
U =(cid:20)A B
C D(cid:21) : (cid:20)X
U(cid:21) →(cid:20)X
U(cid:21)
with block matrix entries A, B, C, D satisfying the relations
A∗A = AA∗ = IX , B = AC∗, D + D∗ = CC∗ (= B∗B)
(3.2)
such that
F (ζ) = D + C(I − P (ζ)A)−1P (ζ)B.
(3.3)
Proof. (1)⇒(2): Use that F ∈ HA(Dd,L(U)) if and only if S(ζ) := [F (ζ)−I][F (ζ)+
I]−1 is in the Schur -- Agler class SA(Dd,L(U)). Then S has a Schur -- Agler decom-
position
I − S(ω)∗S(ζ) =
dXk=1
(1 − ωkζk)eKk(ω, ζ)
for L(U)-valued positive kernels eK1, . . . , eKd on Dd. A routine computation gives
I − S(ω)∗S(ζ) = 2[F (ω)∗ + I]−1 (F (ω)∗ + F (ζ)) [F (ζ) + I]−1.
This leads us to the Dd-Herglotz -- Agler decomposition (3.1) with
Kk(ω, ζ) =
1
2
[F (ω)∗ + I] eKk(ω, ζ) [F (ζ) + I].
Notice that Agler in [1] proved the implications (1)⇒(2) in Theorems 1.2 and 3.1
simultaneously, while we show here that they are, in fact, equivalent.
ζ1I fX1
decomposition (3.1) in the form
(2)⇒(3): Since each kernel Kk is positive, each Kk has a Kolmogorov decom-
H(ζ) = Colk=1,...,d[Hk(ζ)] be the associated block column matrix function defining
position, i.e., there is a function Hk : Dd → L(U , eXk) (for some auxiliary Hilbert
space eXk) so that we have the factorization Kk(ω, ζ) = Hk(ω)∗Hk(ζ). We let
a function on Dd with values in L(U , eX ), where we set eX = eX1 ⊕ ··· ⊕ eXd (written
. We now may rewrite the Agler
as columns). We set P (ζ) =
eG = span
u : u ∈ U , ζ ∈ Dd
F (ω)∗ + F (ζ) = H(ω)∗(I eX − P (ω)∗P (ζ))H(ζ).
eX
eX
We consider the subspace
H(ζ)
F (ζ)
⊂
U
U
...
ζdI fXd
P (ζ)H(ζ)
IU
(3.4)
22
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
where the ambient space eX ⊕ U ⊕ eX ⊕ U is given the Kreın-space inner product
induced by the signature matrix eJ given by
−I eX
0
0
0
0
0
0
IU
0
0
I eX
0
0
IU
0
0
.
eJ =
Then one can check
IU
IU
0
IU
0
0
P (ζ)H(ζ)
P (ω)H(ω)
0
0
0
IU
H(ζ)
F (ζ)
0
0
I eX
0
H(ω)
F (ω)
We claim that
−I eX
0
0
0
u′+
u,
where the last equality follows as a consequence of the Agler decomposition (3.4).
= h[−H(ω)∗H(ζ) + F (ω)∗ + H(ω)∗P (ω)∗P (ζ)H(ζ) + F (ζ)] u, u′i = 0
*
We conclude that eG is a eJ -isotropic subspace. We then extend eG to a J -Lagrangian
subspace G, where the ambient space eX ⊕ U ⊕ eX ⊕ U extended to a space of the
form X ⊕ U ⊕ X ⊕ U with X ⊃ eX and the Kreın-space Gramian matrix J of the
same block form as eJ above.
Indeed, suppose that(cid:20) x
and ζ ∈ Dd,
0 =*J(cid:20) x
x + u, u′iU .
Hence u = H(ζ)∗P eX
x for all ζ ∈ Dd; in particular, u = H(0)∗P eX
element of G has the form(cid:20)
0 =(cid:28)J(cid:20)
0(cid:21) ∈ G. As G is isotropic, we then must have, for all u′ ∈ U
0(cid:21) ," H(ζ)
# u′+ = h−H(ζ)∗P eX
(cid:21). We must also have
(cid:21)(cid:29) = −kxk2
(cid:21) ,(cid:20)
{0}# = {0}.
G ∩" X
U
{0}
x. Thus our
F (ζ)
P (ζ)H(ζ)
H(0)∗P fX x
H(0)∗P fX x
IU
x
0
0
x
0
0
H(0)∗P fX x
u
0
u
0
x
0
0
which enables us to conclude that x = 0 and hence also(cid:20)
claim follows.
X
x
0
0
H(0)∗P fX x
(cid:21) = 0, and the
We are now able to conclude that G is a graph space:
A&B
C&D
0
IX
0
IU
G =
for some closed linear operator (cid:2) A&B
construction the vector(cid:2) P (ζ)H(ζ)u
D = (cid:2) D1
Lemma 3.4 from [12] to conclude that in fact (cid:2) A&B
(cid:20)x
u(cid:21) : (cid:20)x
C&D(cid:3) with domain D ⊂ [ X
(cid:3) is in D for all ζ ∈ Dd and u ∈ U; in particular,
U (cid:3) for some linear manifold D1 ⊂ X . We are now in position to apply
C&D(cid:3) = [ A B
u ] ∈ D for all u ∈ U and hence D splits:
C D ] is bounded with
by setting ζ = 0 ∈ Dd we see that [ 0
u(cid:21) ∈ D
]. Furthermore, by
U
u
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
23
domain equal to all of [ X
U
produced a colligation matrix U = [ A B
], and moreover the identities (3.2) hold. We have now
In particular, it follows that, for any u ∈ U, there is a corresponding(cid:2) x′
u′(cid:3) in [ X
that
U
] so
C D ] so that
A B
C D
0
IX
0
IU
G =
u =
(cid:20)A B
C D(cid:21)(cid:20)P (ζ)H(ζ)u
(cid:21) =(cid:20)H(ζ)u
(cid:20)X
U(cid:21) .
(cid:20)x′
u′(cid:21) .
F (ζ)u(cid:21) for all u ∈ U
A B
C D
0
IX
0
IU
H(ζ)
F (ζ)
P (ζ)H(ζ)
IU
u
From the bottom two components of (3.5) we read off
u′ = u.
x′ = P (ζ)H(ζ)u,
Then the top two components of (3.5) give
(3.5)
(3.6)
which we rewrite as a linear system of operator equations
Solving the first equation in (3.6) for H(ζ) gives
CP (ζ)H(ζ) + D = F (ζ).
(cid:26) AP (ζ)H(ζ) + B = H(ζ)
H(ζ) = (IX − AP (ζ))−1B.
(We note that the inverse on the right hand side exists since A is unitary and
kP (ω)k < 1 for ω ∈ Dd.) Plugging this last expression into the second of equations
(3.6) then yields
F (ζ) = D + CP (ζ)(I − AP (ζ))−1B
= D + C(I − P (ζ)A)−1P (ζ)B
and condition (3) in the statement of Theorem 3.1 follows.
(3)⇒(2): We assume that F has the representation (3.3) where the coefficient
matrices A, B, C, D satisfy the relations (3.2). Then we compute
F (ζ)∗ + F (ω) = D∗ + B∗(I − P (ζ)∗A∗)−1P (ζ)∗C∗ + D + CP (ω)(I − AP (ω))−1B
= [D + D∗] + B∗(I − P (ζ)∗A∗)−1P (ζ)∗A∗B + B∗AP (ω)(I − AP (ω))−1B
= B∗(I − P (ζ)∗A∗)−1X(I − AP (ω))−1B
where we have set X equal to
X = (I − P (ζ)∗A∗)(I − AP (ω)) + P (ζ)∗A∗(I − AP (ω)) + (I − P (ζ)∗A∗)AP (ω)
= I − P (ζ)∗A∗ − AP (ω) + P (ζ)∗P (ω)
+ P (ζ)∗A∗ − P (ζ)∗P (ω) + AP (ω) − P (ζ)∗P (ω)
= I − P (ζ)∗P (ω).
We conclude that the Agler decomposition (3.1) holds with
Kk(ζ, ω) = B∗(I − P (ζ)∗A∗)−1Pk(I − AP (ω))−1B,
24
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
where Pk = P eXk
Theorem 3.1 follows.
is the orthogonal projection of eX onto eXk, and condition (2) in
(2)⇒(1): Given an Agler decomposition (3.1), we may rewrite it in the form
(3.4). From the proof of (2)⇒(3)⇒(2) we see that we may assume that H(ζ) is
holomorphic in ζ ∈ Dd.
If T = (T1, . . . , Td) is a commutative d-tuple of strict
contractions on K, it is straightforward to verify that the formula (3.4) leads to
F (T )∗ + F (T ) = H(T )∗ (IX⊗K − P (T )∗P (T )) H(T ).
k=1 ζkPk, we see that P (T ) =Pd
k=1 Pk⊗Tk has
kP (T )k < 1. Hence F (T )∗ + F (T ) ≥ 0 and we conclude that F ∈ HA(Dd,L(U)).
From the diagonal form of P (ζ) =Pd
(cid:3)
Remark 3.2. Given a representation for F ∈ HA(Dd,L(U)) as in (3.3) and (3.2),
let us separate out the selfadjoint and skew-adjoint parts of F (0) = D to rewrite
the formula (3.3) as
where Re F (0) = 1
(3.2) we see that
F (ζ) = Re F (0) + C(I − P (ζ)A)−1P (ζ)B + R
2 (D + D∗) and we set R = 1
2 (D− D∗) = −R∗. From the relations
1
2
1
2
B∗(cid:0)I + 2A(I − P (ζ)A)−1P (ζ)(cid:1) B
F (ζ) − R =
B∗(cid:0)I + 2AP (ζ)(I − AP (ζ))−1(cid:1) B
=
= V ∗(I − AP (ζ))−1(I + AP (ζ))V,
(3.7)
where V := 1√2
B is such that V ∗V = Re F (0).
We note that Agler [1] obtained the representation (3.7) for a function in the
Herglotz -- Agler class HA(Dd,L(U)) starting with the realization S(ζ) = D + C(I −
P (ζ)A)−1P (ζ)B (with U = [ A B
C D ] unitary as in part (3) of Theorem 1.2) for the
associated function
S(ζ) = (F (ζ) − I)(F (ζ) + I)−1
(3.8)
in the Schur -- Agler class SA(Dd,L(U)) as follows. The fact that S arises as the
Cayley transform of a function in the Herglotz -- Agler class implies that I − S(ζ) =
2(F (ζ) + I)−1 is injective: in particular, D = S(0) has the property that I − D is
invertible; moreover, the fact that U is unitary implies that the U0 := A + B(I −
D)−1C is also unitary. This follows from the two identities,
I
(cid:20)A B
C D(cid:21)(cid:20)
(cid:20)A∗ C∗
B∗ D∗(cid:21)(cid:20)
(I − D)−1C(cid:21) =(cid:20)
(I − D∗)−1B∗(cid:21) =(cid:20)
I
U0
(I − D)−1C(cid:21) ,
(I − D∗)−1B∗(cid:21) ,
U∗0
which, when combined with that fact that [ A B
C D ] is unitary, imply that both U0
and U∗0 are isometries. In terms of our notation, the result of Agler [1, Proof of
theorem 1.8] is that then F has the representation (3.7) with
A = U0,
V =
1
√2
B.
(3.9)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
25
The formula (3.7) can be further adjusted as follows:
F (ζ) − R = V ∗(I − AP (ζ))−1(I + AP (ζ))V
= V ∗(U − P (ζ))−1(U + P (ζ))V
(3.10)
where we set U = A∗. Here we still have V ∗V = Re F (0) and U = A∗ is unitary.
We shall use this representation for a Dd-Herglotz -- Agler function in Section 6.
For an alternate direct derivation of the realization (3.10) for a Dd-Herglotz -- Agler
function, see [11], where the result is given for the rational matrix-valued case with
the additional constraint that F have zero real part on the unit circle; in this case
one can arrange that the state space X is finite-dimensional.
4. The Schur -- Agler class over the right polyhalfplane
In this section we present our realization results for the Schur -- Agler class over
the right polyhalfplane Πd.
Theorem 4.1. Given a function s : Πd → L(U ,Y), the following are equivalent.
(1) s is in the right polyhalfplane Schur -- Agler class SA(Πd,L(U ,Y)).
(2) s has a Πd-Schur -- Agler decomposition, i.e., there exist positive kernels
K1, . . . , Kd on Πd such that
I − s(z)∗s(w) =
dXk=1
(zk + wk)Kk(z, w).
(4.1)
for all z, w ∈ Πd.
(3) There exists a state space X and a d-fold positive decomposition of IX
(Y1, . . . , Yd) (see Section 2.1) together with a Π-scattering-conservative sys-
tem node (see Example 2.5)
so that
C&D(cid:21) : D ⊂(cid:20)X
U =(cid:20)A&B
U(cid:21) →(cid:20)X
Y(cid:21)
s(w) = C&D(cid:20)(cid:0)(Y (w) − A)X(cid:1)−1
IU
B
(cid:21) ,
(4.2)
where we have set Y (w) = w1Y1 + ··· + wdYd.
Proof. (1)⇒(2): The proof uses the component-wise multivariable Cayley trans-
form from Πd to Dd; for this purpose it is convenient to use the condensed notation
(1.4) for the point 1+ζ
point ζ = (ζ1, . . . , ζd) ∈ Πd.
function S ∈ SA(Dd,L(U ,Y)) in the Schur -- Agler class over Dd via
1−ζd(cid:17) in the polydisk Dd associated with the
For s ∈ SA(Πd,L(U ,Y)) in the Schur -- Agler class over Πd we associate the
, . . . , 1+ζd
1−ζ =(cid:16) 1+ζ1
1−ζ1
S(ζ) = s(cid:18) 1 + ζ
1 − ζ(cid:19)
Then by Theorem 1.2 S has a Dd-Schur -- Agler decomposition
IU − S(ω)∗S(ζ) =
dXk=1
(1 − ωkζk)eKk(ω, ζ).
26
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
Using the relation
(where we use the convention (1.5)), we next get that
w − 1(cid:19)
s(w) = S(cid:18) w + 1
dXk=1(cid:18)1 −
zk − 1
zk + 1 ·
IU − s(z)∗s(w) =
wk − 1
wk + 1(cid:19) bKk(z, w)
w + 1(cid:19) .
w − 1
where we set
This in turn leads to
,
z + 1
bKk(z, w) = eKk(cid:18) z − 1
dXk=1
IU − s(z)∗s(w) =
(zk + wk)Kk(z, w)
with Kk given by
and (2) follows.
Kk(z, w) = 2
1
zk + 1bKk(z, w)
1
wk + 1
Schur -- Agler decomposition (4.1) in the form
(2)⇒(3): We use the Kolmogorov decompositions Kk(z, w) = Hk(z)∗Hk(w)
I − s(z)∗s(w) = H(z)∗(P (z)∗ + P (w))H(w)
(where Hk : Πd → L(U , eXk)) of the positive kernels Kk(z, w) to rewrite the Πd-
. We intro-
where we set H(w) = Col1≤k≤d Hd(w) with P (w) =
eX0 := span(cid:8)Ran H(w) : w ∈ Πd(cid:9) ⊂
dMk=1 eXk = eX
duce the subspace
wdI fXd
w1I fX1
(4.4)
(4.3)
...
and we introduce operators eY1, . . . ,eYd on eX0 by
Pk eX0
eYk = P eX0
where P eX0
Pk = P eXk
is the orthogonal projection of eX onto its subspace eX0 (4.4) and where
is the orthogonal projection of eX onto eXk.
It is easily verified that
We next view (4.3) as the statement that the subspace
eY1, . . . ,eYd form a positive decomposition of the identity on eX0.
⊂
eX0
eX0
u : w ∈ Πd, u ∈ U
eG := span
eY (w)H(w)
s(w)
H(w)
IU
Y
U
=: eK
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
27
0
IY
0
0
(4.5)
0
0
I eX0
0
0
I eX0
0 IU
product induced by the indefinite Gramian matrix
To this end we first check the necessary condition that
.
We next check that eG can be expressed as a graph space
0
I eX0
0
0
0
0
0 −IU
is an isotropic subspace of eK, where eK is considered as a Kreın space with inner
eJ :=
eD0
eG =
eA&eB
eC&eD
eC& eDi with dense domain eD0 ⊂ eX0 ⊕ U
associated with a closed operator eU = h eA& eB
mapping into eX0 ⊕ Y.
= {0}
eG ∩
eX
0(cid:21)(cid:29) = ky′k2
0 =(cid:28)eJ(cid:20) x′
0(cid:21) ,(cid:20) x′
0 =*eJ" eY (w)H(w)
as follows. We suppose that x′ ⊕ y′ ⊕ 0 ⊕ 0 ∈ eG. As eG is isotropic, it follows that
from which we see that y′ = 0 and x′ ⊕ 0 ⊕ 0 ⊕ 0 ∈ eG. As eG is isotropic, we must
0(cid:21)+ = hH(w)u, x′i eX0
for all u ∈ U and w ∈ Πd. We now use the condition (4.4) to conclude that
necessarily x′ = 0 as well.
# u,(cid:20) x′
0
0
then also have
Y
{0}
{0}
y′
0
y′
0
s(w)
H(w)
IU
We next observe that (4.3) for z = w = te, where e = (1, . . . , 1), becomes
I − s(te)∗s(te) = 2tH(te)∗H(te),
so for any u ∈ U we have
kH(te)uk2 =
1
2t
(kuk2 − ks(te)uk2) → 0 as t → ∞.
Therefore
spanu,wn(cid:20)H(w)
and we now see thatnh H(w)
IU (cid:21) uo ⊃ spanu,w,tn(cid:20)H(w) − H(te)
(cid:21) uo ⊃ spanu,wn(cid:20)H(w)
0 (cid:21) uo
{0}(cid:21) ,
=(cid:20)fX0
IU i uo has dense span in eX0 ⊕ U. We conclude that eG is
indeed a graph space as claimed.
0
28
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
K := X ⊕ Y ⊕ X ⊕ U where we may arrange that X is a Hilbert space containing
By Proposition 2.5 in [12], we may embed eG into a J -Lagrangian subspace G of
eX and where we set
(4.6)
Furthermore, it is argued there that one can arrange that this (possibly) enlarged
Lagrangian subspace G is also a graph space:
0
IY
0
0
0
0
IX
0
.
X
Y
{0}
{0}
0
IX
0
0
0
0
0 −IU
J =
G ∩
= {0}.
U =(cid:20)A&B
C&D(cid:21) : D(U) ⊂(cid:20)X
U(cid:21) →(cid:20)X
Y(cid:21)
D(U).
I(cid:21)D(U) =
G =(cid:20)U
A&B
C&D
0
IX
0
IU
Hence there is a closed operator
so that
As discussed in Example 2.5, U is a Π-scattering-conservative system node. It is
shown in the proof of Proposition 4.9 from [12] that then the main operator A of U
(given by (2.5) with U in place of Σ) is maximal dissipative (as defined in Example
2.2 above). We let Y1, . . . , Yd be a positive decomposition of the identity on X
which extends eY1, . . . ,eYd; e.g., one way to do this is
YkX⊖ eX0
X⊖ eX0
= I
,
= Yk for k = 1, . . . , d
Y1X⊖ eX0
Yk eX0
= 0 for k = 2, . . . , d,
and extend by linearity. We then set Y (w) = w1Y1 + ··· + wdYd.
As G contains eG, we conclude that
Y (w)H(w)u
for each w ∈ Πd and u ∈ U. Thus, for each such w and u there is ah xw,u
w,ui ∈ D(U)
so that
u′
(4.7)
(4.8)
s(w)u
H(w)u
u
Y (w)H(w)u
s(w)u
H(w)u
u
∈
=
A&B
C&D
0
IU
0
IU
A&B
C&D
0
IX
0
IU
D(U)
(cid:20)xw,u
u′w,u(cid:21) .
From the bottom two rows of (4.8) we read off
H(w)u = xw,u,
u = u′w,u.
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
29
Plugging these back into the first two rows of (4.8) gives
AX H(w)u + Bu = Y (w)H(w)u
C&D(cid:20)H(w)u
u
(cid:21) = s(w)u.
(4.9)
Since U is a Π-scattering-conservative system node, a consequence of Proposition
2.10 is that Y (w) − A is invertible for each w ∈ Πd with an extension (Y (w) −
A)X : X → X−1 having the property that ((Y (w) − A)X )−1 : X−1 → X . We may
therefore solve the first of equations (4.9) for H(w)u to get
and furthermore
H(w)u =(cid:0)(Y (w) − A)X(cid:1)−1
Bu.
(4.10)
(cid:20)(cid:0)(Y (w) − A)X(cid:1)−1
u
Bu
(cid:21) ∈ D(C&D) for each w ∈ Πd.
We can now substitute (4.10) into the second of equations (4.9) to arrive at the
desired realization formula (4.2) for s, and (3) follows.
(3)⇒(2): Assume that s has a realization as in (4.2). For w ∈ Πd, set
I
B
B.
H(w) =(cid:0)(Y (w) − A)X(cid:1)−1
(cid:21) u
AX H(w)u + Bu = A&B(cid:20)(cid:0)(Y (w) − A)X(cid:1)−1
= AX(cid:0)(Y (w) − A)X(cid:1)−1
= −Bu + P (w)(cid:0)(Y (w) − A)X(cid:1)−1
= Y (w)(cid:0)(Y (w) − A)X(cid:1)−1
(cid:21) =(cid:20)A&B
C&D(cid:21)(cid:20)H(w)u
U(cid:20)H(w)u
(cid:21) =(cid:20)Y (w)H(w)u
= Y (w)H(w)u.
Bu + Bu
s(w)u
Bu
u
u
Bu + Bu
(cid:21) .
(4.11)
Observe that
Combining this with (4.3) gives
By Proposition 4.9 from [12], the fact that U is a Π-scattering-conservative system
node tells us that the graph of U is a Lagrangian subspace of X ⊕ Y ⊕ X ⊕ U in
the Kreın-space inner product induced by J as in (4.6). In particular it holds that
0 =*J
=*J
A&B
C&D
0
IX
0
IU
Y (w)H(w)u
s(w)u
H(w)u
u
(cid:21)+
u
(cid:20)H(w)u
,
(cid:21) ,
A&B
C&D
0
IX
0
IU
Y (z)H(z)u′
s(z)u′
H(z)u′
u′
u′
(cid:20)H(z)u′
+
= h[H(z)∗(Y (z)∗ + Y (w))H(w) + (s(z)∗s(w) − IU )] u, u′i.
(4.12)
30
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
By the arbitrariness of u, u′ ∈ U, we conclude that
H(z)∗ dXk=1
(zk + wk)Yk! H(w) = IU − s(z)∗s(w).
(4.13)
Then (4.13) leads to the Agler decomposition (4.1) with
Kk(z, w) = H(z)∗YkH(w) = [Y 1/2
k H(z)]∗[Y 1/2
k H(w)].
(2)⇒(1): We write the Agler decomposition (4.1) in the form (4.3). Then, if
A = (A1, . . . , Ad) is a commutative d-tuple of strictly accretive operators, the
functional calculus gives
I − s(A)∗s(A) = H(A)∗(P (A)∗ + P (A))H(A).
From the diagonal form P (w) = w1P1 + ··· + wdPd of P (w), we see that P (A)∗ +
P (A) ≥ 0, and hence I − s(A)∗s(A) ≥ 0, or ks(A)k ≤ 1. We conclude that
s ∈ SA(Πd,L(U ,Y)) and (1) follows.
(cid:3)
5. The Herglotz -- Agler class over the polyhalfplane
In this section we present our realization results for a restricted class of Herglotz --
Agler functions over the right polyhalfplane Πd, where a growth condition (5.1) is
imposed at infinity.
Theorem 5.1. Given a function f : Πd → L(U), the following are equivalent:
(1) f ∈ HA(Πd,L(U)) and also f satisfies the growth condition at +∞:
lim
t→+∞
t−1f (te)u = 0 for each u ∈ U .
(5.1)
where e = (1, . . . , 1) ∈ Πd.
itive kernels K1, . . . , Kd on Πd such that
(2) f has a Πd-Herglotz -- Agler decomposition, i.e., there exist L(U)-valued pos-
f (z)∗ + f (w) =
dXk=1
(zk + wk)Kk(z, w)
(5.2)
and in addition f satisfies the growth condition (5.1).
(3) There exists a Hilbert state space X and a positive decomposition of the
identity (Y1, . . . , Yd) on X along with an Π-impedance-conservative system
node (see Example 2.6)
C&D(cid:21) : D(Y) ⊂(cid:20)X
Y =(cid:20)A&B
U(cid:21) →(cid:20)X
U(cid:21)
f (w) = C&D(cid:20)(cid:0)(Y (w) − A)X(cid:1)−1
(cid:21) .
IU
B
such that
Proof. We show
(5.3)
(1)⇒(2)⇒(3)⇒(2)⇒(1).
(1)⇒(2): We note that f is in the Herglotz -- Agler class HA(Πd,L(U)) if and
only if s = (f − I)(f + I)−1 is in the Schur -- Agler class SA(Πd,L(U)). Thus by
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
31
Theorem 4.1 this s has a Πd-Schur -- Agler decomposition
I − s(z)∗s(w) =
dXk=1
(zk + wk)eKk(z, w)
for some L(U)-valued positive kernels eKk on Πd. A standard computation gives
I − s(z)∗s(w) = 2(f (z)∗ + I)−1[f (z)∗ + f (w)](f (w) + I)−1
from which we see that (5.2) holds with
Kk(z, w) =
1
2
(f (z)∗ + I)eKk(z, w)(f (w) + I).
(2)⇒(3): We use the Kolmogorov decompositions Kk(z, w) = Hk(z)∗Hk(w)
of the positive kernels Kk (with Hk : Πd → L(U , eXk) say) to rewrite the Agler
decomposition (5.2) in the form
f (z)∗ + f (w) = H(z)∗(P (z)∗ + P (w))H(w)
(5.4)
where we have set
.
wdI eXd
Just as in the proof of Theorem 4.1, we introduce the subspace
. . .
...
H1(w)
Hd(w)
w1I eX1
H(w) =
, P (w) =
eX0 := span{Ran H(w) : w ∈ Πd} ⊂
and introduce operators eY1, . . . ,eYd on eX0 by
eYk = P eX0
We then view (5.2) as the statement that the subspace
Pk eX0
.
dMk=1 eXk =: eX
(5.5)
(5.6)
(5.7)
is an isotropic subspace of eK when eK is given the Kreın-space inner product induced
by the Gramian matrix
⊂
eX0
eX0
U
U
=: eK
eG = span
eY (w)H(w)
f (w)
H(w)
IU
u : u ∈ U , w ∈ Πd
.
0
I eX0
0 −IU
0
0
0
0
0
0
0
0
0
I eX0
0 −IU
eJ =
We show next that eG is a graph space, i.e.,
= {0}.
eG ∩
U
{0}
{0}
eX0
32
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
Toward this end, suppose that x ⊕ u ⊕ 0 ⊕ 0 ∈ eG. Our goal is to show that then
necessarily x = 0 and u = 0. As eG is eJ -isotropic, we necessarily have, for all u′ ∈ U,
+
f (w)u′
H(w)u′
,
0 =*eJ
eY (w)H(w)u′
=(cid:28)(cid:20)H(w)u′
−u′ (cid:21) ,(cid:20)x
u(cid:21)(cid:29)
= hu′, H(w)∗x − ui
u′
x
u
0
0
(5.8)
(5.9)
from which we conclude that
We note that, as a consequence of (5.4),
u = H(w)∗x for all w ∈ Πd.
H(te)∗H(te) =
f (te)∗ + f (te)
2t
.
The growth assumption (5.1) then implies that H(te) → 0 strongly as t → +∞.
To conclude the proof of (5.7), we now need only specialize (5.9) to the case
w = te and take a weak limit as t → +∞ to show that u = 0. It then follows from
(5.8) that x is orthogonal to span{H(w)u : w ∈ Πd, u ∈ U}. As a consequence of
(5.5), this in turn forces x = 0 and (5.7) follows.
We next embed eG into a J -Lagrangian subspace G of X ⊕ U ⊕ X ⊕ U, where X
is a Hilbert space containing eX as a subspace and the indefinite Gramian matrix
has the same form as eJ in (5.6) above:
J =
0
0
0
0
0
IX
0 −IU
0
IX
0 −IU
0
0
0
0
in such a way that G is still a graph subspace. That this is possible follows via a
minor adjustment of Proposition 2.5 in [12] as indicated in the proof of Theorem
4.12 there. We also note that condition (2) in Example 2.6 is automatic since the
(5.10)
subspace eG satisfies this condition by construction. It then follows that G has the
form (4.7) for a closed operator
U =(cid:20)A&B
C&D(cid:21) : D(U) ⊂(cid:20)X
U(cid:21) →(cid:20)X
Y(cid:21) .
That U is a Π-impedance-conservative system node follows from the fact that G is
J -Lagrangian (with J given by (5.10)). We also extend the positive decomposition
on X just as in the proof of (2)⇒(3) in Theorem 4.1 above. That we recover f (w)
as the transfer-function for the system node U, i.e., the formula (5.3) holds, now
follows exactly as in the proof of Theorem 4.1.
of the identityeY1, . . . ,eYd on eX0 to a positive decomposition of the identity Y1, . . . , Yd
(3)⇒(2): We follow the proof of (3)⇒(2) in Theorem 4.1. If we define H(w) =
i (well-defined by Proposition 2.10), we arrive at (4.11) and
A&Bh (Y (w)−AX )−1
B
I
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
33
(4.12), but with J given by (5.10) rather than by (4.5), leading to the adjusted
final conclusion
0 = h[H(z)∗(Y (z)∗ + Y (w))H(w) − (f (z)∗ + f (w))] u, u′i.
This leads to the Agler decomposition (5.2) with
Kk(z, w) = H(z)∗YkH(w) = [Y 1/2
k H(z)]∗[Y 1/2
k H(w)]
as in the proof of (3)⇒(2) in Theorem 4.1.
It remains to show that the growth condition (5.1) necessarily holds if f has a
realization (5.3) from a Π-impedance-conservative system node. To see this, we note
that then the single-variable Herglotz function f (se) (s ∈ Π) has an impedance-
conservative system-node realization. That the growth condition (5.1) holds now
follows from the result for the single-variable case (see Theorem 7.4 in [35]).
(2)⇒(1): The proof is parallel to the proofs of (2)⇒(1) in Theorems 3.1 and
4.1. Write the Agler decomposition (5.2) in the form (5.4) and observe that the
functional calculus gives
f (A)∗ + f (A) = H(A)∗(P (A)∗ + P (A))H(A).
If A is a strictly accretive commutative d-tuple, from the diagonal form of P (w)
we see that P (A)∗ + P (A) ≥ 0. We conclude that f ∈ HA(Πd,L(U)), and (1)
follows.
(cid:3)
In [3] a criterion was given for when a Πd-Herglotz -- Agler function has a real-
ization involving the structured resolvent (P (w) − A)−1 coming from a spectral
decomposition (P1, . . . , Pd) (so P (w) = w1P1 + ··· + wdPd) rather than just a pos-
itive decomposition (Y1, . . . , Yd) of the identity). We give our version of a result
of this type, with realization in terms of a Π-impedance-conservative system node
rather than in the form presented in [3].
Theorem 5.2. Suppose that f : Πd → L(U) is a Πd-Herglotz -- Agler function satis-
fying the growth condition (5.1). Then the following are equivalent:
(1) There exists a Hilbert space X and a d-fold spectral decomposition (P1, . . . ,
Pd) of IX along with a Π-impedance-conservative system node
such that
C&D(cid:21) : D(U) ⊂(cid:20)X
Y =(cid:20)A&B
U(cid:21) →(cid:20)X
U(cid:21)
f (w) = C&D(cid:20)(cid:0)(P (w) − A)X(cid:1)−1
(cid:21) .
1−ζ(cid:17) − Iihf(cid:16) 1+ζ
1−ζ(cid:17) + Ii−1
1−ζd(cid:17) if ζ = (ζ1, . . . , ζd) ∈ Dd), then S is in the
(where by convention
, . . . , 1+ζd
(1.4) 1+ζ
Schur -- Agler class SA(Dd,L(U)) and S has a realization as in (1.2) where
U = [ A B
C D ] is unitary with the additional property that 1 is not in the point
spectrum of U.
(2) If S(ζ) = C(f )(ζ) :=hf(cid:16) 1+ζ
1−ζ = (cid:16) 1+ζ1
1−ζ1
B
IU
Proof. (1)⇒(2): We suppose that we are given a Π-impedance-conservative system
node Y as in condition (1). We set
H(w) =(cid:0)(P (w) − A)X(cid:1)−1
B : U → X
34
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
(well-defined by Proposition 2.10) and verify that
(cid:20)A&B
C&D(cid:21)(cid:20)H(w)u
u
(cid:21) =(cid:20)P (w)H(w)u
f (w)u
(cid:21)
(5.11)
for all u ∈ U. Furthermore, working as in the proof of (3) ⇒ (2) in Theorem 5.1,
we see that H(w) so defined provides an Agler decomposition for f :
f (z)∗ + f (w) =
(zk + wk)H(z)∗PkH(w).
dXk=1
If we set eY = h −(A&B)
C&D i, then the fact that Y = (cid:2) A&B
C&D(cid:3) is an impedance-
conservative system node means that eY is skew-adjoint (see Corollary 2.7): eY∗ =
−eY. Moreover we can rewrite the identity (5.11) in the form
eY(cid:20)H(w)u
As eY is skew-adjoint, easily verified properties of the Cayley transform imply
that U := (eY − I)(eY + I)−1 is unitary and the point 1 is not in the point spectrum
of U (this is another version of Proposition 2.3).
U =: [ A B
(cid:21) =(cid:20)−P (w)H(w)u
C D ] provides a Dd-scattering-conservative realization of S := C(f ).
It remains only to check that
f (w)u
(5.12)
(cid:21)
u
We rewrite (5.12) in terms of U as
or equivalently
We reorganize this using linearity to get
u
f (w)u
(I + U)(I − U)−1(cid:20)H(w)u
(I + U)(cid:20)H(w)u
U(cid:20)(P (w) − I)H(w)u
(cid:21) =(cid:20)−P (w)H(w)u
(cid:21)
(cid:21) = (I − U)(cid:20)−P (w)H(w)u
(cid:21)
−(f (w) + I)u (cid:21) =(cid:20)(P (w) + I)H(w)u
−(f (w) − I)u (cid:21) .
f (w)u
u
1
w + 1(cid:19) (f (w) + I).
(wk − 1)PkH(w)
wk + 1eHk(cid:18) w − 1
dXk=1
dXk=1
wk + 1eHk(ζ)(f (w) + I)
dXk=1
ζkeHk(ζ)(f (w) + I)
= P (ζ)eH(ζ)(f (w) + I)
wk − 1
=
=
We introduce L(U ,Xk)-valued functions eHk on the polydisk Dd according to the
relation (where we again use the convention (1.5))
Hk(w) := PkH(w) =
(5.13)
Then we note that
(P (w) − I)H(w) =
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
35
where we set
and
ζ =
w − 1
w + 1
(as in (1.5)) for w ∈ Πd
Similarly one can verify that
(P (w) + I)H(w) =
dXk=1
and we have arrived at the pair of identities
dXk=1 eHk(ζ).
eH(ζ) =
wk + 1eHk(ζ)(f (w) + I) = eH(ζ)(f (w) + I),
wk + 1
Fix a vector v ∈ U. Define uw = (I + f (w))−1v so v = (I + f (w))uw. Then (5.13)
with u = uw can be rewritten in the form
(5.14)
(P (w) − I)H(w) = P (ζ)eH(ζ)(I + f (w))
(P (w) + I)H(w) = eH(ζ)(I + f (w)).
(cid:21) =(cid:20) eH(ζ)v
−S(ζ)v(cid:21) .
U(cid:20)P (ζ)eH(ζ)v
−v
Writing out U = [ A B
C D ], one can now solve (5.14) in the standard way to arrive at
S(ζ) = D + CP (ζ)(I − AP (ζ))−1B,
i.e., the unitary colligation matrix U with the additional property that U does
not have 1 as an eigenvalue provides a Dd-scattering-conservative realization for
S = C(f ).
(2)⇒(1): We suppose that S = C(f ) has a Dd-scattering-conservative realization
C D ] does not have 1
(1.2) where the associated unitary colligation matrix U = [ A B
as an eigenvalue. Then we know that U also has the defining property
U(cid:20)P (ζ)eH(ζ)
(cid:21) u =(cid:20)eH(ζ)
S(ζ)(cid:21) u
for all u ∈ U and ζ ∈ Dd where eH(ζ) =
provides a Dd-Schur -- Agler
decomposition (1.1). Since 1 is not an eigenvalue of U by assumption we may form
the Cayley transform
eH1(ζ)
...
(5.15)
eHd(ζ)
I
eY := (I + U)(I − U)−1 with D(eY) = Ran(I − U).
By construction we have
By Proposition 2.3, eY is skew-adjoint:
eY : (I − U)(cid:20)x
eY = −eY∗.
u(cid:21) 7→ (I + U)(cid:20)x
u(cid:21) .
(5.16)
(5.17)
36
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
From (5.15) we note that
Notice that (I − S(ζ)) = 2(f (w) + I)−1 is invertible. Hence (5.17) leads to
(I − U)(cid:20)P (ζ)eH(ζ)
(I + U)(cid:20)P (ζ)eH(ζ)
eY(cid:20)(P (ζ) − I)eH(ζ)(I − S(ζ))−1
I
I
I
(I − S(ζ))u
(cid:21) u =(cid:20)(P (ζ) − I)eH(ζ)u
(cid:21) ,
(cid:21) u =(cid:20)(P (ζ) + I)eH(ζ)
(I + S(ζ))u (cid:21) .
(cid:21) u =(cid:20)(P (ζ) + I)eH(ζ)(I − S(ζ))−1u
wk + 1eHk(cid:18) w − 1
w + 1(cid:19) (I + f (w))
(I + S(ζ))(I − S(ζ))−1u
1
Let us set
where we use the convention (1.5) as usual, and note that
Then we compute
I + f (w) = 2(I − S(ζ))−1.
Hk(w) =
(cid:21) .
From these identities we see that the identity (5.18) is equivalent to
eY(cid:20)−H(w)
I
u ] ∈ D(eY).
In particular, for u ∈ U then there is an xu = −H(w)u ∈ X so that [ xu
As we have already observed that eY is skew-adjoint (see (5.16)), we now have
all the hypotheses needed in order to apply Corollary 2.7 to conclude that Y =
C&D(cid:3) is a Π-impedance-conservative system node. Moreover the
eY(cid:2) −I 0
0 I(cid:3) =: (cid:2) A&B
relation (5.19) leads to the relation
Y(cid:20)H(w)
I (cid:21) u =(cid:20)P (w)H(w)
f (w)
(cid:21) u.
(5.18)
(5.19)
(P (ζ) − I)eH(ζ)(I − S(ζ))−1 =
and similarly
(P (ζ) + I)eH(ζ)(I − S(ζ))−1 =
2
= −
= −
wk + 1 − 1(cid:19) eHk(ζ)(I − S(ζ))−1
wk + 1eHk(ζ)(I − S(ζ))−1
Hk(w) = −H(w)
dXk=1(cid:18) wk − 1
dXk=1
dXk=1
dXk=1(cid:18) wk − 1
dXk=1
wk + 1eHk(ζ)(I − S(ζ)−1
dXk=1
(cid:21) u =(cid:20)P (w)H(w)
wkHk(w) = P (w)H(w).
+ 1(cid:19) eHk(ζ)(I − S(ζ))−1
(cid:21) u.
wk + 1
f (w)
2wk
=
=
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
37
We now recover f (w) as the transfer-function for the Π-impedance-conservative
system node Y exactly as in the proof of (2)⇒(3) in Theorem 5.1.
(cid:3)
Remark 5.3. The implication (2)⇒(1) in Theorem 5.2 is due essentially to Agler --
McCarthy -- Young [3] (without explicit reference to Π-impedance-conservative sys-
tem nodes), where the result is worked out for the scalar-valued case in the Nevan-
linna -- Agler (rather than Herglotz -- Agler) setting.
6. Bessmertnyı long resolvent representations for Herglotz -- Agler
functions
Bessmertnyı long-resolvent representations were introduced by Bessmertnyı in
connection with the study of general rational matrix functions of several variables,
with a special symmetrized form of such a representation handling functions f in the
Πd-Herglotz -- Agler class having an extension to Ωd :=Sλ∈T(λΠ)d ⊂ Cd satisfying
additional symmetry conditions (see [17, 18, 19, 20, 21, 28] and our companion
paper [11] for more detail). In [9], a relaxation of the symmetrized Bessmertnyı
long-resolvent representation was proposed which handles more general Herglotz --
Agler-class functions (i.e., the homogeneity property is discarded).
Given a 2 × 2-block operator pencil
(cid:20)V11(w) V12(w)
V21(w) V22(w)(cid:21) =: V(w) = V0 + w1V1 + ··· + wdVd ∈ L(U ⊕ X )
(6.1)
we define the transfer function fV(w) associated with the operator pencil V by
fV(w) := V11(w) − V12(w)V22(w)−1V21(w).
(6.2)
wherever the formula makes sense. Let us say that a representation f (w) = fV(w)
of the form (6.2) for a given f is a B-realization if the operator pencil V(w) satisfies
(1) the B-symmetry condition V(w) = −V(−w)∗, namely
V0 + V∗0 = 0, Vk = V∗k for k = 1, . . . , d,
(6.3)
and (2) the B-positivity condition
Vk = V∗k ≥ 0 for k = 1, . . . , d with
dXk=1
V22,k strictly positive definite.
(6.4)
It was shown in [9] that any function f of the form (6.2) with pencil V(w) satisfying
conditions (6.3) and (6.4) is in the Herglotz -- Agler class HA(Πd,L(U)).
a B-realization with operator pencil V(w) = V0 + w1V1 + ··· + wdVd given by
We note that the realization (5.3) in Theorem 5.1 can formally be considered as
V0 =(cid:20) D&C
−(B&A)(cid:21) = −V∗0, Vk =(cid:20)0
0
0 Yk(cid:21) = V∗k ≥ 0 for k = 1, . . . , d
meeting all the constraints (6.3) -- (6.4) for a B-realization of f with the exception
that V0 is unbounded. We now make precise how a general Herglotz -- Agler-class
function (i.e., with the growth condition at infinity (5.1) removed) can be com-
pletely characterized in terms of possession of a nonhomogeneous Bessmertnyı-type
representation of the form (6.2) subject to (6.3) and (6.4) but with possibly un-
bounded skew-adjoint operator V0.
38
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
We shall consider unbounded pencils of the form (6.1) but with the constant
term
a possibly unbounded operator on(cid:20)U
Agler system node properties:
V0,21 V0,22(cid:21)
V0 =(cid:20)V0,11 V0,12
X(cid:21) satisfying what we shall call the Herglotz --
The gist of conditions (HA1) and (HA2) is that the reorganized colligation matrix
].
(HA1) V0 is skew-adjoint on [ U
X
(HA2) For each u ∈ U there is an xu ∈ X so that [ u
0(cid:21)
(cid:20)A&B
C&D(cid:21) =(cid:20) 0 −IX
0 (cid:21) V0(cid:20) 0
IX
IU
IU
xu ] ∈ D(V0).
is an impedance-conservative system node as discussed in Example 2.6. To reduce
the number of subscripts and to suggest the connection with system nodes in the
work of Staffans et al.
[29, 35, 36, 12], we shall use the notation encoded in the
following formal definition.
Definition 6.1. We shall say that the pencil (6.1) is a Herglotz -- Agler operator
pencil if the following conditions hold:
(1) V0 has the form
I
−(B&A)(cid:21) :=(cid:20) 0
V0 =(cid:20) D&C
C&D(cid:21) is an impedance-conservative system node as in Example 2.6.
where(cid:20)A&B
(2) The homogeneous part of the pencil, VH(w) :=Pd
0(cid:21)(cid:20)A&B
C&D(cid:21)(cid:20)0
k=1 wkVk has each
0(cid:21)
−I
I
I
Vk,21 Vk,22(cid:21)
Vk =(cid:20)Vk,11 Vk,12
dXk=1(cid:20)Vk,11 Vk,12
Vk,21 Vk,22(cid:21) =(cid:20)VH,11(e)
0
0
I(cid:21)
a bounded positive semidefinite operator on U ⊕ X with the sum having
the form
VH(e) =
(6.5)
where we use the notation e = (1, . . . , 1) as in (5.1).
Let us suppose that
V(w) = V0 + VH(w) =(cid:20) D&C
is a Herglotz -- Agler operator pencil. Thus in particular (cid:20)0 −I
−(B&A)(cid:21) +(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21)
0 (cid:21) V0(cid:20)0
I
I
all the properties delineated in Definition 2.4, with the additional property that
V0 = −V∗0 (see Example 2.6). In particular we may write
I
0(cid:21) has
Here AX : X → X−1 is the extension of the skew-adjoint operator A defined by
.
B&A = (cid:2)B AX(cid:3)(cid:12)(cid:12)D(V0)
x(cid:21) ∈ D(V0)(cid:27) with Ax = −V0(cid:20)0
x(cid:21)
D(A) =(cid:26)x : (cid:20)0
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
39
and with B : U → X−1 constructed via
Bu = B&A(cid:20) u
xu(cid:21) − AX x
where xu is any choice of vector in X such that [ u
xu ] ∈ D(V0); it can be shown that
B is well defined, i.e., the formula for Bu is independent of the choice of xu ∈ X .
Furthermore, the domain of V0 is the same as the domain of B&A and has the
precise characterization
D(V0) = D(B&A) =(cid:26)(cid:20)u
x(cid:21) ∈(cid:20)U
X(cid:21) : Bu + AX x ∈ X(cid:27) .
The following proposition gives some additional key properties for Herglotz -- Agler
operator pencils.
Proposition 6.2. Suppose that
V(w) = V0 + VH(w) :=(cid:20) D&C
−(B&A)(cid:21) +(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21)
is a Herglotz -- Agler operator pencil. Then, for a given w ∈ Πd, the formal Bess-
mertnyı transfer function (6.2),
fV(w) = V11(w) − V12(w)V22(w)−1V21(w)
= (D + VH,11(w)) − (C + VH,12(w))(VH,22(w) − A)−1(−B + VH,21(w)),
can be interpreted as a bounded operator on U equal to the sum of five well-defined
bounded terms
where
fV(w) = fV,1(w) + fV,2(w) + fV,3(w) + fV,4(w) + fV,5(w)
(6.6)
I
fV,1(w) = D&C(cid:20)
(VH,22(w) − A)−1B(cid:21) ,
fV,2(w) = VH,12(w)(cid:0)(VH,22(w) − A)X(cid:1)−1
fV,3(w) = −C(VH,22(w) − A)−1VH,21(w),
fV,4(w) = −VH,12(w)(VH,22(w) − A)−1VH,21(w),
fV,5(w) = VH,11(w).
VH,21(w),
(6.7)
Directly in terms of the Bessmertnyı pencil V(w), we have, for each w ∈ Πd,
(cid:20)
−V22(w)−1V21(w)(cid:21) ∈ L(U ,D(V0)),
I
V11(w)&V12(w) ∈ L(D(V0),U),
(6.8)
(6.9)
(6.10)
and we recover fV(w) as the composition of bounded operators
fV(w) = V11(w)&V12(w) ·(cid:20)
−V22(w)−1V21(w)(cid:21) ∈ L(U).
I
Proof. Analysis of fV,1: It follows from part (4) of Proposition 2.10 that the oper-
atorh ((VH,22(w)−A)X )−1
that h ((VH,22(w)−A)X )−1
IU
IU
i maps D(V0) into D(V0). Furthermore, one can check
i is bounded as an operator from U to D(V0) (with
D(V0) equipped with the graph norm of VH,22(w)). As D&C has domain equal to
B
B
40
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
B
IU
D(V0), we see that D&C maps D(V0) into U. Moreover, part (4) of Definition 2.4
assures us that D&C is bounded as an operator from D(V0) into U. We conclude
that fV,1(w) = D&C ·h ((VH,22(w)−A)X )−1
Analysis of fV,2: By part (2) of Proposition 2.10,(cid:0)(VH,22(w) − A)X(cid:1)−1
it follows that fU,2(w) = VH,12(w) ·(cid:0)(VH,22(w) − A)X(cid:1)−1
B maps
U boundedly into X . As VH,12(w) is bounded as an operator from X into U,
B is bounded as an
i ∈ L(U).
operator on U.
Analysis of fV,3: It is a consequence of part(2) of Proposition 2.10 that the
operator (VH,22(w)− A)−1 is bounded from X into X1. From the definition (2.8) of
C, we see that C maps X1 boundedly into U. Since also VH,21(w) is bounded as an
operator from U to X , it follows that fV,3(w) = −C · (VH,22(w) − A)−1 · VH,21(w)
defines a bounded operator on U.
Analysis of fV,4: Note that VH,21(w) ∈ L(U ,X ), by part (3) of Proposition
2.10 (VH,22(w) − A)−1 is bounded as an operator from X into X1 and hence as an
operator from X into itself, and VH,12(w) is bounded as an operator from X into
U. It follows that fV,4(w), as a composition of bounded operators, is bounded as
an operator on U.
Analysis of fV,5: This is the easiest term: VH,11(w) is a bounded operator on
U from the definition of Herglotz-Agler pencil (Definition 6.1).
Verification of formula (6.10): We first write out V(w) in terms of constant term
and homogeneous part:
Thus
V(w) = V0 + VH,k(w) =(cid:20) D&C
−(B&A)(cid:21) +(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21) .
V11(w)&V12(w) = D&C +(cid:2)VH,11(w) VH,12(w)(cid:3) .
The first term maps D(U0) boundedly into U while the second term is bounded
from the larger space [ U
] into U. It follows that the sum indeed is bounded from
X
D(U0) into U, verifying property (6.9). Similarly,
(cid:20)
I
−V22(w)−1V21(w)(cid:21) =(cid:20)
=(cid:20)
I
(VH,22(w) − A)−1(B − VH,21(w)(cid:21)
(VH,22(w) − A)−1B(cid:21) −(cid:20)
I
Proposition 2.10 while the second term maps U boundedly into (cid:2) 0
The first term maps U boundedly into D(U0) as a consequence of part (4) of
(and hence also boundedly into D(U0)) by part (2) of Proposition 2.10; this verifies
property (6.8). Thus the composition in the formula (6.10) defines a bounded
operator on U. Working out the various pieces in detail, we see that the result
agrees with the formula for fV(w) in (6.6).
(cid:3)
0
(VH,22(w) − A)−1VH,21(w)(cid:21) .
X1(cid:3) ⊂ D(U0)
The following is the main result of this section.
Theorem 6.3. Given a function f : Πd → L(U), the following are equivalent:
(1) f is in the Herglotz -- Agler class HA(Πd,L(U)).
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
41
(2) f has a Πd-Herglotz -- Agler decomposition, i.e., there exist L(U)-valued pos-
itive kernels K1, . . . , Kd on Πd such that
f (z)∗ + f (w) =
(3) There exists a Hilbert space X and a Herglotz -- Agler pencil
dXk=1
V(w) = V0 + VH(w) =(cid:20) D&C
−(B&A)(cid:21) +(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21)
(zk + wk)Kk(z, w).
such that f (w) = fV(w) (with fV(w) as in (6.6) -- (6.7) or (6.10)).
Proof. (1) ⇔ (2): This is already done in the proof of Theorem 5.1.
F over the polydisk Dd:
(1) or (2) ⇒ (3): We start with the representation for a Herglotz -- Agler function
F (ζ) = R + V ∗(U − P (ζ))−1(U + P (ζ))V
(6.11)
where U is unitary, P (ζ) = ζ1P1 + ··· + ζdPd is a spectral decomposition of the
identity on the state space X , R = F (0)−F (0)∗
as in formula
(3.10). We use the tuple version of the mutually inverse Cayley changes of variable
(1.3):
and V ∗V = F (0)+F (0)∗
2
2
1 + ζ
1 − ζ ∈ Πd, w ∈ Πd 7→
w − 1
w + 1 ∈ Dd,
ζ ∈ Dd 7→
in the Herglotz -- Agler class over the polydisk Dd and hence we can represent F (ζ)
with the conventions (1.4) and (1.5) in force. If we set F (ζ) = f(cid:16) 1+ζ
as in (6.11). Moreover, we recover f (w) from F (ζ) via f (w) = F(cid:16) w−1
1−ζ(cid:17), then F is
w+1(cid:17). This
leads to the formula
f (w) = R + V ∗M (w)V
(6.12)
where we set
M (w) =(cid:18)U − P(cid:18) w − 1
w + 1(cid:19)(cid:19)−1(cid:18)U + P(cid:18) w − 1
w + 1(cid:19)(cid:19) .
We compute further
M (w) =(cid:0)U − (P (w) − I)(P (w) + I)−1(cid:1)−1(cid:0)U + (P (w) − I)(P (w) + I)−1(cid:1)
= (P (w)U + U − P (w) + I)−1 (P (w)U + U + P (w) − I)
= (P (w)(U − I) + (U + I))−1 (P (w)(U + I) + (U − I))
(6.13)
Let us split out the eigenspace of U for eigenvalue z = 1 (if any) by writing U in
the form
U =(cid:20)I
0 U0(cid:21)
0
with respect to the decomposition X = X (1) ⊕X (0) (X (1) equal to the 1-eigenspace
for U and X (0) equal to the orthogonal complement of X (1) in X ). Then U0 is
unitary but does not have 1 as an eigenvalue. We have
U − I =(cid:20)0
0 U0 − I(cid:21) , U + I =(cid:20)2I
0 U0 + I(cid:21) .
0
0
42
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
and hence
M (w) =(cid:18)P (w)(cid:20)0
0 U0 − I(cid:21) +(cid:20)2I
0
0
0 U0 + I(cid:21)(cid:19)−1
·(cid:18)P (w)(cid:20)2I
T = (I + U0)(I − U0)−1.
Let us set
0
0 U0 + I(cid:21) +(cid:20)0
0 U0 − I(cid:21)(cid:19)
0
Then T is a possibly unbounded skew-adjoint operator on X0 (by Proposition 2.3)
and we have the following relations between T and U0:
U0 = (T − I)(T + I)−1 = I − 2(T + I)−1
We may then continue the computation (6.13) to get
M (w) =(cid:20) 1
0
0
0
0
2 I
0
(T + I).
⇒ U0 − I = −2(T + I)−1,
(U0 − I)−1(cid:21)(cid:18)P (w)(cid:20)0
(U0 − I)−1 = −
I(cid:21) +(cid:20)I
(U0 + I)(U0 − I)−1(cid:21) +(cid:20)0
0 I(cid:21) +(cid:20)I
0 I(cid:21)(cid:19)(cid:20)I
0 −(T + I)∗(cid:21)
(U0 + I)(U0 − I)−1(cid:21)(cid:19)−1
I(cid:21)(cid:19)(cid:20)2I
0 −T(cid:21)(cid:19)−1
0 −(T + I)−1(cid:21)
·(cid:18)P (w)(cid:20)I
=(cid:20)I
0 −(T + I)(cid:21)(cid:18)P (w)(cid:20)0
·(cid:18)P (w)(cid:20)I
0 −T(cid:21) +(cid:20)0
=(cid:20)I
0 −(T + I)(cid:21) N (w)(cid:20)I
0 U0 − I(cid:21)
0
0
0
0
0
0
0
0
0
0
0
0
0
·
0
1
2
·
(6.14)
0
0
0
0
0
If we write out the block matrix decomposition
where, due to the identity −(T + I)∗ = T − I arising from T = −T ∗, N (w) is given
by
N (w) =(cid:18)P (w)(cid:20)0
I(cid:21) +(cid:20)I
0 −T(cid:21)(cid:19)−1
·(cid:18)P (w)(cid:20)I
0 −T(cid:21) +(cid:20)0
0 I(cid:21)(cid:19) ·(cid:20)I
Pk,01 Pk,00(cid:21)
wk(cid:20)Pk,11 Pk,10
P (w) =(cid:20)P11(w) P10(w)
P01(w) P00(w)(cid:21) =
dXk=1
X (0)(cid:21) of X , we can write out
of P (w) with respect to the decomposition X = (cid:20)X (1)
(cid:18)P (w)(cid:20)0
=(cid:20)I
0 P00(w) − T(cid:21)−1
(I − T 2)−1(cid:21) .
0 −T(cid:21)(cid:19)−1
I(cid:21) +(cid:20)I
more explicitly
P10(w)
(6.15)
0
0
0
0
0
=(cid:20)I −P10(w)(P00(w) − T )−1
(P00(w) − T )−1
0
(cid:21) ,
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
43
P (w)(cid:20)I
0 −T(cid:21) +(cid:20)0
0
and from (6.15) we see that N (w) is given by
N (w) =(cid:20)I −P10(w)(P00(w) − T )−1
(P00(w) − T )−1
0
0
P01(w)
0 I(cid:21) =(cid:20)P11(w) −P10(w)T
I − P00(w)T(cid:21)
(I − P00(w)T )(I − T 2)−1(cid:21) .
−P10(w)T (I − T 2)−1
(cid:21)(cid:20)P11(w)
P01(w)
(6.16)
At this stage it is convenient to introduce the Gelfand triple (or rigging) of X (0)
associated with the (possibly unbounded) skew-adjoint operator T :
1
X (0)
:= Dom T = Ran(I − T )−1,
X (0)
−1 := completion of X (0) in X (0)
Then we see that (I − T )−1 is well defined as an element of L(X (0),X (0)
L(X (0)
−1 ,X (0)). A careful inspection of the formula (6.16) for N (w) shows that
−1 -norm: kxk−1 = k(I − T )−1xkX (0) for x ∈ X (0).
) and of
1
(6.17)
(6.18)
1 (cid:21)
−1(cid:21) →(cid:20)X (1)
N (w) : (cid:20)X (1)
X (0)
X (0)
J := −T (I − T 2)−1,
from which it follows that the formula (6.14) gives sense for M (w) as an element
of L(X (1) ⊕ X (0)). However the formula (6.16) (and (6.15)) lacks symmetry. To fix
this we introduce the operator J by
Thus J = −J∗ and we can consider J as an element of L(X (0),X (0)
as L(X (0)
L(X (0)
) as well
−1 ,X (0)). While the operator I − P00(w)T makes sense as an element of
,X (0)) (as well as L(X (0),X (0)
−1 )), the individual terms in the additive de-
1
1
composition
I − P00(w)T = (I − T 2) + (T 2 − P22(w)T )
make sense only as elements in L(X (0)
(6.19)
−1 ). Nevertheless, we proceed to get a
more symmetric formula for N (w) as follows. Note first that the decomposition
(6.19) leads to
,X (0)
1
(I − P00(w)T )(I − T 2)−1 = I − (P00(w) − T )T (I − T 2)−1
0
P10(w)J
(P00(w) − T )−1
From (6.16) and the definition (6.18) of J, we then have
N (w) =(cid:20)I −P10(w)(P00(w) − T )−1
= I + (P00(w) − T )J : X (0) → X (0).
(cid:21)(cid:20)P11(w)
I + (P00(w) − T )J(cid:21)
=(cid:20)P11(w) − P10(w)(P00(w) − T )−1P01(w) −P10(w)(P00(w) − T )−1
(P00(w) − T )−1 + J (cid:21)
−1(cid:21) to(cid:20)X (1)
X (0)(cid:21) rather than
which a priori makes sense only as an operator from(cid:20)X (1)
X (0)
1 (cid:21) as in (6.17), due to the decoupling of an ∞ − ∞ cancellation occurring
to(cid:20)X (1)
X (0)
(P00(w) − T )−1P01(w)
in the application of the decomposition (6.19). This in turn leads to difficulties
P01(w)
(6.20)
44
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
in understanding M (w) as a bounded operator on X (1) ⊕ X (0) from the formula
(6.14).
Continuation of the analysis with an extra assumption: Assuming for
the moment that T is bounded (as is the case for the situation where the state space
X is finite-dimensional as in the setting discussed in [11]), this difficulty does not
occur and we may continue as follows. From (6.20) we see that
N (w) =(cid:20)P11(w)
0
0
J(cid:21) −(cid:20)P10(w)
−I (cid:21) (P00(w) − T )−1(cid:2)P01(w)
If we block-decompose the operator V : U → X = (cid:20)X (1)
combine (6.12) with (6.14) and (6.21) while noting the simplification
(6.21)
I(cid:3) .
V0(cid:21) and then
X (0)(cid:21) as V =(cid:20)V1
(I + T )J(I + T )∗ = (I + T )[−(I + T )−1T (I − T )−1](I + T )∗ (by (6.18))
= −T (since T ∗ = −T ).
we arrive at
f (w) = R + V ∗1 P11(w)V1 − V ∗0 T V0
− [V ∗1 P10(w) + V ∗0 (I + T )](P00(w) − T )−1[P01(w)V1 − (I + T )∗V0].
We have arrived at a Bessmertnyı long-resolvent representation for f
(6.22)
where the operator pencil V(w) =(cid:20)V11(w) V12(w)
V21(w) V22(w)(cid:21) is given by
f (w) = fV(w)
V11(w) = R + V ∗0 (I + T )J(I + T )∗V0 + V ∗1 P11(w)V1,
V12(w) = V ∗0 (I + T ) + V ∗1 P10(w),
V21(w) = −(I + T )∗V0 + P01(w)V1,
V22(w) = −T + P00(w).
Thus the associated linear pencil
V21(w) V22(w)(cid:21) = V0 + w1V1 + ··· + wdVd
(6.23)
(6.24)
(6.25)
has coefficients
−(I + T )∗V0
V(w) =(cid:20)V11(w) V12(w)
V0 =(cid:20) R − V ∗0 T V0
=(cid:20)R 0
0(cid:21) +(cid:20)V ∗0
Vk =(cid:20)V ∗1 Pk,11V1 V ∗1 Pk,10
k=1 Pk,00 = IX (0) (sincePd
Pk,01V1
0
0
0
−T
V ∗0 (I + T )
(cid:21)
I(cid:21)(cid:20) −T
−(I + T )∗ −T (cid:21)(cid:20)V0
Pk,00 (cid:21) for k = 1, . . . , d.
I + T
0
0
I(cid:21) ,
Moreover, it is easily checked that V0 is skew-adjoint and that Pk ≥ 0 for each
k with Pd
k=1 Pk = IX ), and hence V(w) is a Herglotz --
Agler pencil and Theorem 6.3 is completely proved in case T = (I + U0)(I − U0)−1
is bounded on X (0).
Back to the general case: For the general case (where T is allowed to be
unbounded), the formula (6.25) for Vk still makes good sense and the Vk's meet
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
45
property (2) in Definition 6.1. The next step is to make sense of the formula (6.24)
for V0.
The first term in the formula (6.24) for V0 can always be added in later so we
focus on the second term V′0:
V′0 =(cid:20)V ∗0
0
0
I(cid:21)(cid:20) −T
−(I + T )∗ −T (cid:21)(cid:20)V0
I + T
0
0
I(cid:21) .
We view V′0 as a possibly unbounded operator with dense domain in(cid:20)U
X(cid:21) given by
Then the flip of V′0, namely
D(V′0) =(cid:26)(cid:20)u
x(cid:21) ∈(cid:20)U
0 (cid:21) V′0(cid:20)0
0(cid:21) =(cid:20)I
X(cid:21) : x − V0u ∈ D(T )(cid:27) .
0 V ∗0(cid:21)(cid:20) T
−T (cid:21)(cid:20)I
I − T
I + T
0
I
I
(cid:20)0 −I
I
0
0 V0(cid:21)
has exactly the form of the model Π-impedance-conservative system node given in
Proposition 2.8. We can now conclude that V(w) given by (6.23), (6.24), (6.25) is
indeed a Herglotz -- Agler pencil. It remains only to verify that f (w) = fV(w).
From the formula (6.12) for f combined with the formula (6.14) for M (w) and
the formula (6.20) for N (w), we know that f (w) has the representation
f (w) = R + V ∗1 P11(w)V1 − V ∗1 P10(w)(P00(w) − T )−1P01(w)V1
− V ∗0 (I + T )(P00(w) − T )−1P01(w)V1
+ V ∗1 P10(w)[T (I + T )−1 + (P00(w) − T )−1(I − P00(w)T )(I + T )−1]V0
+ V ∗0 (I + T )(P00(w) − T )−1(I − P00(w)T )(I + T )−1V0.
(6.26)
On the other hand the Bessmertnyı transfer function associated with the pencil V
(6.23) can be written as
fV(w) = [V11(w)&V12(w)](cid:20)
−V22(w)−1V21(w)(cid:21)
I
= R + V ∗0 (cid:2)−T + (I + T )(P00(w) − T )−1(I − T )(cid:3) V0
+ V ∗1 P11(w)V1 − V ∗1 P10(w)(P00(w) − T )−1P01(w)V1
− V ∗0 (I + T )(P00(w) − T )−1P01(w)V1
+ V ∗1 P10(w)(P00(w) − T )−1(I − T )V0.
(6.27)
−1 and V ∗0 ∈ L(X (0),U) has no extension to X (0)
Note that care must be taken in writing the first term of the expression after the R
term: the individual expressions −V ∗0 T V0 and V ∗0 (I + T )(P00(w) − T )−1(I − T )V0
make no sense since the operators −T and (I + T )(P00(w)− T )−1(I − T ) map X (0)
into X (0)
−1 ; as we shall see in detail
below, the combination −T + (I + T )(P00(w)− T )−1(I − T ) fortuitously maps X (0)
back into itself so that the combined term V ∗0 (−T +(I +T )(P00(w)−T )−1(I−T ))V0
makes good sense as a bounded operator on U for each w ∈ Πd; roughly speaking,
this is where we couple back together the ∞ − ∞ cancellation introduced earlier
to make our formulas once again make sense. Note also that the formula (6.27)
agrees with the formula (6.22) once one takes care to rearrange the terms in (6.22)
so that the result makes sense as a well-defined bounded operator on U defining
the operator f (w).
46
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
By the analysis done above with the extra assumption imposed, we see that the
two expressions (6.26) and (6.27) agree in the special case where the skew-adjoint
operator T is bounded. Once the operator −T +(I +T )(P00(w)−T )−1P01(w)(I−T )
is exhibited more explicitly as a bounded operator on X (0) (even in the case where
T itself is unbounded), it is possible to verify the equality of the two expressions
(6.26) and (6.27) by approximating the unbounded case by the bounded case and
then taking limits. As this is really about algebra, however, perhaps more satisfying
is to verify the equality between (6.26) and (6.27) directly by brute-force algebra.
Toward this goal, we note that each term in (6.26) can be paired with an identical
term in (6.27) once we establish the validity of the two identities:
(I + T )(P00(w) − T )−1(I − P00(w)T )(I + T )−1
= −T + (I + T )(P00(w) − T )−1(I − T )
P10(w) [T (I + T )−1 + (P00(w) − T )−1(I − P00(w)T )(I + T )−1]
= P10(w)(P00(w) − T )−1(I − T ).
(6.28)
In particular, (6.28) demonstrates how the expression −T + (I + T )(P00(w) −
T )−1(I − T ) actually defines a bounded operator on X . These two identities can be
verified directly by brute-force algebra; we leave the details to the reader (or as an
exercise for MATHEMATICA). This concludes the proof of (1) or (2) ⇒ (3) in Theorem
6.3.
(3) ⇒ (2): We assume that f (w) = fV(w) for a Herglotz -- Agler pencil
V(w) =(cid:20)V11(w)&V12(w)
V21(w)&V22(w)(cid:21) =(cid:20) D&C
−(B&A)(cid:21) +(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21) .
Thus, as explained in Proposition 6.2, if we set xw = −V22(w)−1V21(w) for w ∈ Πd,
then for each u ∈ U we have
and
We may also compute
(cid:20) u
xwu(cid:21) ∈ D(cid:18)(cid:20)V11(w)&V12(w)
V21(w)&V22(w)(cid:21)(cid:19)
f (w)u = (V11(w)&V12(w))(cid:20) u
xwu(cid:21)
−V22(w)−1V21(w)u(cid:21)
−V22(w)−1V21(w)u(cid:21) (as a vector in X (0)
u
u
−1 )
(6.29)
(V21(w)&V22(w))(cid:20)
=(cid:2)V21(w) V22(w)(cid:3)(cid:20)
= V21(w)u − V21(w)u = 0.
Thus (6.29) can be expanded to the identity
(cid:20)V11(w)&V12(w)
V21(w)&V22(w)(cid:21)(cid:20) u
xwu(cid:21) =(cid:20)f (w)
0 (cid:21) u.
(6.30)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
47
In addition to u ∈ U and w ∈ Πd, choose another pair u′ ∈ U and z ∈ Πd and
consider the sesquilinear form
Q(z, w)[u, u′] :=
As a consequence of (6.30) we see that
(cid:28)(cid:20)V11(w)&V12(w)
V21(w)&V22(w)(cid:21)(cid:20) u
xwu(cid:21) ,(cid:20) u′
Q(z, w)[u, u′] =(cid:28)(cid:20)f (w)u
xzu′(cid:21)(cid:29) +(cid:28)(cid:20) u
(cid:21) ,(cid:20) u′
xzu′(cid:21)(cid:29) +(cid:28)(cid:20) u
xwu(cid:21) ,(cid:20)V11(z)&V12(z)
V21(z)&V22(z)(cid:21)(cid:20) u′
xwu(cid:21) ,(cid:20)f (z)u′
(cid:21)(cid:29)
0
0
xzu′(cid:21)(cid:29) .
= h(f (z)∗ + f (w)) u, u′i.
(6.31)
On the other hand, from the decomposition of V(w) as V(w) = V0 + VH(w) with
V0 = −V∗0, we have
Q(z, w)[u, u′] =DV(w)(cid:20) u
=D (VH(z)∗ + VH(w))(cid:20)
xwu(cid:21) ,(cid:20) u′
xzu′(cid:21)E +D(cid:20) u
−V22(w)−1V21(w)(cid:21) u,(cid:20)
I
xwu(cid:21) , V(z)(cid:20) u′
xzu′(cid:21)E
I
−V22(z)−1V21(z)(cid:21) u′E
(zj + wj)hH(z)∗VjH(w)u, u′i
(6.32)
where we set H(w) =(cid:20)
(6.32) gives us
−V22(w)−1V21(w)(cid:21) ∈ L(U ,U ⊕ X ). Combining (6.31) and
I
f (z)∗ + f (w) =
(zj + wj )H(z)∗VjH(w)
=
dXj=1
dXj=1
where Vj ≥ 0 on U ⊕ X by assumption, and (2) follows.
(cid:3)
We next illustrate Theorem 6.3 by looking at some special cases.
Special case 1: V1 = 0. We note that the case V1 = 0 in the proof of Theorem
6.3 is exactly the case where the representation (6.27) for f collapses to
f (w) = R + V ∗0 [−T + (I + T )(P00(w) − T )−1(I − T )]V0.
(6.33)
where P00(w) = w1P1,00 + ··· + wdPd,00 is a positive decomposition of IX (0) , i.e.,
f has a representation exactly as in part (3) of Theorem 5.1. In general, from the
property (6.5) for a Herglotz -- Agler pencil, we have
VH(te) = tVH(e) = t(cid:20)VH,11(e)
0
0
I(cid:21) .
Thus, in the general representation (6.27) for f , f (te) takes on the simplified form
f (te) = R + V ∗0 [−T + (I + T )(tI − T )−1(I − T )]V0 + tV ∗1 V1.
We have already seen that
1
t
lim
t→+∞
[R + V ∗0 (−T + (I + T )(tI − T )−1(I − T ))V0] = 0.
We conclude that in general
1
t
lim
t→+∞
f (te) = V ∗1 V1.
48
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
Thus the growth condition at ∞ (5.1) is equivalent to the condition that V1 = 0.
In this way we arrive at Theorem 5.1 as a corollary of Theorem 6.3.
Special case 2: X (1) = {0}. This corresponds to the case where Pk = Pk,00 and
P00(w) = w1P1,00 + ··· + wdPd,00 is a spectral decomposition (not just a positive
decomposition) of IX (0) . Then the representation (6.27) collapses again to (6.33),
but this time with the stronger property that w1P1,00 + ··· + Pd,00wd is a spectral
decomposition of IX (0) , i.e., exactly the conclusion of part (1) of Theorem 5.2. On
the one hand, the condition X (1) = {0} means that the unitary operator U in the
Herglotz representation (3.10) for the function F ∈ HA(Dd,U) given by
1 − ζ(cid:19)
F (ζ) = f(cid:18) 1 + ζ
6.3 is that the colligation matrix U = (cid:20)A B
C D(cid:21) in the unitary Givone -- Roesser
does not have 1 as an eigenvalue. On the other hand, condition (1) in Theorem
representation
S(ζ) = D + C(I − P (ζ)A)−1P (ζ)B
(6.34)
for the SA(Dd,U)-class function
S(ζ) =(cid:18)f(cid:18) 1 + ζ
1 − ζ(cid:19) − I(cid:19)(cid:18)f(cid:18) 1 + ζ
1 − ζ(cid:19) + I(cid:19)−1
= (F (ζ) − I)(F (ζ) + I)−1 (6.35)
does not have 1 as an eigenvalue. To see that these conditions match up, we recall
from the discussion in Remark 2.11 (note formulas (3.7), (3.9) and (3.10)) that
we have the following connection between the unitary operator U in the Herglotz
representation (3.10) for F (ζ) and the unitary colligation matrix U = (cid:20)A B
C D(cid:21)
generating the Givone -- Roesser representation (6.34) for S(ζ) given by (6.35):
U∗ = U0 := A − B(I − D)−1C.
As was observed in Remark 3.2, I − D is injective since I − D = I − S(0) =
2(F (0) + I)−1. By the theory of Schur complements, given that I − D is injective,
then I − U is injective if and only if its Schur complement U0 = I − B(I − D)−1C
is injective. As U0 is unitary, this in turn is equivalent to U = (U0)∗ not having
1 as an eigenvalue. In this way we recover Theorem 5.2 as a corollary of Theorem
6.3.
Special case 3: X (0) = {0}. In this case, M (w) = P (w), and the representation
(6.27) collapses to
f (w) = V(w) = R + V ∗P (w)V.
Special case 4: R = 0, U = U∗, V0 = 0. In this case, T = T ∗ = −T ∗, hence
T = 0, and the linear pencil A(w) is homogeneous:
Moreover, V(e) =(cid:20)V ∗1 V1
0
0
0
0
0
0
I(cid:21) P (w)(cid:20)V1
V(w) = VH(w) =(cid:20)V ∗1
I(cid:21) .
I(cid:21), as in (6.5). The representation
f (w) = V ∗1 (P11(w) − P10(w)P00(w)−1P01(w))V1
= VH,11(w) − VH,12(w)VH,22(w)−1VH,21
(6.36)
(6.37)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
49
is then Bessmertnyı's long-resolvent representation in the infinite-dimensional set-
ting, as in [28], i.e., f belongs to the Bessmertnyı class Bd(U).
Remark 6.4. Notice that for a function f ∈ Bd(U) one can always find a ho-
mogeneous Bessmertnyı pencil VH(w) satisfying the condition (6.5). First of all,
Indeed, if ker VH,22(e) 6= {0}, then the maximum
VH,22(e) must be invertible.
principle and positivity of the coefficients Vk, k = 1, . . . , d, force VH(w) to have
the form
VH(w) =
in contradiction with condition (6.5) for a Herglotz-Agler pencil. Therefore, one
can replace the pencil VH(w) by the pencil
U
Ran VH,22(e)
Ker VH,22(e)
,
0
0
U
0
0
0
VH,11(w)
Ran VH,22(e)
Ker VH,22(e)
:
→
0 VH,22(e)−1(cid:21) ,
eVH,12(w)
eVH,22(w)
0 VH,22(e)−1(cid:21) VH(w)(cid:20)I
eVH,21(w)
bVH(w) =(cid:20)I
f (w) = bVH,11(w) − bVH,12(w)bVH,22(w)−1bVH,21.
0
0
which satisfies the condition (6.5) and provides another Bessmertnyı transfer-function
realization of f :
In fact, the special case described in Remark 6.4 covers the whole class Bd(U).
In the following theorem, we collect the characterizations of this class from [28],
together with the two additional characterizations: via its image in the class
HA(Dd,L(U)) under the Cayley transform over the variables, and as this special
case of Theorem 6.3.
Theorem 6.5. Given a function f : Πd → L(U), the following are equivalent:
(1) f is in the Herglotz -- Agler class HA(Πd,L(U)) and can be extended to a
(λΠ)d which satisfies the
holomorphic L(U)-valued function on Ωd := Sλ∈T
following conditions:
(a) Homogeneity: f (λw1, . . . , λw2) = λf (w1, . . . , wd) for every λ ∈ C\{0}
(b) Real Symmetry: f (w1, . . . , wd) = f (w1, . . . , wd)∗ for every w ∈ Ωd.
kernels K1, . . . , Kd on Πd such that the identity
(2) f has a Πd-Bessmertnyı decomposition, i.e., there exist L(U)-valued positive
and w = (w1, . . . , wd) ∈ Ωd.
f (w) =
wkKk(z, w),
z, w ∈ Πd,
holds, or equivalently, the following two identities hold:
f (z)∗ ± f (w) =
(zk ± wk)Kk(z, w),
z, w ∈ Πd.
dXk=1
dXk=1
(3) f belongs to the Bessmertnyı class Bd(U), i.e., there exist a Hilbert space
X and a Bessmertnyı pencil
V(w) = VH(w) =(cid:20)VH,11(w) VH,12(w)
VH,21(w) VH,22(w)(cid:21) =
wkVk,
dXk=1
50
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
where Vk ∈ L(U ⊕ X ) are positive semidefinite operators, so that (6.5)
holds and f (w) = fV(w) (with fV(w) as in (6.37)).
(4) The double Cayley transform of f ,
S(ζ) = C(f ) :=hf(cid:16) 1 + ζ
1 − ζ(cid:17) − Iihf(cid:16) 1 + ζ
1 − ζ(cid:17) + Ii−1
,
belongs to the Schur -- Agler class SA(Dd,L(U)) and has a unitary Givone --
Roesser realization (1.2) with a unitary and self-adjoint colligation matrix
U = U∗ = U−1.
(5) The Cayley transform of f over the variables,
F (ζ) = f(cid:16) 1 + ζ
1 − ζ(cid:17),
belongs to the Herglotz -- Agler class HA(Dd,L(U)) and has a representation
(6.11) with R = 0, U = U−1 = U∗, and Ran V contained in the 1-eigenspace
of U .
(6) There exist a Hilbert space X , its subspaces X (0), X (1) with X = X (1)⊕X (0),
a decomposition of IX , P (w) = w1P1 + ··· + wdPd = (cid:20)P11(w) P10(w)
P01(w) P00(w)(cid:21)
(with respect to the two-fold decomposition of X ), and V1 ∈ L(U ,X (1)),
such that f has the form (6.36).
Proof. The equivalence of statements (1), (2), (3), and (4) has been proved in [28].
(5)⇒(6) has been shown above, in the first paragraph of Special case 4. This is
an application of the construction in the proof of Theorem 6.3 to this special case.
(6)⇒(3) is obvious.
(4)⇒(5). The function F is related to S = C(f ) as in (3.8). Let U =(cid:20)A B
C D(cid:21)
be a unitary colligation matrix providing the transfer-function realization for S
as in (4), i.e., U∗ = U, together with the spectral decomposition of IX , P (ζ) =
ζ1P1 + ··· + ζdPd. As it was shown in Remark 3.2, F has a representation (3.10)
with skew-adjoint R, unitary U = U∗0 = A∗ − C∗(I − D∗)−1B∗ and V = 1√2
B.
Now we also have that R = R∗ = −R∗, which means that R = 0, and that
U = U∗ = A − B(I − D)−1B∗.
It follows that the space X has an orthogonal
decomposition X = X (1) ⊕ X (0), where X (1) is the 1-eigenspace of U and X (0) is
the (-1)-eigenspace of U . With respect to this decomposition, let us write
U0 =(cid:20) R − V ∗0 T V
−(I + T )∗V0
(cid:21) ,
V ∗0 (I + T )
−T
0
U = U0 =(cid:20)I
0 −I(cid:21) , A =(cid:20)A11 A10
Since A00 is a self-adjoint and contractive, the identity
−I = A00 + B0(I − D)−1B∗0
A01 A00(cid:21) , B =(cid:20)B1
B0(cid:21) .
is possible only if B0 = 0. Thus im V ⊆ X (1), which completes the proof.
Special case 5: the single-variable case. When we specialize Theorem 6.3 to
the single-variable case, some simplifications occur. In particular the Herglotz pencil
(6.23) involves only two operators, namely the flip of a Π-impedance-conservative
system node which can be assumed to be the canonical form (6.24)
(cid:3)
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
51
along with a single positive operator necessarily of the diagonal form
U1 =(cid:20)V ∗1 V1
0
Then the pencil U(w) has the form
U(w) =(cid:20) R − V ∗0 T V0
−(I + T )∗V0
V ∗0 (I + T )
−T
0
I(cid:21) .
(cid:21) + w(cid:20)V ∗1 V1
0
0
I(cid:21)
and the Bessmertnyı transfer-function realization for f ∈ H(Π,L(U)) becomes
f (w) = R + wV ∗1 V1 + V ∗0 (−T + (I + T )(wI − T )−1(I + T )∗)V0.
(6.38)
For simplicity, let us now assume that U = C (so f is scalar-valued). Then R ∈ C
is just a purely imaginary number and V ∗0 V0 is an operator on C and so can be
identified with a nonnegative real number α (the image of the operator V ∗0 V0 acting
on 1 ∈ C). From the representation (6.38), we see that there is no harm in cutting
the state space X (0) (on which T is acting) down to the smallest subspace reducing
for T which contains the range of the rank-1 operator V0, i.e., we may assume that
V0 · 1 is a cyclic vector for T . Then the spectral theorem tells us that there is a
measure ν on the imaginary line iR so that T is unitarily equivalent to
T = M−ζ : f (ζ) 7→ −ζf (ζ)
1
1−ζ . As this function must be in L2(ν), we conclude that
1+ζ2
acting on L2(ν). Without loss of generality we take the cyclic vector V0 · 1 to be the
function 1
1−ζ ∈ L2(ν),
dν(ζ) is a finite measure. When this is done then we see that the
i.e., that
adjoint operator V ∗0 : L2(ν) 7→ C is given by
1
1 − ζ
V ∗0 : f (ζ) 7→ZiR
f (ζ) dν(ζ).
1
Then the Bessmertnyı realization (6.38) for f (w) collapses to the integral represen-
tation formula
f (w) = αw + R +ZiR(cid:20)
ζ
1 + ζ2 +
1
ζ + w(cid:21) dν(ζ).
This agrees with the classical Nevanlinna integral representation for holomorphic
functions taking the right halfplane into itself. Actually the formula is usually
stated for holomorphic functions taking the upper halfplane into itself (see [24,
Theorem 1 page 20]); however the correspondence ef (ω) 7→ f (w) := −ief(iw) be-
tween ef in the Nevanlinna class and f in the Herglotz class enables one to easily
convert one integral representation to the other. We also point out that our proof
(starting with the Herglotz representation (3.10) for the Herglotz function F (ζ)
on the disk and then separating out the 1-eigenspace of the unitary operator U
in that representation) is just an operator-theoretic analogue of the proof of the
integral representation formula in [24], where one starts with the integral Herglotz
representation
F (ζ) = R +ZT
t + ζ
t − ζ
dµ(t)
and then separates out any point mass of µ at the point 1 on the circle.
52
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
Remark 6.6. In recent work [4], Agler -- Tully-Doyle -- Young obtain a realization
formula for the most general scalar-valued Nevanlinna -- Agler function on the upper
polyhalfplane. It is a straightforward matter to adjust the formulas to the right
polyhalfplane setting which we have here and to extend the results to the operator-
valued case. The result amounts to combining our formulas (6.12), (6.14) and
(6.15), i.e.,
f (w) =R + V ∗(cid:20)I
0
0 −(I + T )(cid:21)(cid:18)P (w)(cid:20)0
0 I(cid:21)(cid:19)(cid:20)I
0 −T(cid:21) +(cid:20)0
0 I(cid:21) +(cid:20)I
0 −T(cid:21)(cid:19)−1
0 −(I − T )−1(cid:21) V.
·(cid:18)−P (w)(cid:20)I
0
0
0
0
·
(6.39)
0
R
However the associated Bessmertnyı pencil
V(w) ="
(−P (w)(cid:2) I 0
0 −T(cid:3) + [ 0 0
0 I ])h I
0 −T(cid:3)#
V ∗(cid:2) I
0 −(I+T )(cid:3)
0 −(I−T )−1i V P (w) [ 0 0
0 I ] +(cid:2) I 0
0
0
lacks the symmetry properties of what we are calling a Herglotz -- Agler operator
pencil (see Definition 6.1). Hence there is no easy analogue of the proof of (3) ⇒
(2) in Theorem 6.3 and it is not all transparent from the presentation (6.39) why
the resulting function (6.39) has a Πd-Herglotz -- Agler decomposition (5.2); indeed,
it takes several pages of calculations in [4] (see Propositions 3.4 and 3.5 there) to
arrive at this result.
There are other results in [4] and [5] using realization theory to characterize
various types of boundary behavior of the function f at infinity; we do not go into
this topic here.
References
[1] J. Agler, On the representation of certain holomorphic functions defined on a polydisc, In
Topics in operator theory: Ernst D. Hellinger Memorial Volume, Oper. Theory Adv. Appl.,
Vol. 48, pp. 47 -- 66, Birkhauser, Basel, 1990.
[2] J. Agler and J.E. McCarthy, Nevanlinna -- Pick interpolation on the bidisk, J. Reine
Angew. Math. 506 (1999), 191 -- 204.
[3] J. Agler, J.E. McCarthy, and N.J. Young, Operator monotone functions and Lowner func-
tions of several variables, Annals of Mathematics 176 (2012), 1783 -- 1826.
[4] J. Agler, R. Tully-Doyle, and N.J. Young, Nevanlinna representations in several variables,
arXiv:1203.2261v2.
[5] J. Agler, R. Tully-Doyle, and N.J. Young, Boundary behavior of analytic functions of two
variables via generalized models, Indag. Math. (N.S.) 23 (2012) no. 4, 995 -- 1027.
[6] Y. Arlinskii, S. Belyi, and E. Tsekanovskii, Conservative Realizations of Herglotz -- Nevanlinna
Functions, Operator Theory: Advances and Applications 217, Birkhauser, 2011.
[7] D.Z. Arov and M.A. Nudelman, Passive linear stationary dynamical scattering systems with
continuous time, Integral Equations and Operator Theory 24 (1996), 1 -- 45.
[8] T.A. Azizov and I..S. Iokhvidov, Linear Operators in Spaces with an Indefinite Metric, Wiley,
1980.
[9] J. A. Ball. Multidimensional circuit synthesis and multivariable dilation theory, Multidi-
mens. Syst. Signal Process 22 (2011) no. 1-3, 27 -- 44.
[10] J.A. Ball and V. Bolotnikov, Canonical de Branges -- Rovnyak model transfer-function re-
alization for multivariable Schur-class functions, in: Hilbert Spaces of Analytic Functions
(Ed. J. Mashreghi, T. Ransford, and K. Seip), CRM Proceedings & Lecture Notes 51,
Amer. Math. Soc., Providence, 2010.
[11] J.A. Ball and D.S. Kaliuzhnyi-Verbovetskyi, Rational Cayley inner Herglotz -- Agler functions:
positive-kernel decompositions and transfer-function realizations. Linear Algebra Appl. 456
(2014), 138 -- 156.
SCHUR -- AGLER AND HERGLOTZ -- AGLER CLASSES
53
[12] J.A. Ball and O.J. Staffans, Conservative state-space realizations of dissipative system be-
haviors, Integral Equations and Operator Theory 54 (2006), 151 -- 213.
[13] J. A. Ball and T. Trent, Unitary colligations, reproducing kernel Hilbert spaces, and
Nevanlinna-Pick interpolation in several variables, J. Funct. Anal. 157 (1998), 1 -- 61.
[14] J. Behrndt, S. Hassi, and H. de Snoo, Functional models for Nevanlinna families, Opuscula
Mathematica 28 (2008) no. 3, 233 -- 245.
[15] J. Behrndt, S. Hassi, and H. de Snoo, Boundary relations, unitary colligations, and functional
models, Complex Analysis and Operator Theory 3 (2009), 57 -- 98.
[16] S. Belyi, S. Hassi, H. de Snoo, and E. Tsekanovskii, A general realization theorem for matrix-
valued Herglotz -- Nevanlinna functions, Linear Algebra and its Applications 419 (2006) no. 2 --
3, 331 -- 358.
[17] M. F. Bessmertnyı. Functions of Several Variables in the Theory of Finite Linear Structures.
Ph. D. Thesis, Kharkov University, Kharkov, 1982. (Russian).
[18] M. F. Bessmertnyı. On realizations of rational matrix functions of several complex variables.
In: Interpolation Theory, Systems Theory and Related Topics: The Harry Dym Anniversary
Volume (D. Alpay, I. Gohberg, and V. Vinnikov, eds.), Oper. Theory Adv. Appl., Vol. 134,
pp. 157 -- 185, Birkhauser-Verlag, Basel, 2002.
[19] M. F. Bessmertnyı. On realizations of rational matrix functions of several complex vari-
ables. II. Translated from the Russian by V. Katsnelson. In: Reproducing Kernel Spaces and
Applications, Oper. Theory Adv. Appl., Vol. 143, pp. 135 -- 146, Birkhauser, Basel, 2003.
[20] M. F. Bessmertnyı. On realizations of rational matrix functions of several variables. III. In:
Current Trends in Operator Theory and Its Applications, Oper. Theory Adv. Appl., Vol. 149,
pp. 133 -- 138, Birkhauser, Basel, 2004.
[21] M. F. Bessmertnyı. Functions of several variables in the theory of finite linear structures. I.
Analysis. In: Operator Theory, Systems Theory and Scattering Theory: Multidimensional
Generalizations, Oper. Theory Adv. Appl. 157, pp. 91 -- 106, Birkhauser, Basel, 2005.
[22] J. Bogn´ar, Indefinite Inner Product Spaces, Springer-Verlag, New York-Heidelberg-Berlin,
1974.
[23] D. Bors and M. Majewski, On the existence of an optimal solution of the Mayer prob-
lem governed by 2D continuous counterpart of the Fornasini-Marchesini model, Multidi-
mens. Syst. Signal Process 24 (2013) no. 4, 657 -- 665.
[24] W.F. Donoghue, Monotone Matrix Functions and Analytic Continuation, Die Grundlehren
der mathematischen Wissenshaften in Einzeldarstellungen Band 207, Springer, New York,
1974.
[25] E. Fornasini and G. Marchesini, Doubly-indexed dynamical systems: state-space models and
structural properties, Math. Systems Theory 12 (1978(, 59 -- 72.
[26] P.A. Fuhrmann, Linear Operators and Systems in Hilbert Space, McGraw-Hill, New York,
New York, 1981.
[27] J.W. Helton, Systems with infinite-dimensional state space: the Hilbert space approach, Re-
cent Trends in System Theory, Proc. IEEE 64 (1976) no. 1, 145 -- 160.
[28] D. S. Kalyuzhnyı-Verbovetzkiı. On the Bessmertnyı class of homogeneous positive holomor-
phic functions of several variables. In: Current Trends in Operator Theory and Its Applica-
tions, Oper. Theory Adv. Appl. 149, pp. 255-289, Birkhauser, Basel, 2004.
[29] J. Malinen, O.J. Staffans and G. Weiss, When is a linear system conservative?
Quart. Appl. Math. 64 (2006) no. 1, 61 -- 91.
[30] V. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge Studies in Ad-
vanced Mathematics 78, Cambridge University Press, 2002.
[31] R.S. Phillips, Dissipative operators and hyperbolic systems of partial differential equations,
Trans. Amer. Math. Soc. 90 (1959), 193 -- 254.
[32] D. Salamon, Infinite dimensional linear systems with unbounded control and observation: a
functional analytic approach, Trans. Amer. Math. Soc. 300 (1987), pp. 383 -- 431.
[33] Y.L. Smuljan, Invariant subspaces of semigroups and the Lax-Phillips scheme, deposited in
VINITI, N 8009-1386, Odessa, 49 pp., 1986.
[34] O.J. Staffans, Passive and conservative continuous-time impedance and scatterings systems.
Part I: well-posed systems, Math. Control Signals Systems 15 (2002), 291 -- 315.
54
J.A. BALL AND D.S. KALIUZHNYI-VERBOVETSKYI
[35] O.J. Staffans, Passive and conservative infinite-dimensional impedance and scattering sys-
tems (from a personal point of view), in Mathematical Systems in Biology, Communica-
tion, Computation and Finance (MTNS2002 Notre Dame, Indiana) (ed. J. Rosenthal and
D.S. Gilliam) pp. 373 -- 414, IMA Volume 314 Springer, 2003.
[36] O.J. Staffans, Well-posed Linear Systems, Encyclopedia of Mathematics and Its Applications
103, Cambridge University Press, 2005.
[37] O.J. Staffans, On scattering passive system nodes and maximal scattering dissipative opera-
tors, Proc. Amer. Math. Soc. 141 no. 4 (2013), 1377 -- 1383.
[38] O.J. Staffans and G. Weiss, A physically motivated class of scattering passive linear systems,
SIAM J. Control Optim. 50 No. 5 (2012), 3083 -- 3112.
[39] G. Weiss, Transfer functions of regular linear systems. Part I: characterizations of regularity,
Trans. Amer. Math. Soc. 3342 (1994), pp. 827 -- 854.
[40] G. Weiss and O.J. Staffans, Maxwell's equations as a scattering passive linear system, SIAM
J. Control Optim. 51 (2013), 3722 -- 3756.
Department of Mathematics, Virginia Tech, Blacksburg, VA, 24061
E-mail address: [email protected]
Department of Mathematics, Drexel University, 3141 Chestnut St., Philadelphia,
PA, 19104
E-mail address: [email protected]
|
1909.06634 | 1 | 1909 | 2019-09-14T17:22:42 | Decay estimates connected with Toeplitzness of composition operators | [
"math.FA"
] | In a higher dimensional version of an earlier conjecture of Nazarov and Shapiro, the truth of which would imply that any composition operator on the second Hardy space is weakly asymptotically Toeplitz, Shayya proved that the arithmetic means of the main diagonal entries of the matrix of coefficients connected with the conjecture, decay like 1/log N. But the proof does not extend to other diagonal entries. We prove that the arithmetic means of the entries along each diagonal decay like 1/(log N loglog N) uniformly. | math.FA | math |
DECAY ESTIMATES CONNECTED WITH
TOEPLITZNESS OF COMPOSITION OPERATORS
Faruk F. Abi-Khuzam
Abstract. Let f ∈ L∞(T d) with kf kL∞ (T d) ≤ 1, ν ∈ Zd, n, k ∈ Z and
put bn,n−k = RE f (x)ne−2πi(n−k)ν ·xdx, E = {x ∈ T d : f (x) = 1}. Shayya
conjectured that, if f (ξ) = 0 for all ξ in a half-space S of lattice points, and
ν ∈ −S, and f (0) 6= 0, then limn→∞ bn,n−k = 0, k ∈ Z. This is a higher
dimensional version of an earlier conjecture of Nazarov and Shapiro, the truth
of which would imply that any composition operator is weakly asymptotically
Toeplitz on the Hardy space H 2. Shayya proved that the arithmetic means
of the main diagonal entries, {bn,n}, decay like {log N }−1, but the proof does
not extend to other diagonal entries. Using an elementary method, we extend
and improve all known results.
In particular, we show that the arithmetic
means of the entries of each diagonal, {bn,n−k }, decay like {log N log log N }−1
uniformly in k ∈ Z.
1. INTRODUCTION
Let φ be a function holomorphic in the open unit disk U of the complex plane
and satisfying the two conditions
φ(0) 6= 0, φ(z) < 1, z ∈ U.
The composition operator Cφ : H(U ) → H(U ), defined via φ by Cφf = f ◦ φ,
restricts to a bounded linear operator on the Hardy space H 2(U ), and so has an
infinite matrix (cα.β) associated to it and defined by
φ(z)αz−β dz
2πiz
.
cα,β =Z∂U
If S is the forward shift operator on H 2(U ), and S∗ is its adjoint, then Cφ is said to
be weakly asymptotically Toeplitz , or WAT, if the sequence {S∗nCφSn} of bounded
linear operators on H 2(U ) converges in the weak operator topology. Equivalently,
Cφ is WAT, if
If E = {z ∈ ∂U : φ(z) = 1}, and we put
lim
n→∞
cn,n−k = 0, k ∈ Z.
bn,n−k =ZE
φ(z)nzk−n dz
2πiz
,
2000 Mathematics Subject Classification. Primary 42B05; Secondary 47B33.
Key words and phrases. Toeplitz operator, weakly asymptotically Toeplitz.
1
2
FARUK F. ABI-KHUZAM
then, clearly, cn,n−k → 0, if and only if bn,n−k → 0, as n → ∞.The WAT conjecture
on the circle, formulated by Nazarov and Shapiro [4], is the statement that
(1.1)
lim
n→∞
bn,n−k = 0, k ∈ Z.
At present, it is not known if (1.1) is true, even when k = 0. But the arithmetic
means are more tame. In [5], it was shown that
(1.2)
lim
N→∞
1
N
bn,n−k2 = 0, k ∈ Z.
N
Xn=1
An infinite series carrying the full sequence {cn,n−k} was considered in [1], where
it was shown, using complex analytic methods, that
cn,n−k2
n
< ∞, k ≥ 0,
∞
Xn=1
from which it follows, through a simple summation by parts argument, that the
arithmetic means of the sequence {cn,n−k2} tend to 0, and also that
∞
Xn=1
1
n(n + 1)
n
Xj=1
cj,j−k2 < ∞, k ≥ 0.
j=1 cj,j−k2 = 0,
The convergence of this later series implies that lim inf n→∞
and this anticipates later work on the arithmetic means of the sequence {bn,n−k2}.
was established as part of a study of these prob-
lems undertaken recently by Shayya [6] in the setting of functions defined on the
d-dimensional torus T d. He proved the existence of a constant C, essentially de-
pendent on φ(0), such that
The convergence ofP∞
n Pn
bn,n−k2
log n
n=1
n
(1.3)
1
N
bn,n2 ≤
C
log N
π
, M ≥ 1, N ≥ 2,
M+N
Xn=M
thereby obtaining a quantitative strengthening of (1.2) , in the case k = 0. Shayya's
method uses a positive Borel measure µ such that
bn,n2 =Z 1/2
−1/2
e−2πinθdµ(θ),
bn,n2
n
n=1
. This measure, obtained as
Shayya's result (1.3) tells us that the lim supn→∞
an application of Herglotz's theorem, is only available when k = 0.
along with the convergence of the series P∞
n Pn
j=1 bn,n2 < ∞, while,
j=1 bn,n2 = 0. So if the limit exists,
by the argument above, lim inf n→∞
it would have to be 0, and a natural question would then be to find the rate of
decay to 0.
In the present work we consider the more general problem of the
arithmetic means of bn,n−k2, k ∈ Z. We obtain, uniformly in k ∈ Z, a quantitative
strengthening of (1.2). In addition, we show that these means actually decay much
faster than suggested by (1.3) and we obtain, among other results, the inequality
n Pn
log n
log n
(1.4)
1
N + 3
bn,n−k2 ≤
C/(1 − 1/ log 3)
log(N + 3) log[ log(N +4)
log 4
]
, k ∈ Z,
for all M ≥ 1, and N ≥ 1, where the constant C depends only on φ(0).
M+N
Xn=M
TOEPLITZNESS OF COMPOSITION OPERATORS
3
2. PRELIMINARIES AND STATEMENT OF RESULTS
We shall consider the problem of the decay of the arithmetic means in the
setting of functions defined on T d, as formulated in [6]. Since we shall be working
on the coefficients bn,n−k2 with k ∈ Z, we will not have available a measure as
in [6], and so we shall proceed along different, rather elementary, lines. However,
there will be two important common ingredients between the present paper and [6],
namely the use of Helson's and Lowdenslager's generalization of Szego's theorem
[3], and the restriction to the set where the boundary values of the given function
have unit modulus. We shall obtain a bound on the sum of the series
∞
Xm=−∞
m6=N
bm,m−k2
m − N
,
where k, N ∈ Z. The usefulness of this stems from the fact that the bound will be
independent of k and N , thereby making it possible to obtain uniform bounds on
the arithmetic means of {bm,m−k2}.
We now give a summary, taken mostly from [6], of material needed in connection
with the Helson-Lowdenslager result.
Definition 1. A set S ⊂ Zd is said to be a half-space of lattice points in Rd if
(i) 0 6∈ S
(ii) ξ 6= 0 implies ξ ∈ S or −ξ ∈ S
(iii) ξ, ξ′ ∈ S implies ξ + ξ′ ∈ S.
Clearly, if S is a half-space of lattice points in Rd, then so is its reflection in
the origin. i.e. −S = {−ξ : ξ ∈ S}.
Also [6], if S1 and S2 are half-spaces of lattice points in Rd1 and Rd2 respectively,
then S1 × Rd2 ∪ Rd1 × S2 contains a half-space of lattice points in Rd1 × Rd2 (e.g.,
S1 × Rd2 ∪ {0} × S2).
We shall need the following generalization of Szego's theorem [3], stated as
corollary A in [6].
Lemma 1. Suppose f ∈ L2(T d), S is a half-space of lattice points in Rd, and
f (ξ) = 0 for all ξ ∈ S. then
log f (x)dx ≥ log f (0).
ZT d
As a simple application of this lemma, suppose S is a half-space of lattice points
in Rd, f ∈ L∞(T d), kf kL∞(T d) ≤ 1, and f (ξ) = 0 for all ξ ∈ S. For 0 < r < 1, and
ν ∈ −S, define F : T d × T d → C by
F (x, y) = e2πi(ν,−ν)·(x,y) − rf (x)f (y).
Then, [6], F (ξ, η) = 0 for all (ξ, η) ∈ S × Rd ∪ Rd × (−S), and so for all (ξ, η) in a
half-space of lattice points in Rd × Rd. Lemma 1 applies and we conclude that
ZT dZT d
log F (x, y)dxdy ≥ log F (0, 0) = log r f (0)2.
4
FARUK F. ABI-KHUZAM
Also, using the equality log u = 2 log+ u − log u, we have, for any measur-
able subset of T d,
(2.1)
log F (x, y)dxdy
ZEZE
= 2ZT dZT d
log F (x, y)dxdy ≤ZT dZT d
log+ F (x, y)dxdy −ZT dZT d
≤ 2 log(1 + r) − log r f (0)2 ≤ log
log F (x, y)dxdy
4
r f (0)2
.
We shall also need two sequences of functions logj x and Lq(x). The first is the
iterated logarithmic function, and the second is a product of different iterates:
If j ≥ 1 is an integer, we define the function logj as follows:
log1 x = log x, logj x = log ◦ logj−1 x.
We will need to use the iterated logarithm on the domain where it is well-
defined and positive, and this, of course, is trivial to determine.In terms of the
iterated logarithm we define the function Lq by
q
Lq(x) =
logj x, q ≥ 1.
Yj=1
Also, for a positive differentiable function g, we define [2] the function a(.; g) by
a(x; g) = xg′(x)/g(x), and note that
a(x; L1) =
log′
q+1(x) =
, a(x; Lq) =
1
log x
1
.
xLq(x)
1
log x
1 +
q
Xj=2
1
logj x
, q ≥ 2.
Clearly, a(x; Lq) decreases monotonically to 0 as x → ∞, and so there exist positive
constants αq and γq , depending only on the integer q, such that 0 < αq < 1, and
logq+1 x > 0, 0 < a(x; Lq) < αq, x ≥ γq.
Theorem 1. Suppose f ∈ L∞(T d) with kf kL∞(T d) ≤ 1, and ν ∈ Zd. For
n, k ∈ Z, define
bn,n−k =ZE
f (x)ne−2πi(n−k)ν·xdx,
where E = {x ∈ T d : f (x) = 1}.When f (0) 6= 0, put C = log
16
f (0)4 .
(i) If f (ξ) = 0 for all ξ in a half-space S of lattice points, ν ∈ −S, f (0) 6= 0,
and N ∈ Z, then
(2.2)
∞
Xm=−∞
m6=N
bm,m−k2
m − N
≤ log
16
f (0)4
, k ∈ Z.
(ii) If f (ξ) = 0 for all ξ in a half-space S of lattice points, ν ∈ −S, f (0) 6= 0,
M ≥ 1, and p ≥ 1, then
(2.3)
1
p + 1
M+p
Xm=M
bm,m−k2 ≤
C
log(p + 1)
, k ∈ Z.
TOEPLITZNESS OF COMPOSITION OPERATORS
5
(iii) Suppose g is positive, and 0 < a(x; g) < α, for all x ≥ γ ≥ 1. If f (ξ) = 0
for all ξ in a half-space S of lattice points, ν ∈ −S, f (0) 6= 0, M ≥ 1, and p ≥ 1,
then
(2.4)
(cid:18)Z p+1
1
dt
tg(t + γ)(cid:19)
M+p
Xm=M
bm,m−k2 ≤
C
1 − α
·
p + γ
g(p + γ)
, k ∈ Z.
(iv) Suppose Lq, αq, and γq are as defined above. If f (ξ) = 0 for all ξ in a
half-space S of lattice points, ν ∈ −S, f (0) 6= 0, M ≥ 1, and p ≥ 1, then
(2.5)
bm,m−k2 <
C/(1 − αq)
Lq(p + γq)[logq+1(p + 1 + γq) − logq+1(1 + γq)]
, k ∈ Z.
1
p + γq
M+p
Xm=M
Also, for a fixed q ≥ 1, there is a constant C′ depending only on q and C, such
that, for all large p,
1
p
M+p
Xm=M
uniformly in k ∈ Z.
bm,m−k2 <
C′
Lq+1(p)
=
C′
log p · log log p · · · logq+1 p
,
The inequality in part (ii) represents a sharpening, and an extension of (1.3).
The inequality in part (iv) shows that, uniformly in k ∈ Z, the decay of the arith-
metic means of the sequence {bm,m−k2} is faster than that of L−1
q+1 , for any q ≥ 1.
3. PROOF OF THEOREM 1
The definition of bn,n−k gives us that
bn.n−k2 =ZEZE(cid:16)f (x)f (y)e−2πiν·(x−y)(cid:17)n
and, for N, k ∈ Z, we have, since f (x) = 1 for x ∈ E,
.e2πikν·(x−y)dxdy,
bn+N,n+N −k2 + b−n+N,−n+N −k2
= 2ZEZE(cid:16)f (x)f (y)e−2πiν·(x−y)(cid:17)N
.e2πikν·(x−y) Re(cid:16)f (x)f (y)e−2πiν·(x−y)(cid:17)n
Multiplying by rn, 0 < r < 1, dividing by n and summing over n ≥ 1, we obtain
dxdy.
bn+N,n+N −k2 + b−n+N,−n+N −k2
n
rn
∞
Xn=1
= 2ZEZE(cid:16)f (x)f (y)e−2πiν·(x−y)(cid:17)N
where F (x, y) = e2πiν·(x−y) − rf (x)f (y) .
integral of log F (x, y) we get that
.e2πikν·(x−y). log
1
F (x, y)
dxdy,
In view of the estimate (2.1) on the
bn+N,n+N −k2 + b−n+N,−n+N −k2
n
rn ≤ log
16
r2 f (0)4
,
∞
Xn=1
and letting r → 1−, we obtain
bn+N,n+N −k2 + b−n+N,−n+N −k2
n
≤ C.
∞
Xn=1
6
FARUK F. ABI-KHUZAM
Reindexing we arrive at
∞
Xm6=N
m=−∞
bm,m−k2
m − N
≤ C, k ∈ Z,
and this completes the proof of part (i).
If M ≥ 1, p ≥ 1, then, for a given N , the partial sums in (2.2) for M ≤ m ≤
M + p, m 6= N , are also bounded by the constant C uniformly in k, N ∈ Z, and we
may sum them over N from M to M + p to get, after an interchange in the order
of summation
(3.1)
M+p
M+p
Xm=M
XN =M
N 6=m
bm,m−k2
m − N
≤
M+p
XN =M
C = C(p + 1), k ∈ Z.
N =M
N 6=m
On the left-hand side of (3.1) we use the inequality PM+p
m − N
bm,m−k 2!
j
≥ log(p + 1) M+p
Xm=M
≥ M+p
Xm=M
XN =M
Xj=1
N 6=m
M+p
so that
1
1
p
M+p
Xm=M
bm,m−k 2!
Xm=M
log(p + 1)
M+p
and the inequality in (2.3) follows.
1
m−N ≥Pp
j=1
1
j :
bm,m−k 2! ,
bm,m−k 2 ≤ C(p + 1), k ∈ Z,
For proof of part (iii), first divide both sides of (2.2) by g(N − M + γ), N ≥
M ≥ 1,and proceed exactly as in the proof of part (ii). In the resulting inequality
the left-hand side is
M+p
(
Xm=M
bm,m−k 2)
M+p
XN =M
N 6=m
1
m − N g(N − M + γ)
M+p
≥ (
Xm=M
bm,m−k 2)
1
jg(j + γ)
p
Xj=1
≥
M+p
Xm=M
while the right-hand side is
bm,m−k 2Z p+1
1
dt
tg(t + γ)
,
C
M+p
XN =M
1
g(N − M + γ)
< C{
1
g(γ)
+Z p
0
dt
g(t + γ)
}.
The conditions on a(x; g), along with the positivity of g, imply that both functions
g and x/g(x) are increasing on [γ, ∞). An application of the Cauchy mean-value
theorem gives a ξ ∈ (γ, x) such that
dt
g(t)
γ
R x
x
g(x) − γ
g(γ)
=
1
1 − a(ξ; g)
<
1
1 − α
.
TOEPLITZNESS OF COMPOSITION OPERATORS
7
Since γ ≥ 1, 0 < α < 1,we have
−(cid:18) 1
1 − α(cid:19) γ
g(γ)
+
1
g(γ)
< 0,
and it follows that
(3.2)
1
g(γ)
+Z p
0
dt
g(t + γ)
<
1
1 − α
·
p + γ
g(p + γ)
,
which gives the required bound and the inequality in (2.4) follows.
If q ≥ 1, and Lq is the function defined in section 2, and αq and γq are the
positive constants associated with it, we may apply the result in part (iii) with
g(x) = Lq(x). This gives us
(cid:18)Z p+1
1
dt
tLq(t + γq)(cid:19)
M+p
Xm=M
bm,m−k2 ≤
C
1 − αq
·
p + γq
Lq(p + γq)
, k ∈ Z.
For the integral on the left-hand side we have
Z p+1
1
dt
tLq(t + γq)
≥Z p+1
1
dt
(t + γq)Lq(t + γq)
= logq+1(p + 1 + γq) − logq+1(1 + γq),
and it follows that
1
p + γq
M+p
Xm=M
for all k ∈ Z.
bm,m−k2 ≤
C/(1 − αq)
Lq(p + γq)[logq+1(p + 1 + γq) − logq+1(1 + γq)]
,
If q ≥ 1 is fixed, since Lq(p + γq)[logq+1(p + 1 + γq) − logq+1(1 + γq)] ∼ Lq+1(p)
as p → ∞, there is a constant C′depending only on q and C, such that, for all
large p,
1
p
M+p
Xm=M
bm,m−k2 <
C′
Lq+1(p)
,
uniformly in k ∈ Z.
In the special case where q = 1, Lq(x) = log x and we may take γ = 3, and
log 3 . Then Theorem 1 gives us that
α = 1
1
p + 3
M+p
Xm=M
bm,m−k2 ≤
C/(1 − 1/ log 3)
log(p + 3)[log log(p + 4) − log log 4]
, k ∈ Z.
There are some interesting questions that present themselves in connection with
the study of the problems considered in this note. One question concerns the rate
. Even in the case k = 0, this
of decay of the tail end of the series P∞
is not known. Another question concerns the series P∞
n
known if this series is convergent, even in the case q = 1, and k = 0.
n=1 Lq(n) bn,n−k2
. It is not
bn,n−k2
n=1
n
8
FARUK F. ABI-KHUZAM
4. REFERENCES
[1] F. ABI KHUZAM AND B. SHAYYA, A remark on the WAT conjecture,
preprint, 2009.
[2]W. HAYMAN, A generalisation of Stirling's formula, J. Reine. Angew.
Math. 196 (1956), 67-95.
[3]H. HELSON AND D. LOWDENSLAGER, Prediction theory and Fourier
series in several variables, Acta Math. 99 (1958), 165-202.
[4] F. NAZAROV AND J. SHAPIRO, On the Toeplitzness of composition op-
erators, Complex Var. Elliptic Wqu. 52 (2007), 193-210.
[5] J. SHAPIRO, Every composition operator is ( mean) asymptotically Toeplitz,
J. Math. Anal. Appl. 333 (2007), 523-529.
[6] B.SHAYYA, The WAT conjecture on the Torus, Proc. Amer. Math. Soc.
139 ( 2011), 3633-3643.
Department of Mathematics, American University of Beirut, Beirut Lebanon
E-mail address: [email protected]
|
1711.00091 | 3 | 1711 | 2019-10-30T08:01:27 | The invertibility of U-fusion cross Gram matrices of operators | [
"math.FA"
] | For applications like the numerical solution of physical equations a discretization scheme for operators is necessary. Recently frames have been used for such an operator representation. In this paper, we apply fusion frames for this task. We interpret the operator representation using fusion frames as a generalization of fusion Gram matrices. We present the basic definition of $U$-fusion cross Gram matrices of operators for a bounded operator $U$. We give sufficient conditions for their (pseudo-)invertibility and present explicit formulas for the inverse. In particular, we characterize fusion Riesz bases and fusion orthonormal bases by such matrices. Finally, we look at which perturbations of fusion Bessel sequences preserve the invertibility of the fusion Gram matrix of operators. | math.FA | math |
The invertibility of U -fusion cross Gram matrices of operators
Mitra Shamsabadia, Ali Akbar Arefijamaala, Peter Balazsb,∗
aDepartment of Mathematics and Computer Sciences, Hakim Sabzevari University, Sabzevar, Iran
bAcoustics Research Institute, Austrian Academy of Sciences, Vienna, Austria
Abstract
For applications like the numerical solution of physical equations a discretization scheme for operators
is necessary. Recently frames have been used for such an operator representation.
In this paper, we
apply fusion frames for this task. We interpret the operator representation using fusion frames as a
generalization of fusion Gram matrices. We present the basic definition of U -fusion cross Gram matrices
of operators for a bounded operator U . We give sufficient conditions for their (pseudo-)invertibility and
present explicit formulas for the inverse.
In particular, we characterize fusion Riesz bases and fusion
orthonormal bases by such matrices. Finally, we look at which perturbations of fusion Bessel sequences
preserve the invertibility of the fusion Gram matrix of operators.
1. Introduction and motivation
For the representation (and modification) of functions a standard approach is using orthonormal bases
(ONBs). It can be hard to find a 'good' orthonormal basis, in the sense that it sometimes cannot fulfill
given properties, as formally expressed e.g.
in the Balian-Low theorem [34]. For solving this problem,
frames were introduced by Duffin and Schaeffer [28] and widely developed by many authors [17, 22, 26, 29].
In recent years, frames have been the focus of active research, both in theory [2, 18, 30] and applications
[11, 19, 24]. Also, several generalizations have been investigated, e.g. [1, 3, 4, 44, 47], among them fusion
frames [20, 21, 32], which are the topic of this paper.
For a numerical treatment of operator equations, used for example for solving integral equations in
acoustics [38], the involved operators have to be discretized to be handled numerically. The (Petrov-
)Galerkin approach [31] is a particular and well-known way for this discretization. For an operator O,
the matrix M defined by Mk,l = hOψl, φki is called the matrix corresponding to the operator O, or the
system matrix. The standard way for the discretization of operators is using bases [33], but recently the
general theory for frames has been developed [8, 10]. Frames were also used in numerics [35], in particular
in an adaptive approach [25, 45]. In [12, 13] sufficient and necessary conditions of the invertibility of such
matrices is investigated. Note that the system matrix of the identity is the cross Gram matrix of the
∗Corresponding author
Email addresses: [email protected] (Mitra Shamsabadi), [email protected] (Peter Balazs)
Preprint submitted to Elsevier
October 31, 2019
two sequences {ψk}k∈I and {φk}k∈I . Therefore, in [15] the concept of matrix representation of operators
using frames is reinterpreted as a generalization of the Gram matrix to investigate the inverses. As
the concept of domain decomposition is a particularly relevant topic in this field, the extension of the
approach to operator representations to fusion frame is very useful [40].
In this paper, we therefore look at those U -fusion cross Gram matrices. In particular, we investi-
gate the (pseudo-)invertibility of U -fusion Gram matrices of operators. In Section 2, we review basic
notations and preliminaries.
In Section 3, we give necessary as well as sufficient conditions for the
(pseudo-)invertibility of the U -cross Gram matrices, characterize fusion orthonormal bases and fusion
Riesz bases by those properties and give formulas for the (pseudo-)inverses. Finally, in Section 4, some
stability results are discussed.
2. Preliminaries and Notations
Throughout this paper, H is a separable Hilbert space, I a countable index set and IH the identity
operator on H and {ei}i∈I an orthonormal basis for H. The orthogonal projection on a subspace V ⊆ H
is denoted by πV . We will denote the set of all linear and bounded operators between Hilbert spaces H1
and H2 by B(H1,H2) and for H1 = H2 = H, it is represented by B(H). We denote the range and the null
space of an operator U by ran (U ) and ker (U ), respectively. For a closed range operator U ∈ B(H1,H2),
the pseudo-inverse of U is defined the unique operator U† ∈ B(H2,H1) satisfying
ker(cid:0)U†(cid:1) = ran (U )⊥ , ran(cid:0)U†(cid:1) = ker (U )⊥ , and U U†U = U.
The operator U has closed range if and only if U∗ has closed range and (U∗)† = (cid:0)U†(cid:1)∗, see e.g.
Lemma 2.4.1, Lemma 2.5.2].
[22,
For 0 < p < ∞, the space of all compact operators (which are the closure of finite-rank operators and
denoted by K(H)) T on H such that its singular values {λn}n∈I belonging to ℓp is called the Schatten
p-class of H. It is denoted by Sp(H) which is a Banach space with the norm
kTkp = Xn
p
λnp! 1
.
(2.1)
An operator T ∈ B(H) is called trace class if trace(T ) := Pi∈I hT ei, eii < ∞, for every orthonormal
2 =Pi∈I kT eik2 < ∞. It
basis {ei}i∈I for H. It is shown that T is trace class if and only if T ∈ S1(H). Also, the class of Hilbert-
Schmidt operators of H is denoted by S2(H), and T ∈ S2(H) if and only if kTk2
is well-known that K(H) and Sp(H) are two sided ∗-ideal of B(H), that is, a Banach algebra under the
norm (2.1) and the finite rank operators are dense in (Sp(H),k.kp). Moreover, for T ∈ Sp(H), one has
kTkp = kT ∗kp and kTk ≤ kTkp. If S1 ∈ B(H,H1) and S2 ∈ B(H2,H), then kS1Tkp ≤ kS1kkTkp and
kT S2kp ≤ kS2kkTkp. For more information about these operators, see [33, 41, 42, 48].
2
2.1. Fusion frames
We now review some definitions and primary results of fusion frames. For more information see
[20, 21, 32].
For each sequence {Wi}i∈I of closed subspaces in H, the space
Xi∈IM Wi!ℓ2
=({fi}i∈I : fi ∈ Wi,Xi∈I
(cid:28){fi}i∈I ,{gi}i∈I(cid:29) =Xi∈I
hfi, gii,
kfik2 < ∞) ,
with the inner product
is a Hilbert space.
We now give the central definition of fusion frames:
Definition 2.1. Let {Wi}i∈I be a family of closed subspaces of H and {ωi}i∈I be a family of weights,
i.e. ωi > 0, i ∈ I. The sequence {(Wi, ωi)}i∈I is called a fusion frame for H if there exist constants
0 < AW ≤ BW < ∞ such that
AWkfk2 ≤Xi∈I
i kπWi fk2 ≤ BWkfk2,
ω2
(f ∈ H).
The constants AW and BW are called fusion frame bounds.
If we have the upper bound, we call
{(Wi, ωi)}i∈I a Bessel fusion sequence. A fusion frame is called tight, if AW and BW can be chosen
to be equal, and Parseval if AW = BW = 1. If ωi = ω for all i ∈ I, the collection {(Wi, ωi)}i∈I is called
it is called a Riesz decomposition of H if for every f ∈ H there is a unique choice of fi ∈ Wi such that
ω-uniform. A fusion frame {(Wi, ωi)}i∈I is said to be a fusion orthonormal basis if H =Li∈I Wi and
f =Pi∈I fi. A family of subspaces is called complete if span (Wi) = H.
It is clear that every fusion orthonormal basis is a Riesz decomposition of H. Moreover, a family
{Wi}i∈I of closed subspaces of H is a fusion orthonormal basis if and only if it is a 1-uniform Parseval
fusion frame [20].
Furthermore, the synthesis operator TW : (Pi∈IL Wi)ℓ2 → H for a Bessel fusion sequence W =
{(Wi, ωi)}i∈I is defined by
TW ({fi}i∈I ) =Xi∈I
ωifi.
The adjoint operator T ∗W : H → (Pi∈IL Wi)ℓ2 which is called the analysis operator is given by
T ∗W f = {ωiπWi f}i∈I ,
(f ∈ H).
Both are bounded by √BW .
If W = {(Wi, ωi)}i∈I is a fusion frame, the fusion frame operator SW : H → H, which is defined by
i πWi f , is bounded (with bound BW ), invertible and positive [20, 32].
SW f = TW T ∗W f =Pi∈I ω2
3
Every Bessel fusion sequence V = {(Vi, υi)}i∈I is called a Gavrut¸a-dual of W = {(Wi, ωi)}i∈I , if
f =Xi∈I
ωiυiπVi S−1
W πWi f,
(f ∈ H),
TV φV W T ∗W = IH,
for more details see [32]. From here on, for simplicity we say dual instead of Gavrut¸a-dual. The sequence
W Wi, ωi(cid:1)(cid:9)i∈I is also a fusion frame for H and a dual of W, called the canonical
dual of W [20, 32]. Rephrasing that, a Bessel fusion sequence V = {(Vi, υi)}i∈I is a dual of a fusion frame
W = {(Wi, ωi)}i∈I if and only if
of subspaces fW :=(cid:8)(cid:0)S−1
where the bounded operator φV W : (Pi∈IL Wi)ℓ2 → (Pi∈IL Vi)ℓ2 is given by
and kφV Wk ≤(cid:13)(cid:13)S−1
W (cid:13)(cid:13). Also, V is called a pseudo-dual of W if TV φV W T ∗W is an invertble operator, see [23]
for discrete case. Another approach to duality [36, 37] uses any bounded operator O : (Pi∈IL Wi)ℓ2 →
(Pi∈IL Vi)ℓ2 . Starting with two fusion frames the duality is defined analogously to (2.2), i.e. TV O T ∗W =
IH. We stick to the Gavrut¸a duals, but all results herein can be adapted to this other definition of duality.
Let {Wi}i∈I be a family of closed subspaces of H and {ωi}i∈I a family of weights. We say that
{(Wi, ωi)}i∈I is a fusion Riesz basis for H if span (Wi) = H and there exist constants 0 < C ≤ D < ∞
such that for each finite subset J ⊆ I and all (fj ∈ Wj, j ∈ J) we have
φV W ({fi}i∈I ) = {πVi S−1
W fi}i∈I
(2.2)
(2.3)
Remark 2.2. Note that weights are not included in the definition of the Riesz decomposition. Obviously,
we have that f =Pi∈I fi with unique fis, if and only if f =Pi∈I wifi with the same uniqueness.
The next theorem explores fusion Riesz bases with respect to local frames and their operators.
Theorem 2.3. [20] Let {Wi}i∈I be a fusion frame for H and {eij}j∈Ji be an orthonormal basis for Wi,
for each i ∈ I. Then the following conditions are equivalent:
(1) {Wi}i∈I is a Riesz decomposition of H.
(2) The synthesis operator TW is one to one.
(3) The analysis operator T ∗W is onto.
(4) {Wi}i∈I is a fusion Riesz basis for H.
(5) {eij}i∈I,j∈Ji is a Riesz basis for H.
For some results we need a version of Theorem 2.3 formulated for general family of subspaces:
Proposition 2.4. Let W = {(Wi, ωi)}i∈I be a family of closed subspaces and {fij}j∈Ji be a Riesz basis
for Wi, for each i ∈ I with bounds Ai and Bi, respectively, such that
0 < inf
i∈I
Then the following conditions are equivalent:
Ai ≤ sup
i∈I
Bi < ∞.
4
CXj∈J
2
kfjk2 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj∈J
ωjfj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ DXj∈J
kfjk2.
(2.4)
(1) W is a Riesz decomposition of H.
(2) The synthesis operator TW is bounded and bijective.
(3) The analysis operator T ∗W is bounded and bijective.
(4) W is a fusion Riesz basis for H.
(5) {ωifij}i∈I,j∈Ji is a Riesz basis for H.
Proof. (1) ⇔ (2) ⇔ (3) by Theorem 2.3 for any family of closed subspaces (as those conditions imply a
fusion frame property).
(4) ⇔ (5) by relating the inequality (2.4) for fusion Riesz basis W to the one for Riesz bases
{wifij}i∈I,j∈Ji see e.g. [22, Theorem 3.3.7].
(4) ⇔ (2) The equation (2.4) is equivalent to TW being injective, having closed range and being
bounded. By the completeness we have (4) ⇒ (2). By the surjectivity of TW we have completeness and
so (2) ⇒ (4).
The following characterizations of fusion Riesz bases will be used frequently in this note, which is a
generalization of [43, Theorem 2.2] to non-uniform fusion frames.
Proposition 2.5. Let W = {(Wi, ωi)}i∈I be a fusion frame in H. Then the following are equivalent:
(1) W is a fusion Riesz basis.
(2) S−1
(3) ω2
W Wi ⊥ Wj for all i, j ∈ I, i 6= j.
i πWi S−1
W πWj = δijπWj for all i, j ∈ I.
Proof. (1) ⇒ (2) was proved in [20, Proposition 4.3].
(2) ⇒ (1) Suppose that {eij}j∈Ji is an orthonormal basis for Wi, for all i ∈ I. Then for any f ∈ H we
have
Xi∈I Xj∈Ji
hf, ωieiji2 =Xi∈I
i kπWi fk2 .
ω2
It easily follows that {eij}i∈I,j∈Ji is a weighted frame [9] (with weights ωi > 0) for H with the frame
operator SW . Moreover, by Proposition 2.4 it is enough to show that {ωieij}i∈I,j∈Ji is a Riesz basis
for H or equivalently, that the sequences {ωieij}i∈I,j∈Ji and {S−1
W ωieij}i∈I,j∈Ji are biorthogonal. This
immediately follows from the reconstruction formula
f =Xi∈I Xj∈Ji
hf, S−1
W ωieijiωieij,
(f ∈ H).
(2) ⇒ (3) Suppose that {(Wi, ωi)}i∈I is a fusion frame in H and f, g ∈ H. By (2) we obtain
hπWi S−1
W πWj f, gi = hS−1
W πWj f, πWi gi = 0
(i 6= j).
In particular, suppose {eij}j∈Ji is an orthonormal basis for Wi, then we know by the argument above that
the sequence {ωieij}i∈I,j∈Ji is a Riesz basis for H with the frame operator SW . Hence, {S−1/2
W ωieij}i∈I,j∈Ji
is an orthonormal basis for H. Let f, g ∈ H and take
πWi f = Xj∈Ji
cij eij,
dijeij ,
πWi g = Xj∈Ji
5
for all i ∈ I, for some {cij}j∈Ji and {dij}j∈Ji in ℓ2. Then, for all i ∈ I we have
i πWi S−1
(cid:10)ω2
W πWi f, g(cid:11) = DS−1/2
= hXj∈Ji
= Xj,k∈Ji
= *Xj∈Ji
W ωiπWi f, S−1/2
W ωiπWi gE
cij S−1/2
W ωieiki
dikS−1/2
W ωieij,Xk∈Ji
cijdikDS−1/2
cijeij,Xk∈Ji
W ωieij, S−1/2
W ωieikE
dikeik+ = hπWi f, gi .
So, ω2
W πWi = πWi .
i πWi S−1
(3) ⇒ (2) Let f ∈ Wi and g ∈ Wj , where i 6= j. Then
hS−1
W f, gi = hS−1
W πWi f, πWj gi
= hπWj S−1
W πWi f, gi = 0.
From the Proposition 2.5, it easily follows that for a fusion Riesz basis W
1
W πWi fi}i∈I = {
ω2
i
i fi is its bounded inverse.
φW W (fi) = {πWi S−1
W fi}i∈I = {πWi S−1
Therefore, it is bounded and the operator fi 7→ w2
fi}i∈I .
(2.5)
All items in Proposition 2.5 include a dependency on weights, for (2) the weight is included in the
definition of SW . So, the question arises how dependent the Riesz property is on the considered weights.
It is worthwhile to mention that, by the fusion frame definition, the family of weights belongs to ℓ∞+
assuming that the subspaces are non-empty, see [37, Remark 2.4]. Moreover, if W = {(Wi, wi)}i∈I is a
fusion Riesz basis, then (2.4) shows that
(i ∈ I).
Using Proposition 2.4 the following lemma immediately follows.
Lemma 2.6. Let W = {(Wi, ωi)}i∈I and V = {(Wi, vi)}i∈I be a family of subspaces in H with different
weights. Then W is a fusion Riesz basis if and only if V is a fusion Riesz basis.
√C ≤ wi ≤
√D,
This could also be seen as direct consequence of Proposition 2.4, as the Riesz decomposition property,
e.g. item (1) is independent of the weight.
In the sequel, for a given fusion Riesz basis W = {(Wi, wi)}i∈I , we denote by W ′ the 1-uniform family
of subspaces {(Wi, 1)}i∈I.
We will use the following criterion for the invertibility of operators.
Proposition 2.7. [33] Let F : H → H be invertible on H. Suppose that G : H → H is a bounded
operator and kGh − F hk ≤ υkhk for all h ∈ H, where υ ∈ [0,
). Then G is invertible on H and
kF −1k
1
G−1 =P∞k=0[F −1(F − G)]kF −1.
6
3. U -fusion cross Gram matrix of operators
In this section, we extend the notion of cross Gram matrices [15] to fusion frames and discuss on their
invertibility.
We interpret the representation of operators using fusion frames [14] as a generalization of the Gram
matrix of operators:
Definition 3.1. Let W = {(Wi, ωi)}i∈I be a Bessel fusion sequence for H and V = {(Vi, υi)}i∈I a fusion
frame for H. For U ∈ B(H), the matrix operator GU,W,V :(cid:0)Pi∈IL Wi(cid:1)ℓ2 →(cid:0)Pi∈IL Wi(cid:1)ℓ2 given by
GU,W,V = φW V T ∗V U TW ,
is called the U -fusion cross Gram matrix. If U = IH, it is called fusion cross Gram matrix and denoted
by GW,V . We use GW for GW,W ; the so called fusion Gram matrix.
Note that
[GU,W,V (fi)]j ="φW V (vkπVk UXi
wifi)k∈I#
j∈I
=Xi
wivjπWj S−1
V πVj U
Bj,i
{z
fi =Xi
Bj,ifi,
}
where Bj,i : Wi → Wj . Therefore GU,W,V is a block-matrix of operators [6]1, which motivates the name
(cross-)Gram matrix.
Clearly, using (2.3), U -fusion cross Gram matrices are well-defined and
kGU,W,V k = kφW V T ∗V U TWk
≤ kφW V kkT ∗V k kUk kTWk
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13)kUkpBW BV
√BW BV
kUk.
≤
AV
We have chosen to use GU,W,V = φW V T ∗V U TW instead of the 'naive' GU,W,V = T ∗V U TW . The reason
for that is that, by this definition, GU,W,V maps Pi∈IL Wi into itself and is a projection as in the Hilbert
space case (albeit an oblique one, see Remark 3.2).
We can represent an operator U ∈ B(H) from its U -fusion cross Gram matrix. Suppose W is a dual
fusion frame of V , then
TWGU,W,V T ∗W S−1
W = TW φW V T ∗V U TW T ∗W S−1
W = U.
(3.1)
From the ideal property of Sp(H) in B(H) it follows that if U is compact, trace class and Hilbert Schmidt,
so is GU,W,V . By using (3.1) we can deduce the converse with some assumptions. Moreover,
φV W G∗U,W,V = GU ∗,V,W φ∗W V .
1This could be called a generalized subband matrix, motivated by system identification applications [39].
7
Remark 3.2. Note that as in the discrete Hilbert space frame case GV,W = G2
V,W and so this an oblique
projection whenever V is a dual fusion frame of W . Also TV GV,W = TV , and GV,W (φV W T ∗W ) = φV W T ∗W ,
but in general the operator is not self-adjoint. Using the above identities and the definition of GV,W we
achieve
ker (GV,W ) = ker (TV )
and
ran (GV,W ) =(cid:8){πVi S−1
W fi}i∈I :
{fi}i∈I ∈ ran (T ∗W )(cid:9) .
In the next result, we are going to characterize Gram matrices of fusion orthonormal bases.
Proposition 3.3. Let W = {(Wi, ωi)}i∈I be a fusion Riesz basis. The following are equivalent.
(1) W is a fusion orthonormal basis.
(2) GW = IPi∈I L Wi .
(3) GW ′ = IPi∈I L W ′
i .
Proof. (1) ⇒ (2) Assume that W is a fusion orthonormal basis. Applying (2.5) for all f = {fi}i∈I ∈
Pi∈IL Wi we have
(2) ⇒ (1) By Proposition 2.5 we obtain
SW ′ = SW ′ TW T −1
W
GW f = φW W T ∗W TW f
1
w2
i
= {πWi fi}i∈I = f .
wjfji∈I
wiπWiXj∈I
wj fji∈I
wiπWiXj∈I
= φW W
=
= Xi∈I
= Xi∈I
= TW φW W T ∗W TW T −1
W
= TWGW T −1
πWi TW T −1
W
W = IH.
w2
i πWi S−1
W πWi TW T −1
W
W ′ is a fusion orthonormal basis.
(1) ⇒ (3) It follows from (1) ⇒ (2) and the fact that W is a fusion orthonormal basis if and only if
(3) ⇒ (1) It is enough to show that W ′ is a Parseval fusion frame by Proposition 3.23 of [20].
S2
W ′ = TW ′ T ∗W ′ TW ′T ∗W ′
Now, the invertibility of SW ′ implies SW ′ = IH.
= TW ′GW ′ T ∗W ′ = SW ′ .
8
3.1. Invertibility of U -cross Gram matrices
We now discuss the relationship between the invertibility of Gram matrices and their associated
operators.
Proposition 3.4. Let W = {(Wi, ωi)}i∈I be a fusion frame in H and U ∈ B(H). The following are
equivalent:
(1) W is a fusion Riesz basis and U is invertible.
(2) GU,W,W is invertible.
(3) GU,W,W is onto.
(4) GU,W,W is one to one.
(5) GU,fW ,W is invertible.
(6) GU,fW ,W is onto.
(7) GU,fW ,W is one to one.
Proof. (1) ⇒ (2) is trivial by the invertibility of φW W , see (2.5).
is a fusion Riesz basis by Theorem 2.3. So, φW W is invertible and then
(2) ⇒ (1) By using the invertibility GU,W,W = φW W T ∗W U TW , the operator TW is injective and so W
U = (T ∗W )−1 φ−1
W WGU,W,W T −1
W
is invertible.
(2) ⇒ (3) and (2) ⇒ (4) are trivial.
(3) ⇒ (1) If GU,W,W = φW W T ∗W U TW is onto, then φ∗W W = φW W is invertible. Hence, T ∗W U TW =
φ−1
W WGU,W,W is onto. Thus, W is a fusion Riesz basis and U is invertible.
(1) ⇒ (5) Using Theorem 2.9 of [5],fW is also a fusion Riesz basis and so by (2.5), TfW is invertible by
Theorem 2.3. Hence, GU,fW ,W = φfW W T ∗W U TfW is invertible by the invertibility of U and φfW W{fi}i∈I =
W fi(cid:9)i∈I , for all {fi}i∈I ∈Pi∈IL Wi.
(cid:8)S−1
GU,fW ,W = φfW W T ∗W U TfW is invertible and therefore fW is a fusion Riesz basis by Theorem 2.3. By
Theorem 2.9 of [5], W is also a fusion Riesz basis. Applying Theorem 2.3, T ∗W and TfW are invertible and
it immediately follows the desired result.
(5) ⇒ (1) Note that φfW W is invertible by the definition of φfW W . Also, TfW is one to one since
(1) ⇔ (4) ⇔ (6) ⇔ (7) can be proved in an analogue way.
Similar to the question, when the inverse of a frame multiplier is a multiplier again [7, 16], we can
show that for fusion Riesz bases the operator keeps its structure. To state our results in a accessible way,
we need a new definition:
Definition 3.5. Let W = {(Wi, wi)}i∈I be a fusion frame and V = {(Vi, vi)}i∈I a Bessel fusion sequence.
The alternate cross-fusion frame operator LV W : H → H is defined by
We denote LW W as LW .
LV W = TV φV W T ∗W .
9
It follows that kLV Wk ≤
√BW BV
AW
and
(LV W )∗ =Xi∈I
πWi S−1
W πVi .
Obviously, LV W is an invertible operator if and only if V is a pseudo-dual of W . In particular, LV W = IH
if and only if V is a dual of W .
In the following we summarize the basic properties of LW .
Proposition 3.6. Let W = {(Wi, wi)}i∈I be a fusion frame. Then LW is positive, self-adjoint and
invertible operator.
Proof. It is easy to see that φW W is self-adjoint and so, LW is self-adjoint. Moreover, for all f ∈ H we
have
w2
i πWi S−1
W πWi f, f+
W πWi f, wi · S−1/2
hLW f, fi = *Xi∈I
= Xi∈IDwi · S−1/2
= Xi∈I
2!
W πWi f(cid:13)(cid:13)(cid:13)
i(cid:13)(cid:13)(cid:13)S−1/2
W (cid:13)(cid:13)(cid:13)
≥ (cid:13)(cid:13)(cid:13)S1/2
(cid:13)(cid:13)(cid:13)S1/2
W (cid:13)(cid:13)(cid:13)
kT ∗W fk2 ≥
AW
w2
−2
W πWi fE
2kfk2,
where AW is a lower bound of the fusion frame W . This shows that LW is positive and an invertible
operator in B(H).
As an easy consequence of Proposition 2.5, we state
Corollary 3.7. Let W = {(Wi, ωi)}i∈I be a fusion Riesz basis. Then LW = SW ′ .
Theorem 3.8. Let W = {(Wi, ωi)}i∈I and V = {(Vi, υi)}i∈I be fusion frames in H and U an invertible
operator in B(H). Then the following assertions hold.
(1) If GU,W,W is invertible, then W is a fusion Riesz basis and
W ′ U −1S−1
W ′ ,W,W .
G−1
U,W,W = GS−1
(2) If GU,V,W is invertible and V is a dual of W , then V is a fusion Riesz basis and G−1
(3) If GU,W,V is invertible, then W is a fusion Riesz basis and
U,V,W = GU −1,V,W .
G−1
U,W,V = G(LW V U SW ′ )−1,W,W ,
Proof.
(1) By Proposition 3.4 W is a fusion Riesz basis. Using Corollary 3.7 follows that
GU,W,WGS−1
W ′ U −1S−1
W ′ ,W,W = φW W T ∗W U TW φW W T ∗W S−1
W ′ U−1S−1
W ′TW
= φW W T ∗W S−1
W ′ TW = I(Pi∈I L fWi)ℓ2
.
With the same we have GS−1
W ′ U −1S−1
W ′ ,W,WGU,W,W = I(Pi∈I L fWi)ℓ2
.
10
(2) According to (2.2) we have TV φV W T ∗W = IH. In addition, the invertibility of GU,V,W implies that
φV W T ∗W has a right inverse. Using Proposition 2.3 V is a fusion Riesz basis, and therefore,
Hence,
φV W T ∗W TV = I(Pi∈I L Vi)ℓ2
.
GU,V,WGU −1,V,W = φV W T ∗W U TV φV W T ∗W U−1TV
= φV W T ∗W TV = I(Pi∈I L Vi)ℓ2
.
Similarly, GU −1,V,WG−1
U,V,W = I(Pi∈I L fWi)ℓ2
.
(3) It follows from Corollary 3.7.
Repeating the previous argument and using (2.2) leads to a characterization for fusion Riesz bases
due to U -fusion cross Gram matrices.
Theorem 3.9. Let V = {(Vi, υi)}i∈I be a dual fusion frame of W = {(Wi, ωi)}i∈I in H. The following
are equivalent:
(1) V is a fusion Riesz basis.
(2) GV,W = I(Pi∈I L Vi)ℓ2
.
(3) GV,W has a left inverse.
Consider W = {(Wi, ωi)}i∈I , V = {(Vi, υi)}i∈I and Z = {(Zi, zi)}i∈I as fusion frames in H and
U1, U2 ∈ B(H), it is obvious to see that
GU1,W,V GU2,W,Z = GU1TW φW Z T ∗
Z U2,W,V .
In particular, if V = {(Vi, υi)}i∈I is a dual of Z = {(Zi, zi)}i∈I , then
GU1,V,WGU2,V,Z = GU1U2,V,W .
As a special case, if U is invertible and V a fusion Riesz basis such that V is a dual of W , then GU,V,W
has an inverse in the form of Gram matrices
The above identity is also proved in Theorem 3.8.
(GU,V,W )−1 = GU −1,V,W .
11
3.2. Pseudo-inverses
In the following we discuss the pseudo-inverse of U -fusion cross Gram matrices with closed range,
and under some conditions we represent their pseudo-inverse as a U -fusion cross Gram matrix again
motivated by the discrete frame case [15]. In the following we state a sufficient condition for a U -fusion
cross Gram matrix having closed range.
Lemma 3.10. Let V = {(Vi, υi)}i∈I be a dual fusion frame of W = {(Wi, ωi)}i∈I in H and U an operator
in B(H) such that U V = {(U Vi, υi)}i∈I is also a fusion frame in H. Then GU,V,W has closed range and
Proof. The dual condition (2.2) follows that φV W T ∗W has closed range. Thus,
ran (GU,V,W ) = ran (φV W T ∗W ) .
ran (GU,V,W ) = ran (φV W T ∗W U TV )
= ran (φV W T ∗W TUV )
= ran (φV W T ∗W ) .
The assumptions in the above lemma are fulfilled, in particular, for U being invertible [32].
Theorem 3.11. Let V = {(Vi, υi)}i∈I be a dual fusion frame of W = {(Wi, ωi)}i∈I in H, also let
U ∈ B(H) and GU,V,W have closed range. The following are equivalent:
(1) G†U,V,W = GU †,V,W .
(2) ran (φV W T ∗W U∗) = ran (T ∗V U∗) and ran (φV W T ∗W U ) = ran (T ∗V U ).
(3) φV W T ∗W U∗ = T ∗V S−1
Proof. (2) ⇒ (1) Applying (2.2) follows that
V U∗ and φV W T ∗W U = T ∗V S−1
V U .
GU,V,WGU †,V,WGU,V,W = φV W T ∗W U TV φV W T ∗W U†TV φV W T ∗W U TV
= φV W T ∗W U U†U TV
= GU,V,W .
In addition, (2.2) is also follows that TW φ∗V W is surjective. Hence, we have
ran(cid:0)GU †,V,W(cid:1) = ran(cid:0)φV W T ∗W U†TV(cid:1)
= ran (φV W T ∗W U∗)
(by (2)) = ran (T ∗V U∗)
= ran (T ∗V U∗TW φ∗V W )
= ran(cid:0)G∗U,V,W(cid:1) .
12
Moreover,
ker(cid:0)GU †,V,W(cid:1) = ker(cid:0)φV W T ∗W U†TV(cid:1)
= ker(cid:0)U†TV(cid:1)
= ran(cid:0)T ∗V (U†)∗(cid:1)⊥
= ran (T ∗V U )⊥
(by (2)) = ran (φV W T ∗W U )⊥
= ker (U∗TW φ∗V W )
= ker(cid:0)G∗U,V,W(cid:1) .
Thus, (1) is obtained. The converse follows immediately from the above identities. To show (2) ⇒ (3)
using Douglas's theorem [27], there is an operator C ∈ B(H) such that
So,
φV W T ∗W U = T ∗V U C.
φV W T ∗W U = T ∗V U C
= T ∗V S−1
= T ∗V S−1
V TV T ∗V U C
V TV φV W T ∗W U = T ∗V S−1
V U.
The other identity is obtained similarly. The converse is clear.
Note that fusion Riesz bases satisfy in the assumptions of Theorem 3.11. These assumptions might
In particular, the failure to fulfill this equality makes the concept of
seem obvious, but are not [32].
duality of fusion frames interesting, so fusion frames are not 'just another' generalization of frames.
Remark 3.12. In the proof of Theorem 3.11(1) we can replace the fusion frame appeared in the right
side by any fusion frame Z such that V is its dual. More precisely, let W and Z be fusion frames and V
is a dual of both W and Z. Then
if and only if φV Z T ∗Z U∗ = T ∗V S−1
G†U,V,W = GU †,V,Z,
V U∗ and φV Z T ∗Z U = T ∗V S−1
V U .
Using a similar argument as Theorem 3.11 we obtain the following.
Theorem 3.13. Let W = {(Wi, ωi)}i∈I be a fusion frame in H, also let U ∈ B(H) and GU,V,W have
closed range. The following are equivalent:
(1) (GU,W,W )† = GL−1
W ,W,W .
(2) The operators φW W T ∗W L−1
W U∗ and T ∗W U∗ have the same range as φW W T ∗W U and T ∗W (L−1
W )∗U ,
W U †L−1
respectively, for all i ∈ I
W U∗ = T ∗W S−1
(3) φW W T ∗W L−1
W U∗ and φW W T ∗W U = T ∗W S−1
W U .
13
4. Stability of U -cross Gram Matrices of Operators
In this section, we state a general stability for the invertibility of U -fusion cross Gram matrices,
compare to the results on the invertibility of multipliers [46].
Theorem 4.1. Let W = {(Wi, υi)}i∈I be a Bessel fusion sequence and U1, U2 ∈ B(H) with kU1 − U2k <
µ. Also, let V = {(Vi, υi)}i∈I and Z = {(Zi, υi)}i∈I be fusion frames on H such that GU1,W,V is invertible
and
2
+ ǫkfk,
(4.1)
,
(4.2)
2
2
υ2
υ2
Xi∈I
for all f ∈ H, in addition
i kπZi fk2! 1
i kπZi f − πVi fk2! 1
+ λ2 Xi∈I
i kπVi fk2! 1
kU2k + kU2k < (cid:13)(cid:13)(cid:13)G−1
U1,W,V(cid:13)(cid:13)(cid:13)
B(cid:13)(cid:13)S−1
V (cid:13)(cid:13)
where ǫ > 0, B = max{BW , BV , BZ} andPi∈I υi2 < ∞. Then GU2,W,Z is also invertible.
≤ λ1 Xi∈I
√B(cid:19)√B(cid:13)(cid:13)S−1
Z (cid:13)(cid:13) Xi∈I
µ +(cid:18)λ1 + λ2 +
υi2! 1
Proof. First note that
−1
υ2
ǫ
2
υ2
k(SV − SZ ) fk = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I
≤ Xi∈I
i (πVi f − πZi f )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2 Xi∈I
υi2! 1
υ2
i k(πVi f − πZi f )k2! 1
2
,
for all f ∈ H. Therefore,
2
i (πVi − πZi ) S−1
(cid:13)(cid:13)(cid:0)S−1
Z (cid:1) f(cid:13)(cid:13)
V − S−1
Z f(cid:13)(cid:13)
= (cid:13)(cid:13)S−1
V (SV − SZ) S−1
Z f(cid:13)(cid:13)
V (cid:13)(cid:13)(cid:13)(cid:13)(SV − SZ ) S−1
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13) Xi∈I
υi2! 1
2 Xi∈I(cid:13)(cid:13)υ2
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13) Xi∈I
υi2! 1
≤ (cid:13)(cid:13)S−1
λ1 Xi∈I
Z f(cid:13)(cid:13)2! 1
i(cid:13)(cid:13)πVi S−1
i(cid:13)(cid:13)πZi S−1
V (cid:13)(cid:13) Xi∈I
υi2! 1
2(cid:16)λ1pBZ + λ2pBV + ǫ(cid:17)(cid:13)(cid:13)S−1
≤ (cid:13)(cid:13)S−1
Z f(cid:13)(cid:13)2! 1
Z f(cid:13)(cid:13)2! 1
Z (cid:13)(cid:13)kfk.
+ λ2 Xi∈I
υ2
υ2
2
2
2
14
Z f(cid:13)(cid:13)
+ ǫ(cid:13)(cid:13)S−1
Using the assumption (4.1) and the above computations imply that
2
υ2
υ2
υ2
υ2
V − S−1
V πVi − S−1
V πVi − S−1
Xi∈I
Z πZi )f(cid:13)(cid:13)2! 1
i(cid:13)(cid:13)(S−1
≤ Xi∈I
V πZi )f(cid:13)(cid:13) +(cid:13)(cid:13)(S−1
i(cid:0)(cid:13)(cid:13)(S−1
≤ Xi∈I
V (cid:13)(cid:13)kπVi f − πZi fk +(cid:13)(cid:13)S−1
i(cid:0)(cid:13)(cid:13)S−1
≤ Xi∈I
kπVi f − πZi fk2! 1
+(cid:13)(cid:13)S−1
i(cid:13)(cid:13)S−1
V (cid:13)(cid:13)2
V (cid:13)(cid:13)(cid:16)λ1pBZ + λ2pBV + ǫ(cid:17)kfk+
≤ h(cid:13)(cid:13)S−1
(cid:13)(cid:13)S−1
V (cid:13)(cid:13)(cid:13)(cid:13)S−1
V (cid:13)(cid:13)(cid:16)λ1pBZ + λ2pBV + ǫ(cid:17)1 +(cid:13)(cid:13)S−1
≤ (cid:13)(cid:13)S−1
Z (cid:13)(cid:13)pBZ Xi∈I
υi2! 1
2
V πZi − S−1
2
2
Z πZi )f(cid:13)(cid:13)(cid:1)2! 1
Z (cid:13)(cid:13)kπZi fk(cid:1)2! 1
Z (cid:13)(cid:13) Xi∈I
V − S−1
υ2
2
i kπZi fk2! 1
2(cid:16)λ1pBZ + λ2pBV + ǫ(cid:17)kfk
υi2! 1
Z (cid:13)(cid:13)pBZ Xi∈I
2kfk.
Finally, applying (4.2) we obtain
kGU1,W,V − GU2,W,Zk
= kφW V T ∗V U1TW − φW Z T ∗Z U2TWk
≤ kφW V T ∗V U1TW − φW V T ∗V U2TWk + kφW V T ∗V U2TW − φW Z T ∗Z U2TWk
V πVi − S−1
υ2
V πVi − S−1
i(cid:13)(cid:13)(S−1
V (cid:13)(cid:13)pBW BV kU1 − U2k + Xi∈I(cid:13)(cid:13)υiπWi (S−1
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13)pBW BV µ +"Xi∈I
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13)pBW BV µ +
≤ (cid:13)(cid:13)S−1
V (cid:13)(cid:13)(cid:16)λ1pBZ + λ2pBV + ǫ(cid:17)1 +(cid:13)(cid:13)S−1
(cid:13)(cid:13)S−1
< (cid:13)(cid:13)(cid:13)G−1
U1,W,V(cid:13)(cid:13)(cid:13)
−1
.
2
Z πZi )(cid:13)(cid:13)2# 1
Z (cid:13)(cid:13)pBZ Xi∈I
2
Z πZi )U2TW(cid:13)(cid:13)2! 1
kU2kpBW
υi2! 1
2 kU2kpBW
Therefore, GU2,W,Z is also invertible by Proposition 2.7.
15
Remark 4.2. It is worthwhile to mention that if we consider in Theorem 4.1
1. the perturbation condition
kπZi f − πVi fk ≤ λ1υi kπZi fk + λ2υi kπVi fk + ǫkfk
we can replace the assumption Pi∈I υi2 < ∞ by bounded weights. Hence, uniform fusion frames
satisfy a slightly different version of this theorem.
2. the condition
Xi∈I
kπZi f − πVi fk2! 1
2
≤ λ1 Xi∈I
i kπZi fk2! 1
2
υ2
+ λ2 Xi∈I
i kπVi fk2! 1
2
υ2
+ ǫkfk,
instead of (3.4) we get the same result substitutingPi∈I υi4 forPi∈I υi2.
As a consequence we obtain the following result.
Corollary 4.3. Let W = {(Wi, υi)}i∈I be a fusion Riesz basis with bounds AW and BW and Z =
{(Zi, υi)}i∈I a fusion frame in H such that (4.1) holds. Also, U ∈ B(H) with kU − IHk < µ. If
µ +(cid:18)λ1 + λ2 +
ǫ
√B(cid:19)√B(cid:13)(cid:13)S−1
Z (cid:13)(cid:13) Xi∈I
2
υi2! 1
B(cid:13)(cid:13)S−1
W (cid:13)(cid:13)
where ǫ > 0 and B = max{BW , BZ} andPi∈I υi2 < ∞. Then GU,W,Z is also invertible.
,
kUk + kUk <
AW
5. Acknowledgement
The first and the second authors were supported in part by the Iranian National Science Foundation
(INSF) under Grant 97018155. The last author was partly supported by the START project FLAME
Y551-N13 of the Austrian Science Fund (FWF) and the DACH project BIOTOP I-1018-N25 of Austrian
Science Fund (FWF). He also thanks Nora Simovich for help with typing.
References
[1] A. Aldroubi, A. Baskakov, I. Krishtal. Slanted matrices, banach frames, and sampling. J. Funct. Anal.,
255(7):1667 -- 1691, 2008.
[2] A. Aldroubi, C. Cabrelli, A. C¸ akmak, U. Molter, A. Petrosyan. Iterative actions of normal operators. J.
Funct. Anal., 272(3):1121 -- 1146, 2017.
[3] S. T. A. Ali, J.-P. Antoine, and J.-P. Gazeau. Continuous frames in Hilbert space. Ann. Physics, 222(1):1 -- 37,
1993.
[4] J.-P. Antoine and P. Balazs. Frames and semi-frames. J. Phys. A , 44:205201, 2011.
16
[5] A. A. Arefijamaal and F. A. Neyshaburi. Some properties of alternate duals and approximate alternate duals
of fusion frames. Turkish J. Math., 41(5):1191 -- 1203, 2018.
[6] F. V. Atkinson, H. Langer, R. Mennicken, and A. A. Shkalikov. The essential spectrum of some matrix
operators. Math. Nachr., 167(1):5 -- 20, 1994.
[7] P. Balazs. Basic definition and properties of Bessel multipliers. J. Math. Anal. Appl., 325(1):571 -- 585, 2007.
[8] P. Balazs. Matrix-representation of operators using frames. Sampl. Theory Signal Image Process., 7(1):39 -- 54,
2008.
[9] P. Balazs, J.-P. Antoine, and A. Grybos. Weighted and controlled frames: Mutual relationship and first
numerical properties. Int. J. Wavelets Multiresolut. Inf. Process., 8(1):109 -- 132, 2010.
[10] P. Balazs and K. Grochenig. A guide to localized frames and applications to Galerkin-like representations
of operators. In I. Pesenson, H. Mhaskar, A. Mayeli, Q. T. L. Gia, and D.-X. Zhou, editors, Frames and
Other Bases in Abstract and Function Spaces, Applied and Numerical Harmonic Analysis series (ANHA).
Birkhauser/Springer, 2017.
[11] P. Balazs, N. Holighaus, T. Necciari, and D. Stoeva. Frame theory for signal processing in psychoacoustics.
In R. Balan, J. J. Benedetto, W. Czaja, and K. Okoudjou, editors, Excursions in Harmonic Analysis Vol. 5,
Springer, 2017.
[12] P. Balazs and G. Rieckh. Oversampling operators: Frame representation of operators. Analele Universitatii
"Eftimie Murgu", 18(2):107 -- 114, 2011.
[13] P. Balazs and G. Rieckh. Redundant representation of operators. arXiv:1612.06130.
[14] P. Balazs, M. Shamsabadi, A. A. Arefijamaal and C. Gardon. Representation of operators using fusion
frames. preprint.
[15] P. Balazs, M. Shamsabadi, A. A. Arefijamaal, and A. Rahimi. U -cross Gram matrices and their invertibility.
J. Math. Anal. Appl., 476(2):367-390, 2019.
[16] P. Balazs and D. Stoeva. Representation of the inverse of a multiplier. J. Math. Anal. Appl., 422:981 -- 994,
2015.
[17] J. Benedetto and P. Ferreira. Modern sampling theory. Mathematics and applications. Birkhauser, 2001.
[18] B. G. Bodmann, P. G. Casazza. The road to equal-norm Parseval frames. J. Funct. Anal., 258(2): 397 -- 420,
2010.
[19] H. Bolcskei, F. Hlawatsch, and H. G. Feichtinger. Frame-theoretic analysis of oversampled filter banks. IEEE
Trans. Signal Processing, 46(12):3256 -- 3268, 1998.
[20] P. G. Casazza and G. Kutyniok. Frames of subspaces. Cont. Math., 2004.
[21] P. G. Casazza, G. Kutyniok, and S. Li. Fusion frames and distributed processing. Appl. Comput. Harmon.
Anal., 254(1):114 -- 132, 2008.
[22] O. Christensen. Frames and Bases. An Introductory Course. Applied and Numerical Harmonic Analysis.
Basel Birkhauser, 2008.
[23] O. Christensen and R. S. Laugesen. Approximately dual frames in Hilbert spaces and applications to Gabor
frames. Sampl. Theory Signal Image Process, 9(3):77 -- 89, 2010.
17
[24] N. Cotfas and J. P. Gazeau. Finite tight frames and some applications. J. Phys. A, 43(19):193001, 2010.
[25] 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.
[26] I. Daubechies, A. Grossmann, and Y. Meyer. Painless non-orthogonal expansions. J. Math. Phys., 27:1271 --
1283, 1986.
[27] R. G. Douglas, On majorization, factorization and range inclusion of operators onHilbert space. Proc. Amer.
Math. Soc., 17(2):413 -- 415, 1996.
[28] R. J. Duffin and A. C. Schaeffer. A class of nonharmonic Fourier series. Trans. Amer. Math. Soc., 72:341 -- 366,
1952.
[29] H. G. Feichtinger and T. Strohmer. Gabor Analysis and Algorithms - Theory and Applications. Birkhauser
Boston, 1998.
[30] H. Fuhr, J. Lemvig. System bandwidth and the existence of generalized shift-invariant frames. J. Funct.
Anal., 276(2):563 -- 601, 2019.
[31] L. Gaul, M. Kogler, and M. Wagner. Boundary Element Methods for Engineers and Scientists. Springer,
2003.
[32] P. Gavrut¸a. On the duality of fusion frames. J. Math. Anal. Appl., 333(2):871 -- 879, 2007.
[33] I. Gohberg, S. Goldberg, and M. A. Kaashoek. Basic Classes of Linear Operators. Birkhauser, 2003.
[34] K. Grochenig. Foundations of Time-Frequency Analysis. Birkhauser Boston, 2001.
[35] H. Harbrecht, R. Schneider, and C. Schwab. Multilevel frames for sparse tensor product spaces. Numer.
Math., 110(2):199 -- 220, 2008.
[36] S. B. Heineken and P. Morillas. Properties of finite dual fusion frames. Linear Algebra and its Applications,
453:1 -- 27, 2014.
[37] S. B. Heineken, P. Morillas, A. Benavente, and M. Zakowicz. Dual fusion frames. Arch. Math. (Basel),
103(4):355 -- 365, 2014.
[38] W. Kreuzer, P. Majdak, and Z. Chen. Fast multipole boundary element method to calculate head-related
transfer functions for a wide frequency range. J. Acoust .Soc. Am., 126(3):1280 -- 1290, 2009.
[39] D. Marelli and M. Fu. Performance analysis for subband identification.
IEEE Trans. Signal Process.,
51(12):3128 -- 3142, 2003.
[40] P. Oswald. Stable space splittings and fusion frames. In Wavelets XIII (V. Goyal, M. Papadakis, D. Van de
Ville, eds.), Proceedings of SPIE San Diego, volume 7446, 2009.
[41] A. Pietsch. Operator Ideals. North-Holland Publishing Company, 1980.
[42] R. Schatten. Norm Ideals of Completely Continuous Operators. Springer Berlin, 1960.
[43] M. Shamsabadi and A. A. Arefijamaal. The invertibility of fusion frame multipliers. Linear Multilinear
Algebra, 65(5):1062 -- 1072, 2016.
[44] M. Speckbacher and P. Balazs. Reproducing pairs and the continuous nonstationary Gabor transform on lca
groups. J. Phys. A, 48:395201, 2015.
18
[45] R. Stevenson. Adaptive solution of operator equations using wavelet frames. SIAM J. Numer. Anal.,
41(3):1074 -- 1100, 2003.
[46] D. T. Stoeva and P. Balazs. Invertibility of multipliers. Appl. Comput. Harmon. Anal., 33(2):292 -- 299, 2012.
[47] W. Sun. G-frames and g-Riesz bases. J. Math. Anal. Appl., 322(1):437 -- 452, 2006.
[48] J. Weidmann. Linear Operators in Hilbert Spaces. Springer New York, 1980.
19
|
1309.4280 | 1 | 1309 | 2013-09-17T12:22:22 | From local to global ideal-triangularizability | [
"math.FA"
] | Let L be a Banach lattice with order continuous norm, and let S be a multiplicative semigroup of ideal-triangularizable positive compact operators on L such that, for every pair A, B in S, the atomic diagonal of the commutator AB-BA is equal to zero. We prove that the semigroup S is ideal-triangularizable. | math.FA | math |
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY
ROMAN DRNOVSEK, MARKO KANDI ´C
Abstract. Let L be a Banach lattice with order continuous norm, and let S be a
multiplicative semigroup of ideal-triangularizable positive compact operators on L such
that, for every pair {S, T } ⊆ S, the atomic diagonal of the commutator ST − T S is equal
to zero. We prove that the semigroup S is ideal-triangularizable.
Math. Subj. Classification (2000): 47A15, 47B65.
Key words: Invariant subspaces, Banach lattices, closed ideals, positive operators, idem-
potents, the atomic diagonal, semigroups of operators.
1. Introduction
Let L be a normed Riesz space, and let L+ denote the positive cone of L. Elements
of L+ are called positive vectors. A linear subspace J of L is an ideal in L whenever
x ≤ y and y ∈ J imply x ∈ J. A band in L is an order closed ideal. A band B in L is
a projection band whenever L = B ⊕ Bd. In this case, there exists a positive projection
onto B with a kernel Bd, and the direct sum is also an order direct sum. A Banach lattice
L is said to have an order continuous norm whenever every decreasing net {xα}α with
infimum 0 of positive vectors in L converges in norm to 0. It is well known that a Banach
lattice with order continuous norm is Dedekind complete, and that its closed ideals are
bands [2]. Moreover, every band in a Dedekind complete Banach lattice is a projection
band.
A non-zero vector a ∈ L+ is an atom in a normed Riesz space L if 0 ≤ x, y ≤ a and
x ∧ y = 0 imply either x = 0 and y = 0, or equivalently, if 0 ≤ x ≤ a implies x = λa for
some λ ≥ 0, i.e., the principal ideal Ba generated by a is one dimensional. It turns out
that Ba is a projection band [13]. The decomposition L = Ba ⊕ Bd
a implies that for an
arbitrary (positive) vector x ∈ L there exist a (positive) scalar λx and a (positive) vector
Date: June 4, 2018.
THE PAPER WILL APPEAR IN LINEAR AND MULTILINEAR ALGEBRA.
1
2
ROMAN DRNOVSEK, MARKO KANDI ´C
yx ∈ Bd
a such that x = λxa + yx. The linear functional ϕa : L → R associated to the atom
a is defined by ϕa(x) = λx. A positive vector x ∈ L is a quasi-interior point in L if the
principal ideal in L generated by x is dense in L.
A (linear) operator T between Riesz spaces L1 and L2 is said to be positive if T maps
the positive cone L+
1 into the positive cone L+
2 . The absolute kernel N (T ) of T is defined
as N (T ) = {x ∈ L1 : T x = 0}.
If it is a zero ideal, then T is said to be strictly
positive. The range ideal R(T ) of T is the ideal generated by the range of T . Recall
that every positive operator T on a Banach lattice is continuous, and its spectral radius
r(T ) belongs to its spectrum σ(T ). Throughout the text we assume that functionals and
operators acting on normed Riesz spaces are continuous.
A family F of operators on a normed Riesz space L is said to be ideal-reducible when-
ever there exists a nontrivial closed ideal in L that is invariant under all operators from
the family F . Otherwise, we say that F is ideal-irreducible. The following very useful
proposition is proved in [8].
Proposition 1.1. Let L be a normed Riesz space, and let S be a semigroup of positive
operators on L. The following statements are equivalent:
(a) S is ideal-reducible;
(b) there exist a nonzero positive functional ϕ ∈ L∗ and a nonzero positive vector
f ∈ L+ such that ϕ(Sf ) = {0};
(c) there exist nonzero positive operators A and B on L such that ASB = {0};
(d) some nonzero semigroup ideal of S is ideal-reducible.
If there is a chain C that is maximal as a chain of closed ideals of L and that has the
property that every ideal in C is invariant under all the operators in a family F , then F
is said to be ideal-triangularizable, and C is an ideal-triangularizing chain for F . Every
maximal chain of closed ideals is also maximal as a chain of closed subspaces of L [6]. Let
I and J be closed ideals in L that are invariant under every operator from a family F . If
I ⊆ J, then F induces a family bF of operators on the quotient normed Riesz space J/I as
follows. For each T ∈ F the operator bT is defined on J/I by bT (x + I) = T x + I. Any such
family bF is called a family of ideal-quotients of the family F . A set P of properties is said
to be inherited by ideal-quotients if every family of ideal-quotients of a family of operators
satisfying P also satisfies the same properties. Ideal-triangularizability of operators on
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY1
3
Banach lattices is in practice often reduced to ideal-reducibility of operators. The details
are contained in the following lemma that was proved in [5].
Lemma 1.2 (The Ideal-triangularization Lemma). Let P be the set of properties inher-
ited by ideal-quotients.
If every family of operators on a Banach lattice of dimension
greater than one which satisfies P is ideal-reducible, then every such family is ideal-
triangularizable.
Let T be an operator on a Banach lattice L with an order continuous norm.
If B1
and B2 are bands of L such that B2 ⊆ B1, then the quotient Banach lattice B1/B2 is
isometrically lattice isomorphic to the band B := B1 ∩ Bd
2, and so the norm of B is also
order continuous. If T leaves the bands B1 and B2 invariant, then the induced operator
T on B1/B2 can be identified with the compression of T on B, that is, with the operator
PBT B, where PB denotes the band projection on B.
For the terminology not explained in the text about normed Riesz spaces or Banach
lattices and operators acting on them we refer the reader to classical textbooks [1], [2],
[16] and [13].
Finally, we explain the organization of the paper. Section 2 deals with the structure
of positive idempotents of finite ranks.
In Section 3 the atomic diagonal operator D
introduced recently in [12] is connected with the known band projection onto the center
of a Dedekind complete Banach lattice. The main result of the paper is Theorem 4.5
(and its finite-dimensional special case Corollary 4.6) in Section 4. It reveals the relation
between the atomic diagonal operator and the ideal-triangularizability of semigroups of
positive operators. We also prove Theorem 4.2 that completes [12, Theorem 3] and extends
[3, Theorem 3.8] and [14, Corollary 27].
2. The structure of positive idempotents of finite rank
The structure result for positive idempotents on Lp-spaces (1 ≤ p < ∞) was given by
Zhong in [19]; see also the book [15, Section 8.7] for a complete treatment. In this section
we extend this result to Banach lattices with order continuous norm.
Let L be a normed Riesz space, and let P be a positive idempotent operator of rank
one on L. Then there exist a positive vector x ∈ L and a positive linear functional ϕ on
L such that ϕ(x) = 1 and P = x ⊗ ϕ, i.e., P y = ϕ(y)x for y ∈ L. The ideal-irreducibility
4
ROMAN DRNOVSEK, MARKO KANDI ´C
of such an idempotent is characterized by the following lemma that is a generalization of
[15, Lemma 8.7.11].
Lemma 2.1. Let L be a normed Riesz space, and let P = x ⊗ ϕ be a positive idempotent
of rank one on L, where x ∈ L is a positive vector and ϕ a positive linear functional on
L. The following statements are equivalent:
(a) P is ideal-irreducible;
(b) The vector x is a quasi-interior point of L, and ϕ is a strictly positive functional
on L;
(c) P and P ∗ are strictly positive operators on L and L∗, respectively.
Proof. To see that (a) implies (b), assume that P is ideal-irreducible. If y is a positive
vector in the absolute kernel N (ϕ) of ϕ, then P y = ϕ(y)x = 0 implies that N (ϕ) is
invariant under P . Since P is ideal-irreducible, we have N (ϕ) = {0}, so that ϕ is strictly
positive on L. If 0 ≤ y ≤ λx with y ∈ L and λ ≥ 0, then P y ≤ P y ≤ λP x = λϕ(x)x
implies that the closure of the principal ideal generated by the positive vector x is invariant
under P . Hence, x needs to be a quasi-interior point in L.
Assume that (b) holds. To prove (a), assume that J is a proper closed ideal that
is invariant under the operator P . By [8, Lemma 1.1], there exists a nonzero positive
functional ψ on L which is zero on J. Then for y ∈ J we have 0 = ψ(P y) = ψ(x)ϕ(y).
Since x is a quasi-interior point, ψ(x) > 0 by [1, Lemma 4.15], and so we conclude that
ϕ(y) = 0. Since ϕ is a strictly positive functional, this implies that y = 0, so that J = {0}.
Therefore, the operator P does not have nontrivial closed invariant ideals.
Assume again that (b) holds. To show (c), assume that P y = 0 and P ∗ψ = 0 for some
positive vectors y and ψ in L and L∗, respectively. Since ϕ is a strictly positive functional
on L, the equality 0 = P y = ϕ(y)x implies that y = 0. Since x is a quasi-interior point
in L, the equality 0 = P ∗(ψ) = ψ(x)ϕ implies that ψ = 0 by [1, Lemma 4.15].
To see that (c) implies (b), suppose that P and P ∗ are strictly positive operators on
L and L∗, respectively. Let ψ be an arbitrary positive functional on L with ψ(x) = 0.
Then P ∗ψ = (ϕ ⊗ x)ψ = ψ(x)ϕ = 0 implies ψ = 0, since P ∗ is strictly positive. It follows
from [1, Lemma 4.15] that x is a quasi-interior point in L. Suppose now that ϕ(y) = 0
for some positive vector y ∈ L. Then P y = (x ⊗ ϕ)(y) = ϕ(y)x = 0. Since P is strictly
positive, we have y = 0, so that ϕ is strictly positive on L.
(cid:3)
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY2
5
Note that in the case when any of the equivalent statements of Lemma 2.1 holds for a
positive ideal-triangularizable idempotent P of rank one, then L is lattice isomorphic to
R, and P is just the identity on R.
Positive idempotents of finite rank are treated in the following propositions that ex-
tend [15, Proposition 8.7.12]. It should be noted that these propositions contain addi-
tional statements regarding the structure of finite rank idempotents that are also ideal-
triangularizable.
Proposition 2.2. Let L be a Banach lattice with order continuous norm and P a positive
idempotent of finite rank r such that P and P ∗ are strictly positive. There exist pairwise
disjoint bands L1, . . . , Lr in L such that L = L1 ⊕ · · · ⊕ Lr and P is of the form
P = P1 ⊕ · · · ⊕ Pr,
where Pj is an ideal-irreducible positive idempotent of rank 1 on Lj. If, in addition, P is
ideal-triangularizable, then r = dim L and P is the identity operator on L.
Proof. If r = 1, then the assertion follows from Lemma 2.1 and the remark following it.
Assume r > 1. Let x and y be linearly independent positive vectors in the range of the
operator P. Replacing y by x + y (if necessary) we may assume 0 ≤ x ≤ y and x 6= y. Let
t0 be the supremum of the nonempty set {t ∈ R : t ≥ 0, ty ≤ x} that is bounded from
above. Since t0y ≤ x ≤ y and x 6= y, we have t0 < 1. Pick s ∈ (t0, 1) and let z = x − sy.
The vectors z+ and z− are both nonzero, since z is neither positive nor negative. From
z = P z = P z+ − P z− ≤ P z+ it easily follows z+ ≤ P z+. Since P is strictly positive,
P (P z+ − z+) = 0 implies P z+ = z+. Let J be the closed ideal generated by the vector
z+. Note that J is invariant under P and it is nontrivial, since z− is nonzero. Therefore,
the operator P can be decomposed as
with respect to the decomposition L = J ⊕ J d.
(cid:20) P1 P2
0 P4 (cid:21)
We claim that P2 = 0. Since P 2 = P , we have P 2
1 = P1, P 2
4 = P4 and P1P2 +P2P4 = P2.
We conclude that P1P2P4 = 0. Since the operator P is strictly positive, the operator P1
is also strictly positive, and so P2P4 = 0. Taking the adjoints we have P ∗
the adjoint P ∗ is strictly positive, so is the operator P ∗
2 = 0. Since
4 on (J d)∗, and hence P2 = 0 as
4 P ∗
claimed. We finish the proof by induction on r.
(cid:3)
(1)
P =
0 XQ XQY
QY
0 Q
0
0
0
=
X
I
0 Q(cid:2) 0 I Y (cid:3) ,
where Q is an idempotent of rank r such that Q and Q∗ are strictly positive operators on B2
and B∗
2, respectively. Moreover, there exist pairwise disjoint bands L1, . . . , Lr in B2 with
B2 = L1 ⊕ · · · ⊕ Lr such that Q = Q1 ⊕ · · · ⊕ Qr, where Qj is a strictly positive idempotent
of rank one on Lj. If, in addition, P is ideal-triangularizable, then the dimension of B2
is finite and Q is the identity operator on B2.
Proof. Let B1 be the absolute kernel of the operator P , and let B be the band generated
by B1 and the range of the operator P . Let B2 = B ∩ Bd
1 and B3 = Bd. Let
6
ROMAN DRNOVSEK, MARKO KANDI ´C
Proposition 2.3. Let L be a Banach lattice with order continuous norm and P a positive
idempotent of finite rank r on L. There exist pairwise disjoint bands B1, B2 and B3 with
L = B1 ⊕ B2 ⊕ B3 such that P can be decomposed as
0 X Z
0 Q Y
0
0
0
be the operator matrix corresponding to the operator P with respect to the decomposition
L = B1 ⊕ B2 ⊕ B2.
Since P is an idempotent, we have Q2 = Q, X = XQ, Y = QY and Z = XY , so that
the decomposition (1) is proved. It also shows that the rank of Q must be r. We claim
that Q and Q∗ are strictly positive operators on B2 and B∗
2, respectively. Suppose that
for some positive vector x ∈ B2 we have Qx = 0. Then Xx = XQx = 0, so that P x = 0
2 satisfies Q∗ϕ = 0. Then ϕ is
implies x = 0. Suppose that a positive functional ϕ in B∗
zero on the range of the operator P which implies ϕ = 0. The last two assertions hold by
Proposition 2.2.
(cid:3)
3. The atomic diagonal operator
Throughout this section, let L be a Dedekind complete Banach lattice. Denote by Lr(L)
the Dedekind complete Riesz space of all regular operators on L, i.e., those operators that
are linear combinations of positive operators.
It is well known that Lr(L) becomes a
Banach lattice algebra with respect to the regular norm defined by kT kr := kT k. The
center Z(L) is the ideal in Lr(L) generated by the identity operator I, i.e.,
Z(L) = {T ∈ Lr(L) : T ≤ λI for some λ ≥ 0}.
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY3
7
If T ∈ Z(L), then the operator norm and the regular norm of T coincide. Since Z(L) is
also a band in Lr(L), we have a band decomposition Lr(L) = Z(L) ⊕ Z(L)d. Let P be
the band projection onto Z(L). By a result of Voigt [18], P is a contraction with respect
to the operator norm. Schep [17] proved that the component P(T ) of a positive operator
T in Z(L) is
(2)
P(T ) = inf( nXi=1
PiT Pi : 0 ≤ Pi ≤ I, P 2
i = Pi,
Pi = I) .
nXi=1
Let A be the band generated by all atoms in L, and let A ⊆ A be the maximal set of
pairwise disjoint atoms of norm one. Given a ∈ A, we denote by Pa the band projection
onto the band Ba. Let T be a positive operator on L. In [12] it is proved that the operator
D(T ) = sup(Xa∈F
PaT Pa : F is a finite subset of A)
exists, since it is a supremum of an increasing net that is bounded from above, and it also
satisfies 0 ≤ D(T ) ≤ T. If L is atomic (i.e., A = L), then D(T ) = P(T ). For a general L
we have
Proposition 3.1. Let T be a positive operator on L. If PA denotes the band projection
onto the band A, then D(T ) = PAP(T ).
Proof. Since PA ∈ Z(L), it commutes with P(T ), and so PAP(T ) = PAP(T )PA. Applying
nXi=1
Pi = I) PA.
Pi = I) =
nXi=1
Qi = PA) .
nXi=1
(PAPi)T (PiPA) : 0 ≤ Pi ≤ I, P 2
i = Pi,
QiT Qi : 0 ≤ Qi ≤ PA, Q2
i = Qi,
PiT Pi : 0 ≤ Pi ≤ I, P 2
i = Pi,
Since 0 ≤ PA ≤ I, PA is order continuous, so that
Schep's formula (2) we obtain
PAP(T ) = PA inf( nXi=1
PAP(T ) = inf( nXi=1
= inf( nXi=1
Since A is an atomic Dedekind complete Banach lattice, PAP(T ) = D(T ).
(cid:3)
We use the equality from Proposition 3.1 to extend the operator D to the operator on
the whole space Lr(L), so we define D(T ) := PAP(T ) for T ∈ Lr(L). This extension is
called the atomic diagonal operator and the operator D(T ) on L is said to be the atomic
diagonal of an operator T ∈ Lr(L).
T =(cid:20) T1 T2
T3 T4 (cid:21) ,
T =(cid:20) T1
T3
T2
T4 (cid:21)
and
PA =(cid:20) I 0
0 0 (cid:21) .
8
ROMAN DRNOVSEK, MARKO KANDI ´C
Proposition 3.2. The following assertions hold for the atomic diagonal operator D:
(a) D is a band projection onto the band
{T ∈ Lr(L) :
T ≤ λPA for some λ ≥ 0}
in Lr(L) that can be identified with the center Z(A);
(b) kD(T )k ≤ kP(T )k ≤ kT k for all T ∈ Lr(L);
(c) P(T ) = D(T ) for every positive compact operator T on L.
Proof. For the proof of (a) just note that, with respect to the band decomposition L =
A ⊕ Ad, an operator T ∈ Lr(L), its modulus T and the band projection PA have the
block forms
To show (b), use Voigt's theorem [18, Theorem 1.4] to obtain that kD(T )k = kPAP(T )k ≤
kP(T )k ≤ kT k.
If PAd is the band projection on Ad, then the operator
P(T ) − D(T ) = PAdP(T ) = PAdP(T )PAd ∈ Z(L)
is dominated by a positive compact operator PAdT PAd, and so it is equal to 0 by [17,
Corollary 1.7]. This proves (c).
(cid:3)
4. The ideal-triangularizability
We first extend [12, Theorem 3] by showing the implication that remained unproven in
[12]. In the proof we will make use of the following extension of Ringrose's Theorem; see
[4, Theorem 2.1]. Recall that if T is a power-compact operator (i.e., some power of T is
a compact operator) on a Banach space X, then the algebraic multiplicity m(T, λ) of a
non-zero complex number λ is the dimension of the subspace ker ((λ−T )k), where k is the
smallest natural number such that ker ((λ − T )k) = ker ((λ − T )k+1). On the other hand,
the geometric multiplicity of λ is the dimension of the subspace ker (λ − T ). We say that
a chain C of closed subspaces of X is a complete chain if it contains arbitrary intersections
and closed linear spans of its members. If a closed subspace M is in a complete chain
C, then the predecessor M− of M in C is defined as the closed linear span of all proper
subspaces of M belonging to C.
Theorem 4.1. Let T be a power-compact operator on a Banach space X, and let C be
a complete chain of closed subspaces invariant under T . Let C ′ be a subchain of C of all
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY4
9
subspaces M ∈ C such that M− 6= M. For each M ∈ C ′, define TM to be the quotient
operator on M/M− induced by T . Then
Moreover, for each non-zero complex number λ we have
σ(T ) \ {0} = [M∈C ′
m(T, λ) = XM∈C ′
σ(TM) \ {0}.
m(TM, λ).
The following theorem completes [12, Theorem 3]. It extends [3, Theorem 3.8], where
the implication (c) ⇒ (a) has been proved for a positive compact operator on a Banach
lattice lp (1 < p < ∞). It is also a generalization of [14, Corollary 27], where positive
trace-class operators acting on Lp(µ) (1 ≤ p < ∞) are considered. A diagonal entry of an
operator T on a Banach lattice L is the number ϕa(T a), where ϕa is the linear functional
associated to an atom a ∈ L+.
Theorem 4.2. Let T be a positive power-compact operator on a Banach lattice L with
order continuous norm. The following conditions are mutually equivalent:
(a) T is ideal-triangularizable;
(b) T − D(T ) is quasinilpotent;
(c) The diagonal entries of T consists precisely of eigenvalues (except maybe zero) of
the operator T repeated according to their algebraic multiplicities.
Proof. It suffices to prove that (c) implies (a), since other implications were proved in
[12]. Let C be a maximal chain of closed ideals invariant under the operator T . We will
prove that, in fact, C is an ideal-triangularizing chain for T . Maximality of C implies that
{0} and L are elements of C, and that C is a complete chain. Therefore, we only need to
prove that for every J ∈ C the dimension of the quotient space J/J− is less than or equal
to one. Maximality of C also implies that the induced operator TJ is ideal-irreducible on
J/J−, and so it is not quasinilpotent by [9, Theorem 1.3] provided dim(J/J−) ≥ 2. Note
that TJ can be identified with the compression of T to the band J ∩ J d
−.
Let B be the set of all ideals J in C with dim(J/J−) ≥ 2. Assume that B is not
empty. Let λ be the largest number among the diagonal entries of the compressions of
the operator T onto J ∩ J d
−, where J runs over B. Pick any ideal J ∈ B such that λ
is the diagonal entry of the compression of the operator T onto J ∩ J d
−. We claim that
λ = r(TJ ). Otherwise we would have λ < r(TJ), and so the number of appearances of
10
ROMAN DRNOVSEK, MARKO KANDI ´C
r(TJ ) as a diagonal entry of the compression of T onto some one-dimensional quotient
space (induced by an ideal in C) would be precisely the algebraic multiplicity of r(TJ ). In
view of Theorem 4.1 this would imply that r(TJ) is not in the spectrum of TJ which is
absurd. So, λ = r(TJ ) as claimed.
Therefore, the spectral radius r(TJ ) appears on the diagonal of the compression of
T onto J ∩ J d
−. By [12, Corollary 1], this compression is ideal-reducible, contradicting
the maximality of the chain C. This implies that B is an empty set, and so C is an
ideal-triangularizing chain for the operator T .
(cid:3)
Now we turn our attention to multiplicative semigroups of ideal-triangularizable positive
operators. We first recall [5, Theorem 4.5].
Theorem 4.3. Every semigroup of quasinilpotent positive compact operators on a Banach
lattice is ideal-triangularizable.
As a consequence we obtain the following result.
Theorem 4.4. Let L be an atomless Banach lattice with order continuous norm. If S
is a semigroup of ideal-triangularizable positive compact operators on L, then it is ideal-
triangularizable.
Proof. By [8, Proposition 4.5], every operator in S is quasinilpotent, and so Theorem 4.3
can be applied.
(cid:3)
Theorems 4.3 and 4.4 do not hold without the assumption that the operators are
compact. Namely,
in [10] there was constructed an irreducible semigroup of square-
zero positive operators on Lp[0, 1) (1 ≤ p < ∞) in which any finite number of elements
generate an ideal-triangularizable semigroup. So, one may seek for extensions of Theorem
4.4 by relaxing the assumption that L is atomless. It will turn out that in this case the
atomic diagonal operator plays important role.
Let L be a Banach lattice with order continuous norm, and let S be a semigroup of
positive operators on L.
If S is ideal-triangularizable, then D(ST ) = D(S)D(T ) for
every pair {S, T } ⊆ S, as was shown in the proof of [12, Proposition 3]. It follows that
D(ST ) = D(T S) or equivalently D(ST − T S) = 0 for every pair {S, T } ⊆ S. The
following main theorem of the paper treats the question when the converse implication
holds. Note also that it is a generalization of Theorem 4.4.
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY5
11
Theorem 4.5. Let S be a semigroup of ideal-triangularizable positive compact operators
on a Banach lattice L with order continuous norm such that D(ST ) = D(T S) for every
pair {S, T } ⊆ S. Then the semigroup S is ideal-triangularizable.
Proof. With no loss of generality we may assume that S is closed under positive scalar
multiplication. Let us prove that the assumptions are satisfied for the closure of S, so that
we may also assume that S is a closed set. To this end, let a sequence {Sk}k ⊆ S converge
to a positive compact operator S. By [17, Theorem 1.6], D(Sk) is a positive compact
operator for each k. Since the diagonal operator D is continuous by Theorem 3.2, the
sequence {D(Sk)}k converges to the operator D(S), so that the sequence {Sk − D(Sk)}k
converges to the operator S − D(S). Since the operators Sk − D(Sk) are all quasinilpotent
by Theorem 4.2, the compact operator S −D(S) is quasinilpotent as well by [15, Corollary
7.2.11.]. Thus, Theorem 4.2 implies that S is ideal-triangularizable. Since D(ST ) =
limk→∞ D(SkT ) = limk→∞ D(T Sk) = D(T S) for every T ∈ S, we conclude (because of
symmetry) that D(ST ) = D(T S) for every pair {S, T } in the closure of S. Therefore, we
have shown that there is no loss of generality in assuming that S is a closed set.
We first show that S is ideal-reducible. Assume otherwise. It follows from Theorem 4.3
that any semigroup consisting of quasinilpotent positive compact operators and multiples
of the identity operator is ideal-triangularizable. Therefore, there exists an operator A ∈ S
such that r(A) = 1 and A is not the identity operator I. Since A is ideal-triangularizable,
[12, Proposition 4] implies that its spectrum σ(A) is contained in [0, 1]. We distinguish
two cases:
Case 1: The geometric multiplicity of the eigenvalue 1 ∈ σ(A) is equal to its algebraic
multiplicity. Then the sequence {Ak}k converges to a nonnegative idempotent E ∈ S of
finite rank satisfying E 6= I. Since E is ideal-triangularizable, it follows from Proposition
2.3 that there exist bands B1, B2 and B3 such that, with respect to the decomposition
L = B1 ⊕ B2 ⊕ B3, E has the block-triangular form
E =
0 ,
0 X XY
0
Y
0
I
0
where only one of the bands B1 and B3 may be equal to zero, and the dimension n =
dim B2 is finite, so that B2 is atomic and isomorphic to Rn by [16, Corollary 1, p.70].
Assume first that B1 = {0} and Y = 0, so that E has the form
E =(cid:20) I 0
0 0 (cid:21) .
12
Let
ROMAN DRNOVSEK, MARKO KANDI ´C
S =(cid:20) S1 S2
S3 S4 (cid:21) ∈ S and
T =(cid:20) T1 T2
T3 T4 (cid:21) ∈ S
be arbitrary operators in S. The operators ES · T E and T E · ES are both in S, and they
have the same diagonal. From
ES · T E =(cid:20) S1T1 + S2T3 0
0 (cid:21)
0
and
T E · ES =(cid:20) T1S1 T1S2
T3S1 T3S2 (cid:21)
we conclude that
Since
D(S1T1) + D(S2T3) = D(T1S1).
tr(D(S1T1)) = tr(S1T1) = tr(T1S1) = tr(D(T1S1)),
we obtain that tr(D(S2T3)) = 0, so that D(S2T3) = 0. Since S is ideal-irreducible, we can
choose an atom a ∈ B2 and an operator T ∈ S such that f = T3a is a nonzero positive
vector in B3.
If ϕa is the linear functional associated to the atom a, then ϕa(Sf ) =
ϕa(S2T3a) + ϕa(S4T3a) = 0, as D(S2T3) = 0 and S4T3a ⊥ a. This is a contradiction with
Proposition 1.1.
The case when B3 = {0} and X = 0 can be handled similarly. Hence, it remains to
consider the case when X and Y are not both zero. We consider only the case X 6= 0,
since the case Y 6= 0 is similar. Let
be an arbitrary operator. Then
S1 S2 S3
S4 S5 S6
S7 S8 S9
∈ S
S =
ES =
SE =
XS4 + XY S7
∗
0
∗
0
0 S4X + S5
0
∗
0
∗
∗
S5 + Y S8 ∗
0
S7XY + S8Y .
∗
∗
and
Since the operator ES is ideal-triangularizable, its (1, 1) block XS4 + XY S7 is ideal-
triangularizable as well, and so is the operator XS4. The equality D(SE) = D(ES)
implies that D(XS4) = 0, and so the operator XS4 is quasinilpotent by Theorem 4.2.
Since σ(XS4)\{0} = σ(S4X)\{0}, S4X is a nilpotent operator on a finite-dimensional
Banach lattice B2. Since the operator SE is ideal-triangularizable, its (2, 2) block S4X +S5
is ideal-triangularizable as well, and so is S4X. It follows that D(S4X) = 0. Since X 6= 0,
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY6
13
there is an atom a ∈ B2 such that f = Xa is a nonzero vector in B1. The vectors S1f
and S7f are disjoint with the atom a, and so ϕa(Sf ) = ϕa(S4f ) = ϕa(S4Xa) = 0, as
D(S4X) = 0. This is again a contradiction with Proposition 1.1.
Case 2: The geometric multiplicity of the eigenvalue 1 ∈ σ(A) is smaller than its
algebraic multiplicity. Then, by the Riesz decomposition theorem, the Banach space L
can be decomposed into a direct sum of a finite-dimensional subspace L1 and a (possibly
zero) closed subspace L2 such that A has the block-diagonal form
A =(cid:20) I + N 0
C (cid:21) ,
0
where N is a nilpotent operator on L1 of the nilpotency index k ≥ 2, and r(C) < 1
provided L2 is nonzero. Now
lim
m→∞
Am
k−1(cid:1) =(cid:20) N k−1 0
(cid:0) m
0
0 (cid:21) ∈ S.
This proves that the semigroup S contains a square-zero operator M 6= 0. Let N (M) and
R(M) denote the absolute kernel and the range ideal of the operator M, respectively. We
claim that R(M) ⊆ N (M). If y is a vector in R(M), then there exist positive vectors
x1, . . . , xn ∈ E and positive scalars λ1, . . . , λn such that
y ≤
Then we have
λjMxj.
0 ≤ My ≤
λjM 2xj = 0,
nXj=1
nXj=1
which proves the claim. Now, if z ∈ N (M)d then Mz ∈ R(M) ⊆ N (M). This shows
that, with respect to the band decomposition L = N (M) ⊕ N (M)d, we have
Let S be an arbitrary operator in S, and let
0
M =(cid:20) 0 X
0 (cid:21) .
S =(cid:20) S1 S2
S3 S4 (cid:21)
be its block operator matrix with respect to the band decomposition L = N (M)⊕N (M)d.
Then
MS =(cid:20) XS3 XS4
0
0
(cid:21)
and
SM =(cid:20) 0 S1X
0 S3X (cid:21) .
14
ROMAN DRNOVSEK, MARKO KANDI ´C
Since D(MS) = D(SM), we have D(XS3) = 0 and D(S3X) = 0. Since MS and
SM are ideal-triangularizable operators, the operators XS3 and S3X are also ideal-
triangularizable. Theorem 4.2 implies that XS3 and S3X are both quasinilpotent, so
that MS and SM are quasinilpotent as well.
Let J be the semigroup ideal in S generated by the operator M. Pick any S ∈ J .
Then S is a product of operators from S, and the operator M appears in this product
at least once. If M appears either at the beginning or at the end of this product, then
S is quasinilpotent by the observation above. Otherwise there exist S1 and S2 in S
such that S = S1MS2. By the well known equality for the spectral radius, we have
r(S) = r(S1MS2) = r(MS2S1) = 0. This implies that J consists of quasinilpotent
compact operators, so that it is ideal-triangularizable by Theorem 4.3. This is again a
contradiction with Proposition 1.1.
We have shown that the semigroup S is ideal-reducible. We will finish the proof by
applying the Ideal-triangularization lemma. The property D(ST ) = D(T S) is inherited by
ideal-quotients. Indeed, let J be a closed ideal in L that is invariant under the semigroup
S. Let SJ denote the restriction of the operator S to J, and let PJ be the band projection
onto J. Then we have
Pa(SJ)(T J )Pa = Xa∈A∩J
Pa(ST )JPa =
D(SJT J) = Xa∈A∩J
= Xa∈A∩J
PaPJST PJ Pa = PJ Xa∈A
PaST Pa! PJ = PJ D(ST )PJ.
This implies that
D(SJT J) = PJ D(ST )PJ = PJ D(T S)PJ = D(T JSJ).
This proves that the property of zero diagonals of commutators of operators from S is
inherited to restrictions on invariant closed ideals. Since the induced operator TJ of the
operator T on the quotient Banach lattice E/J can be identified with the compression of
the operator T to J d, we can (similarly as above) prove that the property that the diagonal
of every commutator of operators from S is zero is inherited by induced operators on the
quotient Banach lattice E/J. Since ideal-triangularizability is inherited by ideal-quotients
by [11, Proposition 2.3], we finish the proof by applying the Ideal-triangularization lemma.
(cid:3)
In the finite-dimensional case Theorem 4.5 can be stated as follows.
FROM LOCAL TO GLOBAL IDEAL-TRIANGULARIZABILITY7
15
Corollary 4.6. Let S be a semigroup of nonnegative n × n matrices. Suppose that for
every matrix S ∈ S there exists a permutation matrix PS (depending on S) such that the
matrix PSSP −1
there exists a permutation matrix P such that every matrix in the semigroup P SP −1 is
is upper triangular. If D(ST ) = D(T S) for every pair S, T ∈ S, then
S
upper triangular.
The following example which was already introduced in [7] shows that Theorem 4.5 and
Corollary 4.6 do not hold for collections of ideal-triangularizable nonnegative matrices
with zero diagonals of commutators.
Example 4.7. Let e1, e2, . . ., en be the standard basis vectors of IRn, where n ≥ 3.
Define ideal-triangularizable nilpotent matrices by Ai = eieT
i+1 for i = 1, 2, . . . , n − 1, and
An = eneT
1 . Then the collection {A1, A2, . . . , An} has the property that D(AiAj) = 0 for
all 1 ≤ i, j ≤ n. We claim that the collection is not ideal-triangularizable. Assume the
contrary. Then the sum S = A1 + A2 + . . . + An is ideal-triangularizable. Since all the
diagonal entries of S are zero, S must be nilpotent which contradicts the fact that Sn = I.
The following example shows that Theorem 4.5 and Corollary 4.6 do not hold without
the assumption that operators are positive.
Example 4.8. Let
A =
0 0
0 0 −1
0
0 0
0 0
0
1 −1
1
0
0
and B =
0 0 0 0
0 0 0 0
1 1 0 0
1 1 0 0
.
Then A2 = B2 = AB = BA = 0, so that S = {0, A, B} is a semigroup of ideal-
triangularizable matrices such that D(S) = 0 for all S ∈ S. However, S is not ideal-
triangularizable, as the diagonal of the matrix
A + B =
is zero, but the matrix is not nilpotent.
0 0 1 1
0 0 1 1
1 1 0 0
1 1 0 0
Acknowledgments. This work was supported in part by the Slovenian Research Agency.
16
ROMAN DRNOVSEK, MARKO KANDI ´C
References
[1] Y. A. Abramovich, C. D. Aliprantis, An Invitation to Operator Theory. American Mathematical
Society, Providence (2002).
[2] C. D. Aliprantis, O. Burkinshaw, Positive operators, Reprint of the 1985 original, Springer, Dor-
drecht (2006).
[3] J. Bernik, L. W. Marcoux, H. Radjavi, Spectral conditions and band reducibility of operators, J.
Lond. Math. Soc. (2) 86 (2012), no. 1, 214 -- 234.
[4] R. Drnovsek, Once more on positive commutators, Studia Math. 211 (2012), no. 3, 241 -- 245.
[5] R. Drnovsek, Common invariant subspaces for collections of operators, Integral Equat. Oper. Th.
39 (2001), 253 -- 266.
[6] R. Drnovsek, Triangularizing semigroups of positive operators on an atomic normed Riesz spaces,
Proc. Edin. Math. Soc. 43 (2000), 43 -- 55.
[7] R. Drnovsek, M. Jesenko, M. Kandi´c, Positive commutators and collections of operators, Oper.
Matrices 6 (3) (2012), 535 -- 542.
[8] R. Drnovsek, M. Kandi´c, Ideal-triangularizability of semigroups of positive operators, Integral
Equat. Oper. Th. 64 (4) (2009), 539 -- 552.
[9] R. Drnovsek, M. Kandi´c, More on positive commutators, J. Math. Anal. Appl. 373 (2011), 580 --
584.
[10] R. Drnovsek, D. Kokol-Bukovsek, L. Livshits, G. MacDonald, M. Omladic, H. Radjavi, An irre-
ducible semigroup of non-negative square-zero operators, Integral Equat. Oper. Th. 42 (2002), no.
4, 449 -- 460.
[11] M. Kandi´c, Ideal-triangularizability of upward directed sets of positive operators, Ann. Funct. Anal.
2 (1) (2011), 206 -- 219.
[12] M. Kandi´c, Multiplicative coordinate functionals and ideal-triangularizability, Positivity,
doi:10.1007/s11117-013-0222-z.
[13] W. A. J. Luxemburg, A. C. Zaanen, Riesz spaces I, North-Holland, Amsterdam (1971).
[14] G. MacDonald, H. Radjavi, Standard triangularization of semigroups of non-negative operators,
J. Funct. Anal. 219 (2005), 161 -- 176.
[15] H. Radjavi, P. Rosenthal, Simultaneous Triangularization, Springer-Verlag, New York, 2000.
[16] H. H. Schaefer, Banach lattices, Springer-Verlag, Berlin-Heidelberg-New York, 1974.
[17] A. R. Schep, Positive diagonal and triangular operators, J. Oper. Theory 3 (1980), 165 -- 178.
[18] J. Voigt, The projection onto the center of operators in a Banach lattice, Math. Z. 199 (1988), no.
1, 115 -- 117.
[19] Y. Zhong, Functional positivity and invariant subspaces of semigroups of operators, Houston J.
Math. 19 (1993), 239 -- 262.
Roman Drnovsek, Marko Kandi´c :
Faculty of Mathematics and Physics
University of Ljubljana
Jadranska 19
1000 Ljubljana
Slovenia
E-mails : [email protected], [email protected]
|
1109.4793 | 1 | 1109 | 2011-09-22T12:59:46 | Structure constants of the Weyl calculus | [
"math.FA"
] | We find some explicit bounds on the ${\mathcal L}(L^2)$-norm of pseudo-differential operators with symbols defined by a metric on the phase space. In particular, we prove that this norm depends only on the "structure constants" of the metric and a fixed semi-norm of the symbol. Analogous statements are made for the Fefferman-Phong inequality. | math.FA | math |
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
WEN DENG
Abstract. We find some explicit bounds on the L(L2)-norm of pseudo-differential
operators with symbols defined by a metric on the phase space. In particular, we
prove that this norm depends only on the "structure constants" of the metric
and a fixed semi-norm of the symbol. Analogous statements are made for the
Fefferman-Phong inequality.
1. Introduction
The class of symbols Sm
1,0 consists of smooth functions a defined on the phase space
Rn × Rn such that for all multi-indices α, β,
(1.1)
(∂α
ξ ∂β
x a)(x, ξ) ≤ Cα,β(1 + ξ)m−α.
1,0. We have
Property A. If a is in S0
The best constants Cα,β in (1.1) are called the semi-norms of the symbol a in the
Fr´echet space Sm
1,0, then a(x, D) defines a bounded operator on L2(Rn).
One might ask some very natural questions: the operator norm ka(x, D)kL(L2(Rn))
is bounded by which constant? Is it a semi-norm of the symbol a? If yes, then which
semi-norm? Questions of the same type might be asked for the constant C in the
following inequality:
Property B (Fefferman-Phong inequality). If a is a non-negative symbol
belonging to S2
1,0, then there exists C > 0 such that, for all u ∈ S(Rn),
(1.2)
Reha(x, D)u, uiL2(Rn) + Ckuk2
L2(Rn) ≥ 0.
We can pose similar questions in many other examples of classes of symbols, such
as the semi-classical symbols, Shubin's class, etc. As a particular example, the class
Σm, defined as the set of smooth functions a on Rn × Rn × R+ such that for all
multi-indices α, β,
(1.3)
∀x, ξ ∈ Rn, τ ∈ R+,
(∂α
ξ ∂β
x a)(x, ξ, τ ) ≤ Cα,β(1 + ξ + τ )m−α,
is useful for Carleman estimates. One would like to check the Property A and
Property B independent of the parameter τ .
Several authors like Bony [1], Boulkhemair [3], Lerner-Morimoto [7], have al-
ready considered these questions and they were able to identify the constants. The
constants in Properties A, B are always a constant Cn times a semi-norm of the
symbol, whose order depends only on the dimension n. Although the problem is
well-understood for a single class of pseudo-differential calculus, including the class
S(m, g) developed by Hormander, we want to address a more general and useful
Date: July 9, 2018.
2000 Mathematics Subject Classification.
Key words and phrases. Weyl calculus, phase space, Fefferman-Phong inequality.
1
2
WEN DENG
question, having in mind the class Σm depending on the non-compact parameter
τ ≥ 0 which is defined in (1.3) and is useful for Carleman estimates.
In this paper, we consider the Weyl quantization for pseudodifferential operators
and we choose the framework with a metric g on the phase space. The metric g is
assumed to be admissible, that is slowly varying, satisfying the uncertainty principle
and is temperate (see Definition 2.1, 2.6 below). The so-called structure constants
of g are closely related to these properties. We can define very general classes of
symbols S(m, g) attached to the metric g and a g-admissible weight m (see Definition
2.3) and we have an effective symbolic calculus. The following results are classical:
(see [5, chapter 18], [6, chapter 2])
(1.4)
(1.5)
L2-boundedness:
Fefferman-Phong:
a ∈ S(1, g)
=⇒ kawkL(L2(Rn)) ≤ C,
a ∈ S(λ2
g, g), a ≥ 0 =⇒ aw + C ≥ 0.
The question that we would like to address is the following: what happens if we
change the metric g but keep the same structure constants?
We intend to show that the constants involved in (1.4), (1.5) depend only on the
structure constants of the metric g and a fixed semi-norm of a. Since it may happen
that the metric g depends on a non-compact parameter with uniform structure
the class Σm), this fact is useful explicitly or implicitly in many
constants (e.g.
examples where these metrics are used and it seems useful to rely on a more stable
argument than referring to "inspection of the proofs".
Remark. An abstract functional analysis argument does not seem to work. Our
method is to follow the proofs, by carefully computing all the constants.
2. Metric on the phase space
In this section, we introduce the definitions of the admissible metric and exhibit
its properties. We use the Weyl quantization which associates to a symbol a the
operator aw defined by
(2.1)
(awu)(x) = ZZ e2iπ(x−y)·ξa(
x + y
2
, ξ)u(y)dydξ.
Consider the symplectic space R2n equipped with the symplectic form σ = Pn
dxj. Given a positive-definite quadratic form Γ on R2n, we define
j=1 dξj∧
(2.2)
Γσ(T ) = sup
Γ(Y )=1
σ(T, Y )2,
which is also a positive-definite quadratic form. Let g be a measurable map from
R2n into the cone of positive-definite quadratic forms on R2n, i.e. for each X ∈ R2n,
gX is a positive definite quadratic form on R2n.
Definition 2.1 (Slowly varying metric). We say that g is a slowly varying metric
on R2n, if there exists C0 ≥ 1 such that for all X, Y, T ∈ R2n,
(2.3)
gX(X − Y ) ≤ C −1
0 =⇒ C −1
0 ≤
gX(T )
gY (T )
≤ C0.
Definition 2.2 (Slowly varying weight). Let g be a slowly varying metric on R2n.
A function m : R2n → (0, +∞) is called a g-slowly varying weight if there exists
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
3
µm ≥ 1 such that for all X, Y ∈ R2n,
(2.4)
gX(Y − X) ≤ µ−1
m =⇒ µ−1
m ≤
m(X)
m(Y )
≤ µm.
Definition 2.3 (Class of symbols). Let g be a slowly varying metric on R2n and m
be a g-slowly varying weight. The class of symbols S(m, g) is defined as the subset
of functions a ∈ C ∞(R2n) satisfying that for all k ∈ N, there exists Ck > 0 such that
for all X, T1, · · · , Tk ∈ R2n,
a(k)(X)(T1, · · · , Tk) ≤ Ckm(X) Y1≤j≤k
gX(Tj)1/2.
For a ∈ S(m, g), l ∈ N, we denote
(2.5)
kak(l)
S(m,g) = max
0≤k≤l
sup
a(k)(X)(T1, · · · , Tk)m(X)−1.
X,Tj∈R2n
gX (Tj )=1
The space S(m, g) equipped with the countable family of semi-norms (k · k(l)
is a Fr´echet space.
S(m,g))l∈N
For a slowly varying metric g on the phase space R2n, we can introduce some
partition of unity related to g. Define the g-ball near X ∈ R2n
(2.6)
UX,r = {Y, gX(X − Y ) ≤ r2},
we have the following theorem, which is Theorem 2.2.7 in [6].
Theorem 2.4 (Partition of unity). Let g be a slowly varying metric on R2n and
C0 > 0 given in (2.3). Then for all r ∈ (0, C −1/2
], there exists a family (ϕY )Y ∈R2n
0
of smooth functions supported in UY,r such that
(2.7)
(2.8)
∀k ∈ N,
sup
Y ∈R2n
kϕY k(k)
S(1,g) ≤ C(k, r, n, C0),
∀X ∈ R2n,
ZR2n
ϕY (X)gY 1/2dY = 1,
where C(k, r, n, C0) is a positive constant depending only on k, r, n, C0 and gY is
the determinant of gY with respect to the standard Euclidean norm.
Proof. As in the proof of Theorem 2.2.7 in [6], let χ0 ∈ C ∞
such that χ0(t) = 1 on t ≤ 1/2, χ0(t) = 0 on t ≥ 1. Define for r ∈ (0, C −1/2
0 (R+; [0, 1]) non-increasing
],
0
Since ωY (X) is supported in UY,r and χ0 is non-increasing, by (2.3) we have
gY 1/2dY.
ω(X, r) = ZR2n
χ0(cid:0)r−2gY (X − Y )(cid:1)
}
{z
gX1/2dY = ZR2n
χ0(cid:0)r−2C0gX(X − Y )(cid:1)C −n
=ωY (X)
0
ω(X, r) ≥ ZR2n
χ0(Z2)dZC −2n
0
r2n,
and an estimate from above of the same type, i.e. there exists a positive constant
C1 = C1(r, n, C0) such that
C −1
1 ≤ ω(X, r) ≤ C1.
4
WEN DENG
Now let us check the derivatives of ωY (X). Using the notation h, iY the inner-product
associated to gY , we have
ω′
Y (X)T = χ′
0(cid:0)r−2gY (X − Y )(cid:1)r−2hX − Y, T iY ,
and by induction, for k ≥ 1, T ∈ R2n, ω(k)
Y (X)T k is a finite sum of terms of type
(2.9)
cp,kχ(p)
0 (cid:0)r−2gY (X − Y )(cid:1)r−2phX − Y, T i2p−k
Y
gY (T )k−p,
where cp,k is a constant depending only on p, k and p ∈ [k/2, k] ∩ N. Since the
support of χ(p)
0 , the term (2.9) can be bounded
0
from above by
is included in [0, 1] and r2 ≤ C −1
cp,kkχ(p)
0 kL∞r−2p(r2)(2p−k)/2C k/2
0 gX(T )k/2,
so that for all k ≥ 1, ω(k)
S(1, g) and moreover,
Y (X)T k ≤ C(k, r, C0)gX(T )k/2. This implies that ωY is in
(2.10)
∀k ∈ N,
sup
Y ∈R2n
kωY k(k)
S(1,g) ≤ C(k, r, C0).
Now we choose a non-negative function χ1 ∈ C ∞
t ≤ 1, then
0 (R+; [0, 1]) such that χ1(t) = 1 on
ω(k)(X, r)T k = (cid:12)(cid:12)ZR2n
≤ sup
Y ∈R2n
ω(k)
Y (X)T kχ1(cid:0)r−2gY (X − Y )(cid:1)gY 1/2dY(cid:12)(cid:12)
S(1,g)gX(T )k/2ZR2n
kωY k(k)
χ1(cid:0)r−2gY (X − Y )(cid:1)gY 1/2dY
≤ C(k, r, n, C0)gX(T )k/2,
which implies that ω(·, r) is a symbol in S(1, g) with kω(·, r)k(k)
Since ω is bounded from below by C −1
S(1,g) ≤ C ′(k, r, n, C0).
1 , the function ω(·, r)−1 is also in S(1, g) and
(2.11)
We define
kω(·, r)−1k(k)
S(1,g) ≤ C ′′(k, r, n, C0).
ϕY (X) = ωY (X)ω(X, r)−1,
then the estimate (2.7) follows from (2.10), (2.11) and moreover, the family (ϕY )Y ∈R2n
satisfies the requirements of Theorem 2.4.
(cid:3)
A direct consequence of Theorem 2.4 is the following.
Proposition 2.5. Let g be a slowly varying metric on R2n and m be a g-slowly
varying weight. Let C0, µm be given in (2.3), (2.4) respectively. Let a be a symbol
in S(m, g). Then for all 0 < r ≤ min(C −1/2
, µ−1/2
m ),
0
a(X) = ZR2n
aY (X)gY 1/2dY,
where aY has support included in UY,r and
(2.12)
∀k ∈ N,
sup
Y ∈R2n
kaY k(k)
S(m(Y ),gY ) ≤ C(k, r, C0, n, µm)kak(k)
S(m,g).
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
5
Proof. Define aY (X) = a(X)ϕY (X). Since ϕY is supported in UY,r, we have, for
k ≥ 0, X ∈ UY,r, T ∈ R2n,
a(k)
Y (X)T k = (cid:12)(cid:12)(cid:12) X0≤l≤k
≤ X0≤l≤k
l(cid:19)a(l)(X)T l · ϕ(k−l)
(cid:18)k
Y
(X)T k−l(cid:12)(cid:12)(cid:12)
ck,lkak(l)
S(m,g)m(X)gX(T )l/2kϕY k(k−l)
S(1,g)gX(T )(k−l)/2
≤ C(k)kak(k)
≤ C(k)µmC k/2
S(m,g)kϕY k(k)
0 kak(k)
S(1,g)m(X)gX(T )k/2
S(m,g)kϕY k(k)
S(1,g)m(Y )gY (T )k/2,
which completes the proof.
(cid:3)
For two positive-definite quadratic forms Γ1, Γ2 on R2n, the harmonic mean Γ1∧Γ2
is defined by
(2.13)
2 )−1,
which is also a positive-definite quadratic form on R2n.
Γ1 ∧ Γ2 = 2(Γ−1
1 + Γ−1
Definition 2.6 (Admissible metric). We say that g is an admissible metric on R2n
if g is slowly varying (see Definition 2.1) and there exist C ′
0 > 0, N0 ∈ N such that
for all X, Y, T ∈ R2n,
(2.14)
uncertainty principle
(2.15)
temperance
X(T ),
gX(T ) ≤ gσ
gX(T ) ≤ C ′
where gσ is given by (2.2) and ∧ given by (2.13).
0gY (T )(cid:0)1 + (gσ
X ∧ gσ
Y )(X − Y )(cid:1)N0,
We may suppose C ′
0 = C0 in the sequel, where C0 is given in (2.3). Then the
constants (C0, N0) appearing in (2.3), (2.15) are called the structure constants of
the metric g.
Definition 2.7 (Admissible weight). Suppose that g is an admissible metric on
R2n. A function m : R2n → (0, +∞) is called a g-admissible weight if m is a g-
slowly varying weight (see Definition 2.2) and there exist µm > 0, νm ∈ N such that
for all X, Y ∈ R2n,
(2.16)
m(X) ≤ µmm(Y )(cid:0)1 + (gσ
X ∧ gσ
Y )(X − Y )(cid:1)νm.
The constants (µm, νm) appearing in (2.4), (2.16) are called the structure constants
of the g-admissible weight m.
Let g be an admissible metric on R2n. We define for X ∈ R2n,
(2.17)
λg(X) = inf
T 6=0(cid:16)gσ
X (T )
gX (T )(cid:17)1/2
.
Then the uncertainty principle (2.14) can be expressed by
gX ≤ λg(X)−2gσ
X,
λg(X) ≥ 1.
Lemma 2.8 ([6, Remark 2.2.17]). For any s ∈ R, λs
with structure constants (µλs
constants of the metric g (C0, N0).
g is an admissible weight,
g) in (2.4), (2.16) depending only on the structure
g , νλs
6
WEN DENG
Proof. We first verify that λs
T ∈ R2n, we have
g is a g-slowly varying weight. For gX(X − Y ) ≤ C −1
0 ,
C −1
0 gX(T ) ≤ gY (T ) ≤ C0gX(T ), C −1
0 gσ
X(T ) ≤ gσ
Y (T ) ≤ C0gσ
X(T ),
which implies
gσ
Y (T )
gY (T )
Taking the infimum with respect to T , we get
gσ
X(T )
gX(T )
C −2
≤
0
≤ C 2
0
gσ
X(T )
gX(T )
.
C −2
0 λg(X)2 ≤ λg(Y )2 ≤ C 2
0 λg(X)2,
so that λg is g-slowly varying with µλg = C0 and so is λs
check that λs
g is temperate. We have for all X, Y, T ∈ R2n,
g with µλs
g = C s
0 . Next we
which gives
gX(T ) ≥ C −1
X(T ) ≤ C0gσ
gσ
X ∧ gσ
X ∧ gσ
0 gY (T )(cid:0)1 + (gσ
Y (T )(cid:0)1 + (gσ
0 λg(Y )2(cid:0)1 + (gσ
Y )(X − Y )(cid:1)−N0,
Y )(X − Y )(cid:1)N0,
Y )(X − Y )(cid:1)2N0.
g with νλs
λg(X)2 ≤ C 2
X ∧ gσ
Thus λg is temperate with νλg = N0 and so is λs
the proof of Lemma 2.8.
g = sN0. This completes
(cid:3)
The composition a♯b of two symbols is defined by awbw = (a♯b)w and we have,
with the notations [X, Y ] = σ(X, Y ), D = (2iπ)−1∂,
(a♯b)(X) = 22nZZR2n×R2n
a(Y )b(Z)e−4iπ[X−Y,X−Z]dY dZ,
For a ∈ S(m1, g), b ∈ S(m2, g), we have the asymptotic expansion
(a♯b)(X) = exp(cid:0)iπ[DY , DZ](cid:1)(cid:0)a(Y )b(Z)(cid:1)Y =Z=X.
a♯b(x, ξ) = X0≤k<p
wk(a, b)(x, ξ) + rp(a, b)(x, ξ),
(2.18)
(2.19)
(2.20)
(2.21)
(2.22)
(2.23)
with wk(a, b) = 2−k Xα+β=k
(−1)β
α!β!
Dα
ξ ∂β
x a Dβ
ξ ∂α
x b ∈ S(m1m2λ−k
g , g),
rp(a, b)(X) = Rp(cid:0)a(X) ⊗ b(Y )(cid:1)Y =X ∈ S(m1m2λ−p
g , g),
Rp = Z 1
0
(1 − θ)p−1
(p − 1)!
exp
θ
4iπ
[∂X , ∂Y ]dθ(cid:16) 1
4iπ
[∂X , ∂Y ](cid:17)p
Notice w1(a, b) = 1
asymptotic (2.20) at p = 2 is
4iπ {a, b}, where {, } denotes the Poisson bracket, so that the
(2.24)
a♯b = ab +
1
4iπ
{a, b} + r2(a, b).
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
7
Definition 2.9 (The main distance function). Let g be an admissible metric on R2n.
Define the main distance function, for r > 0, X, Y ∈ R2n,
(2.25)
δr(X, Y ) = 1 + (gσ
X ∧ gσ
Y )(UX,r − UY,r),
where UX,r is given in (2.6) and
g(U − V ) = inf
X∈U,Y ∈V
g(X − Y ).
Lemma 2.10 ([6, Lemma 2.2.24], Integrability of δr). Let g be an admissible met-
ric with structure constants (C0, N0). Then there exist positive constants N1 =
N1(n, C0, N0), C = C(n, C0, N0) such that for all r ∈ (0, C −1/2
],
0
(2.26)
X∈R2nZR2n
sup
δr(X, Y )−N1gY 1/2dY ≤ C < +∞,
Proof. Suppose r ≤ C −1/2
Y ′ ∈ UY,r, T ∈ R2n, we have
0
. Using the slowness and temperance of g, for X ′ ∈ UX,r,
(gσ
X ∧ gσ
Y )(T ) ≥ C −1
0 (gσ
X ′ ∧ gσ
Y ′)(T ) ≥ C −2
≥ C −3
0 gσ
0 gσ
≥ C −3−N0
0
X ′(T )(cid:0)1 + (gσ
X(T )(cid:0)1 + C0(gσ
X (T )(cid:0)1 + (gσ
X ′ ∧ gσ
X ∧ gσ
X ∧ gσ
Y ′)(X ′ − Y ′)(cid:1)−N0
Y )(X ′ − Y ′)(cid:1)−N0
Y )(X ′ − Y ′)(cid:1)−N0.
gσ
Taking the infimum in X ′ ∈ UX,r, Y ′ ∈ UY,r, we get
(2.27)
We have also
X(T ) ≤ C 3+N0
gσ
0
δr(X, Y )N0(gσ
X ∧ gσ
Y )(T ).
gX(T )
gY (T )
≤ C 2
0
gX ′(T )
gY ′(T )
≤ C 3
≤ C 3
X ′ ∧ gσ
0(cid:0)1 + (gσ
0(cid:0)1 + C0(gσ
(cid:0)1 + (gσ
0
Y ′)(X ′ − Y ′)(cid:1)N0
Y )(X ′ − Y ′)(cid:1)N0
Y )(X ′ − Y ′)(cid:1)N0.
X ∧ gσ
X ∧ gσ
≤ C 3+N0
By taking the infimum in X ′, Y ′, we get the following inequality
(2.28)
Then
gX(T )
gY (T )
≤ C 3+N0
0
δr(X, Y )N0.
1+gX(X − Y ) ≤ 1 + 3gX(X − X ′) + 3gX(X ′ − Y ′) + 3gX(Y ′ − Y )
0
≤ 3C 3+N0
≤ 3C 3+N0
≤ 9C 6+2N0
0
0
δr(X, Y )N0(cid:0)1 + gX(X − X ′) + gX(X ′ − Y ′) + gY (Y ′ − Y )(cid:1) by (2.28)
δr(X, Y )N0(cid:0)1 + 2r2 + gσ
δr(X, Y )2N0(cid:0)1 + (gσ
X(X ′ − Y ′)(cid:1)
Y )(X ′ − Y ′)(cid:1) by (2.27),
X ∧ gσ
so that 1 + gX(X − Y ) ≤ 9C 6+2N0
0
δr(X, Y )2N0+1. In the other hand, we have
gY 1/2
gX1/2 ≤ C n(3+N0)
0
δr(X, Y )nN0,
8
WEN DENG
so that for N1 = nN0 + (n + 1)(2N0 + 1) > 0,
ZR2n
δr(X, Y )−N1gY 1/2dY ≤ C(n, C0, N0)ZR2n
δr(X, Y )−N1+nN0gX1/2dY
≤ C ′(n, C0, N0)ZR2n(cid:0)1 + gX (X − Y )(cid:1)−(n+1)
= C ′(n, C0, N0)ZR2n
(1 + Z2)−(n+1)dZ < +∞.
gX1/2dY
The proof of the lemma is complete.
(cid:3)
3. L2-boundedness
In this section, we prove the L2-boundedness of pseudo-differential operators with
symbol in S(1, g) and make precise the operator norms.
3.1. The constant metric case.
Proposition 3.1. Suppose that g is a positive-definite quadratic form (constant
metric) on R2n with g ≤ gσ. Then there exists a constant C(n) > 0 depending only
on the dimension n such that for all a ∈ S(1, g),
kawkL(L2(Rn)) ≤ C(n)kak(2n+1)
S(1,g) .
Proof. Since g is a constant metric, according to Lemma 4.4.25 in [6], there exist
symplectic coordinates (x, ξ) such that
g = X1≤j≤n
λ−1
j (dxj2 + dξj2),
gσ = X1≤j≤n
λj(dxj2 + dξj2),
with λj > 0. g ≤ gσ is expressed as
min
1≤j≤n
λj ≥ 1.
As a result, we have g ≤ dx2 + dξ2 := Γ0, which implies S(1, g) ⊂ S(1, Γ0) and
for all a ∈ S(1, g),
(3.1)
∀l ∈ N,
kak(l)
S(1,Γ0) ≤ kak(l)
S(1,g).
By Theorem 1.1.4 in [6] and aw = (J 1/2a)(x, D), where J t is introduced in Lemma
4.1.2 in [6], we obtain that
kawkL(L2(Rn)) ≤ C(n)kak(2n+1)
S(1,Γ0),
where C(n) depends only on n. Together with (3.1), we complete the proof of the
proposition.
(cid:3)
3.2. The general case.
Theorem 3.2. Let g be an admissible metric on R2n with structure constants (C0, N0)
(see Definition 2.6). Then there exist C = C(n, C0, N0) > 0 and l = l(n, C0, N0) ∈ N
such that for all a ∈ S(1, g) (see Definition 2.3),
kawkL(L2(Rn)) ≤ Ckak(l)
S(1,g).
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
9
Proof. Using the partition in Proposition 2.5, we write
aw = ZR2n
aw
Y gY 1/2dY,
where aY is supported in UY,r and satisfies (2.12). By Proposition 3.1, we have
supY kaw
S(1,g) < +∞. The following lemma is useful.
Y kL(L2(Rn)) ≤ C(r, n, C0, N0)kak(2n+1)
Lemma 3.3 (Cotlar). Let H be a Hilbert space and (Ω, A, ν) a measured space such
that ν is a σ-finite positive measure. Let (Ay)y∈Ω be a measurable family of bounded
operators on H such that
y∈ΩZΩ
sup
kA∗
yAzk1/2
L(H)dν(z) ≤ M,
y∈ΩZΩ
sup
kAyA∗
zk1/2
L(H)dν(z) ≤ M.
Then for all u ∈ H, we have
ZZΩ×Ω
hAyu, AzuiHdν(y)dν(z) ≤ M 2kuk2
H,
which implies the strong convergence of A = RΩ Aydν(y) and kAkL(H) ≤ M.
In order to apply Cotlar's lemma, we should estimate k¯aw
Y aw
ZkL(L2(Rn)), i.e. a
semi-norm of ¯aY ♯aZ in S(1, gY + gZ). Indeed, the following estimate holds.
Lemma 3.4. Let g, aY be as above. For any k, N ∈ N, there exist C = C(k, N, n) >
0, l = l(k, N, n) ∈ N such that
(3.2)
k¯aY ♯aZk(k)
S(1,gY +gZ ) ≤ Ck¯aY k(l)
S(1,gY )kaZk(l)
S(1,gZ )δr(Y, Z)−N .
We use some biconfinement estimates, which can be found in [6, section 2.3], to
prove Lemma 3.4.
Definition 3.5 (Confined symbols). Let g be a positive-definite quadratic form on
R2n such that g ≤ gσ. Let a be a smooth function on R2n and U ⊂ R2n. We say
that a is g-confined in U, if for all k, N ∈ N, there exits Ck,N > 0 such that for all
X, T ∈ R2n,
We denote
(3.3)
(3.4)
a(k)(X)T k ≤ Ck,N g(T )k/2(cid:0)1 + gσ(X − U)(cid:1)−N/2
.
kak(k,N )
g,U =
sup
X,T ∈R2n,g(T )=1
a(k)(X)T k(cid:0)1 + gσ(X − U)(cid:1)N/2
,
and
a(l)
g,U = max
k≤l
kak(k,l)
g,U .
Theorem 3.6 ([6, Theorem 2.3.2], biconfinement estimate). Let g1, g2 be two positive-
definite quadratic forms on R2n such that gj ≤ gσ
j . Let aj, j = 1, 2 be gj-confined in
Uj, a gj-ball of radius ≤ 1. Then for all k, N ∈ N, for all X, T ∈ R2n,
(3.5)
(a1♯a2)(k)(X)T k ≤ Ak,N (g1+g2)(T )k/2(cid:16)1+(gσ
1 ∧gσ
2 )(X−U1)+(gσ
1 ∧gσ
2 )(X−U2)(cid:17)−N/2
,
with Ak,N = γ(k, N, n)a1(l)
g1,U1a2(l)
g2,U2, l = 2n + 1 + k + N.
Now we begin the proof of Lemma 3.4.
10
WEN DENG
Proof of Lemma 3.4. The symbol aY is gY -confined in UY,r, since aY is supported in
the gY -ball UY,r. Moreover, we have
∀k, N ∈ N,
kaY k(k,N )
gY ,UY,r
=
sup
a(k)(X)T k,
X∈UY,r,T ∈R2n
gY (T )=1
∀l ∈ N,
aY (l)
gY ,UY,r
= max
k≤l
kaY k(k,l)
gY ,UY,r
= kaY k(l)
S(1,gY ).
Applying (3.5) to ¯aY ♯aZ and using the triangular inequality
(gσ
Y ∧ gσ
Z)(X − UY,r) + (gσ
Y ∧ gσ
Z)(X − UZ,r) ≥
1
2
we get
(gσ
Y ∧ gσ
Z)(UY,r − UZ,r),
(¯aY ♯aZ)(k)(X)T k ≤ γ(k, N, n)k¯aY k(l)
×(cid:0)1 +
1
2
(gσ
Y ∧ gσ
S(1,gZ )(gY + gZ)(T )k/2
S(1,gY )kaZk(l)
Z)(UY,r − UZ,r)(cid:1)−N/2
.
Using the definition of the distance δr, we complete the proof of Lemma 3.4.
(cid:3)
End of the proof of Theorem 3.2. Now by Proposition 3.1, Lemma 3.4 and the
estimate (2.12), we obtain that for any N > 0, there exists l = l(N, n) ∈ N such
that
k¯aw
Y aw
ZkL(L2(Rn)) ≤ C(n)k¯aw
Y aw
Zk(2n+1)
S(1,gY +gZ )
≤ C(N, n)k¯aY k(l)
S(1,gY )kaZk(l)
S(1,gZ )δr(Y, Z)−N
≤ C(N, n, C0)(cid:0)kak(l)
S(1,g)(cid:1)2δr(Y, Z)−N
The same inequality holds for aY ♯¯aZ. Choose N = 2N1, where N1 is given in (2.26),
so that
max{k¯aw
Y aw
Zk1/2
L(L2(Rn)), kaw
Y ¯aw
Zk1/2
L(L2(Rn))} ≤ Ckak(l)
S(1,g)δr(Y, Z)−N1,
where C = C(n, C0, N1) > 0, l = l(n, N1) ∈ N. Then together with Lemma 2.10, the
assumptions of Cotlar's lemma are fulfilled with M = Ckak(l)
S(1,g), and this completes
the proof of Theorem 3.2.
(cid:3)
4. Fefferman-Phong inequality
In this section, we prove that the constant in the Fefferman-Phong inequality
depends only on the structure constants of the metric and a fixed semi-norm of the
symbol.
Theorem 4.1 (Fefferman-Phong inequality). Let g be an admissible metric on R2n
with structure constants (C0, N0) (see Definition 2.6). Let a be a non-negative symbol
in S(λ2
g, g) (see Definition 2.3 and (2.17)). Then the operator aw on L2(Rn) is
semi-bounded from below. More precisely, there exist l = l(n, C0, N0) ∈ N, C =
C(n, C0, N0) > 0 such that
(4.1)
aw + Ckak(l)
g ,g) ≥ 0.
S(λ2
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
11
4.1. The constant metric case. For the constant metric case, we use the results
of Sjostrand and refer the readers to [6, page 116] for the detailed proof.
Let 1 = Pj∈Z2n χ0(X −j) be a partition of unity, χ0 ∈ C ∞
χ0(X − j).
c (R2n). Denote χj(X) =
Proposition 4.2 ([6, Proposition 2.5.6]). Suppose a ∈ S(R2n). We say that a
belongs to the class A if ωa ∈ L1(R2n), with ωa(Ξ) = supj∈Z2n F (χja)(Ξ), where F
is the Fourier transform. We have
S0
0,0 ⊂ S0,0;2n+1 ⊂ A ⊂ C 0(R2n) ∩ L∞(R2n),
0,0 = C ∞
where S0
b (R2n) is the space of C ∞ functions on R2n which are bounded as
well as all their derivatives, S0,0;2n+1 is the set of functions defined on R2n such
that (∂α
x a)(x, ξ) ≤ Cαβ for α + β ≤ 2n + 1. A is a Banach algebra for the
multiplication with the norm kakA = kωakL1(R2n).
ξ ∂β
Theorem 4.3 ([6, Theorem 2.5.10]). For all non-negative function a defined on R2n
satisfying a(4) ∈ A, then the operator aw is semi-bounded from below. More precisely,
where Cn depends only on the dimension n.
aw + Cnka(4)kA ≥ 0,
4.2. Proof of Theorem 4.1. We shall use the partition of unity (ϕY )Y ∈R2n given
in Theorem 2.4. Let (ψY )Y ∈R2n be a family of real-valued functions supported in
UY,2r, equal to 1 on UY,r and
(4.2)
kψY k(k)
S(1,g) ≤ C(k, r, C0).
sup
Y ∈R2n
Indeed, with the same notations as in the proof of Theorem 2.4, the function
ψY (X) = χ0(cid:0) 1
we write
2r−2gY (X − Y )(cid:1) satisfies the requirements. Then with aY = ϕY a,
(4.3)
ψY ♯aY ♯ψY = aY + rY .
Lemma 4.4 (Estimate for rY ). For all k, N ∈ N, there exist C = C(k, N, C0) > 0,
l = l(k, N, C0) ∈ N such that for all X ∈ R2n, T ∈ R2n with gY (T ) ≤ 1,
r(k)
Y (X)T k ≤ CkaY k(l)
S(λg (Y )2,gY )(cid:0)1 + gσ
Y (X − UY,2r)(cid:1)−N .
Moreover, there exist C1 = C1(n, C0, N0) > 0, l1 = l1(n, C0, N0) ∈ N such that
(4.4)
(4.5)
rw
(cid:13)(cid:13)ZR2n
Y gY 1/2dY(cid:13)(cid:13)L(L2(Rn)) ≤ C1kak(l1)
S(λ2
g,g),
To prove Lemma 4.4, we use the biconfinement estimate for the remainders, the
proof of which can be found in [6, section 2.3].
Theorem 4.5 ([6, Theorem 2.3.4], biconfinement estimate). Let g1, g2 be two positive-
definite quadratic forms on R2n with gj ≤ gσ
j . Let aj, j = 1, 2 be gj-confined in Uj,
a gj-ball of radius ≤ 1. Recall (2.20)
rp(a1, a2)(X) = (a1♯a2)(X) − X0≤k<p
1
j!(cid:0)iπ[DX1, DX2](cid:1)j(cid:0)a1(X1)a2(X2)(cid:1)X1=X2=X.
12
WEN DENG
Then for all k, l, p ∈ N, for all X, T ∈ R2n, we have
(cid:0)rp(a1, a2)(cid:1)(k)
×(cid:16)1 + (gσ
(4.6)
(X)T k ≤ Ak,N,p(g1 + g2)(T )k/2Λ−p
1,2
1 ∧ gσ
2 )(X − U1) + (gσ
1 ∧ gσ
2 )(X − U2)(cid:17)−N/2
with Ak,N,p = C(k, N, p, n)a1(l)
g2,U2, l = 2n + 1 + k + p + N and
g1,U1a2(l)
g2(T )(cid:17)1/2
1 (T )
T ∈R2n,T 6=0(cid:16)gσ
inf
=
T ∈R2n,T 6=0(cid:16)gσ
2 (T )
g1(T )(cid:17)1/2
inf
.
(4.7)
Λ1,2 =
Now we use Theorem 4.5 to prove Lemma 4.4.
Proof of Lemma 4.4. By the asymptotic formula (2.24), we have
ψY ♯aY = aY +
1
4iπ
{ψY , aY }
+r2(ψY , aY ),
=0
{z
}
since ψY = 1 on the support of aY . The symbol ψY is gY -confined in UY,2r and aY
is gY -confined in UY,r, and moreover, we have
∀l ∈ N,
ψY (l)
gY ,UY,2r
= kψY k(l)
S(1,gY ),
aY (l)
gY ,UY,r
= λg(Y )2kaY k(l)
S(λg(Y )2,gY ).
Applying (4.6) to r2(ψY , aY ), we have for all k, N ∈ N, there exist C(k, N, n) > 0,
l(k, N, n) ∈ N such that for all X, T ∈ R2n,
(cid:0)r2(ψY , aY )(cid:1)(k)(X)T k
≤ C(k, N, n)ψY (l)
≤ C(k, N, n)kψY k(l)
(4.8)
gY ,UY,2r
aY (l)
S(1,gY )kaY k(l)
gY ,UY,r
gY (T )k/2Λ−2
1,2(cid:0)1 + gσ
S(λg(Y )2,gY )gY (T )k/2(cid:0)1 + gσ
Y (X − UY,2r)(cid:1)−N
Y (X − UY,2r)(cid:1)−N
,
noticing here Λ1,2 defined in (4.7) is equal to λg(Y ). An analogous estimate as (4.8)
holds for r2(aY , ψY ). In our case, we write rY , which is defined in (4.3),
rY = (ψY ♯aY − aY )♯ψY + (aY ♯ψY − aY )
= r2(ψY , aY )♯ψY + r2(aY , ψY ).
Then the estimate (4.4) follows from (4.8) and (3.5). Furthermore, for any k, N ∈ N,
there exist C = C(k, N, n, C0) > 0, l = l(k, N, n, C0) ∈ N such that
k¯rY ♯rZk(k)
S(1,gY +gZ ) ≤ CkaY k(l)
S(λg(Y )2,gY )kaZk(l)
S(λg(Z)2,gZ )δ2r(Y, Z)−N .
Thus we can apply Cotlar's lemma and get the estimate (4.5).
(cid:3)
Lemma 4.6 (Estimate for ψY ). For all k, N ∈ N, there exist C = C(k, N, C0) > 0,
l = l(k, N, C0) ∈ N such that for all X ∈ R2n, T ∈ R2n with gY (T ) ≤ 1,
(4.9)
(ψY ♯ψY )(k)(X)T k ≤ C(cid:0)kψY k(l)
S(1,gY )(cid:1)2(cid:0)1 + gσ
Y (X − UY,2r)(cid:1)−N
.
Moreover, there exists C2 = C2(n, C0, N0) > 0 such that
(4.10)
kZ ψw
Y ψw
Y gY 1/2dY kL(L2(Rn)) ≤ C2.
STRUCTURE CONSTANTS OF THE WEYL CALCULUS
13
Proof. The inequality (4.9) follows immediately from (3.5). And it follows from
(3.5), (4.2) and (4.9) that for all k, N ∈ N,
k(ψY ♯ψY )♯(ψZ♯ψZ)k(k)
S(1,gY +gZ) ≤ Cδ2r(Y, Z)−N ,
for some C = C(k, N, n, C0) > 0. Then by choosing N = 2N1 and using Cotlar's
lemma, we get the estimate (4.10).
(cid:3)
End of the proof of Theorem 4.1. The symbol aY is non-negative and uniformly in
S(λg(Y )2, gY ), so that we can apply the Fefferman-Phong inequality (Theorem 4.3)
for the constant metric gY to get
Y + C(n)kaY k(l(n))
aw
By Proposition 2.5 and Lemma 2.8, we have
S(λg (Y )2,gY ) ≥ 0.
kaY k(l(n))
S(λg(Y )2,gY ) ≤ C(n, C0, N0)kak(l(n))
S(λ2
g ,g),
so that
(4.11)
Y + C3kak(l(n))
aw
S(λ2
g,g) ≥ 0.
where C3 = C3(n, C0, N0) > 0, l(n) ∈ N are constants. Combining (4.3), (4.5),
(4.10) and (4.11), we obtain
aw = ZR2n
= ZR2n
aw
Y gY 1/2dY
ψw
Y aw
Y ψw
Y gY 1/2dY −ZR2n
g ,g)ZR2n
ψw
Y ψw
rw
Y gY 1/2dY
≥ −C3kak(l(n))
S(λ2
Y gY 1/2dY − C1kak(l1)
S(λ2
g,g)
≥ −Ckak(l)
g,g),
S(λ2
for some C = C(n, C0, N0) > 0 and l = l(n, C0, N0) ∈ N. The proof of Theorem 4.1
is complete.
(cid:3)
References
[1] Jean-Michel Bony, Sur l'in´egalit´e de Fefferman-Phong, Seminaire: ´Equations aux D´eriv´ees
Partielles, 1998 -- 1999, S´emin. ´Equ. D´eriv. Partielles, ´Ecole Polytech., Palaiseau, 1999, pp. Exp.
No. III, 16. MR 1721321 (2000i:35232)
[2] Jean-Michel Bony and Jean-Yves Chemin, Espaces fonctionnels associ´es au calcul de Weyl-
Hormander, Bull. Soc. Math. France 122 (1994), no. 1, 77 -- 118. MR 1259109 (95a:35152)
[3] Abdesslam Boulkhemair, On the Fefferman-Phong inequality, Ann. Inst. Fourier (Grenoble) 58
(2008), no. 4, 1093 -- 1115. MR 2427955 (2009h:35468)
[4] Nils Dencker, The resolution of the Nirenberg-Treves conjecture, Ann. of Math. (2) 163 (2006),
no. 2, 405 -- 444. MR 2199222 (2006i:35386)
[5] Lars Hormander, The analysis of linear partial differential operators. III, Classics in Math-
ematics, Springer, Berlin, 2007, Pseudo-differential operators, Reprint of the 1994 edition.
MR 2304165 (2007k:35006)
[6] Nicolas Lerner, Metrics on the phase space and non-selfadjoint pseudo-differential operators,
Pseudo-Differential Operators. Theory and Applications, vol. 3, Birkhauser Verlag, Basel, 2010.
MR 2599384
[7] Nicolas Lerner and Yoshinori Morimoto, On the Fefferman-Phong inequality and a Wiener-type
algebra of pseudodifferential operators, Publ. Res. Inst. Math. Sci. 43 (2007), no. 2, 329 -- 371.
MR 2341014 (2009b:47083)
14
WEN DENG
Wen Deng, Institut de Math´ematiques de Jussieu, Universit´e Pierre-et-Marie-
Curie (Paris 6), 4 place Jussieu, 75005 Paris, France.
E-mail address: [email protected]
URL: http://www.math.jussieu.fr/~wendeng/
|
1302.5790 | 1 | 1302 | 2013-02-23T12:09:47 | Crofton formulae and geodesic distance in hyperbolic spaces | [
"math.FA"
] | The geodesic distance between points in real hyperbolic space is a hypermetric, and hence is a kernel negative type. The proof given here uses an integral formula for geodesic distance, in terms of a measure on the space of hyperplanes. An analogous integral formula, involving the space of horospheres, is given for complex hyperbolic space.By contrast geodesic distance in a projective space is not of negative type. | math.FA | math |
CROFTON FORMULAE AND GEODESIC DISTANCE
IN HYPERBOLIC SPACES
GUYAN ROBERTSON
Abstract. The geodesic distance between points in real hyper-
bolic space is a hypermetric, and hence is a kernel negative type.
The proof given here uses an integral formula for geodesic distance,
in terms of a measure on the space of hyperplanes. An analogous
integral formula, involving the space of horospheres, is given for
complex hyperbolic space. By contrast geodesic distance in a pro-
jective space is not of negative type.
Introduction
Motivated by problems in harmonic analysis on homogeneous spaces,
J. Faraut and K. Harzallah showed in [FH] that the geodesic distance
d on a real or complex hyperbolic space of dimension n ≥ 2 is a kernel
of negative type. Equivalently, √d is a Hilbert space distance [HV:
Chapter 5]. On the other hand, in the real hyperbolic plane, there is
an explicit Crofton integral formula for the length L of a rectifiable
curve C in terms of a measure µ on the space of geodesics [S1, Section
3], [GS, Section 4.4]. The formula has the form R n(γ)dµ(γ) = 2L,
where n(γ) is the number of times the geodesic γ meets C and the
integral is over the space of all geodesics. (M.W. Crofton proved the
corresponding formula for the euclidean plane in 1868.)
This paper proves a Crofton formula for geodesic distances in real
hyperbolic space of dimension n, from which one obtains an explicit
geometric proof of the result of Faraut and Harzallah, in the real case.
In this case the measure µ is defined on the space of totally geodesic
submanifolds of codimension one. There is also a Crofton formula in
the complex case. However, the measure is defined on the space of
horospheres, and it is not clear how to deduce that d is of negative
type.
Date: July 28, 1997.
1991 Mathematics Subject Classification. Primary 22D10.
This research was supported by the Australian Research Council.
1
Typeset by AMS-LATEX.
2
G. ROBERTSON
Hyperbolic spaces H n
F can be defined over R, C, H and (in the case
n = 2) the octonions O. In each case the space is two point homoge-
neous: if d(x, y) = d(x′, y′) then there is an isometry g of H n
F such that
gx = x′, gy = y′ [He1, Chapter IX, Proposition 5.1, p.355]. It turns
out that this property, together with the existence of a certain invari-
ant measure is at the heart of the Crofton formulae. We note that the
noncompact riemannian manifolds which are two point homogeneous
are exactly the euclidean spaces and hyperbolic spaces [He1].
The Crofton formula that we prove for H n
R has the consequence that
the geodesic distance d is a hypermetric in the sense of [A]. It follows
R, endowed with the metric √d, embeds
that any finite subset of H n
isometrically in a euclidean sphere (Corollary 3.2).
Finally, we complete the investigation of Faraut and Harzallah [FH]
for symmetric spaces of rank one, by showing explicitly that the geo-
desic distance on a projective space is not of negative type, although
it can be expressed by means of a Crofton formula.
1. motivation and preliminaries
In the course of the proof of [RS, Proposition 1.4], a geometric ar-
gument was given to show that the euclidean distance d(x, y) between
points x, y ∈ Rn is a kernel of negative type. In retrospect, the basis of
that argument is seen to be the classical Crofton formula asserting that
d(x, y) equals the measure of the set of euclidean hyperplanes which
meet the line segment [xy]. The relevant measure is the appropriately
normalized measure on the space of hyperplanes which is invariant un-
der isometries of Rn. This measure lifts to a measure µ on the space of
half spaces of Rn. Then d(x, y) = µ(Sx△Sy), where Sx denotes the set
of half spaces which contain x. It is natural to try to extend this ar-
gument to the rank one symmetric spaces considered by J. Faraut and
K. Harzallah [FH]. It turns out that the only such spaces for which
the geodesic distance is a kernel of negative type are real or complex
hyperbolic spaces, and euclidean spheres.
Let F be one of the (skew-) fields R, C or H. Regard Fn+1 as a right
vector space over F and define a hermitian form on Fn+1 by means of
the formula
hz, wi = −z0w0 + z1w1 + · · · + znwn.
F is the image in the projective space P n
F of
The hyperbolic space H n
the set of negative vectors {z ∈ Fn+1 : hz, zi < 0}.
3
It is convenient to use the same notation for a negative vector x ∈
F . The geodesic distance between
Fn+1 and its equivalence class in P n
points x, y in X is given by cosh d(x, y) = hx,yi
(hx,xihy,yi)
.
1
2
The octionic hyperbolic plane H 2
O requires a more involved definition
[M].
2. real hyperbolic space
Denote by X the hyperbolic space H n
R [CG]. Then X is the image in
the projective space RPn of the set of negative vectors
(cid:8)x ∈ Rn+1 : hx, xi = −(x0)2 + (x1)2 + . . . (xn)2 < 0(cid:9) .
The orthogonal group G = O(1, n) is the subgroup of GL(n + 1, R)
which preserves the form h·,·i, and G acts isometrically and transi-
tively on X. The stabilizer in G of the point x0 = (1, 0, . . . , 0) ∈ X
is the compact group K = O(1) × O(n). Thus X is isomorphic to the
topological homogeneous space G/K. The hyperbolic space H n−1
em-
beds naturally into X = H n
R as the subspace S0 consisting of all points
with last coordinate equal to zero.
R
Every totally geodesic submanifold of codimension one in X is a
G -- translate of S0 [CG, Proposition 2.5.1]. It is convenient simply to
refer to such a submanifold as a hyperplane. The space S of all hyper-
planes may therefore be identified with the topological homogeneous
space G/G(S0), where G(S0) is the subgroup of G consisting of ele-
ments which leave S0 globally invariant. By [CG, Lemma 4.2.1], we
have G(S0) ∼= O(1, n − 1) × O(1). The groups G and G(S0) are both
unimodular [He1, Chapter X, Proposition 1.4]. (A direct proof is given
in [BP, Proposition C.4.11].) It follows [N, Chapter 3, p.140, Corollary
4] that there is a nonzero positive G-invariant measure µS on the space
S of hyperplanes.
We also consider the space H of half spaces in X. These are the G --
translates of the half -- space H0 consisting of points with last coordinate
positive. The group G(H0) ∼= O(1, n − 1) is unimodular and there is a
corresponding invariant measure µH on H = G/G(H0).
There is a natural double covering π : H → S and so by uniqueness of
the measures (up to a positive multiple) [N, Chapter 2, p.95, Corollary]
we may assume that µS = π∗ ◦ µH.
first prove the following Crofton formula.
Let [xy] denote the (unique) geodesic between points x, y ∈ X. We
4
G. ROBERTSON
Proposition 2.1. There is a constant k > 0 such that, if x, y ∈ X
then
µS {S ∈ S : S ∩ [xy] 6= ∅} = kd(x, y).
For the purposes of the proof, we introduce the notation
m(x, y) = µS {S ∈ S : S ∩ [xy] 6= ∅} .
Lemma 2.2. If x, y ∈ X then
{S ∈ S : S ∩ [xy] 6= ∅} is compact;
(a)
(b) m(x, y) < ∞.
Proof: (a) Let g, h ∈ G. We claim that the point gx0 lies on the
hyperplane hS0 if and only if gK ∩ hG(S0) 6= ∅. The crucial fact used
in the proof is that G(S0) acts transitively on S0. Therefore
gx0 ∈ hS0 ⇐⇒ h−1gx0 ∈ S0
⇐⇒ h−1gx0 = g0x0, for some g0 ∈ G(S0)
⇐⇒ g−1
0 h−1g ∈ K, for some g0 ∈ G(S0)
⇐⇒ h−1g = g0k0, for some g0 ∈ G(S0), k0 ∈ K
⇐⇒ gK ∩ hG(S0) 6= ∅.
This proves our assertion. Furthermore, since [xy] is compact, there
exists a compact subset J ⊂ G such that Jx0 = [xy], by [N, p.137,
Lemma 1]. A hyperplane hS0 meets [xy] if and only if JK∩hG(S0) 6= ∅,
that is h ∈ JKG(S0). The set of such hyperplanes hS0 is therefore
compact, being the image of the compact set JK under the quotient
map G → G/G(S0).
(b) This follows immediately from (a).
(cid:3)
Proof of Proposition 2.1. We must prove that m(x, y) =
kd(x, y) where k > 0 is constant. This is based on the following facts:
(a) X is two-point homogeneous with respect to the action of G; (b)
It follows that m(x, y) = m(x′, y′)
the measure µS is G-invariant.
whenever d(x, y) = d(x′, y′). Moreover, by considering a large number
of pairwise disjoint geodesic segments of equal length in some fixed
geodesic segment [ab], we see that m(x, y) → 0 as d(x, y) → 0.
In
particular m(x, x) = 0. If we divide [xy] into s segments [xixi+1] of
equal length (1 ≤ i ≤ s), then m(x, y) = sm(x1, x2). In this way, if
d(x′, y′) = qd(x, y) where q is rational, then m(x′, y′) = qm(x, y). Now
m(x, y) < ∞, by Lemma 2, and so by continuity there is a constant
k ≥ 0 such that m(x, y) = kd(x, y) for all x, y ∈ X. We must check
that k > 0. For this, it is enough to show that m(x, y) > 0 for some
5
x, y. Take x = (1, 0, 0. . . . , ǫ), y = (1, 0, 0. . . . ,−ǫ). Then the hyper-
plane S0 meets [xy] transversally at the interior point x0 = (1, 0, . . . , 0)
of the geodesic segment [xy]. There is an open neighbourhood eV of the
identity in G such that gS0 meets [xy] at an interior point for all g ∈ eV .
The image V of eV in G/G(S0) is an open set which is contained in the
set of all hyperplanes meeting [xy]. Since V has positive measure, this
implies that m(x, y) > 0, as required.
(cid:3)
Remark 2.3. Unlike the case n = 2 in [S1, Section 3], [GS, Section
4.4], the above proof of the Crofton formula in hyperbolic space is not
constructive. However it gives a geometric explanation of why the result
is true.
We now re-interpret Proposition 1, replacing µS by µH, the invariant
measure on the space of all half-spaces. This will allow us to use the
method of [RS] to show that the metric on X is of negative type. Given
x ∈ X, let Σx denote the set of half-spaces containing x.
Lemma 2.4. If x, y ∈ X then
µH(Σx △ Σy) = µS {S ∈ S : S ∩ [xy] 6= ∅} .
Proof: The double covering π : H → S is given by π(H) = ∂H,
the boundary of H. Since a hyperplane S is totally geodesic we have
either # ([xy] ∩ S) ≤ 1 or [xy] ⊆ S. Therefore
Σx △ Σy = π−1 {S ∈ S : S ∩ [xy] 6= ∅ and [xy] 6⊆ S} .
Now µS {S ∈ S : [xy] ⊆ S} = 0. The result follows, since µS = π∗ ◦
µH.
Corollary 2.5. If x, y ∈ X then d(x, y) = kµH(Σx △ Σy) where k > 0
is constant. Hence d is a kernel of negative type.
(cid:3)
Proof: The first assertion is immediate from Lemma 2.4. The
second assertion follows by embedding X into L2(H, µH) via x 7→ vx
where vx = χx − χx0, and χx is the characteristic function of Σx. Then
2 and so √d is a Hilbert space distance. (c.f.
µH(Σx △ Σy) = kvx − vyk2
[RS, Proposition 1.1].)
(cid:3)
Remark 2.6. Faraut and Harzallah show [FH] that the distance func-
tion on real or complex hyperbolic space is a kernel of negative type. See
[HV, Chapitre 6, Th´eor`eme 21]. The corresponding result for quater-
nionic hyperbolic space is false, because the group of isometries Sp(1, n)
has Kazhdan's property (T ) [FH, Th´eor`eme 6.4]. It would be interesting
to have a direct proof of this fact, avoiding the use of property (T ).
6
G. ROBERTSON
Remark 2.7. Corollary 2.5 is stronger than the result of Faraut and
Harzallah. It asserts that the distance d is a measure definite kernel,
in the sense of [RS], a concept which is in general strictly stronger than
that of negative type. See the next section.
Remark 2.8. Since G(H0) is a closed noncompact subgroup of G, it
follows from Moore's ergodicity Theorem [Z, Theorem 2.2.6] that any
lattice subgroup Γ of G acts ergodically on H = G/G(H0). Note that
the group Γ does not have property (T ). Fix an element x ∈ X then
µH(Σx△gΣx) = µH(Σx△Σgx) is a positive constant multiple of d(x, gx),
and hence unbounded. The action of Γ in [RS, Theorem 2.1] may
therefore be chosen to be ergodic. Whether one can find an ergodic
action with similar properties for all groups without property (T ) is an
open question.
Remark 2.9. Consider any Coxeter system (W, S), where W is a Cox-
eter group and S is the canonical generating set. Let ∆ denote the
associated Coxeter complex. By a result of J. Tits [B, Chap. IV, p.41],
the natural distance d(c, c′) between chambers c and c′ is equal to the
number of walls of ∆ separating c and c′. This is precisely the Crofton
formula for distance, relative to counting measure on the space of walls.
Let H denote the space of half spaces ("roots") of ∆ with counting
measure. If c is a chamber of ∆, let Hc be the set of roots containing
c. Then 2d(c, c′) = #(Hc △ Hc′). (See the proof of [HV, Proposition
6.14].) The argument proceeds as before, showing that d is a kernel of
negative type.
The same type of argument is applied in [BS, Proof of Theorem 3] to
a polygonal complex X, which is locally finite, simply connected and of
type (4, 4) or (6, 3). Let B denote the set consisting of all barycentres
of faces of X or of edges of X which are not adjacent to faces. An
unbounded negative definite kernel N of the above form is defined on
B. The kernel N is equivalent to the geodesic distance between elements
of B and is invariant under automorphisms of X. This is used in [BS]
to show that a properly discontinuous group of automorphisms of X
does not have property (T).
3. the hypermetric property for H n
R
A semimetric space (X, d) is said to be hypermetric if it satisfies the
following property: for each finite subset {x1, x2 . . . , xm} of X and inte-
gers {t1, t2 . . . , tm} such thatPm
i,j=1 titjd(xi, xj) ≤
0. The corresponding statement, with t1, t2 . . . , tm real numbers satis-
fying Pm
i=1 ti = 0, says that d is a kernel negative type. It is easy to
i=1 ti = 1, we havePm
7
see that if d has the hypermetric property then d is a kernel negative
type [A, 1.2].
Suppose that (W, µ) is a measure space and that a semimetric d on
a set X is defined by the formula d(x, y) = µ(Sx △ Sy), where for each
x ∈ X, Sx is a measurable set. It was proved in [K, Theorem 3.1] that
d is then a hypermetric. The next result is therefore an immediate
consequence of Corollary 2.5.
Corollary 3.1. The geodesic distance d on H n
R is a hypermetric.
Finite hypermetric spaces have been characterized up to isometry in
[A], and there is a detailed exposition of their properties in [DGL]. In
view of the fact that the space (H n
R, d) has constant negative curvature,
the next result may seem slightly surprising.
Corollary 3.2. Let {x1, x2 . . . , xm} be any finite subset of H n
with the metric √d. Then ({x1, x2 . . . , xm},√d) embeds isometrically
in a euclidean sphere in Rp, where p ≥ log2 m.
from [HV, Proposition 5.14] that ({x1, x2 . . . , xm},√d) embeds isomet-
Proof: Since d is of negative type (being hypermetric) it follows
rically in Rp, for some p. The fact that the image is contained in a
euclidean sphere is then a consequence of [A, Lemme 1.12]. The esti-
mate p ≥ log2 m is provided by [A, Proposition 1.18].
(cid:3)
4. complex hyperbolic space
R, endowed
C of CPn.
Complex hyperbolic space H n
C is constructed in a manner similar to
that of H n
R [CG]. The bihermitian form hz, wi = −z0w0 + z1w1 +· · · +
znwn on Cn defines a set of negative vectors in Cn+1, defined by the
condition hz, zi < 0, which projects to the subspace H n
One might try to mimic the arguments of the preceding section in the
complex case. The space H n
C is two point homogeneous with respect
to the isometry group U(1, n). Natural analogues of hyperplanes are
the equidistant hypersurfaces [GP, Section 4], also known as spinal
surfaces in [M]. These are, by definition, subspaces of the form S =
{z ∈ H n
C. There is a natural
invariant measure on the space of equidistant hypersurfaces, but we
cannot prove an analogue of Lemma 2.2, because the group G(S0) of
elements of U(1, n) which leave a given equidistant hypersurface S0
globally invariant does not act transitively on S0 [M, Section 3.2]. An
additional problem is that equidistant hypersurfaces are not totally
geodesic and a geodesic segment can meet an equidistant hypersurface
more than once, without being contained in the hypersurface. However,
C : d(z, z1) = d(z, z2)}, where z1, z2 ∈ H n
8
G. ROBERTSON
even in the complex case, it is possible to prove a Crofton formula if
we consider horospheres instead of equidistant hypersurfaces.
Horospheres in complex hyperbolic space are described geometri-
cally in [GP, Section 1]. From a group-theoretic point of view they
may be described as follows [He2, Chapter II.1]. Let G = U(1, n),
so that H n
C = G/K where K = U(1) × U(n). Let G = KAN be
the Iwasawa decomposition and M the centralizer of A in K. Thus
MAN is a minimal parabolic subgroup. Let x0 be the origin K in
G/K and ξ0 = Nx0 Then ξ0 is a horosphere and the subgroup of G
which maps ξ0 to itself equals MN. The homogeneous space G/MN
is the space of horocycles. Note that N is isomorphic to the complex
Heisenberg group [He2, Chapter 2, p.215]. The groups G and MN are
therefore unimodular [He1, Chapter X, Proposition 1.4], and there is
a G-invariant measure on G/MN. Since MN acts transitively on ξ0,
the set of horospheres which meet a geodesic segment [xy] is compact,
as in Lemma 2.2. Consider the expression m(x, y) =RG/M N n(h)dµ(h),
where n(h) is the number of times a horocycle h ∈ G/MN meets the
geodesic segment [xy]. Since H n
C is two-point homogeneous, and µ is
G-invariant, m(x, y) depends only on d(x, y). The same argument as in
the proof of Proposition 2.1 establishes the following Crofton formula.
Proposition 4.1. There is a constant k > 0 such that, for all x, y ∈
H n
C,
ZG/M N
n[x, y](h)dµ(h) = kd(x, y)
where n[x, y](h) is the number of times that h meets [xy].
Remark 4.2. Because it is possible to have 1 < n[x, y](h) < ∞, it is
not clear how to use this result to prove that d(x, y) is of negative type,
as we did for the real case in Corollary 2.5. In fact, the example in
the next section indicates that a Crofton formula alone is not enough.
Note that a Crofton formula involving horospheres is clearly also valid
in real hyperbolic space.
5. projective space
The purpose of this section is twofold. Firstly we give an example
to show that the existence of a Crofton formula for a distance func-
tion does not in general imply that the distance is of negative type.
Secondly, this example completes the classification of riemannian sym-
metric spaces of rank one for which the geodesic distance is of negative
type.
The projective spaces P n
F of dimension n ≥ 2 over R, C, H and (in
the case n = 2) the octonions O, are compact two-point homogeneous
spaces.
Proposition 5.1. The geodesic distance d on P n
type.
F is not of negative
9
The projective spaces all contain P 2
it is enough to consider the case of P 2
used for a vector x ∈ R3 and its equivalence class in P 2
distance between points x, y in P 2
R as a geodesic subspace, and so
R. As usual, the same notation is
R. The geodesic
,
R is given by cos d(x, y) = (x,y)
((x,x)(y,y))
1
2
where (x, y) is the usual inner product on R3.
Proposition 5.2. The geodesic distance d on P 2
type.
R is not of negative
Proof: We show that the metric d is not of negative type by ex-
hibiting points x1, x2 . . . , x6 ∈ P 2
R and real numbers t1, t2 . . . , t6 sat-
i=1 ti = 0 such that P6
isfying P6
In fact we
choose (t1, t2 . . . , t6) = (1, 1, 1,−1,−1,−1) and for notational conve-
nience put (x1, x2, x3, x4, x5, x6) = (p1, p2, p3, q1, q2, q3). It is enough to
choose these points so that
i,j=1 titjd(xi, xj) > 0.
X d(pi, pj) +X d(qi, qj) >X d(pi, qj).
(5.1)
We make the following choices. Let p1 = (1, 0, 1), p2 = (1, 0,−1),
p3 = (0, 1, 0), and q1 = (0, 1, 1), q2 = (0, 1,−1), q3 = (1, 0, 0).
Then d(pi, pj) = π/2, d(qi, qj) = π/2, for i, j ∈ {1, 2, 3}. Moreover
d(pi, qj) =
π/3 if both i, j ∈ {1, 2},
π/4 if exactly one of i, j equals 3,
π/2 if i = 3 and j = 3.
We therefore obtainP d(pi, pj)+P d(qi, qj) = 3π/2+3π/2 = 3π and
P d(pi, qj) = 4π/3 + 4π/4 + π/2 = 17π/6, which proves the inequality
(5.1).
(cid:3)
Example. There is a Crofton formula for geodesic distance d on the
space P n
R , but as we have just seen, this distance is not of negative type.
The proof of a Crofton formula for d(x, y) in terms of totally geodesic
hypersurfaces which meet the segment [xy] is exactly the same as for
real hyperbolic space in Section 2, except that by compactness there is
no need for an analogue of Lemma 2.2. The reason that the analogue
of Corollary 2.5 fails is that a totally geodesic hypersurface does not
separate P n
R into two parts.
10
G. ROBERTSON
Remark 5.3. Two point homogeneous riemannian manifolds have been
classified completely [He1, Chapter IX §5]. They are the euclidean
spaces, the circle, and the symmetric spaces of rank one. The sym-
metric spaces of rank one for which the geodesic distance is of negative
type are the spheres and the real or complex hyperbolic spaces. This
follows from the results of [FH], together with our result for projective
spaces. Note that it is easy to give a new proof that the geodesic dis-
tance on a sphere is of negative type along the lines of Corollary 2.5,
using a measure on the space of half-spaces.
Remark 5.4. There is a naturally occurring invariant metric of neg-
ative type on the projective space P n
F [Hi]. For there is an embedding
of P n
F as the set of primitive idempotents in the formally real Jordan
algebra A of hermitian n× n matrices over F. The trace form defines a
euclidean distance dE on A which induces a metric of negative type on
F . It follows from [Hi] that P n
P n
F is also two point homogeneous relative
to this metric.
ADDENDUM
In Remark 2.6 we asked for a direct proof of the fact that the distance
function on quaternionic hyperbolic space is a kernel of negative type.
This is in fact possible, by considering the following 24 points. For
σ ∈ {+1,−1} and ǫ ∈ {±i,±j,±k}, let xσ
ǫ =
(3, 0, 2σ + 2ǫ). Direct calculation shows that
δ ) −X d(xσ
X d(xσ
ǫ , yρ
ǫ and −1 to the yσ
ǫ .
δ ) = 417.03 − 415.77 > 0.
This shows that the condition for negative type fails with the number
+1 being associated to the xσ
ǫ = (3, 2σ + 2ǫ, 0) and yσ
ǫ , xρ
δ) +X d(yσ
ǫ , yρ
References
[A]
P. Assouad, Sur les in´egalit´es valides dans L1, Europ. J. Combinatorics 5
(1984), 99 -- 112.
[BS] W. Ballmann and J. ´Swia
tkowski, On L2- cohomology and property (T) for
'
automorphism groups of polyhedral cell complexes, preprint, Universitat
Bonn, 1996.
N. Bourbaki. Groupes et Alg`ebres de Lie, Chapitres IV, V et VI., Masson,
Paris 1981.
[B]
[BP] R. Benedetti and C. Petronio, Lectures on Hyperbolic Geometry, Springer-
[CG]
Verlag, Berlin-Heidelberg-New York, 1992.
S.S. Chen and L. Greenberg, Hyperbolic spaces, Bers Symposium "Contri-
butions to Analysis", L.V. Ahlfors (ed.), (Academic Press, 1974), 49-87.
[DGL] M. Deza, V.P. Grishukhin and M. Laurent, Hypermetrics in the geometry of
numbers, DIMACS Series in Discrete Mathematics and Computer Science
20, 1995.
11
[FH]
J. Faraut and K. Harzallah, Distances hilbertiennes invariantes sur un es-
pace homog`ene, Ann. Inst. Fourier 24 (1974), 171-217.
[GP] W.M. Goldman and J.R. Parker, Dirichlet polyhedra for dihedral groups
acting on complex hyperbolic space, J. Geometric Analysis 2 (1992), 517 --
554.
I.M. Gelfand and M.M. Smirnov, Lagrangians satisfying Crofton formulas,
Radon transforms, and nonlocal differentials, Advances in Math. 109 (1994),
188 -- 227.
[GS]
[HV] P. de la Harpe and A. Valette, La Propri´et´e (T) de Kazhdan pour les groupes
localement compacts, Ast´erisque 179, Soc. Math. France, 1989.
[He1] S. Helgason, Differential Geometry and Symmetric Spaces, Academic Press,
New York, 1962.
[Hi]
[K]
[M]
[He2] S. Helgason, Geometric Analysis on Symmetric Spaces, Mathematical Sur-
veys and Monographs, vol.39, American Mathematical Society, Providence,
R.I., 1994.
U. Hirzebruch, Uber Jordan-Algebren und kompakte Riemannsche sym-
metrische Raume vom Rang 1, Math. Zeitschr. 90 (1965), 339 -- 354.
J.B. Kelly, Metric inequalities and symmetric differences, Inequalities II, O.
Shisha (ed.), Academic Press (1970), 193 -- 212.
G. Mostow, On a remarkable class of polyhedra in complex hyperbolic space,
Pac. J. Math. 86 (1980), 171-276.
[N]
L. Nachbin, The Haar Integral, van Nostrand, Princeton, New Jersey, 1965.
[RS] G. Robertson and T. Steger, Negative definite kernels and a dynamical
characterization of property (T ) for countable groups, Ergodic Theor. and
Dynam. Systems, to appear.
L.A. Santal´o, Integral geometry on surfaces of constant negative curvature,
Duke Math. J. 10 (1943), 687-704
L.A. Santal´o, Integral Geometry and Geometric Probability, Addison-
Wesley, Reading, Mass., 1976.
R. Zimmer, Ergodic Theory and Semisimple Groups, Monographs in Math.,
vol. 81, Birkhauser, Boston, 1984.
[S2]
[S1]
[Z]
Department of Mathematics, University of Newcastle, NSW 2308,
AUSTRALIA
E-mail address: [email protected]
|
1705.04926 | 1 | 1705 | 2017-05-14T08:14:32 | Frames in Quaternionic Hilbert Spaces | [
"math.FA"
] | In this paper, we introduce and study the frames in separable quaternionic Hilbert spaces. Results on the existence of frames in quaternionic Hilbert spaces have been given. Also, a characterization of frame in quaternionic Hilbert spaces in terms of frame operator is given. Finally, a Paley-Wiener type perturbation result for frames in quaternionic Hilbert space has been obtained. | math.FA | math |
Akxk2≤
hx, xni2≤ Bkxk2,
for all x ∈ H."
(1.1)
∞
Xn=1
FRAMES IN QUATERNIONIC HILBERT SPACES
S.K. Sharma and Shashank Goel
Abstract. In this paper, we introduce and study the frames in separable quaternionic
Hilbert spaces. Results on the existence of frames in quaternionic Hilbert spaces have
been given. Also, a characterization of frame in quaternionic Hilbert spaces in terms of
frame operator is given. Finally, a Paley-Wiener type perturbation result for frames in
quaternionic Hilbert space has been obtained.
Duffin and Schaeffer [11] indroduced frames for Hilbert spaces while working on some
deep problems in non-harmonic Fourier series. They gave the following definition:
1. Introduction
"Let H be a Hilbert space. Then {xn}n∈N ⊂ H is said to be a frame for H if there
exist finite constants A and B with 0 < A ≤ B such that
The positive constants A and B , respectively, are called lower and upper frame bounds
for the frame {xn}n∈N . The inequality (1.1) is called the frame inequality for the frame
{xn}n∈N . A frame {xn}n∈N in H is said to be
• tight if it is possible to choose A = B .
• Parseval if it is a tight frame with A = B = 1.
Frame theory began to spread among researchers when in 1986, Daubechies, Grossmann
and Meyer published a fundamental paper [10] and observed that frames can be also used
to provide the series expansions of functions in L2(R). The main property of frames which
makes them so useful is their redundancy, due to which, representation of signals using
frames is advantageous over basis expansions in a variety of practical applications.
Frame works as an important tool in the study of signal and image processing [2], filter
bank theory [4], wireless communications [13] and sigma-delta quantization [3]. For more
literature on frame theory, one may refer to [5, 8, 6].
Recently, Khokulan, Thirulogasanthar and Srisatkunarajah [14] introduced and studied
frames for finite dimensional quaternionic Hilbert spaces. Sharma and Virender [15]
study some different types of dual frames of a given frame in a quaternionic Hilbert
space and gave various types of reconstructions with the help of dual frame.
In this
paper, we will introduce and study the frames in separable quaternionic Hilbert spaces .
Results on the existence of frames in quaternionic Hilbert spaces have been given. Also,
a characterization of frame in quaternionic Hilbert spaces in terms of frame operator is
given. Finally, a Paley-Wiener type perturbation result for frames in quaternionic Hilbert
space has been obtained.
2. Quaternionic Hilbert space
As the quaternions are non-commutative in nature therefore there are two different
types of quaternionic Hilbert spaces exist, the left quaternionic Hilbert space and the
2010 Mathematics Subject Classification. 42C15, 42A38.
Key words and phrases. Frame, Quaternionic Hilbert spaces.
1
2
Sharma and Goel
right quaternionic Hilbert space depending on positions of quaternions. In this section,
we will study some basic notations about the algebra of quaternions, right quaternionic
Hilbert space and operators on right quaternionic Hilbert spaces.
Throughout this paper, we will denote Q to be a non-commutative field of quaternions,
I be a non empty set of indicies, VR(Q) be a separable right quaternionic Hilbert space,
by the term "right linear operator", we mean a "right Q-linear operator" and B(VR(Q))
denotes the set of all bounded (right Q-linear) operators of VR(Q):
B(VR(Q)) := {T : VR(Q) → VR(Q) : kTk< ∞}.
The non-commutative field of quaternions Q is a four dimensional real algebra with
In Q, 0 denotes the null element and 1 denotes the identity with respect to
unity.
multiplication. It also includes three so-called imaginary units, denoted by i, j, k . i.e.,
Q = {x0 + x1i + x2j + x3k : x0, x1, x2, x3 ∈ R}
where i2 = j 2 = k2 = −1; ij = −ji = k; jk = −kj = i and ki = −ik = j . For each
quaternion q = x0 + x1i + x2j + x3k ∈ Q, define conjugate of q denoted by q as
If q = x0 + x1i + x2j + x3k is a quaternion, then x0 is called the real part of q
and x1i + x2j + x3k is called the imaginary part of q . The modulus of a quaternion
q = x0 + x1i + x2j + x3k is defined as
q = x0 − x1i − x2j − x3k ∈ Q.
q= (qq)1/2 = (qq)1/2 =qx2
0 + x2
1 + x2
2 + x2
3.
For every non-zero quaternion q = x0 + x1i + x2j + x3k ∈ Q, there exists a unique inverse
q−1 in Q as
q−1 =
q
q2 =
x0 − x1i − x2j − x3k
px2
2 + x2
0 + x2
1 + x2
3
.
Definition 2.1. A right quaternionic vector space VR(Q) is a linear vector space under
right scalar multiplication over the field of quaternionic Q, i.e.,
VR(Q) × Q → VR(Q)
(2.1)
and for each u, v ∈ VR(Q) and p, q ∈ Q, the right scalar multiplication (2.1) satisfying
the following properties:
(u, q) → uq
(u + v)q = uq + vq
u(p + q) = up + uq
v(pq) = (vp)q.
Definition 2.2. A right quaternionic pre-Hilbert space or right quaternionic inner product
space VR(Q) is a right quaternionic vector space together with the binary mapping
h..i : VR(Q) × VR(Q) → Q (called the Hermitian quaternionic inner product) which
satisfies following properties:
(a) hv1v2i = hv2v1i for all v1, v2 ∈ VR(Q).
(b) hvvi > 0 if v 6= 0.
(c) hvv1 + v2i = hvv1i + hvv2i for all v, v1, v2 ∈ VR(Q)
(d) hvuqi = hvuiq for all v, u ∈ VR(Q) and q ∈ Q.
In view of Definition 2.2, a right quaternionic inner product space VR(Q) also have the
following properties:
(i) hqvui = qhvui for all v, u ∈ VR(Q) and q ∈ Q
(ii) pv1 + qv2 ∈ VR(Q), for all v1, v2 ∈ VR(Q) and p, q ∈ Q
Frames in quaternionic Hilbert space
3
kuk=phuui, u ∈ VR(Q).
Let VR(Q) be right quaternionic inner product space with the Hermitian inner product
h..i. Define the quaternionic norm k.k: VR(Q) → R+ on VR(Q) by
(2.2)
Definition 2.3. The right quaternionic pre-Hilbert space is called a right quaternionic
Hilbert space, if it is complete with respect to the norm (2.2) and is denoted by VR(Q).
Theorem 2.4 (The Cauchy-Schwarz Inequality). [12] If VR(Q) is a right quaternionic
Hilbert space then
huvi2≤ huuihvvi,
for all u, v ∈ VR(Q).
Moreover, a norm defined in (2.2) satisfy the following properties:
For the non-commutative field of quaternions Q, define the space ℓ2(Q) by
(a) kuqk= kukq, for all u ∈ VR(Q) and q ∈ Q.
(b) ku + vk≤ kuk+kvk, for all u, v ∈ VR(Q).
(c) kuk= 0 for some u ∈ VR(Q), then u = 0.
kqik2< +∞(cid:27)
ℓ2(Q) =(cid:26){qi}i∈I ⊂ Q : Xi∈I
under right multiplication by quaternionic scalars together with the quaternionic inner
product on ℓ2(Q) defined as
hpqi =Xi∈I
piqi, p = {pi}i∈I and q = {qi} ∈ ℓ2(Q).
(2.3)
It is easy to observe that ℓ2(Q) is a right quaternionic Hilbert space with respect to
quaternionic inner product (2.3).
Definition 2.5 ([12]). Let VR(Q) be a right quaternionic Hilbert Space and S be a
subset of VR(Q). Then, define the set:
• S⊥ = {v ∈ VR(Q) : hvui = 0 ∀ u ∈ S}.
• hSi be the right Q-linear subspace of VR(Q) consisting of all finite right Q-linear
combinations of elements of S .
ai as the following element
Let i → ai ∈ R+, i ∈ I be a function on I . Then, define Pi∈I
of R+ ∪ {+∞}:
Xi∈I
ai : J is a non-empty finite subset of I(cid:27).
ai = sup(cid:26)Xi∈J
ai < +∞, then the set of all i ∈ I such that ai
6= 0 is at
most countable. In view of this, given a quaternionic Hilbert space VR(Q) and a map
ui is said to be converges absolutely if Pi∈Ikuik< +∞.
It is clear that, if Pi∈I
i → ui ∈ VR(Q), i ∈ I , the series Pi∈I
Pi∈I
Theorem 2.6 ([12]). Let VR(Q) be a quaternionic Hilbert space and let N be a subset
of VR(Q) such that, for z, z′ ∈ N such that hzz′i = 0 if z 6= z′ and hzzi = 1. Then
the following conditions are equivalent:
If this happens, then only a finite or countable number of ui is non-zero and the series
ui converges to a unique element of VR(Q), independently from the ordering of u′is.
u =Xz∈N
zhzui
where the series Pz∈N
zhzui converges absolutely in VR(Q).
4
Sharma and Goel
huzihzvi.
huvi =Xz∈N
(a) For every u, v ∈ VR(Q), the series Pz∈Nhuzihzvi converges absolutely and
(b) For every u ∈ VR(Q), kuk2=Pz∈Nhzui2 .
(c) N⊥ = 0.
(d) hNi is dense in H .
Definition 2.7 ([12]). Every quaternionic Hilbert space VR(Q) admits a subset N , called
Hilbert basis or orthonormal basis of VR(Q), such that, for z, z′ ∈ N , hzz′i = 0 if z 6= z′
and hzzi = 1 and satisfies all the conditions of Theorem 2.6.
Further, if there are two such sets, then they have the same cardinality. Furthermore,
if N is a Hilbert basis of VR(Q), then for every u ∈ VR(Q) can be uniquely expressed as
Definition 2.8 ([1]). Let VR(Q) be a right quaternionic Hilbert space and T be an
operator on VR(Q). Then T is said to be
• right Q-linear if T (v1α+v2β) = T (v1)α+T (v2)β, for all v1, v2 ∈ VR(Q) and α, β ∈ Q.
• bounded if there exist K ≥ 0 such that kT (v)k≤ Kkvk, for all v ∈ VR(Q).
Definition 2.9 ([1]). Let VR(Q) be a right quaternionic Hilbert space and T be an
operator on VR(Q). Then the adjoint operator T ∗ of T is defined by
hvT ui = hT ∗vui, for all u, v ∈ VR(Q).
Further, T is said to be self-adjoint if T = T ∗ .
Theorem 2.10 ([1]). Let VR(Q) be a right quaternionic Hilbert space and S & T be two
bounded right linear operators on VR(Q). Then
(a) T + S and T S ∈ B(VR(Q)). Moreover:
kT + Sk≤ kTk+kSk and kT Sk≤ kTkkSk
(b) hT vui = hvT ∗ui.
(c) (T + S)∗ = T ∗ + S∗ .
(d) (T S)∗ = S∗T ∗ .
(e) (T ∗)∗ = T.
(f ) I∗ = I , where I is an identity operator on VR(Q).
(g) If T is an invertible operator then (T −1)∗ = (T ∗)−1 .
Theorem 2.11 ([12]). Let VR(Q) be a right quaternionic Hilbert space and let T ∈
B(VR(Q)) be an operator. If T ≥ 0, then there exists a unique operator in B(VR(Q)),
indicated by √T , such that √T ≥ 0 and √T√T = T . Furthermore, it turns out that
√T commutes with every operator which commutes with T and if T is invertible and
self-adjoint, then √T is also invertible and self-adjoint.
Theorem 2.12 ([9]). Let VR(Q) be a right quaternionic Hilbert space and T ∈ B(VR(Q))
and let s ∈ Q be such that kTk< s. Then the operator
(s−1T )ns−1I
Xn≥0
Frames in quaternionic Hilbert space
5
is the right and left algebraic inverse of (sI − T ) and the series converges in the operator
norm.
3. Frames in Quaternionic Hilbert space
We begin this section with the following definition of frames in right quaternionic
Hilbert spaces VR(Q).
Definition 3.1. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a
sequence in VR(Q). Then {ui}i∈I is said to be a frame for VR(Q), if there exist two
finite constants with 0 < A ≤ B such that
(3.1)
Akuk2≤Xi∈I
huiui2≤ Bkuk2, for all u ∈ VR(Q).
The positive constants A and B , respectively, are called lower frame and upper frame
bounds for the frame {ui}i∈I . The inequality (3.1) is called frame inequality for the frame
{ui}i∈I . A sequence {ui}i∈I is called Bessel sequence for right quaternionic Hilbert space
VR(Q) with bound B , if {ui}i∈I satisfies the right hand side of the inequality (3.1). A
frame {ui}i∈I for right quaternionic Hilbert space VR(Q) is said to be
• tight, if it is possible to choose A and B satisfying inequality (3.1) with A = B .
• Parseval frame, if it is tight with A = B = 1.
• exact, if it ceases to be a frame whenever anyone of its element is removed.
Regarding the existence of frames in right quaternionic Hilbert space VR(Q), we have
the following examples:
Example 3.2. Let N be a Hilbert basis for right quaternionic Hilbert space VR(Q) such
that, for each zi, zk ∈ N , i, k ∈ I , we have
hzizki =(0, for i 6= k
1, for i = k.
(1) Tight and Non-Exact. Let {ui}i∈I be a sequence in VR(Q) defined as
ui = ui+1 = zi, i ∈ I.
Then {ui}i∈I is tight and non-exact frame for VR(Q) with bound A = 2. Indeed,
we have
Xi∈I
huiui2 = 2Xi∈I
hziui2
= 2kuk2, for all u ∈ VR(Q).
(2) Non-Tight and Non-exact. Let {ui}i∈I be a sequence in VR(Q) defined as
Then {ui}i∈I is non-tight and non-exact frame for VR(Q). Indeed, we have
huiui2≤ 2kuk2, for all u ∈ VR(Q).
(3) Parseval. Let {ui}i∈I be a sequence in VR(Q) defined as
u1 = z1
uik = uik+1 = uik+2 = . . . = uik+1−1 =
zi√i
,
(u1 = z1
ui = zi−1, i ≥ 2, i ∈ I.
kuk2≤Xi∈I
6
Sharma and Goel
where ik = ik−1 + (k − 1), k ∈ N, i0 = 1.
Then {ui}i∈I is a Parseval frame for VR(Q). Indeed, we have
Xi∈I
= kuk2, for all u ∈ VR(Q).
huiui2=Xi∈I
2
i(cid:12)(cid:12)(cid:12)(cid:12)
(cid:28) zi√i(cid:12)(cid:12)(cid:12)(cid:12)
u(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
for VR(Q).
(4) Exact. Let {zi}i∈I be a Hilbert basis of VR(Q). Then {zi}i∈I is an exact frame
i
ukqk −
j
Xk=1
i
Xk=1
(cid:13)(cid:13)(cid:13)(cid:13)
ukqk(cid:13)(cid:13)(cid:13)(cid:13)
uiqi converges unconditionally.
Next, we show that for a sequence {ui}i∈I in right quaternionic Hilbert space VR(Q)
being a Bessel sequence is a sufficient condition for the series Pi∈I
uiqi , {qi}i∈I ⊂ ℓ2(Q) to
converge unconditionally.
Theorem 3.3. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a Bessel
sequence for VR(Q) with Bessel bound B . Then, for every sequence {qi}i∈I ∈ ℓ2(Q), the
series Pi∈I
Proof. For i, j ∈ I, i > j . Then, we have
ukqk(cid:13)(cid:13)(cid:13)(cid:13)
= (cid:13)(cid:13)(cid:13)(cid:13)
Xk=j+1
(cid:28) i
kvk=1(cid:12)(cid:12)(cid:12)(cid:12)
ukqkv(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
Xk=j+1
Xk=j+1(cid:12)(cid:12)(cid:12)(cid:12)
qkhukvi(cid:12)(cid:12)(cid:12)(cid:12)
≤ (cid:18) i
kgk=1(cid:18) i
qk2(cid:19)1/2
Xk=j+1
Xk=j+1
√B(cid:18) i
qk2(cid:19)1/2
Xk=j+1
≤
Pk=1qk2(cid:27)i∈I
Since {qi}i∈I ∈ ℓ2(Q) and (cid:26) i
(cid:26) i
Pk=1
If view of Theorem 3.3, if {ui}i∈I is a Bessel sequence for VR(Q). Then, the (right)
synthesis operator for {ui}i∈I is a right linear operator T : ℓ2(Q) → VR(Q) defined by
uiqi(cid:27) unconditionally
Hence (cid:26)Pi∈I
is a Cauchy sequence in R. Therefore
hukvi2(cid:19)1/2
is Cauchy sequence in VR(Q).
ukqk(cid:27)i∈I
convergent in VR(Q).
sup
.
= sup
i
= sup
kvk=1
(cid:3)
T ({qi}i∈I) =Xi∈I
uiqi, {qi} ∈ ℓ2(Q).
The adjoint operator T ∗ of right synthesis operator T is called the (right) analysis
operator. Further, the analysis operator T ∗ : VR(Q) → ℓ2(Q) is given by
T ∗(u) = {huiui}i∈I, u ∈ VR(Q).
Frames in quaternionic Hilbert space
7
Infact, for u ∈ VR(Q) and {qi}i∈I ∈ ℓ2(Q), we have
hT ∗(u){qi}i∈Ii = huT ({qi}i∈I)i
uiqi(cid:29)
huuiiqi
= (cid:28)u(cid:12)(cid:12)(cid:12)(cid:12)Xi∈I
= Xi∈I
= (cid:28){huiui}i∈I,{qi}i∈I(cid:29).
Thus
T ∗(u) = {huiui}i∈I, u ∈ VR(Q).
Next, we give a characterization for a Bessel sequence in a right quaternionic Hilbert
space.
Theorem 3.4. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a sequence
in VR(Q). Then, {ui}i∈I is a Bessel sequence for VR(Q) with bound B if and only if the
right linear operator T : ℓ2(Q) → VR(Q) defined by
T ({qi}i∈I) =Xi∈I
uiqi, {qi}i∈I ∈ ℓ2(Q)
is a well defined and bounded operator with kTk≤ √B .
Proof. Let {ui}i∈I be a Bessel sequence for right quaternionic Hilbert space VR(Q). Then,
by Theorem 3.3, the T is a well defined and bounded operator with kTk≤ √B .
Conversely, let T be a well defined and bounded right linear operator with kTk≤ √B .
Then, the adjoint of a bounded right linear operator T is itself bounded and kTk= kT ∗k.
Since, for u ∈ VR(Q), we have
Xi∈I
huiui2 = kT ∗(u)k2
≤ kT ∗k2kuk2
= kTk2kuk2,
it follows that {ui}i∈I is a Bessel sequence for right quaternionic Hilbert space VR(Q)
with bound B .
(cid:3)
Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame for VR(Q).
Then, the (right) frame operator S : VR(Q) → VR(Q) for frame {ui}i∈I is a right linear
operator given by
S(u) = T T ∗(u)
= T ({huiui}i∈I)
= Xi∈I
uihuiui, u ∈ VR(Q).
In next result, we discuss some properties of the frame operator for a frame in right
quaternionic Hilbert space.
8
Sharma and Goel
Theorem 3.5. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame for
VR(Q) with lower and upper frame bounds A and B , respectively and frame operator S .
Then S is positive, bounded, invertible and self adjoint right linear operator on VR(Q).
Proof. For any u ∈ VR(Q), we have
hSuui = (cid:28)Xi∈I
= Xi∈I
= Xi∈I
uihuiui, u(cid:29)
huiuihuiui
huiui2.
This gives
Akuk2≤ hSuui ≤ Bkuk2, u ∈ VR(Q).
Thus
AI ≤ S ≤ BI.
(3.2)
Hence S is positive and bounded right linear operator on VR(Q). Also, 0 ≤ I − B−1S ≤
B − A
I and consequently
B
kI − B−1Sk = sup
kvk=1(cid:12)(cid:12)(cid:12)(cid:12)
B − A
(cid:28)(I − B−1S)v(cid:12)(cid:12)(cid:12)(cid:12)
v(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
≤
< 1.
B
Then, S is invertible. Further for any u, v ∈ VR(Q), we have
hSuvi = (cid:28)Xi∈I
= Xi∈I
= Xi∈I
= (cid:28)u(cid:12)(cid:12)(cid:12)(cid:12)Xi∈I
= huSvi.
uihuiui(cid:12)(cid:12)(cid:12)(cid:12)
v(cid:29)
huiuihuivi
huuiihuivi
uihuivi(cid:29)
Thus S is also a self adjoint right linear operator on VR(Q).
(cid:3)
Corollary 3.6 (The Reconstruction Formula). Let VR(Q) be a right quaternionic Hilbert
space and {ui}i∈I be a frame for VR(Q) with frame operator S . Then, for every
u ∈ VR(Q) can be expressed as
S−1uihuiui.
u =Xi∈I
Frames in quaternionic Hilbert space
Proof. As S is invertible. Therefore, for u ∈ VR(Q), we have
u = S−1S(u)
= Xi∈I
S−1uihuiui.
9
(cid:3)
In the next result, we construct a frame with the help a give frame in a right
quaternionic Hilbert space.
Theorem 3.7. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with lower and upper frame bounds A and B , respectively and frame operator
is also a frame for VR(Q) with bounds B−1 and A−1 and right
S . Then {S−1ui}i∈I
frame operator S−1 .
Proof. For u ∈ VR(Q), we have
hS−1uiui2 = Xi∈I
huiS−1ui2
Xi∈I
≤ BkS−1uk2
≤ BkS−1k2kuk2.
Thus {S−1ui}i∈I is a Bessel sequence for VR(Q). It follows that the right frame operator
for {S−1ui}i∈I is well defined. Therefore, we have
Xi∈I
S−1uihS−1uiui = S−1(cid:18)Xi∈I
uihS−1uiui(cid:19)
= S−1S(S−1u)
= S−1u, u ∈ VR(Q).
(3.3)
Thus, the right frame operator for {S−1ui}i∈I is S−1 . The operator S−1 commutes with
both S and I : VR(Q) → VR(Q) (an identity operator on VR(Q)). Therefore, multiplying
the inequality (3.2) with S−1 , we have
i.e.
B−1I ≤ S−1 ≤ A−1I
B−1kuk2≤ hS−1uui ≤ A−1kuk2, u ∈ VR(Q).
By using (3.3), we get
B−1kuk2≤Xi∈I
hS−1uiui2≤ A−1kuk2, u ∈ VR(Q).
Therefore {S−1ui}i∈I is frame for VR(Q) with bounds
S−1 .
1
B
and
1
A
and right frame operator
(cid:3)
In view of Theorem 3.7, we have a following definition:
Definition 3.8. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with frame operator S . Then, the frame {S−1ui}i∈I is called the canonical
dual of the frame {ui}i∈I .
Next, we construct a Parseval frame with the help of a give frame in a right quaternionic
Hilbert space.
10
Sharma and Goel
Theorem 3.9. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with frame operator S . Then {S−1/2ui}i∈I is a Parseval frame for VR(Q).
Proof. By Theorem 2.11, for any u ∈ VR(Q), we have
u = S−1/2SS−1/2u
uihuiS−1/2ui
S−1/2uihuiS−1/2ui.
Therefore, for any u ∈ VR(Q)
= S−1/2Xi∈I
= Xi∈I
kuk2 = huui
= (cid:28)Xi∈I
= Xi∈I
= Xi∈I
S−1/2uihuiS−1/2ui(cid:12)(cid:12)(cid:12)(cid:12)
u(cid:29)
huiS−1/2uihS−1/2uiui
hS−1/2uiui2.
(3.4)
(by using (3.4)) .
Thus {S−1/2ui}i∈I is a Parseval frame for VR(Q).
Next, we give a characterization of Parseval frames {ui}i∈I
Hilbert space VR(Q).
Theorem 3.10. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with frame operator S . Then {ui}i∈I is a Parseval frame for VR(Q) if and
only if S is an identity operator on VR(Q).
Proof. Let {ui}i∈I be a Parseval frame for VR(Q). Then, for all u ∈ VR(Q), we have
(cid:3)
for right quaternionic
huiui2 = kuk2
Xi∈I
which implies
hSuui = huui.
Thus, S is an identity operator on VR(Q).
Conversely, let S is an identity operator on VR(Q). Thus, for u ∈ VR(Q), we have
u = S(u)
So, we have
uihuiui.
= Xi∈I
kuk2 = huui
= (cid:28)Xi∈I
= Xi∈I
uihuiuiu(cid:29)
huiui2.
(3.5)
Frames in quaternionic Hilbert space
Hence {ui}i∈I is a Parseval frame for right quaternionic Hilbert space VR(Q).
11
(cid:3)
In the next theorem, we give a necessary condition for a Parseval frame {ui}i∈I for
right quaternionic Hilbert space VR(Q).
Theorem 3.11. Let VR(Q) be a right quaternionic Hilbert space and {qi}i∈I be a semi-
normalized sequence of quaternions in Q with bounds a and b. If {uiqi}i∈I is a Parseval
frame for VR(Q), then {ui}i∈I is a frame for VR(Q) with bounds b−2 and a−2 .
Proof. Let {uiqi}i∈I be a Parseval frame for VR(Q). Then, for any u ∈ VR(Q), we have
huiqiui2= kuk2.
Xi∈I
This gives
qi2huiui2= kuk2.
Xi∈I
(3.6)
Since {qi}i∈I is a normalized sequence with bounds a and b, we get
huiui2, u ∈ VR(Q).
qi2huiui2≤ b2Xi∈I
a2Xi∈I
huiui2≤Xi∈I
Therefore, by using (3.5), we have
1
b2kuk2≤Xi∈I
huiui2≤
1
a2 , u ∈ VR(Q).
Hence {ui}i∈I is a frame for VR(Q) with bounds b−2 and a−2 .
(cid:3)
Finally, in this section we give a characterization of frame for right quaternionic Hilbert
space in terms of operators.
Theorem 3.12. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a
sequence in VR(Q). Then {ui}i∈I is a frame for VR(Q) if and only if the right linear
operator T : ℓ2(Q) → VR(Q)
T ({qi}i∈I) =Xi∈I
uiqi, {qi}i∈I ∈ VR(Q)
is a well-defined and bounded mapping from ℓ2(Q) onto VR(Q).
Proof. Let {ui}i∈I be a frame for VR(Q). Then the frame operator S = T T ∗ for {ui}i∈I
is invertible on VR(Q). So T is an onto mapping. Also, by Theorem 3.4, T is well defined
and bounded mapping from ℓ2(Q) to VR(Q).
Conversely, let T be a well defined, bounded right linear operator from ℓ2(Q) onto
VR(Q). Then, by Theorem 3.4, {ui}i∈I is a Bessel sequence for VR(Q). Since T is an
onto, there exists an right linear operator T † : VR(Q) → ℓ2(Q) such that
u = T T †u
= Xi∈I
ui(T †u)i, u ∈ VR(Q),
12
Sharma and Goel
where (T †u)i denotes the i−th coordinate of T †u. Therefore, for u ∈ VR(Q), we have
2
kuk4 = huui2
ui(T †u)i(cid:12)(cid:12)(cid:12)(cid:12)
u(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
= (cid:12)(cid:12)(cid:12)(cid:12)
(cid:28)Xi∈I
(T †u)i2 Xi∈I
≤ Xi∈I
≤ kT †k2kuk2Xi∈I
huiui2.
huiui2
Hence {ui}i∈I is a frame for VR(Q) with bounds kT †k−2 and kTk2 .
(cid:3)
4. Stability of Frames in Right Quaternionic Hilbert Space
Christensen [7] gave a version of Paley-Wiener Theorem for frames in Hilbert spaces.
In this section, we give a similar version for frames in right quaternionic Hilbert spaces.
Theorem 4.1. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with lower and upper frame bounds A and B , respectively. Let {vi}i∈I be a
sequence in VR(Q) and assume that there exist λ, µ ≥ 0 such that (cid:18)λ +
√A(cid:19) < 1 and
µ
(4.1)
µ
.
qi2(cid:19)1/2
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
viqi(cid:13)(cid:13)(cid:13)(cid:13)
µ
√B(cid:19)(cid:19)2
(ui − vi)qi(cid:13)(cid:13)(cid:13)(cid:13)
≤ λ(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
uiqi(cid:13)(cid:13)(cid:13)(cid:13)
+ µ(cid:18)Xi∈J
for all finite quaternions qi ∈ Q, i ∈ J ⊆ I with J< +∞. Then {vi}i∈I is a frame for
√A(cid:19)(cid:19)2
and B(cid:18)1 +(cid:18)λ +
VR(Q) with bounds A(cid:18)1 −(cid:18)λ +
Proof. Let J ⊆ I with J< +∞. Then
uiqi(cid:13)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
(ui − vi)qi(cid:13)(cid:13)(cid:13)(cid:13)
≤ (cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
qi2(cid:19)1/2
uiqi(cid:13)(cid:13)(cid:13)(cid:13)
≤ (1 + λ)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
+ µ(cid:18)Xi∈J
uiqi(cid:12)(cid:12)(cid:12)(cid:12)
kvk=1(cid:12)(cid:12)(cid:12)(cid:12)
v(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:28)Xi∈J
qi2(cid:19)1/2
≤ (cid:18)Xi∈J
kvk=1(cid:18)Xi∈J
qi2(cid:19)1/2
√B(cid:18)Xi∈J
huivi2(cid:19)1/2
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
uiqi(cid:13)(cid:13)(cid:13)(cid:13)
sup
.
= sup
≤
Since
.
This gives
Frames in quaternionic Hilbert space
13
≤(cid:18)(1 + λ)√B + µ(cid:19)(cid:18)Xi∈J
qi2(cid:19)1/2
.
(4.2)
viqi(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
T ({qi}i∈I) =Xi∈J
Let T : ℓ2(Q) → VR(Q) be a right linear operator defined by
viqi, {qi}i∈I ∈ ℓ2(Q).
Then, by (4.2), T is well defined and bounded with kTk≤ (cid:18)(1 + λ)√B + µ(cid:19).
Therefore, by Theorem 3.4, {vi}i∈I is a Bessel sequence for VR(Q) with the upper bound
√B(cid:19)(cid:19)2
B(cid:18)1 +(cid:18)λ +
Let U be the right synthesis operator for the frame {ui}i∈I . Define an operator
W : VR(Q) → ℓ2(Q) by
µ
.
W (u) = U∗(U U∗)−1u, u ∈ VR(Q).
Since {(U U∗)−1ui}i∈I is also a frame for VR(Q) with bounds
1
B
and
1
A
, then
kW (u)k2 = Xi∈I
hS−1uiui2
1
Akuk2, u ∈ VR(Q).
≤
Using (4.1) with {qi}i∈I = W u, u ∈ VR(Q), we get
ku − T W uk ≤ λkuk+µkW uk
≤ (cid:18)λ +
µ
√A(cid:19)kuk, u ∈ VR(Q).
. Since (cid:18)λ +
µ
This gives kT Wk≤ 1 + λ +
invertible and k(T W )−1k≤
µ
√A
1 −(cid:18)λ +
Now, for any u ∈ VR(Q), we have
1
.
µ
√A(cid:19)
u = (T W )(T W )−1u
= Xi∈I
vih(U U∗)−1ui(T W )−1ui.
√A(cid:19) < 1, then the operator T W is
14
Sharma and Goel
Therefore, for each u ∈ VR(Q), we have
2
vih(U U∗)−1ui(T W )−1ui(cid:12)(cid:12)(cid:12)(cid:12)
u(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
kuk4 = (cid:12)(cid:12)(cid:12)(cid:12)
(cid:28)Xi∈I
2(cid:18)Xi∈I
h(U U∗)−1ui(T W )−1ui(cid:12)(cid:12)(cid:12)(cid:12)
≤ Xi∈I (cid:12)(cid:12)(cid:12)(cid:12)
hviui2(cid:19)
A(cid:20)
hviui2(cid:19).
kuk2(cid:18)Xi∈I
1 − (λ + µ√A
)(cid:21)2
≤
1
1
This gives
hviui2≥ A(cid:18)1 −(cid:18)λ +
µ
√A(cid:19)(cid:19)2
Xi∈I
kuk2, u ∈ VR(Q).
Hence {vi}i∈I is a frame for VR(Q) with desire bounds.
Corollary 4.2. Let VR(Q) be a right quaternionic Hilbert space and {ui}i∈I be a frame
for VR(Q) with lower and upper frame bounds A and B , respectively. Let {vi}i∈I be a
sequence in VR(Q) and assume that there exist 0 < R < A such that
(cid:3)
(ui − vi)qik≤
kXi∈I
√R(cid:18)Xi∈I
qi2(cid:19)1/2
, for all {qi}i∈I ∈ ℓ2(Q).
(4.3)
Then {vi}i∈I is also a frame for VR(Q) with bounds (√A − √R)2 and (√B + √R)2 .
Proof. Take λ = 0 and µ = √R in Theorem 4.1.
√A(cid:19) < 1 can not be dropped in the Theorem 4.1. In
Remark 4.3. A condition (cid:18)λ +
Example 4.4. Let VR(Q) be a right quaternionic Hilbert space, {zi}i∈N be a Hilbert
basis of VR(Q), {pi}i∈N be a sequence in Q and {vi}i∈N be a sequence in VR(Q) such
that
this, regard, we give the following example:
µ
(cid:3)
vi = zi + zi+1pi, i ∈ N.
Then, for J ⊆ N with J< ∞ and {qi}i∈N ∈ ℓ2(Q), we have
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
(vi − zi)qi(cid:13)(cid:13)(cid:13)(cid:13)
= (cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
≤ sup
i
zi+1piqi(cid:13)(cid:13)(cid:13)(cid:13)
pi(cid:18)Xi∈I
qi2(cid:19)1/2
.
Thus, if p = sup
(1 − p)2 and (1 + p)2 . On taking pi = 1, for all i ∈ N, we get
pi< 1, then by Theorem 4.1, {ui}i∈N is a frame for VR(Q) with bounds
i
vi = zi + zi+1, i ∈ N.
(4.4)
Frames in quaternionic Hilbert space
15
Then, for any sequence {qi}i∈N ∈ ℓ2(Q) and for J ⊆ N with J< ∞
(cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
(vi − zi)qi(cid:13)(cid:13)(cid:13)(cid:13)
= (cid:13)(cid:13)(cid:13)(cid:13)Xi∈J
≤ (cid:18)Xi∈I
zi+1qi(cid:13)(cid:13)(cid:13)(cid:13)
qi2(cid:19)1/2
.
Thus, the condition (4.1) is satisfied with either (λ, µ) = (1, 0) or (0, 1). So, (cid:18)λ+
1 and {vi}i∈N is not a frame for VR(Q).
µ
√A(cid:19) ≮
References
[1] S.L. Adler, Quaternionic Quantum Mechanics and Quantum Fields, Oxford University Press, New
York, 1995.
[2] R. Balan, P.G. Casazza and D. Edidin, On signal reconstruction without phase, Appl. Comp. Harm.
Anal., 20 (2006), 345-356.
[3] J. Benedetto, A. Powell and O. Yilmaz, Sigma-Delta quantization and finite frames, IEEE Trans.
Inform. Theory, 52 (2006), 1990-2005.
[4] H. Bolcskel, F. Hlawatsch and H.G. Feichtinger, Frame - theoretic analysis of over sampled filter
banks, IEEE Trans.Signal Process. 46 (1998), 3256-3268.
[5] P.G. Casazza, The art of frame theory, Taiwanese J. of Math., 4 (2)(2000), 129-201.
[6] P.G. Casazza and G. Kutyniok, Frames of subspaces, in Wavelets, Frames and Operator Theory
(College Park, MD, 2003), Contemp. Math., 345, Amer. Math. Soc., Providence, RI, 2004, 87- 113.
[7] O. Christensen, A Paley-Wiener theorem for frames, Proc. Amer. Math. Soc., 123 (1995), 2199-2202.
[8] O. Christensen, An introduction to Frames and Riesz Bases, Birkhauser, 2003.
[9] F. Colombo,
I. Sabadini and D.C. Struppa: Noncommutative Functional Calculus, Theory
and applications of slice hyperholomorphic functions, Progress in Mathematics, vol. 289,
Birkhauser/Springer Basel AG, Basel, 2011.
[10] I. Daubechies, A. Grossmann and Y. Meyer, Painless non-orthogonal expansions, J. Math. Physics,
27 (1986), 1271-1283.
[11] R.J. Duffin and A.C. Schaeffer, A class of non-harmonic Fourier series, Trans. Amer. Math. Soc.,
72 (1952), 341-366.
[12] R. Ghiloni, , V. Moretti and A. Perotti, Continuous slice functional calculus in quaternionic Hilbert
spaces, Rev. Math. Phys. 25 (2013), 1350006.
[13] R.W. Heath and A.J. Paulraj, Linear dispersion codes for MIMO systems based on frame theory,
IEEE Trans.Signal Process. 50 (2002), 2429- 2441.
[14] M. Khokulan, K. Thirulogasanthar and S. Srisatkunarajah, Discrete frames on finite dimensional
quaternion Hilbert spaces, arXiv: 1302.2836v1
[15] S.K. Sharma and Virender, Dual frames on finite dimensioal quaternionic Hilbert space, Poincare
J. Anal. Appl., 2016 (2), 79-88.
(S.K. Sharma) Department of Mathematics, Kirori Mal College, University of Delhi,
Delhi 110 007, INDIA.
E-mail address: [email protected]
(Shashank Goel) Department of Mathematics, Amity Institute of Applied Sciences,
Amity University, Noida, U.P-201301, INDIA.
E-mail address: [email protected]
|
1306.0724 | 1 | 1306 | 2013-06-04T10:27:24 | Wandering subspaces of the Bergman space and the Dirichlet space over polydisc | [
"math.FA"
] | Doubly commutativity of invariant subspaces of the Bergman space and the Dirichlet space over the unit polydisc $\mathbb{D}^n$ (with $ n \geq 2$) is investigated. We show that for any non-empty subset $\alpha=\{\alpha_1,\dots,\alpha_k\}$ of $\{1,\dots,n\}$ and doubly commuting invariant subspace $\s$ of the Bergman space or the Dirichlet space over $\D^n$, the tuple consists of restrictions of co-ordinate multiplication operators $M_{\alpha}|_\s:=(M_{z_{\alpha_1}}|_\s,\dots, M_{z_{\alpha_k}}|_\s)$ always possesses wandering subspace of the form \[\bigcap_{i=1}^k(\s\ominus z_{\alpha_i}\s). \] | math.FA | math |
WANDERING SUBSPACES OF THE BERGMAN SPACE AND THE
DIRICHLET SPACE OVER POLYDISC
A. CHATTOPADHYAY, B. KRISHNA DAS, JAYDEB SARKAR, AND S. SARKAR
Abstract. Doubly commutativity of invariant subspaces of the Bergman space and the
Dirichlet space over the unit polydisc Dn (with n ≥ 2) is investigated. We show that for any
non-empty subset α = {α1, . . . , αk} of {1, . . . , n} and doubly commuting invariant subspace
S of the Bergman space or the Dirichlet space over Dn, the tuple consists of restrictions of co-
S) always possesses wandering
ordinate multiplication operators MαS := (Mzα1 S , . . . , Mzα
subspace of the form
k
(S ⊖ zαiS).
k
\i=1
1. Introduction
A closed subspace W of a Hilbert space H is said to be wandering subspace (following
Halmos [Hal]) for an n-tuple of commuting bounded linear operators T = (T1, . . . , Tn) (n ≥ 1)
if
for all (l1, . . . , ln) ∈ Nn \ {(0, . . . , 0)} and
W ⊥ T l1
1 T l2
2 · · · T ln
n W
H = span{T l1
1 T l2
2 · · · T ln
n h : h ∈ W, l1, . . . , ln ∈ N}.
In this case, the tuple T is said to have the wandering subspace property.
The main purpose of this paper is to investigate the following question:
Question: Let (T1, . . . , Tn) be a commuting n-tuple of bounded linear operators on a Hilbert
space H. Does there exists a wandering subspace W for (T1, . . . , Tn)?
This question has an affirmative answer for the restriction of multiplication operator by
the coordinate function Mz, to an invariant subspace of the Hardy space H 2(D) (Beurlings
theorem [Beu]) or the Bergman space A2(D) (Aleman, Richter and Sundberg [ARS]) or the
Dirichlet space D(D) (Richter [Ric] ) over the unit disc D of the complex plane C. For n ≥ 2,
existence of wandering subspaces for general invariant subspaces of the Hardy space H 2(Dn)
over the unit polydisc Dn rather fails spectacularly (cf. [Rud]).
2010 Mathematics Subject Classification. 47A13, 47A15, 47A20, 47L99.
Key words and phrases. Invariant subspace, Beurling's theorem, Bergman space, Dirichlet space, Hardy space,
Doubly commutativity.
1
2
CHATTOPADHYAY, DAS, SARKAR, AND SARKAR
Recall that the Hardy space over the unit polydisc Dn = {z = (z1, . . . , zn) ∈ Cn : zi <
1, i = 1, . . . , n} is denoted by H 2(Dn) and defined by
H 2(Dn) = {f ∈ O(Dn) : sup
0≤r<1ZTn
f (rz)2dθ < ∞},
where dθ is the normalized Lebesgue measure on the torus Tn, the distinguished boundary
of Dn, rz := (rz1, . . . , rzn) and O(Dn) denotes the set of all holomorphic functions on Dn (cf.
[Rud]).
In [Beu], A. Beurling characterize all closed Mz-invariant subspaces of H 2(D) in the follow-
ing sense: Let S 6= {0} be a closed subspace of H 2(D). Then S is Mz-invariant if and only if
S = θH 2(D) for some inner function θ (that is, θ ∈ H ∞(D) and θ = 1 a.e. on the unit circle
T). In particular, Beurlings theorem yields the wandering subspace property for Mz-invariant
subspaces of H 2(D) as follows: if S = θH 2(D) is an Mz-invariant subspace of H 2(D) then
(1)
⊕zmW,
S =Xm≥0
where W is the wandering subspace for MzS given by
W = S ⊖ zS = θH 2(D) ⊖ zθH 2(D) = θC.
In [Rud], W. Rudin showed that there are invariant subspaces M of H 2(D2) which does
not contain any bounded analytic function. In particular, the Beurling like characterization
of (Mz1, . . . , Mzn)-invariant subspaces of H 2(Dn), in terms of inner functions on Dn is not
possible.
On the other hand, the wandering subspace (and the Beurlings theorem) for the shift
invariant subspaces of the Hardy space H 2(D) follows directly from the classical Wold decom-
position [Wol] theorem for isometries. Indeed, for a closed Mz-invariant subspace S(6= {0})
of H 2(D),
Consequently, by Wold decomposition theorem for the isometry MzS on S we have
∞
\m=0
S =Xm≥0
(MzS)mS =
M m
z S ⊆
M m
z H 2(D) = {0}.
∞
\m=0
∞
\m=0
⊕zmW ⊕ (
∞
\m=0
(MzS)mS) =Xm≥0
⊕zmW,
where W = S ⊖ zS and hence, (1) follows. Therefore, the notion of wandering subspaces
is stronger (as well as of independent interest) than the Beurlings characterization of Mz-
invariant subspaces of H 2(D).
To proceed, we now recall the following definition: a commuting n-tuple (n ≥ 2) of bounded
j Ti for all
linear operators (T1, . . . , Tn) on H is said to be doubly commuting if TiT ∗
1 ≤ i < j ≤ n.
j = T ∗
Natural examples of doubly commuting tuple of operators are the multiplication operators
by the coordinate functions acting on the Hardy space or the Bergman space or the Dirichlet
space over the unit polydisc Dn (n ≥ 2).
WANDERING SUBSPACES OF THE BERGMAN SPACE AND THE DIRICHLET SPACE OVER Dn
3
Let H ⊆ O(Dn) be a reproducing kernel Hilbert space over Dn (see [Aro]) and multi-
plication operators {Mz1, . . . , Mzn} by coordinate functions are bounded. Then a closed
(Mz1, . . . , Mzn)-invariant subspace S of H is said to be doubly commuting if the n-tuple
(Mz1S, . . . , MznS) is doubly commuting, that is, RiR∗
j Ri for all 1 ≤ i < j ≤ n, where
Ri = MziS.
j = R∗
In [SSW], third author and Sasane and Wick proved that any doubly commuting invariant
subspace of H 2(Dn) (where n ≥ 2) has the wandering subspace property (see [Man] for n = 2).
Also in [ReT], Redett and Tung obtained the same conclusion for doubly commuting invariant
subspaces of the Bergman space A2(D2) over the bidisc D2.
In this paper we prove that doubly commuting invariant subspaces of the Bergman space
A2(Dn) and the Dirichlet space D(Dn) has the wandering subspace property. Our result on
the Bergman space over polydisc is a generalization of the base case n = 2 in [ReT]. Our
analysis is based on the Wold-type decomposition result of S. Shimorin for operators closed
to isometries [Shi].
The paper is organized as follows: in Section 2, we obtain some general results concerning
the wandering subspaces of tuples of doubly commuting operators. We obtain wandering
subspaces for doubly commuting shift invariant subspaces of the Bergman and Dirichlet spaces
over Dn in Section 3.
2. Wandering subspace for tuple of doubly commuting operators
In this section we prove the multivariate version of the S. Shimorin's result for tuple of
doubly commuting operators on a general Hilbert space. We show that for a tuple of doubly
commuting operators T = (T1, . . . , Tn) on a Hilbert space H, if for any reducing subspace Si
of Ti the subspace Si ⊖ TiSi is a wandering subspace for TiSi, i = 1, . . . , n, then
(H ⊖ TiH)
n
Ti=1
is a wandering subspace for T . We fix for the rest of the paper a natural number n ≥ 2 and
set Λn := {1, . . . , n}.
For a closed subset K of a Hilbert space H, an n-tuple of commuting operators T =
(T1, . . . , Tn) on H and a non-empty subset α = {α1, . . . , αk} ⊆ Λn, we write [K]Tα to denote
the smallest closed joint Tα := (Tα1, . . . , Tαk)-invariant subspace of H containing K. In other
words
(2)
[K]Tα =
∞
_l1,l2,...,lk=0
T l1
α1T l2
α2 · · · T lk
αk (K).
If α is a singleton set {i} then we simply write [K]Ti.
For a non-empty subset α = {α1, . . . , αk} ⊆ Λn, we also denote by Wα the following
subspace of H,
(3)
Wα =
(H ⊖ TαiH).
k
\i=1
4
CHATTOPADHYAY, DAS, SARKAR, AND SARKAR
Again if α is a singleton set {i} then we simply write Wi. Thus with the above notation
Wα = Tαi∈α
Wαi.
A bounded linear operator T on a Hilbert space H is analytic if
concave if it satisfies the following inequality
kT 2xk2 + kxk2 ≤ 2kT xk2
(x ∈ H).
T nH = {0} and it is
∞
Tn=0
Multiplication by coordinate functions on the Dirichlet space over the unit polydisc are con-
cave operators as we show in the next section.
For a single bounded operator T on a Hilbert space H, the following result ensure the
existence of the wandering subspace under certain conditions (see [Ric], [Shi]).
Theorem 2.1. (Richter, Shimorin). Let T be an analytic operator on a Hilbert space H
satisfies one of the following properties:
(i) k T x + y k2≤ 2(k x k2 + k T y k2) (x, y ∈ H),
(ii) T is concave.
Then H ⊖ T H is a wandering subspace for T , that is,
H = [H ⊖ T H]T .
The following proposition is essential in order to generalize the above result for certain
tuples of commuting operators.
Proposition 2.2. Let T = (T1, . . . , Tn) be a doubly commuting tuple of operators on H.
Then Wα is Tj-reducing subspace for all non-empty subset α ⊆ Λn and j /∈ α, where Wα is
as in (3).
Proof. Let α = {α1, . . . , αk} be a non-empty subset of Λn and j /∈ α. First note that
Wl = Ker T ∗
l
for all 1 ≤ l ≤ n and therefore Wα =
y ∈ H. By doubly commutativity of T we have,
k
Ti=1
Ker T ∗
αi. Let x ∈ Wα, αi ∈ α and
hTjx, Tαiyi = hT ∗
αiTjx, yi = hTjT ∗
αix, yi = 0.
Therefore, TjWα ⊥ TαiH and hence TjWα ⊆ Wαi for all αi ∈ α. Thus Wα is an invariant
subspace for Tj. Also, by commutativity of T we have
for all αi ∈ α and y ∈ H. This implies T ∗
This completes the proof.
hT ∗
j x, Tαiyi = hT ∗
j x, yi = hT ∗
αiT ∗
j Wα ⊥ TαiH and then T ∗
αix, yi = 0,
j T ∗
j Wα ⊆ Wαi for all αi ∈ α .
(cid:3)
Now we prove the main theorem in the general Hilbert space operator setting. Below for a
set α, we denote by #α the cardinality of α.
WANDERING SUBSPACES OF THE BERGMAN SPACE AND THE DIRICHLET SPACE OVER Dn
5
Theorem 2.3. Let T = (T1, . . . , Tn) be a doubly commuting tuple of operators on H such
that for any reducing subspace Si of Ti, the subspace
Si ⊖ TiSi
is a wandering subspace for TiSi, i = 1, . . . , n. Then for each non-empty subset α =
{α1, . . . , αk} ⊆ Λn, the tuple Tα = (Tα1, . . . , Tαk ) has the wandering subspace property. More-
over, the corresponding wandering subspace is given by
Wα =
(H ⊖ TαiH).
k
\i=1
Proof. First note that we only need to show H = [Wα]Tα for any non-empty subset α of Λn as
the orthogonal property for wandering subspace is immediate. Now for a singleton set α the
result follows form the assumption that Wi is a wandering subspace for Ti, i = 1, . . . , n. Now
= Wα\{αi} for any αi ∈ α. Because for αi, αj ∈ α,
for #α ≥ 2, it suffices to show that [Wα]Tαi
one can repeat the procedure to get
[Wα]T{αi,αj } = [[Wα]Tαi
]Tαj
= Wα\{αi,αj},
and continue this process until the set α \ {αi, αj} becomes a singleton set and finally apply
the assumption for singleton set.
To this end, let α ⊆ Λn, #α ≥ 2 and αi ∈ α. Consider the set F = Wα\{αi} ⊖ Tαi(Wα\{αi}).
Now by Proposition 2.2, Wα\{αi} is a reducing subspace for Tαi and therefore
F = {x ∈ Wα\{αi} : x ∈ Ker T ∗
αi}
= Wα\{αi} ∩ Wαi
= Wα.
On the other hand, since Wα\{αi} is a reducing subspace for Tαi then by assumption F = Wα
is a wandering subspace for Tαi. Thus
[Wα]Tαi
= Wα\{αi}.
This completes the proof.
(cid:3)
Combining Theorem 2.1 with the above theorem we have the following result, which is the
case of our main interest.
Corollary 2.4. Let T = (T1, . . . , Tn) be a commuting tuple of analytic operators on H such
that T is doubly commuting and satisfies one of the following properties:
(a) Ti is concave for each i = 1, . . . , n
(b) kTix + yk2 ≤ 2(kxk2 + kTiyk2) (x, y ∈ H, i = 1, 2, . . . , n).
Then for any non-empty subset α = {α1, . . . , αk} ⊆ Λn, Wα is a wandering subspace for
Tα = (Tα1, . . . , Tαk), where Wα is as in (3).
Proof. Note that if T satisfies condition (a) or (b) then for any 1 ≤ i ≤ n and reducing
subspace Si of Ti, TiSi also satisfies condition (a) or (b) respectively. Thus by Theorem 2.1,
T satisfies the hypothesis of the above theorem and the proof follows.
(cid:3)
6
CHATTOPADHYAY, DAS, SARKAR, AND SARKAR
We end this section by proving kind of converse of the above result and part of which is a
generalization of [ReT], Theorem 3.
Theorem 2.5. Let T = (T1, . . . , Tn) be an n-tuple of commuting operators on H with the
property
kTix + yk2 ≤ 2(kxk2 + kTiyk2) (x, y ∈ H, i = 1, 2, . . . , n)
or Ti is concave for all i = 1, . . . , n. Then,
(i) T is doubly commuting on H, and
(ii) Ti is analytic for all i = 1, 2, . . . , n
if and only if
(a) for any non-empty subset α = {α1, . . . , αk} ⊆ Λn, Wα is a wandering subspace for Tα =
(Tα1, . . . , Tαk) and for k ≥ 2, [Wα]Tαi
(b) Ti commutes with T ∗
= Wα\{αi} for all αi ∈ α,
j Tj for all 1 ≤ i < j ≤ n.
Proof. The forward direction follows from Theorem 2.3.
For the converse, suppose that (a) and (b) hold. To show (i), let 1 ≤ i ≤ n be fixed. By
assumption (a), Wi = [W{i,j}]Tj for all 1 ≤ i 6= j ≤ n. This shows that Wi is Tj invariant
subspace for all 1 ≤ j 6= i ≤ n. Let x ∈ H. Since [Wi]Ti = H then there exists a sequence xk
converging to x such that
xk =
T m
i xm,k,
NkXm=0
where each Nk is a nonnegative integer and xm,k is a member of Wi. Now for any 1 ≤ j 6= i ≤ n
we have,
T ∗
i Tjxk =
NkXm=0
T ∗
i TjT m
i xm,k =
T ∗
i TjT m
i xm,k.
NkXm=1
On the other hand,
TjT ∗
i xk =
NkXm=0
TjT ∗
i T m
i xm,k =
TjT ∗
i T m
i xm,k =
NkXm=1
T ∗
i TjT m
i xm,k,
NkXm=1
where the last equality follows from (b). So T ∗
TjT ∗
Finally, by Theorem 3.6 of [Shi] we have that
i x. Thus we have (i).
i Tjxk = TjT ∗
i xk and by taking limit T ∗
i Tjx =
H = [Wi]Ti
⊕
T m
i H,
∞
\m=1
for all 1 ≤ i ≤ n. By part (a), [Wi]Ti = H for all 1 ≤ i ≤ n. Thus
1 ≤ i ≤ n and this completes the proof.
∞
Tm=1
T m
i H = {0} for all
(cid:3)
WANDERING SUBSPACES OF THE BERGMAN SPACE AND THE DIRICHLET SPACE OVER Dn
7
3. Wandering subspaces for invariant subspaces of A2(Dn) and D(Dn)
In this section we apply the general theorem proved in the previous section to obtain
wandering subspaces for invariant subspaces of the Bergman space and the Dirichlet space
over polydisc.
The Bergman space over Dn is denoted by A2(Dn) and defined by
A2(Dn) = {f ∈ O(Dn) :ZDn
f (z)2dA(z) < ∞},
where dA is the product area measure on Dn.
Below for any invariant subspace S of A2(Dn) and non-empty set α = {α1, . . . , αk} ⊆ Λn,
we denote by W S
α the following set:
W S
α :=
(S ⊖ zαiS).
k
\i=1
We denote by M = (Mz1, . . . , Mzn) the n-tuple of co-ordinate multiplication operators on
A2(Dn).
Theorem 3.1. Suppose S is a closed joint M-invariant subspace of A2(Dn). If S is dou-
bly commuting then for any non-empty subset α = {α1, . . . , αk} of Λn, W S
α is a wandering
subspace for MαS = (Mzα1 S, . . . , Mzαk
S).
The proof of the above theorem follows if we show the tuple of operators
(Mz1S . . . , MznS)
on S satisfies all the hypothesis of Corollary 2.4. First note that by analyticity of A2(Dn),
all the co-ordinate multiplication operators Mzi, i = 1, . . . , n on A2(Dn) are analytic. Then
MziS is also analytic for all i = 1, . . . , n. Thus the only hypothesis remains to verify is either
condition (a) or (b) of Corollary 2.4. In the next lemma we show that the tuple (Mz1, . . . , Mzn)
satisfies condition (b) and therefore so does (Mz1S . . . , MznS).
Lemma 3.2. For all 1 ≤ i ≤ n the operators
Mzi : A2(Dn) → A2(Dn),
f 7→ zif,
satisfies
for all f, g ∈ A2(Dn).
kMzif + gk2
A2(Dn) + kMzigk2
A2(Dn) ≤ 2(cid:16)kf k2
A2(Dn)(cid:17) ,
Before we prove this lemma, we recall a well known fact regarding the norm of a function in
A2(Dn) and prove an inequality. If f is in A2(Dn) with the following power series expansion
corresponding to i-th variable:
f (z1, · · · , zn) =
akzk
i ,
∞
Xk=0
8
CHATTOPADHYAY, DAS, SARKAR, AND SARKAR
where ak ∈ A2(Dn−1) for all k ∈ N, then
kf k2
A2(Dn) =
kakk2
A2(Dn−1)
(k + 1)
.
∞
Xk=0
Next we prove the following inequality (can be found in [DuS], page 277, we include the
proof for completeness) for any z, w ∈ C and k ∈ N \ {0},
(4)
z + w2
k + 1
≤ 2(cid:18)z2
k
+
w2
k + 2(cid:19) .
To this end, note that 2k(k + 2) Re(z ¯w) ≤ (k + 2)2z2 + k2w2, which follows from the
inequality (k + 2)z − kw2 ≥ 0. Now
z2 + w2
z2(k + 2)/k + w2k/(k + 2)
z2 + w2 + 2 Re(z ¯w)
k + 1
z + w2
k + 1
=
≤
+
k + 1
= 2(cid:18)z2
k
+
w2
k + 2(cid:19) .
k + 1
∞
Pk=0
Now we prove the lemma.
Proof. Let 1 ≤ i ≤ n be fixed. Let f (z1, · · · , zn) =
be the power series expansions of two functions f, g ∈ A2(Dn) with respect to zi-th variable,
where ak, bk ∈ A2(Dn−1) for all k ∈ N. Then
akzk
i and g(z1, · · · , zn) =
bkzk
i
∞
Pk=0
(Mzif + g) =
akzk+1
i +
∞
Xk=0
∞
Xk=0
bkzk
i = b0 +
Now
(ak−1 + bk)zk
i .
∞
Xk=1
kMzif + gk2
A2(Dn) = kb0k2
A2(Dn−1) +
A2(Dn−1)
∞
∞
kak−1 + bkk2
(k + 1)
Xk=1
Xk=1 kak−1k2
Xk=1
A2(Dn)(cid:17) .
+
∞
kakk2
A2(Dn−1)
kbk−1k2
A2(Dn−1)
k + 1
k + 1
A2(Dn) + kMzigk2
≤ kb0k2
A2(Dn−1) + 2
= 2 ∞
Xk=0
= 2(cid:16)kf k2
A2(Dn−1)
k
+
kbkk2
A2(Dn−1)
k + 2
! ( by (4))
!
(cid:3)
Thus the result.
WANDERING SUBSPACES OF THE BERGMAN SPACE AND THE DIRICHLET SPACE OVER Dn
9
Now we turn our discussion to the Dirichlet space over polydisc. The Dirichlet space over
D is denoted by D(D) and defined by
D(D) := {f ∈ O(D) : f ′ ∈ A2(D)}.
(k + 1)ak2. The Dirichlet space over Dn is denoted
For any f =
by D(Dn) and defined by
∞
∞
D =
Pk=0
akzk ∈ D(D), kf k2
Pk=0
D(Dn) := {f ∈ O(cid:0)D; D(Dn−1)(cid:1) : f =
Another way of expressing D(Dn) is the following
D(Dn) := D(D) ⊗ · · · ⊗ D(D)
.
n-times
{z
}
∞
Xk=0
akzk,
∞
Xk=0
(k + 1)kakk2
D(Dn−1) < ∞}.
In the above, ak ∈ D(Dn−1) for all k ∈ N. The co-ordinate multiplication operators on D(Dn)
are also denoted by Mzi, i = 1, . . . , n. Since D(Dn) contains holomorphic functions on Dn
then Mzi is analytic for all i = 1, . . . , n. Let 1 ≤ i ≤ n be fixed. Now for f ∈ D(Dn), let
i be the Taylor expansion of f corresponding to zi-th variable, where ak ∈ D(Dn−1)
f =
akzk
∞
Pk=0
for all k ∈ N. Then kf k2 =
(k + 1)kakk2
D(Dn−1) and
kM 2
zif k2 + kf k2 =
=
(k + 1)kak−2k2
D(Dn−1) +
(k + 1)kakk2
D(Dn−1)
(k − 1)kakk2
D(Dn−1) +
(k + 1)kakk2
D(Dn−1)
∞
∞
Xk=0
Xk=0
∞
Pk=0
∞
∞
Xk=2
Xk=0
Xk=1
∞
= 2
kkakk2
D(Dn−1)
= 2kMzif k2.
Therefore Mzi is concave for all i = 1, . . . , n. Thus again by Corollary 2.4, we have proved
the following result of wandering subspaces for invariant subspaces of Dirichlet space over
polydisc.
Theorem 3.3. Suppose S is a closed joint (Mz1, . . . , Mzn)-invariant subspace of D(Dn). If
S is doubly commuting then for any non-empty subset α = {α1, . . . , αk} of Λn, W S
α is a
wandering subspace for MαS = (Mzα1 S, . . . , Mzαk
S), where
W S
α =
(S ⊖ zαiS).
k
\i=1
10
CHATTOPADHYAY, DAS, SARKAR, AND SARKAR
We conclude the paper with the remark that since the hypothesis of Corollary 2.4 and The-
orem 2.5 are same then the same conclusion as in Theorem 2.5 holds for invariant subspaces
of Bergman space or Dirichlet space over polydisc.
Acknowledgment: First two authors are grateful to Indian Statistical Institute, Bangalore
Centre for warm hospitality. The fourth author was supported by UGC Centre for Advanced
Study.
References
[Aro] N. Aronszajn, Theory of reproducing kernels, Trans. of American Mathematical Soc. 68 (1950), 337-
404.
[ARS] A. Aleman, S. Richter and C. Sundberg, Beurling's theorem for the Bergman space, Acta. Math. 177
(1996), 275 -- 310.
[Beu] A. Beurling, On two problems concerning linear transformations in Hilbert space, Acta. Math. 81
(1949), 239 -- 255
[DuS] P. Duren and A. Schuster, Bergman Spaces, Math. Surveys and Monographs, 100, Amer. Math. Soc.
Providence.
[Hal] P. R. Halmos, Shifts on Hilbert spaces, J. Reine Angew. Math. 208 (1961) 102 -- 112.
[Man] V. Mandrekar, The validity of Beurling theorems in polydiscs, Proc. Amer. Math. Soc. 103 (1988),
145 -- 148.
[ReT] D. Redett and J. Tung, Invariant subspaces in Bergman space over the bidisc Proc. Amer Math Soc,
138 (2010) , 2425 -- 2430.
S. Richter, Invariant subspaces of the Dirichlet shift, J. Reine Angew. Math. 386 (1988), 205 -- 220.
[Ric]
[Rud] W. Rudin, Function theory in polydiscs, Benjamin, New York, 1969.
[Shi]
S. Shimorin, Wold-type decompositions and wandering subspaces for operators close to isometries, J.
Reine Angew. Math. 531 (2001), 147 -- 189.
[SSW] J. Sarkar, A. Sasane and B. Wick, Doubly commuting submodules of the Hardy module over polydiscs,
preprint, arXiv:1302.5312.
[Wol] H. Wold, A study in the analysis of stationary time series, Almquist and Wiksell, Uppsala, 1938.
(A. Chattopadhyay) Indian Statistical Institute, Statistics and Mathematics Unit, 8th
Mile, Mysore Road, Bangalore, 560059, India
E-mail address: [email protected], [email protected]
(B. K. Das) Indian Statistical Institute, Statistics and Mathematics Unit, 8th Mile, Mysore
Road, Bangalore, 560059, India
E-mail address: [email protected], [email protected]
(J. Sarkar) Indian Statistical Institute, Statistics and Mathematics Unit, 8th Mile, Mysore
Road, Bangalore, 560059, India
E-mail address: [email protected], [email protected]
(S. Sarkar) Department of Mathematics, Indian Institute of Science, Bangalore, 560 012,
India
E-mail address: [email protected]
|
1608.03080 | 1 | 1608 | 2016-08-10T08:32:26 | The classical theory of calculus of variations for generalized functions | [
"math.FA",
"math.DG"
] | We present an extension of the classical theory of calculus of variations to generalized functions. The framework is the category of generalized smooth functions, which includes Schwartz distributions while sharing many nonlinear properties with ordinary smooth functions. We prove full connections between extremals and Euler-Lagrange equations, classical necessary and sufficient conditions to have a minimizer, the necessary Legendre condition, Jacobi's theorem on conjugate points and Noether's theorem. We close with an application to low regularity Riemannian geometry. | math.FA | math |
THE CLASSICAL THEORY OF CALCULUS OF VARIATIONS
FOR GENERALIZED FUNCTIONS
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Abstract. We present an extension of the classical theory of calculus of vari-
ations to generalized functions. The framework is the category of generalized
smooth functions, which includes Schwartz distributions while sharing many
nonlinear properties with ordinary smooth functions. We prove full connec-
tions between extremals and Euler-Lagrange equations, classical necessary and
sufficient conditions to have a minimizer, the necessary Legendre condition, Ja-
cobi's theorem on conjugate points and Noether's theorem. We close with an
application to low regularity Riemannian geometry.
1. Introduction and motivations
Singular problems in the calculus of variations have longly been studied both in
mathematics and in relevant applications (see, e.g., [15, 5, 39, 23] and references
therein). In this paper, we introduce an approach to variational problems involving
singularities that allows the extension of the classical theory with very natural
statements and proofs. We are interested in extremizing functionals which are either
distributional themselves or whose set of extremals includes generalized functions.
Clearly, distribution theory, being a linear theory, has certain difficulties when
nonlinear problems are in play.
To overcome this type of problems, we are going to use the category of generalized
smooth functions, see [9, 10, 12, 13]. This theory seems to be a good candidate, since
it is an extension of classical distribution theory which allows to model nonlinear
singular problems, while at the same time sharing many nonlinear properties with
ordinary smooth functions, like the closure with respect to composition and several
non trivial classical theorems of the calculus. One could describe generalized smooth
functions as a methodological restoration of Cauchy-Dirac's original conception of
generalized function, see [6, 26, 22].
In essence, the idea of Cauchy and Dirac
(but also of Poisson, Kirchhoff, Helmholtz, Kelvin and Heaviside) was to view
generalized functions as suitable types of smooth set-theoretical maps obtained from
ordinary smooth maps depending on suitable infinitesimal or infinite parameters.
For example, the density of a Cauchy-Lorentz distribution with an infinitesimal
scale parameter was used by Cauchy to obtain classical properties which nowadays
are attributed to the Dirac delta, cf. [22].
In the present work, the foundation of the calculus of variations is set for func-
tionals defined by arbitrary generalized functions. This in particular applies to any
Schwartz distribution and any Colombeau generalized function, and hence justifies
the title of the present paper.
2000 Mathematics Subject Classification. 49-XX, 46F-XX, 46F30, 53B20.
Key words and phrases. Calculus of variations, Schwartz distributions, generalized functions
for nonlinear analysis, low regular Riemannian geometry.
A. Lecke has been supported by the uni:doc fellowship programme of the University of Vienna.
L. Luperi Baglini has been supported by grant M1876-N35 of the Austrian Science Fund FWF.
P. Giordano has been supported by grants P25311-N25 and P25116-N25 of the Austrian Science
Fund FWF.
1
2
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
For example, during the last years, the study of low regularly Riemannian
and Lorentzian geometry was intensified and made a huge amount of progress
(cf. [25, 24, 37, 31, 27, 35]). It was shown that the exponential map is a bi-Lipschitz
homeomorphism when metrics g ∈ C1,1 are considered, [32, 25], or that Hawking's
singularity theorem still holds when g ∈ C1,1, see [24]. However, calculus of vari-
ations in the classical sense may cease to hold when metrics with C1,1 regularity,
or below, are considered [17, 28]. This motivates the search for an alternative. In
fact, if p, q ∈ Rd and Ω(p, q) denotes the set of all Lipschitz continuous curves
connecting p and q, the natural question about what curves γ ∈ Ω(p, q) realize
the minimal g-length leads to the corresponding geodesic equation, but the Jacobi
equation is not rigorously defined. To be more precise: The Riemannian curvature
tensor exists only as an L∞
loc function on Rd and is evaluated along γ. However, the
image Im(γ) of γ has Lebesgue-measure zero, if d > 1. Thus we cannot state the
Jacobi equations properly.
In order to present a possible way out of the aforementioned problems, the
singular metric g is embedded as a generalized smooth function. In this way, the
embedding ι(g) has derivatives of all orders, valued in a suitable non-Archimedean
ring1 ρ(cid:101)R ⊇ R, and behaves very closely to a standard smooth function. We apply
our extended calculus of variations to the generalized Riemannian space (ρ(cid:101)Rd, ι(g)),
and sketch a way to translate the given problem into the language of generalized
smooth functions, solve it there, and translate it back to the standard Riemannian
space (Rd, g). Clearly, the process of embedding the singular metric g using ι(g)
introduces infinitesimal differences. This is typical in a non-Archimedean setting,
but the notion of standard part comes to help:
standard real number s, i.e. x− s ≤ r for all r ∈ R>0, then the standard part of x
is exactly s. We then show that (assuming that (Rd, g) is geodesically complete) the
standard part of the minimal length in the sense of generalized smooth functions
is the minimal length in the classical sense, and give a simple way to check if a
given (classical) geodesic is a minimizer of the length functional or not.
In this
way, the framework of generalized smooth functions is presented as a method to
solve standard problems rather than a proposal to switch into a new setting.
if x ∈ ρ(cid:101)R is infinitely close to a
The structure of the present paper is as follows. We start with an introduction
into the setting of generalized smooth functions and give basic notions concerning
generalized smooth functions and their calculus that are needed for the calculus
of variations (Section 2). The paper is self-contained in the sense that it contains
all the statements required for the proofs of calculus of variations we are going to
present. If proofs of preliminaries are omitted, we clearly give references to where
they can be found. Therefore, to understand this paper, only a basic knowledge of
distribution theory is needed.
In Section 3, we obtain some preliminary lemmas regarding the calculus of variations
with generalized smooth functions. The first variation and the notion of critical
point will be defined and studied in section 4. We prove the fundamental lemma
of calculus of variations and the full connection between critical points of a given
functional and solutions of the corresponding Euler-Lagrange equation. In section
5, we study the second variation and define the notion of local minimizer. We
also extend to generalized functions classical necessary and sufficient conditions to
have a minimizer, and we give a proof of the Legendre condition.
In Section 6,
we introduce the notion of Jacobi field and extend to generalized functions the
definition of conjugate points, so as to prove the corresponding Jacobi theorem. In
Section 7, we extend the classical Noether's theorem. We close with an application
to C1,1 Riemannian geometry in Section 8.
1I.e. a ring that contains infinitesimal and infinite numbers.
CALCULUS OF VARIATIONS FOR GF
3
Note that the work [23] already established the calculus of variations in the
setting of Colombeau generalized functions by using a comparable methodological
approach. Indeed, generalized smooth functions are related to Colombeau gener-
alized functions, and one could say that the former is a minimal extension of the
latter so as to get more general domains for generalized functions and hence the
closure with respect to composition and a better behaviour on unbounded sets.
However, there are some conceptual advantages in our approach.
(i) Whereas generalized smooth functions are closed with respect to composi-
tion, Colombeau generalized functions are not. This forces [23] to consider
only functionals defined using compactly supported Colombeau generalized
functions, i.e. functions assuming only finite values, or tempered generalized
function.
(ii) The authors of [23] are forced to consider the so called compactly supported
points c(Ω) (i.e. finite points in Ω ⊆ Rn), where the setting of generalized
smooth functions gives the possibility to consider more natural domains like
the interval [a, b] ⊆ ρ(cid:101)R. This leads us to extend in a natural way the state-
still hold when we take as a, b ∈ ρ(cid:101)R two infinite numbers such that a < b, or
as boundary points two unbounded points p, q ∈ ρ(cid:101)Rd.
ments of classical results of calculus of variations. Moreover, all our results
(iii) Furthermore, the theory of generalized smooth functions was developed to
be very user friendly, in the sense that one can avoid cumbersome "ε-wise"
proofs quite often, whereas the proofs in [23] frequently use this technique.
Thus, one could say that some of the proofs based on generalized smooth
functions are more "intrinsic" and close to the classical proofs in a standard
smooth setting. This allows a smoother approach to this new framework.
(iv) The setting of generalized smooth functions depends on a fixed infinitesimal
net (ρε)ε∈(0,1] ↓ 0, whereas the Colombeau setting considers only ρε = ε.
This added degree of freedom allows to solve singular differential equations
that are unsolvable in the classical Colombeau setting and to prove a more
general Jacobi theorem on conjugate points.
In [23] only the notion of global minimizer is defined, whereas we define
the notion of local minimizer as in [8] using a natural topology in space of
generalized smooth curves.
(v)
(vi) We obtain more classical results like the Legendre condition, and the classical
(vii)
results about Jacobi fields and conjugate points.
In addition, note that the Colombeau generalized functions can be embedded
into generalized smooth functions. Thus our approach is a natural extension
of [23].
2. Basic notions
The new ring of scalars. In this work, I denotes the interval (0, 1] ⊆ R and we will
always use the variable ε for elements of I; we also denote ε-dependent nets x ∈ RI
simply by (xε). By N we denote the set of natural numbers, including zero.
We start by defining the new simple non-Archimedean ring of scalars that extends
the real field R. The entire theory is constructive to a high degree, e.g. no ultrafilter
or non-standard method is used. For all the proofs of results in this section, see
[13, 9, 12].
Definition 1. Let ρ = (ρε) ∈ RI be a net such that limε→0 ρε = 0+, then
(i)
ε ) a ∈ R>0} is called the asymptotic gauge generated by ρ. The
I(ρ) := {(ρ−a
net ρ is called a gauge.
4
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
If P(ε) is a property of ε ∈ I, we use the notation ∀0ε : P(ε) to denote
(ii)
∃ε0 ∈ I ∀ε ∈ (0, ε0] : P(ε). We can read ∀0ε as for ε small.
(iii) We say that a net (xε) ∈ RI is ρ-moderate, and we write (xε) ∈ Rρ if
∃(Jε) ∈ I(ρ) : xε = O(Jε) as ε → 0+.
(iv) Let (xε), (yε) ∈ RI , then we say that (xε) ∼ρ (yε) if ∀(Jε) ∈ I(ρ) : xε =
) as ε → 0+. This is a congruence relation on the ring Rρ of
yε + O(J−1
moderate nets with respect to pointwise operations, and we can hence define
ε
ρ(cid:101)R := Rρ/ ∼ρ,
which we call Robinson-Colombeau ring of generalized numbers, [34, 4]. We
denote the equivalence class x ∈ ρ(cid:101)R simply by x =: [xε] := [(xε)]∼ ∈ ρ(cid:101)R.
In the following, ρ will always denote a net as in Def. 1. The infinitesimal ρ
can be chosen depending on the class of differential equations we need to solve
for the generalized functions we are going to introduce, see [14]. For motivations
exists (zε) ∈ RI such that (zε) ∼ρ 0 (we then say that (zε) is ρ-negligible) and
xε ≤ yε + zε for ε small. Equivalently, we have that x ≤ y if and only if there
exist representatives [xε] = x and [yε] = y such that xε ≤ yε for all ε. Clearly,
concerning the naturality of ρ(cid:101)R, see [12].
We can also define an order relation on ρ(cid:101)R by saying that [xε] ≤ [yε] if there
ρ(cid:101)R is a partially ordered ring. The usual real numbers r ∈ R are embedded in ρ(cid:101)R
considering constant nets [r] ∈ ρ(cid:101)R.
max ([xε], [yε]) := [max(xε, yε)] and the absolute value [xε] := [xε] ∈ ρ(cid:101)R. Note,
tomary notation ρ(cid:101)R∗ for the set of invertible generalized numbers. Our notations for
intervals are: [a, b] := {x ∈ ρ(cid:101)R a ≤ x ≤ b}, [a, b]R := [a, b]∩ R, and analogously for
segments [x, y] := {x + r · (y − x) r ∈ [0, 1]} ⊆ ρ(cid:101)Rn and [x, y]Rn = [x, y] ∩ Rn. Fi-
Even if the order ≤ is not total, we still have the possibility to define the
infimum min ([xε], [yε]) := [min(xε, yε)], and analogously the supremum function
e.g., that x ≤ z and −x ≤ z imply x ≤ z. In the following, we will also use the cus-
nally, we write x ≈ y to denote that x−y is an infinitesimal number, i.e. x−y ≤ r
for all r ∈ R>0. This is equivalent to limε→0+ xε − yε = 0 for all representatives
x = [xε] and y = [yε].
Topologies on ρ(cid:101)Rn. On the ρ(cid:101)R-module ρ(cid:101)Rn, we can consider the natural extension
of the Euclidean norm, i.e. [xε] := [xε] ∈ ρ(cid:101)R, where [xε] ∈ ρ(cid:101)Rn. Even if this
generalized norm takes values in ρ(cid:101)R, it shares several properties with usual norms,
to consider on ρ(cid:101)Rn topologies generated by balls defined by this generalized norm
Definition 2. Let R ∈(cid:110)ρ(cid:101)R∗
like the triangular inequality or the property y · x = y·x. It is therefore natural
(cid:111)
, c ∈ ρ(cid:101)Rn and x, y ∈ ρ(cid:101)R, then:
(cid:111)
(cid:110)
x ∈ ρ(cid:101)Rn x − c <R r
and a set of radii R:
≥0, R>0
for each r ∈ R.
(i) We write x <R y if ∃r ∈ R : r ≤ y − x.
(ii) BR
(iii) B E
r (c) :=
r (c) := {x ∈ Rn x − c < r}, for each r ∈ R>0, denotes an ordinary
Euclidean ball in Rn.
(cid:111)
is a base for a topology on ρ(cid:101)Rn. The
der relation a ≤ b and a (cid:54)= b (that we will never use) because for R ∈(cid:110)ρ(cid:101)R∗
r (c) r ∈ R, c ∈ ρ(cid:101)Rn(cid:111)
topology generated in the case R = ρ(cid:101)R∗
The relation <R has better topological properties as compared to the usual strict or-
≥0, R>0
≥0 is called sharp topology, whereas the one
with the set of radii R = R>0 is called Fermat topology. We will call sharply open
the set of balls
(cid:110)
BR
CALCULUS OF VARIATIONS FOR GF
5
≥0
set any open set in the sharp topology, and large open set any open set in the
Fermat topology; clearly, the latter is coarser than the former. The existence of
infinitesimal neighborhoods implies that the sharp topology induces the discrete
topology on R. This is a necessary result when one has to deal with continuous
generalized functions which have infinite derivatives. In fact, if f(cid:48)(x0) is infinite,
we have f (x) ≈ f (x0) only for x ≈ x0 , see [11, 10]. With an innocuous abuse of
language, we write x < y instead of x <ρ(cid:101)R∗
For example, ρ(cid:101)R∗
y and x <R y instead of x <R>0 y.
intervals are defined using the relation <, i.e. (a, b) := {x ∈ ρ(cid:101)R a < x < b}.
Lemma 3. Let x ∈ ρ(cid:101)R. Then the following are equivalent:
≥0 = ρ(cid:101)R>0. We will simply write Br(c) to denote an open ball in
The following result is useful to deal with positive and invertible generalized
r (c) for an open ball in the Fermat topology. Also open
the sharp topology and B F
numbers (cf. [16]).
x is invertible and x ≥ 0, i.e. x > 0.
(i)
(ii) For each representative (xε) ∈ Rρ of x we have ∀0ε : xε > 0.
(iii) For each representative (xε) ∈ Rρ of x we have ∃m ∈ N∀0ε : xε > ρm
topology is dense in [a, b].
We will also need the following result.
Lemma 4. Let a, b ∈ ρ(cid:101)R such that a < b, then the interior int ([a, b]) in the sharp
and bounded sets in ρ(cid:101)Rn is by using a net (Aε) of subsets Aε ⊆ Rn. We have two
ways of extending the membership relation xε ∈ Aε to generalized points [xε] ∈ ρ(cid:101)R:
Internal and strongly internal sets. A natural way to obtain sharply open, closed
ε
Definition 5. Let (Aε) be a net of subsets of Rn, then
(i)
(cid:110)
[xε] ∈ ρ(cid:101)Rn ∀0ε : xε ∈ Aε
(cid:111)
(ii)
is called the internal set generated by
[Aε] :=
the net (Aε). See [33] for the introduction and an in-depth study of this
notion in the case ρε = ε.
Let (xε) be a net of points of Rn, then we say that xε ∈ε Aε, and we read it
(cid:110)
(cid:111)
[xε] ∈ ρ(cid:101)Rn xε ∈ε Aε
ε) ∼ρ (xε), then
as (xε) strongly belongs to (Aε), if ∀0ε : xε ∈ Aε and if (x(cid:48)
also x(cid:48)
,
exists r ∈ ρ(cid:101)R>0 such that K ⊆ Br(0). Analogously, a net (Aε) is sharply
and we call it the strongly internal set generated by the net (Aε).
ε ∈ Aε for ε small. Moreover, we set (cid:104)Aε(cid:105) :=
(iii) Finally, we say that the internal set K = [Aε] is sharply bounded if there
bounded the internal set [Aε] is sharply bounded.
Therefore, x ∈ [Aε] if there exists a representative [xε] = x such that xε ∈ Aε for
ε small, whereas this membership is independent from the chosen representative in
the case of strongly internal sets. Note explicitly that an internal set generated by
a constant net Aε = A ⊆ Rn is simply denoted by [A].
The following theorem shows that internal and strongly internal sets have dual
topological properties:
Theorem 6. For ε ∈ I, let Aε ⊆ Rn and let xε ∈ Rn. Then we have
(i)
if and only if [d(xε, Aε)] = 0 ∈ ρ(cid:101)R.
[xε] ∈ [Aε] if and only if ∀q ∈ R>0 ∀0ε : d(xε, Aε) ≤ ρq
ε. Therefore [xε] ∈ [Aε]
[xε] ∈ (cid:104)Aε(cid:105) if and only if ∃q ∈ R>0 ∀0ε : d(xε, Ac
ε := Rn\Aε.
Therefore, if (d(xε, Ac
ε)] > 0.
[Aε] is sharply closed and (cid:104)Aε(cid:105) is sharply open.
[Aε] = [cl (Aε)], where cl (S) is the closure of S ⊆ Rn. On the other hand
(cid:104)Aε(cid:105) = (cid:104)int(Aε)(cid:105), where int (S) is the interior of S ⊆ Rn.
ε)) ∈ Rρ, then [xε] ∈ (cid:104)Aε(cid:105) if and only if [d(xε, Ac
(ii)
(iii)
(iv)
ε, where Ac
ε) > ρq
6
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Generalized smooth functions and their calculus. Using the ring ρ(cid:101)R, it is easy to
probability density by f (x, σ), and if we set σ = [σε] ∈ ρ(cid:101)R>0, where σ ≈ 0, we
Definition 7. Let X ⊆ ρ(cid:101)Rn and Y ⊆ ρ(cid:101)Rd be arbitrary subsets of generalized
consider a Gaussian with an infinitesimal standard deviation.
If we denote this
obtain the net of smooth functions (f (−, σε))ε∈I . This is the basic idea we are
going to develop in the following
points. Then we say that
f : X −→ Y is a generalized smooth function
if there exists a net fε ∈ C∞(Ωε, Rd) defining f in the sense that X ⊆ (cid:104)Ωε(cid:105), f ([xε]) =
[fε(xε)] ∈ Y and (∂αfε(xε)) ∈ Rd
ρ for all x = [xε] ∈ X and all α ∈ Nn. The space
of generalized smooth functions (GSF) from X to Y is denoted by ρGC∞(X, Y ).
Let us note explicitly that this definition states minimal logical conditions to ob-
tain a set-theoretical map from X into Y and defined by a net of smooth functions.
In particular, the following Thm. 8 states that the equality f ([xε]) = [fε(xε)] is
meaningful, i.e. that we have independence from the representatives for all deriva-
tives [xε] ∈ X (cid:55)→ [∂αfε(xε)] ∈ ρ(cid:101)Rd, α ∈ Nn.
Theorem 8. Let X ⊆ ρ(cid:101)Rn and Y ⊆ ρ(cid:101)Rd be arbitrary subsets of generalized points.
Let fε ∈ C∞(Ωε, Rd) be a net of smooth functions that defines a generalized smooth
map of the type X −→ Y , then
ε) ∈ Rn
ρ :
(i)
(ii)
(iii) For all α ∈ Nn, the GSF g : [xε] ∈ X (cid:55)→ [∂αfε(xε)] ∈ (cid:101)Rd is locally Lipschitz
∀α ∈ Nn ∀(xε), (x(cid:48)
∀[xε] ∈ X ∀α ∈ Nn ∃q ∈ R>0 ∀0ε : supy∈BE
εq (xε) ∂αuε(y) ≤ ε−q.
that g(x) − g(y) ≤ Lx − y for all x, y ∈ U and some L ∈ ρ(cid:101)R.
in the sharp topology, i.e. each x ∈ X possesses a sharp neighborhood U such
(iv) Each f ∈ ρGC∞(X, Y ) is continuous with respect to the sharp topologies in-
ε] ∈ X ⇒ (∂αuε(xε)) ∼ρ (∂αuε(x(cid:48)
[xε] = [x(cid:48)
ε)).
duced on X, Y .
(v) Assume that the GSF f is locally Lipschitz in the Fermat topology and that
its Lipschitz constants are always finite: L ∈ R. Then f is continuous in the
Fermat topology.
f : X −→ Y is a GSF if and only if there exists a net vε ∈ C∞(Rn, Rd)
defining a generalized smooth map of type X −→ Y such that f = [vε(−)]X .
(vii) Subsets S ⊆ ρ(cid:101)Rs with the trace of the sharp topology, and generalized smooth
(vi)
maps as arrows form a subcategory of the category of topological spaces. We
will call this category ρGC∞, the category of GSF.
The differential calculus for GSF can be introduced showing existence and unique-
ness of another GSF serving as incremental ratio.
Theorem 9 (Fermat-Reyes theorem for GSF). Let U ⊆ ρ(cid:101)Rn be a sharply open set,
let v = [vε] ∈ ρ(cid:101)Rn, and let f ∈ ρGC∞(U, ρ(cid:101)R) be a generalized smooth map generated
map r ∈ ρGC∞(T, ρ(cid:101)R), called the generalized incremental ratio of f along v,
by the net of smooth functions fε ∈ C∞(Ωε, R). Then
(i)
There exists a sharp neighborhood T of U × {0} and a generalized smooth
∀(x, h) ∈ T : f (x + hv) = f (x) + h · r(x, h).
such that
U × {0}.
(cid:104) ∂fε
∂vε
(cid:105)
(ii) Any two generalized incremental ratios coincide on a sharp neighborhood of
(iii) We have r(x, 0) =
(xε)
v := ∂f
∂v (x) := r(x, 0), so that ∂f
for every x ∈ U and we can thus define Df (x)·
∂v ∈ ρGC∞(U, ρ(cid:101)R).
CALCULUS OF VARIATIONS FOR GF
7
j
If U is a large open set, then an analogous statement holds replacing sharp neigh-
borhoods by large neighborhoods.
Note that this result permits to consider the partial derivative of f with respect
to an arbitrary generalized vector v ∈ ρ(cid:101)Rn which can be, e.g., infinitesimal or
(cid:16)ρ(cid:101)Rn(cid:17)j −→ ρ(cid:101)Rd over the ring ρ(cid:101)R will be denoted by
Lj(ρ(cid:101)Rn, ρ(cid:101)Rd). For A = [Aε(−)] ∈ Lj(ρ(cid:101)Rn, ρ(cid:101)Rd), we set A := [Aε], the generalized
infinite. Using this result, we can also define subsequent differentials Djf (x) as
j−multilinear maps, and we set Djf (x) · hj := Djf (x)(h,
. . . . . . , h). The set of
all the j−multilinear maps
number defined by the operator norms of the multilinear maps Aε ∈ Lj(Rn, Rd).
The following result follows from the analogous properties for the nets of smooth
functions defining f and g.
Theorem 10. Let U ⊆ ρ(cid:101)Rn be an open subset in the sharp topology, let v ∈ ρ(cid:101)Rn
and f , g : U −→ ρ(cid:101)R be generalized smooth maps. Then
∀r ∈ ρ(cid:101)R
∂v + ∂g
∂v · g + f · ∂g
(i)
(ii)
(iii)
(iv) For each x ∈ U , the map df (x).v := ∂f
(v)
∂(f +g)
∂v = ∂f
∂(r·f )
∂v = r · ∂f
∂v (x) ∈ ρ(cid:101)R is ρ(cid:101)R-linear in v ∈ ρ(cid:101)Rn
∂(f·g)
∂v = ∂f
Let U ⊆ ρ(cid:101)Rn and V ⊆ ρ(cid:101)Rd be open subsets in the sharp topology and g ∈
ρGC∞(V, U ), f ∈ ρGC∞(U, ρ(cid:101)R) be generalized smooth maps. Then for all
x ∈ V and all v ∈ ρ(cid:101)Rd, we have ∂(f◦g)
Theorem 11. Let f ∈ ρGC∞(U, ρ(cid:101)R) be a generalized smooth function defined in
the sharply open set U ⊆ ρ(cid:101)Rn. Let a, x ∈ ρ(cid:101)Rn such that the line segment [a, x] ⊆ U .
We also have a generalization of the Taylor formula:
(x) = df (g(x)) . ∂g
∂v (x).
∂v
∂v
∂v
∂v
Then, for all n ∈ N we have
∃ξ ∈ [a, x] : f (x) =
Djf (a)
· (x − a)j +
Dn+1f (ξ)
(n + 1)!
· (x − a)n+1.
(2.1)
If we further assume that all the n components (x − a)k ∈ ρ(cid:101)R of x − a ∈ ρ(cid:101)Rn are
invertible, then there exists ρ ∈ ρ(cid:101)R>0, ρ ≤ x − a, such that
j=0
j!
n(cid:88)
n(cid:88)
∀k ∈ Bρ(0)∃ξ ∈ [a− k, a + k] : f (a + k) =
Djf (a)
· kj +
Dn+1f (ξ)
(n + 1)!
· kn+1 (2.2)
j!
j=0
Dn+1f (ξ)
(n + 1)!
· kn+1 ≈ 0.
(2.3)
Formula (2.1) corresponds to a direct generalization of Taylor formulas for ordi-
nary smooth functions with Lagrange remainder. On the other hand, in (2.2) and
(2.3), the possibility that the differential Dn+1f may be infinite at some point is
considered, and the Taylor formulas are stated so as to have infinitesimal remainder.
The following local inverse function theorem will be used in the proof of Jacobi's
theorem (see [9] for a proof).
Theorem 12. Let X ⊆ ρ(cid:101)Rn, let f ∈ ρGC∞(X, ρ(cid:101)Rn) and suppose that for some x0
in the sharp interior of X, Df (x0) is invertible in L(ρ(cid:101)Rn, ρ(cid:101)Rn). Then there exists
a sharp neighborhood U ⊆ X of x0 and a sharp neighborhood V of f (x0) such that
f : U → V is invertible and f−1 ∈ ρGC∞(V, U ).
8
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
+(a) := limt→a
a<t
GSF is based on the following
We can define right and left derivatives as e.g. f(cid:48)(a) := f(cid:48)
which always exist if f ∈ ρGC∞([a, b], ρ(cid:101)Rd). One dimensional integral calculus of
Theorem 13. Let f ∈ ρGC∞([a, b], ρ(cid:101)R) be a generalized smooth function defined
in the interval [a, b] ⊆ ρ(cid:101)R, where a < b. Let c ∈ [a, b]. Then, there exists one and
only one generalized smooth function F ∈ ρGC∞([a, b], ρ(cid:101)R) such that F (c) = 0 and
F (cid:48)(x) = f (x) for all x ∈ [a, b]. Moreover, if f is defined by the net fε ∈ C∞(R, R)
and c = [cε], then F (x) =
for all x = [xε] ∈ [a, b].
f(cid:48)(t),
fε(s) ds
(cid:104)´
(cid:105)
xε
cε
´
We can thus define
´
(cid:17)(cid:48)
(−)
c
(i)
Definition 14. Under the assumptions of Theorem 13, we denote by
f (s) ds ∈ ρGC∞([a, b], ρ(cid:101)R) the unique generalized smooth function such that:
´
(cid:16)´
c
c f = 0
f
u f (s) ds = f (x) for all x ∈ [a, b].
All the classical rules of integral calculus hold in this setting:
Theorem 15. Let f ∈ ρGC∞(U, ρ(cid:101)R) and g ∈ ρGC∞(V, ρ(cid:101)R) be generalized smooth
functions defined on sharply open domains in ρ(cid:101)R. Let a, b ∈ ρ(cid:101)R with a < b and c,
(x) = d
dx
(−)
c
(−)
u
f :=
(ii)
´
x
d ∈ [a, b] ⊆ U ∩ V , then
(i)
´
´
´
´
´
´
´
´
∀λ ∈ ρ(cid:101)R
´
d
d
d
c (f + g) =
c f +
c g
´
´
d
d
c λf = λ
c f
e f for all e ∈ [a, b]
´
d
e
c f =
c f +
c f = −
d
c
d f
c f(cid:48) = f (d) − f (c)
c f(cid:48) · g = [f · g]d
c −
´
c f · g(cid:48)
d
d
d
d
(ii)
(iii)
(iv)
(v)
(vi)
Theorem 16. Let f ∈ ρGC∞(U, ρ(cid:101)R) and ϕ ∈ ρGC∞(V, U ) be generalized smooth
functions defined on sharply open domains in ρ(cid:101)R. Let a, b ∈ ρ(cid:101)R, with a < b, such
that [a, b] ⊆ V , ϕ(a) < ϕ(b), [ϕ(a), ϕ(b)] ⊆ U . Finally, assume that ϕ([a, b]) ⊆
[ϕ(a), ϕ(b)]. Then
ϕ(b)
f (t) dt =
ϕ(a)
a
b
f [ϕ(s)] · ϕ(cid:48)(s) ds.
Embedding of Schwartz distributions and Colombeau functions. We finally recall
two results that give a certain flexibility in constructing embeddings of Schwartz
distributions. Note that both the infinitesimal ρ and the embedding of Schwartz
distributions have to be chosen depending on the problem we aim to solve. A trivial
example in this direction is the ODE y(cid:48) = y/ dε, which cannot be solved for ρ = (ε),
but it has a solution for ρ = (e−1/ε). As another simple example, if we need the
property H(0) = 1/2, where H is the Heaviside function, then we have to choose
the embedding of distributions accordingly. See also [14, 30] for further details.
If ϕ ∈ D(Rn), r ∈ R>0 and x ∈ Rn, we use the notations r (cid:12) ϕ for the function
x ∈ Rn (cid:55)→ 1
These notations permit to highlight that (cid:12) is a free action of the multiplicative
group (R>0,·, 1) on D(Rn) and ⊕ is a free action of the additive group (R>0, +, 0)
on D(Rn). We also have the distributive property r (cid:12) (x ⊕ ϕ) = rx ⊕ r (cid:12) ϕ.
(cid:1) ∈ R and x ⊕ ϕ for the function y ∈ Rn (cid:55)→ ϕ(y − x) ∈ R.
Lemma 17. Let b ∈ ρ(cid:101)R be a net such that limε→0+ bε = +∞. Let d ∈ (0, 1). There
exists a net (ψε)ε∈I of D(Rn) with the properties:
rn · ϕ(cid:0) x
r
CALCULUS OF VARIATIONS FOR GF
9
´
supp(ψε) ⊆ B1(0) for all ε ∈ I.
ψε = 1 for all ε ∈ I.
´
∀j ∈ N∀0ε : 1 ≤ α ≤ j ⇒
ψε ≤ 1 + η.
∀η ∈ R>0 ∀0ε :
If n = 1, then the net (ψε)ε∈I can be chosen so that
(i)
(ii)
(iii) ∀α ∈ Nn ∃p ∈ N : supx∈Rn ∂αψε(x) = O(bp
(iv)
(v)
(vi)
In particular ψb
ε (cid:12) ψε satisfies (ii) - (v).
xα · ψε(x) dx = 0.
ε := b−1
´
ε) as ε → 0+.
´
0
−∞ ψε = d.
It is worth noting that the condition (iv) of null moments is well known in the
study of convergence of numerical solutions of singular differential equations, see
e.g. [38, 7, 19] and references therein.
Concerning embeddings of Schwartz distributions, we have the following result,
where c(Ω) := {[xε] ∈ [Ω] ∃K (cid:98) Ω∀0ε : xε ∈ K} is called the set of compactly
supported points in Ω ⊆ Rn.
Theorem 18. Under the assumptions of Lemma 17, let Ω ⊆ Rn be an open set
and let (ψb
ε) be the net defined in 17. Then the mapping
Ω : T ∈ E(cid:48)(Ω) (cid:55)→(cid:2)(cid:0)T ∗ ψb
ε
(cid:1) (−)(cid:3) ∈ ρGC∞(c(Ω), ρ(cid:101)R)
ιb
uniquely extends to a sheaf morphism of real vector spaces
ιb : D(cid:48) −→ ρGC∞(c((−)), ρ(cid:101)R),
(−) : C∞(−) −→ ρGC∞(c((−)), ρ(cid:101)R)
and satisfies the following properties:
(i)
´
If b ≥ dρ−a for some a ∈ R>0, then ιbC∞
is a sheaf morphism of algebras;
If T ∈ E(cid:48)(Ω) then supp(T ) = supp(ιb
ιb commutes with partial derivatives, i.e. ∂α(cid:0)ιb
limε→0+
T ∈ D(cid:48)(Ω) and α ∈ N.
Ω(T ));
Ω ιb
(ii)
(iii)
(iv)
Ω(T )(cid:1) = ιb
Ω(T )ε · ϕ = (cid:104)T, ϕ(cid:105) for all ϕ ∈ D(Ω) and all T ∈ D(cid:48)(Ω);
Ω (∂αT ) for each
Concerning the embedding of Colombeau generalized functions, we recall that the
special Colombeau algebra on Ω is defined as the quotient Gs(Ω) := EM (Ω)/N s(Ω)
of moderate nets over negligible nets, where the former is
EM (Ω) := {(uε) ∈ C∞(Ω)I ∀K (cid:98) Ω∀α ∈ Nn ∃N ∈ N : sup
x∈K
∂αuε(x) = O(ε−N )}
and the latter is
N s(Ω) := {(uε) ∈ C∞(Ω)I ∀K (cid:98) Ω∀α ∈ Nn ∀m ∈ N : sup
x∈K
∂αuε(x) = O(εm)}.
Using ρ = (ε), we have the following compatibility result:
Theorem 19. A Colombeau generalized function u = (uε) + N s(Ω)d ∈ Gs(Ω)d
defines a generalized smooth map u : [xε] ∈ c(Ω) −→ [uε(xε)] ∈ (cid:101)Rd which is locally
assignment provides a bijection of Gs(Ω)d onto ρGC∞(c(Ω), ρ(cid:101)Rd) for every open set
Lipschitz on the same neighborhood of the Fermat topology for all derivatives. This
Ω ⊆ Rn.
2.1. Extreme value theorem and functionally compact sets. For GSF, suit-
able generalizations of many classical theorems of differential and integral calculus
hold: intermediate value theorem, mean value theorems, a sheaf property for the
Fermat topology, local and global inverse function theorems, Banach fixed point
theorem and a corresponding Picard-Lindelof theorem, see [13, 12, 29, 9].
Even though the intervals [a, b] ⊆(cid:101)R, a, b ∈ R, are neither compact in the sharp
nor in the Fermat topology (see [13, Thm. 25]), analogously to the case of smooth
functions, a GSF satisfies an extreme value theorem on such sets. In fact, we have:
10
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Theorem 20. Let f ∈ GC∞(X,(cid:101)R) be a generalized smooth function defined on the
subset X of (cid:101)Rn. Let ∅ (cid:54)= K = [Kε] ⊆ X be an internal set generated by a sharply
bounded net (Kε) of compact sets Kε (cid:98) Rn , then
∃m, M ∈ K ∀x ∈ K : f (m) ≤ f (x) ≤ f (M ).
(2.4)
We shall use the assumptions on K and (Kε) given in this theorem to introduce
a notion of "compact subset" which behaves better than the usual classical notion
of compactness in the sharp topology.
Definition 21. A subset K of (cid:101)Rn is called functionally compact, denoted by K (cid:98)f
(cid:101)Rn, if there exists a net (Kε) such that
(i) K = [Kε] ⊆(cid:101)Rn
If, in addition, K ⊆ U ⊆(cid:101)Rn then we write K (cid:98)f U . Finally, we write [Kε] (cid:98)f U if
(ii)
(iii) ∀ε ∈ I : Kε (cid:98) Rn
(Kε) is sharply bounded
(ii), (iii) and [Kε] ⊆ U hold.
We motivate the name functionally compact subset by noting that on this type of
subsets, GSF have properties very similar to those that ordinary smooth functions
have on standard compact sets.
Remark 22.
By [33, Prop. 2.3], any internal set K = [Kε] is closed in the sharp topology.
In particular, the open interval (0, 1) ⊆ (cid:101)R is not functionally compact since
(i)
(ii)
(iv)
the GSF f (x) = x does not satisfy the conclusion (2.4) of Prop. 20.
In the present paper, we need the following properties of functionally compact sets.
it is not closed.
If H (cid:98) Rn is a non-empty ordinary compact set, then [H] is functionally
compact. In particular, [0, 1] = [[0, 1]R] is functionally compact.
(cid:101)Rn is not functionally compact since it is not sharply bounded.
(iii) The empty set ∅ =(cid:101)∅ (cid:98)f (cid:101)R.
(v) The set of compactly supported points(cid:101)Rc is not functionally compact because
Theorem 23. Let K ⊆ X ⊆ (cid:101)Rn, f ∈ GC∞(X,(cid:101)Rd). Then K (cid:98)f (cid:101)Rn implies
f (K) (cid:98)f (cid:101)Rd.
Corollary 24. If a, b ∈(cid:101)R and a ≤ b, then [a, b] (cid:98)f (cid:101)R.
Let us note that a, b ∈ (cid:101)R can also be infinite, e.g. a = [−ε−N ], b = [ε−M ] or
Theorem 25. Let K (cid:98)f (cid:101)Rn and H (cid:98)f (cid:101)Rd, then K × H (cid:98)f (cid:101)Rn+d. In particular, if
ai ≤ bi for i = 1, . . . , n, then(cid:81)n
a = [ε−N ], b = [ε−M ] with M > N . Finally, in the following result we consider the
product of functionally compact sets:
As a corollary of this theorem and Rem. (22).(ii) we get
i=1[ai, bi] (cid:98)f (cid:101)Rn.
A theory of compactly supported GSF has been developed in [9], and it closely
resembles the classical theory of LF-spaces of compactly supported smooth func-
tions. It establishes that for suitable functionally compact subsets, the correspond-
ing space of compactly supported GSF contains extensions of all Colombeau gen-
eralized functions, and hence also of all Schwartz distributions.
CALCULUS OF VARIATIONS FOR GF
11
3. Preliminary results for calculus of variations with GSF
In this section, we study extremal values of generalized functions at sharply
interior points of intervals [a, b] ⊆ ρ(cid:101)R. As in the classical calculus of variations, this
variational problems. Since the new ring of scalars ρ(cid:101)R has zero divisors and is not
will provide the basis for proving necessary and sufficient conditions for general
totally ordered, the following extension requires a more refined analysis than in the
classical case.
b
a
d
ds
b
a
∂
∂s
f (τ, s) dτ =
The following lemma shows that we can interchange integration and differentia-
tion while working with generalized functions.
Lemma 26. Let a, b, c, d ∈ ρ(cid:101)R, with a < b and c < d. Let f ∈ ρGC∞(X, Y ) and
assume that [a, b] × [c, d] ⊆ X ⊆ ρ(cid:101)R2 and Y ⊆ ρ(cid:101)Rd. Then for all s ∈ [c, d],we have
(3.1)
Proof. We first note that f (·, s) ∈ ρGC∞([a, b], Y ) by the closure of GSF with respect
to composition. Therefore, ∂
(3.1) is well defined as an integral of a GSF. In order to show that also the left hand
side of (3.1) is well-defined, we need to prove that also σ ∈ [c, d] (cid:55)→
∂s f (·, s) ∈ ρGC∞([a, b], ρ(cid:101)Rd), and the right hand side of
´
ρ(cid:101)Rd is a GSF. Let f be defined by the net fε ∈ C∞(cid:0)Ωε, Rd(cid:1), with X ⊆ (cid:104)Ωε(cid:105). Let
a f (τ, σ) dτ ∈
ρ(cid:101)R2 and the extreme value theorem 20 applied
(cid:12)(cid:12)(cid:12)(cid:12) ∂n
[σε] ∈ [c, d], then [a, b] × {[σε]} (cid:98)f
to ∂nf
∂σn yields the existence of N ∈ R>0 such that
(cid:12)(cid:12)(cid:12)(cid:12) dτ ≤ ρ−N
∂σn fε(τ, σε)
· (bε − aε).
fε(τ, σε) dτ
f (τ, s) dτ.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dn
dσn
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
bε
aε
bε
aε
b
This proves that also the left hand side of (3.1) is well-defined as a derivative of
a GSF. From the classical derivation under the integral sign, the Fermat-Reyes
theorem 9, and Thm. 13 about definite integrals of GSF, we obtain
d
ds
b
f (τ, s) dτ =
d
ds
a
ε
(cid:35)
(cid:35)
b
a
(cid:34)
(cid:20) ∂
∂s
aε
bε
aε
[fε(τ, s)] dτ
bε
fε(τ, s) dτ
fε(τ, s) dτ
(cid:21)
fε(τ, s)
dτ
=
=
=
=
d
ds
d
ds
(cid:34)
b
b
a
a
∂
∂s
f (τ, s) dτ.
(cid:3)
The next result will frequently be used in the following
Lemma 27. Let (D,≥) be a directed set and let f : D −→ ρ(cid:101)R be a set-theoretical
map such that f (d) ≥ 0 for all d ∈ D, and ∃ limd∈D f (d) ∈ ρ(cid:101)R in the sharp topology.
Then limd∈D f (d) ≥ 0.
Proof. Note that the internal set [0, +∞) = [[0, +∞)R] is sharply closed by Thm. 6.(iii).
(cid:3)
12
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
(i)
Remark 28.
If x ∈ ρ(cid:101)R, then x ≥ 0 if and only if ∃A ∈ R>0 ∀a ∈ R>A : x ≥ −dρa. Indeed,
it suffices to let a → +∞ in f (a) = x + dρa.
(ii) Assume that x, y ∈ ρ(cid:101)Rn and
Then taking s → 0 in f (s) = sy − x we get x = 0.
∃s0 ∈ ρ(cid:101)R>0 ∀s ∈ ρ(cid:101)R>0 : s ≤ s0 ⇒ x ≤ sy.
Definition 29. We call x = (x1, . . . , xd) ∈ ρ(cid:101)Rd componentwise invertible if and
only if for all k ∈ {1, . . . , d} we have that xk ∈ ρ(cid:101)R is invertible.
Lemma 30. Let f ∈ ρGC∞(U, Y ) where Y ⊆ ρ(cid:101)R and U ⊆ ρ(cid:101)Rd is a sharply open
Proof. By Lem. 3, it follows that for V ⊆ ρ(cid:101)R, the set of invertible points in V,
i.e. V ∩ ρ(cid:101)R∗ ⊆ V is dense in V (with respect to the sharp topology). This implies
that U ∩ (ρ(cid:101)Rd)∗ ⊆ U is dense. By Thm. 8.(iv), f is sharply continuous, so Lem. 27
subset. Then f ≥ 0 if and only if f (x) ≥ 0 for all componentwise invertible x ∈ U .
yields that f (x) ≥ 0. The other direction is obvious.
(cid:3)
Analogously to the classical case, we say that x0 ∈ X is a local minimum of
f ∈ ρGC∞(X) if there exists a sharply open neighbourhood (in the trace topology)
Y ⊆ X of x0 such that f (x0) ≤ f (y) for all y ∈ Y . A local maximum is defined
accordingly. We will write f (x0) = min!, which is a short hand notation to denote
that x0 is a (local) minimum of f .
Lemma 31. Let X ⊆ ρ(cid:101)R and let f ∈ ρGC∞(X, ρ(cid:101)R). If x0 ∈ X is a sharply interior
GSF with respect to composition. Let r ∈ ρ(cid:101)R>0 be such that B2r(0) =: U ⊆ X and
f (0) = min! over U . Take any x ∈ ρ(cid:101)R such that 0 < x < r, so that [−x,x] ⊆ U .
Proof. Without loss of generality, we can assume x0 = 0, because of the closure of
local minimum of f then f(cid:48)(x0) = 0.
Thus, if x > 0, by Taylor's theorem 11 there exists ξ ∈ [0, x] such that
Set K := [Brε(0)] (cid:98)f B2r(0) ⊆ U and M := maxx∈K f(cid:48)(cid:48)(x) ∈ ρ(cid:101)R≥0. Due to the
2
f (x) = f (0) + f(cid:48)(0) · x +
f(cid:48)(cid:48)(ξ)
· x2.
fact that f (0) is minimal, we have
f(cid:48)(cid:48)(ξ)
2
(3.2)
2 x = M
As a corollary of Lem. 4 and Thm. 8.(iv), we have
2 x. Therefore, f(cid:48)(0) ≤ M
f(cid:48)(0) · x +
2 x2 and −f(cid:48)(0) ≤ M
Thus −f(cid:48)(0) · x ≤ M
x < 0, we get f(cid:48)(0) ≤ − M
follows by Rem. 28.(ii).
· x2 = f (x) − f (0) ≥ 0.
2 x since x > 0. Analogously, if we take
2 x and the conclusion
(cid:3)
Lemma 32. Let a, b ∈ ρ(cid:101)R with a < b and let f ∈ ρGC∞([a, b], ρ(cid:101)Rd) such that
Lemma 33. Let a, b ∈ ρ(cid:101)R with a < b and let f ∈ ρGC∞([a, b], ρ(cid:101)R) such that
f (x) = 0 for all sharply interior points x ∈ [a, b]. Then f = 0 on [a, b].
Now, we are able to prove the "second - derivative - test" for GSF.
f (x0) = min! for some sharply interior x0 ∈ [a, b]. Then f(cid:48)(cid:48)(x0) ≥ 0. Vice versa,
if f(cid:48)(x0) = 0 and f(cid:48)(cid:48)(x0) > 0, then f (x0) = min!.
CALCULUS OF VARIATIONS FOR GF
Proof. As above, we can assume x0 = 0. Let r ∈ ρ(cid:101)R>0 be such that B2r(0) =: U ⊆
X and f (0) = min! over U . Take any x ∈ ρ(cid:101)R such that 0 < x < r, so that [0, x] ⊆ U ,
and set K := [Brε(0)] (cid:98)f B2r(0) ⊆ U and M := maxx∈K f(cid:48)(cid:48)(cid:48)(x) ∈ ρ(cid:101)R≥0. By
13
Taylor's theorem 11, we obtain that for some ξ ∈ [0, x]
f(cid:48)(cid:48)(0)x2 +
1
6
1
2
By assumption, we have that for all a ∈ R>0
f (x) = f (0) + f(cid:48)(0)x +
f(cid:48)(cid:48)(cid:48)(ξ)x3.
By Lemma 31, we know that f(cid:48)(0) = 0. Thus, we obtain for all a ∈ R>0
0 ≤ f (x) − f (0) + dρa.
1
2
1
6
f (x) − f (0) =
f(cid:48)(cid:48)(0)x2 +
f(cid:48)(cid:48)(cid:48)(ξ)x3 ≥ −dρa.
2 f(cid:48)(cid:48)(0)x2 + 1
assuming that a > A and dρA < r. We get f(cid:48)(cid:48)(0) ≥ −(cid:0)2 + M
Therefore, also 1
conclusion follows from Lem. 27 as a → +∞.
by Lem. 3. Since f(cid:48)(0) = 0, for all x ∈ Br(0), Taylor's formula gives
Now assume that f(cid:48)(0) = 0 and f(cid:48)(cid:48)(0) > 0, so that f(cid:48)(cid:48)(0) > dρa for some a ∈ R>0
6 M x3 ≥ −dρa. In this inequality we can set x = dρa/3,
(cid:1) dρa/3, and the
3
f(cid:48)(cid:48)(cid:48)(ξx)x3,
1
6
2 dρa + 1
6 f(cid:48)(cid:48)(cid:48)(ξx)x(cid:1). Now
Mx → 0 as x → 0.
1
2
(cid:12)(cid:12)(cid:12)(cid:12) 1
f(cid:48)(cid:48)(0)x2 +
f (x) − f (0) =
where ξx ∈ [0, x]. Therefore, f (x) − f (0) > x2(cid:0) 1
(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1
We can hence write f (x)−f (0) > x2(cid:0) 1
∃s ∈ ρ(cid:101)R>0 : s < r, ∀x ∈ Bs(0) : − 1
2 dρa − 1
f(cid:48)(cid:48)(cid:48)(ξx)x
Thus
6
6
which proves that x = 0 is a local minimum.
1
6
dρa <
4 dρa(cid:1) = x2 1
4
1
4
dρa.
f(cid:48)(cid:48)(cid:48)(ξx)x <
4 dρa ≥ 0 for all x ∈ Bs(0),
(cid:3)
For the generalization of Lem. 31 and Lem. 33 to the multivariate case, one can
proceed as above, using the ideas of [23]. Note, however, that we do not need this
generalization in the present work.
4. First variation and critical points
In this section, we define the first variation of a functional and prove that some
classical results have their counterparts in this generalized setting, for example the
fundamental lemma (Lem. 37) or the connection between critical points and the
Euler-Lagrange equations (Thm. 38).
Definition 34. If a, b ∈ ρ(cid:101)R and a < b, we define
ρGC∞
0 (a, b) :=
(cid:111)
(cid:110)
η ∈ ρGC∞(ρ(cid:101)R, ρ(cid:101)Rd) : η(a) = 0 = η(b)
0 (a, b) is an ρ(cid:101)R-module.
.
0 . We also note here that ρGC∞
When the use of the points a, b is clear from the context, we adopt the simplified
notation ρGC∞
One of the positive features of the use of GSF for the calculus of variations is
their closure with respect to composition. For this reason, the next definition of
functional is formally equal to the classical one, though it can be applied to arbitrary
generalized functions F and u.
14
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Theorem 35. Let a, b ∈ ρ(cid:101)R with a < b. Let u ∈ ρGC∞([a, b], ρ(cid:101)Rd) and let F ∈
ρGC∞([a, b] × ρ(cid:101)Rd × ρ(cid:101)Rd, ρ(cid:101)R) and define
b
I(u) :=
F (t, u, u) dt.
(4.1)
Let η ∈ ρGC∞
0 , then
δI(u; η) :=
d
ds
I(u + sη)
a
=
a
(cid:18)
b
η
(cid:12)(cid:12)(cid:12)(cid:12)s=0
(cid:19)
F u(t, u, u)
dt.
Fu(t, u, u) − d
dt
Proof. We have (we use Thm. 10, Thm. 15 and Lemma 26)
(cid:12)(cid:12)(cid:12)(cid:12)s=0
d
ds
I(u + sη)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)(cid:12)s=0
dt
b
=
=
=
d
ds
b
b
a
a
F (t, u + sη, u + s η) dt
a
∂
∂s
F (t, u + sη, u + s η)
ηFu(t, u, u) + ηF u(t, u, u) dt
(cid:18)
b
η
a
(cid:18)
= [ηF u(t, u, u)]b
a +
b
=
η
a
Fu(t, u, u) − d
dt
Fu(t, u, u) − d
dt
(cid:19)
F u(t, u, u)
dt.
(cid:19)
F u(t, u, u)
dt
We call δI(u; η) the first variation of I. In addition we call u ∈ ρGC∞([a, b], ρ(cid:101)Rd)
a critical point of I if δI(u; η) = 0 for all η ∈ ρGC∞
0 .
(cid:3)
To prove the fundamental lemma of the calculus of variations, Lem. 37, we first
show that every GSF can be approximated using generalized strict delta nets.
Lemma 36. Let a, b ∈ ρ(cid:101)R be such that a < b and let f ∈ ρGC∞([a, b], ρ(cid:101)R). Let x ∈
[a, b] and R ∈ ρ(cid:101)R>0 be such that BR(x) ⊆ [a, b]. Assume that Gt ∈ ρGC∞(ρ(cid:101)R, ρ(cid:101)R)
satisfy
´
−R Gt = 1 for t ∈ ρ(cid:101)R>0 small.
R
(i)
(ii) For t small, (Gt)t∈ρ(cid:101)R>0
∀δ ∈ ρ(cid:101)R>0 ∃ρ ∈ ρ(cid:101)R>0 ∀t ∈ Bρ(0) ∩ ρ(cid:101)R>0 ∀y ∈ [−R,−δ] ∪ [δ, R] : Gt(y) = 0. (4.2)
(iii) ∃M ∈ ρ(cid:101)R>0 ∃ρ ∈ ρ(cid:101)R∀t ∈ Bρ(0) :
´
−R Gt(y) dy ≤ M .
is zero outside every ball Bδ(0), 0 < δ < R, i.e.
R
R
Then
Moreover
lim
t→0+
´
−R
−R f (x − y)Gt(y) dy =
R
f (x − y)Gt(y) dy = f (x).
´
x−R f (y)Gt(x − y) dy.
x+R
Proof. We only have to generalize the classical proof concerning limits of convolu-
tions with strict delta nets. We first note that
R
−R
f (x − y)Gt(y) dy =
f (y)Gt(x − y) dy
x+R
x−R
CALCULUS OF VARIATIONS FOR GF
15
so that these integrals exist because (x − R, x + R) = BR(x) ⊆ [a, b]. Using (i), for
t small, let's say for 0 < t < S ∈ ρ(cid:101)R>0, we get
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
f (x − y)Gt(y) dy − f (x)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
−R
R
R
−R
R
[f (x − y) − f (x)] Gt(y) dy
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤
f (x − y) − f (x) · Gt(y) dy.
−R
we have
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
For each r ∈ ρ(cid:101)R>0, sharp continuity of f at x yields f (x − y) − f (x) < r for all y
such that y < δ ∈ ρ(cid:101)R>0, and we can take δ < R. By (ii), for 0 < t < min(ρ, S),
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ r
The right hand side of (4.3) can be taken arbitrarily small in ρ(cid:101)R>0 because [−δ, δ] (cid:98)f
ρ(cid:101)R, (iii) and because of the extreme value theorem 20 applied to the GSF Gt. (cid:3)
Lemma 37 (Fundamental Lemma of the Calculus of Variations). Let a, b ∈ ρ(cid:101)R
such that a < b, and let f ∈ ρGC∞([a, b], ρ(cid:101)R). If
f (x − y)Gt(y) dy − f (x)
Gt(y) dy.
(4.3)
−R
−δ
+δ
R
b
f (t)η(t) dt = 0 for all η ∈ ρGC∞
0 ,
(4.4)
a
then f = 0.
Proof. Let x ∈ [a, b]. Because of Thm. 8.(iv) and Lem. 4, without loss of generality
we can assume that x is a sharply interior point, so that BR(x) ⊆ [a, b] for some
, where
R ∈ ρ(cid:101)R>0. Let ϕ ∈ D[−1,1](R) be such that
x ∈ R and t ∈ ρ(cid:101)R>0, and Gt(x) := [Gt,ε(xε)] for all x ∈ ρ(cid:101)R. Then, for t sufficiently
small, we have Gt(x − .) ∈ ρGC∞
a f (y)Gt(x − y) dy = 0. For t
small, we both have that Gt(x− .) = 0 on [a, x− R]∪ [x + R, b] and the assumptions
of Lem. 36 hold. Therefore
ϕ = 1. Set Gt,ε(x) := 1
tε
0 and (4.4) yields
(cid:16) x
(cid:17)
´
´
ϕ
tε
b
a
R
=
−R
0 =
b
f (y)Gt(x − y) dy =
f (x − y)Gt(y) dy,
x+R
x−R
f (y)Gt(x − y) dy =
(cid:3)
and Lem. 36 hence yields f (x) = 0.
Thus we obtain the following
Theorem 38. Let a, b ∈ ρ(cid:101)R such that a < b, and let u ∈ ρGC∞([a, b], ρ(cid:101)Rd). Then
u solves the Euler-Lagrange equations
for I given by (4.1), if and only if δI(u; η) = 0 for all η ∈ ρGC∞
u is a critical point of I.
0 , i.e. if and only if
Fu − d
dt
F u = 0
(4.5)
5. second variation and minimizers
As in the classical case (see e.g. [8]), thanks to the extreme value theorem 20 and
the property of the interval [a, b] of being functionally compact, we can naturally
define a topology on the space ρGC∞([a, b], ρ(cid:101)Rd):
16
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Definition 39. Let a, b ∈ ρ(cid:101)R, with a < b. Let m ∈ N0 and v ∈ ρGC∞([a, b], ρ(cid:101)Rd).
(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) ,
(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
∈ ρ(cid:101)R,
(cid:107)v(cid:107)m := max
n≤m
1≤i≤d
max
dtn vi(Mni)
dtn vi(mni)
Then
where Mni, mni ∈ [a, b] satisfy
dn
∀t ∈ [a, b] :
dtn vi(mni) ≤ dn
dtn vi(t) ≤ dn
dtn vi(Mni).
The following result permits to calculate the (generalized) norm (cid:107)v(cid:107)m using any
net (vε) that defines v.
(i)
(cid:20)
Lemma 40. Under the assumptions of Def. 39, let a = [aε] and b = [bε] be such
that aε < bε for all ε. Then we have:
If the net (vε) defines v, then (cid:107)v(cid:107)m =
(cid:107)v(cid:107)m ≥ 0;
ε(t)(cid:12)(cid:12)(cid:21)
∀c ∈ ρ(cid:101)R : (cid:107)c · v(cid:107)m = c · (cid:107)v(cid:107)m;
For all u ∈ ρGC∞([a, b], ρ(cid:101)Rd), we have (cid:107)u+v(cid:107)m ≤ (cid:107)u(cid:107)m+(cid:107)v(cid:107)m and (cid:107)u·v(cid:107)m ≤
cm · (cid:107)u(cid:107)m · (cid:107)v(cid:107)m for some cm ∈ ρ(cid:101)R>0.
(ii)
(iii) (cid:107)v(cid:107)m = 0 if and only if v = 0;
(iv)
(v)
max n≤m
1≤i≤d
(cid:12)(cid:12) dn
maxt∈[aε,bε]
dtn vi
;
Proof. By the standard extreme value theorem applied ε-wise, we get the existence
of ¯mniε, ¯Mniε ∈ [aε, bε] such that
dn
dtn vi
ε( ¯mniε) ≤ dn
ε(t) ≤ dn
ε( ¯Mniε).
dtn vi
dtn vi
Hence(cid:12)(cid:12) dn
dtn vi
max
n≤m
1≤i≤d
max
n≤m
1≤i≤d
max
t∈[aε,bε]
∀t ∈ [aε, bε] :
ε(t)(cid:12)(cid:12) ≤ max(cid:0)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) dn
dtn vi
dtn vi
ε(t)
ε(t)
(cid:12)(cid:12)(cid:12)(cid:12)
max
t∈[aε,bε]
dtn vi
max
dtn vi
ε( ¯mniε)
dtn vi
n≤m
1≤i≤d
ε( ¯Mniε)(cid:12)(cid:12)(cid:1). Thus
ε( ¯mniε)(cid:12)(cid:12) ,(cid:12)(cid:12) dn
(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) ≤ max
(cid:12)(cid:12)(cid:12)(cid:12) ,
max
=
(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) ,
(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) dn
(cid:12)(cid:12)(cid:12)(cid:12) dn
dtn vi( ¯mni)
n≤m
1≤i≤d
dtn vi
(cid:12)(cid:12)(cid:12)(cid:12) ,
ε( ¯mniε)
max
max
= max
n≤m
1≤i≤d
But ¯mniε, ¯Mniε ∈ [aε, bε], so
dtn vi
ε( ¯Mniε)
dtn vi
ε( ¯Mniε)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
.
dtn vi( ¯Mni)
.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19) =
This proves both that (cid:107)v(cid:107)m is well-defined, i.e. it does not depend on the particular
choice of points mni, Mni as in Def. 39, and the claim (i). The remaining properties
(ii) - (v) follows directly from (i) and the usual properties of standard Cm-norms. (cid:3)
Using these ρ(cid:101)R-valued norms, we can naturally define a topology on the space
ρGC∞([a, b], ρ(cid:101)Rd).
Definition 41. Let a, b ∈ ρ(cid:101)R, with a < b. Let m ∈ N, u ∈ ρGC∞([a, b], ρ(cid:101)Rd),
r ∈ ρ(cid:101)R>0, then
(cid:110)
v ∈ ρGC∞([a, b], ρ(cid:101)Rd) (cid:107)v − u(cid:107)m < r
∀u ∈ U ∃m ∈ N∃r ∈ ρ(cid:101)R>0 : Bm
If U ⊆ ρGC∞([a, b], ρ(cid:101)Rd), then we say that U is a sharply open set if
r (u) ⊆ U.
r (u) :=
(cid:111)
Bm
(ii)
(i)
CALCULUS OF VARIATIONS FOR GF
17
As in [13, Thm. 2], one can easily prove that sharply open sets form a topology on
ρGC∞([a, b], ρ(cid:101)Rd). Using this topology, we can define when a curve is a minimizer
numbers a, b ∈ ρ(cid:101)R, a < b. E.g. they can also both be infinite.
Definition 42. Let a, b ∈ ρ(cid:101)R, with a < b and let u ∈ ρGC∞([a, b], ρ(cid:101)Rd), then
of the functional I. Note explicitly that there are no restrictions on the generalized
For all p, q ∈ ρ(cid:101)Rd, we set
(cid:110)
v ∈ ρGC∞([a, b], ρ(cid:101)Rd) v(a) = p, v(b) = q
ρGC∞
bd(p, q) :=
(cid:111)
(i)
.
(ii) We say that u is a local minimizer of I in ρGC∞
Note that ρGC∞
values".
bd(0, 0) = ρGC∞
∃r ∈ ρ(cid:101)R>0 ∃m ∈ N∀v ∈ Bm
0 . The subscript "bd" stands here for "boundary
bd(p, q) if u ∈ ρGC∞
bd(p, q) : I(v) ≥ I(u)
r (u) ∩ ρGC∞
bd(p, q) and
(5.1)
(iii) We define the second variation of I in direction η ∈ ρGC∞
0 as
δ2I(u; η) :=
Note also explicitly that the points p, q ∈ ρ(cid:101)Rd can have infinite norm, e.g. pε →
+∞ as ε → 0. We calculate, by using the standard Einstein's summation conven-
tions
I(u + sη).
(cid:12)(cid:12)(cid:12)(cid:12)0
d2
ds2
(cid:12)(cid:12)(cid:12)(cid:12)0
∂2
∂s2
a
b
(cid:12)(cid:12)(cid:12)(cid:12)0
δ2I(u; η) =
=
=
d2
ds2
b
a
b
a
F (t, u + sη, u + s η) dt
F (t, u + sη, u + s η) dt
Fuiuj (t, u, u)ηiηj + 2Fui uj (t, u, u)ηi ηj + F ui uj (t, u, u) ηi ηj dt,
which we abbreviate as
b
δ2I(u; η) =
a
Fuu(t, u, u)ηη + 2Fu u(t, u, u)η η + F u u(t, u, u) η η dt.
(i)
(ii)
bd(p, q). Then
The following results establish classical necessary and sufficient conditions to
decide if a function u is a minimizer for the given functional (4.1).
δI(u; η) = 0 for all η ∈ ρGC∞
0 ;
δ2I(u; η) ≥ 0 for all η ∈ ρGC∞
0 .
Theorem 43. Let a, b ∈ ρ(cid:101)R with a < b, let F ∈ ρGC∞([a, b] × ρ(cid:101)Rd × ρ(cid:101)Rd, ρ(cid:101)R), let
p, q ∈ ρ(cid:101)Rd and let u be a local minimizer of I in ρGC∞
0 , the map s ∈ ρ(cid:101)R (cid:55)→
Proof. Let r ∈ ρ(cid:101)R>0 be such that (5.1) holds. Since η ∈ ρGC∞
the sharp topology in its codomain. Therefore, we can find ¯r ∈ ρ(cid:101)R>0 such that
bd(p, q) is well defined and continuous with respect to the trace of
This shows that the GSF s ∈ B¯r(0) (cid:55)→ I(u + sη) ∈ ρ(cid:101)R has a local minimum at
bd(p, q) for all s ∈ B¯r(0). We hence have I(u + sη) ≥ I(u).
s = 0. Now, we employ Lem. 31 and Lem. 33 and thus the claims are proven. (cid:3)
Theorem 44. Let a, b ∈ ρ(cid:101)R with a < b and p, q ∈ ρ(cid:101)Rd. Let u ∈ ρGC∞
u + sη ∈ ρGC∞
u + sη ∈ Bm
r (u) ∩ ρGC∞
bd(p, q) be
such that
(i)
(ii)
δI(u; η) = 0 for all η ∈ ρGC∞
0 .
δ2I(v; η) ≥ 0 for all η ∈ ρGC∞
r ∈ ρ(cid:101)R>0 and m ∈ N.
0 and for all v ∈ Bm
r (u) ∩ GC∞
bd (p, q), where
18
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
2r(u) ∩ GC∞
Moreover, if δ2I(v; η) > 0 for all η ∈ ρGC∞
Then u is a local minimizer of the functional I in ρGC∞
bd (p, q), then I(v) > I(u) for all v ∈ Bm
v ∈ Bm
(cid:107)v − u(cid:107)m > 0.
bd (p, q), we set ψ(s) := I(u + s(v − u)) ∈ ρ(cid:101)R for
r (u) ∩ GC∞
Proof. For any v ∈ Bm
all s ∈ B1(0) so that u + s(v − u) ∈ Bm
r (u). Since (v − u)(a) = 0 = (v − u)(b),
we have v − u ∈ ρGC∞
0 , and properties (i), (ii) yield ψ(cid:48)(0) = δI(u; v − u) = 0 and
ψ(cid:48)(cid:48)(s) = δ2I(u + s(v − u); v − u) ≥ 0 for all s ∈ B1(0). We claim that s = 0 is a
minimum of ψ. In fact, for all s ∈ B1(0) by Taylor's theorem 11
such that (cid:107)η(cid:107)m > 0 and for all
bd (p, q) such that
bd(p, q).
r (u) ∩ GC∞
0
ψ(s) = ψ(0) + sψ(cid:48)(0) +
ψ(cid:48)(cid:48)(ξ)
s2
2
2 ψ(cid:48)(cid:48)(ξ) ≥ 0. Finally,
for some ξ ∈ [0, s]. But ψ(cid:48)(0) = 0 and hence ψ(s) − ψ(0) = s2
Lem. 27 yields
lim
2r(u) ∩ GC∞
r (u) ∩ GC∞
bd (p, q), and take v ∈ Bm
s→1− ψ(s) = I(v) ≥ ψ(0) = I(u),
bd (p, q), then ψ(cid:48)(cid:48)(ξ) = 0 and hence I(v) = I(u).
Now, assume that δ2I(v; η) > 0 for all η ∈ ρGC∞
r (u) ∩ GC∞
which is our conclusion. Note explicitly that if δ2I(v; η) = 0 for all η ∈ ρGC∞
0 and
for all v ∈ Bm
0 such that (cid:107)η(cid:107)m > 0 and for all
As above set ψ(s) := I(u + s(v − u)) ∈ ρ(cid:101)R for all s ∈ B3/2(0) so that u + s(v − u) ∈
bd (p, q) such that (cid:107)v − u(cid:107)m > 0.
v ∈ Bm
2r(u). We have ψ(cid:48)(0) = 0 and ψ(cid:48)(cid:48)(s) = δ2I(u + s(v − u); v − u) > 0 for all s ∈
Bm
B3/2(0) because (cid:107)v−u(cid:107)m > 0. Using Taylor's theorem, we get ψ(1) = ψ(0)+ 1
2 ψ(cid:48)(cid:48)(ξ)
for some ξ ∈ [0, 1]. Therefore ψ(1) − ψ(0) = I(v) − I(u) = 1
(cid:3)
Lemma 45. Let (ak)k∈N, (bk)k∈N and (ck)k∈N be sequences in ρ(cid:101)R>0. Assume that
f ∈ ρGC∞([a1, b1], ρ(cid:101)R). Finally, let ak < t < bk for all k ∈ N, then it holds that
→ 1 in the sharp topology as k → +∞. Let
both (ak)k, (bk)k → 0 and
2 ψ(cid:48)(cid:48)(ξ) > 0.
ck
ak+bk
Proof. We can apply the integral mean value theorem for each ε and each defining
net (fε) of f to get the existence of τk ∈ [t − ak, t + bk] such that
Now, we take the limit for k → ∞, and the claim follows by assumption and by
(cid:3)
Thm. 8.(iv), i.e. by sharp continuity of f .
We now derive the so-called necessary Legendre condition:
Theorem 46. Let a, b ∈ ρ(cid:101)R with a < b and let u ∈ ρGC∞([a, b], ρ(cid:101)Rd) be a minimizer
of the functional I. Then
is positive semi definite for all t ∈ [a, b], i.e.
F u u(t, u(t), u(t))
∀λ = (λ1, . . . , λd) ∈ ρ(cid:101)Rd.
Proof. Let λ = [λε] ∈ ρ(cid:101)Rd and k, h ∈ N be arbitrary. Let t = [tε] ∈ [a, b]. We
F ui uj (t, u(t), u(t))λiλj ≥ 0,
(5.2)
can assume that t is a sharply interior point, because otherwise we can use sharp
continuity of the left hand side of (5.2) and Lem. 27. We can also assume that λ
is componentwise invertible because of Lem. 30. We want to mimic the classical
f (t) = lim
k→∞
1
ck
t+bk
f (s) ds.
t−ak
f (τk) =
=
1
bk + ak
ck
bk + ak
t+bk
t−ak
1
ck
t+bk
t−ak
f (s) ds
f (s) ds.
CALCULUS OF VARIATIONS FOR GF
19
proof of [21, Thm. 1.3.2], but considering a "regularized" version of the triangular
function used there (see Fig. 5.1). In particular: (1) the smoothed triangle must
have an infinitesimal height which is proportional to λ, and we will take ρk
ε as
this infinitesimal; (2) in the proof we need that the derivative at t is equal to λ,
and this justifies the drawing of the peak in Fig. 5.1; (3) to regularize the singular
points of the triangular function, we need a smaller infinitesimal, and we can take
e.g. ρ2k
ε . So, consider a net of smooth functions ϑε on [aε, bε], such that the following
properties hold:
ϑε(x) = 0, for x ≤ tε − ρk
(i)
ϑε(x) = 0, for x ≥ tε + ρk
(ii)
(iii) ϑε(x) = λ(x − tε) + ρk
(iv) ϑε(x) = −λ(x − tε) + ρk
ϑε(x) ≤ ρk
ε · λ + 2ρ2k
(v)
ϑε(x) ≤ 2λ for all x
(vi)
ε − ρ2k
ε .
ε + ρ2k
ε .
ε λ for x ∈ [tε − ρk
ε λ for x ∈ [tε + ρ2k
ε λ.
ε , tε].
ε , tε + ρk
ε − ρ2k
ε ].
ε + ρ2k
2ρ2k
ε
λερk
ε
2ρ2k
ε
2ρ2k
ε
a
tε − ρk
ε
tε
tε + ρk
ε
b
Figure 5.1. This figure illustrates the function ϑε we are consid-
ering (blue). The dotted green triangle symbolizes the function
which is used in the classical proofs of the Legendre necessary con-
dition (cf. [21, Thm. 1.3.2]).
The net (ϑε) defines a GSF ϑ := [ϑε(−)] ∈ ρGC∞
point. Setting for simplicity ak := dρk + dρ2k, by assumption we have
0 ≤ δ2I(u, ϑ) =
(5.3)
Now, setting M := max[a,b] Fuu(t, u, u) and N := max[a,b] Fu u(t, u, u), by (v) we
Fuu(t, u, u)ϑϑ + 2F uu(t, u, u) ϑϑ + F u u(t, u, u) ϑ ϑ dt.
0 because t is a sharply interior
t−ak
t+ak
t+ak
Fuu(t, u, u)ϑϑ dt
where we used the evident notation Gk = O(cid:0)dρk(cid:1) to denote that there exists some
A ∈ ρ(cid:101)R>0 such that Gk ≤ A · dρk for all k ∈ N. Using (v) and (vi), we analogously
t−ak
(cid:12)(cid:12)(cid:12)(cid:12) ≤ M · ϑ(t)2 · 2ak = O(cid:0)dρ3k(cid:1) ,
(cid:12)(cid:12)(cid:12)(cid:12) ≤ 4N · ϑ(t) · ak · λ = O(cid:0)dρ2k(cid:1) .
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
have
t+ak
t−ak
F uu(t, u, u) ϑϑ dt
20
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Note that there always exists C ∈ ρ(cid:101)R such that λ ≤ Cdρk. Therefore
lim
k→+∞
1
t+ak
2dρk
t−ak
Fuu(t, u, u)ϑϑ + 2F uu(t, u, u) ϑϑ dt = 0.
(5.4)
Using Lemma 45, (5.4), (5.3) and Lem. 27, we obtain that
F u u(t, u(t) u(t)) ϑ(t) ϑ(t) = lim
k→+∞
F u u(t, u, u) ϑ ϑ dt ≥ 0.
But (iii) yields ϑ(t) = λ, and this concludes the proof.
(cid:3)
1
t+ak
2dρk
t−ak
As in the classical case, Thm. 43.(ii) motivates to define the accessory integral
6. Jacobi fields
b
ψ(t, η, η) dt ∀η ∈ ρGC∞
0 ,
Q(η) :=
a
where
(6.1)
(6.2)
for all t ∈ [a, b] and (l, v) ∈ ρ(cid:101)Rd × ρ(cid:101)Rd. Note that if u minimizes I, then
ψ(t, l, v) := Fuu(t, u, u)ll + 2Fu u(t, u, u)lv + F u u(t, u, u)vv
Q(η) ≥ 0 ∀η ∈ ρGC∞
0 .
As usual, we note that η = 0 is a minimizer of the functional Q and we are
interested whether there are others. In order to solve this problem, we consider the
Euler-Lagrange equations for Q, which are given by
d
dt
ψ η(t, η, η) = ψη(t, η, η),
(6.3)
in other words
d
dt
{F u u(t, u, u) η + Fu u(t, u, u)η} = Fu u(t, u, u) η + Fuu(t, u, u)η.
Since u is given, (6.4) is an ρ(cid:101)R-linear system of second order equations in the
unknown GSF η and with time dependent coefficients in ρ(cid:101)R. We call (6.4) the
(6.4)
Jacobi equations of I with respect to u. As in the classical setting, we define
Definition 47. A solution η ∈ ρGC∞
field along u0 = u.
0 of the Jacobi equations (6.4) is called a Jacobi
The following result confirms that the intuitive interpretation of a Jacobi field as
the tangent space of a smooth family of solutions of the Euler-Lagrange equation
still holds in this generalized setting.
Lemma 48. Let u ∈ ρGC∞([−δ, δ] × [a, b], ρ(cid:101)Rd), where δ ∈ ρ(cid:101)R>0. We write
us := u(s,−) for all s ∈ [−δ, δ]. Assume that each us satisfies the Euler-Lagrange
equations (4.5):
d
dt
Then
F u(t, us, us) = Fu(t, us, us) ∀s ∈ [−δ, δ].
η(t) :=
d
ds
is a Jacobi field along u.
us(t) ∀t ∈ [a, b]
(cid:12)(cid:12)(cid:12)(cid:12)0
CALCULUS OF VARIATIONS FOR GF
Proof. A straight forward calculation gives:
F u(t, us, us) − Fu(t, us, us)
(cid:19)
21
(cid:3)
(cid:12)(cid:12)(cid:12)(cid:12)0
(cid:18) d
dt
0 =
=
d
ds
d
dt
(F u u(t, u, u) η + Fu u(t, u, u)η) − Fu u(t, u, u) η − Fuu(t, u, u)η.
6.1. Conjugate points and Jacobi's theorem. The classical key result con-
cerning Jacobi fields relates conjugate points and minimizers. The main aim of the
present section is to derive this theorem in our generalized framework by extending
the ideas of the proof of [21, Thm. 1.3.4].
A crucial notion is hence that of piecewise GSF:
Definition 49. We call piecewise GSF an n-tuple (f1, . . . , fn) such that
(i)
For all i = 1, . . . , n there exist ai, ai+1 ∈ ρ(cid:101)R such that ai < ai+1 and fi ∈
ρGC∞([ai, ai+1], ρ(cid:101)Rd). Note that [a, b] = [a(cid:48), b(cid:48)] implies a = a(cid:48) and b = b(cid:48)
because the relation ≤ is antisymmetric. Therefore, the points ai, ai+1 are
uniquely determined by the set-theoretical function fi.
For all i = 1, . . . , n, we have fi(ai+1) = fi+1(ai+1).
(ii)
Every pointwise GSF (f1, . . . , fn) defines a set-theoretical function:
(iii) For all t ∈(cid:83)n
We also use the arrow notation (f1, . . . , fn) : (cid:83)n
i=1[ai, ai+1], we set (f1, . . . , fn)(t) := fi(t) if t ∈ [ai, ai+1].
i=1[ai, ai+1] −→ ρ(cid:101)Rd to say that
both (i) and (ii) hold.
Remark 50.
(i)
(v)
(ii)
(iii)
Clearly, t ∈ [ai, ai+1] ∩ [ai+1, ai+2] implies t = ai+1, so that condition (ii)
yields that the evaluation (iii) is well defined.
Since the order relation ≤ is not a total one, we do not have that [ai, ai+1] ∪
[ai+1, ai+2] = [ai, ai+2].
piecewise GSF (f1, f2), then neither the GSF fi nor the points ai are uniquely
determined by ν. For this reason, we prefer to stress our notations with
piecewise GSF.
(g1, . . . , gn) + (f1, . . . , fn) := (g1 + f1, . . . , gn + fn) and r · (f1, . . . , fn) :=
If ν : [a1, a2]∪ [a2, a3] −→ ρ(cid:101)Rd is a set-theoretical function originating from a
symbols like (f1, f2)(t) ∈ ρ(cid:101)Rd.
(iv) Every GSF f ∈ ρGC∞([a1, a2], ρ(cid:101)Rd) can be seen as a particular case of a
If (g1, . . . , gn), (f1, . . . , fn) : (cid:83)n
i=1[ai, ai+1] −→ ρ(cid:101)Rd and r ∈ ρ(cid:101)R, then also
(r · f1, . . . , r · fn) are piecewise GSF, and we hence have a structure of ρ(cid:101)R-
If (f1, . . . , fn) :(cid:83)n
i=1[ai, ai+1] −→ ρ(cid:101)Rd and F ∈ ρGC∞(ρ(cid:101)Rd, ρ(cid:101)Rn), then we can
define the composition F◦(f1, . . . , fn) := (F◦f1, . . . , F◦fn) :(cid:83)n
ρ(cid:101)Rn.
i=1[ai, ai+1] −→
Definition 51. Let x = [xε] ∈ ρ(cid:101)R, then
ν(x) := sup(cid:8)b ∈ R xε = O(ρb
(iii) dρ(x) := dρ− log xe ∈ ρ(cid:101)R>0.
It is worth noting that −e : ρ(cid:101)R −→ R≥0 induces an ultrametric on ρ(cid:101)R that gener-
Piecewise GSF inherit from their defining components a well-behaved differential
and integral calculus. The former is even more general and taken from [1].
ε)(cid:9) ∈ R ∪ {+∞}.
xe := e−ν(x) ∈ R≥0.
module.
(i)
(ii)
(vi)
ates exactly the sharp topology, see e.g. [2, 11] and references therein. However, we
22
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
will not use this ultrametric structure in the present paper, and we only introduced
it to get an invertible infinitesimal dρ(x) that goes to zero with x: it is in fact easy
to show that
x
lim
x→0
dρ(x)
= 1
in the sharp topology.
Definition 52. Let T ⊆ ρ(cid:101)R and let f : T −→ ρ(cid:101)Rd be an arbitrary set-theoretical
function. Let t0 ∈ T be a sharply interior point of T . Then we say that f is
differentiable at t0 if2
∃m ∈ ρ(cid:101)Rd :
f (t + h) − f (t0) − m · h
lim
h→0
dρ(h)
= 0.
(6.5)
In this case, using Landau little-oh notation, we can hence write
as h → 0.
As in the classical case, (6.6) implies the uniqueness of m ∈ ρ(cid:101)Rd, so that we can
f (t + h) = f (t0) + m · h + o(dρ(h))
define f(cid:48)(t0) := f (t0) := m, and the usual elementary rules of differential calculus.
By the Fermat-Reyes theorem, this definition of derivative generalizes that given
for GSF.
In particular, this notion of derivative applies to the set-theoretical function
induced by a piecewise GSF (f1, . . . , fn). We therefore have that (f1, . . . , fn)(−) is
differentiable at each ai < t < ai+1, and (f1, . . . , fn)(cid:48)(t) = f(cid:48)
i (t), but clearly there
is no guarantee that (f1, . . . , fn)(−) is also differentiable at each point ai.
The notion of definite integral is naturally introduced in the following
(6.6)
i=1[ai, ai+1] −→ ρ(cid:101)Rd be a piecewise GSF, then
Definition 53. Let (f1, . . . , fn) :(cid:83)n
an+1
(f1, . . . , fn)(t) dt :=
ai+1
fi(t) dt.
n(cid:88)
a1
i=1
ai
Since our main aim in using piecewise GSF is to prove Jacobi's theorem, we do
not need to prove that the usual elementary rules of integration hold, since we will
always reduce to integrals of GSF.
Having a notion of derivative and of definite integral, also for piecewise GSF we
can consider functionals
ν := (f1, . . . , fn), a1 = a, an = b =⇒ I(ν) :=
F (t, ν(t), ν(t))dt ∈ ρ(cid:101)R,
b
(6.7)
and the concept of piecewise GSF (global) minimizer : I(ν) ≤ I(ν) for all ν ∈ ρGC∞
0 .
For the proof of Jacobi's theorem, we will only need this particular notion of global
minimizer. Note explicitly that in (6.7) we only need existence of right and left
derivatives of GSF, because of Def. 53 and of Def. 14 of definite integral of GSF.
a
Classically, several proofs of Jacobi's theorem use both some form of implicit
function theorem and of uniqueness of solution for linear ODE.
Theorem 54 (Implicit function theorem). Let U ⊆ ρ(cid:101)Rn, V ⊆ ρ(cid:101)Rd be sharply open
sets. Let F ∈ ρGC∞(U × V, ρ(cid:101)Rd) and (x0, y0) ∈ U × V . If ∂2F (x0, y0) is invertible
in L(ρ(cid:101)Rd, ρ(cid:101)Rd), then there exists a sharply open neighbourhood U1 × V1 ⊆ U × V of
(x0, y0) such that
∀x ∈ U1 ∃!yx ∈ V1 : F (x, yx) = F (x0, y0).
(6.8)
2This definition is based on [1, Def. 2.2].
CALCULUS OF VARIATIONS FOR GF
23
Moreover, the function f (x) := yx for all x ∈ U1 is a GSF f ∈ ρGC∞(U1, V1) and
satisfies
Df (x) = − (∂2F (x, f (x)))
−1 ◦ ∂1F (x, f (x)).
(6.9)
Proof. The usual deduction of the implicit function theorem from the inverse func-
tion theorem in Banach spaces can be easily adapted using Thm. 12 and noting
(cid:3)
that det [∂2F (−,−)] is a GSF such that det [∂2F (x0, y0)] ∈ ρ(cid:101)R>0.
In the next theorem, the dependence of the entire theory on the initial infinites-
imal net ρ = (ρε) ↓ 0 plays an essential role. Indirectly, the same important role
will reverberate in the final Jacobi's theorem.
Theorem 55 (Solution of first order linear ODE). Let A ∈ ρGC∞([a, b], ρ(cid:101)Rd×d),
where a, b ∈ ρ(cid:101)R, a < b, and t0 ∈ [a, b], y0 ∈ ρ(cid:101)Rd. Assume that
(cid:12)(cid:12)(cid:12)(cid:12) ≤ −C · log dρ ∀t ∈ [a, b],
where C ∈ R>0. Then there exists one and only one y ∈ ρGC∞([a, b], ρ(cid:101)Rd) such that
(cid:12)(cid:12)(cid:12)(cid:12)
A(s) ds
(6.10)
t0
t
(cid:40)
Moreover, this y is given by y(t) = exp
A(s) ds
y(cid:48)(t) = A(t) · y(t)
y(t0) = y0
if t ∈ [a, b]
(cid:16)´
(cid:17)
t
t0
(cid:104)
(cid:17) · y0 for all t ∈ [a, b].
(cid:16)´
(cid:17)(cid:105)
(cid:16)´
t
t0
(6.11)
exp
tε
t0ε
t
t0
(cid:19)
A(s) ds
=
exp
A(s) ds
Aε(s) ds
(cid:16)´
(cid:18)
, where t = [tε],
Proof. We first note that exp
≤ e−C log dρ ≤ dρ−C.
because for all t ∈ [a, b], we have
t0 = [t0ε] and A(s) = [Aε(sε)] ∈ ρ(cid:101)Rd×d. This exponential matrix in ρ(cid:101)Rd×d is a GSF
(cid:17) · y0 are ρ-moderate. Analogously,
(cid:16)−
Therefore, all values of y(t) = exp
one can prove that also y(k)(t) are moderate for all k ∈ N and t ∈ [a, b]. Considering
that derivatives can be calculated ε-wise, we have that this GSF y satisfies (6.11),
and this proves the existence part.
To show uniqueness, we can proceed as in the smooth case. Assume that z ∈
for all t ∈ [a, b].
Since h(cid:48) = −A · h, we have (hz)(cid:48) = h(cid:48)z + hz(cid:48) = −Ahz + hAz = −Ahz + Ahz = 0.
From uniqueness of primitives of GSF, Thm. 13, we have that h· z = h(t0)· z(t0) =
y0. Therefore z = h−1 · y0.
(cid:3)
ρGC∞([a, b], ρ(cid:101)Rd) satisfies (6.11), and set h(t) := exp
If α, β ∈ ρ(cid:101)R, we write α = OR(β) to denote that there exists C ∈ R>0 such
A(s) ds
A(s) ds
(cid:17)
´
t
t0
t
t0
that α ≤ C · β. Therefore, assumption (6.10) can be written as
OR(log dρ). Note that this assumption is weaker, in general, than
A(s) ds =
´
t
t0
(b − a) · max
t∈[a,b]
A(t) = OR(log dρ).
theorem.
The following result is the key regularity property we need to prove Jacobi's
Lemma 56. Let a, a(cid:48), b ∈ ρ(cid:101)R, with a < a(cid:48) < b, and let K ∈ ρGC∞([a, b] × ρ(cid:101)Rd ×
ρ(cid:101)Rd, ρ(cid:101)R). Let ν = (η, β) : [a, a(cid:48)] ∪ [a(cid:48), b] −→ ρ(cid:101)Rd be a piecewise GSF which satisfies
the Euler-Lagrange equation
Ku(t, ν(t), ν(t)) − d
dt
K u(t, ν(t), ν(t)) = 0 ∀t ∈ [a, a(cid:48)) ∪ (a(cid:48), b].
(6.12)
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
24
Finally, assume that det(cid:0)K ui uj (a(cid:48), η(a(cid:48)), η(a(cid:48)))i,j=i,...,d
(cid:1) ∈ ρ(cid:101)R is invertible, then
ν(t) = η(a(cid:48)).
(6.13)
ν(t) = lim
t→a(cid:48)
a(cid:48)<t
In particular, if β ≡ 0[a(cid:48),b], then η(a(cid:48)) = 0.
lim
t→a(cid:48)
t<a(cid:48)
Proof. Set Φ(t, l, v, q) := K u(t, l, v) − q for all t ∈ [a, b] and all l, v, q ∈ ρ(cid:101)Rd. For
simplicity, set (t0, l0, v0, q0) := (a(cid:48), η(a(cid:48)), η(a(cid:48)), K u(a(cid:48), η(a(cid:48)), η(a(cid:48))). Our assumption
on the invertibility of K u u(a(cid:48), η(a(cid:48)), η(a(cid:48))) = ∂vΦ(t0, l0, v0, q0) makes it possible to
apply the implicit function Thm. 54 to conclude that there exists a neighbourhood
T × L × V × Q of (t0, l0, v0, q0) such that
∀(t, l, q) ∈ T × L × Q∃!v ∈ V : Φ(t, l, v, q) = Φ(t0, l0, v0, q0).
(6.14)
But Φ(t0, l0, v0, q0) = K u(a(cid:48), η(a(cid:48)), η(a(cid:48))) − q0 = 0. Moreover, the unique function
ϕ defined by Φ(t, l, ϕ(t, l, q), q) = 0 for all (t, l, q) ∈ T × L × Q is a GSF ϕ ∈
ρGC∞(T × L × Q, V ). Now, for all t ∈ [a, a(cid:48)) ∪ (a(cid:48), b], we have
Φ(t, ν(t), ν(t), K u(t, ν(t), ν(t))) = K u(t, ν(t), ν(t)) − K u(t, ν(t), ν(t)) = 0.
Therefore, uniqueness in (6.14) yields
ν(t) = ϕ (t, ν(t), K u(t, ν(t), ν(t)))
∀t ∈ [a, a(cid:48)) ∪ (a(cid:48), b].
(6.15)
We now integrate the Euler-Lagrange equation (6.12) on [a, t], obtaining
K u(t, ν(t), ν(t)) =
Ku(s, ν(s), ν(s)) ds + K u(a, η(a), η(a)) ∀t ∈ [a, a(cid:48)) ∪ (a(cid:48), b].
t
a
t
(cid:18)
This entails that we can write
ν(t) = ϕ
t, ν(t),
a
Ku(s, ν(s), ν(s)) ds + K u(a, η(a), η(a))
But the function t ∈ [a, a(cid:48)) ∪ (a(cid:48), b] (cid:55)→
on the left and on the right of a(cid:48) because on [a, a(cid:48)) and on (a(cid:48), b] it is a GSF; in fact
for t < a(cid:48) we have
a Ku(s, ν(s), ν(s)) ds ∈ ρ(cid:101)Rd has equal limits
´
t
∀t ∈ [a, a(cid:48)) ∪ (a(cid:48), b].
(6.16)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
a
t
Ku(s, ν(s), ν(s)) ds −
a(cid:48)
a
(cid:19)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
Ku(s, ν(s), ν(s)) ds
≤ max
t∈[a,a(cid:48)]
Ku(s, η(s), η(s)) · t − a(cid:48),
and this goes to 0 as t → a(cid:48), t < a(cid:48). Analogously we can proceed for t > a(cid:48) using
β. Therefore
t
lim
t→a(cid:48)
t<a(cid:48)
Ku(s, ν(s), ν(s)) ds = lim
t→a(cid:48)
t>a(cid:48)
a
a
t
Ku(s, ν(s), ν(s)) ds.
ν(t) = η(a(cid:48)) = limt→a(cid:48)
a(cid:48)<t
Applying this equality in (6.16), we get limt→a(cid:48)
t<a(cid:48)
claimed. Finally, if β ≡ 0[a(cid:48),b], then limt→a(cid:48)
a(cid:48)<t
In the following definition and below, we use the complete notation ρGC∞
(see Def. 34).
Definition 57. Let a, a(cid:48), b ∈ ρ(cid:101)R, where a < a(cid:48) < b. We call a(cid:48) conjugate to a w. r. t.
the variational problem (4.1) if there exists a non identically vanishing Jacobi field
η ∈ ρGC∞
0 (a, a(cid:48)) along u[a,a(cid:48)] such that η(a) = 0 = η(a(cid:48)), where ψ is given by (6.2).
ν(t) as
(cid:3)
0 (a, a(cid:48))
ν(t) = 0.
CALCULUS OF VARIATIONS FOR GF
25
In order to prove the important Jacobi's theorem in the present generalized
context, which shows that we cannot have minimizers if there are interior points
which are conjugate to a, we finally need the following
Lemma 58. Let u ∈ ρGC∞([a, b], ρ(cid:101)Rd) and let a(cid:48) ∈ (a, b). Let η ∈ ρGC∞
0 (a, a(cid:48)) be a
Jacobi field along u[a,a(cid:48)], with η(a) = 0 = η(a(cid:48)). Then
a(cid:48)
ψ(t, η, η) dt = 0.
Proof. Since ψ is ρ(cid:101)R-homogeneous of second order in (η, η),we have
2ψ(t, η, η) = ψη(t, η, η)η + ψ η(t, η, η) η.
Thus we calculate:
a
a(cid:48)
2
ψ(t, η, η) dt =
a
ηψη(t, η, η) + ηψ η(t, η, η) dt
(cid:19)
a(cid:48)
(cid:18)
a
a(cid:48)
η
=
a
ψη(t, η, η) − d
dt
ψ η(t, η, η)
dt by integration by parts
= 0 since η is a Jacobi field.
(cid:3)
After these preparations we can finally prove
Theorem 59 (Jacobi). Let a, b ∈ ρ(cid:101)R be such that a < b. Suppose that F ∈
ρGC∞([a, b] × ρ(cid:101)Rd × ρ(cid:101)Rd, ρ(cid:101)R) and u ∈ ρGC∞([a, b], ρ(cid:101)R) are such that
a(cid:48) ∈ (a, b) is conjugate to a
det F u u(t, u(t), u(t)) ∈ ρ(cid:101)R is invertible for all t ∈ [a, b].
(i)
(ii)
(iii) For all t ∈ [a, a(cid:48)]
Fu u(s, u(s), u(s)) − Fuu(s, u(s), u(s))
(cid:20) d
ds
t
a(cid:48)
u u (s, u(s), u(s)) ·
−F −1
t
(cid:21)
ds =
= OR(log dρ)
b
a
a(cid:48)
a
b
a(cid:48)
a(cid:48)
−F −1
u u (s, u(s), u(s)) · d
ds
F u u(s, u(s), u(s)) ds = OR(log dρ).
Then u cannot be a local minimizer of I. Therefore, for any r ∈ ρ(cid:101)R>0 there exists
bd(u(a), u(b)) and m ∈ N such that (cid:107)v − u(cid:107)m < r but I(u) (cid:54)≤ I(v).
v ∈ ρGC∞
Proof. By contradiction, assume that u is a local minimizer, and let η ∈ ρGC∞
0 (a, a(cid:48))
be a Jacobi field along u[a,a(cid:48)] such that the conditions from Def. 57 hold for η. We
want to prove that η ≡ 0. Define ν := (η, 0[a(cid:48),b]), which is a piecewise GSF since
η(a(cid:48)) = 0. Since also η(a) = 0, Lem. 58 and homogeneity of ψ yield
Q(ν) =
ψ(t, ν(t), ν(t)) dt =
ψ(t, η(t), η(t)) dt +
ψ(t, 0, 0) dt = 0.
Therefore, Thm. 43 (necessary condition for u being a minimizer) gives Q(ν) ≥
0 = Q(ν) for all ν ∈ ρGC∞
0 (a, b). Thus, ν is a minimizer of the functional Q. Since
ν is only a piecewise GSF, we cannot directly apply Thm. 38 (Euler-Lagrange
equations). But for all ϕ ∈ ρGC∞
0 (a, b) and all s ∈ ρ(cid:101)R, we have
Q(ν + sϕ) =
b
ψ(t, ν + sϕ, ν + s ϕ) dt
a(cid:48)
a
=
ψ(t, η + sϕ, η + s ϕ) dt +
a
b
a(cid:48)
ψ(t, sϕ, s ϕ) dt.
(6.17)
(cid:12)(cid:12)(cid:12)(cid:12)0
26
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
This shows that s ∈ ρ(cid:101)R (cid:55)→ Q(ν + sϕ) ∈ ρ(cid:101)R is a GSF, and hence s = 0 is a minimum
for this function. By Lem. 31 and (6.17), we get
δQ(ν, ϕ) = 0 =
a(cid:48)
a(cid:48)
a
a
=
=
(cid:18)
(cid:18)
Q(ν + sϕ)
d
ds
ψη(t, η, η) − d
dt
ψη(t, η, η) − d
dt
(cid:19)
(cid:19)
b
a(cid:48)
ϕ dt +
ψ η(t, η, η)
ψ η(t, η, η)
ϕ dt.
(ϕψη(t, 0, 0) + ϕψ η(t, 0, 0)) dt
By the fundamental Lem. 37, this implies that η satisfies the Euler-Lagrange equa-
tions for ψ in the interval [a, a(cid:48)). Therefore, ν satisfies the same equations in
[a, a(cid:48)) ∪ (a(cid:48), b]. Moreover, ψ η η(a(cid:48), η(a(cid:48)), η(a(cid:48))) = F u u(a(cid:48), u(a(cid:48)), u(a(cid:48))) is invertible by
assumption (ii). Thus, all the hypotheses of the regularity Lem. 56 hold, and we
derive that η(a(cid:48)) = 0.
(cid:21)
Fu u(t, u, u) − Fuu(t, u, u)
, and
(cid:20) d
For all t ∈ [a, b], we define
ξ(t) := −F −1
u u ·
dt
u u · d
ϑ(t) := −F −1
dt
(cid:32)
(cid:33)
´
order ODE
η
z
y(a(cid:48)) =
=
η(a(cid:48))
η(a(cid:48))
(cid:33)
(cid:32)
(cid:32)
y :=
F u u(t, u, u),
(cid:33)
(cid:32)
(cid:33)
0
1
ξ ϑ
·
η
z
= 0
so that we can re-write the Jacobi equations (6.4) for η on [a, a(cid:48)] as a system of first
=: A · y
∀t ∈ [a, a(cid:48)]
By assumptions (iii), we obtain
hence apply Thm. 55 obtaining y ≡ 0 and thus η ≡ 0.
a(cid:48) A(t) = OR(log dρ) for all t ∈ [a, a(cid:48)], and we can
(cid:3)
t
Note that if one of the quantities in ((iii)) depends even only polynomially on ε,
then we are forced to take e.g. ρε = ε1/ε to fulfill this assumption. This underlines
the importance of the parameter ρ making the entire theory dependent on the pa-
rameter ρ., in order to avoid unnecessary constraints on the scope of the functionals
we look upon.
7. Noether's theorem
We first note that any X ∈ ρGC∞ (J × X, Y ), where J ⊆ ρ(cid:101)R, can also be considered
In this section, we state and prove Noether's theorem following the lines of [3].
as a family in GSF which smoothly depends on the parameter s ∈ J. In this case,
we hence say that (Xs)s∈J is a generalized smooth family in ρGC∞(X, Y ).
In
particular, we can reformulate in the language of GSF the classical definition of
one-parameter group of generalized diffeomorphisms of X as follows:
(i)
(ii)
(iii) X0(x) = x for all x ∈ X,
(Xs)s∈ρ(cid:101)R is a generalized smooth family in ρGC∞(X, X),
For all s ∈ ρ(cid:101)R, the map Xs : X −→ X is invertible, and X−1
s ∈ ρGC∞(X, X),
(iv) Xs ◦ Xt = Xs+t for all s, t ∈ ρ(cid:101)R.
In our proofs, we will in fact only use properties (i) and (iii).
The proof of Noether's theorem is classically anticipated by the following time-
independent version, which the general case is subsequently reduced to.
CALCULUS OF VARIATIONS FOR GF
Theorem 60. Let K ∈ ρGC∞(cid:16)
L × V, ρ(cid:101)R(cid:17)
, where L, V ⊆ ρ(cid:101)Rn are sharply open
27
sets. Let w ∈ ρGC∞((a, b), L) be a solution of the Euler-Lagrange equation corre-
sponding to K, i.e. for all t ∈ (a, b)
w(t) ∈ V , Ku(w(t), w(t)) =
Suppose that 0 is a sharply interior point of J ⊆ ρ(cid:101)R and (Xs)s∈J is a generalized
K u(w(t), w(t)).
d
dt
(7.1)
smooth family in ρGC∞(L, L), such that for all t ∈ (a, b)
∂t Xs(w(t)) ∈ V ,
(i)
(ii) X0(w(t)) = w(t),
(iii) K is invariant under (Xs)s∈J along w, i.e.
∂
(cid:18)
K(w(t), w(t)) = K
Then, the quantity
Xs(w(t)),
∂
∂t
Xs(w(t))
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂
∂s
s (w(t)) ∈ ρ(cid:101)R
X j
K uj (w(t), w(t))
(cid:19)
∀s ∈ J.
(7.2)
(cid:12)(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂
∂s
(cid:12)(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂
∂s
∂
∂s
is constant in t ∈ (a, b).
be arbitrary but fixed. Since s = 0 ∈ J is a sharply interior point, we can consider
Proof. We first note that both sides of (7.2) are in ρGC∞((a, b), ρ(cid:101)R). Let τ ∈ (a, b)
(cid:12)(cid:12)s=0. We obtain
(cid:12)(cid:12)(cid:12)(cid:12)s=0
Xs(w),
Xs(w)
(cid:18)
(cid:19)
(7.2)
=
∂
∂t
d
ds
K
0
∂
∂s
τ
(ii)
=
Ku(w, w)
a
∂
∂s
Xs(w) + K u(w, w)
Xs(w) dt.
∂
∂t
∂
∂s
Since the Euler-Lagrange equations (7.1) for K are given by Ku(w, w) = d
we have
dt K u(w, w),
0 =
=
d
dt
d
dt
(K u(w, w))
(cid:18)
K u(w, w)
Xs(w) + K u(w, w)
(cid:19)
∂
∂t
Xs(w)
.
Xs(w)
(cid:3)
Which is our conclusion by the uniqueness - part of Thm. 13.
Theorem 61 (E. Noether).
We are now able to prove Noether's theorem. For the convenience of the reader,
in its statement and proof we use the variables t, T , l, L, v, V so as to recall tempus,
locus, velocitas resp.
Let a, b ∈ ρ(cid:101)Rd, with a < b, and F ∈ ρGC∞(cid:16)
[a, b] × ρ(cid:101)Rd × ρ(cid:101)Rd, ρ(cid:101)R(cid:17)
ρGC∞([a, b], ρ(cid:101)Rd) be a solution of the Euler-Lagrange equation (4.5) corresponding to
F . Suppose that 0 is a sharply interior point of J ⊆ ρ(cid:101)R and (Xs)s∈J is a generalized
smooth family in ρGC∞((a, b)×ρ(cid:101)Rd, (a, b)×ρ(cid:101)Rd). We denote by Ts(t, l) := X 1
s (t, l) ∈ ρ(cid:101)Rd for all (t, l) ∈ (a, b) × ρ(cid:101)Rd, the two projections
s (t, l) ∈
of Xs on (a, b) and ρ(cid:101)Rd resp. We assume that for all t ∈ (a, b)
(cid:104)
∂t Ts(t, u(t)) ∈ ρ(cid:101)R is invertible,
T0(t, u(t)) = t and L0(t, u(t)) = u(t),
Ts(t, u), Ls(t, u),
∂t Ts(t, u) for all s ∈ J.
(a, b) and Ls(t, l) := X 2
(iii) F (t, u(t), u(t)) = F
(cid:105) · ∂
. Let u ∈
(i)
(ii)
∂
∂
∂t Ls(t,u)
∂
∂t Ts(t,u)
28
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
Then, the quantity
F uj (t, u(t), u(t))
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂
∂s
Lj
s(t, u(t))+
+(cid:2)F (t, u(t), u(t)) − F uk (t, u(t), u(t)) uk(t)(cid:3) ∂
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂s
Ts(t, u(t))
(7.3)
is constant in t ∈ [a, b].
Proof. Since (7.3) is a GSF in t ∈ [a, b], by sharp continuity it suffices to prove
the claim for all t ∈ (a, b). Set L := (a, b) × ρ(cid:101)Rd, V := ρ(cid:101)R∗ × ρ(cid:101)Rd (we recall
that ρ(cid:101)R∗ denotes the set of all invertible generalized numbers in ρ(cid:101)R). Define K ∈
ρGC∞(L × V, ρ(cid:101)R) by
· p ∀(t, l) ∈ L∀(p, v) ∈ V,
K(t, l; p, v) := F
(7.4)
and w ∈ ρGC∞((a, b), L) by w(t) := (t, u(t)) for all t ∈ (a, b). We note that L,
V ⊆ ρ(cid:101)Rd+1 are sharply open subsets and that w(t) = (1, u(t)) ∈ V . The notations
t, l,
for partial derivatives used in the present work result from the symbolic writing
K(u1, . . . , ud+1; u1, . . . , ud+1), so that the variables used in (7.4) yield
(cid:18)
(cid:19)
v
p
(cid:16)
Kt(t, l; p, v) = Ft
Kp(t, l; p, v) = F
(cid:16)
Kvj (t, l; p, v) = F uj−1
Klj (t, l; p, v) = Fuj−1
t, l, v
p
t, l, v
p
(cid:17) · p
(cid:17) · p
(cid:16)
(cid:16)
(cid:17) − F uk
(cid:16)
(cid:17)
(cid:20) d
t, l, v
p
t, l, v
p
(cid:17) vk
p
t, l, v
p
if j = 1
if j = 2, . . . , d + 1,
(7.5)
if j = 1
if j = 2, . . . , d + 1.
(7.6)
(cid:21)
· uk
Kuj (t, l; p, v) =
and
K uj (t, l; p, v) =
From these, for all t ∈ (a, b) and all j = 2, . . . , d + 1, it follows that
F uk (t, u, u) − Fuk (t, u, u)
Ku1 (w, w) − d
dt
Kuj (w, w) − d
dt
K u1(w, w) =
K uj (w, w) = Fuj−1(t, u, u) − d
dt
dt
F uj−1 (t, u, u).
Therefore, since u satisfies the Euler-Lagrange equations for F , this entails that w
is a solution of the analogous equations for K in (a, b). Now, (i) gives
(cid:18) ∂
(cid:19)
∂
∂t
Xs(w(t)) =
Ts(t, u(t)),
∂
∂t
Ls(t, u(t))
∂t
(cid:18)
K(w, w) = F (t, u, u)
(cid:34)
K
Xs(w),
Xs(w)
= F
Ts(t, u), Ls(t, u),
(cid:19)
∈ ρ(cid:101)R∗ × ρ(cid:101)Rd = V.
(cid:35)
∂
∂t Ls(t, u)
∂
∂t Ts(t, u)
· ∂
∂t
Ts(t, u).
Moreover, (ii) gives X0(w(t)) = (T0(t, u(t)), L0(t, u(t))) = w(t). Finally
We can hence apply Thm. 60, and from (7.5), (7.6) we get that
K uj (w, w)
∂
∂s
X j
s (w) = F uj (t, u(t), u(t))
+(cid:2)F (t, u(t), u(t)) − F uk (t, u(t), u(t)) uk(t)(cid:3) ∂
s(t, u(t))+
Lj
(cid:12)(cid:12)(cid:12)(cid:12)s=0
∂
∂s
is constant in t ∈ (a, b).
(cid:12)(cid:12)(cid:12)(cid:12)s=0
Ts(t, u(t))
(cid:3)
∂s
∂
∂t
(cid:12)(cid:12)(cid:12)(cid:12)0
CALCULUS OF VARIATIONS FOR GF
29
8. Application to C1,1 Riemannian metric
In this section, we fix any embedding (ιb
In the following, we apply what we did so far to the problem of length-minimizers
in (Rd, g), where g ∈ C1,1 is a Riemannian metric. Furthermore, we assume that
(Rd, g) is geodesically complete. Note that the seeming restriction of considering
only Rd as our manifold weighs not so heavy. Indeed, the question of length min-
imizers can be considered to be a local one, since it is not guaranteed that global
minimizers exist at all, whereas local minimizers always exist. Additionally, note
that it was shown that it suffices to consider smooth manifolds (cf. [18, Thm. 2.9])
instead of Ck manifolds with 1 ≤ k < +∞. Therefore, there is no need to consider
non-smooth charts.
some a ∈ R>0, and where Ω ⊆ Rd is an arbitrary open set, see Thm. 18. Actually,
the embedding also depends on the dimension d ∈ N>0, but to avoid cumbersome
notations, we denote embeddings always with the symbol ι.
Ω)Ω, where b ∈ ρ(cid:101)R satisfies b ≥ dρ−a for
ij be the Christoffel symbols of gε, and set Γij := [Γε
ij(−)] ∈ ρGC∞(ρ(cid:101)Rd × ρ(cid:101)Rd, ρ(cid:101)R), then for all ε,
By [24, Rem. 2.6.2], it follows that we can always find a net of smooth functions
(gε
ij) such that setting g := ι(g) =: [gε
ij(−)] ∈ ρGC∞(ρ(cid:101)Rd, ρ(cid:101)Rd).
ij → gij in C0 norm.
ij is a Riemannian metric. By Thm. 18.(iii) it follows that gε
gε
A curve γ ∈ ρGC∞(J, ρ(cid:101)Rd), J being a sharply open subset of ρ(cid:101)R, is said to be a
Let Γε
geodesic of (ρ(cid:101)Rd, g) if
Definition 62. We say that (ρ(cid:101)Rd, g) is geodesically complete if every solution of
the geodesic equation belongs to ρGC∞(ρ(cid:101)R, ρ(cid:101)Rd), i.e. if for all p ∈ ρ(cid:101)Rd and all
v ∈ ρ(cid:101)Rd there exists a geodesic γ ∈ ρGC∞(ρ(cid:101)R, ρ(cid:101)Rd) of (ρ(cid:101)Rd, g) such that γ(0) = p
any geodecis γ ∈ ρGC∞(ρ(cid:101)R, ρ(cid:101)Rd) induces a Colombeau generalized function γc(R) ∈
This definition includes also the possibility that the point p or the vector v could
be infinite. By Thm. 19, it follows that if we consider only finite p and v, then
Gs(R)d. Therefore, the space (c(Rd), gc(Rd)×c(Rd)) is geodesically complete in the
sense of [36]. We recall that c(Ω) is the set of compactly supported (i.e. finite)
generalized points in Ω (see Thm. 18).
γ(t) + Γij(γ(t)) γi(t) γj(t) = 0 ∀t ∈ J.
and γ(0) = v.
(8.1)
The definition of length of a (non singular) curve needs the following
√
Lem. 3 readily implies that
xε] ∈ ρ(cid:101)R>0.
Remark 63. We set√− = (−)1/2 : x = [xε] ∈ ρ(cid:101)R>0 (cid:55)→ [
√− ∈ ρGC∞(ρ(cid:101)R>0, ρ(cid:101)R>0). Therefore, the square root is
defined on every strictly positive infinitesimal, but it cannot be extended to ρ(cid:101)R≥0.
(cid:111)
(cid:110)
λ ∈ ρGC∞([0, 1], ρ(cid:101)Rd) λ(0) = p, λ(1) = q, λ(t) > 0 ∀t ∈ [0, 1]
Let p, q ∈ ρ(cid:101)Rd, then
(i)
ρGC∞
Definition 64.
.
>0(p, q) :=
Moreover, for λ ∈ ρGC∞
>0(p, q), we set
(cid:0)gij(α(t)) αi(t) αj(t)(cid:1)1/2
1
dt ∈ ρ(cid:101)R.
Lg(λ) :=
Let x = [xε] ∈ ρ(cid:101)Rn, then we set st(x) := limε→0 xε ∈ Rd, if this limit exists.
0
(ii)
Note that x ≈ st(x) in this case.
30
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
We are interested only in global minimizers of the functional Lg, i.e. curves
Note that (8.1) are the usual geodesic equations for the generalized metric g,
whose derivation is completely analogous to that in the smooth case. Thus they
are the Euler-Lagrange equations of Lg.
λ0 ∈ X(p, q) such that Lg(λ0) ≤ Lg(λ) for all λ ∈ ρGC∞
Lemma 65. Let p, q ∈ Rd and p, q ∈ ρ(cid:101)Rd such that st(p) = p and st(q) = q. Let
>0(p, q) :=(cid:8)w ∈ C1([0, 1]R, Rd) w(0) = p, w(1) = q, w(t) > 0 ∀t ∈ [0, 1]R(cid:9)
λ = [λε(−)] ∈ ρGC∞
¯λ ∈ C1
such that λε → ¯λ in C1 as ε → 0, then
>0(p, q) be such that there exists
>0(p, q).
Proof. We calculate:
(cid:17)1/2 −(cid:16)
ij(λε) λi
gε
ε
λj
ε
(cid:16)
(cid:12)(cid:12)(cid:12)(cid:12)
1
0
By assumption, (gε
such that
ε
ij(λε) λi
(cid:17)1/2 −(cid:16)
λj
ε
(cid:16)
(cid:12)(cid:12)(cid:12)(cid:12)
1
0
ij(λε) λi
gε
ε
(cid:12)(cid:12)(cid:12)(gε
1
0
≤ C
st(Lg(λ)) = Lg(¯λ).
(cid:12)(cid:12)(cid:12)(cid:12) =
dt
gij(¯λ) ¯λi ¯λj(cid:17)1/2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
gij(¯λ) ¯λi ¯λj(cid:17)1/2
=
1
0
(cid:12)(cid:12)(cid:12)(cid:12)
dt
ε − gij(¯λ) ¯λi ¯λj
λj
ij(λε) λi
gε
ε)1/2 + (gij(¯λ) ¯λi ¯λj)1/2
λj
ε
(gε
ij(λε) λi
ε
ε)1/2 → (gij(¯λ) ¯λi ¯λj)1/2, so that there exists C ∈ R>0
λj
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) .
dt
ij(λε) − gij(λε) + gij(λε) − gij(¯λ)) λi
ε
λj
ε + gij(¯λ)( λi
ε
ε − ¯λi ¯λj)
λj
(cid:12)(cid:12)(cid:12) dt.
We hence obtain the claim by the triangle inequality and by convergence of λε, λε
(cid:3)
and gε
ij to ¯λ, ¯λ and gij respectively.
Now, we consider p, q ∈ Rd with p (cid:54)= q. Let
u ∈(cid:8)u ∈ C2,1([0, 1], Rd) u(0) = p, u(1) = q(cid:9)
Let c0 := u(0). Obviously, u is also the unique solution of
p = u(0)
q = u(1).
be a solution of the geodesic equationu = −Γij(u) ui uj
u = −Γij(u) ui uj
y = −Γε
p = u(0)
c0 = u(0).
p = y(0)
c0 = y(0).
ij(y) yi yj
Using these initial conditions, for each fixed ε we can solve the following problem
for a unique yε ∈ C∞([−dε, dε]R, Rd) and some dε ∈ R>0.
Lemma 66. Let u and yε be as above. Then
(i)
For ε sufficiently small, the solution yε can be extended to a solution yε ∈
C∞([0, 1]R, Rd) of (8.4) such that yε(1) = q.
(8.2)
(8.3)
(8.4)
CALCULUS OF VARIATIONS FOR GF
31
yε → u in C2.
(ii)
(iii) The net (yε) defines a GSF, i.e. y := [yε(−)] ∈ ρGC∞
Proof.
ij → Γij locally uniformly. Thus, we
Claim (i), (ii): For all i, j , we have that Γε
obtain these claims by (8.2) and by continuous dependence on parameters in ODE,
see e.g. [25, Lemma 2.3].
>0(p, q).
We have to show that for all ε all derivatives of yε are moderate. This is obviously
true for yε, yε and yε. The claim follows now from the fact that
Claim (iii) I: y := [yε(−)] ∈ ρGC∞([0, 1], ρ(cid:101)Rd)
(cid:0)Γεk
dn+2
(cid:0)Γεk
dn
dtn
ij (y) yi
ε yj
ε
dtn+2 yε = − dn
(cid:18)
(cid:1) = P
yε, . . . ,
dtn
yε,
d
dt
so that there exists a polynomial P such that
(cid:1)
ij (y) yi
ε yj
ε
dn+1
dtn+1 yε, Γεk
ij , DΓεk
ij , . . . , DnΓεk
ij
(cid:19)
.
Claim (iii) II: y(t) > 0 for all t ∈ [0, 1]
By (ii), we have that yε → u in C2. Furthermore, gε → g in C1 by assumption,
and we know that g( u, u) = c > 0 for some c ∈ R>0 since u is a g-geodesic (cf. [20,
Lemma 1.4.5]). Therefore, we obtain that gε( yε, yε) > c/2 > 0 for ε > 0 small
(cid:3)
enough.
Finally, the standard part of the generalized length of y is the length of u:
Theorem 67. Let u and yε be as above. We conclude (using Lemma 65) that
st(Lg(y)) = Lg(u).
Proposition 68. Let y = [yε(−)] be as above. In addition, assume that each yε is
Lgε-minimizing. Then Lg(y) is minimal.
Proof. Let λ = [λε(−)] ∈ ρGC∞
Lg(y) = [Lgε(yε)]. By assumption, for all ε we have
Lgε(λε) ≥ Lgε(yε).
>0(p, q). We have that Lg(λ) = [Lgε(λε)] and that
Therefore, Lg(λ) ≥ Lg(y), as claimed.
Corollary 69. Let λ ∈ ρGC∞
small, yε is Lgε-minimizing. Then Lg(y) = Lg(λ).
>0(p, q) be a minimizer of Lg and assume that for ε
(cid:3)
This Corollary 69 gives us a way to answer the question if a certain classical
geodesic between two given classical points p and q is a length-minimizer.
Furthermore, we are able to prove the following theorem, relating GSF-minimizers
to classical minimizers.
Theorem 70. Let p, q ∈ Rd and let γ ∈ ρGC∞
Assume that st(Lg(γ)) exists and that there exists w ∈ C1
st(Lg(γ)).
>0(p, q) such that Lg(γ) is minimal.
>0(p, q) such that Lg(w) =
Then w is g - minimizing and a g - geodesic.
Proof. Assume to the contrary that there exists a curve σ ∈ C 2 connecting p and
q (w. l. o. g. σ is a g - geodesic) such that
Lg(σ) < Lg(w).
Now we construct (as done above) gε, σε and set σ := [σε]. Then:
st (Lg(σ)) = Lg(σ) < Lg(w) = st (Lg(γ)) .
32
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
But, by assumption we have that Lg(γ) ≤ Lg(σ), which implies
st (Lg(γ)) ≤ st (Lg(σ)) < st (Lg(γ)) .
This is a contradiction.
(cid:3)
9. Conclusions
We can summarize the present work as follows
(i)
The setting of GSF allows to treat Schwartz distributions more closely to
classical smooth functions. The framework is so flexible and the extensions
of classical results are so natural in many ways one may treat it like smooth
functions.
(ii) One key step of the theory is the change of the ring of scalars into a non-
Archimedean one and the use of the strict order relation < to deal with
topological properties. So, the use of < and of ρ(cid:101)R-valued norms allows a nat-
ural approach to topology, even of infinite dimensional spaces (cf. Def. 41).
On the other hand, the use of a ring with zero divisors and a non-total order
relation requires a more refined and careful analysis. However, as proved in
the present work, very frequently classical proofs can be formally repeated
in this context, but paying particular attention to using the relation <, us-
ing invertibility instead of being non zero in R and avoiding the total order
property.
(iii) Others crucial properties are the closure of GSF with respect to composition
and the use of the gauge ρ, because they do not force to narrow the theory
into particular cases.
(iv) The present extension of the classical theory of calculus of variations shows
that GSF are a powerful analytical technique. The final application shows
how to use them as a method to address problems in an Archimedean setting
based on the real field R.
Concerning possible future developments, we can note that:
(v) A generalization of the whole construction to piecewise GSF seems possible.
(vi) A more elegant approach to integration of piecewise GSF could use the exis-
tence of right and left limits of (f1, . . . , fn)(−) and hyperfinite Riemann-like
sums, i.e. sums
∈ ρ(cid:101)Rd
, where int(−) is the integer
N(cid:88)
f (x(cid:48)
(cid:34) Nε(cid:88)
i)(xi − xi−1) :=
(cid:110)
[int(xε)] [xε] ∈ ρ(cid:101)R(cid:111)
(cid:35)
i,ε)(xi,ε − xi−1,ε)
fε(x(cid:48)
i=1
extended to N ∈ (cid:101)N :=
i=1
part function.
The present work could lay the foundations for further works concerning the possi-
bility to extend other results of the calculus of variations in this generalized setting.
References
[1] Aragona, J., Fernandez, R., Juriaans, S.O., A discontinuous Colombeau differential calculus,
Monatsh. Math. 144, 13–29 (2005).
[2] Aragona, J., Juriaans, S.O., Some structural properties of the topological ring of Colombeau's
generalized numbers, Comm. Algebra 29 (2001).
[3] Avez, A., Differential Calculus, John Wiley & Sons Inc., 1986.
[4] Colombeau, J.F., Multiplication of distributions - A tool in mathematics, numerical engi-
neering and theoretical Physics. Springer-Verlag, Berlin Heidelberg (1992).
[5] Davie, A. M., Singular minimisers in the calculus of variations in one dimension. Arch. Ra-
tional Mech. Anal. 101(2), 161–177, 1988.
[6] Dirac, P.A.M., The physical interpretation of the quantum dynamics, Proc. R. Soc. Lond.
A, 113 , 1926–27, 621–641.
CALCULUS OF VARIATIONS FOR GF
33
[7] Engquist, B., Tornberg, A.K., Tsai, R., Discretization of Dirac delta functions in level set
methods. Journal of Computational Physics, 207:28–51, 2005.
[8] Gelfand, I.M., Fomin, S.V., Calculus of variations, Dover Publications, 2000.
[9] Giordano, P., Kunzinger, M., Inverse Function Theorems for Generalized Smooth Functions.
Invited paper for the Special issue ISAAC - Dedicated to Prof. Stevan Pilipovic for his 65
birthday. Eds. M. Oberguggenberger, J. Toft, J. Vindas and P. Wahlberg, Springer series
"Operator Theory: Advances and Applications", Birkhaeuser Basel, 2016.
[10] Giordano, P., Kunzinger, M., A convenient notion of compact sets for generalized func-
tions. Accepted in Proceedings of the Edinburgh Mathematical Society, 2016. See arXiv
1411.7292v1.
[11] Giordano, P., Kunzinger, M., 'New topologies on Colombeau generalized numbers and the
Fermat-Reyes theorem'. Journal of Mathematical Analysis and Applications 399 (2013)
229–238.
[12] Giordano P., Kunzinger M., Steinbauer R., A new approach to generalized functions for
mathematical physics. See http://www.mat.univie.ac.at/giordap7/GenFunMaps.pdf.
[13] Giordano, P., Kunzinger, M., Vernaeve, H., Strongly internal sets and generalized smooth
functions. Journal of Mathematical Analysis and Applications, volume 422, issue 1, 2015, pp.
56–71.
[14] Giordano, P., Luperi Baglini, L., Asymptotic gauges: Generalization of Colombeau type
algebras. Math. Nachr. 289, 2-3, 1–28, (2015).
[15] Graves, L.M., Discontinuous solutions in the calculus of variations. Bull. Amer. Math. Soc.
36, 831–846, 1930.
[16] Grosser, M., Kunzinger, M., Oberguggenberger, M., Steinbauer, R., Geometric theory of
generalized functions, Kluwer, Dordrecht (2001).
[17] Hawking, S.W., Ellis, G., The large scale structure of space-time, Cambridge University Press
(1976)
[18] Hirsch, M.W., Differential Topology, Springer (1976).
[19] Hosseini, B. , Nigam, N., Stockie, J.M., On regularizations of the Dirac delta distribution,
Journal of Computational Physics, Volume 305, 2016, Pages 423–447.
[20] Jost, J., Riemannian Geometry and Geometric Analysis, Springer, 2011.
[21] Jost, J., Li-Jost, X., Calculus of variations, Cambridge Studies in Advanced Mathematics 64,
1998.
[22] Katz, M.G., Tall, D., A Cauchy-Dirac delta function. Foundations of Science, 2012. See
http://dx.doi.org/10.1007/s10699-012-9289-4 and http://arxiv.org/abs/1206.0119.
[23] Konjik, S., Kunzinger, M., Oberguggenberger, M.: Foundations of the Calculus of Variations
in Generalized Function Algebras. Acta Applicandae Mathematicae 103 n. 2, 169–199 (2008)
[24] Kunzinger, M., Steinbauer, R., Stojkovi´c, M., Vickers, J.A., A regularisation approach to
causality theory for C1,1-Lorentzian metrics, Gen. Relativ. Gravit. 46 (2014).
[25] Kunzinger, M., Steinbauer, R., Stojkovi´c, M., The exponential map of a C1,1-metric, Diff.
Geom. Appl. 34, 14 –24 (2014).
[26] Laugwitz, D., Definite values of infinite sums: aspects of the foundations of infinitesimal
analysis around 1820. Arch. Hist. Exact Sci. 39 (1989), no. 3, 195–245.
[27] Lecke, A., Steinbauer, R., Svarc, R., The regularity of geodesics in impulsive pp-waves, Gen.
Relativ. Gravit. 46 (2014).
[28] Lecke, A., Non-smooth Lorentzian Geometry and Causality Theory, PhD Thesis, Universitat
Wien (2016).
[29] Luperi Baglini, L., Giordano, P., Fixed point iteration methods for arbitrary generalized
ODE, preprint.
[30] Luperi Baglini, L., Giordan, P., The category of Colombeau algebras. In revision for Monat-
shefte fur Mathematik. See arXiv 1507.02413.
[31] Lytchak, A., Yaman, A., On Hoelder continuous Riemannian and Finsler metrics, Trans.
Amer. Math. Soc. 358 (2006).
[32] Minguzzi, E., Convex neighborhoods for Lipschitz connections and sprays, Monatshefte fur
Mathemati, Volume 177, Issue 4, pp 569-625 (2015)
[33] Oberguggenberger, M., Vernaeve, H., Internal sets and internal functions in Colombeau the-
ory, J. Math. Anal. Appl. 341 (2008) 649–659.
[34] Robinson, A., Function theory on some nonarchimedean fields, Amer. Math. Monthly 80 (6)
87–109; Part II: Papers in the Foundations of Mathematics (1973).
[35] Samann, C., Steinbauer, R., Lecke, A., Podolsk´y, J.,Geodesics in nonexpanding impulsive
gravitational waves with Λ, part I, Classical and Quantum Gravity 33 (2016).
[36] Samann,C. , Steinbauer, R., Geodesic Completeness of Generalized Space-times. In Eds S.
Pilipovi´c, J., Pseudo-Differential Operators and Generalized Functions, Volume 245 of the
series Operator Theory: Advances and Applications pp 243-253, 2015.
34
ALEXANDER LECKE, LORENZO LUPERI BAGLINI, AND PAOLO GIORDANO
[37] Stojkovi´c, M., Causality theory for C1,1 - metrics, PhD Thesis, Universitat Wien (2015).
[38] Tornberg, A.K., Engquist, B., Numerical approximations of singular source terms in differ-
ential equations, Journal of Computational Physics 200 (2004) 462–488.
[39] Tuckey, C., Nonstandard methods in the calculus of variations, Pitman Research Notes in
Mathematics Series 297. Longman Scientific & Technical, Harlow, 1993.
University of Vienna, Austria
E-mail address: [email protected]
University of Vienna, Austria
E-mail address: [email protected]
E-mail address: [email protected]
|
1806.10752 | 4 | 1806 | 2019-01-28T10:47:49 | Paley-Wiener properties for spaces of entire functions | [
"math.FA"
] | We deduce Paley-Wiener results in the Bargmann setting, which give characterisations of Pilipovi{\'c} spaces of low orders, extending the characterisation of a Gr{\"o}chenig test function space, deduced earlier by the third author. | math.FA | math |
PALEY-WIENER PROPERTIES FOR SPACES OF
ENTIRE FUNCTIONS
ELMIRA NABIZADEH, CHRISTINE PFEUFFER, AND JOACHIM TOFT
Abstract. We deduce Paley-Wiener results in the Bargmann set-
ting. At the same time we deduce characterisations of Pilipović
spaces of low orders. In particular we improve the characterisation
of the Gröchenig test function space H♭1
= SC , deduced in [12].
0. Introduction
Paley-Wiener theorems characterize functions and distributions with
certain restricted supports in terms of estimates of their Fourier-Laplace
transforms. For example, let f be a distribution on Rd and let Br0(0) ⊆
Rd be the ball with center at origin and radius r0. Then f is supported
in Br0(0) if and only if
bf (ζ) . hζiN er0 Im(ζ),
bf (ζ) . hζi−N er0 Im(ζ),
ζ ∈ Cd,
ζ ∈ Cd,
for some N ≥ 0. Furthermore, f is supported in Br0(0) and smooth, if
and only if
for every N ≥ 0 (see e. g. [6, Section 7]).
A similar approach for ultra-regular functions of Gevrey types and
corresponding ultra-distribution spaces can be done. In fact, let s > 1,
D′
s(Rd) be the set of all Gevrey distributions of order s and let Es(Rd)
be the set of all smooth functions with Gevrey regularity s. (See [6]
s(Rd)
and Section 1 for notations.) Then it can be proved that f ∈ D′
is supported in Br0(0), if and only if
bf (ζ) . er0 Im(ζ)+rζ
bf (ζ) . er0 Im(ζ)−rζ
1
s ,
ζ ∈ Cd,
1
s ,
ζ ∈ Cd,
for every r > 0. Furthermore f ∈ Es(Rd) is supported in Br0(0), if and
only if
for some r > 0.
We observe that s in the latter result can not be pushed to be smaller,
because if s ≤ 1, it does not make any sense to discuss compact support
properties of D′
s(Rd) and Es(Rd).
2010 Mathematics Subject Classification. primary 46F05; 32A25; 32A36; sec-
ondary 35Q40; 30Gxx.
Key words and phrases. Bargmann transform.
1
In the paper we consider analogous Paley-Wiener properties when
the Fourier-Laplace transform above is replaced by the reproducing ker-
nel ΠA of the Bargmann transform, and the image spaces are replaced
by suitable subspaces of entire functions on Cd. These subspaces were
considered in [4, 12] and are given by
A♭σ (Cd) =[r>0
As(Cd) =[r>0
Ar,♭σ(Cd),
Ar,s(Cd),
A0,♭σ (Cd) =\r>0
A0,s(Cd) =\r>0
Ar,♭σ(Cd),
Ar,s(Cd),
(0.1)
when σ > 0 and 0 < s < 1
of all entire F on Cd such that
2, where Ar,♭σ (Cd) and Ar,s(Cd) are the sets
F (z) . erz
2σ
σ+1
respective
F (z) . er(loghzi)
1
1−2s .
The spaces in (0.1) appear naturally when considering the Bargmann
transform images of an extended class of Fourier invariant Gelfand-
Shilov spaces, called Pilipović spaces (see [4, 12]).
If (z, w) is the scalar product of z, w ∈ Cd, then the reproducing
kernel of the Bargmann transform is given by
(ΠAF )(z) = π−dhF, e(z, · )− · 2
i,
when F is a suitable function or (ultra-)distribution. If
z 7→ F (z)eRz−z2
∈ L1(Cd), R > 0
(0.2)
holds and dλ(w) is the Lebesgue measure on Cd, then
(ΠAF )(z) = π−dZCd
F (w)e(z,w)−w2
dλ(w).
A recent Paley-Wiener result with respect to the transform ΠA and
to the image spaces (0.1) is given in [12], where it is proved that if
c (Cd) = E ′(Cd) ∩ L∞(Cd), then
L∞
ΠA(E ′(Cd)) = ΠA(L∞
(0.3)
Evidently, L∞
c ⊆ E ′, and the gap between these spaces is rather large.
It might therefore be somehow surprising that the first equality holds
in (0.3).
c (Cd)) = A♭1(Cd).
In Section 2 we improve (0.3) in different ways. Firstly we show that
A,c(Cd) given
c (Cd) in (0.3) with the smaller space L∞
we may replace L∞
by
where L∞
A,c(K) is the set of all F · χK, where F is analytic in a neigh-
bourhood of K and χK is the characteristic function of K. Secondly,
L∞
A,c(Cd) =[ L∞
A,c(K),
2
s(Cd) of all compactly supported
we may replace E ′(Cd) with the set E ′
Gevrey distributions of order s > 1. Summing up we improve (0.3) into
(0.3)′
s(Cd)) = ΠA(L∞
s > 1.
A,c(Cd)) = A♭1(Cd),
ΠA(E ′
In Section 2 we also deduce various kind of related mapping prop-
erties when A♭1(Cd) in (0.3) is replaced by any of the spaces in (0.1).
More precisely, let χ ∈ L∞
c (Cd) be non-negative, radial symmetric in
each complex variable zj and bounded from below by a positive con-
stant near the origin. Then we prove
σ
σ
0,♭σ0
2σ − 1
ΠA(A′
2 , 1), σ0 =
ΠA(A♭σ0
ΠA(As(Cd) · χ) = As(Cd),
(Cd) · χ) = A♭σ (Cd), σ ∈ ( 1
(Cd) · χ) = A♭σ (Cd), σ ∈ (0, 1
s ∈ [0, 1
2).
Some related properties are deduced for σ = 1
2, as well as when A♭σ and
As are replaced by A0,♭σ and A0,s, respectively. (Cf. Theorems 2.2 -- 2.8
and Propositions 2.9 -- 2.11.)
Finally, in Section 3 we use the results in Section 2 to deduce char-
1 − 2σ
2), σ0 =
,
,
acterizations of Pilipović spaces of small orders.
1. Preliminaries
In this section we recall some basic facts. We start by discussing
Pilipović spaces and some of their properties. Then we recall some facts
on modulation spaces. Finally we discuss the Bargmann transform and
some of its mapping properties, and introduce suitable classes of entire
functions on Cd.
1.1. The Pilipović spaces. The definition of Pilipović spaces is based
on the Hermite functions, which are given by
hα(x) = π− d
4 (−1)α(2αα!)− 1
2 e
x2
2 (∂αe−x2
), α ∈ Nd.
The Hermite functions are eigenfunctions for the Fourier transform,
and for the Harmonic oscillator Hd ≡ x2 − ∆ which acts on functions
and (ultra-)distributions defined on Rd. More precisely, we have
Hdhα = (2α + d)hα.
It is well-known that the set of Hermite functions is a basis of S (Rd)
and an orthonormal basis of L2(Rd) (cf. [10]). In particular, if f ∈
L2(Rd), then
where
L2(Rd) = Xα∈Nd ch(f, α)2,
kfk2
f (x) = Xα∈Nd
ch(f, α)hα
3
(1.1)
+ and s ∈ R♭ we set
e−( 1
d ),
rαα!− 1
2σ ,
1
1 +···+ 1
2s
rd
1
2s
·α
·α
r1
rα,
·α
1
1 +···+ 1
2s
rd
·α
1
2s
d ),
r1
e( 1
rαα!
1
2σ ,
s ∈ R+ \ { 1
2},
s = ♭σ,
s = 1
2,
α ∈ Nd
s ∈ R+ \ { 1
2},
s = ♭σ,
(1.3)
(1.4)
and
ϑr,s(α) ≡
ϑ′
r,s(α) ≡
is the Hermite seriers expansion of f , and
ch(f, α) = (f, hα)L2(Rd)
is the Hermite coefficient of f of order α ∈ Rd.
should belong to the extended set
In order to define the full scale of Pilipović spaces, their order s
(1.2)
of positive real numbers, with extended inequality relations as
R♭ = R+ ∪ { ♭σ ; σ ∈ R+ },
s1 < ♭σ < s2
and ♭σ1 < ♭σ2
when s1 < 1
2 ≤ s2 and σ1 < σ2. (Cf. [12].)
For such r ∈ Rd
rα,
s = 1
2 ,
α ∈ Nd.
Definition 1.1. Let s ∈ R♭ = R♭ ∪ {0}, and let ϑr,s and ϑ′
(1.3) and (1.4).
(1) H0(Rd) consists of all Hermite polynomials, and H′
sists of all formal Hermite series expansions in (1.1);
(2) if s ∈ R♭, then Hs(Rd) (H0,s(Rd)) consists of all f ∈ L2(Rd)
0(Rd) con-
r,s be as in
such that
ch(f, hα) . ϑr,s(α)
holds true for some r ∈ Rd
s(Rd) (H′
(3) if s ∈ R♭, then H′
series expansions in (1.1) such that
+ (for every r ∈ Rd
+);
0,s(Rd)) consists of all formal Hermite
ch(f, hα) . ϑ′
r,s(α)
holds true for every r ∈ Rd
+ (for some r ∈ Rd
+).
The spaces Hs(Rd) and H0,s(Rd) are called Pilipović spaces of Roumieu
respectively Beurling types of order s, and H′
0,s(Rd) are
called Pilipović distribution spaces of Roumieu respectively Beurling
types of order s.
s(Rd) and H′
4
Remark 1.2. Let Ss(Rd) and Σs(Rd) be the Fourier invariant Gelfand-
Shilov spaces of order s ∈ R+ of Roumieu respective Beurling types
(see [12] for notations). Then it is proved in [8, 9] that
H0,s(Rd) = Σs(Rd) 6= {0},
H0,s(Rd) 6= Σs(Rd) = {0},
Hs(Rd) = Ss(Rd) 6= {0},
s >
s ≤
s ≥
Hs(Rd) 6= Ss(Rd) = {0},
s <
,
,
1
2
1
2
1
2
1
2
.
and
In Proposition 1.3 below we give further characterisations of Pilipović
spaces.
Next we recall the topologies for Pilipović spaces. Let s ∈ R♭, r > 0,
and let kfkHs;r and kfkH′
s ∈ R♭
s ∈ R♭,
(1.5)
(1.6)
s;r be given by
α∈Nd ch(f, α)ϑ′
r,s(α),
kfkHs;r ≡ sup
and
kfkH′
s;r ≡ sup
α∈Nd ch(f, α)ϑr,s(α),
when f is a formal expansion in (1.1). Then Hs;r(Rd) consists of all
expansions (1.1) such that kfkHs;r is finite, and H′
s;r(Rd) consists of
all expansions (1.1) such that kfkH′
s;r is finite. It follows that both
s;r(Rd) are Banach spaces under the norms f 7→ kfkHs;r
Hs;r(Rd) and H′
and f 7→ kfkH′
s;r, respectively.
We let the topologies of Hs(Rd) and H0,s(Rd) be the inductive re-
spectively projective limit topology of Hs;r(Rd) with respect to r > 0.
0,s(Rd) are the pro-
In the same way, the topologies of H′
s(Rd) and H′
jective respectively inductive limit topology of H′
s;r(Rd) with respect
to r > 0. It follows that all the spaces in Definition 1.1 are complete,
s(Rd) are Fréchet spaces with semi-norms
and that H0,s(Rd) and H′
f 7→ kfkHs;r and f 7→ kfkH′
s;r, respectively.
[12]. The proof is therefore omitted.
0(Rd). Then
Proposition 1.3. Let s ∈ R+ ∪ {0} and let f ∈ H′
f ∈ H0,s(Rd) (f ∈ Hs(Rd)), if and only if f ∈ C ∞(Rd) and satis-
fies kH N
d fkL∞ . hN N!2s for every h > 0 (for some h > 0).
The following characterisations of Pilipović spaces can be found in
From now on we let
φ(x) = π− d
4 e− x2
2 .
5
(1.7)
1.2. Spaces of entire functions and the Bargmann transform.
Let Ω ⊆ Cd be open. Then A(Ω) is the set of all analytic functions in Ω.
If instead Ω ⊆ Cd is closed, then A(Ω) is the set of all functions which
are analytic in an open neighbourhood of Ω. In particular, if z0 ∈ Cd
is fixed, then A({z0}) is the set of all complex-valued functions which
are defined and analytic near z0.
We shall now consider the Bargmann transform which is defined by
the formula
(Vdf )(z) = π− d
when f ∈ L2(Rd) (cf. [1]). Here
exp(cid:16) −
1
2
(hz, zi + y2) + 2
1
2hz, yi(cid:17)f (y) dy,
hz, wi =
zjwj, when z = (z1, . . . , zd) ∈ Cd
4ZRd
dXj=1
and w = (w1, . . . , wd) ∈ Cd,
and otherwise h · , · i denotes the duality between test function spaces
and their corresponding duals. It is evident that Vdf is the entire func-
tion on Cd, given by
(Vdf )(z) =ZRd
Ad(z, y)f (y) dy,
or
where the Bargmann kernel Ad is given by
(Vdf )(z) = hf, Ad(z, · )i,
(1.8)
Ad(z, y) = π− d
4 exp(cid:16) −
1
2
(hz, zi + y2) + 2
1
2hz, yi(cid:17).
(Rd)
We note that the right-hand side in (1.8) makes sense when f ∈ S ′
and defines an element in A(Cd), since y 7→ Ad(z, y) can be interpreted
(Rd) with values in A(Cd). Here and in what follows,
as an element in S 1
A(Ω) denotes the set of analytic functions on the open set Ω ⊆ Cd.
It was proved in [1] that f 7→ Vdf is a bijective and isometric map
from L2(Rd) to the Hilbert space A2(Cd) ≡ B2(Cd) ∩ A(Cd), where
B2(Cd) consists of all measurable functions F on Cd such that
1
2
2
kFkB2 ≡(cid:16)ZCd F (z)2dµ(z)(cid:17) 1
2 < ∞.
(1.9)
Here dµ(z) = π−de−z2 dλ(z), where dλ(z) is the Lebesgue measure on
Cd. We recall that A2(Cd) and B2(Cd) are Hilbert spaces, where the
scalar product is given by
(F, G)B2 ≡ZCd
F (z)G(z) dµ(z), F, G ∈ B2(Cd).
(1.10)
If F, G ∈ A2(Cd), then we set kFkA2 = kFkB2 and (F, G)A2 = (F, G)B2.
6
Furthermore, Bargmann showed that there is a convenient reproduc-
ing formula on A2(Cd). More precisely, let
F (w)e(z,w) dµ(w),
(1.11)
(ΠAF )(z) ≡ZCd
dXj=1
when z 7→ F (z)eRz−z2 belongs to L1(Cd) for every R ≥ 0. Here
(z, w) =
zjwj, when z = (z1, . . . , zd) ∈ Cd
and w = (w1, . . . , wd) ∈ Cd,
is the scalar product of z ∈ Cd and w ∈ Cd. Then it is proved in
[1, 2] that ΠA is the orthogonal projection of B2(Cd) onto A2(Cd). In
particular, ΠAF = F when F ∈ A2(Cd).
In [1] it is also proved that
Vdhα = eα, where
eα(z) ≡
zα
√α!
,
z ∈ Cd.
(1.12)
In particular, the Bargmann transform maps the orthonormal basis
{hα}α∈Nd of L2(Rd) bijectively into the orthonormal basis {eα}α∈Nd of
monomials of A2(Cd). Hence, there is a natural way to identify formal
Hermite series expansion by formal power series expansions
c(F, α)eα(z),
(1.13)
F (z) = Xα∈Nd
Xα∈Nd
ch(f, α)eα(z).
by letting the series (1.1) be mapped into
(1.14)
(1.15)
It follows that if f, g ∈ L2(Rd) and F, G ∈ A2(Cd), then
(f, g)L2(Rd) = Xα∈Nd
(F, G)A2(Cd) = Xα∈Nd
ch(f, α)ch(g, α),
c(F, α)c(G, α).
Here and in what follows, ( · , · )L2(Rd) and ( · , · )A2(Cd) denote the scalar
products in L2(Rd) and A2(Cd), respectively. Furthermore,
ch(f, α) = c(F, α) when F = Vdf.
(1.16)
We now recall the following spaces of power series expansions given
in [12].
Definition 1.4. Let s ∈ R♭ = R♭ ∪ {0}, and let ϑr,s and ϑ′
(1.3) and (1.4).
r,s be as in
7
(1) A0(Cd) consists of all analytic polynomials on Cd, and A′
0(Cd)
consists of all formal power series expansions on Cd in (1.13);
(2) if s ∈ R♭, then As(Cd) (A0,s(Cd)) consists of all F ∈ L2(Cd)
such that
c(F, hα) . ϑr,s(α)
holds true for some r > 0 (for every r > 0);
0,s(Cd)) consists of all formal power
(3) if s ∈ R♭, then A′
s(Cd) (A′
series expansions in (1.13) such that
c(F, hα) . ϑ′
r,s(α)
holds true for every r > 0 (for some r > 0).
Let f ∈ H′
0(Rd) with formal Hermite series expansion (1.1). Then the
Bargmann transform Vdf of f is defined to be the formal power series
expansion (1.14). It follows that Vd agrees with the earlier definition
when acting on L2(Rd), that Vd is linear and bijective from H′
0(Rd) to
A′
0(Cd), and restricts to bijections from the spaces
H0,s(Rd), Hs(Rd), H′
s(Rd) and H′
0,s(Rd)
to
A0,s(Cd), As(Cd), A′
s(Cd) and A′
0,s(Cd)
(1.17)
(1.18)
If s ∈ R♭, f ∈ Hs(Rd), g ∈ H′
respectively, when s ∈ R♭. We also let the topologies of the spaces in
(1.18) be inherited from the spaces in (1.17).
s(Cd),
then (f, g)L2(Rd) and (F, G)A2(Cd) are defined by the formula (1.15).
It follows that (1.16) holds for such choices of f and g. Furthermore,
the duals of Hs(Rd) and As(Cd) can be identified with H′
s(Rd) and
s(Cd), respectively, through the forms in (1.15). The same holds true
A′
with
s(Rd), F ∈ As(Cd) and G ∈ A′
H0,s, H′
0,s, A0,s,
and A′
0,s
in place of
Hs, H′
s, As,
and A′
s,
respectively, at each occurrence.
8
In order to identify the spaces of power series expansions above with
spaces of analytic functions, we let
1
1
2σ
z2
M 0
r1(loghz1i)
r1z1
2 − (r1z1
M1,r,s(z) =
1,r,s(z) =(M1,r,s(z),
M2,r,s(z) =
2,r,s(z) =(M2,r,s(z),
M 0
2σ
r1z12 + · · · + rdzd2,
r1z1
σ−1 + · · · + rdzd
2 + (r1z1
z2
1
2σ
σ−1 ,
1
s ),
s + · · · + rdzd
s 6= 1
2,
s = 1
2,
r1z12 + · · · + rdzd2,
1
1−2s ,
s < 1
2,
s = ♭σ, σ > 0,
s ≥ 1
2,
1−2s + · · · + rd(loghzdi)
σ+1 + · · · + rdzd
2σ
σ+1 ,
1
s ),
s + · · · + rdzd
s 6= 1
2,
s = 1
2,
s = ♭σ, σ > 1,
s ≥ 1
2,
(1.19)
when r ∈ Rd
+ and z ∈ Cd. For conveniency we set Mr = M1,r,♭1.
By [12] we have the following. The proof is therefore omitted.
+. Then
2,r,s be as in (1.19)
1,r,s, M2,r,s and M 0
Proposition 1.5. Let M1,r,s, M 0
when s ∈ R♭ and r ∈ Rd
A0,s(Cd) = { F ∈ A(Cd) ; F e−M 0
1,r,s ∈ L∞(Cd) for every r ∈ Rd
As(Cd) = { F ∈ A(Cd) ; F e−M1,r,s ∈ L∞(Cd) for some r ∈ Rd
A′
s(Cd) = { F ∈ A(Cd) ; F e−M2,r,s ∈ L∞(Cd) for every r ∈ Rd
0,s(Cd) = { F ∈ A(Cd) ; F e−M 0
A′
2,r,s ∈ L∞(Cd) for some r ∈ Rd
A′
♭1(Cd) = A(Cd)
Next we recall the link between the Bargmann transform and certain
types of short-time Fourier transforms. The short-time Fourier trans-
1/2(Rd) with respect to the window φ ∈ S1/2(Rd) \ 0 is
form of f ∈ S ′
defined by
0,♭1(Cd) = A({0}).
+ },
+ },
+ },
+ },
and A′
s > 0,
s > 0,
s > ♭1,
s > ♭1,
Vφf (x, ξ) ≡ hf, φ( · − x)e−ih · ,ξii.
We assume from now on that φ is given by (1.7),
Let S be the dilation operator, defined by
(SF )(x, ξ) = F (2− 1
2 x,−2− 1
2 ξ),
9
(1.20)
when F ∈ L1
that
loc(R2d). Then it follows by straight-forward computations
(Vdf )(z) = (Vdf )(x + iξ) = (2π)
d
2 e
1
2 (x2+ξ2)e−ihx,ξiVφf (2
1
2 x,−2
1
2 ξ)
= (2π)
d
2 e
1
2 (x2+ξ2)e−ihx,ξi(S−1(Vφf ))(x, ξ),
(1.21)
or equivalently,
Vφf (x, ξ) = (2π)− d
2 e− 1
4 (x2+ξ2)e−ihx,ξi/2(Vdf )(2− 1
2 x,−2− 1
2 ξ)
= (2π)− d
2 e−ihx,ξi/2S(e− · 2
2 (Vdf ))(x, ξ).
(1.22)
We observe that (1.21) and (1.22) can be formulated as
V ◦ Vd = Vφ,
Vd = UV ◦ Vφ,
and U −1
where UV is the linear, continuous and bijective operator on D ′(R2d) ≃
D ′(Cd), given by
(UVF )(x, ξ) = (2π)
d
2 e
1
2 (x2+ξ2)e−ihx,ξiF (2
Let Dd,r(z0) be the polydisc
1
2 x,−2
1
2 ξ).
(1.23)
{ z = (z1, . . . , zd) ∈ Cd ; zj − z0,j < rj, j = 1, . . . , d }
with center and radii given by
z0 = (z0,1, . . . , z0,d) ∈ Cd
Then
and r = (r1, . . . , rd) ∈ Rd
+.
A(Cd) = ∩r∈Rd
+
A(Dd,r(z)),
A({0}) = ∪r∈Rd
+
A(Dd,r(z0)).
1.3. Hilbert spaces of power series expansions and analytic
functions. The spaces in Definition 1.4 can also be dscribed by related
unions and intersections of Hilbert spaces of analytic functions and
power series expansions as follows. (See also [12].)
A weight ω on a Borel set Ω is a positive function in L∞
loc(Ω) such
loc(Ω). Let ϑ be a weight on Nd, ω be a weight
that 1/ω belongs to L∞
on Cd, and let
when F ∈ A′
0(Cd) is given by (1.13), and
2
,
[ϑ](Cd) ≡ Xα∈Nd c(F, α)ϑ(α)2! 1
2 z)2 dµ(z)(cid:19) 1
(ω)(Cd) ≡(cid:18)ZCd F (z)ω(2
1
2
(1.24)
,
(1.25)
kFkA2
kFkA2
when F ∈ A(Cd). We let A2
is finite, and A2
that kFkA2
0(Cd) such
[ϑ](Cd) be the set of all F ∈ A′
(ω)(Cd) be the set of all F ∈ A(Cd) such
[ϑ]
10
that kFkA2
under these norms.
(ω)
is finite. It follows that these spaces are Hilbert spaces
If ϑ and ω are related to each others as
ϑ(α) = 1
α!ZRd
+
ω0(r)2rα dr! 1
2
,
α ∈ Nd
(1.26)
and
ω(z) = e
z2
2 ω0(z12, . . . ,zd2),
z ∈ Cd,
(1.27)
for some suitable weight ω0 on Rd
+, then the following multi-dimensional
version of [7, Theorem (4.1)] shows that A2
(ω)(Cd) with
equal norms. Here we identify entire functions with their power series
expansions at origin.
Theorem 1.6. Let eα be as in (1.12), α ∈ Nd, and let ω0 be a positive
measurable function on Rd
+. Also let ϑ and ω be weights on Nd and Cd,
respectively, related to each others by (1.26) and (1.27), and such that
[ϑ](Cd) = A2
rα
(α!)
1
2
. ϑ(α), α ∈ Nd,
[ϑ](Cd) = A2
(1.28)
holds for every r > 0. Then A2
norms.
(ω)(Cd) with equality in
We omit the proof of Theorem 1.6, since the result is an immediate
consequence of [12, Theorem 3.5].
In our situation, the involved weights should satisfy a split condition.
In one dimension, (1.26), (1.27) and (1.28) take the forms
ϑj(αj) =(cid:18) 1
ω0,j(r)2rαj dr(cid:19) 1
2
αj!ZR+
2 ω0,j(zj2),
zj 2
ωj(zj) = e
and
rαj
(αj!)
1
2
. ϑj(αj),
,
αj ∈ N,
zj ∈ C
(1.26)′
(1.27)′
r > 0, αj ∈ N.
(1.28)′
Lemma 1.7. For j = 1, . . . , d, let ω0,j be weights on R+, ωj be weights
on C and ϑj be weights on N such that (1.26)′ -- (1.28)′ hold, j =
j=1 ω0,j(zj), z ≡ (z1, . . . , zd) ∈ Cd. If ϑ
1, . . . , d, and set ω0(z) ≡ Qd
and ω are given by (1.26), and (1.27), then
ϑ(α) =
dYj=1
ϑj(αj),
α = (α1, . . . , αd) ∈ Nd
(1.29)
11
and
ω(z) =
dYj=1
ωj(zj),
z = (z1, . . . , zd) ∈ Cd.
(1.30)
Proof. By [12, Theorem 3.5] and its proof, it follows that ω0,j · rαj ∈
L1(R+) for all j ∈ {1, . . . , d} and αj ∈ N. Hence, Fubini's theorem
gives
ϑ(α) = 1
α!ZRd
+
ω0(r)2rα dr! 1
2
= 1
α!
dYj=1Z ∞
0
ω0,j(rj)2rαj
2
j drj! 1
dYj=1
=
ϑj(αj),
and (1.29) follows. The assertion (1.30) follows from the definitions. (cid:3)
1.4. A test function space introduced by Gröchenig. In this sec-
tion we recall some comparison results deduced in [12], between a test
function space, SC(Rd), introduced by Gröchenig in [5] to handle mod-
ulation spaces with elements in spaces of ultra-distributions.
The definition of SC(Rd) is given as follows. Here we notice that the
continuity of the map Vφ from S1/2(Rd) to S1/2(R2d) implies that its
dual, V ∗
1/2(Rd) is continuous.
φ from S ′
1/2(R2d) to S ′
Definition 1.8. Let φ(x) = π− d
sist of all f ∈ S ′(Rd) such that f = V ∗
E ′(Rd) and F ∈ E ′(Rd), respectively.
4 e− x2
2 . Then SC(Rd) and SG(Rd) con-
φ F , for some F ∈ L∞(Rd) ∩
It follows that f ∈ SC(Rd), if and only if
2 x−y2
f (x) = (2π)− d
F (y, η)e− 1
2ZZR2d
eihx,ηi dydη,
(1.31)
for some F ∈ L∞(Rd) ∩ E ′(Rd).
Remark 1.9. By the identity (Vφh, F ) = (h, V ∗
φ F ) and the fact that the
map (f, φ) 7→ Vφf is continuous from S (Rd) × S (Rd) to S (R2d), it
follows that f = V ∗
φ F is uniquely defined as an element in S ′(Rd) when
F ∈ S ′(R2d) (cf. [3]). In particular, the space SG(Rd) in Definition 1.8
is well-defined, and it is evident that SC(Rd) ⊆ SG(Rd).
The following is a restatement of [12, Lemma 4.9]. The result is
essential when deducing the characterizations of Pilipović spaces in
Section 3.
12
Lemma 1.10. Let F ∈ L∞(Cd) ∪ E ′(Cd). Then the Bargmann trans-
form of f = V ∗
φ F is given by ΠAF0, where
√2ξ)e
1
2 (x2+ξ2)e−ihx,ξi,
x, ξ ∈ Rd.
F0(x + iξ) = (2π3)
d
4 F (√2x,−
(1.32)
Moreover, the images of SC(Rd) and SG(Rd) under the Bargmann
transform are given by
{ ΠAF ; F ∈ L∞(Cd) ∩ E ′(Cd) } and { ΠAF ; F ∈ E ′(Cd) }, (1.33)
respectively.
The next result is a straight-forward consequence of [12, Theorem
4.10]. The proof is therefore omitted.
Proposition 1.11. It holds SC(Rd) = SG(Rd) = H♭1(Rd).
Due to the image properties of the Bargmann transform, the next
result is equivalent with the previous one.
Proposition 1.12. The sets in (1.33) are equal to A♭1(Cd).
In the next section we extend Propositions 1.11 and 1.12 by proving
that the conclusions in Proposition 1.12 hold for suitable smaller and
larger sets than those in (1.33). We also deduce similar identifications
for other Pilipović spaces and their Bargmann images.
2. Paley-Wiener properties for Bargmann-Pilipović spaces
In this section we consider spaces of compactly supported functions
s(Cd). We show that the images of
with interiors in As(Cd) or in A′
such functions under the reproducing kernel ΠA are equal to As(Cd),
for some other choice of s ≤ ♭1. In the first part we state the main
results given in Theorems 2.2 -- 2.8. They are straight-forward conse-
quences of Propositions 2.9 -- 2.12, which in some sense contain more
information. Thereafter we deduce results which are needed for their
proofs. Depending of the choice of s, there are several different situa-
tions for characterizing As(Cd). This gives rise to a quite large flora of
main results, where each one takes care of one situation.
In order to present the main results, we need the following definition.
Definition 2.1. Let t1, t2 ∈ Rd
+ be such that t1 ≤ t2. Then the function
χ ∈ L∞(Cd) is called positive, bounded and radial symmetric with
respect to t1 and t2, if the following conditions are fulfilled:
• χ ∈ L∞(Cd) ∩ E ′(Dt2(0)) is non-negative;
• χ(z1, . . . , zd) = χ0(z1, . . . ,zd) for some function χ0;
• χ ≥ c on Dt1(0) for some constant c > 0.
13
The set of positive, bounded and radial symmetric functions with re-
t1,t2(Cd), and R∞(Cd) is defined
spect to t1 and t2 is denoted by R∞
by
R∞(Cd) ≡ [t1≤t2∈Rd
+
t1,t2(Cd).
R∞
2.1. Main results. We begin with characterizing the largest spaces in
our investigations, which appears when s = ♭1, and then proceed with
spaces of decreasing order. First we recall that elements in As(Cd) and
A0,s(Cd) fulfill conditions of the form
F (z) . er1z1
when s = ♭σ and of the form
2σ
σ+1 +···+rdzd
2σ
σ+1 ,
(2.1)
(2.2)
F (z) . er1(loghz1i)
1
1−2s +···+rd(loghzdi)
1
1−2s ,
when s ∈ [0, 1
2).
Theorem 2.2. Let F ∈ A(Cd), σ = 1 and s > 1. Then the following
conditions are equivalent:
(1) F ∈ A♭1(Cd);
(2) The estimate (2.1) holds for some r ∈ Rd
+;
(3) For some r0 ∈ Rd
+ and every r ∈ Rd
+ with r0 ≤ r and every χ ∈
r0,r(Cd), there exists F0 ∈ A(Dr(0)) such that F = ΠA(F0· χ);
R∞
+ with r0 < r and some χ ∈
R∞
r0,r(Cd), there exists F0 ∈ A(Dr(0)) such that F = ΠA(F0· χ);
+ and every r ∈ Rd
(4) For some r0 ∈ Rd
(5) There exists F0 ∈ E ′(Cd) ∩ L∞(Cd) such that F = ΠAF0;
(6) There exists F0 ∈ E ′
s(Cd) such that F = ΠAF0.
Remark 2.3. Since
E ′(Cd) ∩ L∞(Cd) ⊆ E ′(Cd) ⊆ E ′
s(Cd),
Theorem 2.2 still holds true after E ′
Theorem 2.4. Let F ∈ A(Cd), σ = 1 and χ ∈ R∞(Cd). Then the
following is true:
s has been replaces by E ′ in (6).
(i) The following conditions are equivalent:
(1) F ∈ A0,♭1(Cd);
(2) The estimate (2.1) holds for every r ∈ Rd
+;
(3) There exists F0 ∈ A(Cd) such that F = ΠA(F0 · χ);
morphism.
(ii) The map F 7→ ΠA(F · χ) from A(Cd) to A0,♭1(Cd) is a homeo-
The next result deals with the case when s = ♭σ with σ ∈ ( 1
2, 1).
14
Theorem 2.5. Let F ∈ A(Cd), χ ∈ R∞(Cd), σ ∈ ( 1
2, 1) and let
Then the following is true:
σ0 =
σ
.
2σ − 1
(i) The following conditions are equivalent:
(1) F ∈ A♭σ(Cd) (F ∈ A0,♭σ(Cd));
(2) The estimate (2.1) holds for some (for every) r ∈ Rd
+;
(3) There exists F0 ∈ A′
(Cd) (F0 ∈ A′
0,♭σ0
♭σ0
(Cd)) such that
F = ΠA(F0 · χ);
(ii) The mappings F 7→ ΠA(F · χ) from A′
0,♭σ0
(Cd) to A♭σ (Cd) and
from A′
♭σ0
(Cd) to A0,♭σ(Cd) are homeomorphisms.
The next result deals with the case when s = ♭σ with σ = 1
2 .
Theorem 2.6. Let F ∈ A(Cd), σ = 1
following is true:
2 and χ ∈ R∞(Cd). Then the
(i) The following conditions are equivalent:
(1) F ∈ A♭σ(Cd) (F ∈ A0,♭σ(Cd));
(2) The estimate (2.1) holds for some (for every) r ∈ Rd
+;
(3) There exists F0 ∈ A′
0,1/2(Cd) (F0 ∈ A0,1/2(Cd)) such that
0,1/2(Cd) to A♭1/2(Cd) and
(ii) The mappings F 7→ ΠA(F · χ) from A′
from A0,1/2(Cd) to A0,♭1/2(Cd) are homeomorphisms.
F = ΠA(F0 · χ);
The next result deals with the case when s = ♭σ with σ ∈ (0, 1
2).
2) and let
Theorem 2.7. Let F ∈ A(Cd), χ ∈ R∞(Cd), σ ∈ (0, 1
Then the following is true:
σ0 =
σ
.
1 − 2σ
(i) The following conditions are equivalent:
(1) F ∈ A♭σ(Cd) (F ∈ A0,♭σ(Cd));
(2) The estimate (2.1) holds for some (for every) r ∈ Rd
+;
(3) There exists F0 ∈ A♭σ0
(Cd) (F0 ∈ A0,♭σ0
(Cd)) such that
F = ΠA(F0 · χ);
(ii) The mappings F 7→ ΠA(F · χ) from A♭σ0
from A0,♭σ0
(Cd) to A0,♭σ (Cd) are homeomorphisms.
(Cd) to A♭σ (Cd) and
In the next result we consider the case when s ∈ [0, 1
2) is real.
Theorem 2.8. Let F ∈ A(Cd), χ ∈ R∞(Cd), s ∈ [0, 1
following is true:
2). Then the
(i) The following conditions are equivalent:
15
(1) F ∈ As(Cd) (F ∈ A0,s(Cd));
(2) The estimate (2.2) holds for some (for every) r ∈ Rd
+;
(3) There exists F0 ∈ As(Cd) (F0 ∈ A0,s(Cd)) such that F =
(ii) The mappings F 7→ ΠA(F · χ) from As(Cd) to As(Cd) and from
ΠA(F0 · χ);
A0,s(Cd) to A0,s(Cd) are homeomorphisms.
The previous theorems are essentially consequences of Propositions
2.9 -- 2.11, where more detailed information about involved constants are
given.
Proposition 2.9. Let F ∈ A(Cd), σ = 1, s > 1 and r ∈ Rd
the following conditions are equivalent:
+. Then
+ such that r0 < r, (2.1) holds with r0 in place
of r;
(1) For some r0 ∈ Rd
(2) For some t1 ∈ Rd
t2 < r and every χ ∈ R∞
such that F = ΠA(F0 · χ);
(3) For some t1 ∈ Rd
t1 ≤ t2 < r, there exist χ ∈ R∞
that F = ΠA(F0 · χDr(0));
L∞(Cd) such that F = ΠAF0;
(4) For some r0 ∈ Rd
(5) For some r0 ∈ Rd
such that F = ΠAF0.
+ such that t1 < r, every t2 ∈ Rd
+ such that t1 < r and every t2 ∈ Rd
+ with t1 ≤
t1,t2(Cd), there exists F0 ∈ A(Dr(0))
+ with
t1,t2(Cd) and F0 ∈ A(Dr(0)) such
+ such that r0 < r, there exists F0 ∈ E ′(Dr(0))∩
+ such that r0 < r, there exists F0 ∈ E ′
s(Dr0(0))
Theorems 2.5 -- 2.7 essentially follows from the following proposition.
Proposition 2.10. Let τ > 1
let χ ∈ R∞
2 , r, t1, t2 ∈ Rd
t1,t2(Cd). Then the following is true:
+ be such that t1 ≤ t2, and
(1) Let F ∈ A(Cd) be such that
F (z) . er0,1z1
2τ +1 +···+r0,dzd
2
2τ +1
(2.3)
holds for some r0 ∈ Rd
+ such that r0 < r. Then for some r0 ∈
+ such that r0 < r, there exists F0 ∈ A(Cd) such that F =
Rd
ΠA(F0 · χ) and
2
2
16
where
2
2τ −1 ,
2τ −1 +···+R0,dzd
F0(z) . eR0,1z1
2 (cid:18) 2r0
2τ − 1
R0 =
2τ −1
2τ + 1(cid:19) 2τ +1
− 4
t
1
2τ −1
;
(2.4)
(2.5)
(2) Let r0, R0 ∈ Rd
+ be such that r0 < r and
R0 =
2τ − 1
2 (cid:18) 2r0
2τ + 1(cid:19) 2τ +1
2τ −1
− 4
t
2
2τ −1
,
(2.6)
F0 ∈ A(Cd) be such that (2.4) holds and let F = ΠA(F0 · χ).
Then F ∈ A(Cd) and satisfies (2.4) for some r0 ∈ Rd
+ such that
r0 < r.
Theorem 2.8 follows from the following two propositions, where the
first one concerns the case when s > 0 and the second one make a
more detailed explanation of the case s = 0, i. e. the case of analytic
polynomials.
Proposition 2.11. Let s ∈ (0, 1
the following is true:
+, and let χ ∈ R∞(Cd). Then
2), r ∈ Rd
(1) Suppose F ∈ A(Cd) satisfies
F (z) . er0,1(loghz1i)
1
for some r0 ∈ Rd
such that F = ΠA(F0 · χ) and
1−2s +···+r0,d(loghzdi)
(2.7)
+ such that r0 < r. Then there is an F0 ∈ A(Cd)
1−2s
1
F0(z) . er0,1(loghz1i)
1
1−2s +···+r0,d(loghzdi)
1
1−2s
(2.8)
+ such that r0 < r;
(2) Suppose F0 ∈ A(Cd) satisfies (2.8) for some r0 ∈ Rd
for some r0 ∈ Rd
+ such that
r0 < r, and let F = ΠA(F0 · χ). Then F ∈ A(Cd) and satisfies
(2.7) for some r0 ∈ Rd
+ such that r0 < r.
Proposition 2.12. Let χ ∈ R∞(Cd) and let N ≥ 0 be an integer.
Then the following is true:
(1) Suppose F ∈ A(Cd) is given by
c(F, α)zα,
(2.9)
F (z) = Xα≤N
F0(z) = Xα≤N
where {c(F, α)}α≤N ⊆ C. Then there is an F0 ∈ A(Cd) such
that F = ΠA(F0 · χ) and
c(F0, α)zα,
(2.10)
where {c(F0, α)}α≤N ⊆ C and satisfies c(F0, α) = 0 when
c(F, α) = 0;
(2) Suppose F0 ∈ A(Cd) satisfies (2.10) for some {c(F0, α)}α≤N ⊆
C, and let F = ΠA(F0 · χ). Then F ∈ A(Cd) and satisfies
(2.9) for some {c(F, α)}α≤N ⊆ C such that c(F, α) = 0 when
c(F0, α) = 0.
17
2.2. Preparing results and their proofs. For the proofs of Proposi-
tions 2.9 -- 2.11 and thereby of Theorems 2.2 -- 2.8 we need some prepara-
tory results. Because the proof of Proposition 2.9 needs some room,
we put parts of the statement in the following separate proposition. At
the same time we slightly refine some parts concerning the image of
compactly supported elements in L∞ under the map ΠA.
Proposition 2.13. Let s > 1, r0, r ∈ Rd
suppose that either
+ be such that r0 < r and
s(Dr0(0)), F0 ∈ E ′
F0 ∈ E ′
Then F = ΠAF0 ∈ A(Cd) and satisfies
0,s(Dr0(0)), F0 ∈ E ′(Dr0(0)) or F0 ∈ L∞(Dr(0)).
F (z) . er1z1+···+rdzd.
Proof. By the inclusions
E ֒→ E ′
0,s
0,s+ε ֒→ E ′
s ֒→ E ′
Let r2 = r. First suppose that F0 ∈ E ′
when ε > 0, it suffices to consider the case when F0 ∈ E ′
F0 ∈ L∞(Dr(0)) hold.
s(Dr0(0)) holds, choose r1 ∈ Rd
+
such that r0 < r1 < r2, Ψ(y, η) = e−(y2+η2) and let Φz(y, η) = e(z,y+iη).
By identifying Cd with R2d and using the fact that F0 ∈ E ′
s(Dr0(0)) we
obtain
s(Dr0(0)) or
ΠAF0(z) = π−dhF0, ΦzΨi . sup
α∈N2d kDα(ΦzΨ)kL∞(Dr1 (0))
hα
1 α!s
! (2.11)
for every h1 > 0. We also have Ψ ∈ E1/2(R2d) ֒→ E0,s(R2d), which
implies
kDαΨkL∞(Dr1 (0)) . hα
2 α!s
for every h2 > 0. Furthermore,
DαΦz(y, η) = mα(z)e(z,y+iη) ≤ mα(z)er1,1z1+···+r1,dzd,
where
y+iη ∈ Dr1(0),
mα(z) =
dYj=1
zαj +αd+j
j
,
z ∈ Cd, α ∈ N2d.
18
By choosing h1 = 4 and h2 = 1 above, and letting y + iη ∈ Dr1(0),
Leibnitz rule gives
4−αα!−sDα(ΦzΨ)(y, η)
≤ 4−αα!−sXγ≤α(cid:18)α
. 4−αα!−s Xγ≤α(cid:18)α
. 4−α Xγ≤α(cid:18)α
γ≤α dYj=1
γ(cid:19)DγΦz(y, η) Dα−γΨ(y, η)
γ(cid:19)mγ(z)(α − γ)!s! er1,1z1+···+r1,dzd
γ(cid:19)mγ(z)γ!−s! er1,1z1+···+r1,dzd
γd+j!s ! er1,1z1+···+r1,dzd.
· zjγd+j
k(cid:1) ≤ 2k when n, k are non-negative
sums are bounded by 2α, and that(cid:0)n
γj! !s
γj!s = (zj
α∈N2d kDα(ΦzΨ)kL∞(Dr1 (0))
! . e2s(z1
In the last inequality we have used that the number of terms in the
integers such that k ≤ n.
By combining (2.12) with the estimate
s )er1,1z1+···+r1,dzd
zjγj
γj!s
1
s ,
≤ eszj
≤ sup
we get
sup
1
s +···+zd
1
zjγj
1
s )γj
(2.12)
hα
1 α!s
In the last inequality we have used the fact that r1 < r2 and s > 1.
From the latter estimate and (2.11) we obtain
. er2,1z1+···+r2,dzd.
ΠAF0(z) . er2,1z1+···+r2,dzd,
s(Dr0(0)).
Suppose instead that F0 ∈ L∞(Q) holds, where Q = Dr2(0) ⊆ Cd,
and the result follows when F0 ∈ E ′
and let Qj = Dr2,j (0) ⊆ C. Then
ΠAF0(z) .ZQ F0(w)e(z,w) dλ(w)
dYj=1 ZQj
≤ kF0kL∞
ezj wj dλ(wj)! . er2,1z1+···+r2,dzd,
and the result follows in this case as well.
(cid:3)
19
In the next lemma we give options on compactly supported functions
which are mapped on the basic monomials, eα by the operator ΠA.
Lemma 2.14. Let t1, t2 ∈ Rd
and let χ0 be such that χ0(z1, . . . ,zd) = χ(z1, . . . , zd). If
+ be such that t1 ≤ t2, χ ∈ R∞
t1,t2(Cd),
Fα,χ(z) = ςαzαχ(z),
with
ςα = 2−dα!
1
2 Z∆t2
χ0(u)e−u2
u2αu1 · · · ud du!−1
then the following is true:
, ∆t = { u ∈ Rd
+ ; u ≤ t},
(2.13)
(1) ΠAFα,χ = eα;
(2) for some constant C > 0 which only depends on kχkL∞, c in
Definition 2.1 and the dimension d, it holds
C −1 dYj=1
2,j (αj + 1)! t−2α
t−2
2 α!
1
2 ≤ ςα ≤ Cet12 dYj=1
1,j (αj + 1)! t−2α
t−2
1 α!
1
2
(2.14)
Proof. By using polar coordinates in each complex variable when inte-
grating we get
wαχ(w)e(z,w)−w2
dλ(w)
Iα(u, z)e−u2
uαχ0(u)u1 · · · ud du,
(2.15)
ezjuj e−iθj! dθ =
dYj=1
Iαj (uj, zj)
(2.16)
(ΠAFα,χ)(z) = π−dςαZCd
= π−dςαZ∆t2
eihα,θi dYj=1
where
with
Iα(u, z) =Z[0,2π)d
Iαj (uj, zj) =Z 2π
Iαj (uj, zj) =Z 2π
0
0
By Taylor expansions we get
eiαj θj ezj uje−iθj dθj.
j e−ikθj
k!
! dθj
zk
j uk
eiαj θj ∞Xk=0
∞Xk=0 (cid:18)Z 2π
=
0
20
ei(αj −k)θj dθj(cid:19) zk
j uk
j
k! ! =
j uαj
2πzαj
αj!
j
By inserting this into (2.15) and (2.16) we get
(ΠAFα,χ)(z) = π−dςαZ∆t2
2Z∆t2
= 2dςαα!− 1
and (1) follows.
(2π)d uαzα
α!
e−u2
uαχ0(u)u1 · · · ud du
e−u2
u2αχ0(u)u1 · · · ud du! eα(z) = eα(z)
Since χ is non-negative and fullfils χ ≥ c on Dt1(0) we get
ςα . et12
α!
= et12
α!
1
2
1
u2αu1 · · · u2 du!−1
2 Z∆t1
dYj=1 t2αj +2
2αj + 2!−1
≍ et12
1,j
α!
1
2 dYj=1
1,j (αj + 1)),! t−2α
1
(t−2
which gives the right inequality in (2.14). By the support properties of
χ we also have
ςα & α!
1
2 Z∆t2
u2αu1 · · · u2 du!−1
,
and the left inequality in (2.14) follows by similar arguments, and (2)
follows.
(cid:3)
The next lemma shows that we may estimate entire functions by
different Lebesgue norms. We omit the proof, since the result follows
from [11, Theorem 3.2].
Lemma 2.15. Suppose s, τ ∈ R and r, r0 ∈ Rd
+ are such that
1
τ > −
2
Let p, q ∈ [1,∞], F ∈ A(Cd) and set
s <
1
2
,
and r0 < r.
M1,r(z) = r1z1
2τ +1 + · · · + rdzd
2
2
2τ +1
and
Then
M2,r(z) = r1(loghz1i)
1
1−2s + · · · + rd(loghzdi)
1
1−2s .
kF · e−Mj,rkLp(Cd) . kF · e−Mj,r0kLq(Cd).
The next lemma relates Lebesgue estimates of entire functions with
estimates on corresponding Taylor coefficients. Here we let the Gamma
21
function on Rd
+ be defined by
Γd(x1, . . . , xd) =
Γ(xj),
dYj=1
1
where Γ is the Gamma function on R+.
Lemma 2.16. Let τ > − 1
2, M1,r be the same as in Lemma 2.15, ω(z) =
2 z2−M1,r(z), j = 1, 2, z ∈ Cd, and let α0 = (1, . . . , 1) ∈ Nd. Also let
e
(cid:19)(cid:19) 1
ϑr(α) =(cid:18)(2τ + 1)(2r)−(2τ +1)(α+α0)(cid:18) Γ ((2τ + 1)(α + α0))
.
2 Xα∈Nd c(F, α)ϑr(α)2! 1
If F ∈ A(Cd) is given by (1.13), then
kF · e−M1,rkL2(Cd) = π
α!
,
d
2
2
[ϑr](Cd) = A2
(ω)(Cd) with equality in norms.
and A2
Proof. Since
e−M1,r(z) =
e−rjzj
2
2τ +1 ,
dYj=1
Lemma 1.7 shows that we may assume that d = 1, giving that r = r1
and α0 = 1.
In view of Theorem 1.6 we have A2
in norms, when
[ϑ](Cd) = A2
(ω1)(Cd) with equality
ϑ(α) =(cid:18) 1
α!Z ∞
0
e−2rt
2
1
2τ +1 tα dt(cid:19) 1
,
provided it is verified that (1.28) holds for every r > 0. By taking
u = 2rt
2τ +1 as new variables of integration we obtain
1
ϑ(α) =(cid:18)(2τ + 1)(2r)−(2τ +1)(α+1) 1
α!Z ∞
0
=(cid:18)(2τ + 1)(2r)−(2τ +1)(α+1)(cid:18)Γ ((2τ + 1)(α + 1))
α!
2
e−uuα(2τ +1)+2τ du(cid:19) 1
(cid:19)(cid:19) 1
2
.
Since Γ((2τ + 1)(α + 1)) & rα for every r > 0, by Stirling's formula,
(1.28) holds for every r > 0. This implies that ϑ = ϑ1,r, and the result
follows.
(cid:3)
We also need the following version of Stirling's formula.
Lemma 2.17. Let α ≥ 0 be an integer and let τ > − 1
2 . Then
Γ ((2τ + 1)(α + 1))
α!
≍ (2τ + 1)(2τ +1)·α (α + 1)τ · α!2τ .
22
(2.17)
Lemma 2.17 follows by repeated applications of Stirling's formula
and the standard limit
lim
t→∞(cid:16)1 +
x
t(cid:17)t
= ex
for every x ∈ R. In order to be self-contained we present the arguments.
Proof. The result is obviously true for α = 0. For α ≥ 1 we have
ατ ≍ (α + 1)τ . A combination of the latter relations and Stirling's
formula gives
2
α2τ (α+1)
Γ ((2τ + 1)(α + 1))
α!
≍
((2τ + 1)α + 2τ )(2τ +1)·α+2τ + 1
2
eα+2τ (α+1)
eα
αα+ 1
2
·
=
2τ
e2τ (α+1)
(2τ + 1)(2τ +1)·α+2τ + 1
2(cid:18)1 +
≍ (2τ + 1)(2τ +1)·α α2τ (α+1)
(2τ + 1)α(cid:19)(2τ +1)·α+2τ + 1
(2τ + 1)α(cid:19)(2τ +1)α
·(cid:18)1 +
≍ (2τ + 1)(2τ +1)·α α2τ (α+1)
e2τ α ≍ (2τ + 1)(2τ +1)·α (α + 1)τ α!2τ .
α(cid:17)x2α
x1, x2 ∈ R+,
(cid:16)1 +
e2τ α
x1
2τ
,
In the forth relation we have used the fact that
(2.18)
increases with α and has the limit ex1x2 as α tends to infinity. This
gives the result.
(cid:3)
Proposition 2.9 essentially follows from the following lemma.
Lemma 2.18. Let τ > − 1
A(Cd). Then the following is true:
2 , r0, r ∈ Rd
+ be such that r0 < r and F ∈
(1) If
then
(2) If
then
2
2τ + 1(cid:19) 2τ +1
c(F, α) .(cid:18) 2r0
F (z) . er1z1
2
2τ +1 +···+rdzd
·α
α!−τ ,
2
2τ +1 ;
2
2τ +1 +···+r0,dzd
2
2τ +1
F (z) . er0,1z1
c(F, α) .(cid:18) 2r
2
2τ + 1(cid:19) 2τ +1
23
·α
α!−τ
(2.19)
(2.20)
(2.21)
(2.22)
Proof. Let ϑr be the same as in Lemma 2.16. First we prove (1). Sup-
pose that (2.19) holds, let r1 ∈ Rd
+ be such that r0 < r1 < r and let
α0 = (1, . . . , 1) ∈ Nd. Also let M1,r be the same as in Lemma 2.15.
Then Lemmata 2.15 and 2.16 give
kF · e−M1,rkL∞(Cd) . kF · e−M1,r1kL2(Cd)
2
2
1
2
≍ Xα∈Nd c(F, α)ϑr1(α)2! 1
=Xα∈Nd(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
r1(cid:19)(α+α0) 2τ +1
c(F, α)ϑr0(α)(cid:18)r0
2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
α∈Nd(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dYj=1
α∈Nd c(F, α)(cid:18)2τ + 1
2r0 (cid:19) 2τ +1
2!(cid:18) r0
Xα∈Nd dYj=1
r1(cid:19)(2τ +1)·(α+α0)!
c(F, α)ϑr0(α)
(αj + 1)− τ
≍ sup
(αj + 1)
. sup
2
τ
·α
α!τ! .
Here the second inequality follows from the fact that
is convergent since r0 < r1, and the fifth relation follows from Lemma
2.17. This implies that (2.20) holds and (1) follows.
Next assume that (2.21) holds. By Lemmata 2.15 and 2.16 we get
kF · e−M1,r0kL∞(Cd) & kF · e−M1,rkL2(Cd)
2
α∈Nd
(c(F, α) ϑr(α))
≥ sup
= Xα∈Nd c(F, α)ϑr(α)2! 1
α∈Nd c(F, α)(cid:18)2τ + 1
2r (cid:19) 2τ +1
α∈Nd c(F, α)(cid:18)2τ + 1
2 α dYj=1
≥ sup
2 α!τ!
2r (cid:19) 2τ +1
(αj + 1)
& sup
τ
2 α
α!τ! ,
where the third inequality follows from Lemma 2.17. This gives (2). (cid:3)
Proposition 2.11 mainly follows from the following result.
+ be such that r0 < r, s ∈ (0, 1
Lemma 2.19. Let r, r0 ∈ Rd
F ∈ A(Cd). Then the following is true:
24
2) and let
(1) if (2.7) holds, then
1
2s +···+Rdαd
c(F, α) . e−(R1α1
(2) if R0 ∈ Rd
c(F, α) . e−(R0,1α1
+ is given by
1
2s +···+R0,dαd
1
2s
;
r (cid:19) 1−2s
2s ), R = s(cid:18)1 − 2s
2s ), R0 = s(cid:18)1 − 2s
r0 (cid:19) 1−2s
2s
1
(2.23)
,
(2.24)
then (2.7) holds with r in place of r0.
Proof. Let r1, r2 ∈ Rd
+ be such that r1 < r < r2. By Lemma 1.7 we
may assume that d = 1, and by Lemma 2.15 the result follows if we
prove
sup
α∈N(cid:16)c(F, α)eR2·α
1
2s 2(cid:17) .ZC F (z)e−r(loghzi)
1
1−2s 2 dλ(z)
α∈N(cid:16)c(F, α)eR1·α
. sup
1
2s 2(cid:17) ,
(2.25)
where R1, R2 ∈ R+ satisfy
By Theorem 1.6 we have
Rj =(cid:18)1 − 2s
1
,
2s
j = 1, 2.
rj (cid:19) 1−2s
1−2s 2 dλ(z) ≍Xα∈Nc(F, α)ϑr(α)2,
2α!Z ∞
1−2s tα dt(cid:19) 1
e−r(loghti)
.
0
1
2
ZC F (z)e−r(loghzi)
ϑr(α) =(cid:18) π
where
Let
θ =
1
1 − 2s
and gr,α(t) = e−r(log t)θ
tα.
In order to prove (2.25), we need to prove
eR2α
1
2s . ϑr(α) . eR1α
1
2s ,
(2.26)
which shall be reached by modifying the proof of (15) in [4].
We have
ϑr(α)2 .Z ∞
e
e(r1−r)(log t)θ
gr1,α(t) dt
. sup
t≥e
(gr1,α(t))Z ∞
e
25
e(r1−r)(log t)θ
dt ≍ sup
t≥e
(gr1,α(t)).
By straight-forward computations it follows that gr,α(t) attains its
global maximum for
tr,α = exp(cid:18)(cid:16) α
2s (cid:19) ,
θr(cid:17) 1−2s
(2.27)
and that
gr1,α(tr1,α) = exp(cid:16)r
1− 1
2s
1
θ− 1
2s (θ − 1)α
1
2s(cid:17) = e2R1·α
1
2s ,
and the second inequality in (2.26) follows.
In order to prove the first inequality in (2.26), we claim that for some
c which is independent of α we have
(cid:18)1 −
c
tr2,α(cid:19)α
& e(r−r2)(log tr2,α)θ
.
(2.28)
In fact, by (2.27) and Taylor's formula, (2.28) follows if we prove
1 −
αc
tr2,α
& e(r−r2)((1−2s)α/r2)
1
2s
(2.29)
for some c > 0 which is independent of α. In view of (2.27) it follows
that tr2,α tends to infinity as α turns to infinity, which implies that the
left-hand side of (2.29) has the limit one and the right-hand side has
the limit zero as α tends to infinity. Hence, (2.28) holds.
By (2.28) and the fact that 1
2s > 1 we get
ϑr(α)2 &
1
α!Z tr2,α
tr2,α−c
e−r(log t)θ
tα dt &
1
α!
e−r(log tr2,α)θ
(tr2,α − c)α
&
1
α!
e−r2(log tr2,α)θ
tα
r2,α =
e2R2·α
1
2s & e2R3·α
1
2s .
1
α!
This gives the result.
(cid:3)
2.3. Proofs of main results. Next we prove Proposition 2.9 and
thereby Theorem 2.2.
Proof of Proposition 2.9. It is clear that (2) ⇒ (3) ⇒ (4) ⇒ (5). By
Proposition 2.13 it follows that (5) implies (1). We need to prove that
(1) implies (2).
Suppose (1) holds and let r4 = r. Choose r1, r2, r3 ∈ Rd
+ such that
r2,r3(Cd), and let Fα,χ be
r0 < r1 < r2 ≤ r3 < r4 and r1r4 < r2
as in Lemma 2.14 with r2 and r3 in place of r1 and r2. If τ = 1, then
Lemma 2.18 (2) gives
2, χ ∈ R∞
Let
1 α!− 1
2 .
c(F, α) . rα
F0(z) = Xα∈Nd
26
c(F, α)ςαzα.
(2.30)
(2.31)
We claim that the series in (2.31) is uniformly convergent with respect
to z in DR(0).
In fact, if zj ≤ r4,j, then (2.14) gives
rα
c(F, α)ςαzα . hαidr−2α
1 rα
2
4 = hαidρα,
+ satisfies ρj = r1,j r4,j
where ρ ∈ Rd
gent, Weierstrass theorem shows that (2.31) is uniformly convergent
and defines an analytic function in Dr4,j (0). Hence, F0 ∈ A(Dr4,j (0)).
Furthermore, by the support of χ we get
< 1. SincePα∈Ndhαidρα is conver-
r2
2,j
ΠA(F0 · χ) = ΠA Xα∈Nd
= Xα∈Nd
c(F, α)ςαzα · χ!
c(F, α)ςαΠA(zα · χ) = Xα∈Nd
Hence (2) holds, and the proof is complete.
c(F, α)eα = F.
(2.32)
(cid:3)
For future references we observe that if ςα and χ are the same as in
Lemma 2.14, then (2.31) shows that the relationship between c(F, α)
and c(F0, α) is given by
c(F0, α) = c(F, α)ςαα!
1
2 .
(2.33)
Proof of Theorem 2.4. The equivalence between (1) and (2) is clear. It
is also obvious that (3) implies (2) in view of Proposition 2.9, and that
(3) implies (4). We shall prove that (1) implies (3) and (4) implies (1).
+. Let r, R ∈
r,R(Cd), Fα,χ be as in Lemma 2.14 and let F0 be given by
Suppose (1) holds. Then (2.30) holds for every r1 ∈ Rd
Rd
+, χ ∈ R∞
(2.31). By (2.30) we have
for every r1 ∈ Rd
+, giving that
c(F, α)ςα . hαidr−2αrα
1
c(F, α)ςα . rα
0
for every r0 ∈ Rd
+. This implies that the series in (2.31) is locally
uniformly convergent with respect to z and defines an entire function on
Cd. Hence F0 ∈ A(Cd). Moreover, by (2.32) it follows that ΠA(F0χ) =
F , and we have proved that (1) implies (3).
Next suppose that (4) holds. Then
c(F0, α) . rαα!
1
2
for every r ∈ Rd
+. By (2.14) and (2.33) we get
2 . rας −1
α α!− 1
c(F, α) = c(F0, α)ς −1
α . rαR2αα!− 1
2 .
Since r ∈ Rd
+ can be chosen arbitrarily small we get
c(F, α) . rα
27
0 α!− 1
2
for every r0 ∈ Rd
proved that (4) implies (1), and the result follows.
+. This implies that F ∈ A0,♭1(Cd). That is, we have
(cid:3)
Next we prove Proposition 2.10.
Proof of Proposition 2.10. Suppose (2.3) holds, and let r1, r2 ∈ Rd
such that r0 < r1 < r2 < r. Then Lemma 2.18 gives
+ be
2
2τ + 1(cid:19) 2τ +1
·α
α!−τ .
By Lemma 2.14 and (2.33) we get
c(F, α) .(cid:18) 2r1
2τ + 1(cid:19) 2τ +1
2τ + 1(cid:19) 2τ +1
·α
·α
2
2
c(F0, α) .(cid:18) 2r1
.(cid:18) 2r2
when
α!−τhαidα!t−2α
1
α!−(τ −1)t−2α
1 =(cid:18) 2R0
2τ − 1(cid:19) 2τ −1
2
·α
α!−(τ −1),
R0 =
This proves (1).
2τ − 1
2 (cid:18) 2R2
2τ + 1(cid:19) 2τ +1
2τ −1
− 4
t
1
2τ −1
.
Suppose instead that r0, R0, r1, r2 ∈ Rd
+ are such that (2.6) hold
and r0 < r1 < r2 < r, and that F0 ∈ A(Cd) satisfies (2.4). Also let
F = ΠA(F0 · χ). Then
By combining the latter estimate with Lemma 2.18 we get
2
2τ + 1(cid:19) 2τ +1
t−2
2 .
2
(cid:18) 2R0
2τ − 1(cid:19) 2τ −1
c(F0, α) .(cid:18) 2r1
c(F, α) .(cid:18) 2r1
<(cid:18) 2r1
2τ + 1(cid:19) 2τ +1
2τ + 1(cid:19) 2τ +1
·α
2
2
t−2α
2 α!1−τ .
·α
α!−τ
Hence Lemma 2.14 and (2.33) give
and Lemma 2.18 again implies that (2.3) holds with r2 in place of r0.
This gives the result.
(cid:3)
Proof of Theorems 2.5 -- 2.7. First suppose σ ∈ ( 1
and τ = 1
2, 1), and let σ0 = σ
2σ−1
2σ . Then
2σ
=
2
2τ + 1
and
σ + 1
2σ0
σ0 − 1
=
2
.
2τ − 1
Theorem 2.5 now follows from these observations and Proposition 2.10
in the case τ ∈ ( 1
2, 1).
28
In the same way, Theorem 2.6 follows by choosing τ = 1 in Proposi-
tion 2.10.
Finally, suppose σ ∈ (0, 1
2 ), and let σ0 = σ
2
2σ
2σ0
=
σ + 1
2τ + 1
and
=
σ0 + 1
1−2σ and τ = 1
2σ . Then
2
,
2τ − 1
and Theorem 2.5 follows from these observations and Proposition 2.10
in the case τ > 1.
(cid:3)
Next we prove Propositions 2.11 and 2.12 and thereby Theorem 2.8.
Proof of Propositions 2.11 and 2.12. Let r, rj, Rj ∈ Rd
be such that
+, j = 0, 1, 2, 3,
r0 < r1 < r2 < r3 < r
and Rj = s(cid:18)1 − 2s
rj (cid:19) 1−2s
2s
.
First suppose that F ∈ A(Cd) satisfies (2.7) and let F0 be the formal
power series expansion with coefficients given by (2.33). Then F =
ΠA(F0 · χ).
By Lemma 2.19 we get
c(F, α) . e−(R1,1α1
Hence Lemma 2.14 and (2.33) give
1
2s +···+R1,dαd
1
2s ).
c(F0, α) . α!t−2α
1
e−(R1,1α1
1
2s +···+R1,dαd
1
2s )
. e−(R2,1α1
1
2s +···+R2,dαd
1
2s ).
In the last inequality we have used the fact that s < 1
that R2 < R1 and
2 , which implies
t−2α
1 α! . e(R1,1−R2,1)α1
1
2s +···+(R1,d−R2,d)αd
1
2s .
By applying Lemma 2.19 again it follows that F0 satisfies (2.8) with
r3 in place of r0. This gives (1).
Suppose instead that F0 ∈ A(Cd) satisfies (2.8) and let F = ΠA(F0 ·
χ). Then Lemma 2.19 gives
c(F0, α) . e−(R1,1α1
1
2s +···+R1,dαd
1
2s ),
and it follows from Lemma 2.14 and (2.33) that
c(F, α) . α!−1t2α
2 e−(R1,1α1
1
2s +···+R1,dαd
1
2s )
. e−(R1,1α1
1
2s +···+R1,dαd
1
2s ).
By applying Lemma 2.19 again we deduce (2.7) with r2 in place of r0,
and Proposition 2.11 follows.
Proposition 2.12 is a straight-forward consequence of (2.33). The
(cid:3)
details are left for the reader.
29
3. Characterizations of Pilipović spaces
In this section we combine Lemma 1.10 with Theorems 2.2 -- 2.8 to
get characterizations of Pilipović spaces.
We begin with the following characterization of H♭1. The result is
a straight-forward combination of Lemma 1.10 and Theorem 2.2. The
details are left for the reader.
Proposition 3.1. Let φ be as in (1.7), r ∈ Rd
+, χr be the characteristic
function for Dr(0) and let s > 0. Then the following conditions are
equivalent:
(1) f ∈ H♭1(Rd);
(2) f = V ∗
(3) f = V ∗
(4) f = V ∗
s(R2d);
φ F for some F ∈ E ′
φ F for some F ∈ E ′(R2d) ∩ L∞(R2d);
φ F for some F ∈ E ′(R2d) ∩ L∞(R2d) which satisfies
F (x, ξ) = F0(x − iξ)e−ihx,ξiχ(x, ξ)
+, χ = χr and F0 ∈ A(Dr(0)).
for some r ∈ Rd
(3.1)
Remark 3.2. It is clear that χ in Proposition 3.1 can be chosen as any
χ ∈ R∞(Cd) with suitable support properties.
2 (x2+ξ2) which is absent
Remark 3.3. In (1.32) there is a factor e
in (3.1). We notice that this factor is not needed in (3.1) because
t1,t2(Cd) is invariant under multiplications of such functions.
R∞
The next result follows from Lemma 1.10 and Theorems 2.4 -- 2.8. The
1
details are left for the reader.
Proposition 3.4. Let φ be as in (1.7) and χ ∈ R∞(Cd). Then the
following conditions are equivalent:
(1) f ∈ H0,♭1(Rd);
(2) f = V ∗
for some F0 ∈ A(Cd).
φ F for some F ∈ E ′(R2d)∩ L∞(R2d) which satisfies (3.1)
Proposition 3.5. Let φ be as in (1.7), χ ∈ R∞(Cd), σ ∈ ( 1
let
2, 1) and
σ0 =
σ
.
2σ − 1
Then the following conditions are equivalent:
(1) f ∈ H♭σ (Rd) (f ∈ H0,♭σ (Rd));
(2) f = V ∗
φ F for some F ∈ E ′(R2d)∩ L∞(R2d) which satisfies (3.1)
for some F0 ∈ A′
0,♭σ0
(Cd) (F0 ∈ A′
♭σ0
(Cd)).
Proposition 3.6. Let φ be as in (1.7), χ ∈ R∞(Cd) and let σ = 1
2.
Then the following conditions are equivalent:
(1) f ∈ H♭σ (Rd) (f ∈ H0,♭σ (Rd));
30
(2) f = V ∗
φ F for some F ∈ E ′(R2d)∩ L∞(R2d) which satisfies (3.1)
for some F0 ∈ A′
0,1/2(Cd) (F0 ∈ A0,1/2(Cd)).
Proposition 3.7. Let φ be as in (1.7), χ ∈ R∞(Cd), σ ∈ (0, 1
let
2) and
Then the following conditions are equivalent:
σ
σ0 =
.
1 − 2σ
(1) f ∈ H♭σ (Rd) (f ∈ H0,♭σ (Rd));
(2) f = V ∗
φ F for some F ∈ E ′(R2d)∩ L∞(R2d) which satisfies (3.1)
for some F0 ∈ A♭σ0
(Cd) (F0 ∈ A0,♭σ0
(Cd)).
Proposition 3.8. Let φ be as in (1.7), χ ∈ R∞(Cd) and let s ∈ (0, 1
2).
Then the following conditions are equivalent:
(1) f ∈ Hs(Rd) (f ∈ Hs(Rd));
(2) f = V ∗
φ F for some F ∈ E ′(R2d)∩ L∞(R2d) which satisfies (3.1)
for some F0 ∈ As(Cd) (F0 ∈ A0,s(Cd)).
References
[1] V. Bargmann On a Hilbert space of analytic functions and an associated inte-
gral transform, Comm. Pure Appl. Math., 14 (1961), 187 -- 214.
[2] V. Bargmann On a Hilbert space of analytic functions and an associated in-
tegral transform. Part II. A family of related function spaces. Application to
distribution theory., Comm. Pure Appl. Math., 20 (1967), 1 -- 101.
[3] E. Cordero, S. Pilipović, L. Rodino, N. Teofanov Quasianalytic Gelfand-Shilov
spaces with applications to localization operators, Rocky Mt. J. Math. 40
(2010), 1123-1147.
[4] C. Fernandez, A. Galbis, J. Toft The Bargmann transform and powers of har-
monic oscillator on Gelfand-Shilov subspaces, RACSAM 111 (2017), 1 -- 13.
[5] K. H. Gröchenig Foundations of Time-Frequency Analysis, Birkhäuser, Boston,
2001.
[6] L. Hörmander The Analysis of Linear Partial Differential Operators, vol I -- III,
Springer-Verlag, Berlin Heidelberg NewYork Tokyo, 1983, 1985.
[7] A.M.E.M. Janssen, S.J.L. Eijndhoven Spaces of type W, growth of Hermite
coefficients, Wigner distribution, and Bargmann transform, J. Math. Anal.
Appl. 152 (1990), 368 -- 390.
[8] S. Pilipović Generalization of Zemanian spaces of generalized functions which
have orthonormal series expansions, SIAM J. Math. Anal. 17 (1986), 477-484.
[9] S. Pilipović Tempered ultradistributions, Boll. U.M.I. 7 (1988), 235 -- 251.
[10] M. Reed, B. Simon Methods of modern mathematical physics, Academic Press,
London New York, 1979.
[11] J. Toft The Bargmann transform on modulation and Gelfand-Shilov spaces,
with applications to Toeplitz and pseudo-differential operators, J. Pseudo-
Differ. Oper. Appl. 3 (2012), 145 -- 227.
[12] J. Toft Images of function and distribution spaces under the Bargmann trans-
form, J. Pseudo-Differ. Oper. Appl. 8 (2017), 83 -- 139.
31
Department of Mathematics, Linnaeus University, Växjö, Sweden
E-mail address: [email protected]
Department of Mathematics, University of Regensburg, Regensburg,
Germany
E-mail address: [email protected]
Department of Mathematics, Linnaeus University, Växjö, Sweden
E-mail address: [email protected]
32
|
1704.06950 | 1 | 1704 | 2017-04-23T17:09:39 | Self-Adjoint Operators in Extended Hilbert Spaces $H\oplus W$: An Application of the General GKN-EM Theorem | [
"math.FA"
] | We construct self-adjoint operators in the direct sum of a complex Hilbert space $H$ and a finite dimensional complex inner product space $W$. The operator theory developed in this paper for the Hilbert space $H\oplus W$ is originally motivated by some fourth-order differential operators, studied by Everitt and others, having orthogonal polynomial eigenfunctions. Generated by a closed symmetric operator $T_{0}$ in $H$ with equal and finite deficiency indices and its adjoint $T_{1}$, we define \textit{families} of minimal operators $\{\widehat{T}_{0}\}$ and maximal operators $\{\widehat{T}_{1}\}$ in the extended space $H\oplus W$ and establish, using a recent theory of complex symplectic geometry, developed by Everitt and Markus, a characterization of self-adjoint extensions of $\{\widehat{T}_{0}\}$ when the dimension of the extension space $W$ is not greater than the deficiency index of $T_{0}$. A generalization of the classical Glazman-Krein-Naimark (GKN) Theorem - called the GKN-EM Theorem to acknowledge the work of Everitt and Markus - is key to finding these self-adjoint extensions in $H\oplus W.$ We consider several examples to illustrate our results. | math.FA | math | SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES H ⊕ W : AN
APPLICATION OF THE GENERAL GKN-EM THEOREM
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
Abstract. We construct self-adjoint operators in the direct sum of a complex Hilbert space H and
a finite dimensional complex inner product space W . The operator theory developed in this paper
for the Hilbert space H ⊕ W is originally motivated by some fourth-order differential operators,
studied by Everitt and others, having orthogonal polynomial eigenfunctions. Generated by a closed
symmetric operator T0 in H with equal and finite deficiency indices and its adjoint T1, we define
families of minimal operators { bT0} and maximal operators { bT1} in the extended space H ⊕ W and
establish, using a recent theory of complex symplectic geometry, developed by Everitt and Markus,
a characterization of self-adjoint extensions of { bT0} when the dimension of the extension space W
is not greater than the deficiency index of T0. A generalization of the classical Glazman-Krein-
Naimark (GKN) Theorem - called the GKN-EM Theorem to acknowledge the work of Everitt and
Markus - is key to finding these self-adjoint extensions in H ⊕ W. We consider several examples to
illustrate our results.
7
1
0
2
r
p
A
3
2
]
.
A
F
h
t
a
m
[
1
v
0
5
9
6
0
.
4
0
7
1
:
v
i
X
r
a
Contents
Introduction
1.
2. The von-Neumann Formulas and the GKN Theorem
3. Complex Symplectic Geometry and a Generalization of the GKN Theorem
4. Maximal and Minimal Operators in H ⊕ W
5. Examples
5.1. Example 1: The Legendre Type Self-Adjoint Operator
5.2. Example 2: A Simple First-Order Differential Operator
5.3. Example 3: Variations on the Fourier Self-Adjoint Operator
References
1
4
8
13
21
21
25
27
31
1. Introduction
In [6, p. 105], the authors list ten open problems related to orthogonal polynomial eigenfunctions
of differential equations. The one pertinent to this present paper is the following (paraphrased to
simplify the original notation):
The GKN theory provides a recipe, in theory, for determining all self-adjoint exten-
sions in the Hilbert space L2(I; w) of formally symmetric differential expressions of
Date: April 23, 2017 [wellmanlittlejohnEMfinalversion.tex].
1991 Mathematics Subject Classification. Primary 05C38, 15A15; Secondary 05A15, 15A18.
Key words and phrases. symmetric operator, self-adjoint operator, differential operator, maximal operator, mini-
mal operator, Glazman-Krein-Naimark theory, symplectic GKN theorem, orthogonal polynomialsct .
1
2
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
the form
(1.1)
ℓ2r[y](u) =
1
w(u)
rXj=0
(−1)j (qj(u)y(j)(u))(j)
(u ∈ I)
on some open interval I = (a, b); we assume here that w > 0 and each coefficient qj
is sufficiently differentiable on I. This theory works well in developing the spectral
theory for the second-order classical differential equations of Jacobi, Laguerre, and
Hermite.1 However, for nonclassical symmetric differential equations (1.1) with
orthogonal polynomial solutions, the appropriate right-definite setting is a Hilbert-
Sobolev space S with orthogonalizing Sobolev inner product
(1.2)
hf, gi =Z b
a
f (u)g(u)w(u)du +
pXj=0(cid:16)αjf (j)(a)g(j)(a) + βjf (j)(b)g(j)(b)(cid:17) .
The Sobolev space S has the form L2(I; w)⊕ Ck for some k ≤ 2p. Develop a general
GKN-type theory for this setting; in particular, provide a 'recipe' for determining
the self-adjoint operator having the orthogonal polynomials as eigenfunctions.
In this paper, we answer this question. In fact, we will see that we can provide a recipe for all
self-adjoint operators, generated by ℓ2r[·], in this Sobolev setting. Our result is a generalization
of the Glazman-Krein-Naimark (GKN) theory of self-adjoint extensions of Lagrangian symmetric
ordinary differential expressions in a weighted Hilbert space L2(I; w), where I is an interval of the
real line R.
This work is originally motivated by fourth-order differential equations having non-classical or-
thogonal polynomials as eigenfunctions. In each of these fourth-order examples, the orthogonalizing
inner product has the form
hf, gi = Af (a)g(a) +Z b
a
f (u)g(u)w(u)du + Bf (b)g(b),
where A, B ≥ 0. Indeed, H. L. Krall [13, 14] classified, up to a complex linear change of variable,
these orthogonal polynomials which were subsequently named the Legendre type, Laguerre type
and Jacobi type polynomials and studied extensively by A. M. Krall in [12]. Following the work of
the two Kralls, other contributions connecting orthogonal polynomial eigenfunctions to higher-order
differential equations have emerged; all known examples have polynomial eigenfunctions orthogonal
with respect to an inner product of the form (1.2). These various contributions are far too numerous
to list in this manuscript but we refer to the Erice and Patras reports [5, 6] for further details and
the references therein contained.
In [3, 4], the authors construct the Legendre type self-adjoint operator, generated by the fourth-
order Legendre type differential expression
(1.3)
L2
in the Hilbert space L2
ℓLT [y](u) := (u2 − 1)2y(4)(u) + 8u(u2 − 1)y(3)(u) + (4A + 12)(u2 − 1)y′′(u) + 8Auy′(u),
µ[−1, 1] = {f : [−1, 1] → C f is Lebesgue measurable and hf, fiµ =Z[−1,1] f2 dµ < ∞},
µ[−1, 1]. Here
1The GKN theory is applicable as well to developing the self-adjoint operator theory associated with exceptional
orthogonal polynomials.
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
3
where
hf, giµ = Af (−1)g(−1) +Z 1
−1
f (u)g(u)du + Af (1)g(1).
µ[−1, 1] is isometrically isomorphic to L2(−1, 1)⊕ C2. The classical Glazman-Krein-
We note that L2
Naimark (GKN) theory of self-adjoint extensions of Lagrangian symmetric differential expressions
is not (immediately) applicable in this situation. To develop the appropriate operator theory in
L2
µ[−1, 1], Everitt and Littlejohn studied properties of functions in the maximal domain ∆ of ℓLT [·]
in the base space L2(−1, 1). They prove the surprising smoothness condition
f ∈ ∆ ⇒ f ′′ ∈ L2(−1, 1)
from which it follows that f, f ′ ∈ AC[−1, 1]. Using standard operator-theoretic methods, they then
prove that the operator T : D(T ) ⊂ L2
µ[−1, 1] defined by
µ[−1, 1] → L2
(T f )(u) =
D(T ) = ∆
−8Af ′(−1) u = −1
ℓLT [f ](u)
−1 < u < 1
8Af ′(1)
u = 1
is self-adjoint. It is remarkable that ∆ is the domain of the self-adjoint operator T in L2
µ[−1, 1].
Indeed, the expression ℓLT [·] is in the limit-3 case at both singular endpoints x = ±1 in L2(−1, 1)
so every self-adjoint operator in L2(−1, 1), generated by ℓLT [·], is necessarily determined by two
appropriate boundary restrictions on the space ∆.
We will re-examine this Legendre type example in Section 5.1 as an application of the results
developed in this paper. To this end, let (W,h·,·iW ) be a finite-dimensional complex inner product
space and assume (H,h·,·iH ) is a complex Hilbert space. Then H ⊕ W , the direct sum of H and
W, is the Hilbert space defined by
(1.4)
with inner product
(1.5)
and associated norm
H ⊕ W = {(x, a) x ∈ H, a ∈ W}
h(x, a), (y, b)iH⊕W := hx, yiH + ha, biW
k(x, a)k2
H⊕W = kxk2
H + kak2
W .
Throughout this paper, we refer to H ⊕ W as an extended Hilbert space and call H the base space
and W the extension space.
Our starting point in this paper - assumptions we keep throughout this article - is a closed,
symmetric operator T0 in H having equal and finite deficiency indices, denoted by their common
value def(T0), and adjoint operator T1 satisfying the inclusions
We call T0 the minimal operator and T1 the maximal operator in H. Then, under the essential
assumption that
T ∗
1 = T0 ⊆ T ∗
0 = T1.
def(T0) ≤ dim W,
we construct one-parameter families {bT0} of minimal operators and associated maximal operators
{bT1} in H ⊕ W, generated by T0 and T1 in H, satisfying the properties
def (bT0) = def(T0)
(bT0 ∈ {bT0})
4
and
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
(cid:16)bT1(cid:17)∗
= bT0 ⊆(cid:16)bT0(cid:17)∗
= bT1
(bT0 ∈ {bT0}, bT1 ∈ {bT1}).
W.
Both families {bT0} and {bT1} are parametrized by an arbitrary, fixed self-adjoint operator B : W →
With the constructions of {bT0} and {bT1} in place, we then appeal to a general theory of complex
symplectic algebra, with important applications and implications to boundary value problems in
ordinary and partial differential equations, which was developed by Everitt and Markus in a series
of remarkable papers [7, 8, 9, 10]. An important consequence of their theory is a generalized GKN
theory - which we call GKN-EM theory after the contributions of Everitt and Markus - that we
apply to characterize all self-adjoint extensions (respectively, restrictions) of bT0 ∈ {bT0} (respectively,
of bT1 ∈ {bT1}).
The contents of this paper are as follows. In Section 2, we briefly discuss the Stone-von Neumann
theory of self-adjoint extensions of symmetric operators in a Hilbert space as well as the now classic
GKN theory, including a statement of the GKN Theorem (Theorem 2.3). Section 3 deals with key
complex symplectic geometric results developed by Everitt and Markus and culminates in the GKN-
EM Theorem (Theorem 3.2). The families {bT0} and {bT1} of minimal and maximal operators in
H⊕W, generated by T0 and T1 in the base space H, are developed in Section 4. Another key notion,
the symplectic form [·,·]H⊕W in the extended space H⊕W , essential to our application of the GKN-
EM theory, is defined in Section 4. Also, in this section, we apply Theorem 3.2 to characterize all
self-adjoint extensions {bT} of bT0 ∈ {bT0}; see the summary theorem given in Theorem 4.4. Lastly,
Section 5 deals with several examples to illustrate our results. These examples include another look
at the Legendre type example where further light is shed on this particular example. Indeed, we
show that, remarkably, continuity is a GKN-EM boundary condition.
Notation: R, C and N will denote, respectively, the sets of real numbers, the complex numbers and
the positive integers. All inner products in this paper will be denoted by h·,·i , properly subscripted
indicating the particular underlying vector space. Ordered pairs in H ⊕ W will be written as (·,·) ;
if dim W > 1, then an ordered pair in H ⊕ W will have the form (·, (·,·)). Our base space will be a
complex Hilbert space (H,h·,·iH), our extension space will be finite-dimensional complex Hilbert
space (W,h·,·iW ) and the extended space will be the direct sum space (H ⊕ W,h·,·iH⊕W ). Linear
operators in the base space H will be denoted by T0, T1, T, etc. while operators in the extended
space H ⊕ W will be hatted: bT0, bT1, bT , etc. The notation
x has property P (x ∈ A)
means that property P holds for all x in the set A. Lastly, the cardinality of a set A is denoted by
card(A) whereas the dimension of subspace W of some vector space will be written dim W.
2. The von-Neumann Formulas and the GKN Theorem
Standard references for topics discussed in this section are [1, 2, 15, 16, 17, 18, 19, 21].
Throughout this paper, the linear operator T0 : D(T0) ⊆ H → H will be an arbitrary closed,
symmetric operator in H while T1 : D(T1) ⊆ H → H is a linear operator satisfying the operator
inclusions
(2.1)
T ∗
1 = T0 ⊆ T ∗
0 := T1;
in particular, we see that T0 and T1 are adjoints of each other. Because of the inclusions in (2.1),
we call T0 the minimal operator and T1 the maximal operator. Specific reasons for this notation
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
5
will be discussed below in this section (see also Remark 2.1). Notice that if T0 has a self-adjoint
extension T in H, then
so T necessarily has the same form as T1; that is,
T0 ⊆ T = T ∗ ⊆ T ∗
0 = T1,
T x = T1x (x ∈ D(T )).
The general theory of self-adjoint extensions of the minimal operator T0 (equivalently, self-adjoint
restrictions of the maximal operator T1) in a Hilbert space - called the Stone-von Neumann theory
- is discussed in depth in [2, Chapter XII, Section 4]. Of central importance in this theory are two
particular subspaces X± of D(T1), defined by
X± := {x ∈ D(T1) T1x = ±ix} ,
where i = √−1. These spaces are called the positive and negative deficiency spaces of T0. The first
von Neumann formula decomposes the maximal domain D(T1) into linearly independent subman-
ifolds:
Theorem 2.1 (The First von Neumann Formula). D(T1) = D(T0) + X+ + X−.
In fact, the sum in this formula is actually an orthogonal direct sum. Indeed, under the graph inner
product
(2.2)
and associated norm
(2.3)
hx, yi∗
H := hx, yiH + hT1x, T1yiH (x, y ∈ D(T1))
(kxk∗
H )2 = kxk2
H + kT1xk2
H ≥ kxk2
H ,
D(T1) is a Hilbert space and, with this inner product, D(T0), X+ and X− are closed, orthogonal
subspaces of D(T1); see [2, Chapter XII]. Notice that if x ∈ D(T1) and
x = x0 + x+ + x−,
where x0 ∈ D(T0) and x± ∈ X±, then
(2.4)
(kxk∗
H)2 =(cid:16)(cid:13)(cid:13)x0(cid:13)(cid:13)∗
H(cid:17)2
+(cid:0)(cid:13)(cid:13)x+(cid:13)(cid:13)∗
H(cid:1)2
+(cid:0)(cid:13)(cid:13)x−(cid:13)(cid:13)∗
H(cid:1)2
.
The dimensions of X±, denoted by dim(X±), are called the positive and negative deficiency indices
of T0. A key result in the Stone-von Neumann theory is that the equality of these deficiency indices
is equivalent to the existence of self-adjoint extensions T of T0 in H. Moreover, if dim(X+) =
dim(X−) = 0, T0 = T1 is self-adjoint and is, in fact, the only self-adjoint extension of T0 in H.
In the case that dim(X+) = dim(X−), we refer to this common value as the deficiency index and
denote it by def(T0). In addition to requiring the equality of these deficiency indices for the entirety
of this paper, we assume the deficiency indices are also finite. Thus, another key assumption in
this paper is:
Condition 2.1. 1 ≤ def(T0):= dim(X+) = dim(X−) < ∞.
The second von Neumann formula gives a description of the domain of any self-adjoint extension
T of T0 in H:
6
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
Theorem 2.2 (The Second von Neumann Formula). Let T : D(T ) ⊆ H → H be a self-adjoint
extension of T0. Then there exists an isometric isomorphism V : X+ → X− from the positive
deficiency space X+ onto the negative deficiency space X− such that
(2.5)
T x = T1x
(2.6)
Conversely, if T and its domain D(T ) are defined through (2.5) and (2.6) for some isometric
isomorphism V : X+ → X− , then T is a self-adjoint extension of T0.
D(T ) = {x + x+ + V x+ x ∈ D(T0), x+ ∈ X+}.
The Glazman-Krein-Naimark (GKN) theory is both a refinement and an application of the
Stone-von Neumann theory to self-adjoint operator extensions of ordinary differential expressions.
Excellent expositions of this theory can be found in Akhiezer and Glazman [1, Volume II, Chapter 8]
and Naimark [15, Part II, Chapter V]. To describe this theory we assume, for the sake of simplicity,
that ℓ[·] is a real, 2n-th order Lagrangian symmetrizable differential expression of the form
(2.7)
ℓ[y](u) =
1
(−1)j(cid:16)qj(u)y(j)(u)(cid:17)(j)
nXj=0
w(u)
(u ∈ I),
where each coefficient qj : I → R in (2.7) is j-times continuously differentiable on I (noting,
however, that general 'quasi-differentiable' conditions can be placed on these coefficients; see also
[20]). The setting for the study of ℓ[·] is the Hilbert space
L2(I; w) = {f : I → C f is Lebesgue measurable and ZI f2 wdu < ∞}
endowed with the standard inner product
hf, giw =Z b
a
f (u)g(u)w(u)du (f, g ∈ L2(I; w)).
Here I ⊆ R is an open interval and w is a positive (a.e.) Lebesgue measurable function on I. The
maximal operator L1 : D(L1) ⊆ L2(I; w) → L2(I; w), generated by ℓ[·], is defined to be
L1f = ℓ[f ]
f ∈ D(L1) = {f : I → C f (j) ∈ ACloc(I) (j = 0, 1, . . . , 2n − 1); f, ℓ[f ] ∈ L2(I; w)}.
In this setting, the term 'maximal' is appropriate; indeed, D(L1) - which is called the maximal
domain - is the largest subspace of L2(I; w) for which the expression ℓ[·] acts on and maps into
L2(I; w). It is clear that L1 is a densely defined operator. We denote the adjoint of L1 by L0; it
is natural then to call L0 the minimal operator generated by ℓ[·]. The GKN theory shows that, in
fact, L1 and L0 are adjoint to each other and L0 is a closed symmetric operator in L2(I; w). More
explicitly, L∗
1 = L0 and
L0 = L0 ⊆ L∗
0 = L1.
Remark 2.1. The operators T0 and T1, defined earlier, are analogous to the minimal operator L0
and maximal operator L1, respectively. Because of this, we call T0 and T1, respectively, the minimal
and maximal operators even though, in the general situation, the terms maximal and minimal may
not seem as appropriate as they do in the GKN theory. Likewise, we shall call their respective
domains the minimal domain D(T0) and the maximal domain D(T1).
The domain D(L0) of the minimal operator is given explicitly by
(2.8)
D(L0) = {f ∈ D(L1) [f, g]wb
a = 0 for all g ∈ D(L1)},
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
7
where [·,·]wb
(2.9)
a is the skew-symmetric bilinear form obtained from the classic Green's formula
hL1f, giw − hf, L1giw = [f, g]wb
a
(f, g ∈ D(L1)).
Moreover, we note that Condition (2.1) is automatically satisfied in this setting.
deficiency indices of L0 are equal since ℓ has real coefficients and thus
Indeed, the
Moreover, in this case,
ℓ[f ] = if ⇐⇒ ℓ[f ] = ℓ[f ] = −if .
0 ≤ def(L0) ≤ 2n.
We are now in position to state the GKN Theorem. Notice that this theorem provides a 'recipe' for
constructing all self-adjoint extensions of the minimal operator L0 in L2(I; w) by specifying certain
restrictions (boundary conditions), using the bilinear form [·,·]w, on the maximal domain D(L1).
We emphasize, however, that the original GKN Theorem is valid only for the minimal operator
L0 associated with a real Lagrangian symmetrizable differential expressions of even order in the
specific Hilbert space L2(I; w). Compare the statement of the GKN Theorem below with that of
the GKN-EM Theorem (Theorem 3.2) at the end of the next section.
Theorem 2.3 (The GKN Theorem). Suppose L0 and L1 are, respectively, the minimal and maximal
operators in L2(I; w), generated by the differential expression ℓ[·], given in (2.7). In addition, let
m = def(L0) so 0 ≤ m ≤ 2n.
(i) Suppose the set {gj j = 1, . . . , m} ⊆ D(L1) satisfies the two conditions
(a)
mXj=1
αjgj ∈ D(L0) =⇒ αj = 0 (j = 1, . . . , m) and
(b)
[gi, gj]wb
a = 0
(i, j = 1, . . . , def(L0)).
Define the operator S : D(S) ⊆ L2(I; w) → L2(I; w) by
Sf = L1f
(2.10)
(2.11)
(2.12)
(2.13)
f ∈ D(S) = {f ∈ D(L1) [f, gj]wb
(j = 1, . . . , def(L0))}.
Then S is a self-adjoint extension of the minimal operator L0 in L2(I; w).
a = 0
(ii) Conversely, if S : D(S) ⊆ L2(I; w) → L2(I; w) is a self-adjoint extension of the minimal
operator L0 in L2(I; w), then there exists a set {gj j = 1, . . . , m} ⊆ D(L1) satisfying the
conditions (2.10) and (2.11) such that S is given explicitly by (2.12) and (2.13).
Remark 2.2. A collection of vectors {gj j = 1, . . . , m} ⊆ D(L1) satisfying condition (2.10)
are said to be linearly independent modulo D(T0) while those that satisfy (2.11) are said to satisfy
Glazman symmetry conditions. Further light, as well as a generalization, into these concepts be
made in the next section.
Remark 2.3. Each of the conditions
given in (2.13), is called a 'boundary condition'. In the case that def(L0) = 0, then L1 (= L0) is
the only self-adjoint extension of L0 and, in this case, there are no boundary conditions.
[f, gj]wb
a = 0
(j = 1, . . . , def(L0)),
8
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
The GKN-EM Theorem, which we discuss in the next section in Theorem 3.2, is a generalization
of the GKN Theorem but, remarkably, is valid in an arbitrary Hilbert space for an arbitrary closed
symmetric operator with equal and finite deficiency indices. This theorem is a highlight application
of the general complex symplectic theory developed by Everitt and Markus.
3. Complex Symplectic Geometry and a Generalization of the GKN Theorem
In a series of papers [7, 8, 9, 10], Everitt and Markus developed an extensive theory of complex
symplectic geometry with applications to linear ordinary and partial differential equations. Their
work was motivated by their interest in boundary value problems. In this section, we report on
their results that pertain to this manuscript. A highlight application of their investigations is an
important, and remarkable, generalization of Theorem 2.3; see Theorems 3.1 and 3.2 below. This
generalization is key to the results we establish in the next section.
Definition 3.1. A complex symplectic space S is a complex vector space together with a conjugate
bilinear (sesquilinear) complex-valued function [· : ·] : S×S→ C satisfying the properties
(i) [c1x1 + c2x2 : y] = c1[x1 : y] + c2[x2 : y],
(ii) [x : y] = −[y : x],
(iii) [x : S] = 0 =⇒ x = 0 (non-degenerate condition).
We call [· : ·] a (non-degenerate) symplectic form.
Complex symplectic spaces are non-trivial generalizations (not merely complexifications) of clas-
sical real symplectic spaces of Lagrangian and Hamiltonian mechanics (see [11]). Indeed, complex
symplectic spaces have a much wider scope and admit new applications. For example, whereas
real symplectic spaces cannot be odd dimensional, it is the case that, for every n ∈ N, there exists
complex symplectic spaces of dimension n.
Along with their real symplectic counterparts, complex symplectic spaces support the notion of
Lagrangian subspaces (see [10] equation (1.10)).
Definition 3.2. A subspace L of a complex symplectic space S is called Lagrangian if [L : L] = 0;
that is to say, when
A Lagrangian L ⊆ S is called a complete Lagrangian when
x ∈ S and [x : L] = 0 ⇒ x ∈ L.
[x : y] = 0
(x, y ∈ L).
We can characterize complete Lagrangian subspaces as follows. This characterization is key for
later results.
Lemma 3.1. A Lagrangian subspace L ⊆ S is a complete Lagrangian if and only if
(3.1)
L = {x ∈ S [x : y] = 0 (y ∈ L)}.
Proof. Suppose L is a complete Lagrangian subspace of S. By definition of complete, it is clear that
{x ∈ S [x : y] = 0 (y ∈ L)} ⊆ L. On the other hand, if x ∈ L then [x : y] = 0 for all y ∈ L since L
is Lagrangian. Hence L ⊆ {x ∈ S [x : y] = 0 (y ∈ L)}. Conversely, if L is Lagrangian and given by
(3.1), then it is clear that L is a complete.
(cid:3)
An essential step in the work of Everitt and Markus is a natural generalization of the skew-
symmetric bilinear form [·,·]wb
Definition 3.3. [x, y]H := hT1x, yiH − hx, T1yiH for x, y ∈ D(T1) .
a given by Green's formula (2.9).
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
9
Following the work of Everitt and Markus, we will see below that [·,·]H can be identified with
a degenerate symplectic form. We also note that [·,·]H coincides with [·,·]wb
a in the case T1 is the
maximal differential operator, generated by ℓ[·] (see (2.7)), in the weighted Hilbert space L2(I; w).
As shown in [10], the quotient space
S′ := D(T1)/D(T0),
Notice that, from Theorem 2.1 and Condition 2.1, S
(3.2)
with zero element 0 = D(T0), is a complex symplectic space when endowed with the form [·,·]H ;
we outline the specific details below.
has dimension 2def(T0). Indeed one may
view S
as an isomorphic copy of the orthogonal sum of the deficiency spaces X± of T0. Everitt
and Markus call the space S
are, of course, cosets x
= {x + D(T0)} (x ∈ D(T1)). In this case, we call the vector x a representative vector of the coset
{x + D(T0)}.
the boundary space of T0. The elements of S
′
′
′
′
We now consider the natural projection φ : D(T1) → S
′
defined by
φ(x) = {x + D(T0)}
(x ∈ D(T1)) .
The following proposition makes clear the connection between a basis of a subspace of S
and the
notion of linear independence modulo D(T0) which we first encountered in Theorem 2.3 and Remark
2.2.
Lemma 3.2. A collection of cosets {φtj}d
dimension d of the boundary space S
j=1 ⊆ D(T1), is a basis for a subspace of
j=1, where {tj}d
if and only if the representative vectors {tj}d
j=1 satisfy
′
′
dXj=1
αjtj ∈ D(T0) =⇒ αj = 0
(j = 1, 2, . . . , d);
that is to say, {tj}d
Proof. The equation Pd
j=1 is linearly independent modulo D(T0).
j=1 αjφtj = 0 is equivalent to Pd
j=1 αjtj ∈ D(T0).
The following lemma generalizes the characterization of the domain of the minimal operator; see
(cid:3)
(2.8).
Lemma 3.3. D(T0) = {x ∈ D(T1) [x, y]H = 0 (y ∈ D(T1))} .
Proof. Fix x ∈ D(T1) and suppose
[x, y]H = 0 (y ∈ D(T1)).
Since [x, y]H = −[y, x]H, we see that hT1y, xiH = hy, T1xiH so x ∈ D(T ∗
x ∈ D(T0). Since T ∗
0 = T1 and T0x = T1x, we see that
1 ) = D(T0). Conversely, let
hT1x, yiH = hT0x, yiH = hx, T1yiH (y ∈ D(T1));
that is, for each y ∈ D(T1),
[x, y]H = hT1x, yiH − hx, T1yiH = 0.
(cid:3)
This result allows the boundary space S
′
to be equipped with a complex symplectic form.
Definition 3.4 (Boundary Space Symplectic Form).
(3.3)
[φx : φy]S′ := [x, y]H (x, y ∈ D(T1)).
10
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
Lemma 3.3 assures Definition 3.4 above is independent of the choice of representative vectors.
Moreover, Lemma 3.3 establishes the non-degeneracy property of Definition 3.1. From the definition
of a Lagrangian subspace, the following extension of Lemma 3.2 is clear.
Proposition 3.1. A collection of cosets {φtj}d
subspace of the boundary space S′ if and only if the representative vectors {tj}d
j=1 form a basis for a d-dimensional Lagrangian
j=1 satisfy
(a)
(3.4)
(b)
(3.5)
and
dXj=1
αjtj ∈ D(T0) =⇒ αj = 0 (j = 1, . . . , d);
[ti, tj]H = 0
(i, j = 1, . . . , d).
Notice that the properties (3.4) and (3.5) are identical to those conditions discussed in Theorem
2.3. Because of their importance in the special case when d = def (T0) , we incorporate these two
properties into the following definition.
Definition 3.5. A collection of vectors {tj j = 1, . . . , def (T0)} ⊆ D(T1) is called a GKN set for
T0 if
(i) the set {tj j = 1, . . . , def (T0)} is linearly independent modulo the minimal domain D(T0);
that is to say
(3.6)
and
if
def(T0)Xj=1
αjtj ∈ D(T0) then αj = 0 for j = 1, . . . , def (T0) ;
(ii) the set {tj j = 1, . . . , def (T0)} satisfies the symmetry conditions
[ti, tj]H = 0
(3.7)
Remark 3.1. Observe that if G ⊆ D(T1) is a GKN set for T0, then any non-empty, proper subset
P ⊆ G is linearly independent modulo D(T0) and satisfies the symmetry conditions in (3.7). We
refer to P as a partial GKN set. However, we note that the only partial GKN sets P that we use
in this manuscript are those which satisfy card(P ) = dim(W ) ≤ def(T0), where W is a complex
finite-dimensional extension space; see Condition 4.1 in Section 4.
(i, j = 1, . . . , def (T0)).
We now turn our attention to characterizing complete Lagrangians. A key result of Everitt and
Markus in this setting is that not only do complete Lagrangians L exist (see [10, Equations (1.54)
and (1.61)]) but their dimensions are precisely that of the deficiency index; that is,
(3.8)
dim L = def(T0)
(see [10, Equation (3.9)]). Moreover,
Lemma 3.4. With def(T0) < ∞, a Lagrangian subspace L ⊆ S′ is complete if and only if each of
the two conditions hold:
(i) dim L = def(T0);
(ii) L = {φx [φx : φtj]S′ = 0 (j = 1, 2, . . . ,def(T0))} for some GKN set {tj j = 1, 2, . . . ,def(T0)}.
Moreover, in this case,
(3.9)
φ−1L = {x ∈ D(T1) [x, tj]H = 0 (j = 1, 2, . . . , def(T0))}.
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
11
Proof. Suppose L ⊆ S
3.1,
′
(3.10)
is complete. Then, by (3.8), dim L = def(T0) establishing (i). By Lemma
L = {φx [φx : φy]S′ = 0 (φy ∈ L)}.
Let {φtj j = 1, 2, . . . ,def(T0)} be a basis for L. Then, by Proposition 3.1, {tj j = 1, 2, . . . ,def(T0)}
is a GKN set for T0. It follows from (3.10), that
(3.11)
L = {φx [φx : φtj]S′ = 0 (j = 1, 2, . . . , def(T0))},
proving (ii). Lastly, using the identification in (3.3) along with the identity in (3.11), (3.9) is clear.
Conversely, suppose (i) and (ii) hold.
It is straightforward to show that L is a subspace of S
.
Clearly (3.9) follows from (ii). Moreover, since {tj j = 1, 2, . . . ,def(T0)} is a GKN set for T0, we
see that
′
[φti : φtj]S′ = [ti, tj]H = 0.
It follows by taking linear combinations that L is Lagrangian. Finally, from (3.8), we see that L is
complete.
(cid:3)
The authors in [10, Theorem 1.14 and Remark 1.15] establish the following characterization of
self-adjoint extensions of T0 in terms of complete Lagrangian subspaces L of S
′
.
Theorem 3.1 (The Finite-Dimensional GKN-EM Theorem). Let T0 and T1 be, respectively, the
minimal and maximal operators as defined in Section 2 and let S
be given by (3.2). There exists a
one-to-one correspondence between the set {T} of all self-adjoint extensions of T0 and the set {L}
of all complete Lagrangians L ⊆ S
. More specifically,
′
′
(a) if T is a self-adjoint operator with T0 ⊆ T ⊆ T1, then
x ∈ D(T )}
L := {φx ∈ S
′
(3.12)
is a complete Lagrangian subspace of S
′
(b) If L is a complete Lagrangian subspace of S
of dimension def(T0). Moreover, φ−1L = D(T ).
, then L has dimension def(T0). Define
′
D(T ) = {x ∈ D(T1) φx ∈ L}.
Then T : D(T ) ⊂ H → H given by
T x = T1x
x ∈ D(T )
is a self-adjoint operator satisfying T0 ⊆ T ⊆ T1. Moreover, φ−1 L = D(T ).
Combining Theorem 3.1 with Lemmas 3.1 and 3.4, we are now in position to state and prove
an important consequence of Theorem 3.1 which, for our purposes, is key to the results developed
in the next section and in the examples of Section 5. We note that the next theorem is an exact
generalization of the GKN theorem stated in Theorem 2.3.
Theorem 3.2 (The Finite-Dimensional Symplectic GKN-EM Theorem). Suppose T0 and T1 are
linear operators satisfying the conditions set forth in Section 2 and [·,·]H is the symplectic form
defined in Definition 3.3. In particular, we assume T0 has equal and finite deficiency indices denoted
by def (T0).
12
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
(i) If the operator T : D(T ) ⊆ H → H is self-adjoint and satisfies
T0 ⊆ T ⊆ T1
then there exists a GKN set {tj j = 1, . . . , def (T0)} ⊆ D(T1) of T0 such that
(3.13)
D(T ) = {x ∈ D(T1) [x, tj]H = 0 (j = 1, . . . , def (T0))}.
(ii) If {tj j = 1, . . . , def (T0)} ⊆ D(T1) is a GKN set for T0 then the operator T : D(T ) ⊆ H →
H given by
(3.14)
(3.15)
T x = T1x
x ∈ D(T ) = {x ∈ D(T1) [x, tj]H = 0 (j = 1, . . . , def(T0))}
is self-adjoint and satisfies
T0 ⊆ T ⊆ T1.
Proof. (i) Suppose T : D(T ) ⊆ H → H is self-adjoint and satisfies T0 ⊆ T ⊆ T1. By Theorem 3.1,
(3.16)
′
L = {φx ∈ S
′
x ∈ D(T )}
is a complete Lagrangian subspace of S
(3.17)
Moreover, by Lemma 3.4, there exists a GKN set {tj j = 1, 2, . . . ,def(T0)} for T0 such that
of dimension def (T0) from which it follows that
φ−1L = D(T ).
L = {φx [φx : φtj]S′ = 0 (j = 1, 2, . . . , def(T0))}.
and
(3.18)
φ−1L = {x ∈ D(T1) [x, tj]H = 0 (j = 1, 2, . . . , def(T0))}.
Comparing (3.17) with (3.18), we obtain (3.13).
(ii) Suppose {tj j = 1, 2, . . . ,def(T0)} is a GKN set for T0. Let
(3.19)
L = {φx [φx : φtj]S′ = 0 (j = 1, 2, . . . , def(T0))}.
By Lemma 3.4, L is a complete Lagrangian subspace of S
(3.14) and (3.15). Then, from (3.15) and (3.19), we see that
L = {φx x ∈ D(T )}
so that
′
of dimension def(T0). Define T as in
D(T ) = φ−1L = {x ∈ D(T1) φx ∈ L}.
By Theorem 3.1, T is self-adjoint and T0 ⊆ T ⊆ T1.
Remark 3.2. In the case that H = L2(I; w) and T0 and T1 are, respectively, the minimal and
maximal operators L0 and L1, generated by the ordinary differential expression (2.7), Theorem
3.2 is identical to the classical GKN theorem given in Theorem 2.3. Again, it is remarkable that
the GKN theorem extends verbatim to a general Hilbert space with an arbitrary closed symmetric
operator having equal deficiency indices. As in the classical GKN setting, we also call the conditions
(cid:3)
[x, tj]H = 0 for j = 1, . . . , def (T0)
'boundary conditions'. Lastly, we note that, as in Remark 2.3, if def(T0) = 0, there are no such
boundary conditions and, in this case, the only self-adjoint extension of T0 is the maximal operator
T1 (= T0).
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
13
Remark 3.3. Everitt and Markus discuss other important applications of their results to ordinary
and partial differential operators. We refer the reader to Sections 2.1, 2.2 and 4.2 in [10]. They
outline the argument given above in Theorem 3.2 for Sturm-Liouville problems (see [10, Section 2,
equations (2.23), (2.24), and (2.25)]) as well as for general Shin-Zettl quasi-differential operators
(see [10, Section 4; in particular equations (4.57)–(4.61)]). The authors are certain that the most
general situation (when T0 has finite and equal deficiency indices), which we prove in Theorem 3.2,
was known to Everitt and Markus but we cannot find an exact reference in their joint work.
4. Maximal and Minimal Operators in H ⊕ W
deficiency indices def(T0) and adjoint operator T1 satisfying T ∗
by, respectively, the minimal operator T0 and the maximal operator T1 in the base space H. We
We remind the reader that T0 : D(T0) ⊆ H → H is a closed, symmetric operator with equal, finite
0 = T1. In this section,
we identify a family of minimal operators bT0 : D(bT0) ⊆ H ⊕ W → H ⊕ W and an associated family
of maximal operators bT1 : D(bT1) ⊆ H ⊕ W → H ⊕ W in the extended space H ⊕ W generated
show that each bT0 is a closed, symmetric operator in H ⊕ W with equal deficiency indices and
def(bT0) = def(T0). Moreover, the operators bT0 and bT1 are adjoints of each other just as in the
A fundamental assumption in our development of the maximal and minimal operators in H ⊕ W
classical case with T0 and T1.
1 = T0 ⊆ T ∗
is the following dimensionality requirement for the extension space:
Condition 4.1. dim(W ) 6 def (T0) .
Fix a partial GKN set
(4.1)
{tj j = 1, . . . , dim(W )} ⊆ D(T1);
recall, from Remark 3.1 and Condition 4.1, that this set exists and satisfies the two conditions
(4.2)
and
(4.3)
dim(W )Xj=1
αjtj ∈ D(T0) =⇒ αj = 0
(j = 1, . . . , dim(W ))
[ti, tj]H = 0
(i, j = 1, . . . , dim(W )).
It is clear that the maximal operator T1 in the base space is symmetric on
∆0 := D(T0) + span{tj j = 1, . . . , dim(W ) ⊆ D(T1).
(4.4)
Now let
be an orthonormal basis of W and define Ψ : ∆0 → W by
{ξj j = 1, . . . , dim(W )} ⊆ W
Ψ (tj) = ξj
Ψ (s) = 0
(4.5)
and extend Ψ to ∆0; that is to say
Ψ
s +
dim(W )Xj=1
(j = 1, . . . , dim(W ))
(s ∈ D(T0)).
αjtj
=
dim(W )Xj=1
αjξj.
Note the key fact that Ψ maps the partial GKN set {tj j = 1, . . . , dim(W )} onto W.
14
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
generated by T0.
Lastly, fix an arbitrary self-adjoint operator B : W → W in the extension space W . With these
definitions and conditions in place, we are now in position to define a minimal operator cT0 in H⊕W
Definition 4.1. The minimal operator bT0 : D(cid:16)bT0(cid:17) ⊆ H ⊕ W → H ⊕ W is defined to be
bT0(x, a) = (T1x, Ba)
(x, a) ∈ D(bT0) := {(x, Ψx) x ∈ ∆0} .
in fact, a densely defined operator which is both closed and symmetric. Moreover, in Theorem 4.2,
At this point, it is unclear why we call bT0 the minimal operator generated by T0; we will justify
this terminology in Remark 4.1. In Theorem 4.1 below we show that the minimal operator bT0 is,
where it is shown that (bT0)∗ = bT1, we introduce the important linear transformation Ω : D(T1) →
W defined by
(4.8)
Ωx :=
dim(W )Xj=1
[x, tj]Hξj
(x ∈ D(T1)).
(4.9)
Observe, by definition of the partial GKN set {tj j = 1, . . . , dim W} and Lemma 3.3, that
With this transformation Ω, we are now ready to introduce the maximal operator bT1.
Definition 4.2. The maximal operator bT1 : D(cid:16)bT1(cid:17) ⊆ H ⊕ W → H ⊕ W is defined by
Ωx = 0 (x ∈ ∆0).
(4.6)
(4.7)
(4.10)
(4.11)
that is
(4.12)
bT1(x, a) = (T1x, Ba − Ωx)
(x, a) ∈ D(bT1) := {(x, a) x ∈ D(T1), a ∈ W} .
Note that if (x, Ψx) ∈ D(bT0), then (x, Ψx) ∈ D(bT1). Moreover, in this case, Ωx = 0 by (4.9) so
bT1(x, Ψx) = (T1x, BΨx − Ωx) = (T1x, BΨx) = bT0(x, Ψx);
Remark 4.1. The term 'maximal' is appropriate; indeed, observe that D(bT1) is the largest linear
manifold in H ⊕ W on which an operator representation of T1 makes sense. Moreover, once we
establish the fact that (bT0)∗ = bT1, we see that the term 'minimal' is appropriate for the operator bT0.
Proposition 4.1. The extension J : D(J) ⊆ H → H of the minimal operator T0, defined by
bT0 ⊆ bT1.
Jx := T1x
x ∈ D(J) := ∆0,
is a closed symmetric operator.
Proof. Since T0 is densely defined and D(T0) ⊆ ∆0, it is clear that D(J) is dense in H. Now, from
Lemma 3.3 and (3.7), we see that
Hence, from Definition 3.3,
[x, y]H = 0 (x, y ∈ ∆0 ⊆ D(T1)).
0 = [x, y]H = hT1x, yiH − hx, T1yiH = hJx, yiH − hx, JyiH
(x, y ∈ D(J)),
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
15
establishing that J is symmetric in H. To show that J is closed, suppose first that dim(∆0)
(mod(D (T0))) = 1; that is
(4.13)
D(J) = D(T0) + span{t1},
where t1 ∈ D(T1)(cid:31)D(T0). Consider a sequence {xj} ⊆ ∆0 and vectors x, y ∈ H such that xj → x
and T1xj = Jxj → y where the convergence of both sequences is in H. Of course, we need to show
(4.14)
x ∈ D(J)
and
(4.15)
Jx = y.
Since D(J) ⊆ D(T1) and T1 is closed, we know that x ∈ D(T1) and T1x = y. Hence (4.15) will be
established once we show (4.14). Now, by (4.13) and Theorem 2.1, we can write
(4.16)
xj = vj + αjt1
where t1 = t0
x± ∈ X±. Since xj → x and T1xj → T1x, we see that
1 and vj, t0
1 + t−
1 + t+
1 + vj) + αjt+
1 + αjt−
1
(j ∈ N)
= (αjt0
1 ∈ D(T0), t±
1 ∈ X± and x = x0 + x+ + x−, where x0 ∈ D(T0) and
kxj − xk∗
H → 0 as j → ∞,
where k·k∗
H is the graph norm given in (2.3). Since
(cid:0)kxj − xk∗
H(cid:1)2 =(cid:16)(cid:13)(cid:13)αjt0
H(cid:17)2
1 + vj − x0(cid:13)(cid:13)∗
we see, from (2.3) and (2.4), that
(4.17)
(4.18)
(4.19)
1 − x+(cid:13)(cid:13)∗
H(cid:1)2
+(cid:0)(cid:13)(cid:13)αjt−
1 − x−(cid:13)(cid:13)∗
H(cid:1)2
,
+(cid:0)(cid:13)(cid:13)αjt+
1 + vj → x0 in H
1 → x+ in H
αjt+
αjt−
1 → x− in H.
αjt0
1 and t−
Since t+
1 ∈ D(T0), contradicting our choice of t1),
we see from either (4.18) or (4.19) that there exists α ∈ C with αj → α. It follows that αjt1 → αt1
in H. Then, from (4.16) and (4.17), we see that
1 both cannot be zero (otherwise, t1 = t0
xj = vj + αjt1
Hence we see that x = (x0 − αt0
follows by induction on dim(W ).
1
1 + vj + αjt1 − αjt0
= αjt0
→ (x0 − αt0
1) + αt1 ∈ D(J), as required. The general proof of this proposition
1) + αt1 ∈ D(T0) + span{t1}.
(cid:3)
Remark 4.2. Proposition 4.1 shows that, on ∆0, the maximal operator T1 is a closed, symmetric
operator. Of course, T1 is not, in general, symmetric on D(T1).
Theorem 4.1. The operator bT0 is a closed, densely defined symmetric operator in H ⊕ W .
16
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
Proof. (i) bT0 is Hermitian:
Let (x, Ψx), (y, Ψy) ∈ D(bT0). Then, by Proposition 4.1 and the fact that B is symmetric in W , we
see that
DbT0(x, Ψx), (y, Ψy)EH⊕W
= h(T1x, BΨx), (y, Ψy)iH⊕W
= hT1x, yiH + hBΨx, ΨyiW
= hx, T1yiH + hΨx, BΨyiW
=D(x, Ψx), bT0(y, Ψy)EH⊕W
.
Hence bT0 is Hermitian.
(ii) D(bT0) is dense in H ⊕ W :
Since D(T0) is dense in H and Ψ is surjective, it is clear that D(bT0) is dense in H ⊕ W.
(iii) bT0 is symmetric in H ⊕ W :
(iv) bT0 is closed in H ⊕ W :
Suppose that {(xn, Ψxn)} ⊆ D(bT0) is such that
This follows immediately from (i) and (ii).
(4.20)
These conditions in (4.20) and (4.21) are equivalent to
We need to show that (x, a) ∈ D(bT0) and bT0(x, a) = (y, b); that is to say, we need to prove:
(4.24)
(xn, Ψxn) → (x, a) in H ⊕ W
bT0(xn, Ψxn) → (y, b) in H ⊕ W.
xn → x and T1xn → y in H
Ψxn → a and BΨxn → b in W.
x ∈ ∆0
T1x = y
Ψx = a
and
(4.21)
(4.22)
and
(4.23)
(4.25)
(4.26)
and
(4.27)
Since {xn} ⊆ ∆0, we see that T1xn = Jxn so, by Proposition 4.1,
Ba = b.
x ∈ ∆0 and Jx = T1x = y,
establishing (4.24) and (4.25). For the remainder of this proof, write
(4.28)
and
(4.29)
x = x0 + t
xn = xn,0 + tn,
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
17
where x0, xn,0 ∈ D(T0),
and
t =
dim WXj=1
αjtj,
tn =
dim WXj=1
αn,jtj.
From (4.29) and the definition of Ψ, we see that
Ψxn = Ψ(xn,0 + tn) = Ψtn =
dim WXj=1
αn,jξj
so that, from (4.23),
It follows that
(4.30)
so that T1tn → T1bt. Notice that
(4.31)
αn,j = (Ψxn, ξj)W → ha, ξjiW .
dim WXj=1
tn →bt :=
ha, ξjiW tj
Ψtn → Ψbt =
dim WXj=1
ha, ξjiW ξj = a.
From (4.22), (4.29) and (4.30), we deduce that
xn,0 = xn − tn → x −bt = x0 + t −bt in H
T0xn,0 = T1xn − T1tn → y − T1bt in H.
Since T0 is closed, we see that x0 + t−bt ∈ D(T0) and, in particular, that t−bt ∈ D(T0). By definition
of the partial GKN set {tj 1 ≤ j ≤ dim W}, we must have
(4.32)
Combining (4.31) and (4.32), we obtain
establishing (4.26). Finally,
Ψxn = Ψtn → Ψbt = Ψt = Ψ(x − x0) = Ψx = a;
t =bt.
BΨxn → BΨx = Ba
This brings us to the proof of the fundamental relation between the maximal and minimal
so, by (4.23), Ba = b which proves (4.27). This completes the proof that bT0 is closed.
operators bT1 and bT0.
Theorem 4.2. (bT0)∗ = bT1.
(cid:3)
18
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
Proof. For (x, a) , (y, b) ∈ D(bT1), a calculation shows that
DbT1(x, a), (y, b)EH⊕W −D(x, a), bT1(y, b)EH⊕W
(4.33)
(see Definition 4.50). Notice that when y = tk and b = ξk, we obtain
= [x, y]H − hΩx, biW + ha, ΩyiW .
(4.34)
hΩx, ξjiW =
dim WXk=1
h[x, tk]Hξk, ξjiW = [x, tj]H
since {ξj j = 1, . . . , dim W} is an orthonormal basis of W. Suppose now that (y, b) ∈ D(bT0) so
y ∈ ∆0 and b = Ψy. Then
where y0 ∈ D(T0),
y = y0 +et,
dim WXj=1
et :=
αjtj
(4.35)
(4.37)
(4.38)
(4.39)
(4.40)
(4.41)
and
dim WXj=1
By (4.12), bT0(y, Ψy) = bT1(y, Ψy) so, from (4.33), we obtain
b = Ψy = Ψet =
αjξj.
DbT1(x, a), (y, Ψy)EH⊕W −D(x, a), bT0(y, Ψy)EH⊕W
= [x, y0]H + [x,et]H − hΩx, Ψy0iW −(cid:10)Ωx, Ψet(cid:11)W + ha, ΩyiW .
= [x, y]H − hΩx, biW + ha, ΩyiW
We now deal with each of the five terms in (4.35). First, from Lemma 3.3,
(4.36)
From (4.34), we see that
[x, y0]H = 0.
[x,et]H −(cid:10)Ωx, Ψet(cid:11)W =
dim WXj=1
αj{[x, tj]H − hΩx, ξjiW} = 0.
From (4.5), Ψy0 = 0 so
Likewise, from (4.9), we see that Ωy = 0 so
hΩx, Ψy0iW = 0.
ha, ΩyiW .
=D(x, a), bT0(y, Ψy)EH⊕W
bT1 ⊆ (bT0)∗.
Together, (4.36), (4.37), (4.38) and (4.39) show that
DbT1(x, a), (y, Ψy)EH⊕W
and, hence, we obtain
((x, a) ∈ D(bT1), (y, Ψy) ∈ D(bT0))
To show (bT0)∗ ⊆ bT1, let (x, a) ∈ D((bT0)∗) and set (x∗, a∗) = (bT0)∗(x, a). Then for (y, Ψy) ∈ D(bT0),
h(x∗, a∗), (y, Ψy)iH⊕W =D(bT0)∗(x, a), (y, Ψy)EH⊕W
=D(x, a), bT0(y, Ψy)EH⊕W
(4.42)
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
19
since bT0 is closed. Written out, the identity in (4.42) gives
(4.43)
In particular, if y ∈ D(T0), then (4.43) reduces to
hx∗, yiH + ha∗, ΨyiW = hx, T0yiH + ha, BΨyiW .
Thus x ∈ D(T ∗
(4.44)
0 ) = D(T1) and def
hx∗, yiH = hx, T0yiH .
T1x = x∗.
Substituting (4.44) into (4.43) and recalling that B is symmetric in W yields
ha∗, ΨyiW = hx, T0yiH + ha, BΨyiW − hT1x, yiH
(4.45)
In particular, let y = tk so Ψy = ξk. From (4.34), we see that [x, tk]H = hΩx, ξkiW . Hence, we find
that 4.45) becomes
= −[x, y]H + hBa, ΨyiW .
ha∗, ξkiW = −hΩx, ξkiW + hBa, ξkiW (k = 1, 2, . . . , dim W ).
(4.46)
Since {ξk k = 1, . . . , dim W} is a basis for W, we can conclude from (4.46) that
(4.47)
a∗ = Ba − Ωx.
Consequently, from (4.44) and (4.47), we see that
so
(x∗, a∗) = (T1x, Ba − Ωx) = bT1(x, a)
(cid:3)
(4.48)
Together Theorem 4.1 and Theorem 4.2 establish the following fundamental operator relationship
(bT0)∗ ⊆ bT1.
Combining (4.41) and (4.48), we obtain (bT0)∗ = bT1.
between bT0 and bT1.
Theorem 4.3. bT0 = bT0 ⊆ (bT0)∗ = bT1.
Consequently we may apply the Stone-von Neumann theory to the minimal operator bT0. Accord-
ingly we define the positive and negative deficiency spaces associated with bT0 in H ⊕ W
Definition 4.3 (Deficiency Spaces in the Extended Space H ⊕ W ).
Y± := {(x, a) ∈ D(bT1) bT1(x, a) = ±i(x, a)}.
Remarkably, as we shall see in the next result, the deficiency spaces Y± ⊆ H ⊕ W and X± ⊆ H
are isomorphic. We note that, since B : W → W is self-adjoint, then B ± iI is invertible.
Lemma 4.1. (x, a) ∈ Y± if and only if x ∈ X± and a = (B ∓ iI)−1 Ωx. Moreover, the deficiency
indices of bT0 are equal and finite and satisfy def(bT0) = def(T0).
Proof. Let (x, a) ∈ Y±. Then T1x = ±ix and Ba − Ωx = ±ia. Therefore x ∈ X± and a =
(B ∓ iI)−1 Ωx. Conversely if x ∈ X± and a = (B ∓ iI)−1 Ωx then Ba − Ωx = ±ia so bT1(x, a) =
±i(x, a). We see that the mappings X± → Y± given by x → (x, (B ∓ iI)−1 Ωx) are vector space
isomorphisms. In particular, dim (X±) = dim (Y±) . This shows that the deficiency indices of the
minimal operator bT0 are finite and equal with
(4.49)
dim (Y+) = dim (Y−) = def(bT0) < ∞.
20
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
In particular, equation (4.49) guarantees the GKN-EM theorem applies to bT0. We now define
the (degenerate) symplectic form in H ⊕ W associated with the operators bT0 and bT1. We remark
that, in equation (4.33), we actually already computed this symplectic form.
Definition 4.4 (General Symplectic Form).
(cid:3)
((x, a), (y, b) ∈ D(bT1)),
[(x, a), (y, b)]H⊕W := [x, y]H − hΩx, biW + ha, ΩyiW
(4.50)
where [·,·]H is the symplectic form defined in (3.3) and where the mapping Ω is defined in (4.8).
We are now in position to apply the GKN-EM Theorem (Theorem 3.2) to the minimal operator
bT0 in H ⊕ W and, as a result, characterize all self-adjoint extensions (respectively, restrictions) of
bT0 (respectively, the maximal operator bT1).
Theorem 4.4 (GKN-EM Theorem in H ⊕ W ). We have the following assumptions/definitions:
(i) T0 and T1 are, respectively, the minimal and maximal operators in (H,h·,·iH ) , called the
base (complex) Hilbert space, with domains D(T0) and D(T1); T0 is a closed, symmetric
operator satisfying T0 ⊆ T1 with T ∗
(ii) The deficiency indices of T0 are assumed to be equal and finite and denoted by def(T0);
(iii) [·,·]H is the symplectic form given by
0 = T1 and T ∗
1 = T0;
[x, y]H = hT1x, yiH − hx, T1yiH
(x, y ∈ D(T1)),
(see Definition 3.3);
def(T0) (Condition 4.1) and orthonormal basis {ξj j = 1, . . . , dim W};
(iv) (W,h·,·iW ), the extension space, is a finite dimensional complex Hilbert space with dimW ≤
(v) B : W → W is a self-adjoint operator;
(vi) H ⊕ W , the extended space, is the Hilbert space defined in (1.4) with inner product (1.5);
(vii) P = {tj j = 1, . . . , dim W} is a partial GKN set (see (4.2) and (4.3));
(viii) ∆0 = D(T0) + span{tj j = 1, . . . , dim W} (see (4.4));
(ix) Ψ : ∆0 → W is defined to be
Ψ
dim WXj=1
x0 +
(x) Ω : D(T1) → W is given by
αjtj
=
(x0 ∈ D(T0))
dim WXj=1
(see (4.5));
αjξj
Ωx =
dim WXj=1
[x, tj]H ξj
(see (4.8));
(xi) bT0 : D(bT0) ⊆ H ⊕ W → H ⊕ W is the minimal operator in H ⊕ W defined by
bT0(x, a) = (T1x, Ba)
(x, a) ∈ D(bT0) = {(x, Ψx) x ∈ ∆0}
(see Definition 4.1);
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
21
(xii) bT1 : D(bT1) ⊆ H ⊕ W → H ⊕ W is the maximal operator in H ⊕ W defined by
bT1(x, a) = (T1x, Ba − Ωx)
D(cT1) = {(x, a) x ∈ D(T1); a ∈ W}
(see 4.2);
(xiii) [·,·]H⊕W is the symplectic form given by
[(x, a), (y, b)]H⊕W := [x, y]H − hΩx, biW + ha, ΩyiW
(see 4.50).
((x, a), (y, b) ∈ D(bT1),
Under these definitions and assumptions, we obtain the following results:
Theorems 4.1 and 4.2);
(a) bT0 is a closed, symmetric operator satisfying bT0 ⊆ bT1 with (bT0)∗ = bT1 and (bT1)∗ = bT0 (see
(b) The deficiency indices of bT0 are equal and finite and def(bT0) = def(T0) (see Lemma 4.1);
(c) Suppose bT is a self-adjoint extension of bT0 (equivalently, bT is a self-adjoint restriction of bT1)
satisfying bT0 ⊆ bT ⊆ bT1. Then there exists a GKN set {(xj, aj) j = 1, . . . , def(T0)} ⊆ D(bT1)
(α) {(xj, aj) j = 1, . . . , def(T0)} is linearly independent modulo D(bT0),
(β) [(xj, aj), (xk, ak)]H⊕W = 0 for j, k = 1, . . . , def(T0)
such that
(see Remark 3.1) satisfying the two conditions
(4.51)
(4.52)
bT (x, a) = (T1x, Ba − Ωx)
D(bT ) = {(x, a) ∈ D(bT1) [(x, a), (xj , aj)]H⊕W = 0 (j = 1, . . . , def(T0))}.
(d) If bT is defined by (4.51) and (4.52) where {(xj , aj) j = 1, . . . , def(T0)} ⊆ D(bT1) is a GKN
set satisfying conditions (α) and (β), then bT is a self-adjoint extension of bT0 (equivalently,
bT is a self-adjoint restriction of bT1) in H ⊕ W.
5. Examples
5.1. Example 1: The Legendre Type Self-Adjoint Operator. Throughout this example,
we let H = L2(−1, 1) (with its usual inner product) and W = C2, endowed with the weighted
Euclidean inner product
(5.1)
h(a1, b1), (a2, b2)iW =
a1a2 + b1b2
A
((a1, b1), (a2, b2) ∈ W );
here A is a fixed, positive constant. Let {ξ1, ξ2} be the orthonormal basis in W given by
(5.2)
and, for this example, suppose B : W → W is the zero self-adjoint operator.
defined earlier in (1.3); that is,
ξ1 = (√A, 0) and ξ2 = (0,√A)
In [3, 4, 12], the authors discuss the spectral analysis of the Legendre type differential expression
(5.3)
ℓLT [y](u) :=(cid:0)(1 − u2)2y′′(u)(cid:1)′′
−(cid:0)(8 + 4A(1 − u2))y′(u)(cid:1)′
where A is the same constant appearing in (5.1). This differential expression was first discovered
by H. L. Krall [13, 14]. When
(5.4)
λn = n(n + 1)(n2 + n + 4A − 2)
(n ∈ N0),
22
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
the equation ℓLT [y] = λny has a polynomial solution y = Pn,A(u) of degree n; that is,
(5.5)
The sequence {Pn,A}∞
sequence in the Hilbert space L2
µ[−1, 1], where
ℓLT [Pn,A] = λnPn,A (n ∈ N0).
n=0 is called the Legendre type polynomials; they form a complete orthogonal
µ[−1, 1] = {f : [−1, 1] → C f is Lebesgue measurable with Z[−1,1] f2 dµ < ∞},
L2
with inner product
hf, giµ =
f (−1)g(−1)
A
+Z 1
−1
f (u)g(u)du +
f (1)g(1)
A
,
and where dµ is the Lebesgue-Stieltjes measure given by
dµ = dx +
1
A
δ(x + 1) +
1
A
δ(x − 1).
When y = Pn,A, we see from (5.3) and (5.5) that
(5.6)
∓ 8AP ′
n,A(±1) = λnPn,A(±1).
Various properties of the Legendre type polynomials can be found in [12].
Because the measure µ has jumps at u = ±1, the classic GKN theory is not immediately appli-
µ[−1, 1]. In order to construct
cable in finding a self-adjoint operator representation T of ℓLT [·] in L2
T, Everitt and Littlejohn [3, 4] first studied properties of functions in the maximal domain
D(T1) = {x : (−1, 1) → C x, x′, x′′, x′′′ ∈ ACloc(−1, 1); x, ℓLT [x] ∈ L2(−1, 1)},
where T1 is the maximal operator, generated by ℓLT [·], in the Hilbert space L2(−1, 1). They establish
the remarkable smoothness property
(5.7)
x ∈ D(T1) =⇒ x′′ ∈ L2(−1, 1)
and hence, upon making the natural identifications
(5.8)
we can say that
x(±1) = lim
u→±1
x(u)
x′(±1) = lim
u→±1
x′(u),
Moreover, they prove that the associated sesquilinear form has the simple formulation
x ∈ D(T1) =⇒ x, x′ ∈ AC[−1, 1].
(5.9)
[x, y]H = 8(x(1)y′(1) − x′(1)y(1) + x′(−1)y(−1) − x(−1)y′(−1))
(x, y ∈ D(T1)).
Considering this last formula and Lemma 3.3, it is apparent that the minimal domain associated
with ℓLT [·] is explicitly given by
(5.10)
The deficiency index of the minimal operator T0, generated by ℓLT [·], in L2(−1, 1) is def(T0) = 2.
This follows since each endpoint u = ±1 is in the limit-3 case which can be shown by a Frobenius
analysis. We emphasize that we are not seeking to find self-adjoint extensions of T0 in L2(−1, 1)
but instead we want to find a self-adjoint representation of ℓLT [·] in L2
µ[−1, 1] which produces the
D(T0) =(cid:8)x ∈ D(T1) x(±1) = x′(±1) = 0(cid:9) .
(5.11)
T x(u) =
x ∈ D(T ) := D(T1)
−8Ax′(−1) u = −1
ℓLT [x](u)
−1 < u < 1
8Ax′(1)
u = 1
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
23
Legendre type polynomials {Pn,A}∞
and Littlejohn show that the operator T : D(T ) ⊆ L2
n=0 as eigenfunctions. By analyzing functions in D(T1), Everitt
µ[−1, 1] → L2
µ[−1, 1] defined by
is self-adjoint, has the Legendre type polynomials {Pn,A}∞
spectrum
n=0 as eigenfunctions, and has discrete
σ(T ) = σp(T ) = {λn n ∈ N0},
where each λn is given in (5.4). It is surprising that the maximal domain D(T1) is the domain
of a self-adjoint operator in L2
µ[−1, 1]. By the GKN Theorem, D(T1) cannot be the domain of a
self-adjoint extension of T0 in L2(−1, 1).
We now show, using the results developed in this paper, how to construct the self-adjoint operator
T given in (5.11) in the direct sum space H ⊕ W. Indeed, below, we construct a self-adjoint operator
bT that is, essentially, the operator T defined in (5.11). With this alternative approach, we will see
how continuity is a GKN-EM boundary condition that produces the Legendre type self-adjoint
operator bT .
µ[−1, 1] is isometrically isomorphic
The first step in our analysis is to observe that the space L2
to the direct sum
H ⊕ W = {(x, (a, b)) x ∈ H; (a, b) ∈ W} .
Next define tj ∈ D(T1) (j = 1, 2) by
t1(u) =(cid:26) √A u near − 1
u near 1
0
t2(u) =(cid:26) 0
u near − 1
√A u near 1;
we remark that such functions in D(T1) exist by Naimark's Patching Lemma [15, Lemma 2, Sec-
tion 17.3]. It is straightforward to see, using (5.9) and (5.10), that {t1, t2} is a GKN set for T0.
Consequently, we see that
(5.12)
where ∆0 is defined in (4.4). Moreover,
∆0 = {x0 + c1t1 + c2t2 x0 ∈ D(T0); c1, c2 ∈ C},
Ψ(x0 + c1t1 + c2t2) = c1ξ1 + c2ξ2 = hc1√A, c2√Ai,
(5.13)
where {ξ1, ξ2} is defined in (5.2) and where Ψ : ∆0 → W is the map defined in (4.5). Using (5.9),
calculations show that
[x, t1]H = 8√Ax′(−1), [x, t2]H = −8√Ax′(1)
(x ∈ D(T1)).
It follows that
Ωx = 8√Ax′(−1)ξ1 − 8√Ax′(1)ξ2 = (8Ax′(−1),−8Ax′(1))
(5.14)
(x ∈ D(T1)),
where Ω : D(T1) → W is the mapping defined in (4.8).
The minimal operator bT0 : D(bT0) ⊆ H ⊕ W → H ⊕ W , in this example, is given by
(5.15)
bT0(x, Ψx) = (T1x, BΨx) = (ℓLT [x], (0, 0))
D(bT0) = {(x, Ψx) x ∈ ∆0}.
(5.16)
From the theory we established in Section 4, bT0 is a closed, symmetric operator in H ⊕ W with
defbT0 = 2.
24
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
(5.17)
Using (5.14), we see that the associated maximal operator bT1 : D(bT1) ⊆ H ⊕ W → H ⊕ W is
given explicitly by
(5.18)
From (5.1), (5.9) and (5.14), a calculation shows that the symplectic form [·,·]H⊕W , defined in
(4.50), is given by
bT1(x, (a, b)) = (ℓLT [x], (−8Ax′(−1), 8Ax′(1)))
D(bT1) = {(x, (a, b)) x ∈ D(T1); a, b ∈ C}.
(5.19)
[(x, (a1, b1)), (y, (a2, b2))]H⊕W
= [x, y]H − hΩx, (a2, b2)iW + h(a1, b1), ΩyiW
= 8(x(1)y′(1) − x′(1)y(1) + x′(−1)y(−1) − x(−1)y′(−1))
− 8x′(−1)a2 + 8x′(1)b2 + 8Ay′(−1)a1 − 8Ay′(1)b1
for (x, (a1, b1)), (y, (a2, b2)) ∈ D(bT1). Define xj ∈ D(T1), for j = 1, 2, by
, x2(u) =(cid:26) √A(u + 1) u near − 1
u near 1.
0
x1(u) =(cid:26) 0
u near − 1
√A(u − 1) u near 1
From (5.9) and (5.10), we see that {x1, x2} is a GKN set for T0. Calculations also show that
(5.20)
[x, x1]H = 8√Ax(1) and [x, x2]H = −8√Ax(−1)
(x ∈ D(T1)).
In addition, we see that
(5.21)
We now claim that
Ωx1 = (0,−8A√A) and Ωx2 = (8A√A, 0).
Proposition 5.1. {(x1, (0, 0)), (x2, (0, 0))} is a GKN set for bT0.
Proof. Suppose that
c1(x1, (0, 0)) + c2(x2, (0, 0)) = (c1x1 + c2x2, (0, 0)) ∈ D(bT0).
By definition of D(bT0), we see that Ψ(c1x1 + c2x2) = (0, 0). This implies that c1x1 + c2x2 ∈ D(T0).
However since {x1, x2} is a GKN set for T0, we must have c1 = c2 = 0. Hence {(x1, (0, 0)),hx2, (0, 0))}
is linearly independent modulo D(bT0). Next, using (5.19), we see that
[(x1, (0, 0)), (x2, (0, 0))]H⊕W = [x1, x2]H = 0.
Calculations also show that
This completes the proof of the Proposition.
(cid:3)
[(xj, (0, 0)), (xj , (0, 0))]H⊕W = 0
(j = 1, 2).
We now find the appropriate self-adjoint operator bT in H ⊕ W having the Legendre type poly-
nomial vectors {(Pn,A, (Pn,A(−1), Pn,A(1)))}∞
(d), the operator bT : D(bT ) ⊆ H ⊕ W → H ⊕ W , defined by
bT (x, (a, b)) = (ℓLT [x], (−8Ax′(−1), 8Ax′(1)))
n=0 as eigenfunctions. Indeed, using Theorem 4.4 part
D(bT ) = {(x, (a, b)) ∈ D(bT1) [(x, (a, b)), (xj , (0, 0))]H⊕W = 0 (j = 1, 2)}
is self-adjoint.
(5.22)
(5.23)
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
25
We now investigate each of the two boundary conditions in (5.23). From (5.19), (5.20) and
(5.21), a calculation shows that
0 = [(x, (a, b)), (x1, (0, 0))]H⊕W
= [x, x1]H − hΩx, (0, 0)iW + h(a, b), Ωx1iW
= [x, x1]H + h(a, b), (0,−8A√A)iW
= 8√Ax(1) −
= 8√Ax(1) − 8√Ab,
8A√Ab
A
implying
(5.24)
b = x(1).
A similar calculation, using (5.20) and (5.21), yields
0 = [(x, (a, b)), (x2, (0, 0))]H⊕W = −8√Ax(−1) + 8√Aa
which establishes
(5.25)
a = x(−1).
Hence the domain of bT , given in (5.23), simplifies to
(5.26)
Notice that this domain (5.26) extends the continuity of each x ∈ D(T1) from (−1, 1) to the
closure [−1, 1]. It is remarkable that, in this sense, continuity is a GKN-EM boundary condition.
Furthermore, from (5.5), (5.6) and (5.22), notice that
D(bT ) = {(x, (x(−1), x(1))) x ∈ D(T1)}.
bT (Pn,A, (Pn,A(−1), Pn,A(1))) = (ℓLT [Pn,A], (−8AP ′
= (λnPn,A, (λnPn,A(−1), λnPn,A(1)))
= λn(Pn,A, (Pn,A(−1), Pn,A(1))).
n,A(−1), 8AP ′
n,A(1)))
Moreover, bT is the same operator as T defined in (5.11).
Remark 5.1. If B : W → W is an arbitrary self-adjoint operator in W, the operator bS : D(bS) ⊆
H ⊕ W → H ⊕ W, defined by
bS(x, (a, b)) = (ℓLT [x], B(a, b) + (−8Ax′(−1), 8Ax′(1)))
is self-adjoint in H ⊕ W . However, it is the case that the Legendre type polynomial vectors
{(Pn,A, (Pn,A(−1), Pn,A(1)))}∞
5.2. Example 2: A Simple First-Order Differential Operator. Let H = L2[0, 1] be endowed
with the standard L2 inner product
D(bS) = {(x, (x(−1), x(1))) x ∈ D(T1)},
n=0 are eigenfunctions of bS if and only if B = 0.
hx, yiH =Z 1
−1
x(u)y(u)du
and let W = C have the usual Euclidean inner product
ha, biW := ab
(a, b ∈ W ).
26
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
In this example, we show how to construct a self-adjoint operator in H ⊕ W generated by the
first-order Lagrangian symmetric differential expression
ℓ[x](u) = ix′(u).
Our construction can be modified to find numerous other self-adjoint operators in H ⊕ W generated
by ℓ[·].
The maximal and minimal domains in H associated with ℓ[·] are respectively given by
T1x = ix′
and
(5.27)
D(T1) = {x : [0, 1] → C x ∈ AC[0, 1]; x′ ∈ H},
T0x = ix′
D(T0) = {x ∈ D(T1) x(0) = x(1) = 0};
see [17, Chapter 13, Example 13.4]. The symplectic form [·,·]H associated with T1is given by
[x, y]H = ix(1)y(1) − ix(0)y(0).
An elementary calculation shows that the deficiency indices of T0 are equal with def(T0) = 1.
Choose {ξ1 = 1} as the orthonormal basis for W . All self-adjoint operators B : W → W have
the form Ba = αa for some real number α; we fix one such an operator. Define t1 ∈ D(T1) by
and note that
(5.28)
t1(u) ≡ 1 (0 ≤ u ≤ 1)
[x, t1]H = i(x(1) − x(0)).
It is clear, from (5.27), that t1 is not the minimal domain; moreover, from (5.28), we see that
[t1, t1]H = 0;
this shows that {t1} is a GKN set for T0 in H. Moreover, with this GKN set, the reader can readily
verify, using Theorem 2.3, that the operator T : D(T ) ⊂ H → H defined by
T x = ix′
D(T ) = {x ∈ D(T1) x(0) = x(1)}
is self-adjoint in H; see [17, Chapter 13, Example 13.4] for an interesting direct proof of the self-
adjointness of T.
The operators Ω : D(T1) → W and Ψ : ∆0 → W from Section 4 are given by
Ωx = [x, t1]H ξ1 = i(x(1) − x(0))
and
The maximal operator bT1 : D(bT1) ⊂ H ⊕ W → H ⊕ W and the minimal operator bT0 : D(bT0) ⊂
H ⊕ W → H ⊕ W can now both be defined. Indeed,
Ψ(t0 + at1) = a (t0 ∈ D(T0)).
bT1(x, a) = (ix′, Ba − Ωx) = (ix′, αa − i(x(1) − x(0)))
D(bT1) = {(x, a) x ∈ D(T1), a ∈ W}
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
27
and
bT0(x, Ψx) = (ix′, BΨx) = (ix′, αΨx)
D(bT0) = {(x, Ψx) x ∈ ∆0}.
The symplectic form [·,·]H⊕W associated with bT1 is given by
[(x, a), (y, b)]H⊕W = [x, y]H − hΩx, biW + ha, ΩyiW
(5.29)
= i(x(1)y(1) − x(0)y(0)) − i(x(1) − x(0))b − ai(y(1) − y(0))
for (x, a), (y, b) ∈ D(bT1).
Define x1 ∈ D(T1) by
from (5.29), we see that
x1(u) =(cid:26) 1 u near 1
0 u near 0;
(5.30)
(5.31)
[(x1, 1/2), (x1, 1/2)]H⊕W = 0;
[(x, a), (x1, 1/2)]H⊕W = ix(1) − i(x(1) − x(0))/2 − ai.
We now show that {(x1, 1/2)} is a GKN set for cT0. From (5.30), note that
We claim that (x1, 1/2) /∈ D(bT0); indeed, otherwise, we have
for some t0 ∈ D(T0). However, choosing u = 0 or u = 1 shows that (5.32) is not possible. It now
follows that {(x1, 1/2)} is a GKN set for bT0.
From (5.30), we see that
(u ∈ [0, 1])
x1(u) = t0(u) +
(5.32)
1
2
t1(u)
(5.33)
[(x, a), (x1, 1/2)]H⊕W = 0 if and only if a =
x(0) + x(1)
2
.
It now follows, from Theorem 3.2 and (5.33), that the operator bT : D(bT ) ⊂ H ⊕ W → H ⊕ W
defined by
(5.34)
is self-adjoint.
bT (x, (x(0) + x(1))/2) = (ix′, α(x(0) + x(1))/2 − i(x(1) − x(0)))
D(bT ) = {(x, (x(0) + x(1))/2) x ∈ D(T1)}
5.3. Example 3: Variations on the Fourier Self-Adjoint Operator. For this example, we
consider the well known Fourier differential expression
(5.35)
ℓF [y](u) = −y′′(u)
(u ∈ [a, b])
where [a, b] is a compact interval. Here, the Hilbert space is H = L2[a, b] and, in the sub-examples
below, we will consider W to be either C or C2 with a weighted Euclidean inner product.
The maximal operator T1 : D(T1) ⊂ H → H is defined by
T1x = ℓF [x]
D(T1) = {x : [a, b] → C x, x′ ∈ AC[a, b]; x′′ ∈ L2[a, b]}
28
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
while the minimal operator T0 : D(T0) ⊂ H → H is given by
T0x = ℓF [x]
D(T0) = {x ∈ D(T1)x(a) = x′(a) = x(b) = x′(b) = 0}.
[x, y]H = x(b)y′(b) − x′(b)y(b) + x′(a)y(a) − x(a)y′(a)
The symplectic form [·,·]H associated with T1 is given by
(5.36)
Because ℓF [·] is regular, the deficiency index of T0 is def(T0) = 2. Consequently, by the GKN
Theorem, every self-adjoint extension of T0 in H will be a certain restriction of the maximal
operator defined by two appropriate boundary conditions. One such self-adjoint operator is the
classical Fourier trigonometric self-adjoint operator in H, generated by ℓF [f ], with domain
(x, y ∈ D(T1)).
(cid:8)x ∈ D(T1) x′(a) = x′(b), x(a) = x(b)(cid:9) .
We list several examples of self-adjoint operators, generated by ℓF [·] in H ⊕ W .
5.3.1. One Dimensional Extension Spaces. Consider the one dimensional extension space W, with
basis {ξ1 = 1}, given by
Every self-adjoint operator B : W → W has the form Bz = αz for some α ∈ R; for this example, we
fix such a B. Observe the Hilbert space H ⊕ W is suggested in a natural way by the inner product
W = C
hz1, z2iW = z1z2
(z1, z2 ∈ W ).
x(u)y(u)dx + x(b)y(b).
Z b
a
With this particular inner product in mind, H ⊕ W is isomorphic to the Lebesgue-Stieltjes inte-
gration space generated by the discontinuous Lebesgue-Stieltjes measure
Example 3.1 With the partial GKN set {t1} for T0 in H given by
dµ = du + δ(u − b).
t1(u) =(cid:26) 1 u near b
0 u near a,
a calculation shows that {(x1, 0), (x2, 0)} is a GKN set for cT0 in H ⊕ W where
u near b
u − a u near a.
x1(u) =(cid:26) u − b u near b
and x2(u) =(cid:26) 0
u near a
0
We leave it to the reader to check that
From these equations, we see that the operator
(i) [(x, a), (x1, 0)]H⊕W = x(b) − a
(ii) [(x, a), (x2, 0)]H⊕W = −x(a).
bT (x, x(b)) = (−x′′, αx(b) + x′(b))
D(bT ) = {(x, x(b)) x ∈ D(T1); x(a) = 0}
is self-adjoint in H ⊕ W.
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
29
Example 3.2 By picking the partial GKN set {t1} for T0 in H, where
and the GKN set {(x1, 0), (x2, 0)} for cT0 in H ⊕ W, where
x1(u) =(cid:26) 0 u near b
1 u near a
t1(u) =(cid:26) 0
u near b
u − a u near a,
and x2(u) =(cid:26) u − b u near b
u near a,
0
the reader can check that the operator
bT (x, x′(a)) = (−x′′, αx′(a) + x(a))
D(bT ) =(cid:8)(x, x′(b)) x ∈ D(T1); x(b) = 0(cid:9)
is self-adjoint in H ⊕ W.
5.3.2. Two Dimensional Extension Spaces. For the last three examples, let W = C2 have the
weighted inner product
(cid:10)(z1, z2), (z′
1, z′
2)(cid:11)W =
z1z′
1
M
+
z2z′
2
N
,
where M, N > 0. Let {ξ1 = (√M , 0), ξ2 = (0,√N)} be a basis for W. The reader can check that
the most general form of a self-adjoint operator B : W → W , using this inner product, has the
matrix representation
and
and
(5.39)
yielding
With z = (z1, z2) ∈ W, calculations show that
(5.38)
Ωx1 = (M 3/2, 0) and Ωx2 = (0,−N 3/2).
0 = [(x, z), (x1, 0)]H⊕W = −x(a)√M + z1√M
0 = [(x, z), (x2, 0)]H⊕W = x(b)√N − z2√N
z1 = x(a) and z2 = x(b).
B =(cid:18)
α
βN/M γ (cid:19) ,
β
(5.37)
where α, β ∈ R and β ∈ C.
Example 3.3 Define ti, xi ∈ D(T1) (i = 1, 2) by
,
t1(u) =(cid:26) √M u near a
x1(u) =(cid:26) √M (u − a) u near a
u near b
u near b
0
0
,
t2(u) =(cid:26) 0
x2(u) =(cid:26) 0
,
u near a
√N u near b
√N (u − b) u near b.
u near a
It is the case that {t1, t2} is a GKN set for T0 in H and {(x1, 0), (x2, 0)} is a GKN set for bT0 in
H ⊕ W. Moreover, using (5.36), we see that
[x, t1]H = x′(a)√M ; [x, t2]H = −x′(b)√N ; [x, x1]H = −x(a)√M ; [x, x2]H = x(b)√N
In particular,
Ωx = [x, t1]ξ1 + [x, t2]ξ2 = (x′(a)M,−x′(b)N )
(x ∈ D(T1)).
30
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
That is, the boundary conditions expressed in (5.38) and (5.39) yield continuity of functions in the
domain of the self-adjoint operator bT : D(bT ) ⊂ H ⊕ W → H ⊕ W defined by
bT (x, (x(a), x(b))) = (−x′′, B(x(a), x(b)) − (x′(a)M,−x′(b)N ))
where B is given in (5.37) and
D(bT ) = {(x, (x(a), x(b))) x ∈ D(T1)},
B(x(a), x(b)) :=(cid:18)
βN/M γ (cid:19)(cid:18) x(a)
x(b) (cid:19) .
α
β
In this case, the setting H ⊕ W can be identified with the Hilbert function space L2
σ[a, b], given by
L2
σ[a, b] = {f : [a, b] → C f is Lebesgue measurable on [a, b] and kfkH⊕W < ∞},
where k·kH⊕W is the norm generated by the inner product
hx, yiH⊕W =Z b
a
x(u)y(u)du +
x(a)y(a)
M
+
x(b)y(b)
N
and dσ is the Lebesgue-Stieltjes measure generated by the distribution function
σ(u) =
1
M
−
u
1
N
+ b
+ a u ≤ a
a < u < b
u ≥ b.
Example 3.4 For this example, we switch the roles of {t1, t2} and {x1, x2} which are defined in
Example 5.3.2. Note that, in this case, {(t1, 0), (t2, 0)} is a GKN set for bT0 and {x1, x2} is a GKN
set for T0. The calculations given in the previous example hold with the exception
Again, with z = (z1, z2), we find that
and
so that
Ωx = (−x(a)M, x(b)M )
(x ∈ D(T1)).
0 = [(x, z), (t1, 0)]H⊕W = x′(a)√M − z1√M
0 = [(x, z), (t2, 0)]H⊕W = −x′(b)√N + z2√N
z1 = x′(a) and z2 = x′(b).
In this case, the operator bT : D(bT ) ⊂ H ⊕ W → H ⊕ W, defined by
D(bT ) =(cid:8)(x, (x′(a), x′(b))) x ∈ D(T1)(cid:9)
bT (x, (x′(a), x′(b))) = (−x′′, B(x′(a), x′(b)) − (−x(a)M, x(b)M )),
with domain
is self-adjoint. In this case, for x, y ∈ D(bT ), the inner product on H ⊕ W simplifies to the discrete
Sobolev inner product
(5.40)
hx, yiH⊕W =Z b
a
x(u)y(u)du +
x′(a)y′(a)
x′(b)y′(b)
M
+
.
N
Notice that, because of the derivatives in the discrete part of (5.40), the closure of D(bT ) is not a
function space and there is no positive Borel measure generating this inner product.
SELF-ADJOINT OPERATORS IN EXTENDED HILBERT SPACES
Example 3.5 For our last example, we consider a variation of the last two examples.
define ti, xi ∈ D(T1) (i = 1, 2) by
t1(u) =(cid:26) √M u near a
x1(u) =(cid:26) 0
u near a
√N u near b
u near b
0
,
,
u near a
t2(u) =(cid:26) 0
√N (u − b) u near b
x2(u) =(cid:26) √M (u − a) u near a
u near b
0
,
31
Indeed,
In this case {t1, t2} is a GKN set for T1 and {(x1, 0), (x2, 0)} is a GKN set for bT0. Moreover, a
calculation shows
Ωx = (x′(a)M, x(b)N ).
With z = (z1, z2), the two boundary conditions
yield
0 = [(x, z), (x1, 0)]H⊕W = −x′(b)√N + z2√N
0 = [(x, z), (x2, 0)]H⊕W = −x(a)√M + z1√M
z1 = x(a) and z2 = x′(b).
These calculation show that the operator bT , given by
with domain
bT (x, (x(a), x′(b))) = (−x′′, B(x(a), x′(b)) − (x′(a)M, x(b)N ))
D(bT ) =(cid:8)(x, (x(a), x′(b))) x ∈ D(T1)(cid:9) ,
is self-adjoint in H ⊕ W. For each x, y ∈ D(bT ), the 'mixed' inner product in H ⊕ W reduces to
x′(b)y′(b)
x(a)y(a)
x(u)y(u)du +
+
M
.
N
hx, yiH⊕W =Z b
a
As in the last example, no positive Borel measure generates this inner product and the closure of
D(bT ) in the topology from h·,·iH⊕W is not a function space.
References
[1] N. I. Akhiezer and I. M. Glazman, Theory of Linear Operators in Hilbert space, Parts I and II, Scottish Academic
Press, Pitman Advanced Publishing Program, London, U.K. 1981.
[2] N. Dunford and J. T. Schwartz, Linear Operators, Part II, John Wiley Publishers, New York, 1963.
[3] W. N. Everitt, A. M. Krall, and L. L. Littlejohn, On some properties of the Legendre type differential expression,
Quaestiones Math., 13(1), 1990, 83-116.
[4] W. N. Everitt and L. L. Littlejohn, Differential operators and the Legendre type polynomials, Differential and
Integral Equations, 1(1), 1988, 97-116.
[5] W. N. Everitt and L. L. Littlejohn, Orthogonal polynomials and spectral theory: a survey, Proceedings of the III
International Symposium on Orthogonal Polynomials and Applications, Erice, Italy, 1990. IMACS Annals on
Computing and Applied Mathematics 9(1991), 21-55; J. C. Baltzer AG, Basel, Switzerland, 1991, 21-55.
[6] W. N. Everitt, L. L. Littlejohn and R. Wellman, Orthogonal polynomial solutions of linear ordinary differential
equations, Proceedings of the Fifth International Symposium on Orthogonal Polynomials, Special Functions and
their Applications (Patras, 1999), J. Comput. Appl. Math. 133 (2001), no. 1-2, 85–109.
[7] W. N. Everitt and L. Markus, The Glazman-Krein-Naimark Theorem for Ordinary Differential Operators, Op-
erator Theory: Advances and Applications 98(1997), 118-130.
[8] W. N. Everitt and L. Markus, Complex Symplectic Geometry with Applications to Ordinary Differential Opera-
tors, Trans. Amer. Math. Soc., 351(12), 1999, 4905-4945.
[9] W. N. Everitt and L. Markus, Boundary Value Problems and Symplectic Algebra for Ordinary Differential and
Quasi-Differential Operators, Math. Surveys Monogr., Volume 61, American Mathematical Society, Providence,
Rhode Island, 1999.
32
LANCE L. LITTLEJOHN AND RICHARD WELLMAN
[10] W. N. Everitt and L. Markus, Complex Symplectic Spaces and Boundary Value Problems, Bull. Amer. Math.
Soc. (N.S.) 42 (2005), no. 4, 461–500.
[11] V. Guillemin and S. Sternberg, Symplectic techniques in physics, Cambridge University Press, 1990.
[12] A. M. Krall, On Orthogonal Polynomials Satisfying Fourth Order Differential Equations, Proc. Roy. Soc. Edinb.
(A), 1981, 271-288.
[13] H. L. Krall, Certain Differential Equations for Tchebycheff polynomials, Duke Math J., 4. 1938, 705-718
[14] H. L. Krall, Orthogonal polynomials satisfying a Certain fourth order differential equation, The Pennsylvania
State College Studies, No. 6, The Pennsylvania State College, State College, PA. 1940.
[15] M. A. Naimark, Linear Differential Operators, Part II, Ungar Publishing Co., New York, 1968.
[16] M. Reed and B. Simon, Methods of Modern Mathematical Physics I. Functional Analysis, Academic Press, New
York, 1972.
[17] W. Rudin, Functional Analysis (2nd edition), McGraw-Hill Publishers, New York (1991).
[18] M. Stone, Linear transformations in Hilbert space, American Mathematical Society Colloquium Publications 15,
Providence, RI, 1932.
[19] J. Weidmann, Linear operators in Hilbert space, Springer-Verlag, Heidelberg, 1980.
[20] A. Zettl, Formally self-adjoint quasi-differential operators, Rocky Mountain J. Math., 5(1975), 453-474.
[21] A. Zettl, Sturm-Liouville theory, Mathematical Surveys and Monographs, 121. American Mathematical Society,
Providence, RI, 2005.
Department of Mathematics, Baylor University, Waco, TX 76798-7328, Department of Mathematics
and Computer Science, Westminster College, Salt Lake City, UT 84105
E-mail address: [email protected]
|
1707.05669 | 3 | 1707 | 2018-01-17T17:16:52 | Quantum stochastic Lie-Trotter product formula II | [
"math.FA",
"math-ph",
"math-ph",
"math.PR"
] | A natural counterpart to the Lie-Trotter product formula for norm-continuous one-parameter semigroups is proved, for the class of quasicontractive quantum stochastic operator cocycles whose expectation semigroup is norm continuous. Compared to previous such results, the assumption of a strong form of independence of the constituent cocycles is overcome. The analysis is facilitated by the development of some quantum Ito algebra. It is also shown how the maximal Gaussian component of a quantum stochastic generator may be extracted - leading to a canonical decomposition of such genarators, and the connection to perturbation theory is described. Finally, the quantum Ito algebra is extended to quadratic form generators, and a conjecture is formulated for the extension of the product formula to holomorphic quantum stochastic cocycles. | math.FA | math |
QUANTUM STOCHASTIC
LIE–TROTTER PRODUCT FORMULA II
J. MARTIN LINDSAY
Abstract. A natural counterpart to the Lie–Trotter product formula for norm-contin-
uous one-parameter semigroups is proved, for the class of quasicontractive quantum
stochastic operator cocycles whose expectation semigroup is norm continuous. Com-
pared to previous such results, the assumption of a strong form of independence of
the constituent cocycles is overcome. The analysis is facilitated by the development of
some quantum Ito algebra. It is also shown how the maximal Gaussian component of
a quantum stochastic generator may be extracted - leading to a canonical decomposi-
tion of such generators, and the connection to perturbation theory is described. Finally,
the quantum Ito algebra is extended to quadratic form generators, and a conjecture is
formulated for the extension of the product formula to holomorphic quantum stochastic
cocycles.
Contents
Introduction
1. Quantum Ito algebra
2. Quantum stochastics
3. Product Formula
4. Maximal Gaussian component of a QS generator
5. Perturbation of QS cocycles
6. Holomorphic QS Cocycles
References
1
2
8
11
15
18
19
22
Introduction
The Lie product formula in a unital Banach algebra states that
(ea/neb/n)n → ea+b as n → ∞.
Trotter extended this to C0-semigroups on a Banach space where it holds under com-
patibility assumptions on the generators, convergence being in the strong operator sense
([Tro], see e.g. [Dav]). It has been further refined, notably by Chernoff ([Che]) and Kato
([Kat]). These product formulae are widely used in mathematical physics and probability
theory – for example in establishing positivity preservation of semigroups, and they have
an intimate connection to Feynman–Kac formulae (see e.g. [ReS]). Given that quantum
stochastic cocycles may be analysed from their associated semigroups ([LW3]), it is nat-
ural to seek product formulae in this context. Further motivation comes from the fact that
such cocycles are quantum counterparts to stochastic semigroups in the sense of Skoro-
hod ([Sko]). Product formulae have been obtained in a variety of quantum stochastic
settings ([PaS], [LS1,3], [DLT], [DGS]). The earliest of these dates from before the advent
of quantum stochastic calculus ([HuP]).
In all of these works the constituent cocycles
2000 Mathematics Subject Classification. 46L53 (primary); 81S25, 47D06, 46N50 (secondary).
Key words and phrases. Lie–Trotter product formula, quantum stochastic cocycle, one-parameter
semigroup, series product, concatenation product, quantum Ito algebra, quantum stochastic analysis.
1
2
MARTIN LINDSAY
enjoy a strong independence property, namely their respective noise dimension spaces are
mutually orthogonal.
In this paper a Lie–Trotter product formula is established for quasicontractive element-
ary (i.e. Markov-regular) quantum stochastic operator cocycles, with no independence
assumption on the driving quantum noise. It is a direct generalisation of the product for-
mula proved in [LS1], and is proved by quite different means. Properties of the composition
law on the set of quantum stochastic generators that is realised by the stochastic product
formula established here are also elucidated. Known in the setting of quantum control
theory as the series product ([GoJ]), it is more commonly associated with the perturbation
of quantum stochastic cocycles ([EvH]). The composition of stochastic generators also
corresponds to the operator product (i.e. standard composition) of the generators of the
quantum random walks whose scaled embeddings approximate the constituent cocycles
([BGL]). Analysis of this composition leads to left and right series decompositions of a
quantum stochastic generator. A decomposition for such generators in terms of the so-
called concatenation product is also given; this yields the generator's maximal Gaussian
part.
It is conjectured here that, as in the case of orthogonal noises ([LS3]), the more gen-
eral Lie–Trotter product formula given in this paper has an extension to the class of
holomorphic quantum stochastic cocycles ([LS2]). By contrast, without orthogonality of
noise dimension spaces there seems to be no sensible formulation of a Lie–Trotter product
formula for quantum stochastic mapping cocycles.
The plan of the paper is as follows. In Section 1 some quantum Ito algebra is developed,
for studying the series product on the class of stochastic generators considered here. In
Section 2 the relevant quantum stochastic analysis is recalled. The quantum stochastic
Lie–Trotter product formula is proved in Section 3. In Section 4 the maximal Gaussian
component of a quantum stochastic generator is extracted by means of the concatenation
product. In the short Section 5, the connection to perturbation theory is described, and
in Section 6 the quantum Ito algebra is extended to quadratic form generators and a
conjecture for quasicontractive holomorphic quantum stochastic cocycles is formulated.
Notation. For a vector-valued function f : R+ → V and subinterval J of R+, fJ
denotes the function R+ → V which agrees with f on J and vanishes elsewhere. For
Hilbert spaces h and h′, B(h; h′) denotes the space of bounded operators from h to h′ and
B(h; h′)1 denotes its closed unit ball, abbreviated to B(h) and B(h)1 respectively when
h′ = h. For an operator T ∈ B(h), its real and imaginary parts are denoted Re T and
Im T respectively, thus T is dissipative if and only if Re T 6 0. The selfadjoint part of
a subset A of an involutive space is denoted Asa. The predual of B(h), that is the space
of ultraweakly continuous linear functionals on B(h), is denoted B(h)∗. Algebraic and
ultraweak tensor products are denoted by ⊗ and ⊗ respectively and, for vectors ζ, η ∈ h,
ωζ,η ∈ B(h)∗ denotes the functional given by T 7→ hζ, T ηi. The symbol ⊂⊂ is used to
denote finite subset.
1. Quantum Ito algebra
For this section take Hilbert spaces h and H. The block matrix decomposition enjoyed
by operators in B(h⊕H) is frequently appealed to below. With respect to the distinguished
orthogonal projection
∆ := P{0h}⊕H =(cid:20) 0h
0h;H
0H;h
IH (cid:21) ∈ B(h ⊕ H),
the composition law on B(h ⊕ H) given by
F1 ✁ F2 := F1 + F1∆F2 + F2,
QS LIE–TROTTER PRODUCT FORMULA
3
with useful alternative asymmetric expressions
F1(I + ∆F2) + F2 and F1 + (I + F1∆)F2,
has both
and
F1 + F2 + F3 +(cid:0)F1∆F2 + F1∆F3 + F2∆F3(cid:1) + F1∆F2∆F3,
F1 ✁ F3 + (I + F1∆) F2 (∆F3 + I),
(1.1a)
(1.1b)
as common expression for F1 ✁ (F2 ✁ F3) and (F1 ✁ F2) ✁ F3. Moreover the composition
✁ has 0h⊕H as identity element, and the operator adjoint as involution since
(F1 ✁ F2)∗ = F ∗
2 ✁ F ∗
1 .
The notation ✁ is taken from the quantum control theory literature, where the composi-
tion is called the series product.
Let b(h ⊕ H) := (B(h ⊕ H), ✁) denote the resulting *-monoid (i.e. involutive semigroup-
with-identity), let β ∈ R, and consider the following subsets of b(h ⊕ H):
qcβ(h ⊕ H) := c(h ⊕ H) + β∆⊥,
qc(h ⊕ H) := c(h ⊕ H) + R+∆⊥,
z(h ⊕ H) :=(cid:8)K ⊕ 0H : K ∈ B(h)(cid:9) = ∆⊥ b(h ⊕ H) ∆⊥,
c(h ⊕ H) :=(cid:8)F ∈ b(h ⊕ H) : F ∗ ✁ F 6 0},
i(h ⊕ H) :=(cid:8)F ∈ b(h ⊕ H) : F ∗ ✁ F = 0},
i(h ⊕ H)∗ :=(cid:8)F ∗ : F ∈ i(h ⊕ H)},
u(h ⊕ H) := i(h ⊕ H) ∩ i(h ⊕ H)∗,
and for F ∈ qc(h ⊕ H) set
β0(F ) := inf{β ∈ R : F − β∆⊥ ∈ c(h ⊕ H)}.
These classes are relevant to the characterisation of the stochastic generators of quantum
stochastic cocycles which are respectively contractive, quasicontractive (with exponential
growth bound β0(F )), isometric, coisometric and unitary (see Theorem 2.1 below). Note
that, for F ∈ qc(h ⊕ H),
F ∈ qcβ(h ⊕ H) if and only if β > β0(F ).
Remarks. Further classes are relevant to the characterisation of quantum stochastic cocy-
cles which are nonnegative, selfadjoint, partially isometric or projection-valued ([Wil]).
The characterisation of the generators of 'pure-noise' (or 'local') nonnegative contraction
cocycles (for which h = C) plays an important role in the identification of the minimal
dilation of a quantum dynamical semigroup ([Bha]).
Let F1, F2 and F in b(h ⊕ H) have respective block matrix formsh K1 M1
and(cid:2) K M
L Q−I(cid:3). Then
L2 Q2−Ii
L1 Q1−Ii,h K2 M2
F1 ✁ F2 =(cid:20)K1 + K2 + M1L2 M1Q2 + M2
Q1Q2 − I (cid:21) , and
F ∗ ✁ F =(cid:20)K ∗ + K + L∗L L∗Q + M
Q∗Q − I(cid:21) ,
M ∗ + Q∗L
L1 + Q1L2
(1.2)
4
MARTIN LINDSAY
moreover, for Z ∈ z(h ⊕ H), β ∈ R and X ∈ b(h ⊕ H),
(F + Z)∗ ✁ (F + Z) = Z ∗ + F ∗ ✁ F + Z
(F − β∆⊥)∗ ✁ (F − β∆⊥) = F ∗ ✁ F − 2β∆⊥, and
F ∗ ✁ X ✁ F = F ∗ ✁ F + (I + ∆F )∗X(I + ∆F ).
These identities imply the following relations:
(G ✁ F )∗ ✁ (G ✁ F )
= F ∗ ✁ F + G∗ + G if G ∈ z(h ⊕ H),
6 F ∗ ✁ F + 2β∆⊥
= F ∗ ✁ F
if G ∈ qcβ(h ⊕ H),
if G ∈ i(h ⊕ H).
The basic algebraic properties of these subsets of b(h ⊕ H) are collected in the following
proposition; using the above observations, their proof is straightforward. Recall that an
operator T ∈ B(h) is dissipative if it satisfies Re T 6 0, that is
Rehu, T ui 6 0
(u ∈ h).
Proposition 1.1. In the *-monoid b(h ⊕ H) the following hold.
(a) Its group of invertible elements is given by
b(h ⊕ H)× =(cid:8)(cid:2) ∗ ∗
the identity element being 0 =(cid:2) 0 0
(cid:20)M Q−1L − K −M Q−1
∗ Q(cid:3) − ∆ ∈ b(h ⊕ H) : Q ∈ B(H)×(cid:9);
0 I(cid:3) − ∆, and the inverse of(cid:2) K M
L Q(cid:3) − ∆ being
I (cid:21) Q−1(cid:2)−L I(cid:3) −(cid:20)K 0
I(cid:21) .
0
−Q−1L
Q−1 (cid:21) − ∆ =(cid:20)−M
(b) Its centre is z(h ⊕ H).
(c) Denoting the class of dissipative operators on h by D(h),
qc(h ⊕ H) = c(h ⊕ H) + z(h ⊕ H), and
c(h ⊕ H) ∩ z(h ⊕ H) = {T ⊕ 0H : T ∈ D(h)}.
(d) qc(h⊕ H), c(h⊕ H), i(h⊕ H) and u(h⊕ H) are submonoids; their groups of invertible
elements are given by
i(h ⊕ H)× = c(h ⊕ H)× = u(h ⊕ H), and
qc(h ⊕ H)× = u(h ⊕ H) + z(h ⊕ H).
Moreover,
i(h ⊕ H) ∩ z(h ⊕ H) = u(h ⊕ H) ∩ z(h ⊕ H) = i z(h ⊕ H)sa.
Remark. Clearly u(h ⊕ H) is also closed under taking adjoints, and so is a sub-*-monoid
of b(h ⊕ H).
In fact qc(h ⊕ H) and c(h ⊕ H) are sub-*-monoids too; this is not immediately obvious.
It follows from Part (e) of Theorem 1.3 below, but there are now also direct proofs. The
one given below is based on an elegant argument of Wills, arising as a biproduct of his
analysis of partially isometric quantum stochastic cocycles ([Wil]).
Proposition 1.2. Let F ∈ b(h ⊕ H) and T ∈ z(h ⊕ H)sa. Then
(∆F + I)∗(F ✁ F ∗)(∆F + I) = F ∗ ✁ F + (F ∗ ✁ F ) ∆ (F ∗ ✁ F ),
(1.3)
and
F ∗ ✁ F 6 T if and only if F ✁ F ∗ 6 T.
QS LIE–TROTTER PRODUCT FORMULA
5
Proof. By the associativity of ✁, setting F1 = F ∗, F2 = F ✁ F ∗ and F3 = F in the
expression (1.1b) for F1 ✁ F2 ✁ F3 we see that
F ∗ ✁ F + LHS(1.3) = F ∗ ✁ (F ✁ F ∗) ✁ F
= (F ∗ ✁ F ) ✁ (F ∗ ✁ F ) = RHS(1.3) + F ∗ ✁ F,
so identity (1.3) holds. Suppose now that F ✁ F ∗ 6 T . Then, since ∆T = 0 = T ∆,
LHS(1.3) 6 T and so, since ∆ > 0 and F ∗ ✁ F is selfadjoint,
F ∗ ✁ F = (1.3) − (F ∗ ✁ F )∆(F ∗ ✁ F ) 6 (1.3) 6 T.
The converse implication follows by exchanging F and F ∗.
Consider the following possible block matrix forms for F ∈ b(h ⊕ H):
2 (L∗L + A2)
L
(cid:20)βI + iH − 1
(cid:20) βI + iH − 1
2 (M M ∗ + B2)
−CM ∗ − (I − CC ∗)1/2E∗B
(cid:21) , and
−L∗C − AD(I − C ∗C)1/2
C − I
M
C − I(cid:21) ,
(cid:3)
(1.4a)
(1.4b)
in which β ∈ R, H ∈ B(h)sa, A, B ∈ B(h)+, C ∈ B(H)1 and D, E ∈ B(H; h)1, so that
H = Im K.
Theorem 1.3 (Cf. [LW1], [GL+]). Let β ∈ R and F ∈ b(h ⊕ H).
(a) The following are equivalent:
(i) F ∈ qcβ(h ⊕ H).
(ii) F has block matrix form (1.4a).
(iii) F has block matrix form (1.4b).
(iv) F ∗ ∈ qcβ(h ⊕ H).
(b) F ∈ i(h ⊕ H) if and only if F has block matrix form (1.4a), with β = 0, A = 0
and C isometric.
(c) F ∈ i(h ⊕ H)∗ if and only if F has block matrix form (1.4b), with β = 0, B = 0
and C coisometric.
(d) Let H1 ⊕ H2 be an orthogonal decomposition of H. Then, with respect to the
inclusion J1 : h ⊕ H1 → h ⊕ H,
J ∗
1 qcβ(h ⊕ H)J1 = qcβ(h ⊕ H1).
(1.5)
(e) Set H′ := H⊕(H⊕h). Then, in terms of the inclusion J : h⊕H → (h⊕H)⊕(H⊕h) =
h ⊕ H′,
c(h ⊕ H) = J ∗u(h ⊕ H′)J.
Proof. The proof exploits the fact that, for an operator T ∈ B(h ⊕ H), T > 0 if and only
0H;H⊕h(cid:21) ∈ B(h ⊕ H; h ⊕ H′).
i where (X, Z, V ) ∈ B(h)+ × B(H)+ ×
X
Z 1/2V ∗X 1/2
X 1/2V Z 1/2
Z
0H1;H2(cid:21) ∈ B(h ⊕ H1; h ⊕ H), and J = Ih ⊕(cid:20) IH
Remark. In block matrix form
J1 = Ih ⊕(cid:20) IH1
if T has block matrix formh
B(H; h)1.
β = 0 and, setting S := (I − C ∗C)1/2,
−G∗ ✁ G =(cid:20) A2
SD∗A
ADS
S2 (cid:21) > 0,
(a) Set G = F − β∆⊥. Suppose first that (ii) holds. Then G is given by (1.4a) with
6
MARTIN LINDSAY
so G ∈ c(h ⊕ H). Thus (ii) implies (i). Conversely, suppose that (i) holds and let(cid:2) K M
L C−I(cid:3)
be the block matrix decomposition of G. Then G ∈ c(h ⊕ H) so
(cid:20)K ∗ + K + L∗L M + L∗C
C ∗C − I(cid:21) = G∗ ✁ G 6 0.
M ∗ + C ∗L
Thus (K ∗ + K + L∗L) 6 0, kCk 6 1 and, for some contraction operator D,
M = −L∗C − [−(K ∗ + K + L∗L)]1/2D(I − C ∗C)1/2
so G is given by (1.4a) with β = 0 and H = Im K. Thus (i) implies (ii).
Therefore (i) and (ii) are equivalent; taking adjoints we see that (iv) and (iii) are
equivalent too. The equivalence of (i) and (iv) follows from Proposition 1.2.
(b) This is an immediate consequence of (1.2).
(c) This follows from Part (b), by taking adjoints.
(d) For F1 ∈ qcβ(h ⊕ H1),
F1 = J ∗
1 F J1, where F =(cid:20)F1 0
0(cid:21) ∈ qcβ(h ⊕ H),
0
so RHS ⊂ LHS in (1.5). For the reverse inclusion, setting ∆1 := 0h ⊕IH1 and ∆ := 0h ⊕IH,
∆ − J1∆1J ∗
1 = 0h ⊕ IH1 ⊕ IH2 − 0h ⊕ IH1 ⊕ 0H2 = 0h ⊕ 0H1 ⊕ IH2 > 0.
Thus, for F ∈ c(h ⊕ H), the operator F1 := J ∗
1 F J1 is in c(h ⊕ H1) since
1 F ∗J1 + J ∗
1 ✁ F1 = J ∗
F ∗
1 c(h ⊕ H)J1 ⊂ c(h ⊕ H1) and so, since J ∗
1 F J1 + J ∗
1 F ∗J1∆1J ∗
Therefore J ∗
1 (β ∈ R), LHS ⊂ RHS.
(e) In view of (d), it suffices to show that J ∗u(h ⊕ H′)J ⊃ c(h ⊕ H). Accordingly, let
F ∈ c(h ⊕ H). Then, by what we have proved already, F has block matrix form (1.4a)
with β = 0. Setting
1 β∆⊥J1 = β∆⊥
1 F J1 6 J ∗
1 (F ∗ ✁ F )J1 6 0.
F ′ :=
iH − 1
2 (L∗L + A2) −(L∗C + ADS) L∗T − ADC ∗ −AR
L
C − I
−T
D∗A
RA
S
0
C ∗ − I
0
,
0
0
0
where R = (I − DD∗)1/2, S = (I − C ∗C)1/2, and T = (I − CC ∗)1/2, it is now easily
verified that F ′ ∈ u(h ⊕ H′). Since J ∗F ′J = F , (e) follows and the proof is complete. (cid:3)
Remarks. The dilation property (e) is effectively proved in [GL+], under the assumption
that h and H are separable. If H = h ⊗ k for a Hilbert space k (as it is in the application
to QS analysis, where k is the noise dimension space) then, in (e), H′ = h ⊗ k′ where
k′ = k ⊕ k ⊕ C.
In terms of its block matrix form (cid:2) K M
L C−I(cid:3), and bound β0 := β0(F ), the equivalent
conditions for F ∈ b(h ⊕ H) to be in qc(h ⊕ H) read respectively as follows:
M ∗ + C ∗L
(cid:20)2(Re K − β0Ih) + L∗L M + L∗C
(cid:20)2(Re K − β0Ih) + M M ∗ L∗ + M C ∗
C ∗C − IH(cid:21) 6 0, and
CC ∗ − IH(cid:21) 6 0.
L + CM ∗
and set β0 = β0(F ). Then, setting
Proposition 1.4 (Left and right series decomposition). Let F =(cid:2) K M
L C−I(cid:3) ∈ qc(h ⊕ H)
0(cid:21) ,
F1 =(cid:20)β0Ih + i Im K 0
3 :=(cid:20)Re K + 1
0 (cid:21) , F r
2 =(cid:20)− 1
C − I (cid:21) , and F r
2 L∗L −L∗
L
2 L∗L − β0Ih M + L∗C
0
3 =(cid:20)− 1
2 :=(cid:20)Re K + 1
2 M M ∗ M
−M ∗
0(cid:21) , F ℓ
C − I(cid:21) ,
2 M M ∗ − β0Ih
L + CM ∗
F ℓ
0
0
QS LIE–TROTTER PRODUCT FORMULA
7
the following hold:
F1 ∈ z(h ⊕ H), F ℓ
2 , F r
3 ∈ u(h ⊕ H), F r
2 , F ℓ
3 , ∈ c(h ⊕ H),
and
F1 ✁ F ℓ
2 ✁ F ℓ
3 = F = F1 ✁ F r
2 ✁ F r
3 ,
Proof. It follows from Theorem 1.3, and the above block matrix inequalities characterising
membership of qc(h ⊕ H), that
F ℓ
2 , F r
3 ∈ u(h ⊕ H) and F ℓ
2 ∈ c(h ⊕ H).
3 , F r
The two series decompositions follow from the identities
F1∆F ℓ
2 = 0, F ℓ
2 ∆F ℓ
The rest is clear.
3 =(cid:20)0 L∗ − L∗C
0
0
(cid:21) , F1∆F r
2 = 0 and F r
2 ∆F r
3 =(cid:20)
0
M ∗ − CM ∗ 0(cid:21) .
0
(cid:3)
Remarks. (i) The left and right series decompositions are related via the adjoint operation
as follows:
(F ∗)1 = (F1)∗, (F ∗)ℓ
2 = (F r
3 )∗ and (F ∗)ℓ
3 = (F r
2 )∗.
(ii) Let F =(cid:2) K M
L C−I(cid:3) ∈ qc(h ⊕ H) and set H := Im K. Then F ∈ i(h ⊕ H) if and only
if F has left series decomposition
with C isometric, whereas F ∈ i(h ⊕ H)∗ if and only if F has right series decomposition
with C coisometric.
Now suppose that H has an orthogonal decomposition H1 ⊕ H2. Define injections
0
i(cid:20)H 0
i(cid:20)H 0
0(cid:21) ✁(cid:20)− 1
0(cid:21) ✁(cid:20)0
0
0
0
2 L∗L −L∗
L
2 M M ∗ M
−M ∗
0 (cid:21) ✁(cid:20)0
0 C − I(cid:21) ,
0(cid:21) ,
0 C − I(cid:21) ✁(cid:20)− 1
L N(cid:21) 7→ F ⊕ 0H2 =
L N(cid:21) 7→ Σ(cid:0)F ⊕ 0H1(cid:1) =
ι : b(h ⊕ H1) → b(h ⊕ H), F =(cid:20)K M
ι′ : b(h ⊕ H2) → b(h ⊕ H), F =(cid:20)K M
K M 0
L N 0
0
0
0 , and
L 0 N ,
K 0 M
0
0
0
(1.7)
Σ being the sum-flip map B(h ⊕ H2 ⊕ H1) → B(h ⊕ H), and define the composition
b(h ⊕ H1) × b(h ⊕ H2) → b(h ⊕ H),
(F1, F2) 7→ F1 ⊞ F2 := ι(F1) + ι′(F2),
known as the concatenation product in quantum control theory ([GoJ]). Thus
L2 N2(cid:21) =
(cid:20)K1 M1
L1 N1(cid:21) ⊞(cid:20)K2 M2
K1 + K2 M1 M2
0
N1
0 N2
L1
L2
.
Note that
In view of the identity
ι(F1) = F1 ⊞ 0H2 and ι′(F2) = 0H1 ⊞ F2.
ι(F1) ∆ ι′(F2) = 0h⊕H = ι′(F2) ∆ ι(F1),
for F1 ∈ b(h ⊕ H1) and F2 ∈ b(h ⊕ H2), the concatenation product is effectively a special
case of the series product:
F1 ⊞ F2 = ι(F1) ✁ ι′(F2).
(1.8)
8
MARTIN LINDSAY
We end this section with a significant representation of the quantum Ito algebra. It is
relevant to the realisation of QS cocycles as time-ordered exponentials ([Hol]), and also
to the convergence of a class of scaled quantum random walks to QS cocycles ([BGL]).
Set S := Ih⊕H⊕h + A, where A is the following subalgebra of B(h ⊕ H ⊕ h):
thus S consists of the elements of B(h ⊕ H ⊕ h) having block matrix form(cid:20) Ih ∗ ∗
respect to (operator composition and) the involution given by
0 0 Ih(cid:21). With
0 ∗ ∗
(cid:8)T ∈ B(h ⊕ H ⊕ h) : Ph⊕H⊕{0}T = T = T P{0}⊕H⊕h(cid:9);
T 7→ T ⋆ := ΞT ∗Ξ, where Ξ :=
0 ,
= Ih⊕H⊕h +
L Q − I(cid:21) 7→
(cid:20)K M
Ih M K
0 Q L
0
Ih
0 M K
0 Q − I L
0
0
IH
0
0
0
Ih
Ih
0
0
0
0
S is a sub-*-monoid of B(h ⊕ H ⊕ h), and the following is readily verified.
Proposition 1.5 ([Hol], [Bel]). The prescription
defines an isomorphism of *-monoids φ : b(h ⊕ H) → S.
2. Quantum stochastics
L2(R+; k). The quantum Ito algebra developed in the previous section is applied below
For the rest of the paper fix Hilbert spaces h and k, and set bk := C ⊕ k and K :=
with h = h and H = k ⊗ h, so that h ⊕ H =bk ⊗ h. In this context, the operator
∆ = ∆k := P{0C}⊕k =(cid:20)0
0
0 Ik(cid:21) ∈ B(bk)
is ubiquitous; below it is freely ampliated.
In this section we collect the quantum stochastic (QS) facts needed below. For more
detail, see [L]; for further background, see [Par] and [Mey]. Let F denote the symmetric
Fock space over K. We use normalised exponential vectors
(g) := e−kgk2/2ε(g) where ε(g) :=(cid:0)(n!)−1/2g⊗n(cid:1)n>0
in terms of which the Fock–Weyl operators are the unitary operators on F determined by
the identity
W (f )(g) = e−i Imhf,gi(f + g)
(f, g ∈ K),
and the second quantisation operators are the contractions from F to F ′ determined by
the identity
Γ(C)ε(g) := ε(Cg)
(2.1)
where F ′ is the symmetric Fock space over a Hilbert space K′. Note that Γ(C) is isometric
if C is, and Γ(C ∗) = Γ(C)∗. There is a useful family of slice maps:
Ω(g′, g) := idB(h;h′) ⊗ ω(g′),(g) : B(h; h′)⊗ B(F) → B(h; h′)
(g′, g ∈ K),
(C ∈ B(K; K′)1, g ∈ K),
(g ∈ K),
amongst which E := Ω(0, 0) is referred to as the vacuum expectation. (The Hilbert spaces
h and h′ are determined by context.) Moreover, the action of the slice maps extends to
the class of unbounded operators from h⊗ F to h′ ⊗ F whose domains include h⊗ E, where
E := Lin{(g) : g ∈ K}. The following obvious identity is exploited below:
We also need two families of endomorphisms of (the von Neumann algebra) B(F):
Ω(g′, g)(T ) = E(cid:2)(Ih′ ⊗ W (g′))∗ T (Ih ⊗ W (g))(cid:3).
(2.2)
ρt : T 7→ RtT Rt and σt : T 7→ IF[0,t[ ⊗ StT S∗
t
(t ∈ R+),
QS LIE–TROTTER PRODUCT FORMULA
9
where Rt := Γ(rt) and St := Γ(st) for the unitary operators rt and st defined as follows
rt : K → K,
(rtf )(s) =(f (t − s)
f (s)
if s ∈ [0, t[
if s ∈ [t, ∞[
,
st : K → K
[t,∞[,
(stf )(s) = f (s − t),
where K[t,∞[ := L2([t, ∞[; k). The time-reversal maps (cid:0)ρt(cid:1)t>0 are obviously involutive:
ρ2
t = idB(F ), and the time-shift maps (σt)t∈R+ form a semigroup, known as the CCR flow
of index k. Both are freely ampliated to act on B(h; h′)⊗ B(F), for Hilbert spaces h and
h′. For use below, note the identity
r(g′[r,∞[), s∗
Ω(cid:0)g′
[r,∞[, g[r,∞[(cid:1) ◦ σr = Ω(cid:0)s∗
Let X i =(cid:0)X i
0 + Λt(H i)(cid:1)t>0 for a QS integrand (hi, hi−1)-process H i and bounded op-
erator X i
0 ∈ B(hi; hi−1) ⊗ IF , where i = 1, 2. Suppose that all of the processes H 1, H 2, X 1
and X 2 are bounded, that is they consist of bounded operators, and H 1 and H 2 are
strongly continuous. Then the quantum Ito product formula ([HuP]) reads
(g, g′ ∈ K, r ∈ R+).
(2.3)
r(g[r,∞[)(cid:1)
X 1
0 X 2
t X 2
t = X 1
with ∆ abbreviating ∆ ⊗ Ih1⊗F .
0 + Λt(H) where H =(cid:0)H 1
s(cid:1)s>0, (2.4)
Let X =(cid:0)Λt(H)(cid:1)t>0 for a bounded QS integrand (h1, h2)-process H with block matrix
s ) + (Ibk ⊗ X 1
s (Ibk ⊗ X 2
∗ ∗ ], then
form [ K ∗
s + H 1
s ∆H 2
s )H 2
ds E[Ks]v
(v ∈ h1, t ∈ R+),
(2.5)
E[Xt] v =Z t
0
and, if the process X is bounded then, for bounded operators R ∈ B(h2; h3) ⊗ IF and
S ∈ B(h0; h1) ⊗ IF ,
RΛt(H)S = Λt(cid:0)(Ibk ⊗ R)H·(Ibk ⊗ S)(cid:1)
(t ∈ R+).
(2.6)
A bounded QS (left) cocycle on h, with noise dimension space k, is a bounded process
V on h which satisfies
V0 = Ih⊗F and Vs+t = Vsσs(Vt)
(s, t ∈ R+).
It is a QS right cocycle if instead it satisfies Vs+t = σs(Vt)Vs for s, t ∈ R+. If V is a QS
left cocycle then, V ∗ := (V ∗
t )t>0 and V r :=(cid:0)ρt(Vt)(cid:1)t>0 define QS right cocycles, and
(2.7)
V ♯ :=(cid:0)ρt(V ∗
t ) = ρ(Vt)∗(cid:1)t>0
defines a QS left cocycle, called the dual cocycle of V ([Jou]). This said, we work exclusively
with QS left cocycles in this paper. Following standard terminology of semigroup theory
([HiP]),
β0(V ) := inf{β ∈ R : supt>0 ke−βtVtk < ∞}
If V is strongly continuous then
is referred to as the exponential growth bound of V .
β0(V ) < ∞; V is called quasicontractive if, for some β ∈ R, the QS cocycle (e−βtVt)t>0
is contractive. When V is locally uniformly bounded, it is called elementary (or Markov
regular ) if its expectation semigroup(cid:0)E[Vt](cid:1)t>0 is norm continuous.
Note that, given a QS cocycle V , the two-parameter family Vr,t := σr(Vt−r) satisfies
V0,t = Vt (t ∈ R+), the evolution equation
Vr,r = Ih⊗F and Vr,t = Vr,sVs,t
(t > s > r),
and the biadaptedness property
Vr,t ∈(cid:0)B(h) ⊗ I[0,r[(cid:1)⊗(cid:0)B(F[r,t[) ⊗ I[t,∞[(cid:1)
(t > r > 0).
10
MARTIN LINDSAY
For use below, note that in the notation (2.2), these QS evolutions satisfy the time-
covariance identity
Ω(c′
[r,t[, c[r,t[)(Vr,t) = Ω(c′
[0,t−r[, c[0,t−r[)(Vt−r)
by (2.3), and further evolution identity
for c′, c ∈ k, g′, g ∈ L2
Ω(g′
[r,t[, g[r,t[)(Vr,t) = Ω(g′
loc(R+; k) and t > s > r > 0.
[r,s[, g[r,s[)(Vr,s)Ω(g′
[s,t[, g[s,t[)(Vs,t)
(2.8a)
(2.8b)
Remark. Note that the relations (2.8a) and (2.8b) also hold when V is a product of such
QS evolutions.
equation dXt = Xt·F dΛt, X0 = I has a unique weakly regular weak solution, denoted X F ,
moreover X F is a quasicontractive, elementary QS cocycle, with exponential growth bound
at most β, strongly satisfying its QS differential equation. The resulting map F 7→ X F is
Theorem 2.1 ([Fag], [LW1,2). ] For F ∈ qcβ(bk⊗ h) where β ∈ R, the (left) QS differential
bijective from qc(bk ⊗ h) to the class of quasicontractive elementary QS cocycles; it restricts
to bijections from c(bk ⊗ h), i(bk ⊗ h) and i(bk ⊗ h)∗ to the respective subclasses of contractive,
isometric and coisometric, elementary QS cocycles, moreover it satisfies X F ∗
The unique operator F associated with a quasicontractive elementary QS cocycle V in
= (X F )♯.
this way is referred to as the stochastic generator of V .
There is a basic class of QS cocycles which plays an important role.
Example 2.2. For each c ∈ k, the (Fock) Weyl cocycle W c is given by
Proposition 2.3. Let F ∈ qc(bk ⊗ h) and G ∈ qc(bk) ⊗ Ih ⊂ qc(bk ⊗ h). Then
Proof. First note that, since X G is of the form (cid:0)Ih ⊗ Vt(cid:1)t>0 for a quasicontractive QS
X F X G = X F ✁G and X G X F = X G✁F .
cocycle V on C,
s ) = F ⊗ Vs = (Ibk ⊗ X G
Therefore, by the quantum Ito product formula (2.4),
(F ⊗ IF )(Ibk ⊗ X G
s )(F ⊗ IF )
(s ∈ R+).
d(X F
t X G
t ) =(cid:2)(Ibk ⊗ X F
t )(F ⊗ IF )(Ibk ⊗ X G
t ) + (Ibk ⊗ X F
t X G
t )(F ⊗ IF )(Ibk ⊗ X G
+ (Ibk ⊗ X F
= X F
t X G
t
· (F ✁ G) dΛt.
t )(G ⊗ IF )
t )(∆G ⊗ IF )(cid:3) dΛt
0 X G
Since X F
solutions of the QS differential equation
0 = I and F ✁ G ∈ qc(bk ⊗ h), uniqueness for strongly continuous bounded
(see Theorem 2.1) implies that X F X G = X F ✁G. Since F ∗ ∈ qc(bk⊗h) and G∗ ∈ qc(bk)⊗Ih,
too. The second identity now follows by duality:
dXt = Xt · (F ✁ G) dΛt, X0 = Ih⊗F
it follows that X F ∗✁G∗
= X F ∗
X G∗
=(cid:0)X F ∗✁G∗(cid:1)♯ =(cid:0)X F ∗
X G∗(cid:1)♯ =(cid:0)X G∗(cid:1)♯ (cid:0)X F ∗(cid:1)♯ = X G X F .
X G✁F = X (F ∗✁G∗)∗
These are QS unitary cocycles (both left and right), and X = W c satisfies
moreover X is elementary with stochastic generator
W c :=(cid:0)Ih ⊗ W (c[0,t[)(cid:1)t>0
(c ∈ k).
X r = X and X ♯ = W −c = X ∗,
Fc :=(cid:20)− 1
2 kck2 −hc
ci
0 (cid:21) ⊗ Ih ∈ B(bk ⊗ h).
QS LIE–TROTTER PRODUCT FORMULA
11
(cid:3)
Remarks. (i) More generally, if two quasicontractive QS cocycles commute on their initial
space then their (pointwise) product is also a cocycle.
(ii) In the light of the identities
Fc ✁ Fd = Fc+d − i Imhc, di∆⊥ and F−c = F ∗
c
(c, d ∈ k),
Proposition 2.3 contains Weyl commutation relations (see [BrR], or [L]) as a special case.
The connection to quantum random walks mentioned in the introduction is as follows
(for details, in particular the precise meaning of the terminology used, see [BGL]).
Theorem 2.4 ([BGL]). For i = 1, 2, let Fi ∈ qc(bk ⊗ h) and let (Gi(h))h>0 be a family in
B(bk ⊗ h) satisfying
sup(cid:8)kGi(h)kn : n ∈ N, h > 0, nh 6 T(cid:9) < ∞
(cid:0)h−1/2∆⊥ + ∆(cid:1)(cid:0)Gi(h) − ∆⊥(cid:1)(cid:0)h−1/2∆⊥ + ∆(cid:1) → Fi + ∆ as h → 0.
Then, for all ϕ ∈ B(F)∗ and T ∈ R+,
(T ∈ R+), and
sup
06t6T(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)X hhi
t − X F1✁F2
t
(cid:1)(cid:13)(cid:13) → 0 as h → 0,
)t>0 denotes the h-scale embedded left quantum random walk
where, for h > 0, (X hhi
generated by G1(h)G2(h).
t
3. Product Formula
For proving the quantum stochastic Lie–Trotter product formula it is convenient to
define a constant associated with a pair of quantum stochastic generators. Thus, for
F1, F2 ∈ b(bk ⊗ h) with respective block matrix formsh K1 M1
C(F1, F2) : = k∆⊥F1∆⊥k + k∆⊥F2∆⊥k + k∆⊥F1∆k · k∆F2∆⊥k
L1 N1i andh K2 M2
L2 N2i, set
= kK1k + kK2k + kM1kkL2k.
Note that this depends only on the first row of F1 and first column of F2:
C(F1, F2) = C(∆⊥F1, F2∆⊥).
Proposition 3.1. Let Fi ∈ qcβi(bk ⊗ h) with block matrix formh Ki Mi
t − X F1✁F2
t X F2
t
Li Nii, for i = 1, 2. Then
(t ∈ R+).
(cid:13)(cid:13)E(cid:2)X F1
(cid:3)(cid:13)(cid:13) 6 t2 et(β1+β2)C(F1, F2)2
Proof. Set V (1) = X F1, V (2) = X F2 and V = X F1✁F2. Since
e−t(β1+β2)(V (1)
t V (2)
t − Vt) = X G1
t X G2
t − X G1✁G2
t
(t ∈ R+),
where Gi = Fi − βi∆⊥ ∈ c(bk ⊗ h) (i = 1, 2), it suffices to assume that β1 = β2 = 0, so
that F1, F2, F1 ✁ F2 ∈ c(bk ⊗ h), and the QS cocycles V (1), V (2) and V are all contractive.
Fix t ∈ R+. By the quantum Ito product formula (2.4), V (1)V (2) − V = Λ·(H) where
H = H (1) + H (2) + H (3) for the processes given by
:= (Ibk ⊗ V (1)
s
)(F1 ⊗ IF )(Ibk ⊗ V (2)
s
) − (Ibk ⊗ Vs)(F1 ⊗ IF )
s
H (1)
H (2)
H (3)
s
s
:=(cid:0)Ibk ⊗ (V (1)
:= (Ibk ⊗ V (1)
s − Vs)(cid:1)(F2 ⊗ IF )
)(F1∆ ⊗ IF )(Ibk ⊗ V (2)
s V (2)
s
s
)(∆F2 ⊗ IF ) − (Ibk ⊗ Vs)(F1∆F2 ⊗ IF ).
12
MARTIN LINDSAY
Therefore, by (2.5), E[V (1)
processes given by
t V (2)
t − Vt] =R t
0 ds E[Ys] where Y = Y (1) + Y (2) + Y (3) for the
(K1 ⊗ IF )V (2)
s − Vs(K1 ⊗ IF )
Y (1)
s
Y (2)
s
Y (3)
s
s
:= V (1)
:= (V (1)
:= V (1)
s
s V (2)
s − Vs)(K2 ⊗ IF )
(M1 ⊗ IF )(Ik ⊗ V (2)
s
)(L2 ⊗ IF ) − Vs(M1L2 ⊗ IF ).
Note that, since V (1), V (2) and V are contractive,
kE[Yr]k 6 2(cid:0)kK1k + kK2k + kM1kkL2k(cid:1) = 2C(F1, F2)
In turn, the quantum Ito formula, together with the identities (2.5) and (2.6), imply that,
for i = 1, 2, 3 and s ∈ R+, E[Y (i)
r ] for processes given by
(r ∈ R+).
(3.1)
] =R s
s
0 dr E[Z (i)
r − Vr((K1)2 ⊗ IF )+
(K2 ⊗ IF ) − Vr(K1K2 ⊗ IF )+
((K1)2 ⊗ IF )V (2)
(K1 ⊗ IF )V (2)
(M1 ⊗ IF )(Ik ⊗ K1 ⊗ IF )(Ik ⊗ V (2)
r
r
Z (1)
r
:= V (1)
r
V (1)
r
V (1)
r
r
Z (2)
Z (3)
r
:= Yr(K2 ⊗ IF ), and
:= V (1)
r
V (1)
r
V (1)
r
(K1M1 ⊗ IF )(Ik ⊗ V (2)
(M1 ⊗ IF )(Ik ⊗ V (2)
(M1 ⊗ IF )(Ik ⊗ M1 ⊗ IF )(Ik⊗k ⊗ V (2)
)(L2 ⊗ IF )V (2)
r
r
)(L2 ⊗ IF ) − Vr(M1L2K1 ⊗ IF ),
r
)(Ik ⊗ K2 ⊗ IF )(L2 ⊗ IF ) − Vr(K2M1L2 ⊗ IF )+
r − Vr(K2M1L2 ⊗ IF )+
)(Ik ⊗ L2 ⊗ IF )(Ik⊗k ⊗ V (2)
)(L2 ⊗ IF )
− Vr((M1L2)2 ⊗ IF ).
r
Now, by further use of the contractivity of the processes V (1), V (2) and V , together with
the estimate (3.1),
r
kE[Z (1)
kE[Z (2)
kE[Z (3)
r
r
and so
]k 62C(F1, F2)kK2k, and
]k 62(cid:0)kK1k2 + kK1kkK2k + kM1kkK1kkL2k(cid:1) = 2kK1kC(F1, F2),
]k 62(cid:0)kK1kkM1kkL2k + kM1kkK2kkL2k + kM1k2kL2k2(cid:1) = 2kM1kkL2kC(F1, F2),
kE[V (1)
t V (2)
dr E[Z (1)
r + Z (2)
r + Z (3)
r
dsZ s
t − Vt]k =(cid:13)(cid:13)(cid:13)Z t
dsZ s
6Z t
0
0
0
0
](cid:13)(cid:13)(cid:13)
dr 2 C(F1, F2)2 = t2 C(F1, F2)2,
as required.
(cid:3)
r,t
Then
r,t X F2
r,t − X F1✁F2
Lemma 3.2. Let F1 ∈ qcβ1(bk ⊗ h), F2 ∈ qcβ2(bk ⊗ h), c′, c ∈ k and r, t ∈ R+ with t > r.
(cid:13)(cid:13)Ω(cid:0)c′
(cid:1)(cid:13)(cid:13) 6 (t − r)2 e(t−r)(β1+β2) C(cid:0)F ∗
c′ ✁ F1, F2 ✁ Fc(cid:1)2
[r,t[, c[r,t[(cid:1)(cid:0)X F1
Proof. By definition of the evolutions (cid:0)X G
r,t(cid:1)06r6t, for G = F1, F2 and F1 ✁ F2, and
(cid:1)(cid:13)(cid:13)
t−r − X F1✁F2
In this case, by the
and so we may suppose without loss of generality that r = 0.
identity (2.2), Example 2.2 and the associativity of ✁, the estimate follows immediately
from Propositions 2.3 and 3.1.
(cid:3)
[0,t−r[, c[0,t−r[(cid:1)(cid:0)X F1
LHS =(cid:13)(cid:13)Ω(cid:0)c′
identity (2.8a) along with the remark following it,
t−rX F2
t−r
.
QS LIE–TROTTER PRODUCT FORMULA
13
Now let S denote the subspace of L2(R+; k) consisting of step functions, whose right-
Then
C(g′, F1, F2, g) := maxnC(cid:0)F ∗
continuous versions we use when evaluating. For F1, F2 ∈ b(bk ⊗ h) and g′, g ∈ S, set
c′ ✁ F1, F2 ✁ Fc(cid:1) : c′ ∈ Ran g′, c ∈ Ran go.
Lemma 3.3. Let F1 ∈ qcβ1(bk ⊗ h), F2 ∈ qcβ2(bk ⊗ h), g′, g ∈ S and r, t ∈ R+ with t > r.
(cid:13)(cid:13)Ω(cid:0)g′
(cid:1)(cid:13)(cid:13) 6 (t − r)2 e(t−r)(β1+β2) C(g′
Proof. Set β := β1 + β2 and let {r = t0 < t1 < · · · < tn < tn+1 = t} be such that g and
g′ are constant, say ci and di respectively, on the interval [ti, ti+1[, and, for i = 0, · · · , n,
set Ωi := Ω(g′
[r,t[, g[r,t[(cid:1)(cid:0)X F1
[r,t[, F1, F2, g[r,t[)2.
r,t − X F1✁F2
r,t X F2
r,t
[ti,ti+1[, g[ti,ti+1[)
X 1
i := X F1
ti,ti+1
, X 2
i := X F2
ti,ti+1 and Xi := X F
ti,ti+1
,
noting that Ωi := Ω(ci
sequent remark, we see that
[ti,ti+1[, di
[ti,ti+1[). Then, again using identity (2.8b) and the sub-
−→Y06i6n
Ωi(X 1
i X 2
i ) −
−→Y06i6n
Ωi(Xi).
Therefore, using Lemma 3.2,
r,t
r,t X F2
(cid:1) =
r,t − X F1✁F2
Ω(cid:0)g′
[r,t[, g[r,t[(cid:1)(cid:0)X F1
LHS =(cid:13)(cid:13)(cid:13)
nXi=0(cid:16) −→Y06p<i
nXi=0
eβ(ti−r)(cid:13)(cid:13)Ωi(X 1
nXi=1
(ti+1 − ti)2C(cid:0)F ∗
6eβ(t−r)
i X 2
Ωp(Xp) Ωi(X 1
6
i X 2
i − Xi)
−→Yi<p6n
i − Xi)(cid:13)(cid:13)eβ(t−ti+1)
ci ✁ F1, F2 ✁ Fdi(cid:1)2
Ωp(X 1
p X 2
p )(cid:17)(cid:13)(cid:13)(cid:13)
6 RHS .
(cid:3)
By a partition P of R+ we mean a sequence (sn) in R+ which is strictly increasing
and tends to infinity. Where convenient we identify a partition with its set of terms. For
S ⊂⊂ R+, let S denote the mesh of {0} ∪ S, that is, max{ti − ti−1 : 1 6 i 6 k}, where
S = {t1 < · · · < tn} and t0 := 0.
Let P be a partition of R+, let {t1 < · · · < tN −1} = P ∩ ]r, t[ and set t0 := r and tN := t.
Then
Lemma 3.4. Let F1 ∈ qcβ1(bk ⊗ h), F2 ∈ qcβ2(bk ⊗ h), g′, g ∈ S and r, t ∈ R+ with t > r.
(cid:13)(cid:13)(cid:13)Ω(cid:0)g′
[r,t[, g[r,t[(cid:1)(cid:16) −→Y16j6N
Proof. Set β := β1 + β2 and, for j = 1, · · · , N , set Ωj := Ω(cid:0)g′
6(cid:12)(cid:12)P ∩ [r, t](cid:12)(cid:12) (t − r) e(t−r)(β1+β2) C(g′
[tj−1,tj[, g[tj−1,tj[(cid:1),
[r,t[, F1, F2, g[r,t[)2.
tj−1,tj − X F1✁F2
(cid:17)(cid:13)(cid:13)(cid:13)
X F1
X F2
tj−1,tj
r,t
X 1
j := X F1
tj−1,tj
, X 2
j := X F1
tj−1,tj and Xj := X F1✁F2
tj−1,tj
.
14
MARTIN LINDSAY
Then, using Lemma 3.3 (and arguing as in its proof), we see that
Ωn(X 1
nX 2
n)(cid:17)(cid:13)(cid:13)(cid:13)
LHS =(cid:13)(cid:13)(cid:13)
NXj=1(cid:16) −→Y16n<j
NXj=1
eβ(tj−1−r)(cid:13)(cid:13)Ωj(X 1
NXj=1
6eβ(t−r)
6
Ωn(Xn) Ωj(X 1
j X 2
j − Xj)
−→Yj<n6N
j − Xj)(cid:13)(cid:13)eβ(t−tj+1)
j X 2
(tj − tj−1)2C(g′
[tj−1,tj[, F1, F2, g[tj−1,tj [)2
6eβ(t−r)(cid:12)(cid:12)P ∩ [r, t](cid:12)(cid:12) (t − r) C(g′
[r,t[, F1, F2, g[r,t[)2,
as required.
We may now prove the main result.
Theorem 3.5. Let V 1 and V 2 be quasicontractive elementary QS cocycles on h with noise
dimension space k and respective stochastic generators F1 and F2, and let T ∈ R+ and
ϕ ∈ B(F)∗. Then
sup
06r6t6T(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)V P,1✁2
r,t
− X F1✁F2
r,t
(cid:1)(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0.
The notation here is as follows. For a partition P of R+ and 0 6 r 6 t,
V P,1✁2
r,t
:=
−→Y16j6N
V 1
tj−1,tj V 2
tj−1,tj
in which {t1 < · · · < tN −1} = P ∩ ]r, t[, t0 := r and tN := t.
Proof. Set β := β0(V 1) + β0(V 2) and V := X F1✁F2. We may suppose without loss of
generality that ϕ is of the form ω(h′),(h) for some h′, h ∈ S, since such functionals are
total in the Banach space B(F)∗ and (V P,1✁2
− Vr,t) is uniformly bounded by 2eT max{0,β}
for [r, t] ⊂ [0, T ].
r,t
For any partition P of R+, Lemma 3.4 implies that there is a constant C = C(ϕ, F1, F2)
such that, for all subintervals [r, t] of [0, T ],
(cid:3)
(cid:3)
Proposition 3.1 and Lemmas 3.2, 3.3 and 3.4 extend to the case of QS generators
The result follows.
Fi =h Ki Mi
where
r,t
(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)V P,1✁2
− Vr,t(cid:1)(cid:13)(cid:13) 6(cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) T eT β C 2.
Li Nii (i = 1, · · · , n), the constant C(F1, F2) being replaced by
0 0(cid:21) ∈ B(bC).
fi =(cid:20)kKik kMik
∆⊥f1 ✁(cid:0)f2 ✁ · · · ✁ fn−1(cid:1) ✁ fn ∆⊥
kNik(cid:21) ∈ B(bC) and ∆⊥ = ∆⊥
− I =Z t
X F1✁···✁Fn
·(cid:0)F1 ✁ · · · ✁ Fn(cid:1) dΛs.
C =(cid:20)1 0
X F1✁···✁Fn
kLik
0
s
t
In this case the quantum Ito product formula allows X F1
t − I to be expressed as
a sum of 2n − 1 QS integrals, each of which may be paired with one of the 2n − 1 terms
arising from the expansion of F1 ✁ · · · ✁ Fn in the identity
· · · X Fn
t
QS LIE–TROTTER PRODUCT FORMULA
15
As a consequence, Theorem 3.5 extends to quasicontractive elementary QS cocycles
V 1, · · · , V n, with respective stochastic generators F1, · · · , Fn, as follows:
sup
06r6t6T(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)V P,1✁···✁n
r,t
in the notation
− X F1✁···✁Fn
r,t
(cid:1)(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0,
V P,1✁···✁n
r,t
:=
V 1
tj−1,tj · · · V n
tj−1,tj .
−→Y16j6N
Remarks. (i) Note that the QS cocycle V = X F1✁···✁Fn is contractive (respectively, iso-
metric, or coisometric) provided that all of V 1, · · · , V n are.
(ii) Whilst the convergence of Trotter products holds in the above hybrid norm-ultraweak
topology, if V is isometric and V 1, · · · , V n are contractive (in particular, if V 1, · · · , V n
are isometric) then convergence also holds in the strong operator topology:
sup
06r6t6T(cid:13)(cid:13)(cid:0)V P,1✁···✁n
r,t
− X F1✁···✁Fn
r,t
(cid:1)ξ(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0
This follows from the uniform continuity of the function
(T ∈ R+, ξ ∈ h ⊗ F).
[0, T ]2
6 → h ⊗ F ,
(r, t) 7→ Vr,tξ
(ξ ∈ h ⊗ F , T ∈ R+).
We now revisit the case of QS cocycles with independent driving noise and show how to
view it as a special case of the above theorem. Suppose therefore that the noise dimension
space has orthogonal decomposition k = k1 ⊕ · · · ⊕ kn. Denoting the symmetric Fock
space over L2(R+; ki) by F i, second quantisation of the natural isometry from L2(R+; ki)
to L2(R+; k) (see (2.1)), followed by ampliation, gives an isometry h ⊗ F i → h ⊗ F which
in turn induces a normal *-algebra monomorphism
Υi : B(h ⊗ F i) → B(h ⊗ F)
(i = 1, · · · , n).
Corollary 3.6 ([LS1]). For i = 1, · · · , n, let V i be a quasicontractive elementary QS
cocycle on h with noise dimension space ki and stochastic generator Fi. Then, for all
T ∈ R+ and ϕ ∈ B(F)∗,
The notation here is as follows. For a partition P of R+ and 0 6 r 6 t, setting {t1 <
· · · < tN −1} = P ∩ ]r, t[, t0 := r and tN := t,
− X F1⊞ ··· ⊞Fn
r,t
(cid:1)(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0.
r,t
sup
06r6t6T(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)V P,1⊞ ··· ⊞n
−→Y16j6N
V P,1⊞ ··· ⊞n
r,t
:=
Υ1(V 1
tj−1,tj ) · · · Υn(V n
tj−1,tj ),
s,s′) = X ι′(F2)
Proof. Suppose first that n = 2. Note that, in the notation (1.7), Υ1(V 1
and
(s, s′ ∈ R+, s < s′). Therefore, in view of identity (1.8), for this case
Υ2(V 2
the corollary follows immediately from Theorem 3.5. The general case follows similarly
via obvious extension of the notation (1.7) and corresponding identity (1.8).
(cid:3)
s,s′) = X ι(F1)
s,s′
s,s′
4. Maximal Gaussian component of a QS generator
In this section it is shown that every QS generator F ∈ qc(bk ⊗ h) enjoys a unique
decomposition F1 ⊞ F2 in which F2 is 'pure Gaussian' and F1 is 'wholly non-Gaussian', in
senses to be defined below. This amounts to extracting a maximal Gaussian component
of the generator and demonstrating its uniqueness.
It chimes well with the way that
Hunt's formula decomposes the generator of a L´evy process on a compact Lie group into
a maximal Gaussian component and a jump part (albeit not exactly uniquely).
16
MARTIN LINDSAY
To each contraction C ∈ B(k ⊗ h) we associate the following closed subspaces of k:
and
kC
g :=(cid:8)c ∈ k : (C − I)Ec = 0(cid:9), where Ec := ci ⊗ Ih,
p := Lin(cid:8)(Ik ⊗ hu)ξ : ξ ∈ Ran(C − I)∗, u ∈ h(cid:9).
kC
The former captures the subspace of all 'directions' in which C acts as the identity op-
erator; the latter is complementary (see below). The subscripts denote 'Gaussian' and
'preservation' parts.
Letting J C
p denote the inclusion kC
p → k, set
Cp := (J C
p ⊗ Ih)∗C(J C
p ⊗ Ih),
the compression of C to kC
p ⊗ h.
Lemma 4.1. Let C ∈ B(k ⊗ h) be a contraction. Then the following hold:
(a) k = kC
p ⊕ kC
g .
(b) C has block matrix formh Cp 0
g i, where I C
0 I C
p = kC
p .
g := IkC
g ⊗h.
(c) (kC
p )Cp
Proof. Set I C
p := IkC
p ⊗h.
(a) Let c ∈ k. Then, for all u ∈ h,
hc, (Ik ⊗ hu)(C − I)∗ξi = 0 for all ξ ∈ k ⊗ h ⇐⇒ (C − I)(c ⊗ u) = 0.
It follows that c ∈ (kC
p )⊥ if and only if c ∈ kC
g , so (a) follows.
(b) Contraction operators T on a Hilbert space H enjoy the following elementary prop-
erty. In terms of the orthogonal decomposition H = K⊥ ⊕ K where K := Ker(T − I), T has
g ⊗ h ⊂ Ker(C − I) it follows that, in terms
p ⊗ h) ⊕ (kC
g ⊗ h), C has the claimed block
diagonal block matrix form(cid:2) ∗ 0
0 IK(cid:3). Since kC
of the orthogonal decomposition k ⊗ h = (kC
matrix form.
(c) By (a) it suffices to show that (kC
p )Ee = 0 and so, by (b), (C − I)EJ C
p ∩ kC
g = {0}, by (a). Since J C
p )Cp
g = {0}. Accordingly let e ∈ (kC
p e = 0, in other words J C
g . Then
g . Thus
p is an inclusion this implies that e = 0, as
(cid:3)
p )Cp
p e ∈ kC
(Cp − I C
p e ∈ kC
J C
required.
∗
[ ∗ ∗
∗ 0 ];
∗ C−I ]. We say that
(a) F is Gaussian if kC
(b) F is wholly non-Gaussian if kC
g = k, equivalently C = I so that F has block matrix form
p = k, equivalently Cp = C;
1
2 L∗L −L∗
L
Definition 4.2. Let F ∈ qc(bk ⊗ h) with block matrix form [ ∗
0 i, for some L ∈ B(h; k⊗
(c) F is pure Gaussian if it has block matrix formh −
0 i, for some M ∈ B(k ⊗ h; h);
h), equivalentlyh −
0 C−I(cid:3);
(d) F is pure preservation if it has block matrix form(cid:2) 0
g(bk ⊗ h), wn-g(bk ⊗ h), pg(bk ⊗ h) and pp(bk ⊗ h),
for the respective classes of Gaussian, wholly non-Gaussian, pure Gaussian and pure
preservation generator.
0 0 ], in other words F ∈ z(bk ⊗ h).
(e) F is pure drift if it has block matrix form [ ∗ 0
1
2 M M ∗ M
−M ∗
Write
0
Remarks. (i) The terminology has its origins in the work of Schurmann on L´evy–Khintchin-
type decompositions for L´evy processes on bialgebras such as algebraic quantum groups
QS LIE–TROTTER PRODUCT FORMULA
17
([Sch]). In the presence of a minimality condition, the Gaussian property for the gener-
ator of an elementary unitary QS cocycle U may alternatively be expressed in terms of
the cocycle itself as follows:
t−1E(cid:2)(U 1
t − I)u1
v1 (U 2
t − I)u2
v2 (U 3
t − I)u3
v3(cid:3) → 0 as t → 0
for some K ∈ B(h) and L ∈ B(h; k ⊗ h).
for all u1, · · · , v3 ∈ h and all choices of U 1, U 2 and U 3 from {U, U ∗} where, for X ∈
B(h ⊗ F) and u, v ∈ h, X u
v := (ωu,v⊗ idB(F ))(X) = (hu ⊗ IF )X(vi ⊗ IF ) (see [SSS]
and [Sch]).
determines a decomposition F = F1 ⊞ F ′
(ii) By Theorem 1.3 (a), F ∈ g(bk ⊗ h) if and only if it has block matrix form(cid:2) K −L∗
L 0 (cid:3),
(ii) Given a Gaussian QS generator(cid:2) K −L∗
L 0 (cid:3), any orthogonal decomposition k = k1 ⊕ k2
decomposition L =h L1
L2i, F1 is pure Gaussian with
F1 =(cid:20)− 1
0 (cid:21) and F ′
1 in which, corresponding to the block matrix
1 =(cid:20)K + 1
0 (cid:21) .
1L1 −L∗
1
1L1 −L∗
2
2 L∗
L2
2 L∗
L1
Since L∗L = L∗
1L1 + L∗
2L2, F ′
1 is pure Gaussian too if and only F is.
(iii) In the case of no noise (k = {0}), pp(bk ⊗ h) = ∅ and g(bk ⊗ h), wn-g(bk ⊗ h), pg(bk ⊗ h)
and z(bk ⊗ h) all equal b(bk ⊗ h). Otherwise, when k 6= {0}, all of g(bk ⊗ h), wn-g(bk ⊗ h),
pg(bk ⊗ h), pp(bk ⊗ h) and z(bk ⊗ h) are selfadjoint and the following relations are easily
verified:
Also, in view of the series decompositions (Proposition 1.4),
g(bk ⊗ h) ∩ wn-g(bk ⊗ h) = ∅;
pg(bk ⊗ h) ∩ z(bk ⊗ h) = {0} = g(bk ⊗ h) ∩ pp(bk ⊗ h);
pg(bk ⊗ h) ⊂ u(bk ⊗ h) and pp(bk ⊗ h) ⊂ c(bk ⊗ h);
g(bk ⊗ h) ✁ g(bk ⊗ h) = g(bk ⊗ h) and pp(bk ⊗ h) ✁ pp(bk ⊗ h) = pp(bk ⊗ h);
z(bk ⊗ h) ✁ pg(bk ⊗ h) = g(bk ⊗ h) = pg(bk ⊗ h) ✁ z(bk ⊗ h).
z(bk ⊗ h) ✁ pg(bk ⊗ h) ✁ pp(bk ⊗ h) ⊃ i(bk ⊗ h), and
pp(bk ⊗ h) ✁ pg(bk ⊗ h) ✁ z(bk ⊗ h) ⊃ i(bk ⊗ h)∗.
qc(bk ⊗ h) = [k=k1⊕k2
wn-g(bk1 ⊗ h) ⊞ pg(bk2 ⊗ h).
The next result implies that
Theorem 4.3. Let F ∈ qc(bk ⊗ h). Then F enjoys a unique decomposition
Fwn-g ⊞ Fmg
where, for some orthogonal decomposition k1 ⊕ k2 of k,
LC
g := (J C
g ⊗ Ih)∗L, LC
Proof. Let(cid:2) K M
Fwn-g ∈ wn-g(bk1 ⊗ h) and Fmg ∈ pg(bk2 ⊗ h).
L C−I(cid:3) be the block matrix decomposition of F , and set
Fwn-g =(cid:20)K + 1
Cp − I(cid:21) and Fmg =(cid:20)− 1
Then F = Fwn-g ⊞ Fmg where
2 (LC
LC
p
p ⊗ Ih)∗L, M C
g := M (J C
p := (J C
g M C
p
g )∗LC
g )∗LC
2(LC
LC
g
g ⊗ Ih) and M C
g M C
g
0 (cid:21) .
p := M (J C
p ⊗ Ih).
18
MARTIN LINDSAY
By Remark (ii) above, M C
that Fwn-g is wholly non-Gaussian. This proves existence.
g = −(LC
g )∗ so Fmg is pure Gaussian. It follows from Lemma 4.1
For uniqueness, suppose that F = F1 ⊞ F2 where F1 ∈ wn-g(bk1 ⊗ h) and F2 ∈ pg(bk2 ⊗
h) for an orthogonal decomposition k = k1 ⊕ k2. Then F1 and F2 have block matrix
decompositions
moreover
2L2 −L∗
2
2 L∗
L2
0 (cid:21) ,
F1 =(cid:20)K1 M1
L1 C1 − I(cid:21) and F2 =(cid:20)− 1
0(cid:1) ∈ kC
p = k1 and kC
kC1
g ⊃ k2.
g . Then (C − I)E(e1
Suppose that e1 ∈ k1 and(cid:0)e1
F1 is wholly non-Gaussian this implies that e1 = 0. Thus kC
that C1 = Cp and L2 = LC
0 ) = 0, so (C1 − I)Ee1 = 0. Since
p = k1. It follows
g . This implies that F1 = Fwn-g and F2 = Fmg, as required. (cid:3)
g = k2 and kC
Remarks. (i) Clearly Fmg is the maximal pure Gaussian component of F .
(ii) Let F ∈ i(bk ⊗ h). Then the above decomposition of F takes the form
0 (cid:21) ∈ i(bk1 ⊗ h) ⊞ u(bk2 ⊗ h)
L1 W − I(cid:21) ⊞(cid:20)− 1
F =(cid:20)K1 −L∗
1L1, W isometric and kW
2L2 −L∗
2
2 L∗
L2
with Re K1 = − 1
1W
2 L∗
p = k1; W being unitary if and only if
F ∈ u(bk ⊗ h).
5. Perturbation of QS cocycles
In case the second of two elementary QS cocycles V 1, V 2 is isometric, there is another,
more standard, way of realising the QS cocycle X F1✁F2 where F1 and F2 are the stochastic
generators of V 1 and V 2.
Theorem 5.1 (Cf. [EvH], [BLS]). Let F1 ∈ qc(bk ⊗ h) and F2 ∈ i(bk ⊗ h) Then
X F1✁F2 = X j2,F1X F2
where j2 is the normal *-monomorphic QS mapping cocycle on (the von Neumann algebra)
B(h) given by
t (x) = X F2
j2
t (x ⊗ IF )(X F2
t )∗
(x ∈ B(h), t ∈ R+),
and X j2,F1 is the unique strong solution of the QS differential equation dXt = Xt · Gt dΛt,
X0 = I, for the integrand process G :=(cid:0)(idB(bk) ⊗ j2
t )(F1)(cid:1)t>0.
Proof. Given the existence of X j2,F1 and its quasicontractivity ([BLS]), the result follows
easily from the quantum Ito product formula and uniqueness for weak solutions, which
are bounded with locally uniform bounds, of the QS differential equation dXt = Xt ·
(F1 ✁ F2) dΛt, X0 = Ih⊗F (see Theorem 2.1).
(cid:3)
Remark. In [BLS] we worked in the equivalent category of QS right cocycles.
Applying this result to the Gaussian/non-Gaussian decomposition of QS generators
(Theorem 4.3) yields the following result. Recall the injections (1.7) associated with
realising the concatenation product in terms of the series product (1.8).
Corollary 5.2. Let F ∈ qc(bk ⊗ h) with block matrix form [ ∗
X F = X j2,F1X F2 ,
where, for the orthogonal decomposition k ⊗ h = (kC
p ⊗ h) ⊕ (kC
g ⊗ h),
∗
∗ C−I ]. Then
F1 = ι(Fwn-g) and F2 = ι′(Fmg).
QS LIE–TROTTER PRODUCT FORMULA
Proof. The inclusions
ensure that X F2 is unitary and so Theorem 5.1 applies.
ι′(cid:0)pg(ckC
g ⊗ h)(cid:1) ⊂ pg(bk ⊗ h) ⊂ u(bk ⊗ h)
19
(cid:3)
6. Holomorphic QS Cocycles
In this section the setting is extended to holomorphic QS cocycles ([LS2]). Before
formulating the conjecture, the corresponding Ito algebra is investigated, mirroring Sec-
tion 1. As is customary, we identify each bounded Hilbert space operator T with its
associated quadratic form qT , given by qT [ξ] := hξ, T ξi. We also use the notation q(·, ·)
for the sesquilinear form associated with a quadratic form q[ · ] by polarisation.
Fix Hilbert spaces h and H. Let Q(h ⊕ H) denote the class of quadratic forms Γ on
h ⊕ H having the following structure:
from h to H, and
(cid:26) Dom Γ = D ⊕ H
Γ[ξ] = γ[u] −(cid:2)hζ, Lui + heLu, ζi + hζ, (C − I)ζi(cid:3),
Dom γ = Dom L = DomeL = D.
where D is a subspace of h, C ∈ B(H), γ is a quadratic form on h, L and eL are operators
Write Γ ∼ (γ, L,eL, C), and refer to (γ, L,eL, C) as the components of Γ. Also define an
For reasons which will become apparent, D is not assumed to be dense in h.
ζ(cid:1) ∈ Dom Γ,
for ξ =(cid:0)u
associated operator on h ⊕ H by
F ∆
Γ :=(cid:20)0
L C − I(cid:21) .
0
Thus
Dom F ∆
Γ = Dom Γ and Ran F ∆
Γ ⊂ {0} ⊕ H = Ran ∆
where, as usual, ∆ := 0h ⊕ IH. The inclusion obviously implies that
Γ ⊂ D′ ⊕ H for any subspace D′ of h.
Ran F ∆
(6.1)
Note that if Γ ∈ Q(h ⊕ H) with components (γ, L,eL, C), then the adjoint form Γ∗ belongs
to Q(h ⊕ H) too, with
hξ, F ∆
Γ ξi = −Γ(∆ξ, ξ) and hF ∆
Γ∗ ξ, ξi = −Γ(ξ, ∆ξ)
(ξ ∈ Dom Γ).
(6.2)
Definition 6.1. For Γi ∈ Q(h ⊕ H), with components (γi, Li,eLi, Ci) (i = 1, 2), define
Γ1 ✁ Γ2, Γ1∆Γ2 ∈ Q(h ⊕ H) by
Γ1 ✁ Γ2 ∼ (γ, L,eL, C), where
γ[u] = γ1[u] + γ2[u] − heL1u, L2ui,
L = L1 + C1L2, eL = C ∗
Γ1∆Γ2 ∼ (γ, L,eL, C), where
γ[u] = −heL1u, L2ui,
L = (C1 − I)L2, eL = (C ∗
2eL1 +eL2 and C = C1C2;
2 − I)eL1 and C = (C1 − I)(C2 − I) + I.
Thus, in terms of the associated sesquilinear form,
Γ∗ ∼ (γ∗,eL, L, C ∗) and F ∆
Γ∗ =(cid:20)0
0
eL C ∗ − I(cid:21) .
20
Thus
MARTIN LINDSAY
Γ1 ✁ Γ2 = Γ1 + Γ2 + Γ1∆Γ2,
Dom(Γ1 ✁ Γ2) = Dom(Γ1∆Γ2) = Dom Γ1 ∩ Dom Γ2,
(Γ1∆Γ2)[ξ] = −hF ∆
Γ∗
1
F ∆
ξ, F ∆
Γ1∆Γ2 =(cid:0){0} ⊕ (C1 − I)(cid:1)F ∆
Γ2 ξi
Γ2 .
(ξ ∈ Dom Γ1 ∩ Dom Γ2), and
(6.3a)
(6.3b)
(6.3c)
Lemma 6.2. The prescription (Γ1, Γ2) 7→ Γ1∆Γ2 defines an associative and bilinear
composition on the vector space Q(h ⊕ H) which is also involutive:
(Γ1∆Γ2)∗ = Γ∗
2∆Γ∗
1.
(6.4)
Proof. Let Γi ∈ Q(h ⊕ H) with domain Di ⊕ H (i = 1, 2, 3). Bilinearity follows from the
evident linearity of the map Γ 7→ F ∆
Γ , and (6.4) holds since, for ξ ∈ (D1 ∩ D2) ⊕ H,
(Γ1∆Γ2)∗[ξ] = (Γ1∆Γ2)[ξ] = −hF ∆
Γ2 ξ, F ∆
Γ∗
1
ξi = (Γ∗
2∆Γ∗
1)[ξ].
Clearly Dom(cid:0)(Γ1∆Γ2)∆Γ3(cid:1) =(cid:0)D1 ∩ D2 ∩ D3(cid:1) ⊕ H = Dom(cid:0)Γ1∆(Γ2∆Γ3)(cid:1) and, for ξ ∈
(D1 ∩ D2 ∩ D3) ⊕ H, (6.3c) implies that
Γ2∆Γ3 ξi = hF ∆
Γ∗
1
hF ∆
Γ∗
1
ξ, F ∆
ξ,(cid:0){0} ⊕ (C2 − I)(cid:1)F ∆
2 − I)(cid:1)F ∆
Γ3 ξi = hF ∆
= h(cid:0){0} ⊕ (C ∗
= hF ∆
Γ∗
ξ, F ∆
Γ∗
1
2∆Γ∗
1
ξ, F ∆
(Γ1∆Γ2)∗ ξ, F ∆
Γ3 ξi
Γ3 ξi
Γ3 ξi.
Thus ∆ is associative, by (6.3b).
(cid:3)
Proposition 6.3. The composition (Γ1, Γ2) 7→ Γ1 ✁Γ2 endows Q(h⊕H) with the structure
of a *-monoid whose identity element is Γ0 ∼ (0, 0, 0, I), in particular,
(Γ1 ✁ Γ2)∗ = Γ∗
2 ✁ Γ∗
1.
Proof. Let Γ1, Γ2, Γ3 ∈ Q(h ⊕ H). In view of Lemma 6.2,
(Γ1 + Γ2 + Γ3) + (Γ1∆Γ2 + Γ2∆Γ3 + Γ1∆Γ3) + Γ1∆Γ2∆Γ3
is a common expression for Γ1 ✁ (Γ2 ✁ Γ3) and (Γ1 ✁ Γ2) ✁ Γ3, which have common
domain (D1 ∩ D2 ∩ D3) ⊕ H. It is easily seen that the element Γ0 ∼ (0, 0, 0, I) satisfies
Γ0 ✁ Γ = Γ = Γ ✁ Γ0 for all Γ ∈ Q(h ⊕ H). The fact that the adjoint operation defines
an involution on the resulting monoid follows from its additivity on Q(h ⊕ H) and the
identity (6.4).
(cid:3)
For the following lemma, recall the range observation (6.1).
Lemma 6.4. Let Γi ∈ Q(h ⊕ H) with domain Di ⊕ H (i = 1, 2, 3). Then
(Γ1 ✁ Γ2 ✁ Γ3)[ξ] = (Γ1 ✁ Γ3)[ξ] + Γ2(cid:0)(I + F ∆
Proof. Let ξ ∈ D1 ∩ D2 ∩ D3. Then, since
Γ∗
1
Γ3)ξ(cid:1)
)ξ, (I + F ∆
(ξ ∈ D1 ∩ D2 ∩ D3).
Γ1 ✁ Γ2 ✁ Γ3 − Γ1 ✁ Γ3 = Γ2 + Γ2∆Γ3 + Γ1∆Γ2 + Γ1∆Γ2∆Γ3,
the lemma follows by several applications of the identities (6.2) and (6.3c):
Γ∗
1
)ξ, (I + F ∆
Γ3 )ξ(cid:1)
Γ3 ξ(cid:1) + Γ2(cid:0)F ∆
Γ2(cid:0)(I + F ∆
Γ3 ξ(cid:1)
= Γ2[ξ] + Γ2(cid:0)ξ, F ∆
= Γ2[ξ] +(cid:0)Γ2∆Γ3(cid:1)[ξ] +(cid:0)Γ1∆Γ2(cid:1)[ξ] +(cid:0)Γ1∆Γ2∆Γ3(cid:1)[ξ].
ξ, ξ(cid:1) + Γ2(cid:0)F ∆
Γ∗
1
Γ2 ξi − hF ∆
Γ∗
1
= Γ2[ξ] − hF ∆
Γ∗
2
Γ3 ξi − hF ∆
Γ∗
1
Γ∗
1
ξ, F ∆
ξ, F ∆
ξ, F ∆
ξ,(cid:0){0} ⊕ (I − C2)(cid:1)F ∆
Γ3 ξi
(cid:3)
QS LIE–TROTTER PRODUCT FORMULA
21
Proposition 6.5. Let Γ1, Γ2 ∈ Q(h ⊕ H) and β1, β2 ∈ R, and set Γ = Γ1 ✁ Γ2 and
β = β1 + β2.
(a) Suppose that, for i = 1, 2,
Γ∗
i ✁ Γi > 2βi∆⊥ on Dom Γi.
Then Γ∗ ✁ Γ > 2β∆⊥ on Dom Γ.
(b) Suppose that, for i = 1, 2,
Γ∗
Then Γ∗ ✁ Γ = 0 on Dom Γ.
i ✁ Γi = 0 on Dom Γi.
Proof. By associativity and Lemma 6.4,
for all ξ ∈ Dom Γ1∩Dom Γ2. The result therefore follows since ∆⊥F ∆
Γ2
= 0 on Dom Γ2. (cid:3)
Remark. Thus
(Γ∗ ✁ Γ)[ξ] = (Γ∗
2 ✁ Γ2)[ξ] + (Γ∗
1 ✁ Γ1)(cid:2)(I + F ∆
Γ2)ξ(cid:3)
(cid:8)Γ ∈ Q(h ⊕ H) : Γ∗ ✁ Γ = 0 = Γ ✁ Γ∗(cid:9)
forms a subgroup of the group of invertible elements of (Q(h ⊕ H), ✁).
To complete the discussion of the algebra of the series product on quadratic forms, here
is the form generalisation of Proposition 1.2.
Proposition 6.6. Let Γ ∈ Q(h ⊕ H). Then
(Γ∗ ✁ Γ)[ξ] = (Γ ✁ Γ∗)(cid:2)(I + F ∆
Γ )ξ(cid:3) +(cid:13)(cid:13)F ∆
Γ∗✁Γξ(cid:13)(cid:13)2
Let V ∈ Q(h ⊕ H) be of the form ν ⊕ 0H where ν ∈ Q(h)sa. Then
Γ∗ ✁ Γ > V if and only if Γ ✁ Γ∗ > V.
(ξ ∈ Dom Γ).
(6.5)
Proof. Let ξ =(cid:0)u
share the same domain. On the one hand, setting Γ1 = Γ2 = Γ∗ ✁ Γ in (6.3a) and (6.3b)
yields
ζ(cid:1) ∈ Dom Γ. Note that Γ, Γ∗, Γ∗ ✁ Γ, Γ ✁ Γ∗ and Γ∗ ✁ Γ ✁ Γ∗ ✁ Γ all
Γ∗✁Γξ(cid:13)(cid:13)2
(cid:0)Γ∗ ✁ Γ ✁ Γ∗ ✁ Γ(cid:1)[ξ] −(cid:0)Γ∗ ✁ Γ(cid:1)[ξ] =(cid:0)Γ∗ ✁ Γ(cid:1)[ξ] −(cid:13)(cid:13)F ∆
(cid:0)Γ∗ ✁ Γ ✁ Γ∗ ✁ Γ(cid:1)[ξ] −(cid:0)Γ∗ ✁ Γ(cid:1)[ξ] =(cid:0)Γ ✁ Γ∗(cid:1)(cid:2)(I + F ∆
Γ )ξ(cid:3).
On the other hand, setting Γ1 = Γ∗, Γ2 = Γ ✁ Γ∗ and Γ3 = Γ in Lemma 6.4 yields
Thus (6.5) holds.
.
Now suppose that Γ ✁ Γ∗ > V. Then, since
V(cid:2)(I + F ∆
Γ )ξ(cid:3) = ν[u] = V[ξ],
(6.5) implies that (Γ∗ ✁ Γ)[ξ] > V[ξ]. Thus Γ∗ ✁ Γ > V. The converse implication follows
by exchanging Γ and Γ∗.
(cid:3)
Now we return to the Hilbert spaces h and k. Let Xhol(h) denote the class of quadratic
forms γ on h which are closed, densely defined and satisfy the accretive and semisectorial
conditions
Re γ + β > 0 and (cid:12)(cid:12)Im γ[u](cid:12)(cid:12) 6 α(cid:0) Re γ[u] + kuk2(cid:1)
for some β ∈ R and α ∈ R+, and let Shol(h) denote the class of holomorphic semigroups
they generate (see e.g. [Ouh]).
(u ∈ Dom γ)
In [LS2], a quasicontractive QS cocycle V on h is called holomorphic if its expectation
semigroup belongs to Shol(h). Denoting this class of QS cocycle by QSChol(h, k), it is
shown there that the correspondence Xhol(h) → Shol(h) extends to a bijection
X4
hol(h, k) → QSChol(h, k), Γ 7→ X Γ,
22
MARTIN LINDSAY
in which X4
hol(h, k) denotes the subclass of Q(bk ⊗ h) = Q(h ⊕ (k ⊗ h)) consisting of forms
Γ ∼ (γ, L,eL, C) such that γ ∈ Xhol(h) and Γ∗✁Γ+2β∆⊥ > 0, for some β ∈ R. We speak of
the stochastic form generator of the holomorphic cocycle. If Γ ∼ (γ, L,eL, C) ∈ X4
then it follows from Proposition 6.6 that Γ∗ ∈ X4
C ∈ B(k ⊗ h) is a contraction, and X Γ∗
hol(h, k)
hol(h, k) and in [LS2] it is also shown that
= (X Γ)♯, the dual QS cocycle defined in (2.7).
The bijection extends the above form-semigroup correspondence as follows:
if Γ ∼
hol(h, k) and X Γ = (Pt ⊗ IF )t>0 where P is the
(γ, 0, 0, I) where γ ∈ Xhol(h) then Γ ∈ X4
holomorphic semigroup with form generator γ. It also extends that of Theorem 2.1, in
the sense that if F ∈ qc(bk ⊗ h) with block matrix form(cid:2) K M
hol(h, k) given by
Γ ∼ (q−K , L, M ∗, C).
form in X4
L C−I(cid:3) then X F = X Γ for the
If Γ1, Γ2 ∈ X4
hol(h, k) then Γ1 ✁ Γ2 ∈ X4
hol(h, k) provided only that Dom Γ1 ∩ Dom Γ2 is
dense. This neatly extends the fact that if γ1, γ2 ∈ Xhol(h) then γ1 +γ2 ∈ Xhol(h) provided
only that Dom γ1 ∩ Dom γ2 is dense in h.
Conjecture 6.7. Let V 1 and V 2 be quasicontractive holomorphic QS cocycles on h with
noise dimension space k and respective stochastic form generators Γ1 and Γ2, and suppose
that Dom Γ1 ∩ Dom Γ2 is dense inbk ⊗ h. Then (in the notation of Theorem 3.5), for all
T ∈ R+, ϕ ∈ B(F)∗ and u ∈ h,
sup
06r6t6T(cid:13)(cid:13)(cid:0) idB(h) ⊗ ϕ(cid:1)(cid:0)V P,1✁2
r,t
− X Γ1✁Γ2
r,t
(cid:1) u(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0.
Moreover, if the QS cocycle X Γ1✁Γ2 is isometric and V 1 and V 2 are contractive then, for
all T ∈ R+ and ξ ∈ h ⊗ F,
sup
06r6t6T(cid:13)(cid:13)(cid:0)V P,1✁2
r,t
− X Γ1✁Γ2
r,t
(cid:1) ξ(cid:13)(cid:13) → 0 as (cid:12)(cid:12)P ∩ [0, T ](cid:12)(cid:12) → 0.
Remarks. The conjecture has three special cases where it is proven. Theorem 3.5 covers
the case where Γ1 and Γ2 are bounded. In the semigroup case, where Γi ∼ (γi, 0, 0, I) for
γi ∈ Xhol(h) (i = 1, 2), it reduces to a version of a celebrated result of Kato – as extended
by Simon ([Kat]). For the case of independent driving noises a version of the holomorphic
counterpart to Corollary 3.6, which includes the Kato–Simon theorem, is proved in [LS3].
Acknowledgements. I am grateful to Mateusz Jurczy´nski and Micha l Gnacik for useful
comments on an earlier draft of the paper. Support from the UK-India Education and Re-
search Initiative (UKIERI), under the research collaboration grant Quantum Probability,
Noncommutative Geometry & Quantum Information, is also gratefully acknowledged.
References
[Bel] V.P. Belavkin, A new form and a *-algebraic structure of quantum stochastic integrals in Fock
space, Rend. Sem. Mat. Fis. Milano 58 (1988), 177–193.
[BGL] A.C.R. Belton, M. Gnacik and J.M. Lindsay, Strong convergence of quantum random walks via
semigroup decomposition, arXiv :1712.02848 [math.PR].
[BLS] A.C.R. Belton, J.M. Lindsay and A.G. Skalski, Quantum Feynman-Kac perturbations, J. London
Math. Soc. (2 ) 89 (2014) no. 1, 275–300.
[Bha] B.V.R. Bhat, Cocycles of CCR flows, Mem. Amer. Math. Soc. 149 (2001), no. 709.
[BrR] O. Bratteli and D.W. Robinson, "Operator Algebras and Quantum Statistical Mechanics II: Equi-
librium states. Models in quantum statistical mechanics," 2nd Edition, Springer-Verlag, Berlin,
1997.
[Che] P. R. Chernoff, Product formulas, non-linear semigroups and addition of unbounded operators,
Mem. Amer. Math. Soc. 140, 1974.
[DGS] B. Das, D. Goswami and K.B. Sinha, A homomorphism theorem and a Trotter product formula
for quantum stochastic flows with unbounded coefficients, Comm. Math. Phys. 330 (2014) no. 2,
435–467.
QS LIE–TROTTER PRODUCT FORMULA
23
[DLT] B.K. Das, J.M. Lindsay and O. Tripak, Sesquilinear quantum stochastic analysis in Banach space,
J. Math. Anal. Applic. 409 (2014) no. 2, 1032–1051.
[Dav] E.B. Davies, "One-Parameter Semigroups," Academic Press, London, 1980.
[EvH] M.P. Evans and R.L. Hudson, Perturbations of quantum diffusions, J. London Math. Soc. (2 )
(1990), no. 2, 373–384.
[Fag] F. Fagnola, Characterization of isometric and unitary weakly differentiable cocycles in Fock space,
in, "Quantum Probability & Related Topics," QP-PQ VIII, (Ed. L. Accardi), World Scientific,
Singapore, 1993, pp. 143–164.
[GL+] D. Goswami, J.M. Lindsay, K.B. Sinha and S.J. Wills, Dilation of Markovian cocycles on a von
Neumann algebra, Pacific J. Math. 211 (2003) no. 2, 221–247.
[GoJ] J. Gough and M. James, The series product and its application to quantum feedforward and
feedback networks, IEEE Trans. Automat. Control 54 (2009) no. 11, 2530–2544.
[HiP] E. Hille and R.S. Philips, "Functional Analysis and Semigroups," Coll. Publ. 31, American Math-
ematical Society, Providence, R.I., 1957.
[Hol] A.S. Holevo, Stochastic representation of quantum dynamical semigroups (Russian), [Translated in
Proc. Steklov Math. Inst. (1992) no. 2, 145–154.] Trudy Mat. Inst. Steklov 191 (1989), 130–139.
[HuP] R.L. Hudson and K.R. Parthasarathy, Quantum Ito's formula and stochastic evolution, Comm.
Math. Phys. 93 (1984) no. 3, 301–323.
[Jou] J.-L. Journ´e, Structure des cocycles markoviens sur l'espace de Fock, Probab. Theory Related Fields
75 (1987) no. 2, 291–316.
[L]
[Kat] T. Kato, Trotter's product formula for an arbitrary pair of self-adjoint contraction semigroups,
in, "Topics in Functional Analysis (essays dedicated to M.G. Krein on the occasion of his 70th
birthday)," Adv. in Math. Suppl. Stud. 3, Academic Press, London 1978, pp. 185–195.
J.M. Lindsay, Quantum stochastic analysis - an introduction, in, "Quantum Independent Incre-
ment Process, I: From Classical Probability to Quantum Stochastic Calculus" (Eds. U. Franz & M.
Schurmann), Lecture Notes in Mathematics 1865, Springer-Verlag, Heidelberg 2005, pp. 181–271.
[LS1] J.M. Lindsay and K. B. Sinha, A quantum stochastic Lie–Trotter product formula, Indian J. Pure
Appl. Math. 41 (2010) no. 1, 313–325.
[LS2] - - , Holomorphic quantum stochastic contraction cocycles, Preprint.
[LS3] - - , Trotter–Kato product formulae for quantum stochastic cocycles, in preparation.
[LW1] J.M. Lindsay and S.J. Wills, Existence, positivity, and contractivity for quantum stochastic flows
with infinite dimensional noise, Probab. Theory Rel. Fields 116 (2000) no. 4, 505–543.
[LW2] - - , Markovian cocycles on operator algebras, adapted to a Fock filtration, J. Funct. Anal. 178
(2000), no. 2, 269–305.
[LW3] - -, Quantum stochastic cocycles and completely bounded semigroups on operators spaces, Int.
Math. Res. Notices (2014) no. 11, 3096–3139.
[Mey] P.-A. Meyer, "Quantum Probability for Probabilists" (2nd Edn.), Lecture Notes in Math. 1538,
Springer-Verlag, Berlin 1995.
[Ouh] E.M. Ouhabaz, "Analysis of Heat Equations on Domains," London Mathematical Society Mono-
graphs 31, Princeton University Press, Princeton, 2005.
[Par] K.R. Parthasarathy, "An Introduction to Quantum Stochastic Calculus," Monographs in Mathem-
atics 85, Birkhauser, Basel 1992.
[PaS] K.R. Parthasarathy and K.B. Sinha, A random Trotter–Kato product formula, "Statistics and Prob-
ability: Essays in Honor of C.R. Rao," eds. G. Kallianpur, Paruchuri R. Krishnaiah & J.K. Ghosh,
North-Holland, Amsterdam, 1982, pp. 553-566.
[ReS] M. Reed and B. Simon, "Methods of Modern Mathematical Physics, I: Functional Analysis (2nd
Edn.), II: Fourier Analysis, Self-Adjointness," Academic Press, New York, 1980, 1975.
[SSS] L. Sahu, M. Schurmann and K.B. Sinha, Unitary processes with independent increments and rep-
resentations of Hilbert tensor algebras, Publ. Res. Inst. Math. Sci. 45 (2009) no. 3, 745–785.
[Sch] M. Schurmann, "White noise on bialgebras," Lecture Notes in Math. 1544, Springer-Verlag, Berlin
1993.
[Sko] A.V. Skorohod, "Asymptotic Methods in the Theory of Stochastic Differential Equations," Trans-
lations of Mathematical Monographs 78, American Mathematical Society, Providence, 1989.
[Tro] H.F. Trotter, On the product of semi-groups of operators, Proc. Amer. Math. Soc. 10 (1959),
545–551.
[Wil] S.J. Wills, On the generators of quantum stochastic operator cocycles, Markov Process. Related
Fields 13 (2007) no. 1, 191–211.
Department of Mathematics & Statistics, Lancaster University, Lancaster LA1 4YF, UK
E-mail address: [email protected]
|
1210.4715 | 1 | 1210 | 2012-10-17T12:32:09 | Hadamard differentiability via G\^ ateaux differentiability | [
"math.FA"
] | Let $X$ be a separable Banach space, $Y$ a Banach space and $f: X \to Y$ a mapping.
We prove that there exists a $\sigma$-directionally porous set $A\subset X$ such that if $x\in X \setminus A$, $f$ is Lipschitz at $x$, and $f$ is G\^ateaux differentiable at $x$, then $f$ is Hadamard differentiable at $x$.
If $f$ is Borel measurable (or has the Baire property) and is G\^ ateaux differentiable at all points, then $f$ is Hadamard differentiable at all points except a set which is $\sigma$-directionally porous set (and so is Aronszajn null, Haar null and $\Gamma$-null). Consequently, an everywhere G\^ ateaux differentiable $f: \R^n \to Y$ is Fr\' echet differentiable except a nowhere dense $\sigma$-porous set. | math.FA | math | HADAMARD DIFFERENTIABILITY VIA G ATEAUX
DIFFERENTIABILITY
LUDEK ZAJ´I CEK
Abstract. Let X be a separable Banach space, Y a Banach space and f : X → Y a
mapping. We prove that there exists a σ-directionally porous set A ⊂ X such that if
x ∈ X \ A, f is Lipschitz at x, and f is Gateaux differentiable at x, then f is Hadamard
differentiable at x. If f is Borel measurable (or has the Baire property) and is Gateaux
differentiable at all points, then f is Hadamard differentiable at all points except a set
which is σ-directionally porous set (and so is Aronszajn null, Haar null and Γ-null).
Consequently, an everywhere Gateaux differentiable f : Rn → Y is Fr´echet differentiable
except a nowhere dense σ-porous set.
2
1
0
2
t
c
O
7
1
]
.
A
F
h
t
a
m
[
1
v
5
1
7
4
.
0
1
2
1
:
v
i
X
r
a
1. Introduction
Hadamard derivative of a mapping f : X → Y between Banach spaces, which is stronger
than Gateaux derivative but weaker that Fr´echet derivative, was applied many times in the
literature. For three formally different but equivalent definitions of Hadamard derivative
see Lemma 2.1 below.
If X is finite-dimensional, then Hadamard derivative coincides
with Fr´echet derivative, but for infinite-dimensional spaces Fr´echet derivative is “much
stronger”, even for Lipschitz f .
On the other hand, if f is Lipschitz, then Hadamard derivative coincides with Gateaux
derivative; in this sense these two types of derivatives are rather close. Using this fact, we
obtain the well-known result that if f is everywhere Gateaux differentiable and has the
Baire property, than f is Hadamard differentiable at all points except a nowhere dense
set. Indeed, [12, Corollary 3.9] easily implies that, under above assumptions, f is locally
Lipschitz on a dense open set U. However, an easy example (see Remark 3.10) shows that
in general (even if X = Y = R) we cannot find U with “measure null” complement.
We prove (Theorem 3.9) that if X is separable and f has the Baire property and is
everywhere Gateaux differentiable, then f is Hadamard differentiable at all points except
a σ-directionally porous set. This is an interesting additional information, since each
σ-directionally porous subset of a separable Banach spaces X is “measure null”:
it is
Aronszajn (=Gauss) null, and so also Haar null, (see [2, p. 164 and Chap. 6]) and also
Γ-null (see [8, Remark 5.2.4]).
As an easy consequence of Theorem 3.9, we obtain (see Theorem 3.12) that an ev-
erywhere Gateaux differentiable f : Rn → Y is Fr´echet differentiable except a σ-porous
set.
The main ingredient of the proof of Theorem 3.9 is Theorem 3.4. It asserts that if X is
separable and f : X → Y , then there exists a σ-directionally porous set A ⊂ X such that
if x ∈ X \ A, f is Lipschitz at x, and f is Gateaux differentiable at x, then f is Hadamard
differentiable at x.
2010 Mathematics Subject Classification. Primary: 46G05; Secondary: 26B05, 49J50.
Key words and phrases. Hadamard differentiability, Gateaux differentiability, Fr´echet differentiability,
σ-porous set, σ-directionally porous set, Stepanoff theorem, Aronszajn null set.
The research was supported by the grant GA CR P201/12/0436.
1
2
LUD EK ZAJ´I CEK
Lemma 3.6 implies that if X is separable and f has the Baire property and is everywhere
Gateaux differentiable, then f is Lipschitz at all points except a σ-directionally porous
set. So, Theorem 3.4 together with Lemma 3.6 imply Theorem 3.9.
Further, Theorem 3.4 shows that the infinite-dimensional version of Stepanoff theo-
rem on Gateaux differentiability from [4] holds also for Hadamard differentiability (see
Corollary 3.14).
2. Preliminaries
In the following, by a Banach space we mean a real Banach space. If X is a Banach
space, we set SX := {x ∈ X : kxk = 1}. The symbol B(x, r) will denote the open ball
with center x and radius r.
In a metric space (X, ρ), the system of all sets with the Baire property is the smallest
σ-algebra containing all open sets and all first category sets. We will say that a mapping
f : (X, ρ1) → (Y, ρ2) has the Baire property if f is measurable with respect to the σ-
algebra of all sets with the Baire property. In other words, f has the Baire property, if
and only if f −1(B) has the Baire property for all Borel sets B ⊂ Y (see [6, Section 32]).
Let X, Y be Banach spaces, G ⊂ X an open set, and f : G → Y a mapping.
We say that f is Lipschitz at x ∈ G if lim supy→x
pointwise Lipschitz if f is Lipschitz at all points of G.
ky−xk < ∞. We say that f is
kf (y)−f (x)k
The directional and one-sided directional derivatives of f at x ∈ G in the direction
v ∈ X are defined respectively by
f ′(x, v) := lim
t→0
f (x + tv) − f (x)
t
and f ′
+(x, v) := lim
t→0+
f (x + tv) − f (x)
t
.
The Hadamard directional and one-sided directional derivatives of f at x ∈ G in the
direction v ∈ X are defined respectively by
f ′
H(x, v) := lim
z→v,t→0
t
f (x + tv) − f (x)
It is easy to see that f ′(x, v) (resp. f ′
(resp. f ′
H+(x, −v) = −f ′
H+(x, v)).
and f ′
H+(x, v) := lim
z→v,t→0+
H (x, v)) exists if and only if f ′
f (x + tv) − f (x)
.
t
+(x, −v) = −f ′
+(x, v)
It is well-known and easy to prove that, if f is locally Lipschitz on G, then f ′(x, v) =
f ′
H(x, v) (resp. f ′
The usual modern definition of the Hadamard derivative is the following:
A continuous linear operator L : X → Y is said to be a Hadamard derivative of f at a
H+(x, v)) whenever one of these two derivatives exists.
+(x, v) = f ′
point x ∈ X if
lim
t→0
f (x + tv) − f (x)
t
= L(v)
for each v ∈ X
and the limit is uniform with respect to v ∈ C, whenever C ⊂ X is a compact set. In this
case we set f ′
H (x) := L.
The following fact is well-known (see [11]):
Lemma 2.1. Let X, Y be Banach spaces, ∅ 6= G ⊂ X an open set, x ∈ G, f : G → Y a
mapping and L : X → Y a continuous linear operator. Then the following conditions are
equivalent:
(i) f ′
(ii) f ′
(iii) if ϕ : [0, 1] → X is such that ϕ(0) = x and ϕ′
H(x) := L,
H(x, v) = L(v) for each v ∈ X,
+(0) exists, then (f ◦ ϕ)′
+(0) =
L(ϕ′
+(0)).
HADAMARD DIFFERENTIABILITY VIA G ATEAUX DIFFERENTIABILITY
3
Recall (see, e.g. [11]), that if f is Hadamard differentiable at a ∈ G, then f is Gateaux
differentiable at a, and
(2.1)
if f is locally Lipschitz on G, then also the opposite implication holds.
Further,
(2.2)
if X = Rn, then Hadamard differentiability is equivalent to Fr´echet differentiability.
Definition 2.2. Let X be a Banach space. We say that A ⊂ X is directionally porous
at a point x ∈ X, if there exist 0 6= v ∈ X, p > 0 and a sequence tn → 0 of positive real
numbers such that B(x + tnv, ptn) ∩ A = ∅. (In this case we say that A is porous at x in
direction v.)
We say that A ⊂ X is directionally porous if A is directionally porous at each point
x ∈ A.
We say that A ⊂ X is σ-directionally porous if it is a countable union of directionally
porous sets.
Recall that directional porosity is stronger than (upper) porosity, but
(2.3)
if X is finite-dimensional, then these two notions are equivalent.
Let X be a Banach space, x ∈ X, v ∈ SX and δ > 0. Then we define the open cone
C(x, v, δ) as the set of all y 6= x for which kv − y−x
ky−xk k < δ.
The following easy inequality is well known (see e.g. [9, Lemma 5.1]):
(2.4)
if u, v ∈ X \ {0}, then
(cid:13)(cid:13)(cid:13)(cid:13)
u
kuk
−
v
kvk
2
kuk
ku − vk.
(cid:13)(cid:13)(cid:13)(cid:13) ≤
Because of the lack of a reference we supply the proof of the following easy fact.
Lemma 2.3. Let X be a Banach space, Y a Banach space, G ⊂ X an open set, a ∈ G,
and f : G → Y a mapping. Then the following are equivalent.
H+(a, 0) exists,
H(a, 0) exists,
H(a, 0) = 0,
(i) f ′
(ii) f ′
(iii) f ′
(iv) f is Lipschitz at a.
Proof. The implications (ii) ⇒ (i) and (iii) ⇒ (ii) are trivial.
To prove (iv) ⇒ (iii), suppose that K > 0, δ > 0 and kf (x) − f (a)k ≤ Kkx − ak for
each x ∈ B(a, δ). To prove (iii), let ε ∈ (0, K) be given. Then, for z ∈ X with kzk < ε/K
and 0 < t < δ, we obtain
kf (a + tz) − f (a)k ≤ Kt
ε
K
,
and (iii) follows.
(cid:13)(cid:13)(cid:13)(cid:13)
f (a + tz) − f (a)
t
(cid:13)(cid:13)(cid:13)(cid:13) ≤ ε,
To prove (i) ⇒ (iv), suppose that (iv) does not hold. Then, for each n ∈ N, choose
hn ∈ X such that 0 < khnk < n−2 and kf (a + hn) − f (a)k ≥ n2khnk. Set tn := nkhnk and
zn := n−1khnk−1hn. Then zn → 0, tn → 0, tn > 0, and
(cid:13)(cid:13)(cid:13)(cid:13)
f (a + tnzn) − f (a)
tn
kf (a + hn) − f (a)k
nkhnk
≥ n,
(cid:13)(cid:13)(cid:13)(cid:13) =
which clearly implies that f ′
H+(a, 0) does not exist.
(cid:3)
4
LUD EK ZAJ´I CEK
We will need also the following special case of [12, Lemma 2.4]. It can be proved by the
Kuratowski-Ulam theorem (as is noted in [12]), but the proof given in [12] is more direct.
Lemma 2.4. Let U be an open subset of a Banach space X. Let M ⊂ U be a set residual
in U and z ∈ U . Then there exists a line L ⊂ X such that z is a point of accumulation
of M ∩ L.
3. Results
Proposition 3.1. Let X be a separable Banach space, Y a Banach space, G ⊂ X an open
set, and f : G → Y a mapping. Let A be the set of all points x ∈ G for which there exists
ky−xk < ∞ and f is not Lipschitz at
a cone C = C(x, v, δ) such that lim supy→x,y∈C
x. Then A is a σ-directionally porous set.
kf (y)−f (x)k
Proof. Let {vn : n ∈ N} be a dense subset of SX. For natural numbers k, p, n, m denote
by Ak,p,n,m the set of all points x ∈ A such that
(3.1)
kf (y) − f (x)k
ky − xk
< k whenever 0 < ky − xk <
1
p
and
y − x
ky − xk
− vn
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13) <
1
m
.
Since clearly A is the union of all sets Ak,p,n,m, it is sufficient to prove that each set Ak,p,n,m
is directionally porous. So, suppose that natural numbers k, p, n, m and x ∈ Ak,p,n,m are
given. It is sufficient to prove that Ak,p,n,m is directionally porous at x.
To this end, find a sequence yi → x such that
(3.2)
kf (yi) − f (x)k
kyi − xk
> k(12m + 4) for each i ∈ N.
Set ri := kyi − xk and xi := x − 6mrivn. It is sufficient to prove that there exists i0 ∈ N
such that
(3.3)
B(xi, ri) ∩ Ak,p,n,m = ∅ for each i ≥ i0.
To this end, consider i ∈ N and zi ∈ B(xi, ri) ∩ Ak,p,n,m.
Observe that
6mri − ri ≤ kx − xik − kxi − zik ≤ kx − zik ≤ kx − xik + kxi − zik ≤ 6mri + ri,
6mri − 2ri ≤ kx − zik − ri ≤ kyi − zik ≤ kx − zik + ri ≤ 6mri + 2ri.
These inequalities imply that there exists i0 ∈ N (independent on i) such that
(3.4)
0 < kx − zik <
1
p
and 0 < kyi − zik <
1
p
,
if i ≥ i0.
Applying (2.4), we obtain
(3.5)
x − zi
kx − zik
− vn
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13) =
(cid:13)(cid:13)(cid:13)(cid:13)
x − zi
kx − zik
−
x − xi
kx − xik
(cid:13)(cid:13)(cid:13)(cid:13) ≤ 2
kxi − zik
6mri − ri
<
2ri
6mri − ri
≤
1
m
and
(3.6)(cid:13)(cid:13)(cid:13)(cid:13)
yi − zi
kyi − zik
− vn
(cid:13)(cid:13)(cid:13)(cid:13) =
(cid:13)(cid:13)(cid:13)(cid:13)
yi − zi
kyi − zik
−
x − xi
kx − xik
(cid:13)(cid:13)(cid:13)(cid:13) ≤ 2
kyi − xk + kxi − zik
kyi − zik
< 2
2ri
6mri − 2ri
≤
1
m
.
Since zi ∈ Ak,p,n,m, conditions (3.1), (3.4), (3.5) and (3.6) imply that, if i ≥ i0, then
kf (x) − f (zi)k
kx − zik
< k and
kf (yi) − f (zi)k
kyi − zik
< k.
HADAMARD DIFFERENTIABILITY VIA G ATEAUX DIFFERENTIABILITY
5
So, using also (3.2), we obtain that, if i ≥ i0, then
rik(12m + 4) ≤ kf (x) − f (yi)k ≤ kf (x) − f (zi)k + kf (yi) − f (zi)k
< kkx − zik + kkyi − zik ≤ k(6mri + ri + 6mri + 2ri),
which is impossible. So we have proved (3.3).
(cid:3)
Proposition 3.2. Let X be a separable Banach space, Y a Banach space, G ⊂ X an
open set, and f : G → Y a mapping. Let M be the set of all x ∈ G at which f is Lipschitz
and there exists v ∈ X such that f ′
H+(x, v) does not exist. Then M is
σ-directionally porous.
+(x, v) exists but f ′
Proof. For each k ∈ N, set
Mk := {x ∈ M : kf (y) − f (x)k ≤ kky − xk whenever ky − xk < 1/k}.
It is clearly sufficient to prove that each Mk is directionally porous. So suppose that
k ∈ N and x ∈ Mk. Choose v ∈ X such that f ′
H+(x, v) does not
exist. Lemma 2.3 gives that v 6= 0. We will prove that Mk is porous at x in direction v.
Since f ′
H+(x, v) does not exist, we can choose ε > 0 such that for each δ > 0 there exists
w ∈ B(v, δ) and t ∈ (0, δ) with kt−1(f (x + tw) − f (x)) − yk > ε.
+(x, v) =: y exists but f ′
For each n ∈ N, we choose wn ∈ B(v, 1/n) and tn ∈ (0, 1/n) such that
(3.7)
Set xn := x + tnv and p := ε(3k)−1. It is sufficient to prove that there exists n0 ∈ N such
that
k(tn)−1(f (x + tnwn) − f (x)) − yk > ε.
(3.8)
B(xn, ptn) ∩ Mk = ∅ for each n ≥ n0.
To this end consider n ∈ N and zn ∈ B(xn, ptn) ∩ Mk. Denote x∗
n := x + tnwn. We have
kzn − xnk < ptn,
(3.9)
Thus there exists en0 ∈ N (independent on n) such that the inequality n ≥ en0 implies
kzn − xnk < 1/k, kzn − x∗
nk < 1/k, and thus, since zn ∈ Mk, also
nk + kzn − xnk ≤ tn/n + ptn.
nk ≤ kxn − x∗
kzn − x∗
kf (xn) − f (zn)k ≤ kkxn − znk and kf (x∗
(3.10)
If n ≥ en0, then (3.7), (3.9) and (3.10) imply
n) − f (zn)k ≤ kkx∗
n − znk.
ε <
≤
≤
=
− y
− y
f (x∗
kf (x∗
n) − f (x)
f (xn) − f (x)
f (xn) − f (x)
kf (xn) − f (zn)k
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
− y(cid:13)(cid:13)(cid:13) → 0, we obtain a contradiction, if n > n0, where n0 ≥ en0 is
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13) +
(cid:13)(cid:13)(cid:13)(cid:13) + kp + k(1/n + p)
(cid:13)(cid:13)(cid:13)(cid:13) + (2/3)ε + k/n.
f (xn) − f (x)
n) − f (zn)k
− y
tn
tn
tn
tn
− y
tn
+
tn
Since (cid:13)(cid:13)(cid:13) f (xn)−f (x)
tn
sufficiently large (independent on n). So (3.8) is proved.
(cid:3)
Remark 3.3. The corresponding “two-sided” result which works with f ′(x, v) and f ′
clearly follows from Proposition 3.2.
H (x, v)
6
LUD EK ZAJ´I CEK
Using Proposition 3.2 (together with Remark 3.3) and Lemma 2.1, we immediately
obtain:
Theorem 3.4. Let X be a separable Banach space, Y a Banach space, G ⊂ X an open
set and f : G → Y a mapping. Then the set of all points at which f is Lipschitz and
Gateaux differentiable but is not Hadamard differentiable is σ-directionally porous.
Theorem 3.4 together with [12, Corollary 3.9] have the following consequence.
Corollary 3.5. Let X be a separable Banach space, Y a Banach space, G ⊂ X an open
set, and f : G → Y a pointwise Lipschitz mapping. Then the set of all points from G at
which f is Gateaux differentiable but is not Hadamard differentiable is nowhere dense and
σ-directionally porous.
Lemma 3.6. Let X be a separable Banach space, Y a Banach space, G ⊂ X an open
set, and f : G → Y a mapping. Suppose that f has the Baire property and, for each line
L ⊂ X with G ∩ L 6= ∅, the restriction of f to G ∩ L is continuous. Denote by A the set of
all points x ∈ G for which there exists a set Bx ⊂ SX of the second category in SX such
that
(3.11)
lim sup
t→0+
kf (x + tv) − f (x)k
t
< ∞ for each
v ∈ Bx
and f is not Lipschitz at x. Then A is a σ-directionally porous set.
Proof. By Proposition 3.1 it is sufficient to prove that for each x ∈ A there exists v ∈ SX
and δ > 0 such that
(3.12)
lim sup
y→x,y∈C(x,v,δ)
kf (y) − f (x)k
ky − xk
< ∞.
So fix a point x ∈ A. Find p0 ∈ N with B(x, 1/p0) ⊂ G and, for natural numbers p ≥ p0,
k, denote
Sp,k := {v ∈ SX :
kf (x + tv) − f (x)k
t
≤ k for each 0 < t < 1/p.}
Since Bx is clearly covered by all sets Sp,k, we can find k and p ≥ p0 such that Sp,k is a
second category set (in SX).
Without any loss of generality, we can clearly suppose that x = 0 and f (x) = f (0) = 0.
We will show that S := Sp,k has the Baire property (in SX). Set E := B(0, 1/p) \ {0}
and E∗ := SX × (0, 1/p) (equipped with the product topology). For each (v, t) ∈ E∗
set ϕ((v, t)) := tv. Then ϕ : E∗ → E is clearly a homeomorphism (with ϕ−1(z) =
(z/kzk, kzk) for z ∈ E). Set
M := {z ∈ E :
kf (z)k
kzk
≤ k} = {z ∈ E : kf (z)k − kkzk ≤ 0}.
Since f has the Baire property, using continuity of norms we easily obtain that the real
function z 7→ kf (z)k−kkzk has the Baire property on E, and so M has the Baire property
in E. Consequently M ∗ := ϕ−1(M) and C := E∗ \ M ∗ have the Baire property in E∗.
For any set A ⊂ E∗ and v ∈ SX define the section Av := {t ∈ (0, 1/p) : (v, t) ∈ A} and
the projection π(A) := {v ∈ SX : Av 6= ∅}. It is easy to see that
(3.13)
S = SX \ π(C).
Using the continuity of f on the set {tv : t ∈ (0, 1/p)}, we obtain that
(3.14)
Cv
is open for each v ∈ SX.
HADAMARD DIFFERENTIABILITY VIA G ATEAUX DIFFERENTIABILITY
7
Since C has the Baire property, we can write C = H ∪ T , where H is a Gδ set (in E∗)
and T is a first category set (in E∗). We have
(3.15)
π(C) = π(H) ∪ π(T ) = π(H) ∪ (π(T ) \ π(H)).
The set π(H) is analytic (see, e.g [5, Exercise 14.3]), and so (see [5, Theorem 21.6]) it
has the Baire property. By Kuratowski-Ulam theorem (see, e.g. [5, Theorem 8.41]) there
exists a first category set Z ⊂ SX such that Tv is a first category subset of (0, 1/p) for
each v ∈ SX \ Z.
The inclusion π(T ) \ π(H) ⊂ Z holds.
Indeed, suppose on the contrary that there
exists v ∈ (π(T ) \ π(H)) \ Z. Then Tv is a nonempty first category set and Hv = ∅. Thus
Cv = Hv ∪ Tv = Tv. Using (3.14), we obtain that Cv is a nonempty open first category
set, which is a contradiction.
Consequently π(T ) \ π(H) is a first category set and thus S has the Baire property
by (3.13) and (3.15). Since S is also a second category set, we can choose v ∈ SX and
δ > 0 such that S = Sp,k is residual in SX ∩ B(v, δ). Set U ∗ := (SX ∩ B(v, δ)) × (0, 1/p),
ψ := ϕ ↾U ∗ and U := C(0, v, δ)∩B(0, 1/p). Then ψ : U ∗ → U is clearly a homeomorphism
and ψ(S × (0, 1/p)) ⊂ M. Since S × (0, 1/p) is residual in U ∗ (see [5, Lemma 8.43]), we
obtain that M is residual in U. Now consider an arbitrary z ∈ U. By Lemma 2.4 there
exists a line L ⊂ X and points zn ∈ M ∩ L ∩ U with zn → z. Since the restriction of f
to L ∩ U is continuous, we obtain kf (zn)k
kzk , and consequently z ∈ M. So U ⊂ M,
which implies (3.12).
(cid:3)
kznk → kf (z)k
In fact, we will use only the following immediate consequence of Lemma 3.6.
Lemma 3.7. Let X be a separable Banach space, Y a Banach space, G ⊂ X an open
set, and f : G → Y a mapping. Suppose that f has the Baire property and, for each line
L ⊂ X with G ∩ L 6= ∅, the restriction of f to G ∩ L is continuous. Denote by B the set
of all points x ∈ G at which f is Gateaux differentiable and f is not Lipschitz at x. Then
B is a σ-directionally porous set.
As an immediate application of Lemma 3.7 and Theorem 3.4 we obtain
Proposition 3.8. Let X be a separable Banach space, Y a Banach space, G ⊂ X an
open set, and f : G → Y a mapping. Suppose that f has the Baire property and, for each
line L ⊂ X with G ∩ L 6= ∅, the restriction of f to G ∩ L is continuous. Then the set of
all points x ∈ G at which f is Gateaux differentiable but is not Hadamard differentiable
is σ-directionally porous.
As already mentioned in Introduction, [12, Corollary 3.9] easily implies that if f is
everywhere Gateaux differentiable and has the Baire property, then f is locally Lipschitz
on a dense open set U, and so (see (2.1)) f is Hadamard differentiable at all points except
a nowhere dense set (even if X is nonseparable).
This result together with Proposition 3.8 give the following theorem.
Theorem 3.9. Let X be a separable Banach space, Y a Banach space, G ⊂ X an
open set. Suppose that f : G → Y has the Baire property and is everywhere Gateaux
differentiable. Then f is Hadamard differentiable at all points of G except a nowhere
dense σ-directionally porous set.
Remark 3.10. The assumptions of Theorem 3.9 do not imply that f is locally Lipschitz
on the complement of a closed σ-directionally porous set even in the case X = Y = R.
8
LUD EK ZAJ´I CEK
Indeed, let (an, bn), n ∈ N be the system of bounded intervals such that G := Sn∈N(an, bn)
is dense in R and F := R \ G has positive Lebesgue measure. We can clearly define a
function f on R such that
(i) f (x) = 0 for x ∈ F ,
(ii) f has continuous derivative on each (an, bn),
(iii) f (x) ≤ (x − an)2(x − bn)2 for x ∈ (an, bn),
(iv) for each n ∈ N and δ > 0, the derivative f ′ is unbouded both on (an, an + δ) and
on (bn − δ, bn).
Then f is everywhere differentiable, G is the largest open set on which f is locally Lipschitz
and F is not σ-directionally porous. So Theorem 3.9 cannot be proved simply using
only the fact that Gateaux and Hadamard differentiability coincide for locally Lipschitz
functions.
Remark 3.11. In Theorem 3.9, we cannot omit the assumption that f has the Baire
property. This follows from Lemma 2.3 and Shkarin’s [12] example which shows that
on each separable infinite-dimensional Banach space there exists an everywhere Gateaux
differentiable real function which is discontinuous at all points.
However, if X is finite-dimensional, the assumption that f has the Baire property can
be omitted.
Indeed (as is also noted without a reference in [12, Remark 3.5]), if f is
everywhere Gateaux differentiable on Rn, than f has the Baire property. This follows,
e.g., from the old well-known fact (see [7]) that a partially continuous real function on Rn
is in the (n − 1)st Baire class (whose proof works also for a Banach space valued function).
So, using also (2.2) and (2.3), we obtain the following result.
Theorem 3.12. Let G be an open subset of Rn, Y a Banach space, and f : G → Y an
everywhere Gateaux differentiable mapping. Then f is Fr´echet differentiable at all points
of G except a nowhere dense σ-porous set.
Using Theorem 3.4 and the main result of [4] (which generalizes the result of [10] and
improves the result of [3]), we easily obtain Theorem 3.13 below. In its formulation, the
system A of subsets of a separable Banach space is used. This system which was defined in
[10] is strictly smaller (see [10, Proposition 13]) than the well-known system of Aronszajn
null sets (and also than the system of Γ-null sets, see [13]). We will not recall the (slightly
technical) definition of the system A. Note only, that all members of A are Borel by
definition and that it is easy to see that A is stable with respect to countable unions.
Theorem 3.13. Let X be a separable Banach space, Y a Banach space with the Radon-
Nikod´ym property, G ⊂ X an open set, and f : G → Y a mapping. Then there exists a set
A ∈ A such that if x ∈ G \ A and f is Lipschitz at x, then f is Hadamard differentiable
at x.
Proof. The main result of [4] says that there exists a set A1 ∈ A such that if x ∈ G \ A1
and f is Lipschitz at x, then f is Gateaux differentiable at x. Let A∗
2 be the set of all
x ∈ G at which f is Lipschitz, Gateaux differentiable, but not Hadamard differentiable.
Then A∗
2 is σ-directionally porous by Theorem 3.4. Using [10, Proposition 14] and [10,
Theorem 12] we easily obtain that there exists A2 ∈ A which contains A∗
2. Now it is
clearly sufficient to set A := A1 ∪ A2.
(cid:3)
As a corollary, we obtain an infinite dimensional analogue of Stepanoff theorem on
Hadamard differentiability.
HADAMARD DIFFERENTIABILITY VIA G ATEAUX DIFFERENTIABILITY
9
Corollary 3.14. Let X be a separable Banach space, Y a Banach space with the Radon-
Nikod´ym property, G ⊂ X an open set, and f : G → Y a pointwise Lipschitz mapping.
Then f is Hadamard differentiable at all points of G except a set belonging to A.
Remark 3.15. Corollary 3.14 and (2.2) imply the following result.
Let Y a Banach space with the Radon-Nikod´ym property, G ⊂ Rn an open set, and
f : G → Y a pointwise Lipschitz mapping. Then f is Fr´echet differentiable at all points
of G except a set belonging to A.
However, it is not probably an improvement of the classical Stepanoff theorem, since
from some unpublished results follows that each Borel Lebesgue null subset of Rn belongs
to A. For n = 2 this follows from [1, Theorem 7.5] (via [10, Theorem 12]), which is stated
without a proof, and for n ≥ 3 from the corresponding theorem recently announced by
M. Csornyei and P. Jones.
References
[1] G. Alberti, M. Csornyei, D. Preiss, Structure of null sets in the plane and applications, European
Congress of Mathematics, 3-22, Eur. Math. Soc., Zurich, 2005.
[2] Y. Benyamini, J. Lindenstrauss, Geometric Nonlinear Functional Analysis, Vol. 1, Colloquium
Publications 48, American Mathematical Society, Providence, 2000.
[3] D. Bongiorno, Stepanoff ’s theorem in separable Banach spaces, Comment. Math. Univ. Carolin.
39 (1998), 323-335.
[4] J. Duda, On Gateaux differentiability of pointwise Lipschitz mappings, Canad. Math. Bull. 51
(2008), 205-216.
[5] A.S. Kechris, Classical Descriptive Set Theory, Springer, New York, 1994.
[6] K. Kuratowski, Topology, vol. I, Academic Press, New York, 1966,
[7] H. Lebesgue, Sur les fonctions repr´esentable analytiquement, J. Math. Pure Appl. (6), vol. 1
(1905), 139–212.
[8] J. Lindenstrauss, D. Preiss, J. Tiser, Fr´echet differentiability of Lipschitz maps and porous sets
in Banach spaces, Princeton University Press, Princeton 2012.
[9] J. L. Massera, J. J. Schaffer, Linear differential equations and functional analysis, I, Ann. of
Math. (2) 67 (1958), 517–573.
[10] D. Preiss, L. Zaj´ıcek, Directional derivatives of Lipschitz functions, Israel J. Math. 125 (2001),
1-27.
[11] A. Shapiro, On concepts of directional differentiability, J. Optim. Theory Appl. 66 (1990), 477-
487.
[12] S.A. Shkarin, Points of discontinuity of Gateaux-differentiable mappings, (Russian), Sibirsk.
Mat. Zh. 33 (1992), 176–185.
[13] L. Zaj´ıcek, On sets of non-differentiability of Lipschitz and convex functions, Math. Bohem. 132
(2007), 75-85.
E-mail address: [email protected]
Charles University, Faculty of Mathematics and Physics, Sokolovsk´a 83, 186 75 Praha
8-Karl´ın, Czech Republic
|
1811.10987 | 1 | 1811 | 2018-11-27T14:00:13 | Frechet algebras in abstract Harmonic analysis | [
"math.FA"
] | We provide a survey of the similarities and differences between Banach and Frechet algebras including some known results and examples. We also collect some important generalizations in abstract harmonic analysis; for example, the generalization of the concepts of vector-valued Lipschitz algebras, abstract Segal algebras, Arens regularity, amenability, weak amenability, ideal amenability, etc. | math.FA | math |
FR´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
Z. ALIMOHAMMADI AND A. REJALI
Abstract. We provide a survey of the similarities and differences between
Banach and Fr´echet algebras including some known results and examples.
We also collect some important generalizations in abstract harmonic anal-
ysis; for example, the generalization of the concepts of vector-valued Lips-
chitz algebras, abstract Segal algebras, Arens regularity, amenability, weak
amenability, ideal amenability, etc.
0. Introduction
The class of Fr´echet algebras which is an important class of locally convex
algebras has been widely studied by many authors. For a full understanding of
Fr´echet algebras, one may refer to [27, 36]. In the class of Banach algebras, there
are many concepts which were generalized to the Fr´echet case. Examples of these
concepts are as follows.
Biprojective Banach algebras were studied by several authors, notably A. Ya.
Helemskii [35] and Yu. V. Selivanov [62]. A number of papers appeared in the
literature concerned with extending the notion of biprojectivity and its results to
the Fr´echet algebras; see for example [48, 63].
Flat cyclic Fr´echet modules were investigated by A. Yu. Pirkovskii [49]. Let A
be a Fr´echet algebra and I be a closed left ideal in A. Pirkovskii showed that there
exists a necessary and sufficient condition for a cyclic Fr´echet A-module A♯/I to
be strictly flat. He also generalized a number of characterizations of amenable Ba-
nach algebras to the Fr´echet algebras (ibid.). In 2008, P. Lawson and C. J. Read
[43] introduced and studied the notions of approximate amenability and approxi-
mate contractibility of Fr´echet algebras. Recently, the concepts of ϕ-amenability
and character amenability, ϕ-contractibility and character contractibility, weak
amenability and Ideal amenability for Fr´echet algebras have investigated in sev-
eral papers; see [7, 8, 11, 55].
T. G. Honary [37] gave an extension of Johnson's uniqueness of norm theorem.
He showed that if A and B are Fr´echet algebras, B is semisimple and T : A → B
is a surjective homomorphism with a certain condition, then T is continuous. In
special case, if A is a Banach algebra, then every epimorphism T : A → B is
automatically continuous. Hence, every semisimple Banach algebra has a unique
topology as a Fr´echet algebra. Honary then presented a partial answer to the
well-known Michael's problem posed in [46].
Let (X, d) be a metric space and E be a Banach algebra over C. Take α ∈ R
with α > 0. E. Biyabani and A. Rejali [18] studied the structure and proper-
ties of vector-valued Lipschitz algebras Lipα(X, E) and lipα(X, E) of order α.
Later, these concepts were generalized to the class of Fr´echet algebras by Re-
jali joint with Ranjbari [54]. They also studied the concept of ideal amenability
for vector-valued Fr´echet Lipschitz algebras (ibid.). Furthermore, Rejali et al.
2010 Mathematics Subject Classification. 46H05, 46H25, 46J05,43A20.
Key words and phrases. locally convex space, Fr´echet algebra, Banach algebra.
1
2
Z. ALIMOHAMMADI AND A. REJALI
[9, 10] introduced and studied the notions of Segal and semisimple Segal Fr´echet
algebras. They showed that every continuous linear left multiplier of a Fr´echet
algebra (B, qn)n∈N is also a continuous linear left multiplier of any Segal Fr´echet
algebra (A, pℓ)ℓ∈N in (B, qn)n∈N. Moreover, they proved that if B is a commutative
Fr´echet Q-algebra with an approximate identity, then the spaces of all modular
maximal closed ideals of B and any Segal Fr´echet algebra A in B are homeomor-
phic. Particularly, B is semisimple if and only if A is semisimple [10]. In [2], we
joint with Abtahi introduced and studied the strict, uniform, and compact-open
locally convex topologies on an arbitrary algebra B, by the fundamental system
of seminorms of a locally convex subalgebra (A, pα). In particular, we studied the
strict topology on the multiplier algebra M (A) of a Fr´echet algebra A, and we
showed that A is a M (A)-Segal algebra.
In the present paper we give a list of the significant results which have been
obtained by various authors. Some of the results in this paper were included in
the first author's Ph.D. thesis, which was supported by the center of excellence
for mathematics at the University of Isfahan.
1. Preliminaries
In this section we give some basic definitions and frameworks related to locally
convex spaces which will be required throughout the paper; see [27, 36, 45] for
more information. In this paper, all vector spaces and algebras are assumed to be
over the field C of complex numbers.
A locally convex space E is a topological vector space in which the origin has a
local base of absolutely convex absorbent sets. Throughout the paper, all locally
convex spaces are assumed to be Hausdorff. Every locally convex Hausdorff space
E has a fundamental system of seminorms (pα)α∈Λ, or equivalently a family of
continuous seminorms satisfying the following properties:
(i) For every x ∈ E with x 6= 0, there exists an α ∈ Λ with pα(x) > 0;
(ii) For all α, β ∈ Λ, there exist γ ∈ Λ and C > 0 such that
max(pα(x), pβ (x)) ≤ Cpγ(x)
(x ∈ E).
We denote by (E, pα)α∈Λ, the locally convex space with the fundamental system
of seminorms (pα)α∈Λ.
Suppose that (E, pα)α∈Λ and (F, qβ )β∈B are locally convex spaces. For every
linear mapping T : E → F , the following assertions are equivalent:
(i) T ∈ L(E, F ), i.e. T is continuous;
(ii) T is continuous at 0;
(iii) For each β ∈ B there exist an α ∈ A and C > 0, such that
qβ(T (x)) ≤ Cpα(x)
(x ∈ E).
Following [45], a complete metrizable locally convex space is called a Fr´echet
space. Indeed, a Fr´echet space is a locally convex space which has a countable
fundamental system of seminorms (pℓ)ℓ∈N. It is worth noting to remark that due
to Conway [23], Fr´echet spaces are not assumed to be locally convex. These two
different definitions make some basic differences. Consider Lp(0, 1) (0 < p < 1),
the space of all equivalence classes of measurable functions f : (0, 1) → R such
0 f (x)pdx < ∞. Under the translation metric
thatR 1
d(f, g) =Z 1
0
(f − g)xpdx < ∞
(f, g ∈ Lp(0, 1)),
Lp(0, 1) is a complete metric space. Therefore, according to the definition pre-
sented in [23], Lp(0, 1) is a Fr´echet space. However, Lp(0, 1) has only one nonempty
the seriesP∞
k=1 λkxk converges in A.
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
3
open convex set, namely itself. It follows that due to the definition given in [45],
Lp(0, 1) is not a Fr´echet space. In this paper, our definition of a Fr´echet space is
based on that given in [45].
A Fr´echet space A is called quasinormable if for each zero neighborhood U
there is a zero neighborhood V so that for every ε > 0 there exists a bounded set
B in A such that V ⊆ B + εU .
It should be noted that, by [45, Proposition 25.7], the strong dual space of every
Fr´echet space A is a (DF)-space which is a type of locally convex topological vector
spaces; see [45, page 297] for details. We denote the strong dual of A by A∗.
Following [45, Corollary 24.29], every Fr´echet space A has a web {Cn1,··· ,nk },
that is a family Cn1,··· ,nk , n1, · · · , nk ∈ N, k ∈ N, of absolutely convex subsets of
A with the following properties:
n=1Cn = A;
n=1Cn1,··· ,nk,n = Cn1,··· ,nk for all n1, · · · , nk ∈ N and all k ∈ N;
(i) ∪∞
(ii) ∪∞
(iii) for each sequence (nk)k∈N in N there exists a sequence (λk)k∈N in (0, ∞),
so that for every sequence (xk)k∈N in A with xk ∈ Cn1,··· ,nk,n for all k ∈ N
A barrel set in a locally convex space E is a subset which is convex, balanced,
absorbing and closed. Moreover, E is called barrelled if each barrel in E is a zero
neighborhood. By [45, Proposition 23.23], every Fr´echet space is barrelled.
Let Λ be an index set directed under a (reflexive, transitive, anti-symmetric)
relation ≺, let (Eα)α∈Λ be a family of locally convex spaces, and denote, for α ≺ β,
by τ β
α a continuous linear map of Eβ into Eα. Consider
E =(cid:8)(xα)α∈Λ ∈ Yα∈Λ
Eα : τ β
α (xβ ) = xα whenever α ≺ β(cid:9).
E is called the projective limit of the family (Eα)α∈Λ with respect to the mappings
τ β
α ). The topology of E is the
α (α, β ∈ Λ, α ≺ β), and denoted by lim
←−
projective topology with respect to the family {(Eα, tα, τα)}α∈Λ, where for each
α ∈ Λ, tα is the topology of Eα and τα : E → Eα is the canonical map. The
projective limit E is called reduced if for each α ∈ Λ, τα has dense range. By
applying [61, II.5.4], every Fr´echet space is isomorphic to a projective limit of a
sequence of Banach spaces.
(Eα, τ β
By a similar way, one can introduce the inductive limit of the family (Eα)α∈Λ.
See also [61, II.6.3] and [45, section 24] for more information about the notion of
inductive topology.
A locally convex space E is called bornological (resp. ultra-bornological) if its
topology is the inductive topology corresponding to a system (Eα → E)α∈Λ of
normed (resp. Banach) spaces. Clearly, every ultra-bornological space is bornolog-
ical. Also, every Fr´echet space is ultra-bornological [45, Remark 24.13].
A topological algebra A is an algebra, which is a topological vector space and
the multiplication A × A → A ((a, b) 7→ ab) is separately continuous. Moreover,
A is called a locally convex algebra if A in addition is a locally convex space.
Let (A, pα)α∈Λ be a locally convex algebra. A left approximate identity (left
a.i. for short) in A is a net (ei)i∈I in A, such that
pα(eia − a) −→i 0
(a ∈ A)
for each α ∈ Λ. The left a.i. is bounded (left b.a.i. for short) if the set {ei : i ∈ I}
is bounded. It is called uniformly bounded (left u.b.a.i. for short) if
sup{pα(ei) : i ∈ I, α ∈ Λ} < ∞.
4
Z. ALIMOHAMMADI AND A. REJALI
"Right a.i.", "right b.a.i.", and "right u.b.a.i." are defined similarly. Furthermore,
for two-sided approximate identities, we use of the symbols "a.i.", "b.a.i.", and
"u.b.a.i.".
Following [27], a complete topological algebra is a Fr´echet algebra if its topology
is produced by a countable family of increasing submultiplicative seminorms.
Let A be a Fr´echet algebra. A left Fr´echet A-module is a Fr´echet space E
together with the structure of an A-module such that the bilinear mapping
A × E → E,
(a, x) 7→ ax
is continuous. Similarly, right Fr´echet A-modules and Fr´echet A-bimodules are
defined. Suppose that E and F are left Fr´echet A-modules (resp. right Fr´echet
A-modules, Fr´echet A-bimodules).
In this paper, the space of all continuous
A-module morphisms from E to F is denoted by AL(E, F ) (resp. LA(E, F ),
ALA(E, F )).
2. Comparison between Banach and Fr´echet algebras
In this section, we compare two Banach and Fr´echet algebras and provide a list
of important similarities and differences between them. We use the symbol "(cid:3)"
after each result to separate them from one another in the list.
2.1. Similarities. Following results are marked "S".
S 2.1. Due to [45, Proposition 22.12], there exists a generalization of the Hahn-
Banach theorem and its consequences for locally convex spaces. Let A be a locally
convex space, B be a subspace of A, and p be a continuous seminorm on A. Then,
the following statements hold.
(i) For each f ∈ B∗ there exists an F ∈ A∗ such that F B = f .
(ii) For each a ∈ A there exists an f ∈ A∗ such that f (a) = p(a) and f ≤ p.
(iii) For each a ∈ A with a 6= 0 there exists an f ∈ A∗ such that f (a) 6= 0.
Example. Consider the space CN with the fundamental system of seminorms
pn(x) = max
1≤m≤n
xm
(x = (xn)n∈N ∈ CN).
Let the subspace B consist of all convergent complex-valued sequences. By argu-
ments above, the map T : B → C, defined by
T (x) = lim
n
xn
(x = (xn)n∈N ∈ B),
has a continuous linear extension to CN. This extension is denoted by LIM; see
[33, Theorem 1], [64, III.9] and also [54, section 4] for details. (cid:3)
Before the next result, let us recall some notation. Let A be a locally convex
space. The polars, of nonempty sets M ⊆ A and N ⊆ A∗, are defined by
M ◦ = (cid:8)f ∈ A∗ : f (a) ≤ 1 for all a ∈ M(cid:9) ⊆ A∗, and
◦N = (cid:8)a ∈ A : f (a) ≤ 1 for all f ∈ N(cid:9) ⊆ A.
S 2.2. Following [45, Theorem 22.13], the bipolar theorem is valid in the class
of locally convex spaces. Let A be a locally convex space and B be an absolutely
convex subset of A. Then, ◦(B◦) = B. (cid:3)
S 2.3. We know that if A is a normed space, then by applying Banach-Alaoglu
theorem, the closed unit ball of A∗ is weak∗-compact. The Alaoglu-Bourbaki the-
orem is a generalization by Bourbaki to dual topologies on locally convex spaces.
Let A be a locally convex space. Then, the polar U ◦ of any neighborhood U in A
is weak∗-compact; see [45, Theorem 23.5]. (cid:3)
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
5
S 2.4. Let A and B be Banach spaces. Due to the open mapping theorem,
if the map T : A → B is continuous, linear and surjective, then T is open.
A similar result holds for Fr´echet spaces instead of Banach spaces. Following
[45, Theorem 24.30], the open mapping theorem was generalized to the locally con-
vex spaces. Let A and B be locally convex spaces. If A has a web and B is ultra-
bornological, then every continuous, linear, surjective map T : A → B is open. (cid:3)
S 2.5. Consider the continuous map f : A → B between two metric spaces A and
B. Following [45],
G(f ) =(cid:8)(a, f (a)) : a ∈ A(cid:9) ⊆ A × B,
denote the graph of f , is closed. If A and B are also complete, then the converse
is true. In general, the closed graph theorem holds in the class of locally convex
spaces. Suppose that A and B are locally convex spaces. Exactly like S 2.4, if
A has a web and B is ultra-bornological, then every linear map T : A → B with
a closed graph is continuous [45, Theorem 24.31]. In particular, if A and B are
Fr´echet spaces, then a linear mapping of A into B is continuous if and only if its
graph is closed in A × B. (cid:3)
S 2.6. In addition to the open mapping theorem and closed graph theorem, there
is an extension of the Banach-Steinhaus theorem, known as uniform boundedness
principle. Let A and B be locally convex spaces and A be barrelled. Further, let
H ⊆ L(A, B) be pointwise bounded, i.e., for each a ∈ A, {T a : T ∈ H} is bounded
in B. Then, H is equicontinuous. In other words, for each zero neighborhood V
in B there exists a zero neighborhood U in A with T (U ) ⊆ V for all T ∈ H; see
[19, Theorem III.§4.2.1] and [45, Proposition 23.27].
Corollary. Let A and B be locally convex spaces, A be barrelled and (Tn)n∈N be a
sequence in L(A, B). Also, let (Tna)n∈N be convergent in B for each a ∈ A. Then,
the mapping T : A → B, defined by
is continuous [45, page 274]. (cid:3)
T a = lim
n→∞
Tna
(a ∈ A),
S 2.7. Let A be a Fr´echet algebra, E a right Fr´echet A-module, and F a left
Fr´echet A-module. By applying [65, Theorem 2], the completed projective tensor
product of E and F , denoted by E ⊗F , is again a Fr´echet algebra. Now, denote by
E ⊗AF the completion of the quotient (E ⊗F )/N , where N ⊆ E ⊗F is the closed
linear span of all elements of the form
x · a ⊗ y − x ⊗ a · y
(x ∈ E, y ∈ F, a ∈ A).
Following [49, Proposition 3.2], there exists a vector space isomorphism from
(E ⊗AF )∗ to LA(E, F ∗). (cid:3)
Now, let A be a commutative Fr´echet algebra. For each a ∈ A, let a be
the Gelfand transform of a defined by a(ϕ) := ϕ(a), where ϕ is an arbitrary
nonzero complex-valued algebra homomorphism on A. Note that every complex-
valued algebra homomorphism on a Banach algebra is automatically continuous.
However, the question of whether each homomorphism on a Fr´echet algebra is
necessarily continuous was posed by E. A. Michael [46]. This question has not been
solved. But, only partial answers have been obtained so far; for example see [37] as
mentioned in the introduction. See also [27, chapter 10] for details. In this paper,
the set of all nonzero complex-valued and continuous algebra homomorphisms on
A is denoted by ∆(A). Also, the topology on ∆(A) is the Gelfand topology, that
is the coarsest topology such that all Gelfand transforms are continuous functions
on ∆(A).
6
Z. ALIMOHAMMADI AND A. REJALI
S 2.8. Let A be a commutative unital Fr´echet algebra. Following [27, 3.2], ∆(A)
is never empty. Therefore, the Gelfand-Mazur theorem holds for Fr´echet algebras.
In other words, if every nonzero element in A is invertible, then A = C. (cid:3)
Let B be a Banach algebra with a left b.a.i. (not necessarily u.b.a.i.) and let
E be a left Banach B-module. Due to the Cohen factorization theorem, for given
ε > 0 and x ∈ B · E, there are a ∈ B and y ∈ E such that x = a·y and kx−yk < ε.
Moreover, kak < M , for some M ∈ R+; see [25, page 14]. We mention next an
important generalization by Voigt of the Cohen factorization theorem; see [67] for
details.
S 2.9. Let (A, pℓ) be a Fr´echet algebra and let (ei)i∈I be a left u.b.a.i.
for A.
Suppose that E is a left Fr´echet A-module and B ⊆ E is bounded. Let us recall
aco(B) denote the absolutely convex hull of B, that is
aco(B) =(cid:8) nXi=1
λixi : xi ∈ B, λi ∈ C,
nXi=1
λi ≤ 1(cid:9).
Set B := aco(B) and EB := (lin(B), pB), where pB is the gauge (or Minkowski
functional) of B, i.e., the non-negative real function
x 7→ pB(x) = inf{t > 0 : x ∈ tB}
(x ∈ B).
By [67, Theorem 3], if eix −→i x uniformly for x ∈ B, then (A, E) has the
{B}-strong left factorization property (hereafter abbreviated {B}-SLFP). In other
words, for every zero neighborhood U in E, there exist a ∈ A and a continuous
linear mapping T : EB → E such that
(i) a(T x) = x, for all x ∈ B;
(ii) T x belongs to the closed A-submodule of E generated by x, for all x ∈ B;
(iii) T x − x ∈ U , for all x ∈ B;
(iv) T B : B → E is continuous.
Furthermore, a ∈ A can be chosen to satisfy
(v) if ℓ ∈ N is such that sup{pℓ(ei) : i ∈ I} ≤ 1, then pℓ(a) ≤ 1.
Remark. Let B be a collection of bounded (resp. compact) subsets of E. Following
[67], we say that (A, E) has the B-SLFP if for each B ∈ B, (A, E) has the {B}-
SLFP. In particular, if B is the collection of bounded (resp. compact) subsets of
A, then this property is called the bounded (resp. compact) SLFP.
Note that by [67, Corollary 5], if A has a left u.b.a.i., then A has the compact
SLFP. However, factorization properties are proved for certain Fr´echet algebras
of differentiable functions which do not possess a u.b.a.i.; see [68]. (cid:3)
2.2. Differences. Following results are marked "D".
D 2.10. Every Banach algebra (space) is a Fr´echet algebra (space), but in general
the converse is not true. Let (An, k · kn)n∈N be a sequence of normed spaces and
A =Qn∈N An. Define the metric d on A by
kan − bnkn
d(a, b) :=
∞Xn=1
2n(cid:0)1 + kan − bnkn(cid:1)
(cid:0)a = (an)n∈N, b = (bn)n∈N ∈ A(cid:1).
Following [45, Lemma 5.17], (A, d) is a locally convex metric linear space. Also,
if each (An, k · kn) is complete, then (A, d) is a Fr´echet space. However, (A, d)
is not a Banach space if An 6= {0} for infinitely many n ∈ N. For instance, the
space CN of all sequences in C with the metric
d(x, y) =
∞Xn=1
xn − ynn
2n(cid:0)1 + xn − ynn(cid:1)
(cid:0)x = (xn)n∈N, y = (yn)n∈N ∈ CN(cid:1),
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
7
is a Fr´echet space which is not a Banach space; see [45, Examples 5.18]. (cid:3)
D 2.11. The dual space of a Banach space is also a Banach space. Moreover, if
A is a Banach algebra, then the module actions of A are jointly continuous on
A∗. However, this result is not valid for a Fr´echet case. Let C∞(R), denote the
space of all infinitely differentiable functions on R. Following [66], C∞(R) is a
Fr´echet algebra with the pointwise product and the family of countable seminorms
k · kn, defined by
kf kn :=
1
k!
kf (k)k∞,n
(f ∈ C∞(R)).
nXk=0
The module action of C∞(R) is not jointly continuous on C∞(R)∗. Therefore,
C∞(R)∗ is not a Fr´echet space. By arguments above, we immediately get the
following.
Lemma. Let A be a metrizable locally convex space. The following statements
are equivalent:
(i) A is a Banach space;
(ii) A is a Fr´echet and (DF)-space.
Corollary. Let A be a Fr´echet space. Then, A∗ is a Fr´echet space if and only if
A is a Banach space.
See also [27] and [45]. (cid:3)
D 2.12. For Banach spaces A and B, L(A, B) is also a Banach space; see
[45, Proposition 5.6]. However, by using D 2.11 and [45, page 253], L(A, B) is
not necessarily a Fr´echet space when A and B are Fr´echet spaces. (cid:3)
Before proceeding to the next result, recall that a matrix A = (aj,k)j,k∈N of non-
negative numbers is called a Kothe matrix if it satisfies the following conditions.
(i) For each j ∈ N there exists a k ∈ N with aj,k > 0.
(ii) aj,k ≤ aj,k+1 for all j, k ∈ N.
Now, for 1 ≤ p < ∞ define
Also, for p = ∞ and p = 0 define
λp(A) :=(cid:8)x ∈ CN : kxkk :=(cid:0) ∞Xj=1
:= (cid:8)x ∈ CN : kxkk := sup
:= (cid:8)x ∈ λ∞(A) :
< ∞ for all k ∈ N(cid:9).
xj aj,kp(cid:1)1/p
xjaj,k < ∞ for all k ∈ N(cid:9), and
xj aj,k = 0 for all k ∈ N(cid:9),
lim
j→∞
j∈N
λ∞(A)
c0(A)
respectively. Using the definition above, we can present in D 2.13 an example
of non-distinguished Fr´echet space due to Meise and Vogt [45]. Note that a
metrizable locally convex space A is distinguished if its strong dual is a bar-
relled or bornological locally convex space. Consider the Fr´echet space A and
n )n∈N. By applying [45, Remark 25.13], if A is distinguished,
let A = lim
←−
then A∗ ∼= lim
n, (τ m
n )∗)n∈N as two locally convex spaces. In other words, the
−→
inductive topology coincides with the strong topology on A∗ if and only if A is
distinguished. This is an important advantage of distinguished spaces.
(An, τ m
(A∗
D 2.13. Every Banach space is distinguished [61, Corollary II.7.1]. In general,
however, a Fr´echet space fails to be distinguished. For example, consider the Kothe
matrix A = (ai,j;k)(i,j)∈N2,k∈N such that
(i) for each i, j, k ∈ N with k ≤ i, ai,j;k = ai,j;1 > 0;
8
Z. ALIMOHAMMADI AND A. REJALI
(ii) for each m ∈ N, limj→∞ am,j;ma−1
m,j;m+1 = 0.
By using [45, Corollary 27.18], λ1(A) is a Fr´echet space which is not distinguished.
This example is a generalization of a classical counter-example of Grothendieck
and Kothe; see [42, §31.7]. (cid:3)
D 2.14. Following [45, Proposition 7.5], if A is a reflexive Banach space and
B is a closed subspace of A, then A/B is reflexive. Also, every closed subspace
of a reflexive Fr´echet space is reflexive [45, Proposition 23.26]. But, there are
reflexive Fr´echet spaces with non-reflexive quotients. Consider the Kothe matrix
A = (ai,j;k)(i,j)∈N2,k∈N, where
ai,j;k :=((ki)k
kj
for j < k, i ∈ N
for j ≥ k, i ∈ N.
By [45, Example 27.21], for each infinite subset I of N2 and each n ∈ N there exists
k ∈ N such that inf (i,j)∈I ai,j;na−1
i,j;k = 0. Thus, due to Dieudonn´e-Gomes theorem
[45, Theorem 27.9], λ1(A) is reflexive. Since ai,j;1 = 1 and ai,k;k = a1,k;k, λ1(A)
has a quotient which is isomorphic to ℓ1; see [45, Proposition 27.22]. However, by
applying [45, Corollary 7.10], ℓ1 is non-reflexive. (cid:3)
For the next result, we need to remind the reader that a Montel space A is a
quasi-barrelled space in which each bounded set is relatively compact. Let A be the
Kothe matrix defined as in D 2.14. By [45, Theorem 27.9], λp(A) is a Montel space
for each p ∈ [1, ∞]. For this reason, λp(A)∗ is also a Montel space and so is re-
flexive; see [45, Proposition 24.25]. Following the proof of [45, Proposition 27.22],
the continuous linear map Q : λp(A) → ℓp for p ∈ [1, ∞) defined by
Qx = (
2−jxi,j )i∈N
∞Xj=1
(cid:0)x = (xi,j )i,j ∈ λp(A)(cid:1)
is onto. Now, consider Q∗ : (ℓp)∗ = ℓq → λp(A)∗, where q = p
p−1 . By us-
ing [45, Corollary 27.24], rang(Q∗) is a closed subspace of λp(A)∗ which is non-
reflexive. This example shows that not every closed subspace of a reflexive locally
convex space is reflexive.
D 2.15. Let A be a Banach space and B be a closed subspace of A. Following
[45, page 293], A is finite dimensional if and only if it is a Montel space. There-
fore, B and A/B are also Montel spaces. Clearly, for Fr´echet spaces, every closed
subspace of a Montel space is a Montel space. But, Montel property are in gen-
eral not inherited by quotients. For instance, consider the montel space λ1(A) in
D 2.14 and the space ℓ1 which is not Montel. (cid:3)
Before continuing, we must recall that for a continuous linear map T : A → B
between locally convex spaces A and B, there is a unique map ¯T in L(A/ker(T ), B)
such that T = ¯T ◦ q, where q : A → A/ker(T ) is the quotient map [45, Proposition
22.11]. Now, consider rang(T ) with the relative topology induced by B. The map
T is said to be a topological homomorphism, if ¯T is an isomorphism between
A/ker(T ) and rang(T ).
D 2.16. Consider the short sequence of Fr´echet spaces with continuous linear
maps
(2.1)
{0} → A T→ B S→ C → {0}.
By applying [45, Theorem 26.3], if the short sequence (2.1) is exact, then it is
already topologically exact, that is T and S are topological homomorphisms. Also,
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
9
by [45, Proposition 26.4], the short sequence (2.1) is exact if and only if
(2.2)
{0} → C ∗ S ∗
→ B∗ T ∗
→ A∗ → {0}
is exact. However, in general the short sequence (2.2) is not topologically ex-
act. Consider the Kothe matrix A and Q ∈ L(λp(A), ℓp) defined in D 2.14. By
[45, Remark 27.23], the dual sequence of the short exact sequence
{0} → ker(Q) ֒→ λp(A)
Q
→ ℓp → {0}
of Fr´echet spaces is not topologically exact.
Remark. Let A be a Fr´echet space and B be a closed subspace of A. Consider
the short exact sequence
{0} → B
j
→ A
q
→ A/B → {0},
where j : B → A is the inclusion and q : A → A/B is the quotient map. If the
dual sequence
{0} → (A/B)∗ q∗
→ A∗ j∗
→ B∗ → {0},
is topologically exact, then we obtain the canonical isomorphisms B∗ ∼= A∗/B◦
and (A/B)∗ ∼= B◦; see [45, Remark 26.5]. (cid:3)
In S 2.8, we mentioned an important result concerning the spectrum of Banach
and Fr´echet algebras. In the following, some other results in this field have been
presented.
D 2.17. Let B be a commutative unital Banach algebra. Then, ∆(B) is a nonempty
compact space. Following [27, 3.2], in general the spectrum of a commutative
Fr´echet algebra A is not a compact space. But, it is a nonempty hemicompact
space. In fact, there exists a countable compact exhaustion (Kn)n of A such that
for every compact subset K ⊆ A there is n ∈ N so that K ⊆ Kn. (cid:3)
Before the next result, we recall that an ideal I of a Fr´echet algebra A is called
maximal if I 6= A and I is contained in no other proper ideal of A.
D 2.18. Every maximal ideal in a commutative Banach algebra B is closed.
Hence, there exists a one-to-one correspondence between the maximal ideals of
B and the elements of ∆(B). Opposite to this result maximal ideals of commu-
tative Fr´echet algebras are in general not closed. For example, C(R) has some
maximal ideal which is not closed; see [27, Example 3.2.11]. (cid:3)
D 2.19. Recall that a Hausdorff space is called a k-space if every subset inter-
secting each compact subset in a closed set is itself closed. Following [27, chapter
3], examples of k-spaces are locally compact and first countable spaces. Also, the
spectrum of a commutative Banach algebra is a k-space. But, there are Fr´echet
algebras whose spectrum is not a k-space; see [27, Example 7.1.2]. (cid:3)
Suppose that A is a commutative Fr´echet algebra. Let us recall that a point
ϕ ∈ ∆(A) is said to be a local peak point for A if there exist a neighborhood U
of ϕ in ∆(A) and an element a ∈ A such that a(ϕ) = 1 and
a(ψ) < 1
(ψ ∈ U \ {ϕ}).
Moreover, a is called a peak point for A if U can be taken to be ∆(A).
D 2.20. In the class of Banach algebras, by applying the Rossi's maximum princi-
ple, one can see that every locall peak point is a peak point; see [27, Theorem 1.4.9].
However, this result is not valid for Fr´echet algebras; see also [27, Example 9.2.2].
Therefore, the Rossi's maximum modulus theorem fails to be true for Fr´echet al-
gebras. (cid:3)
10
Z. ALIMOHAMMADI AND A. REJALI
D 2.21. Let B be a commutative unital Banach algebra and let B−1 denote the
set of all invertible elements of B. Due to the Arens-Royden theorem, for every
f ∈ C(∆(B))−1 there exist b ∈ B−1 and g ∈ C(∆(B)) such that f = b exp(g),
It is an open question,
whether the Arens-Royden theorem can be extended to Fr´echet algebras. However,
following [27], this theorem was proved for a special class of Fr´echet algebras. (cid:3)
where exp(g) = P∞
n=0 gn/n!; see [27, Theorem 1.4.8].
3. Some important generalizations in harmonic analysis
In this section, we collect some important generalizations to the Fr´echet al-
gebras which have been obtained in recent years. We mark these results by the
letter "G", and we use the symbol "(cid:3)" after each one of them.
Let B be a Banach algebra. We recall that a left Banach B-module F is called
flat if for every exact admissible sequence
of right Banach B-modules, the sequence
{0} → E1 → E2 → E3 → {0}
{0} → E1 ⊗BF → E2 ⊗BF → E3 ⊗BF → {0}
is exact. Also, B is amenable if for each Banach B-bimodule E every continuous
derivation from B to the dual bimodule E∗ is inner. This definition was introduced
by Johnson in [39]. Later, Helemskii and Sheinberg [34] showed that a Banach
algebra B is amenable if and only if B♯, the unitization of B, is a flat Banach
B-bimodule. In this case, B has a b.a.i. [59, Proposition 2.2.1]. In recent years, a
large number of papers concerning amenability of Banach algebras have appeared;
see for instance [5, 12, 18, 28, 29, 32, 44].
In [49], the arguments above have
generalized as follows:
G 3.1. Due to Pirkovskii [49], a Fr´echet algebra A is amenable (resp. biflat) if
A♯ (resp. A) is a flat Fr´echet A-bimodule. Also, we say that A has a locally
b.a.i. if for each zero neighborhood U ⊆ A there exists C > 0 such that for each
finite subset F ⊆ A and each ε > 0 there exists b ∈ CU with a − ab ∈ εU and
a − ba ∈ εU for all a ∈ F . Note that if A is normable, then the notions of "b.a.i."
and "locally b.a.i." are equivalent [49, Remark 6.4]. By applying [49, Lemma 9.4],
a Fr´echet algebra A with a locally b.a.i.
is amenable if and only if A is biflat.
Furthermore, A is amenable if and only if it is isomorphic to a reduced projective
limit of a sequence of amenable Banach algebras [49, Theorem 9.5]. Let us note
the following theorem proved by A. Pirkovskii [49, Theorem 9.6].
Theorem. Let A be a Fr´echet algebra. The following conditions are equivalent:
(i) A is amenable;
(ii) for each Banach A-bimodule E, every continuous derivation from A to
E∗ is inner;
(iii) for each Fr´echet A-bimodule E, every continuous derivation from A to
E∗ is inner.
Pirkovskii also introduced the concepts of "virtual diagonal", "bounded approxi-
mate diagonal", and "locally bounded approximate diagonal" in the class of Fr´echet
algebras. He then extended Johnson's theorem (see [38, Lemma 1.2] and [38, The-
orem 1.3]; see also [59, Theorem 2.2.4]). Indeed, for a Fr´echet algebra A, the
following statements are equivalent:
(i) A is amenable;
(ii) A has a locally bounded approximate diagonal.
If, in addition, A is quasinormable, then (i) and (ii) are equivalent to the following:
(iii) A has a bounded approximate diagonal;
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
11
(iv) A has a virtual diagonal.
See [49, Theorem 9.7]. (cid:3)
G 3.2. Ghahramani and Loy [26] introduced and studied approximate amenabil-
ity and contractibility of Banach algebras. These notions were investigated in
the class of Fr´echet algebras by Lawson and Read [43]. A Fr´echet algebra A is
called approximately contractible if given any A-bimodule E, and any continuous
derivation D : A → E, there is a net (xα) ⊆ E such that
D(a) = lim
α
(a · xα − xα · a)
(a ∈ A).
Every continuous derivation D with this property is said to be approximately inner.
Also, we recall that a Fr´echet algebra A is approximately amenable if given any
A-bimodule E, every continuous derivation D : A → E∗ is approximately inner.
The authors in [43] generalized the important results of Banach algebras in this
field. Following [26], all known approximately amenable Banach algebras have
b.a.i. It is nonetheless hoped that it would yield Banach algebras without b.a.i.
which had a form of amenability. Note that there are examples of Fr´echet algebras
which are approximately contractible, but which do not have a b.a.i. For a good
many Fr´echet algebras without b.a.i., it was obtained either that the algebra is
approximately amenable, or it is "obviously" not approximately amenable because
it has continuous point derivations; see [43, section 3]. Therefore, the situation
for Fr´echet algebras is quite close to what was hoped for Banach algebras. (cid:3)
Weakly amenable Banach algebras were first introduced by Bade, Curtis and
Dales in [16]. The general definition is due to Johnson [39]. Later, Dales, Ghahra-
mani and Grønbaek [24] introduced and studied the notion of n-weak amenabil-
ity of Banach algebras.
Indeed, a Banach algebra B is n-weakly amenable if
H 1(B, B(n)) = {0}, where B(n) is the n-th dual space of B. In particular, B is
weakly amenable if it is 1-weakly amenable. Recently, several papers dealing with
this subject have appeared; see for example [28, 53, 60].
G 3.3. Let A be a Fr´echet algebra. The notion of weak amenability of A was stud-
ied by Rejali et al. [11]. In fact, the authors showed that some results in the field
of weak amenability of Banach algebras can be generalized for the Fr´echet alge-
bras. For example, if A is weakly amenable, then A is essential [11, Theorem 2.3].
Moreover, if I is a quasinormable closed ideal in the Fr´echet algebra A such
that A/I and I are weakly amenable, then A is weakly amenable, as well; see
[11, Theorem 2.6]. (cid:3)
G 3.4. Let B be a Banach algebra and I be a closed ideal of B. The notions of I-
weak amenability and ideal amenability of B was introduced and studied by Gorgi
and Yazdanpanah in [30].
Indeed, B is I-weakly amenable if every continuous
derivation D : B → I ∗ is inner. Moreover, B is called ideally amenable if B is
I-weakly amenable for every closed ideal I of B. Clearly, every amenable Banach
algebra is ideally amenable, and every ideally amenable Banach algebra is weakly
amenable. However, the concepts of I-weak amenability and ideal amenability are
different from amenability and weak amenability of Banach algebras (ibid.).
In [55], Ranjbari and Rejali generalized these concepts and their results to the
Fr´echet algebras. Later, the concept of vector-valued Lipschitz algebras in the
class of Fr´echet algebras was studied by the same authors in [54]. Especially, they
Proved that Lemma 2.1 of [18] holds for Fr´echet algebras under a certain condition.
Then, they studied ideal amenability of vector-valued Fr´echet Lipschitz algebras;
see [54, section 4] for details. (cid:3)
12
Z. ALIMOHAMMADI AND A. REJALI
Let A be a Banach algebra with a b.a.i. which is contained as closed ideal in a
Banach algebra B. Consider the pseudo-unital Banach A-bimodule E, that is
E = A · E · A = {a · x · b : a, b ∈ A, x ∈ E}.
By using [59, Proposition 2.1.6], E is a Banach B-bimodule in a canonical fachion.
Moreover, for each derivation D : A → E∗, there is a unique extension derivation
eD : B → E∗ of D such that eD is continuous with respect to the strict topology
on B and the weak∗-topology on E∗. In [2], we studied the strict topology on a
Fr´echet algebra and then we obtained a similar result as above.
G 3.5. Let A be a closed ideal in a Fr´echet algebra (B, pℓ)ℓ∈N, satisfying
(cid:8)b ∈ B : b · A = A · b = {0}(cid:9) = {0},
and suppose that B has a b.a.i. with elements in A. Let (E, qn)n∈N be an arbi-
trary pseudo-unital Fr´echet A-bimodule. Then, for every continuous derivation
D : A → E∗, with respect to the topology induced by (pℓ)ℓ∈N on A and the strong
topology on E∗, there exists a unique derivation eD : B → E∗ such that eD is con-
tinuous with respect to the strict topology on B and the weak∗-topology on E∗; see
[2, Theorem 3.2]. (cid:3)
G 3.6. Following [9] and also [10], a Fr´echet algebra (A, pℓ)ℓ∈N is called a Segal
Fr´echet algebra in the Fr´echet algebra (B, qn)n∈N if the following conditions hold:
(1) A is a dense ideal in (B, qn)n;
(2) the identity map (A, pℓ)ℓ → (B, qn)n is continuous;
(3) the map (B, qn)n × (A, pℓ)ℓ → (A, pℓ)ℓ with (a, b) 7→ ab, for all a, b ∈ A,
is jointly continuous.
This definition is in fact a generalization of the definition abstract Segal algebras,
in the Banach case; see [20]. Also in [21, Definition 14], this definition has been
introduced in the field of topological algebras, as follows:
A topological algebra (A, τA) is called a (B, τB)-Segal algebra in the topological
algebra (B, τB) if the following assertions hold:
(1)′ A is a dense left ideal in (B, τB);
(2)′ the identity map from (A, τA) into (A, τB) is continuous;
(3)′ the multiplication
(B, τB) × (A, τA) → (A, τA), ((a, b) 7→ ab),
is continuous.
In [9], Rejali et al. investigated Theorem 1.2 of [20] for Segal Fr´echet algebras
and obtained an analogous result. Indeed, if (A, pℓ)ℓ∈N is a proper Segal Fr´echet
algebra in (B, qn)n∈N which contains an a.i. (ei)i∈I , then (ei)i∈I cannot be bounded
in (A, pℓ)ℓ∈N [9, Proposition 3.5]. Furthermore, the authors characterized closed
ideals of Segal Fr´echet algebras, and showed that the ideal theorem is also valid
for Fr´echet algebras. In other words, every closed ideal of any Segal Fr´echet al-
gebra (A, pℓ)ℓ∈N in (B, qn)n∈N is the intersection of a closed ideal of B with A [9,
Theorem 3.8]. (cid:3)
Suppose that A is a commutative Fr´echet algebra and M (A) is the multiplier
algebra with respect to the strict topology. As mentioned in section 0, A is a
M (A)-Segal algebra; see [2, section 4] for details. In [14], we studied the notion
of multiplier algebra of Fr´echet algebras.
Let B be a Banach algebra and ϕ ∈ ∆(B). Kaniuth, Lau and Pym [40] in-
troduced and studied the concept of (right) ϕ-amenability for Banach algebras as
a generalization of left amenability of Lau algebras. We recall that, B is called
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
13
(right) ϕ-amenable if there exists a continuous linear functional m on B∗ such that
m(ϕ) = 1 and m(f ·a) = ϕ(a)m(f ) for all a ∈ B and f ∈ B∗. Furthermore, the no-
tion of right character amenability of Banach algebras was introduced and studied
by Monfared [47]. Indeed, a Banach algebra B is called right character amenable
if for each ϕ ∈ ∆(B) ∪ {0} and all Banach B-bimodules E with the left module
action a · x = ϕ(a)x, each continuous derivation D : B → E∗ is inner. Alagh-
mandan, Nasr-Isfahani, and Nemati [12] characterized the character amenability
of abstract Segal algebras. Specifically, they showed that ϕ-amenability of B is
equivalent to ϕA-amenability of A, where A is an abstract Segal algebra in B.
Right ϕ-amenability and right character amenability for Fr´echet algebras were
introduced and studied in [7]. In the following, some related results about Segal
Fr´echet algebras are provided.
G 3.7. Let (A, pℓ)ℓ∈N be a Segal Fr´echet algebra in a Fr´echet algebra (B, qn)n∈N.
Precisely, similar to Lemma 2.2 of [12],
∆(A) = {ϕA : ϕ ∈ ∆(B)};
see [7, Lemma 6.7]. Also, if ϕ ∈ ∆(B), then B is right ϕ-amenable if and only if
A is right ϕA-amenable [7, Theorem 6.8]. The following is now immediate from
results above.
Corollary. Let (B, qn)n∈N be a Fr´echet algebra and let (A, pℓ)ℓ∈N be a Segal
Fr´echet algebra in B. Suppose that B is right character amenable. Then, A is
right ϕ-amenable for each ϕ ∈ ∆(A); see [7, Corollary 6.9]. (cid:3)
In addition to the above, we joint with Abtahi [1] generalized the notion of
Connes amenability of Banach algebras. We then proved some of the previous
available results about this notion for Fr´echet algebras. Rahnama and Rejali
[51, 52] also introduced and studied the concepts of amenability modulo an ideal
of Fr´echet algebras and locally bounded approximate diagonal modulo an ideal
of Fr´echet algebras. Finally, the notion of BSE algebras in the class of Fr´echet
algebras was investigated by Rejali et al. in [3, 4, 15].
Before continuing, let us note the following arguments. Let B be a Banach
algebra. We know that on the bidual space B∗∗ of B, there are two multiplications,
called the first and second Arens products, which make B∗∗ into a Banach algebra.
Moreover, B is called Arens regular if these two products coincide on B∗∗. There
are many papers concerning Arens regularity. Also, this notion was investigated
by Rejali et al. in several papers; see for example [6, 17, 41, 56, 57, 58]. Note that
if B∗∗ is Arens regular, then so is B. However, Pym [50] gives an example of an
Arens regular Banach algebra B such that B∗∗ is not Arens regular. Also, look at
the following example.
Example. Consider the Banach algebras An = ℓ1(Z, w 1
(1 + t)α (t ∈ Z) for each α > 0. Let
) (n ∈ N), where wα(t) =
n
A = c0-Yn∈N
B = ℓ∞-Yn∈N
An =(cid:8)(an) ∈ Yn∈N
An =(cid:8)(an) ∈ Yn∈N
An : kankn → 0(cid:9) and
kankn < ∞(cid:9).
An : sup
n∈N
The following statements hold.
(i) by [25, page 35], A is Arens regular.
(ii) by [25, Example 9.2], B is not Arens regular.
(iii) B is a closed sualgebra of A∗∗ and consequently A∗∗ is not Arens regular.
Let G be a locally compact (topological) group with a weight function w on it.
Rejali and Vishki [58] showed that the following statements are equivalent:
14
Z. ALIMOHAMMADI AND A. REJALI
(i) L1(G, w)∗∗ is Arens regular;
(ii) L1(G, w) is Arens regular;
(iii) L1(G) is Arens regular or Ω : G × G → (0, 1], defined by
Ω(x, y) =
w(xy)
w(x)w(y)
(x, y ∈ G),
is 0-cluster, that is limm limn Ω(xm, yn) = 0 = limn limm Ω(xm, yn), for
sequences (xm) and (yn) in G with distinct elements.
Moreover, Baker and Rejali proved that for a semigroup S with a weight function
w on it, ℓ1(S, w) is Arens regular, whenever ℓ1(S) is [17]. Now, the following
question is natural.
Question. Let S be a semigroup and w be a weight function on S. Are the
following statements equivalent?
(i) ℓ1(S, w)∗∗ is Arens regular.
(ii) ℓ1(S, w) is Arens regular.
(iii) ℓ1(S) is Arens regular or Ω is 0-cluster.
Later, Rejali et al. showed that if the set of all idempotents of an inverse semigroup
S is finite, then ℓ1(S) is Arens regular if and only if S is finite; see [6, Theorem 3.8].
Also in this case, if S admits a bounded below weight for which Ω is 0-cluster,
then S is countable; see [41, Theorem 3.4]. By arguments above, another question
arises.
Question. Let S be an inverse semigroup with finitely many idempotents and w
be a weight function on S. Are the following statements equivalent?
(i) ℓ1(S, w)∗∗ is Arens regular.
(ii) ℓ1(S, w) is Arens regular.
(iii) S is finite or Ω is 0-cluster.
Remark. In [6, Example 3.9], the authors showed that there exists an inverse
semigroup S with infinitely many idempotents such that ℓ1(S) is Arens regular.
Arens regularity of Fr´echet algebras was defined by Zivari-Kazempour [69].
Indeed, a Fr´echet algebra A is called Arens regular if the first and second Arens
products coincide on A∗∗. Later, we [13] obtained some results in this field.
G 3.8. In 1966, Gulick [31] studied Arens products on the bidual space of locally
multiplicatively-convex topological (lmc for short) algebras with a hypo-continuous
multiplication. We recall that the product A × A → A, on a lmc algebra A, is
left (resp. right) hypo-continuous if for each zero neighborhood U and for each
bounded set B of A there exists a zero neighborhood V such that V B ⊆ U (resp.
BV ⊆ U ). The multiplication in A is called hypo-continuous if it is both left and
right hypo-continuous.
Now, let A be a Fr´echet algebra. Clearly, A belongs to the class of lmc algebras.
Moreover, by applying [61, III.5.1], the multiplication in A is hypo-continuous.
Therefore, due to Gulick [31] and [45, Proposition 25.9], A∗∗ is always a Fr´echet
algebra under the strong topology and Arens products on A∗∗.
Furthermore, Gulick generalized a theorem of Civin and Yood [22] to the lmc
algebras, especially Fr´echet algebras. He introduced a bicommutative algebra, that
is an lmc algebra whose bidual is commutative. Then, he showed that subalgebras
of bicommutative algebras are also bicommutative [31, Theorem 5.3]. (cid:3)
G 3.9. Following [22], a Banach algebra B has a weak right identity if there exists
a net (eα)α∈Λ in B and an M > 0 such that for every α ∈ Λ, keαk < M and
lim
α
f (aeα − a) = 0
(a ∈ B, f ∈ B∗).
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
15
Civin and Yood proved that B has a weak right identity if and only if B∗∗ has
a right unit with respect to the first Arens product [22, Lemma 3.8]. There is
a generalization of this theorem in the class of lmc algebras especially Fr´echet
algebras. Zivari-Kazempour proved that an lmc algebra A has a right b.a.i. (resp.
left b.a.i) if and only if A∗∗ has a right (resp. left) unit with respect to the first
(resp. second) Arens product [69, Theorem 3.2]. (cid:3)
Acknowledgment. The authors would like to thank the Banach algebra cen-
ter of Excellence for Mathematics, University of Isfahan.
References
[1] F. Abtahi, Z. Alimohammadi, and A. Rejali, Connes amenability of Fr´echet algebras, sub-
mitted.
[2] F. Abtahi, Z. Alimohammadi, and A. Rejali, On the strict, uniform and compact-open topolo-
gies on an algebra, Period. Math. Hungar. 77 (2018), no. 2, 209 -- 223.
[3] F. Abtahi, M. Amiri, and A. Rejali, The Bochner-Schoenberg-Eberline property for commu-
tative Fr´echet algebras, submitted.
[4] F. Abtahi, M. Amiri, and A. Rejali, The Bochner-Schoenberg-Eberline property for Fr´echet
C ∗-algebras and uniform Fr´echet algebras, submitted.
[5] F. Abtahi, M. Azizi, and A. Rejali, Character amenability of the intersection of Lipschitz
algebras, Canad. Math. Bull. 60 (2017), no. 4, 673 -- 689.
[6] F. Abtahi, B. Khodsiani, and A. Rejali, Arens regularity of inverse semigroup algebras, Bull.
Iranian Math. Soc., 40 (2014), no. 6, 1527 -- 1538.
[7] F. Abtahi, S. Rahnama, and A. Rejali, ϕ-amenability and character amenability of Fr´echet
algebras, Forum. Math. (2018), DOI 10.1515/forum-2016-0116.
[8] F. Abtahi and S. Rahnama, ϕ-contractibility and character contractibility of Fr´echet algebras,
Ann. Funct. Anal. 8 (2017), no. 1, 75 -- 89.
[9] F. Abtahi, S. Rahnama, and A. Rejali, Segal Fr´echet algebras, Anal. Math. (2017), DOI
10.1007/s10476-017-0601-y.
[10] F. Abtahi, S. Rahnama, and A. Rejali, Semisimple Segal Fr´echet algebras, Period. Math.
Hungar. 71 (2015), no. 2, 146 -- 154.
[11] F. Abtahi, S. Rahnama, and A. Rejali, Weak amenability of Fr´echet algebras, U.P.B. Sci.
Bull., Series A, 77 (2015), no. 4, 93 -- 104.
[12] M. Alaghmandan, R. Nasr-Isfahani, and M. Nemati, Character amenability and contractibil-
ity of abstract Segal algebras, Bull. Aust. Math. Soc. 82 (2010), no. 2, 274 -- 281.
[13] Z. Alimohammadi, and A. Rejali, On the Arens regularity of Fr´echet algebras and their
biduals, submitted.
[14] Z. Alimohammadi, and A. Rejali, On the multiplier algebra of Fr´echet algebras, submitted.
[15] M. Amiri, A. Rejali, The multiplier algebra and B.S.E. property of the direct sum of Fr´echet
algebras, submitted.
[16] W. G. Bade, P. C. Curtis, and H. G. Dales, Amenability and weak amenability for Bearling
and Lipschitz algebras, Proc. London Math. Soc. 55 (1987), no. 3, 359-377.
[17] J. W. Baker and A. Rejali, On the Arens regularity of weighted convolution algebras, J.
London Math. Soc. (2), 40 (1989), no. 3, 535 -- 546.
[18] E. Biyabani and A. Rejali, Approximate and Character amenability of vector-valued Lips-
chitz algebras, Bull. Korean Math. Soc. 55 (2018), no. 4, 1109 -- 1124.
[19] N. Bourbaki, Topological vector spaces, Elements of mathematics, Springer (1987).
[20] J. T. Burnham, Closed ideals in subalgebras of Banach algebras, Proc. Amer. Math. Soc.
32, (1972), no.2, 551 -- 555.
[21] J. T. Burnham, Segal algebras and dense ideals in Banach algebras, Functional analysis
and its applications, (Internat. Conf., Eleventh Anniversary of Matscience, Madras; 1973
dedicated to Alladi Ramakrishnan), Springer, Berlin, 1974, 33 -- 58 Lecture Notes in Math.,
399.
[22] P. Civin and B. Yood, the second conjugate space of a Banach algebra as an algebra, Pacific
J. Math., 11 (1961), 820 -- 847.
[23] J. B. Conway, A course in functional analysis, Graduate Texts in Math., Vol. 96, Springer-
Verlag, 1985.
[24] H. G. Dales, F. Ghahramani, and N. Grønbaek, Derivations into iterated duals of Banach
algebras, Studia Math. 128 (1998), no. 1, 19 -- 54.
[25] H. G. Dales and A. T.-M. Lau, The second duals of Beurling algebras, Mem. Amer. Math.
Soc., 177 (2005).
16
Z. ALIMOHAMMADI AND A. REJALI
[26] F. Ghahramani and R. J. Loy, Generalized notions of amenability, J. Funct. Anal. 208
(2004), 229 -- 260.
[27] H. Goldmann, Uniform Fr´echet algebras, North-Holland Mathematics Studies, 162. North-
Holand, Amesterdam-New York, 1990.
[28] M. E. Gordji, F. Habibian, and A. Rejali, Module extension of dual Banach algebras, Bull.
Korean Math. Soc. 47 (2010), no. 4, 663 -- 673.
[29] M. E. Gordji, A. Jabbari, and G. Hui Kim, Some characterizations of character amenable
Banach algebras, Bull. Korean Math. Soc. 52 (2015), no. 3, 761 -- 769.
[30] M. E. Gordji and T. Yazdanpanah, Derivations into duals of ideals of Banach algebras,
Proc. Indian Acad. Sci. 114 (2004), no. 4, 399 -- 408.
[31] S. L. Gulick, The bidual of a locally multiplicatively-convex algebra, Pacific J. Math., 17/1
(1966), 71 -- 96.
[32] F. Habibian, H. Samea, and A. Rejali, Ideal amenability and approximate ideal amenability
of matrix Banach algebras, Proc. Ro Acad. Series A, 3 (2011), 173 -- 178.
[33] D. Hajdukovi´c, The functionals of the kind of Banach limits, publications de L'Institut
Mathematique, T. 19(33), 1975.
[34] A. Ya. Helemskii and M. V. Sheinberg, Amenable Banach algebras (Russian), Funktsional.
Anal. i Prilozhen. 13 (1979), no. 1, 42 -- 48; English transl.: Functional Anal. Appl. 13 (1979),
no. 1, 32 -- 37.
[35] A. Ya. Helemskii, On a method for calculating and estimating the global homological di-
mension of Banach algebras, Mat. Sb., 87 (129) (1972), 122 -- 135.
[36] A. Ya. Helemskii, The homology of Banach and topological algebras (Moscow University
Press, English transl: Kluwer Academic Publishers, Dordrecht 1989).
[37] T. G. Honary, Automatic continuity of homomorphisms between Banach algebras and
Fr´echet algebras, Bull. Iranian Math. Soc. 32 (2006), no. 2, 1 -- 11.
[38] B. E. Johnson, Approximate diagonals and cohomology of certain annihilator Banach al-
gebras, Amer. J. Math. 94 (1972), 685 -- 698.
[39] B. E. Johnson, Cohomology in Banach algebras, Mem. Amer. Math. Soc. 127 (1972).
[40] E. Kaniuth, A. T. Lau, and J. Pym, on ϕ-amenability of Banach algebras, Math. Proc.
Combridge Philos. Soc. 144 (2008), no.1, 85 -- 96.
[41] B. Khodsiani, A. Rejali, and H. R. E. Vishki, Arens regularity of certain weighted semigroup
algebras and countability, A. Semigroup Forum 92 (2016), 304 -- 310.
[42] G. Kothe, topological vector space. (I), Spronger-Verlag, New York, 1969.
[43] P. Lawson and C. J. Read, Approximate amenability of Fr´echet algebras, Math. Proc.
Cambridge Philos. Soc., 145 (2008), 403 -- 418.
[44] A. R. Medghalchi and T. Yazdanpanah, n-weak amenability and strong double limit prop-
erty, Bull. Korean Math. Soc. 42 (2005), no. 2, 359 -- 367.
[45] R. Meise and D. Vogt, Introduction to functional analysis, (Oxford Science Publications),
1997.
[46] E. A. Michael, Locally multiplicatively convex topological algebras, Memoirs Amer. Math.
Soc. 11 (1952).
[47] M. S. Monfared, Character amenability of Banach algebras, Math. Proc. Combridge Philos.
Soc. 144 (2008), no. 3, 697 -- 706.
[48] A. Yu. Pirkovskii, Biprojective topological algebras of homological bidimension 1, J. Math.
Sci. (New York) 111 (2002), no. 2, 3476 -- 3495.
[49] A. Yu. Pirkovskii, Flat cyclic Fr´echet modules, Amenable Fr´echet algebras, and approximate
identities, Homology, Homotopy Appl., 11/1 (2009), 81 -- 114.
[50] J. S. Pym, Remarks on the second duals of Banach algebras, J. Nigerian Math. Soc., 2
(1983), 31 -- 33.
[51] S. Rahnama and A. Rejali, Amenability modulo an ideal of Fr´echet algebras, submitted.
[52] S. Rahnama and A. Rejali, Locally bounded approximate diagonal modulo an ideal of Fr´echet
algebras, submitted.
[53] M. Ramezanpour, Weak amenability of the Lau product of Banach algebras defined by a
Banach algebra morphism, Bull. Korean Math. Soc. 54 (2017), no. 6, 1991 -- 1999.
[54] A. Ranjbari and A. Rejali, Fr´echet α-Lipschitz vector-valued operator algebra, U.P.B. Sci.
Bull., Series A, to appear.
[55] A. Ranjbari and A. Rejali, Ideal amenability of Fr´echet algebras, U.P.B. Sci. Bull., Series
A, 79 (2017), no. 4, 51 -- 60.
[56] A. Rejali, The analogue of weighted qroup algebra for semitopological semigroups, J. Sci,
I.R. Iran, 6, no. 2 (1995), 113 -- 120.
[57] A. Rejali, The Arens regularity of weighted semigroup algebras, Sci. Math. Japon, 60, no.
1 (2004), 129 -- 137.
FR ´ECHET ALGEBRAS IN ABSTRACT HARMONIC ANALYSIS
17
[58] A. Rejali, H. R. E. Vishki, Regularity and amenability of the second dual of weighted group
algebras, Proyecciones 26 (2007), 259 -- 267.
[59] V. Runde, Lectures on Amenability, volume 1774 of Lecture Notes in Mathematics. Springer,
Berlin, 2002.
[60] H. Samea, Weak amenability of convolution Banach algebras on compact hypergroups, Bull.
Korean Math. Soc. 47 (2010), no. 2, 307 -- 317.
[61] H. H. Schaefer, Topological vector spaces, Third printing corrected Graduate text in math-
ematics, Vol.3, Springer-Verlag, New York, 1971.
[62] Yu. V. Selivanov, Biprojective Banach algebras, Izv. Akad. Nauk SSSR, Ser. Mat., 43 (1979),
1159 -- 1174.
[63] Yu. V. Selivanov, Biprojective topological algebras (talk in Sheffield, 12 November 1996),
Unpublished manuscript, 1996.
[64] D. R. Sherbert, The structure of ideals and point derivations in Banach algebras of Lipschitz
functions, Trans. Amer. Math. Soc., 111 (1964), 240 -- 272.
[65] H. A. Smith, Tensor products of locally convex algebras, Proc. Am. Math. Soc. 17 (1966),
124 -- 132.
[66] J. L. Taylor, Homology and cohomology for topological algebras, Adv. Math., 9 (1972),
137 -- 182.
[67] J. Voigt, Factorization in Fr´echet algebras, J. London Math. Soc., 29/2 (1984), 147 -- 152.
[68] J. Voigt, Factorization in some Fr´echet algebras of differentiable functions, Studia Math.
77 (1984), 333 -- 348.
[69] A. Zivari-Kazempour, On the Arens product and approximate identity in locally convex
algebras, Filomat, 30 (2016), no. 6, 1493 -- 1496.
Z. Alimohammadi
Department of Mathematics, University of Isfahan, Isfahan, Iran
[email protected]; [email protected]
A. Rejali
Department of Mathematics, University of Isfahan, Isfahan, Iran
[email protected]
|
1608.02047 | 4 | 1608 | 2016-12-21T11:17:00 | Infinitesimal generators of invertible evolution families | [
"math.FA"
] | A logarithm representation of operators is introduced as well as a concept of pre-infinitesimal generator. Generators of invertible evolution families are represented by the logarithm representation, and a set of operators represented by the logarithm is shown to be associated with analytic semigroups. Consequently generally-unbounded infinitesimal generators of invertible evolution families are characterized by a convergent power series representation. | math.FA | math | INFINITESIMAL GENERATORS
OF INVERTIBLE EVOLUTION FAMILIES
YORITAKA IWATA
Abstract. A logarithm representation of operators is introduced as well as a
concept of pre-infinitesimal generator. Generators of invertible evolution fami-
lies are represented by the logarithm representation, and a set of operators rep-
resented by the logarithm is shown to be associated with analytic semigroups.
Consequently generally-unbounded infinitesimal generators of invertible evo-
lution families are characterized by a convergent power series representation.
.
A
F
h
t
a
m
[
4
v
7
4
0
2
0
.
8
0
6
1
:
v
i
X
r
a
1. Introduction
The logarithm of an injective sectorial operator was introduced by Nollau [12] in
1969. After a long time, the logarithm of sectorial operators were studied again from
1990's [2, 13, 14], and its utility was established with respect to the definition of the
logarithms of operators [4, 11] (for a review of sectorial operators, see Hasse [5]).
While the sectorial operator has been a generic framework to define the logarithm
of operators, the sectorial property is not generally satisfied by the evolution family
of operators, where the evolution family correspond to the exponentials of operators
in abstract Banach space framework (for the definition of evolution family in this
paper, see Sec. 2.1).
In this paper we characterize infinitesimal generators of invertible evolution fam-
ilies that has not been settled so far. First of all, by introducing a kind of similarity
transform, a logarithm representation is obtained for such generators. The loga-
rithm representation is utilized to show a convergent power series representation of
invertible evolution families generated by certain unbounded operators, although
the validity of such a representation is not established for any evolution families
generated by unbounded operators. In this context, the concept of pre-infinitesimal
generator is introduced.
2. Mathematical settings
2.1. Evolution family on Banach spaces. Let X and B(X) be a Banach space
with a norm k · k and a space of bounded linear operators on X. The same notation
is used for the norm equipped with B(X), if there is no ambiguity.
For a positive and finite T , let elements of evolution family {U (t, s)}−T ≤t,s≤T
be mappings: (t, s) → U (t, s) satisfying the strong continuity for −T ≤ t, s ≤ T
(for reviews or textbooks, see [1, 6, 10, 15, 17]). The semigroup properties:
(1)
U (t, r) U (r, s) = U (t, s),
2000 Mathematics Subject Classification. 47D03; 35A20; 34K30.
Key words and phrases. invertible evolution family, operator theory, maximal regularity.
The author is grateful to Prof. Emeritus Hiroki Tanabe for valuable comments.
1
2
and
(2)
YORITAKA IWATA
U (s, s) = I,
are assumed to be satisfied, where I denotes the identity operator of X. Both
U (t, s) and U (s, t) are assumed to be well-defined to satisfy
(3)
U (s, t) U (t, s) = U (s, s) = I,
where U (s, t) corresponds to the inverse operator of U (t, s). Since U (t, s) U (s, t) =
U (t, t) = I is also true, the commutation between U (t, s) and U (s, t) follows. Oper-
ator U (t, s) is a generalization of exponential function; indeed the properties shown
in Eqs. (1)-(3) are satisfied by taking U (t, s) as et−s. Evolution family is an ab-
stract concept of exponential function valid for both finite and infinite dimensional
Banach spaces.
Let Y be a dense Banach subspace of X, and the topology of Y be stronger than
that of X. The space Y is assumed to be U (t, s)-invariant. Following the definition
of C0-(semi)group (cf. the assumption H2 in Sec. 5.3 of Pazy [15] or corresponding
discussion in Kato [7, 8]), U (t, s) is assumed to satisfy the boundedness [6, 15];
there exist real numbers M and β such that
(4)
kU (t, s)kB(X) ≤ M eβt,
kU (t, s)kB(Y ) ≤ M eβt.
Inequalities (4) are practically reduced to
kU (t, s)kB(X) ≤ M eβT ,
kU (t, s)kB(Y ) ≤ M eβT ,
when t-interval is restricted to be finite [−T, T ].
2.2. Pre-infinitesimal generator. The counterpart of the logarithm in the ab-
stract framework is introduced. There are two concepts associated with the log-
arithm of operators; one is the pre-infinitesimal generator and the other is t-
differential of U (t, s). For −T ≤ t, s ≤ T , the weak limit
(5)
wlim
h→0
h−1(U (t + h, s) − U (t, s)) u = wlim
h→0
h−1(U (t + h, t) − I) U (t, s) u,
is assumed to exist for u, which is an element of a dense subspace Y of X. A linear
operator A(t) : Y → X is defined by
(6)
A(t)u := wlim
h→0
h−1(U (t + h, t) − I)u
for u ∈ Y and −T ≤ t, s ≤ T , and then let t-differential of U (t, s) in a weak sense
be
(7)
∂tU (t, s) u = A(t)U (t, s) u.
Equation (7) is regarded as a differential equation satisfied by U (t, s)u that implies
a relation between A(t) and the logarithm:
A(t) = ∂tU (t, s) U (s, t).
Let us call A(t) defined by Eq. (6) for a whole family {U (t, s)}−T ≤t,s≤T the pre-
infinitesimal generator. Note that pre-infinitesimal generators are not necessarily
infinitesimal generators; e.g., in t-independent cases, A defined by Eq. (6) is not
necessarily a densely-defined and closed linear operator, while A must be a densely-
defined and closed linear operator with its resolvent set included in {λ ∈ C : Reλ >
β}. On the other hand, infinitesimal generators are necessarily pre-infinitesimal
generators.
In the following a set of pre-infinitesimal generators is denoted by
G(X).
INFINITESIMAL GENERATORS ...
3
2.3. A principal branch of logarithm. The logarithm is defined by the Dunford
integral in this paper. Two difficulties of dealing with logarithm are its singularity
at the origin and its multi-valued property. By introducing a constant κ ∈ C,
the singularity can be handled. Let arg be a function of complex number, which
gives the angle between the positive real axis to the line including the point to the
origin. For the multi-valued property, a principle branch (denoted by "Log") of the
logarithm (denoted by "log") is chosen for any complex number z ∈ C, a branch of
logarithm is defined by
Logz = log z + i arg Z,
where Z is a complex number chosen to satisfy Z = z, −π < arg Z ≤ π, and
arg Z = arg z + 2nπ for a certain integer n.
Lemma 2.1. Let t and s satisfy 0 ≤ t, s ≤ T . For a given U (t, s) defined in Sec. 2,
its logarithm is well defined; there exists a certain complex number κ satisfying
(8)
Log(U (t, s) + κI) = 1
where an integral path Γ, which excludes the origin, is a circle in the resolvent set
of U (t, s) + κI. Here Γ is independent of t and s. Log(U (t, s) + κI) is bounded on
X.
2πiRΓ Logλ (λ − U (t, s) − κ)−1dλ,
Proof. The logarithm Log holds the singularity at the origin, so that it is necessary
to show a possibility of taking a simple closed curve (integral path) excluding the
origin in order to define the logarithm by means of the Dunford-Riesz integral. It
is not generally possible to take such a path in case of κ = 0.
First, U (t, s) is assumed to be bounded for 0 ≤ t, s ≤ T (Eq. (4)), and the
spectral set of U (t, s) is a bounded set in C. Second, for κ satisfying
(9)
κ > M eβT ,
the spectral set of U (t, s) + κI is separated with the origin. Consequently it is
possible to take an integral path Γ including the spectral set of U (t, s) + κI and
excluding the origin. Equation (8) follows from the Dunford-Riesz integral [3].
Furthermore, by adjusting the amplitude of κ, an appropriate integral path always
exists independent of t and s.
Log(U (t, s) + κI) is bounded on X, since Γ is included in the resolvent set of
(cid:3)
(U (t, s) + κI).
According to this lemma, by introducing nonzero κ, the logarithm of U (t, s) + κI
is well-defined without assuming the sectorial property to U (t, s). On the other
hand Eq. (8) is valid with κ = 0 only for limited cases.
3. Main results
3.1. Logarithm representation of pre-infinitesimal generator.
Theorem 3.1. Let t and s satisfy −T ≤ t, s ≤ T , and Y be a dense subspace
of X. For U (t, s) defined in Sec. 2, let A(t) ∈ G(X) and ∂tU (t, s) be determined
by Eqs. (6) and (7) respectively. If A(t) and U (t, s) commute, an evolution family
{A(t)}−T ≤t≤T is represented by means of the logarithm function; there exists a
certain complex number κ 6= 0 such that
(10)
A(t) u = (I + κU (s, t)) ∂tLog (U (t, s) + κI) u,
4
YORITAKA IWATA
where u is an element in Y . Note that U (t, s) defined in Sec. 2 is assumed to be
invertible.
Proof. For U (t, s) defined in Sec. 2, operators Log (U (t, s) + κI) and Log (U (t +
h, s) + κI) are well defined for a certain κ (Lemma 2.1). The t-differential in a weak
sense is formally written by
1
h {Log (U (t + h, s) + κI) − Log (U (t, s) + κI)}
(11)
{(λ − U (t + h, s) − κ)−1 − (λ − U (t, s) − κ)−1}dλ
wlim
h→0
= wlim
h→0
= wlim
h→0
1
1
h
2πiRΓ Logλ
2πiRΓ Logλ
1
{(λ − U (t + h, s) − κ)−1 U(t+h,s)−U(t,s)
(λ − U (t, s) − κ)−1}dλ
h
where Γ, which is possible to be taken independent of t, s and h for a sufficiently
large certain κ, denotes a circle in the resolvent set of both U (t, s) + κI and U (t +
h, s) + κI. A part of the integrand of Eq. (11) is estimated as
k{(λ − U (t + h, s) − κ)−1 U(t+h,s)−U(t,s)
h
(λ − U (t, s) − κ)−1}vkX
(12)
≤ k{(λ − U (t + h, s) − κ)−1kB(X)
k U(t+h,s)−U(t,s)
(λ − U (t, s) − κ)−1}vkX ,
h
for v ∈ X. There are two steps to prove the validity of Eq. (11). In the first step,
the former part of the right hand side of Eq. (12) satisfies
k{(λ − U (t + h, s) − κ)−1kB(X) < ∞,
since λ is taken from the resolvent set of U (t + h, s) − κI. In the same way the
operator (λ − U (t, s) − κ)−1 is bounded on X and Y . Then the continuity of the
mapping t → (λ − U (t, s) − κ)−1 as for the strong topology follows:
k(λ − U (t + h, s) − κ)−1 − (λ − U (t, s) − κI)−1kB(X)
≤ k(λ − U (t + h, s) − κ)−1kB(X)k(U (t + h, s) − U (t, s))(λ − U (t, s) − κ)−1kB(X).
In the second step, the latter part of the right hand side of Eq. (12) is estimated as
(13)
1
h
U(t+h,s)−U(t,s)
(λ − U (t, s) − κ)−1u(cid:13)(cid:13)(cid:13)X
(cid:13)(cid:13)(cid:13)
t A(τ )U (τ, s)(λ − U (t, s) − κ)−1u dτ(cid:13)(cid:13)(cid:13)X
=(cid:13)(cid:13)(cid:13)
hR t+h
hR t+h
≤ 1
t
kA(τ )U (τ, s)kB(Y,X)k(λ − U (t, s) − κ)−1kB(Y )kukY dτ
for u ∈ Y . Because kA(τ )U (τ, s)kB(Y,X) < ∞ is true by assumption, the right hand
side of Eq. (13) is finite. Equation (13) shows the uniform boundedness with respect
to h, then the uniform convergence (h → 0) of Eq. (11) follows. Consequently the
weak limit process h → 0 for the integrand of Eq. (11) is justified, as well as the
commutation between the limit and the integral.
According to Eq. (11), interchange of the limit with the integral leads to
∂tLog(U (t, s) + κI) u = 1
2πiRΓ dλ
(cid:20)(Logλ)(λ − U (t, s) − κ)−1 wlim
h→0 (cid:16) U(t+h,s)−U(t,s)
h
(cid:17) (λ − U (t, s) − κ)−1(cid:21) u
INFINITESIMAL GENERATORS ...
5
for u ∈ Y . Because we are also allowed to interchange A(t) with U (t, s),
∂tLog(U (t, s) + κI) u
= 1
= 1
(14)
2πiRΓ(Logλ)(λ − U (t, s) − κ)−1A(t) U (t, s) (λ − U (t, s) − κ)−1dλ u
2πiRΓ(Logλ) (λ − U (t, s) − κ)−2 U (t, s) dλ A(t) u
for u ∈ Y . A part of the right hand side is calculated as
1
κ
λ (λ − U (t, s) − κ)−1 dλ − 1
λ (λ − U (t, s) − κ)−1 dλ
2πiRΓ
κ
1
λ dλ
1
λ (λ − U (t, s) − κ)−1 U (t, s) dλ
1
λ (λ − U (t, s) − κ)−1 {λ − κ − (λ − U (t, s) − κ)} dλ
= 1
= 1
= 1
= 1
= 1
2πiRΓ
2πiRΓ
2πiRΓ (Logλ) (λ − U (t, s) − κ)−2U (t, s) dλ
2πiRΓ
2πiRΓ
2πiRΓ(λ − U (t, s) − κ)−1 dλ − 1
2πiRΓ(λ − U (t, s) − κ)−1 dλ − 1
2πiRΓ(λ − U (t, s) − κ)−1 dλ
−κ(U (t, s) + κI)−1(cid:8) 1
2πiRΓ
2πiRΓ(λ − U (t, s) − κ)−1 dλ
2πiRΓ(λ − U (t, s) − κ)−1dλ − 1
−κ(U (t, s) + κI)−1(cid:8) 1
2πiRΓ(λ − U (t, s) − κ)−1dλ
2πiRν=rP∞
= (I − κ(U (t, s) + κI)−1)
= (I − κ(U (t, s) + κI)−1)
= I − κ(U (t, s) + κI)−1,
λ (U (t, s) + κI)(λ − U (t, s) − κ)−1dλ(cid:9)
λ dλ(cid:9)
2πiRΓ
U(t,s)n
ν n+1 dν
= 1
n=1
1
1
1
1
due to the integration by parts, where λ − κ = ν = r is a properly chosen circle
large enough to include Γ. (2πi)−1RΓ λ−1dλ = 0 is seen by applying dLogλ/dλ =
n=1 U (t, s)nν−n−1 dν = I follows from the singularity of
1/λ.
ν−n−1.
(2πi)−1Rν=rP∞
Consequently we have
A(t) u = {I − κ(U (t, s) + κI)−1}−1 ∂tLog (U (t, s) + κI) u
= (U (t, s) + κI)U (t, s)−1 ∂tLog (U (t, s) + κI) u
= (I + κU (s, t)) ∂tLog (U (t, s) + κI) u
for u ∈ Y .
(cid:3)
What is introduced by Eq. (10) is a kind of resolvent approximation of A(t)
∂tLog (U (t, s) + κI) = (I + κU (s, t))−1A(t),
in which A(t) is approximated by the resolvent of U (s, t). As seen in the following
it is notable that there is no need to take κ → 0. This point is different from the
usual treatment of resolvent approximations. On the other hand, it is also seen by
Eq. (10) that
∂tLog (U (t, s) + κI) = (U (t, s) + κI)κ=0A(t)(U (t, s) + κI)−1
shows a structure of similarity transform, where (U (t, s) + κ)κ=0 means U (t, s) + κ
satisfying a condition κ = 0.
Under the validity of Theorem 3.1, for −T ≤ t, s ≤ T , let a(t, s) be defined by
a(t, s) := Log(U (t, s) + κI),
6
YORITAKA IWATA
then Eq. (10) is written as A(t) = (I + κU (s, t)) ∂ta(t, s). Since κ is chosen to
separate the spectral set of U (t, s) + κI from the origin, the inverse operator of
(I + κU (s, t)) = U (s, t)(U (t, s) + κI)
is well defined as (I + κU (s, t))−1 = (U (t, s) + κI)−1U (t, s). It also ensures that
∂ta(t, s) is well defined.
Corollary 3.2. Let t and s satisfy 0 ≤ t, s ≤ T . For U (t, s) and A(t) satisfying the
assumption of Theorem 3.1, the exponential of a(t, s) is represented by a convergent
power series:
(15)
ea(t,s) =P∞
n=0
a(t,s)n
n!
,
with a relation ea(t,s) = exp(Log(U (t, s) + κI)) = U (t, s) + κI.
different t and s are further assumed to commute,
If a(t, s) with
(16)
∂tea(t,s)us = ∂ta(t, s) ea(t,s)us
is satisfied for us ∈ Y , where ∂t denotes a t−differential in a weak sense.
Proof. Since a(t, s) is a bounded operator on X (Lemma 2.1), the exponential of
a(t, s) is represented by a convergent power series [9]:
(17)
where
ea(t,s) =P∞
n=0
a(t,s)n
n!
,
ea(t,s) = exp(Log(U (t, s) + κI))
= 1
2πiRΓ eLogλ (λ − U (t, s) − κ)−1dλ
= U (t, s) + κI
is satisfied. Since a(t, s) with different t and s commute,
∂t{a(t, s)}n = n{a(t, s)}n−1 ∂ta(t, s)
leads to
(18)
∂tea(t,s)us = ea(t,s) (∂ta(t, s))us
for us ∈ Y . This is a linear evolution equation satisfied by ea(t,s)us.
(cid:3)
Further calculi on Eq. (18) lead to
∂tea(t,s)us = ∂t(U (t, s) + κI)us = ∂tU (t, s)us,
and
ea(t,s) (∂ta(t, s))us = (U (t, s) + κI)∂t(Log(U (t, s) + κI))us
= (U (t, s) + κI)(U (t, s) + κI)−1U (t, s)A(t)us
= U (t, s)A(t)us = A(t)U (t, s)us,
where Theorem 3.1 is applied. As a result
∂tU (t, s)us = A(t)U (t, s)us
is obtained. Note that ea(t,s) does not satisfy the semigroup property, while U (t, s)
satisfies it.
INFINITESIMAL GENERATORS ...
7
4. Abstract Cauchy problem
4.1. Autonomous case. Logarithmic representation is utilized to solve autonomous
Cauchy problem
(19)
( ∂tu(t) = A(t)u(t)
u(s) = us,
in X, where A(t) ∈ G(X) : Y → X is assumed to be an infinitesimal generator
of U (t, s), −T ≤ t, s ≤ T is satisfied, Y is a dense subspace of X permitting the
representation shown in Eq. (10), and us is an element of X.
As seen in Eq. (16), under the assumption of commutation, a related Cauchy
problem is obtained as
(20)
( ∂tv(t, s) = (∂ta(t, s)) v(t, s)
v(s, s) = ea(s,s)us,
in X, where ∂ta(t, s) = ∂tLog(U (t, s) + κI) is well-defined. It is possible to solve
re-written Cauchy problem, and the solution is represented by
for us ∈ X (cf. Eq. (17)).
v(t, s) = ea(t,s)us =P∞
n=0
a(t,s)n
n! us
Theorem 4.1. Operator ea(t,s) is holomorphic.
Proof. According to the boundedness of a(t, s) on X (Lemma 2.1), ∂n
possible to be represented as
t ea(t,s) [18] is
(21)
for a certain κ, where λneλ does not hold any singularity for any finite λ. Following
the standard theory of evolution equation,
t ea(t,s) = 1
∂n
2πiRΓ λneλ(λ − a(t, s))−1 dλ,
k∂n
t ea(t,s)k ≤ Cθ,n
π(t sin θ)n
is true for a certain constant Cθ,n (n = 0, 1, 2, · · · ), where θ ∈ (0π/2) and arg t <
π/2 are satisfied (for the detail, e.g., see [17]). It follows that
(22)
lim
t→+0
sup tnk∂n
t ea(t,s)k ≤ lim
t→+0
sup tn Cθ,n
π(t sin θ)n < ∞.
Consequently, for z − t < t sin θ, the power series expansion
is uniformly convergent in a wider sense. Therefore ea(t,s) is holomorphic.
(cid:3)
(z−t)n
n! ∂n
t ea(t,s)
n=0
P∞
Theorem 4.2. For us ∈ X there exists a unique solution u(t) ∈ C([−T, T ]; X) of
(19) with a convergent power series representation:
(23)
u(t) = U (t, s)us = (ea(t,s) − κI)us =(cid:16)P∞
where κ is a certain complex number.
n=0
a(t,s)n
n! − κI(cid:17) us,
Proof. The unique existence follows from the assumption for A(t). ea(t,s) is holo-
morphic function (Theorem 4.1) with the convergent power series representation
(Eq. (17)). The solution of the original Cauchy problem is obtained as
a(t,s)n
u(t) = (I + κU (s, t))−1v(t, s) = (I + κU (s, t))−1P∞
n=0
n! us
8
YORITAKA IWATA
for the initial value us ∈ X. Note that A(t) is not assumed to be a generator of
analytic evolution family, but only a generator of invertible evolution family.
(cid:3)
For Iλ denoting the resolvent operator of A(t), the evolution operator defined
by the Hille-Yosida approximation is written by
u(t) = lim
λ→0
s IλA(τ ) dτ )us,
exp(R t
so that more informative representation is provided by Theorem 4.2 compared to
the standard theory based on the Hille-Yosida theorem.
4.2. Non-autonomous case. Series representation in autonomous part leads to
the enhancement of the solvability. Let Y be a dense subspace of X permitting
the representation shown in Eq. (10), and us is an element of X. Let us consider
non-autonomous Cauchy problem
(24)
( ∂tu(t) = A(t)u(t) + f (t)
u(s) = us
in X, where A(t) ∈ G(X) : Y → X is assumed to be the infinitesimal generator
U (t, s), f ∈ L1(−T, T ; X) is locally Holder continuous on [−T, T ]
kf (t) − f (s)k ≤ CH t − sγ
for a certain positive constant CH , γ ≤ 1 and −T ≤ t, s ≤ T . The solution of
non-autonomous problem does not necessarily exist in such a setting (in general,
f ∈ C([−T, T ]; X) is necessary).
Theorem 4.3. Let f ∈ L1(−T, T ; X) be locally Holder continuous on [−T, T ]. For
us ∈ X there exists a unique solution u(t) ∈ C([−T, T ]; X) for (24) such that
u(t) = [P∞
n=0
a(t,s)n
n! − κI]us +R t
s [P∞
n=0
using a certain complex number κ.
a(t,τ )n
n! − κI]f (τ )dτ
Proof. Let us start with cases with f ∈ C([−T, T ]; X). The unique existence follows
from the standard theory of evolution equation. The representation follows from
that of U (t, s) and the Duhamel's principle
(25)
u(t, s) = U (t, s)us +R t
= (ea(t,s) − κI)us +R t
s U (t, τ )f (τ )dτ
s [ea(t,τ ) − κI]f (τ )dτ,
where the convergent power series representation of ea(t,s) is valid (cf. Eq. (17)).
Next let us consider cases with the locally Holder continuous f (t). According to
the linearity of Eq. (24), it is sufficient to consider the inhomogeneous term. For ǫ
satisfying 0 < ǫ << T ,
s
R t+ǫ
[ea(t,τ ) − κI]f (τ )dτ → R t
s [ea(t,τ ) − κI]f (τ )dτ
INFINITESIMAL GENERATORS ...
9
is true by taking ǫ → 0. On the other hand,
(26)
s A(t)U (t, τ )f (τ )dτ
s A(t)U (t, τ )f (t)dτ
∂τ U (t, τ )f (t)dτ
s
s
[ea(t,τ ) − κI]f (τ )dτ =R t+ǫ
A(t)R t+ǫ
s A(t)U (t, τ )(f (τ ) − f (t))dτ +R t+ǫ
=R t+ǫ
=R t+ǫ
s A(t)U (t, τ )(f (τ ) − f (t))dτ −R t+ǫ
=R t+ǫ
=R t+ǫ
−U (t, t + ǫ)f (t) + U (t, s)f (t),
s
s A(t)U (t, τ )(f (τ ) − f (t))dτ − U (t, t + ǫ)f (t) + U (t, s)f (t)
(1 + κU (s, t))∂ta(t, s)[ea(t,τ ) − κI](f (τ ) − f (t))dτ
where ∂τ U (t, τ ) = −A(τ )U (t, τ ) is utilized. The Holder continuity and Eq. (22)
lead to the strong convergence of the right hand of Eq. (26):
[ea(t,τ ) − κI]f (τ )dτ
s (1 + κU (s, t))(∂ta(t, s))[ea(t,τ ) − κI](f (τ ) − f (t))dτ + (U (t, s) − I)f (t)
(due to ǫ → 0) for f ∈ L1(0, T ; X). A(t) is assumed to be an infinitesimal generator,
so that A(t) is a closed operator from Y to X. It follows that
s
A(t)R t+ǫ
→ R t
A(t)R t
=R t
and
s [ea(t,τ ) − κI]f (τ )dτ ∈ Y
R t
s [ea(t,τ ) − κI]f (τ )dτ
s (1 + κU (s, t))(∂ta(t, s))[ea(t,τ ) − κI](f (τ ) − f (t))dτ + (U (t, s) − I)f (t) ∈ X.
The right hand side of this equation is strongly continuous on [−T, T ]. As a result,
(∂ta(t, τ ))ea(t,τ )f (τ )dτ
s (1 + κU (τ, t))−1A(t)(U (t, τ ) + κI)f (τ )dτ
s
s
[ea(t,τ ) − κI]f (τ )dτ
∂tR t+ǫ
= [ea(t,t+ǫ) − κI]f (t + ǫ) +R t+ǫ
→ f (t) +R t
= f (t) +R t
= f (t) + A(t)R t
s A(t)U (t, τ )f (τ )dτ
s [ea(t,τ ) − κI]f (τ )dτ.
s [ea(t,τ ) − κI]f (τ )dτ satisfies Eq. (24), and that it is sufficient to
assume f ∈ L1(0, T ; X) as Holder continuous.
(cid:3)
We see that R t
This result should be compared to the standard theory of evolution equations in
which the inhomogeneous term f is assumed to be continuous on [−T, T ].
If the inhomogeneous term f ∈ Lp([0, T ]; X) is further assumed to be satisfied
for 1 < p < ∞, and Y = D(A(t)) = D(A(0)) and A(·) ∈ C([0, T ], L(Y, X)),
Eq. (20) with such an inhomogeneous term corresponds to the equation exhibiting
the maximal regularity of type Lp [16].
5. Concluding remark
As for the applicability of the theory, the conditions to obtain the logarithmic
representation (conditions shown in Sec. 2) are not so restrictive; indeed, they can
be satisfied by C0-groups generated by t-independent infinitesimal generators. The
most restrictive condition to obtain Eq. (10) is the commutation between A(t) and
U (t, s). Such a commutation is trivially satisfied by t-independent A(t) = A, and
10
YORITAKA IWATA
also satisfied when the variable t is separable (i.e., for an integrable function g(t),
A(t) = g(t)A). In this sense the operators specified in Theorem 3.1 correspond to
a moderate generalization of t-independent infinitesimal generators.
References
[1] W. Arendt, Semigroups and evolution equations:
functional calculus, regularity and kernel
estimates. Handbook of Differential Equations 1 (2002) 185.
[2] K. N. Boyadzhiev, Logarithms and imaginary powers of operators on Hilbert spaces. Collect.
Math. 45 3 (1994) 287-300.
[3] N. Dunford, Spectral theory I, convergence to projections. Trans. Amer. Math. Soc. 54 1943
185-217.
[4] M. Hasse, Spectral properties of operator logarithms. Math. Z. 245 4 (2003) 761-779.
[5] M. Hasse, The functional calculus for sectorial operators. Birkhauser, 2006.
[6] T. Kato, Perturbation Theory for Linear Operators. Springer-Verlag, 1966.
[7] T. Kato, Linear evolution equation of "hyperbolic" type. J. Fac. Sci. Univ. Tokyo 17 (1970)
241-258.
[8] T. Kato, Linear evolution equation of "hyperbolic" type II. J. Math. Soc. Japan 25 4 (1973)
648-666.
[9] T. Kato, A short introduction to perturbation theory for linear operators. Springer-Verlag;
1982.
[10] S. G. Krein, Linear differential equations in Banach space (translated from Russian). Transl.
Math. Monogr. 29, Amer. Math. Soc., 1971.
[11] C. Martinez and M. Sanz, The theory of fractional powers of operators. North-Holland, 2001.
[12] V. Nollau, Uber den Logarithmus abgeschlossener Operatoren in Banachschen Raumen, Acta
Sci. Math. 300 (1969) 161-174.
[13] N. Okazawa, Logarithms and imaginary powers of closed linear operators. Integral Equations
and Operator Theory 38 4 (2000) 458-500.
[14] N. Okazawa, Logarithmic characterization of bounded imaginary powers. Progress in Nonlin-
ear Differential Equations and Their Applications 42 (2000) 229-237.
[15] A. Pazy, Semigroups of linear operators and application to partial differential equations.
Springer-Verlag, 1983.
[16] J. Pruss and R. J. Schnaubelt, Math. Anal. Appl. 256 (2001) 405-430.
[17] H. Tanabe, Equations of evolution. Pitman, 1979.
[18] A. E. Taylor, Spectral theory of closed distributive operators. Acta Math. 84 (1951) 189-224.
(Yoritaka Iwata) Institute of Innovative Research, Tokyo Institute of Technology;
Department of Mathematics, Shibaura Institute of Technology.
E-mail address: iwata [email protected]
|
1603.04897 | 1 | 1603 | 2016-03-15T21:55:51 | Locally piecewise affine functions and their order structure | [
"math.FA"
] | Piecewise affine functions on subsets of $\mathbb R^m$ were studied in \cite{Ovchinnikov:02,Aliprantis:06a,Aliprantis:07a,Aliprantis:07}. In this paper we study a more general concept of a locally piecewise affine function. We characterize locally piecewise affine functions in terms of components and regions. We prove that a positive function is locally piecewise affine iff it is the supremum of a locally finite sequence of piecewise affine functions. We prove that locally piecewise affine functions are uniformly dense in $C(\mathbb R^m)$, while piecewise affine functions are sequentially order dense in $C(\mathbb R^m)$.
This paper is partially based on \cite{Adeeb:14}. | math.FA | math |
LOCALLY PIECEWISE AFFINE FUNCTIONS AND
THEIR ORDER STRUCTURE
S. ADEEB AND V. G. TROITSKY
Abstract. Piecewise affine functions on subsets of Rm were stud-
ied in [11, 7, 8, 9]. In this paper we study a more general concept
of a locally piecewise affine function. We characterize locally piece-
wise affine functions in terms of components and regions. We prove
that a positive function is locally piecewise affine iff it is the supre-
mum of a locally finite sequence of piecewise affine functions. We
prove that locally piecewise affine functions are uniformly dense
in C(Rm), while piecewise affine functions are sequentially order
dense in C(Rm).
This paper is partially based on [3].
1. Piecewise affine functions
We start with a brief overview of affine and piecewise affine functions;
we refer the reader to [11, 7] and [9, Chapter 7] for details.
Fix m ∈ N. By an affine function we mean a function f : Rm → R
of form f (x) = hv, xi+b for some v ∈ Rm and b ∈ R. Its kernel {f = 0}
is an affine hyperplane in Rm. It is easy to see that any two affine
functions which agree on a non-empty open set are equal. We write
A for the vector space of all affine functions on Rm. Clearly, A is a
subspace of C(Rm), the space of all real-valued continuous functions
on Rm.
Let Ω be a convex closed subset of Rm with non-empty interior. Fol-
lowing [9], we call such a set a solid domain. A continuous function
f : Ω → R is called piecewise affine if it agrees with finitely many
Date: December 14, 2017.
2010 Mathematics Subject Classification. Primary: 46A40. Secondary: 46E05.
Key words and phrases. affine function, piecewise affine function, locally piece-
wise affine function, vector lattice, sublattice.
The authors were supported by NSERC grants.
1
2
S. ADEEB AND V. G. TROITSKY
affine functions. That is, there exist f1 . . . , fn ∈ A such that for ev-
ery x ∈ Ω there exists i ≤ n with f (x) = fi(x). Equivalently (see
Definition 7.10 and Theorem 7.13 in [9]), a function f : Ω → R is piece-
wise affine iff there exist distinct affine functions f1, . . . , fn and subsets
S1, . . . , Sn of Ω such that
(1) on Si, f agrees with fi;
(2) Si = Int Si 6= ∅ for every i;
(3) Int Si ∩ Int Sj = ∅ whenever i 6= j;
(4) Sn
i=1 Si = Ω.
In particular, it follows from (2) that the set {f = fi} has non-empty in-
i=1, and (cid:8)(Si, fi)(cid:9)n
terior for every i = 1, . . . , n. The sets {fi}n
i=1
are called the components, regions, and characteristic pairs of f ,
respectively. The characteristic pairs of f are unique up to re-ordering.
We write PA(Ω) for the space of all piecewise affine functions on Ω.
We put PA = PA(Rm).
i=1, {Si}n
Suppose f ∈ PA(Ω) with characteristic pairs (cid:8)(Si, fi)(cid:9)n
i=1. For each
pair (i, j) with 1 ≤ i, j ≤ n and i 6= j, the set (cid:8)x ∈ Rm : fi(x) = fj(x)(cid:9),
or {fi = fj} is a hyperplane in Rm. Let C be the union of all such
hyperplanes. Then C is a closed set with empty interior. It is easy to
see that the set (Int Ω)\C is open, dense in Ω, and naturally partitioned
into a finite number of convex open subsets. We call these subsets the
cells of f .
Let K be a cell of f . Since K is disjoint from C, no two components
agree at any point of K. It follows that there exists a component fi
such that f agrees with fi on K, so that K ⊆ Si. It also follows that for
every j 6= i we either have fj(x) < f (x) for all x ∈ K or fj(x) > f (x)
for all x ∈ K.
Example 1.1. Let m = 1 and f (t) = min(cid:8)t, 1(cid:9). Then f is piecewise
affine on R, with components f1(t) = t, f2(t) = −t, and f3(t) = 1, and
with regions S1 = [0, 1], S2 = [−1, 1], and S3 = (−∞, −1] ∪ [1, +∞).
Note that S3 is not connected. The cells of f are K1 = (−∞, −1),
K2 = (−1, 0), K3 = (0, 1), and K4 = (1, +∞).
Recall that under pointwise order, the space C(Ω) is a vector lattice.
It is easy to see that PA(Ω) is a (vector) sublattice of C(Ω).
LOCALLY PIECEWISE AFFINE FUNCTIONS
3
Let A be a subset of a vector lattice V . Recall that
A∨ = (cid:8)x1 ∨ · · · ∨ xn : n ∈ N, x1, . . . , xn ∈ A(cid:9) and
A∧ = (cid:8)x1 ∧ · · · ∧ xn : n ∈ N, x1, . . . , xn ∈ A(cid:9).
If A is a linear subspace of V then A∨∧ = A∧∨ and this set is exactly
the sublattice generated by A in V ; see, e.g., [9, Lemma 1.21].
Since A ⊆ PA and PA is a sublattice of C(Rm), it follows that
A∧∨ ⊆ PA. In fact, [9, Theorem 7.23] asserts that PA = A∧∨, hence
PA is exactly the sublattice of C(Rm) generated by A. Moreover, if Ω is
a solid domain in Rm and f ∈ PA(Ω) with component set F = {fi}n
i=1
then f ∈ F ∧∨ or, more precisely, f agrees with the restriction to Ω of
a function in F ∧∨. Since every function in F ∧∨ is defined on all of Rm,
it follows that every function in PA(Ω) extends to a function in PA.
We write Bm
∞ for the closed unit ball of ℓm
∞:
xi ≤ 1(cid:9).
∞ = (cid:8)x ∈ Rm : max
Bm
1≤i≤m
For n ∈ N, we write Ωn for the ball of radius n centred at zero in ℓm
∞;
i.e., Ωn = nBm
∞.
2. Locally piecewise affine functions
A function f : Rm → R is said to be locally piecewise affine if its
restriction to any bounded solid domain Ω is in PA(Ω). The following
lemma is straightforward.
Lemma 2.1. Let f : Rm → R. Then f is locally piecewise affine iff
the restriction of f to Ωn is in PA(Ωn) for every n ∈ N.
We write LPA for the collection of all locally piecewise affine func-
tions on Rm. Clearly, LPA is a sublattice of C(Rm) and A ( PA (
LPA.
Let F ⊂ C(Rm). We say that F is locally finite if for every compact
set C all but finitely many functions in F vanish on C. That is, the
set (cid:8)f ∈ F : ∃x ∈ C f (x) 6= 0(cid:9) is finite. Just as in Lemma 2.1, it is
easy to see that F is locally finite iff for every n all but finitely many
functions in F vanish on Ωn. Clearly, any subset of a locally finite set
4
S. ADEEB AND V. G. TROITSKY
of functions is again locally finite. Additionally, if F ⊂ C(Rm) is a
locally finite set of functions, then F ∨, F ∧ and consequently F ∨∧ are
all locally finite. If F is a sequence, then F is termed a locally finite
sequence of functions.
It is a standard fact that C(Rm) is not order complete. Yet we have
the following.
Lemma 2.2. Every locally finite subset of PA has a supremum and an
infimum in C(Rm), which belong to LPA and agree with the pointwise
maximum and minimum, respectively.
Proof. Let F ⊆ PA be a locally finite set. For every point x ∈ Rm, all
but finitely many members of F vanish at x. Hence, h(x) = max(cid:8)f (x) :
f ∈ F(cid:9) is defined. Similarly, on every open ball B in Rm, h is the point-
wise maximum of a finite subcollection of F , so that h is continuous
on B; it follows that h ∈ C(Rm). It follows from the definition that
h = sup F in C(Rm). It also follows that on every ball, h agrees with
a finite subset of F , hence h is locally piecewise affine. The proof for
the infimum is similar.
(cid:3)
We write 1 for the constant one function on Rm.
Lemma 2.3. For every ε > 0 there exists f ∈ PA such that 0 ≤ f ≤ 1,
f equals 1 on Bm
∞, and vanishes outside of (1 + ε)Bm
∞.
Proof. Put g(x) = kxk∞ = maxi≤mxi for x ∈ Rm. Put h : R+ → R+
so that h is continuous, equals 1 on [0, 1], vanishes on [1 + ε, +∞)
and is affine on [1, 1 + ε]. Let f = h ◦ g. Being the composition of
two piecewise affine functions, f is itself piecewise affine. It is easy to
verify that f satisfies the rest of the requirements.
(cid:3)
Theorem 2.4. For every 0 ≤ f ∈ LPA there exists a locally finite
sequence (fi) in PA+ such that f = sup fi.
Proof. Let (cn) be an enumeration of all integer points in Rm. Consider
the "boxes" Kn = cn + Bm
∞ and eKn = cn + 2Bm
∞.
Fix n. Since f is continuous, it is bounded on Kn. Let M > 0 be
such that f (x) < M for all x ∈ Kn. As in Lemma 2.3, we can find a
piecewise affine function gn such that 0 ≤ gn ≤ M 1, gn equals M on
LOCALLY PIECEWISE AFFINE FUNCTIONS
5
Kn, and gn vanishes outside eKn. Put fn = f ∧ gn. We will show that
the sequence (fn) satisfies all the requirements.
By construction, 0 ≤ fn ≤ f , fn agrees with f on Kn and vanishes
outside eKn. Observe that fn is piecewise affine because on eKn it agrees
with f ∧ gn, which is piecewise affine on eKn and fn vanishes outside
of eKn. The sequence (fn) is locally finite because for every n the only
members of the sequence that do not vanish on Kn are fn itself and
the members corresponding to the (finitely many) boxes that meet eKn.
Finally, f = sup fn because 0 ≤ fn ≤ f and fn agrees with f on Kn
for every n.
(cid:3)
3. Characteristic pairs of locally piecewise functions
Recall that every piecewise affine function is determined by its com-
ponents and regions. The following theorem provides a similar charac-
terization for locally piecewise affine functions.
Theorem 3.1. A function f : Rm → R is locally piecewise affine iff
there exist a sequence of distinct affine functions (fi) and a sequence
(Si) of subsets of Rm such that
(1) on Si, f agrees with fi;
(2) Si = Int Si 6= ∅ for every i;
(3) Int Si ∩ Int Sj = ∅ whenever i 6= j;
(4) S∞
(5) every bounded set meets only finitely many Si's.
i=1 Si = Rm;
Proof. Suppose that f satisfies the five conditions; show that it is lo-
cally piecewise affine. First, observe that f is continuous. Indeed, sup-
pose that xn → x in Rm but f (xn) 6→ f (x). Passing to a subsequence,
we can find an ε > 0 such that (cid:12)(cid:12)f (xn) − f (x)(cid:12)(cid:12) > ε and xn ∈ B(x, 1) for
every n. By (5), the ball B(x, 1) meets only finitely many Si's. Hence,
passing to a further subsequence, we may assume that the sequence
(xn) is contained in Si for some i. Since Si is closed, we have x ∈ Si. It
follows that f (xn) = fi(xn) → fi(x) = f (x); a contradiction. Thus, f
is continuous. By (5) again, every bounded solid domain Ω is contained
in the union of finitely many Si's; it follows that the restriction of f to
Ω is in PA(Ω), so that f is locally piecewise affine.
6
S. ADEEB AND V. G. TROITSKY
Suppose now that f is locally piecewise affine. Fix n ∈ N. The
restriction of f to Ωn is in PA(Ωn); consider the affine components of
this restriction. By collecting the affine functions associated with Ωn
for all n ∈ N, we produce a sequence (fi) in A such that for every n
there exists kn such that f agrees with f1, . . . , fkn on Ωn. By dropping
repeated terms, we may assume without loss of generality that all the
members in this sequence are distinct.
For every i, put Ui = Int{f = fi} and Si = Ui. Since fi is a
component of the restriction of f to Ωn for some n, it follows that
Ui 6= ∅. It is easy to see that (1) and (2) are satisfied. To show (3),
suppose that Int Si meets Int Sj for some i and j. This follows that fi
and fj agree on an open set, hence fi = fj and, therefore, i = j.
Fix x ∈ Rm and take n ∈ N such that x ∈ Int(Ωn). Let K1, . . . , Kl
be the cells generated in Ωn by f1, . . . , fkn. Since Sl
j=1 Kp is dense in
Ωn, we have x ∈ Kp for some p ≤ l. Note that Kp is open and f agrees
with some fi on Kp. It follows that Kp ⊆ Ui and, therefore, x ∈ Si.
This yields (4).
Finally, we will show that Int(Ωn) only meets S1, . . . , Skn; this will
immediately yield (5). Suppose that Int(Ωn) meets Si for some i ∈ N.
It follows that Int(Ωn) meets Ui. This implies that Ui meets Kr for
some r ≤ l. Note that f agrees with fj on Kr for some j ≤ kn. It
follows that fj and fi agree on the non-empty open set Kr ∩ Ui, hence
j = i and, therefore, i ≤ kn.
(cid:3)
As before, the sets {fi}∞
i=1, and (cid:8)(Si, fi)(cid:9) are called the
components, regions, and characteristic pairs of f , respectively.
i=1, {Si}∞
4. LPA is uniformly dense in C(Rm)
It is of interest whether every continuous function on Rm can be
approximated in some sense by piecewise affine functions. We will
consider two kinds of denseness here: uniform denseness and order
denseness. We start with uniform denseness. It is easy to see that PA
is not uniformly dense in C(Rm) even for m = 1.
LOCALLY PIECEWISE AFFINE FUNCTIONS
7
Theorem 4.1. LPA is uniformly dense in C(Rm). That is, for every
f ∈ C(Rm) and every ε > 0 there exists h ∈ LPA such that f − h ≤
ε1.
Proof. The proof is analogous to that of Theorem 2.4. Let f ∈ C(Rm)
and ε > 0.
First, we consider the case when f ≥ 0. Let (cn) be an enumeration
∞ and
of all integer points in Rm. Consider the "boxes" Kn = cn + Bm
eKn = cn + 2Bm
∞.
j (x) 6= e∗
j (y), where e∗
Fix n. Observe that PA( eKn) is a sublattice of C( eKn). It clearly
contains constant functions. It is easy to see that it separates points.
Indeed, let x, y ∈ eKn with x 6= y. Then xj 6= yj for some j ≤ m. That
is, e∗
j is the j-th coordinate functional. Clearly,
j is an affine function. Since eKn is compact, it now follows from the
e∗
lattice version of Stone-Weierstrass Theorem that PA( eKn) is uniformly
dense in C( eKn), see, e.g., Theorem 11.3 on page 88 of [4]. It follows
that there is hn ∈ PA( eKn) such that (cid:12)(cid:12)hn(x) − f (x)(cid:12)(cid:12) < ε for all x ∈ eKn.
Extend hn to a piecewise affine function hn in PA.
Since hn is continuous on eKn, it is bounded on eKn. Let Mn > 0
be such that hn(x) < Mn for all x ∈ eKn. As in Lemma 2.3, we can
find gn ∈ PA such that 0 ≤ gn ≤ Mn1, gn equals Mn on Kn, and gn
vanishes outside eKn.
Consider the sequence of functions (cid:0)hn ∧ gn(cid:1). Since gn vanishes
outside of eKn, all but finitely many terms of this sequence vanish on
Kn; it follows that the sequence (cid:0)hn ∧ gn(cid:1) is locally finite. Let h =
hn ∧ gn; by Lemma 2.2 this supremum exists, agrees with the
supn
pointwise maximum, and belongs to LPA.
For every n and every x ∈ Kn, we have
h(x) ≥ (hn ∧ gn)(x) = hn(x) ∧ Mn = hn(x) ≥ f (x) − ε.
It follows that h ≥ f − ε1, so that h − f ≥ −ε1.
On the other hand, for every n and every x ∈ eKn we have
(hn ∧ gn)(x) ≤ hn(x) ≤ f (x) + ε.
8
S. ADEEB AND V. G. TROITSKY
Since hn ∧ gn ≤ gn, it vanishes outside of eKn. It follows that hn ∧ gn ≤
f + ε1. Taking sup over n we get h ≤ f + ε1, so that h − f ≤ ε1.
Therefore, h − f ≤ ε1.
Suppose now that f is an arbitrary function in C(Rm). Then f =
f + − f −, with 0 ≤ f +, f − ∈ C(Rm). By the first part of the proof, we
find h1 and h2 in LPA such that f + − h1 ≤ ε
1.
Put h = h1 − h2, then
1 and f − − h2 ≤ ε
2
2
f − h ≤ f + − h1 + f − − h2 ≤ ε1.
(cid:3)
5. PA is σ-order dense in C(Rm)
Recall that a sublattice F of an Archimedean vector lattice E is said
to be order dense if for every 0 < x ∈ E there exists y ∈ F with
0 < y ≤ x. Equivalently, for every 0 < x ∈ E we have x = sup A where
A = {y ∈ F : 0 ≤ y ≤ x}; see, e.g., [6, Theorem 1.34]. Note that the
set A may be viewed as an increasing net. Therefore, F is order dense
in E iff every element of E+ is the supremum of an increasing net in
F+.
Let f ∈ C(Rm). Suppose that f > 0; that is, f (x) ≥ 0 for every
x ∈ Rm and f (c) > 0 for some c ∈ Rm. It follows easily from Lemma 2.3
that there exists a locally piecewise function h such that 0 < h ≤ f .
This means that PA is an order dense sublattice in C(Rm). It follows
that if 0 ≤ f ∈ C(Rm) then f is the supremum of an increasing net
(fα) in PA, that is, fα ↑ f . This leads to two natural questions.
Question 5.1. Can we replace a net with a sequence?
Question 5.2. How do we deal with non-positive functions?
The answer to the first question is related to the countable sup prop-
erty. Recall that a vector lattice E has the countable sup property
if every increasing net (xα) in E with x = sup xα has an increasing sub-
sequence with the same supremum. Thus, to answer the first question,
it suffices to prove that C(Rm) has the countable sup property.
Recall that for a sequence (xα) in a vector lattice E, the notation
xα ↑ means (xα) is increasing; xα ↑≤ x means xα ↑ and xα ≤ x for
LOCALLY PIECEWISE AFFINE FUNCTIONS
9
every α, and xα ↑ x means xα ↑ and x = supα xα. Note that fα ↑ f
in C(Ω) does not imply pointwise convergence. For f, g ∈ C(Rm), the
notation f < g means f ≤ g and f 6= g. For f ∈ C(Rm) and a solid
domain Ω in Rm, we write fΩ for the restriction of f to Ω; we view it
as an element of C(Ω).
Lemma 5.3. Let (fα) be a net in C(Rm) and f ∈ C(Rm). Then fα ↑ f
in C(Rm) iff fαΩn ↑ fΩn in C(Ωn) for every n.
Proof. Suppose that fα ↑ f in C(Rm). Fix n. It is clear that (fαΩn)
is an increasing net in C(Ωn) and fΩn is an upper bound of this net.
Suppose it is not the least upper bound. Then there exists g ∈ C(Ωn)
such that fαΩn ≤ g for every α but g(t0) < f (t0) for some t0 ∈ Ωn.
Since Ωn is a solid domain, we may assume that t0 ∈ Int Ωn. It follows
easily that there exists a neighborhood V of t0 and ε > 0 such that
V ⊆ Int Ωn and g(t) < f (t) − ε for all t ∈ V . Find 0 < h ∈ C(Rm) such
that h(t0) > 0, h(t) < ε for all t ∈ V , and h vanishes outside V . Then
f − h < f , yet fα ≤ f − h for every α; this contradicts f = sup fα.
Conversely, suppose that fαΩn ↑ fΩn in C(Ωn) for every n. First,
we claim that fα ↑≤ f in C(Rm). Indeed, fix any t0 ∈ Rm. Find n
so that t0 ∈ Ωn. It follows from fαΩn ↑ fΩn that fα(t0) ↑≤ f (t0). In
particular, f is an upper bound of (fα) in C(Rm).
On the other hand, suppose that g is an upper bound of (fα) in
C(Rm). Fix any n. Then gΩn ≥ fαΩn ↑ fΩn yields gΩn ≥ fΩn. This
is true for all n; therefore, g ≥ f . It follows that f = supα fα.
(cid:3)
Theorem 5.4. C(Rm) has the countable sup property.
Proof. Let (fα)α∈Λ be a net in C(Rm) indexed by some directed set Λ
such that fα ↑ f for some f ∈ C(Rm). Fix n ∈ N. By Lemma 5.3, we
have fαΩn ↑ fΩn in C(Ωn). Recall that every vector lattice which ad-
mits a strictly positive functional has the countable sup property; see,
e.g., [5, Theorem 2.6]. Since the Riemann integral is a strictly positive
functional on Ωn, it follows that C(Ωn) has the countable sup prop-
erty. Therefore, the net (fαΩn) has a subsequence whose supremum in
C(Ωn) is fΩn.
10
S. ADEEB AND V. G. TROITSKY
We will now inductively construct a double sequence (αn,k) in Λ
indexed by pairs (n, k) ∈ N2 satisfying the following two conditions
(1) (αn,k) is increasing: if n1 ≤ n2 and k1 ≤ k2 then αn1,k1 ≤ αn2,k2;
(2) for every n we have fαn,k Ωn
↑ fΩn as k → ∞.
For n = 1, let (α1,k)∞
fα1,k Ω1
k=1 be an increasing sequence in Λ such that
↑ fΩ1 as k → ∞; it exists by the preceding paragraph.
Suppose that we have constructed αi,k for all i = 1, . . . , n − 1 and all
k. We will now construct αn,k for all k. Find an increasing sequence
↑ fΩn in C(Ωn) as k → ∞. Now choose
(βk) in Λ such that fβk Ωn
αn,1 to be any index greater than β1 and αn−1,1. Choose αn,2 to be any
index greater than β2, αn−1,2, and αn,1. Proceed inductively: once we
have defined αn,1, . . . , αn,k−1, choose αn,k to be any index greater than
βk, αn−1,k, and αn,k−1. This yields
fΩn ≥ fαn,k Ωn
≥ fβk Ωn
↑ fΩn,
↑ fΩn in C(Ωn) as k → ∞. Also, it is easy to see that
so that fαn,k Ωn
(αn,k) is increasing.
Consider (αk,k)∞
k=1. This is clearly an increasing sequence in Λ. If
↑
n ≤ k then αn,k ≤ αk,k, so that fαn,k ≤ fαk,k ≤ f . This yields fαk,k Ωn
fΩn in C(Ωn) as k → ∞. It now follows from Lemma 5.3 that fαk,k ↑ f
in C(Rm).
(cid:3)
Corollary 5.5. For every 0 ≤ f ∈ C(Rm) there exists a sequence (hn)
of positive piecewise affine functions such that hn ↑ f in C(Rm).
Thus, Question 5.1 has the affirmative answer. We now pass to
Question 5.2. We answer it through order convergence. In literature,
there are two slightly different definitions of an order convergent net
in a vector lattice, which are, generally, not equivalent. We refer the
reader to [10, 1, 2]. In our case, it will be irrelevant which of the two
definitions is used for the following two reasons: first, both definitions
yield the properties that we will be using and, second, by [10], the
two definitions agree in vector lattices which have the countable sup
property; hence, they agree in C(Rm).
Proposition 5.6. Every function in C(Rm) is the order limit of an
order convergent sequence of piecewise affine functions.
LOCALLY PIECEWISE AFFINE FUNCTIONS
11
Proof. Let f ∈ C(Rm). Decompose it as f = f +−f −. By Corollary 5.5,
there exist positive sequences (gn) and (hh) in PA such that gn ↑ f +
and hn ↑ f − in C(Rm). It follows that the sequence (gn − hn) converges
in order to f in C(Rm).
(cid:3)
Acknowledgement. We would like to thank Foivos Xanthos for
valuable discussions.
References
[1] Y. Abramovich and C.D. Aliprantis, An invitation to Operator theory, Vol.
50. Providence, RI: American Mathematical Society, 2002.
[2] Y. Abramovich and G. Sirotkin, On order convergence of nets, Positivity, 9,
2005, 287–292.
[3] S. Adeeb, Locally piece-wise affine functions, M.Sc. Thesis, University of Al-
berta, Department of Mathematical and Statistical Sciences, 2014.
[4] C.D. Aliprantis and O. Burkinshaw, Principles of Real Analysis, Academic
Press, New York and London, 3rd Edition, 1998.
[5] C.D. Aliprantis and O. Burkinshaw, Locally solid Riesz spaces with applica-
tions to economics, 2nd ed., AMS, Providence, RI, 2003.
[6] C.D. Aliprantis and O. Burkinshaw, Positive Operators, 2nd ed., Springer
2006.
[7] C.D. Aliprantis, D. Harris, and R. Tourky, Continuous piecewise linear func-
tions, Macroeconomic Dynamics, 10(1) (2006), 77–99.
[8] C.D. Aliprantis, D. Harris, and R. Tourky. Riesz estimators, J. Econometrics,
136(2) (2007), 431–456.
[9] C.D. Aliprantis and R. Tourky, Cones and Duality, AMS, Providence, RI,
2007.
[10] R.F. Anderson and J.C Mathews, A Comparison of Two Modes of Order
Convergence, Proc. Amer. Math. Soc., 18 (1967), 100–104.
[11] S. Ovchinnikov, Max-min representation of piecewise linear functions, Beit-
rage Algebra Geom. 43 (2002), 297-302.
Department of Civil and Environmental Engineering, University of
Alberta, Edmonton, AB, T6G 1H9, Canada.
E-mail address: [email protected]
Department of Mathematical and Statistical Sciences, University
of Alberta, Edmonton, AB, T6G 2G1, Canada.
E-mail address: [email protected]
|
1403.0720 | 2 | 1403 | 2015-05-12T03:36:04 | Invariant subspaces with no generator and a problem of H. Helson | [
"math.FA"
] | In the almost periodic context, any $H_0^2-$space cannot be generated by one of its elements. Together with cocycle argument, this derives that there exist all kinds of invariant subspaces without single generator, from which we can answer some questions on invariant subspace theory. | math.FA | math |
INVARIANT SUBSPACES WITH NO GENERATOR
AND A PROBLEM OF H. HELSON
JUN-ICHI TANAKA
Dedicated to the memory of Henry Helson
Abstract. In the almost-periodic context, the H 2
0 −space cannot be generated by
one of its elements. Together with a cocycle argument, this implies that there exist
all kinds of invariant subspaces without single generator, from which we answer some
questions on invariant subspace theory.
1. Introduction
The theory of invariant subspaces has been developed in the context of compact
abelian groups with ordered duals, which is a natural generalization of such a theory
on the unit circle T. Many classical results extend to these cases, nevertheless, one also
meets new difficulties. The purpose of this paper is to resolve a longstanding problem
formulated by H. Helson in the 1950s.
Let Γ be a countable dense subgroup of the real line R, endowed with the discrete
topology. Then the dual group K of Γ is a compact abelian group that is metrizable.
For λ in Γ, it is customary to denote by χλ the character on K defined by χλ(x) = x(λ).
Let σ be the normalized Haar measure on K. A function φ in L1(σ) is analytic if its
Fourier coefficients
(1)
aλ(φ) = ZK
φ χλ dσ
vanish for all negative λ in Γ. The Hardy space H p(σ), 1 ≤ p ≤ ∞, is defined to be
the space of all analytic functions in Lp(σ). For technical reasons, it is useful to define
H p
0 (σ) as the subspace of all φ in H p(σ) with a0(φ) = 0. A (weak*-, if p = ∞) closed
subspace M of Lp(σ) is invariant if M contains χλM for all positive λ in Γ. When the
inclusion is strict, M is said to be simply invariant. Of course, both H p(σ) and H p
0 (σ)
are simply invariant subspaces of Lp(σ). If φ is in Lp(σ), and let M[φ] denotes the
smallest invariant subspace of Lp(σ) containing φ, then φ is called a single generator
of M[φ]. Recall that a function of modulus one is said to be unitary and an analytic
Date: Received by editors November 21, 2009, and in revised form, January 17, 2013.
1991 Mathematics Subject Classification. Primary 43A17; Secondary 46J10, 46J15, 28D10.
Key words and phrases. Compact groups with ordered duals, Invariant subspaces, Cocycles, Single
generators.
Partially supported by NSF grant no. 0649765.
1
unitary function is called an inner function. We say a function φ in H p(σ) is outer if
it satisfies that
log a0(φ) = ZK
log φ dσ > −∞ .
Let 1 ≤ q ≤ p ≤ ∞, and let M be a simply invariant subspace of Lp(σ). It follows from
the properties of outer functions that [M ∩ L∞(σ)]q ∩ Lp(σ) = M, where [M ∩ L∞(σ)]q
is the closure of M ∩ L∞(σ) in Lq(σ) (see [4, Chapter V, Section 6] for details). This
fact assures that there is a one-to-one correspondence between the invariant subspaces
in Lp(σ) and those in Lq(σ). Therefore, in dealing with invariant subspaces, we may
restrict our attention to the case of p = 2, in which Hilbert space theory works well. It
follows from Szego's theorem that φ is a single generator of H 2(σ) if and only if φ is
outer in H 2(σ). However, it has been unknown for a long time whether every simply
invariant subspace is singly generated or not. In the literature this has come to be
known as the single generator problem (refer to [6, §5.4], [3, Remark, p. 158] and [4,
p. 138 and p. 177]). The difficulty seems to center on the case of invariant subspace
H 2
0 (σ). In [9, p. 183], it is raised in an equivalent form in connection with stochastic
processes.
Our objective in this note is to show a negative answer to this problem in the almost
periodic settings:
Theorem. The invariant subspace H 2
0 (σ) cannot be generated by one of its elements.
To the best of author's knowledge, H 2
0 (σ) is the first known example of invariant
subspace which cannot be singly generated. On the other hand, by [6, §5.3, Theorem
33], it was shown that every invariant subspace is generated by two of its elements. In
more general setting, we can artificially make H 2
0 -spaces to have a single generator.
For each t in R, let us denote by et the element of K defined by et(λ) = eiλt for λ
in Γ. The map sending t to et embeds R continuously onto a dense subgroup of K.
Define a one-parameter group {Tt}t∈R of homeomorphisms on K by
(2)
Tt x = x + et ,
x ∈ K.
Then the pair (K, {Tt}t∈R) is a strictly ergodic flow, for which σ is the unique invariant
probability measure. The flow (K, {Tt}t∈R) is called an almost periodic flow, because if
φ is continuous on K, then t → φ(x + et) is a uniformly almost periodic function with
exponents in Γ. Let H ∞(dt/π(1+t2)) be the space of all boundary functions of bounded
analytic functions in the upper half-plane H, and let H p(dt/π(1 + t2)), 1 ≤ p < ∞, be
the closure of H ∞(dt/π(1+t2)) in Lp(dt/π(1+t2)). For a function u(x, t) on K × R, the
assertion " t → u(x, t) for σ − a.e. x in K " is sometimes abbreviated to "almost every
t → u(x, t) ". Then φ in Lp(σ) lies in H p(σ) if and only if almost every t → φ(x + et)
lies in H p(dt/π(1 + t2)). This fact enables us to define Hardy spaces on every ergodic
flow (see the end of the next section).
2
Let M be a simply invariant subspace of L2(σ). Set Mλ = χλM for each λ in Γ.
Define
M+ = ^λ<0
Mλ
and M− = _λ>0
Mλ.
Since these spaces are at most one dimension apart, M coincides with either or both
its versions M+ and M−. When M = M+, M is said to be normalized. For φ in
L2(σ), the subspace M[φ] is simply invariant if and only if
(3)
Z ∞
−∞
log φ(x + et)
dt
1 + t2 > −∞ ,
σ − a.e. x ∈ K,
It is well-known that there is a function φ in L2(σ)
(see [6, §3.3, Theorem 22 ]).
satisfying the inequality (3), while log φ does not belong to L1(σ). Our Theorem
asserts that any such function φ must satisfy M[φ]+ = M[φ]−.
A unitary Borel function A(x, t) on K × R is said to be a cocycle on K if A(x, t)
satisfies the cocycle identity
A(x, t + s) = A(x, t) · A(x + et, s) ,
(x, s, t) ∈ K × R × R .
We identify two cocycles which differ only on a set of dσ ×dt−measure zero in K ×R. A
one-to-one correspondence is established between normalized invariant subspaces and
cocycles (as discussed in [6, §2.3]). More precisely, let M be a simply invariant subspace
of L2(σ) with cocycle A(x, t). Then a function φ in L2(σ) lies in M+ if and only if
almost every t → A(x, t)φ(x + et) lies in H 2(dt/π(1 + t2)) (see [6, §3.2]). It is easy to
see that M+ 6= M− if and only if M+ = qH 2(σ) for some unitary function q on K.
Then the cocycle of M has the form q(x) · q(x + et), which is called a coboundary. If a
cocycle is a coboundary multiplied by exp(iαt) for some α in R, then such a cocycle is
said to be trivial. A trivial cocycle exp(iαt) is not a coboundary only if α lies in R \ Γ.
We already know from [8] and [12] that some singly generated subspaces have non-
trivial cocycles, but we can strengthen this fact by noting the following:
Corollary 1. Let M be a simply invariant subspace of L2(σ). If the cocycle of M is
trivial, then M− has no single generator. In other words, if M− is singly generated,
then the cocycle of M is always nontrivial, so that M+ = M−.
A cocycle with values in {−1, 1} is called a real cocycle. It follows from [10] that
there exist real cocycles which are nontrivial.
Corollary 2. Let M be a simply invariant subspace of L2(σ) with real cocycle. Then
M− has no single generator.
A cocycle A(x, t) is said to be analytic if almost every t → A(x, t) lies in H ∞(dt/π(1+
t2)). Then a normalized invariant subspace with analytic cocycle contains always
H 2(σ). We say that an analytic cocycle A(x, t) is a Blaschke or a singular cocycle, if
almost every t → A(x, t) is an inner function of that type in H ∞(dt/π(1 + t2)). Two
cocycles are called cohomologous if one is a coboundary times the other. It is known
that every cocycle is cohomologous to a Blaschke cocycle in some restricted class (see
3
[6, §4.6, Theorem 26] and [17]). This fact makes Blaschke cocycles so important for
the subject. Using our Theorem, we may answer some questions on analytic cocycles:
Corollary 3. In the class of analytic cocycles, the following properties hold:
(a) There is a Blaschke cocycle not being cohomologous to any singular cocycle.
(b) There is a Blaschke cocycle not having exactly the same zeros as any function
in H 2(σ).
It would be helpful to understand the basic idea behind the proof of our Theorem.
On the one hand, we claim that if φ is a single generator of H 2
0 (σ), then φ must have
a very special form. Assume that Γ is the smallest group determined by the nonzero
Fourier coefficients of φ (see below for details). Similarly, let Λ be the smallest group
determined by the nonzero coefficients of φ. Since Λ is a subgroup of Γ, the dual
group of Λ is represented as K/H, where H is the annihilator of Λ in K. Let τ be the
normalized Haar measure on K/H, and fix an element α in Γ with aα(φ) 6= 0. Then
it can be shown that χαφ lies in L2(τ ) and generates the simply invariant subspace of
L2(τ ) with trivial cocycle exp(iαt). We also see that α is independent of Λ, meaning
that nα lies in Λ only for n = 0 in the integer group Z. This implies that K and dσ are
respectively identified with K/H × T and dτ × dθ/2π, since H is regarded as T. Thus,
for each single generator φ of H 2
0 (σ), we derive that Γ 6= Λ. On the other hand, if
H 2
0 (σ) with the property
that Γ = Λ, which contradicts the existence of single generator of H 2
0 (σ) is singly generated, we may construct a generator φ of H 2
0 (σ).
In the next section, we establish some notation and elementary facts about invariant
subspaces in the almost periodic setting. Using group characters, we develop certain
properties of single generators of H 2
0 -spaces in Section 3. In Section 4, the proof of our
Theorem is provided and then Corollaries are proved by using a lemma on cocycles.
We conclude the paper with some remarks in Section 5.
We refer the reader to [1], [4, Chapter VII], [6] and [16, Chapter VIII] for further
details on analyticity on compact abelian groups. Basic results concerning the Hardy
space theory based on uniform algebras can be found in [4, Chapter IV] and [13].
Part of this work was done while the author was visiting the University of North
Carolina at Chapel Hill, and he would like to acknowledge the hospitality of the De-
partment of Mathematics. Especially, he would like to express his sincere gratitude
to Professors Joe Cima and Karl Petersen for helpful discussions. Thanks are due to
the referee as well for his valuable suggestions, which improved the first version of this
paper so much.
4
2. Extension of almost periodic functions
It is easy to show that a function φ in H 2(σ) is outer if and only if a0(φ) 6= 0 and
almost every t → φ(x + et) is outer in H 2(dt/π(1 + t2)). A weak version of this fact
stated below is often used in what follows:
Lemma 1. Let M be a simply invariant subspace of L2(σ) with cocycle A(x, t). A
function φ in L2(σ) generates M− if and only if log φ does not lie in L1(σ) and
almost every t → A(x, t)φ(x + et) is outer in H 2(dt/π(1 + t2)). In particular, H 2
0 (σ) is
singly generated by φ if and only if a0(φ) = 0 and almost every t → φ(x + et) is outer
in H 2(dt/π(1 + t2)).
Proof. Suppose that M[φ] = M− for φ in L2(σ). If log φ lies in L1(σ), then there is
a unitary function q on K such that M[φ] = qH 2(σ) by Szego's theorem. This implies
that M[φ] 6= M−, so log φ cannot lie in L1(σ). Let B(x, t) be the analytic cocycle
defined by the inner part of t → A(x, t)φ(x + et). Let N be the invariant subspace with
cocycle AB(x, t). By [6, §3.2, Theorem 21], we see that N− is contained in M−. On
the other hand, since almost every t → AB(x, t)ψ(x + et) lies in H 2(dt/π(1 + t2)) for
each ψ in M[φ], N+ includes M[φ]. This shows that N+ = M+, so B(x, t) ≡ 1. Then
almost every t → A(x, t)φ(x + et) is outer in H 2(dt/π(1 + t2)).
Conversely, suppose that M[φ] is contained strictly in M−. Then there is a nonzero
function q in M− such that
ZK
ψφq dσ = 0 ,
ψ ∈ H ∞(σ) .
This shows that φq lies in H 1(σ), so almost every t → φq(x+et) lies in H 1(dt/π(1+t2)).
Notice that t → A(x, t)q(x + et) is in H 2(dt/π(1 + t2)). Since
φ(x + et)q(x + et) = A(x, t)φ(x + et)A(x, t)q(x + et),
and since t → A(x, t)φ(x + et) is outer in H 2(dt/π(1 + t2)), we see that almost every
t → A(x, t)q(x + et) is also in H 2(dt/π(1 + t2)). This shows that t → A(x, t)q(x + et) is
constant for σ−a.e. x in K, and so is t → q(x+et). It follows from the ergodic theorem
that q(x) is constant. We then assume q is a unitary function on K. Therefore, A(x, t)
is the coboundary q(x)q(x + et) and M− = qH 2
0 (σ). Thus q does not lie in M−, which
is a contradiction.
The last part of assertion follows from the fact that the cocycle of H 2(σ) equals 1.
Under the assumption that almost every t → φ(x + et) is outer in H 2(dt/π(1 + t2)), we
see easily a0(φ) = 0 if and only if log φ does not lie in L1(σ). Then M[φ] = H 2
0 (σ),
so the proof is complete.
(cid:3)
Let L1(dt) be the usual Lebesgue space on R. Using {Tt}t∈R, one may convolve a
function φ in Lp(σ), 1 ≤ p < ∞, with a function f in L1(dt) by setting
φ(x − et)f (t) dt ,
(φ ∗ f )(x) = Z ∞
−∞
φ(x + et)f (−t) dt = Z ∞
−∞
5
where the integral is a Bochner integral. When p = ∞, the convolution φ ∗ f is defined
in the same way as the weak*-convergent integral. Under the operation of convolution,
Lp(σ) becomes an L1(dt)-module such that
kφ ∗ f kp ≤ kφkpkf k1 ,
φ ∈ Lp(σ) ,
for f in L1(dt). The Fourier transform f of f is defined by the formula
(4)
f (λ) = Z ∞
−∞
f (t)e−iλt dt,
λ ∈ R ,
as usual. We see easily aλ(φ∗f ) = aλ(φ) f (λ), if λ is in Γ. The Poisson kernel Pir(t) for
H is given by Pir(t) = r/π(t2 + r2), for an r > 0. If φ is in L1(σ), then the convolution
φ ∗ Pir is considered as the Poisson integral of t → φ(x + et), that is,
(φ ∗ Pir)(x + es) = Z ∞
−∞
φ(x + et)Pir(s − t) dt .
Lemma 2. Suppose that H 2
properties:
0 (σ) is singly generated. Then we obtain the following
(a) There is a single generator of H 2
(b) If φ is a bounded generator of H 2
with r > 0 and φn for n = 1, 2, · · · .
0 (σ) that is bounded.
0 (σ), then so is each of the functions φ ∗ Pir
0 (σ). Then there is an outer function h in H 2(σ)
Proof. Let ψ be a single generator of H 2
such that h = min(1, ψ−1). It follows from Lemma 1 that the bounded function ψh
generates H 2
0 (σ), thus we obtain (a).
To show (b), we observe that t → (φ ∗ Pir)(x + et) as well as t → φn(x + et) is outer
in H 2(dt/π(1 + t2)) for σ − a.e. x in K. Since a0(φ ∗ Pir) = a0(φn) = 0, (b) follows from
Lemma 1 immediately.
(cid:3)
We next introduce a local product decomposition of K, which is useful for studying
analytic functions on K. Fix a positive γ in Γ, and let Kγ be the closed subgroup of
all x in K such that χγ(x) = 1. Then Kγ × [0, 2π/γ) is identified with K via the map
(y, s) → y + es. Let σ1 be the normalized Haar measure on Kγ. Then the probability
measure (γ/2π)dσ1 × dt on Kγ × [0, 2π/γ) is carried by the map to dσ on K. The
one-parameter group {Tt}t∈R given by (2) is represented as
Tt(y, s) = (y + [(t + s)γ/2π]e2π/γ, t + s − [(t + s)γ/2π]2π/γ)
on Kγ × [0, 2π/γ), where [t] is the largest integer not exceeding t. Define the home-
omorphism T on Kγ by T y = y + e2π/γ. We denote by O(ω, T ) the orbit of a point
ω in (Kγ, T ), that is, the set of all T nω for n in Z. Since O(ω, T ) is dense in Kγ, the
discrete flow (Kγ, T ) is also a strictly ergodic flow, on which σ1 is the unique invariant
probability measure. Since Γ is countable, Kγ is metrizable (see [16, 2.2.6]).
6
A function φ on K has the automorphic extension φ♯ to Kγ × R defined by
φ♯(y, t) = φ(y + [tγ/2π]e2π/γ , t − [tγ/2π]2π/γ).
Since a function f in H 1(dt/π(1 + t2)) extends analytically to H by f (s + ir) =
(f ∗ Pir)(s), we write
φ♯(y, z) = (φ♯ ∗ Pir)(y, s),
z = s + ir ∈ H ,
for each φ in H 1(σ). It is clear that (φ♯ ∗ Pir)(y, s) = (φ ∗ Pir)♯(y, s) on Kγ × R.
The following is due to a property of Lebesgue sets.
Lemma 3. If E1 is a compact subset of Kγ with σ1(E1) > 0, then there is a closed
subset E of E1 with σ1(E1) = σ1(E) such that O(ω, T )∩E is dense in E, for σ1−a.e. ω
in Kγ.
Proof. Recall that the metric density of E1 is 1 at σ1 − a.e. ω in E1, meaning that
lim
δ→0
σ1(E1 ∩ B(ω, δ))
σ1(B(ω, δ))
= 1 ,
where B(ω, δ) is the open ball with center ω and radius δ > 0. Define E to be the
closure of the set of points of E1 at which the metric density of E1 is 1. Clearly, we
have σ1(E1) = σ1(E), since E1 is closed. If σ1(E) = 1, then E = Kγ. Since (Kγ, T )
is strictly ergodic every orbit O(ω, T ) is dense in E. Assume that 0 < σ1(E) < 1. It
follows from the ergodic theorem that there is a σ1−null set N in Kγ outside which
lim
n→∞
1
n
n−1Xj=0
IE(T jω) = σ1(E) ,
where IE denotes the characteristic function of E. Let Hω be the closure of O(ω, T ) ∩E
in Kγ. We claim that if E 6= Hω, then ω lies in N. Indeed, we see that σ1(E \Hω) > 0,
since the metric density of E does not vanish identically on E\Hω. Let p be a continuous
p dσ < σ1(E). Since
function on Kγ such that 0 ≤ p ≤ 1, p ≡ 1 on Hω, and RKγ
p(T jω) = ZKγ
IE(T jω) ≤ lim
n→∞
IE(T jω) = IHω (T jω) for j in Z and since (Kγ, T ) is strictly ergodic, we have
p dσ1 < σ1(E)
1
n
n−1Xj=0
lim sup
n→∞
1
n
n−1Xj=0
by [15, §4.2, Proposition 2.8]. Thus ω has to lie in the null set N.
(cid:3)
For each φ in H ∞(σ), there is a σ1-null set of Kγ outside which z → φ♯(y, z) is
analytic and uniformly bounded on the upper half plane H. Recall that if a family
of analytic functions is uniformly bounded, then it forms a normal family. The next
proposition may be regarded as a strengthened form of Lusin's theorem for analytic
functions on K, so that it has some interest of its own. Here we denote by cl(H) the
closure of H in R2.
7
Proposition 1. Let φ be a function in H ∞(σ), and let ǫ > 0. Then there is a closed
subset E of Kγ with σ1(E) > 1 − ǫ having the following properties:
(a) The convolution (φ♯ ∗ Pir)(y, t) is continuous on E × R, for a given r > 0.
(b) For σ1 − a.e. ω in Kγ, the function (φ♯ ∗ Pir)(T jω, z) on (O(ω, T ) ∩ E) × cl(H)
extends to (φ♯ ∗ Pir)(y, z) on E × cl(H).
Proof. Since φ ∗ Pir lies in H ∞(σ), Lusin's theorem asserts that there is a compact
subset F of K with σ(F ) > 1 − ǫ2 on which φ ∗ Pir is continuous. Regarding F as
a subset of Kγ × [0, 2π/γ), we choose a compact subset E of Kγ with σ1(E) > 1 − ǫ
such that E satisfies the property of Lemma 3 and
(5)
γ
2πZ 2π/γ
0
IF (y, s) ds > 1 − ǫ ,
y ∈ E .
In addition, we assume that z → (φ♯ ∗ Pir/2)(y, z) is analytic on H and
(φ♯ ∗ Pir/2)(y, z) ≤ kφk∞ ,
y ∈ E .
Then the family
F = (cid:8) (φ♯ ∗ Pir/2)(y, z) ; y ∈ E(cid:9)
forms a normal family on H. Let {yn} be a sequence in E tending to y. Since F
is normal, there is a subsequence {yj} of {yn} such that (φ♯ ∗ Pir/2)(yj, z) converges
uniformly on compact subsets of H to a bounded analytic function f (z) on H. Let us
show that f (z) = (φ♯ ∗Pir/2)(y, z). Indeed, we observe by (5) that F ∩ ({y}×[0, 2π/γ))
contains an infinite compact set of the form {y} × J. Since
(φ♯ ∗ Pir)(y, t) = (φ♯ ∗ Pir/2)(y, t + ir/2) = f (t + ir/2),
t ∈ J ,
it follows from the uniqueness principle that f (z) = (φ♯ ∗ Pir/2)(y, z). This shows that
if (yn, tn) tends to (y, t), then (φ♯ ∗ Pir)(yn, tn) tends to (φ♯ ∗ Pir)(y, t). Thus (a) holds.
We notice that (φ♯ ∗ Pir/2)(y, z) is also continuous on E × H.
On the other hand, by Lemma 3, O(ω, T ) ∩ E is dense in E for σ1 − a.e. ω in Kγ.
Since (O(ω, T ) ∩ E) × cl(H) is dense in E × cl(H) and since (φ♯ ∗ Pir)(y, z) is continuous
on E × cl(H), the function (φ♯ ∗ Pir)(T jω, t) on (O(ω, T ) ∩ E) × cl(H) extends to
(φ♯ ∗ Pir)(y, t) on E × H. Thus (b) follows immediately.
(cid:3)
We make some remarks on Proposition 1. Since t → φ♯(y, t) lies in H ∞(dt/π(1 + t2))
for each y in E, we see that (φ♯ ∗ Pir)(y, t + 2π/γ) = (φ♯ ∗ Pir)(T y, t). Then E ∪ T E ∪
· · ·∪T nE also satisfies the properties (a) and (b) and σ1(E ∪T E ∪· · · ∪T nE) converges
to 1, as n → ∞, by the recurrence theorem (see [15, §2.3, Theorem 3.2]). However, to
obtain φ itself, we need a version of Fatou's theorem as discussed in [14, Theorem II].
Denote by O(x, {Tt}t∈R) the orbit of x in (K, {Tt}t∈R). With the notation above, when
x = (y, s) in Kγ × [0, 2π/γ), we see that O(x, {Tt}t∈R) = O(y, T ) × [0, 2π/γ). For x
in K, we say that t → (φ ∗ Pir)(x + et) extends to φ ∗ Pir if, for each ǫ > 0, there is a
compact subset F = F (ǫ, φ) of K with σ(F ) > 1 − ǫ such that φ ∗ Pir is continuous
8
on F and O(x, {Tt}t∈R) ∩ F is dense in F . The above proof may be modified so as to
apply to functions in H 1(σ) as well.
The next lemma is an immediate consequence of Proposition 1.
Lemma 4. Let φ be a function in H ∞(σ), and let r > 0. Then there is an invariant
σ−null set N = N(φ) in K outside which t → (φ ∗ Pir)(x + et) extends to φ ∗ Pir.
Proof. For a given ǫ > 0, let E be a closed subset of Kγ with σ1(E) > 1 − ǫ which
has the property (a) and (b) of Proposition 1. Putting F = E × [0, 2π/γ], we regard
F as a compact subset of K. By (b) of Proposition 1, we choose an invariant null
set N ′ = N ′(φ) in (Kγ, T ) outside which O(ω, T ) ∩ E is dense in E. If we set N =
N ′ × [0, 2π/γ), then the σ−null set N satisfies the desired property.
(cid:3)
Let Ω be a compact metric space on which R acts as a Borel transformation group.
This means that there is a one-parameter group {Ut}t∈R of Borel isomorphisms on Ω
such that the map (ω, t) → Utω of Ω × R to Ω is a Borel map. The pair (Ω, {Ut}t∈R)
is referred to a Borel flow. Especially, (Ω, {Ut}t∈R) is called a continuous flow, if Ut is
a homeomorphism on Ω and the map (ω, t) → Utω is continuous on Ω × R. We often
write ω + t for the translate Utω of ω by t. Let µ be an invariant probability measure
on (Ω, {Ut}t∈R) which is ergodic, meaning that µ(E) = 1 or 0 for each invariant subset
E of Ω. A function φ in L1(µ) is analytic if t → φ(ω + t) lies in H 1(dt/π(1 + t2))
for µ − a.e. ω in Ω. Then the ergodic Hardy space H p(µ), 1 ≤ p ≤ ∞, is defined to
be the space of all analytic functions in Lp(µ). It follows from [13, Theorem I] that
µ is a representing measure for H ∞(µ), for which H ∞(µ) is a weak*-Dirichlet algebra
in L∞(µ). This fundamental result enables us to apply the Hardy space theory based
on uniform algebras, and most of the machinery of invariant subspaces on an almost
periodic flow (K, {Tt}t∈R) can be reconstructed (see [2], [13] and [14] for related topics).
As we mentioned earlier, the H 2
0 −spaces may be singly generated in the situation of
ergodic flows other than almost periodic flows (see [18] and §5 (b)).
Let A(x, t) be a cocycle on an almost periodic flow (K, {Tt}t∈R) and define the Borel
flow (K × T, {St}t∈R) by
(6)
St(x, eiθ) = (Ttx, A(x, t)eiθ),
(x, eiθ) ∈ K × T,
which is called the skew product of K and T induced by A(x, t). Then dσ × dθ/2π is an
invariant probability measure on K ×T. Observe that each function f in L2(dσ×dθ/2π)
is represented as
f (x, eiθ) =
∞Xn=−∞
φn(x)einθ ,
where the coefficients φn are in L2(σ). From this fact, it follows easily that dσ × dθ/2π
is ergodic on (K × T, {St}t∈R) if and only if A(x, t)n is a coboundary only for n = 0
(see [6, §6.2] for details).
9
3. Approximation to generators
We now turn to the structure of compact group K, under the assumption that
H 2
0 (σ) is singly generated by φ in H 2
0 (σ). By multiplying by a suitable outer function,
if necessary, we can assume that φ is a function in L∞(σ) with 1 ≤ kφk∞ < +∞.
Furthermore, we also assume that Γ is the smallest group containing all λ such that
aλ(φ) 6= 0, that is, the smallest group over which Fourier series,
φ(x) ∼ XΓ∋λ>0
aλ(φ)χλ(x) ,
holds. Similarly, denote by Λ the smallest group containing all λ such that aλ(φ) 6= 0.
We observe that the Fourier series of
(cid:0)φ2 + ǫ(cid:1)1/2 = exp(cid:26) 1
2
log (φφ + ǫ)(cid:27) ,
ǫ > 0 ,
is represented on Γ, by considering the Taylor series of z → log z at a large positive.
This shows that Λ is a subgroup of Γ, since
aλ(φ) = lim
ǫ→0
aλ(cid:0)(φ2 + ǫ)1/2(cid:1)
by (1). Since log φ does not lie in L1(σ), the generator φ cannot be periodic in
(K, {Tt}t∈R). Then Γ as well as Λ is a countable dense subgroup of R, endowed with
discrete topology. Let H be the annihilator of Λ, meaning that H is the closed subgroup
of all x in K such that χλ(x) = 1 for all λ in Λ. Then the dual group of Λ is identified
with the quotient group K/H (see [16, 2.1]). We denote by τ the normalized Haar
measure on K/H. Let π be the canonical homomorphism of K onto K/H. For each x
in K, we write ¯x for π(x) = x+H. When a function ψ on K is represented as ψ = ψ◦π
for a function ψ on K/H, we usually identify ψ with ψ, so that ψ(x) = ψ(¯x). Then we
say descriptively that ψ is generated by a function on K/H. If 1 ≤ p ≤ ∞, then Lp(τ )
and H p(τ ) are subspaces of Lp(σ) and H p(σ), respectively.
Since almost every t → φ(x + et) is outer in H ∞(dt/π(1 + t2)) by Lemma 1, we see
that
−∞ < log (φ ∗ Pir)(x) = (log φ ∗ Pir)(x)
for a given r > 0. Since log φ is not in L1(σ) and log φ ≤ kφk∞, Fubini's theorem
shows that
ZK
log φ ∗ Pir dσ = ZK
(log φ ∗ Pir) dσ = ZK
log φ dσ = −∞ .
Let g = φ ∗ Pir. Then Lemma 1 shows that g is also a bounded generator of H 2
0 (σ).
Since Pir(λ) = e−rλ by (4), we obtain aλ(g) = aλ(φ ∗ Pir) = aλ(φ)e−rλ, hence
aλ(φ) 6= 0 if and only if aλ(g) 6= 0. Thus the generator g plays the same role
as φ. For n = 1, 2, . . . , we then denote by φn the outer function in H ∞(σ) with
φn = max(1/n, φ). Since − log n ≤ log φn ≤ kφk∞, each φ−1
is also an outer
n
function in H ∞(τ ). Putting gn = φn ∗Pir, we obtain a sequence {gn} of outer functions
10
in H ∞(τ ) with kgnk∞ ≤ kφk∞. Notice that t → g(x + et) and t → gn(x + et) extend
analytically up to {Re z > −r}. Let us look into the relation between g and gn. Since
gn(x) = exp{(log φn ∗ Pir)(x)} ,
we obtain
g1(x) ≥ g2(x) ≥ · · · ≥ gn(x) −→ g(x) ,
(7)
for σ −a.e. x in K. Although g may not be in L∞(τ ), we observe that gn(x) = gn(¯x)
and g(x) = g(¯x). By (7), it is easy to see that almost every t → (g/gn)(x + et)
n(t) = gn(x+et) and Gx(t) = g(x+et). Let N0 be
converges pointwise to 1 on R. Put Gx
an invariant null set in K outside which the property of Lemma 4 holds simultaneously
for φ and all φn. Moreover, for x in K \ N0, we may assume Gx
n(t) and Gx(t) are outer
functions in H ∞(dt/π(1 + t2)). Then the family of all analytic extensions Gx
n(z) of
Gx
n → ∞ ,
n(t) to {Re z > −r} forms a normal family, since Gx
The following lemma is crucial in our proof of the Theorem.
n(z) ≤ kφk∞.
Lemma 5. For a bounded generator φ of H 2
0 (σ), let Λ, H and τ be as above. Choose
an α in Γ with aα(φ) 6= 0. Then χαφ is generated by a function on K/H, so lies in
L∞(τ ). Consequently, Γ is generated by Λ and α.
Proof. Let {δk} be a decreasing sequence tending to 0. Then there is a sequence {fk}
in L1(dt) such that fk(α) = 1, kfkk1 = 1 and fk = 0 outside (α − δk, α + δk), by
modifying the function t → (1/π) sin2 t/t2 in L1(dt). Since aλ(g) = aλ(φ)e−rλ, we
see that χαφ lies in L2(τ ) if and only if so does χαg. Thus we may replace φ with g in
our argument. Since aλ(g ∗ fk) = aλ(g) fk(λ), we observe that
kg ∗ fk − aα(g)χαk2
aα+λ(g) fk(α + λ)2 → 0 ,
k → ∞ ,
2 = X0<λ<δk
by the Parseval theorem and that
k(g ∗ fk)g − aα(g)(χαg)k2 ≤ kg ∗ fk − aα(g)χαk2 kgk∞ .
From these facts, we conclude that if each (g ∗ fk)g lies in L∞(τ ), then so does χαg.
Since the outer function φn lies in L∞(τ ), so do gn and gn ∗ fk. Then each (gn ∗ fk)gn
lies in L∞(τ ). Let us show that the sequence {(gn ∗ fk)gn} converges to {(g ∗ fk)g}
in L2(σ), from which we obtain that (g ∗ fk)g lies in L∞(τ ). Indeed, in the notation
above, if we fix an x in K \ N0, there is a subsequence {gm} of {gn} such that {Gx
m(t)}
converges pointwise to eiγGx(t) in H ∞(dt/π(1 + t2)) with 0 ≤ γ < 2π, where γ depends
on x and {gm}. This implies that
(gm ∗ fk)(x + et) → e−iγ(g ∗ fk)(x + et),
m → ∞,
pointwise in L∞(dt/π(1 + t2)). Note that every subsequence of {gn} contains such a
subsequence {gm}. Since e−iγeiγ = 1, the sequence {gn} itself satisfies
(gn ∗ fk) gn(x + et) → (g ∗ fk) g(x + et) ,
n → ∞,
11
pointwise in L∞(dt/π(1 + t2)). Since
k(gn ∗ fk) gnk∞ ≤ kgnk2
∞kfkk1 ≤ kφk2
∞kfkk1,
it follows from the bounded convergence theorem that
k(gn ∗ fk) gn − (g ∗ fk) gk2 → 0 ,
n → ∞,
so that (g ∗ fk) g lies in L∞(τ ). Therefore, χα g as well as χα φ is generated by a
function on K/H. On the other hand, by the property of Γ, each element in Γ has the
form λ + nα for λ in Λ and n in Z, thus the proof is complete.
(cid:3)
Recall that K/H coincides with the dual group of Λ. Let α be as in Lemma 5 and
let C(¯x, t) be the trivial cocycle on K/H defined by C(¯x, t) = exp(iαt). Since α is
positive, C(¯x, t) is an analytic cocycle. We denote by (K/H × T, {St}t∈R) the skew
product of K/H and T induced by C(¯x, t), which is the continuous flow obtained by
St(¯x, eiθ) = (Tt ¯x, C(¯x, t)eiθ),
(¯x, eiθ) ∈ K/H × T .
Then dτ × dθ/2π is the invariant probability measure on K/H × T (see the end of the
preceding section). Let us represent the generator g and all the limits of subsequences
of {gn} on K/H × T, which is the smallest product group with such property. Each
function ψ on K/H extends naturally to the one on K/H × T by setting ψ(¯x, eiθ) =
ψ(¯x). Since g and gn are functions on K/H, they belong to L∞(dτ × dθ/2π).
With the above notation, we fix a w in K \ N0. Since Gw
n (t) and Gw(t) are outer
functions in H 2(dt/π(1 + t2)) which extend analytically to {Re z > −r}, we may
n (t) converges pointwise to Gw(t) on R, by multiplying each gn by a
assume that Gw
suitable constant of modulus one. By regarding Lemma 4, the functions Gw
n (t) and
Gw(t) extend to gn and g, respectively. However, we obtain the following:
Lemma 6. For σ − a.e. x in K, Gx
we find two subsequences {gm} and {gk} of {gn} such that Gx
to eiβGx(t) and eiγGx(t) with 0 ≤ β < γ < 2π, respectively.
n(t) never converges pointwise on R. Consequently,
k(t) converge
m(t) and Gx
Proof. Since 1/n ≤ gn(x) ≤ kφk∞, each g−1
n is also an outer function in H ∞(σ). This
implies that almost every t → (g/gn)(x + et) is an outer function in H ∞(dt/π(1 + t2)).
Furthermore, since
a0(g/gn) = ZK
g−1
n dσ = 0 ,
g/gn dσ = ZK
g dσZK
Lemma 1 assures that each g/gn is also a single generator of H 2
Denote by F the invariant set of all x in K for which {Gx
0 (σ).
n(t)} itself converges.
Suppose that F has positive measure. By (7) and the ergodic theorem, (g/gn)(x)
converges to an invariant function on F , so to a constant of modulus one on K. Then
the bounded convergence theorem shows that a0(g/gn) 6= 0 for large n. Such g/gn
cannot be a single generator of H 2
(cid:3)
0 (σ), which contradicts the above observation.
12
Let us mention a few remarks derived from Lemma 6. When 0 ≤ β < 2π, Z(β)
denotes the subgroup of T generated by eiβ, that is,
Z(β) = (cid:8)eijβ ; j ∈ Z(cid:9) .
If β/2π is rational, then the order of Z(β) is finite. Fix two points w and x in K \ N0.
We assume by Lemma 6 that a subsequence {gk} of {gn} satisfies that Gw
k(t)
converge respectively to eijβGw(t) and ei(j+1)βGx(t) for j in Z, by multiplying each gk
by a suitable constant of modulus one. Denote by O( ¯w) the orbit O( ¯w, {Tt}t∈R) of ¯w
in (K/H, {Tt}t∈R). Then g is determined naturally on O( ¯w) × Z(β) and O(¯x) × Z(β)
to represent the limits of the subsequence {gk} of {gn} on them. For each m in Z, we
see also that every limit of {gm
k } is represented on these product subsets.
k (t) and Gx
k(t)} of {Gx
j (t)ℓ} of {Gx
If ℓ is a positive integer, then gℓ as well as φℓ is also a bounded generator of H 2
0 (σ)
by Lemma 2. We choose an invariant null set N(ℓ) including N0 outside which a
n(t)ℓ} converges to eiγGx(t)ℓ with 0 < γ < 2π. Define the
subsequence {Gx
invariant null set N1 by N1 = ∪∞
ℓ=1 N(ℓ) . When ℓ = m !, we take again a subsequence
j (t)} converging to eiβ(m)Gx(t) with eiβ(m)ℓ = eiγ. Then the order of
{Gx
Z(β(m)) is larger than m, so ∪∞
m=1 Z(β(m)) is dense in T. Therefore, to represent g
and all the limits of subsequences of {gn} on each orbit, the product group K/H × T
is the smallest one. Let us explain the meaning more precisely. Under the assumption
of Lemma 5, we put hα = χαg. Then hα lies in L2(τ ). Define the group character Pα
of K/H × T by the projection Pα(¯x, eiθ) = eiθ. Since
(hαPα)(St(¯x, eiθ)) = hα(¯x + et)C(¯x, t)eiθ = hα(¯x + et)eiαteiθ,
the function t → (hαPα)(St(¯x, eiθ)) is an outer function in H ∞(dt/π(1 + t2)) for
dτ × dθ/2π − a.e. (¯x, eiθ) in K/H × T. Then the outer function Gx(t) equals t →
(hαPα)(St(¯x, eiθ)) for some θ with 0 ≤ θ < 2π. In order to represent consistently all
kinds of limits of subsequences {Gx
k(t)}, we require the family of all outer functions
t → (hαPα)(St(¯x, eiθ)) with 0 ≤ θ < 2π.
Lemma 7. Let Γ and Λ be as above. Then Λ cannot be equal to Γ.
Proof. Let α be as in Lemma 5. Then α lies in Λ if and only if Λ = Γ. We suppose,
on the contrary, that α lies in Λ. Since K/H = K, let us consider the skew product
(K × T, {St}t∈R) of K and T induced by the cocycle C(x, t) = eiαt. We use freely the
notation above. Since
F (x, eiθ) = (χα Pα)(x, eiθ) ,
(x, eiθ) ∈ K × T ,
is an invariant function that is not constant, dσ × dθ/2π is not an ergodic measure
on (K × T, {St}t∈R). Now K is represented as the local product decomposition Kα ×
[ 0, 2π/α), in which Kα is the closed subgroup of all x in K such that χα(x) = 1. If
we put
G(x, eiθ) = hα(x) Pα(x, eiθ) ,
(x, eiθ) ∈ K × T,
13
then, for each x = (y, s) in Kα × [0, 2π/α), the equation
(8)
G(St(x, eiθ)) = ei(θ+αt)hα(x + et) = ei(θ−αs)g(x + et)
holds, since ei(θ+αt)χα(y + es + et) = ei(θ−αs) and hα = χαg. By regarding T as the
interval [ 0, 2π/α), K × T is identified with Kα × [ 0, 2π/α) × [ 0, 2π/α). Let E be the
subset of K × T defined by
E = Kα × {(s, s) ; 0 ≤ s < 2π/α}.
Then E is a closed invariant set in (K ×T, {St}t∈R), for which (K, {Tt}t∈R) is isomorphic
to (E, {St}t∈R) via the map (y, s) → (y, s, s). We see also that the ergodic measure
dσ is carried to (α/2π)dσ1 × ds on E by this map, where σ1 is the normalized Haar
measure on Kα. We regard gn, g and hα as the functions on (K × T, {St}t∈R). Recall
that almost every Gx
n(t) and Gx(t) are outer functions in H ∞(dt/π(1 + t2)).
Let x be in K \ N1 and let {gk} be a subsequence of {gn} such that Gx
k(t) converges
pointwise to t → eiαβeiαthα(x + et) with 0 ≤ β < 2π/α. Notice that t → eiαβeiαthα(x +
et) is an outer function in H ∞(dt/π(1 + t2)) and that hα(x + et) = g(x + et). Let
x = (y, s) in Kα × [ 0, 2π/α) as above. Since x may be replaced by any point in the
orbit O(x) of x, we consider x as a function of s on [ 0, 2π/α). It follows from (8) that
eiαβeiαthα(y + es + et) = eiα(β−s)G(St(y + es, eiαs)) ,
(s, t) ∈ [ 0, 2π/α) × R.
Putting t = 0 and replacing y with y + e[sα/2π], if necessary, we observe that
eiα(β−s)G(y + es, eiαs) = eiαβe−iαsGy(s) ,
s ∈ R.
This shows that Gy
k(s) converges pointwise to s → eiαβ(χα g)(y + es), which cannot
be an outer function in H ∞(dt/π(1 + t2)). Hence any subsequence of {Gx
n(t)} cannot
converge to an outer function in H ∞(dt/π(1 + t2)) for σ − a.e. x in K. Thus we have
a contradiction.
(cid:3)
In view of Lemma 7, we know that there are two possibilities in relation to α and
Λ. Either nα lies in Λ only for n = 0 or ℓα lies in Λ for an integer ℓ ≥ 2. We claim
that the latter case cannot occur, meaning that α is independent to Λ.
Lemma 8. Let Λ, H and α be as above. Then nα lies in Λ if and only if n = 0 in Z.
Consequently, H is isomorphic to T, so that K and dσ are identified with K/H × T
and dτ × dθ/2π, respectively.
Proof. Suppose that ℓα lies in Λ for some ℓ ≥ 2. By Lemma 2, φℓ is also a bounded
generator of H 2
0 (σ). It follows from Lemma 6 that χℓα and (χαφ)ℓ lie in L2(τ ), so does
φℓ itself. Let Γℓ and Λℓ be the smallest groups determined by the nonzero Fourier
coefficients of φℓ and φℓ as above. Then they both are subgroups of Λ. On the other
hand, since
aλ(φ) = lim
ǫ→+0
aλ(cid:0)(φℓ + ǫ)1/ℓ(cid:1) ,
14
each λ in Λ with aλ(φ) 6= 0 lies in Λℓ. This implies that Λ = Λℓ = Γℓ. By replacing
φ with φℓ in Lemma 7, this gives a contradiction. Thus nα lies in Λ if and only if
n = 0.
Since C(¯x, t)n is a coboundary only for n = 0, the measure dτ × dθ/2π is ergodic on
(K/H × T, {St}t∈R). Define the isomorphism of Λ × Z onto Γ by
(λ, n) = λ + nα ,
(λ, n) ∈ Λ × Z.
Then the conjugate map ∗ of is given by ∗(x) = (¯x, eiθ) on K, where χα(x) = eiθ.
Indeed, we observe that
χλ(¯x)einθ = h(λ, n), (¯x, eiθ)i = χλ+nα(x) = χλ(¯x)χα(x)n ,
for each (λ, n) in Λ × Z. Via the map ∗ , K is identified with K/H × T, and dτ ×dθ/2π
is carried by the map to dσ on K.
(cid:3)
We notice that the annihilator H of Λ is isomorphic to T, and g(x) as well as φ(x)
is constant on almost every coset ¯x = x + H in K/H.
4. Contradiction to existence
We may now offer our proof of the main result stated in Section 1.
Proof of the Theorem. Suppose, on the contrary, that a bounded function φ generates
H 2
0 (σ). Let Γ and Λ be the dense subgroups of R defined as in Section 3 with respect
to φ and φ, respectively. Choose an α in Γ with aα(φ) 6= 0. It follows from Lemma 8
that α is independent of Λ and Γ is generated by α and Λ. Let 0 < β < 1. Since the
function
(1 + βχα)−1 =
(−β)kχkα
∞Xk=0
lies in H ∞(σ), (1 + βχα)2 is an outer function in H ∞(σ). Define φ1 = (1 + βχα)2 φ .
In view of Lemma 1, φ1 is also a bounded generator of H 2
0 (σ). As above, let Γ1 and Λ1
be the smallest groups determined by the nonzero Fourier coefficients of φ1 and φ1,
respectively. Notice that Γ1 is a subgroup of Γ. We claim that the generator φ1 cannot
satisfy the property of Lemma 7. Indeed, since φ1 = (1 + β2 + βχα + βχα) φ , we
obtain by (1) that
aλ(φ1) = (1 + β2)aλ(φ) + βaλ+α(φ) + βaλ−α(φ).
Since α does not lie in Λ, if λ is in Λ, then aλ+α(φ) = aλ−α(φ) = 0. Then we have
aλ(φ1) = (1 + β2)aλ(φ) and aλ+α(φ1) = βaλ(φ),
for each λ in Λ. These facts imply that Λ1 contains Λ and α, so that Γ = Λ1 = Γ1,
which contradicts Lemma 7.
(cid:3)
The next proof is of independent interest, because it suggests that our Theorem is
regarded essentially as the converse to Corollary 1.
15
Proof of Corollary 1. We consider the case where the cocycle C(x, t) of M has the
form C(x, t) = eiαt. Then M− is the space of all ψ in L2(σ) satisfying that
ψ(x) ∼ XΓ∋λ > −α
aλ(ψ)χλ(x) .
Suppose that M− has a generator φ. Then log φ does not lie in L1(σ) and we may
assume that φ is bounded. If ℓα is in Γ for a positive integer ℓ, then the bounded func-
tion (χαφ)ℓ is a single generator of H 2
0 (σ) by Lemma 1, which is contrary to Theorem.
We next consider the case that
α ∈ R \
∞[n=1
(1/n)Γ .
Since C(x, t)n is a coboundary only for n = 0, the measure dσ × dθ/2π is ergodic on
the skew product (K × T, {St}t∈R) induced by C(y, t), that is,
St(x, eiθ) = (x + et , eiαteiθ),
(x, eiθ) ∈ K × T.
Let Γ1 be the discrete group generated by Γ and α, and let K1 be the dual group of
Γ1. Since (λ, n) = λ + αn is an isomorphism of Γ × Z onto Γ1, the almost periodic
flow on K1 is identified with (K × T, {St}t∈R). Then, via the dual map ∗ of , the
normalized Haar measure dµ on K1 is identified with dσ × dθ/2π. Define the function
φ1 in L2(µ) by φ1(x, eiθ) = φ(x)eiθ. Since log φ does not lie in L1(σ), neither does
log φ1 in L1(µ). Since t → φ1 ◦ St(x, eiθ) is outer in H 2(dt/π(1 + t2)) for µ −a.e. (x, eiθ)
in K × T, Lemma 1 implies that φ1 is a single generator of H 2
0 (µ), which contradicts
our Theorem.
(cid:3)
Proof of Corollary 2. Denote by C(x, t) the real cocycle of M. Suppose that M− has
a generator φ, for which log φ does not lie in L1(σ). It follows from Lemma 1 that
almost every t → C(x, t)φ(x + et) is outer in H 2(dt/π(1 + t2)). We may assume that φ
is bounded. Since C(x, t)2 ≡ 1, φ2 is a single generator of H 2
0 (σ) by Lemma 1, which
contradicts our Theorem.
(cid:3)
By the same way as above, we may show that if C(x, t) takes only finite values, then
M− cannot be singly generated. Indeed, by the cocycle identity, the set of values of
C(x, t) forms a group of order k,
Then if φ generates M−, then φk is a generator of H 2
0 (σ).
Z(2π/k) = (cid:8)ei2πj/k ; j = 0, . . . , k − 1(cid:9) .
Let M be the normalized simply invariant subspace of L2(σ) with cocycle A(x, t).
Recall that ψ lies in M if and only if almost every t → A(x, t)ψ(x + et) lies in
in [6, §3.2]). To prove Corollary 3. we need the following:
H 2(dt/π(1+t2)). Denote byfM the invariant subspace with cocycle A(x, t) (as discussed
Lemma 9. Let M and fM be as above. If M is singly generated, then (fM)− cannot be
singly generated.
16
Proof. Since A(x, t) · A(x, t) ≡ 1, H 2
0 (σ) is the smallest subspace of L2(σ) containing all
ψ1ψ2 with ψ1 in M∩L∞(σ) and ψ2 in (fM)−∩L∞(σ) (see [6, §3.2, Theorem 20]). Suppose
that (fM)− is singly generated. Then Lemma 1 shows that there are bounded single
generators φ1 and φ2 of M and (fM)−, respectively. Thus φ1φ2 is a single generator of
0 (σ), which contradicts our Theorem.
H 2
(cid:3)
Proof of Corollary 3. (a) Let M be a simply invariant subspace with nontrivial cocycle
A(x, t). It follows from [11] that M is singly generated if and only if A(x, t) is coho-
mologous to a singular cocycle. On the other hand, by [6, §4.6, Theorem 26], every
cocycle is cohomologous to a Blaschke cocycle. By virtue of Lemma 9, we obtain easily
a desired Blaschke cocycle.
(b) From Lemma 9, we choose a Blaschke cocycle B(x, t) such that the invariant
subspace N having the cocycle B(x, t) is not singly generated. We claim that B(x, t)
satisfies the desired property. Suppose, on the contrary, that some function ψ in H 2(σ)
has exactly the same zeros as B(x, t). By multiplying by a suitable outer function,
we assume that ψ is bounded. Then ψ generates the invariant subspace with cocycle
B(x, t)S(x, t), where S(x, t) is the singular cocycle determined by the inner part of
t → B(x, t)ψ(x + et) in H 2(dt/π(1 + t2)). On the other hand, it follows from [11] and
Lemma 1 that there is a function h in L2(σ) such that almost every t → S(x, t)h(x+et)
is outer in H 2(dt/π(1 + t2)). Observe that
(hψ)(x + et) = B(x, t) · S(x, t)h(x + et) · B(x, t)S(x, t)ψ(x + et) .
Since the inner part of t → (hψ)(x + et) is t → B(x, t), the subspace N is singly
generated by hψ, thus we have a contradiction.
(cid:3)
In the proof of (b) above, if the singular cocycle S(x, t) is a coboundary, then h is
taken as a unitary function, otherwise log h does not lie in L1(σ).
(a)
It is sometimes useful to study the spectral measures associated with invariant
subspaces. Let M be a simply invariant subspace of L2(σ) and put
5. Remarks
Mλ = ^λ≥ν
χν M.
for each λ in R. Denote by Pλ the orthogonal projection of L2(σ) onto Mλ. By the
property that
^−∞<λ<∞
Mλ = {0}
and
17
_−∞<λ<∞
Mλ = L2(σ) ,
we obtain the contuinity of the spectral resolution of identity {I − Pλ}λ∈R on L2(σ),
where I is the identity map on L2(σ). Let A(x, t) be the cocycle of M. By Stone's
theorem, a unitary group {Vt}t∈R on L2(σ) is defined as
Vt φ(x) = A(x, t)Tt φ(x) = −Z ∞
−∞
eiλt dPλφ(x) ,
φ ∈ L2(σ) ,
where Tt φ(x) = φ(x + et). For a nonzero function φ in L2(σ), −d(Pλφ, φ) is a finite
positive measure on R. On almost periodic flows, by comparing with Lebesgue measure
dλ, the type of such measures is uniquely determined. We then say that each of
M, A(x, t) and {Vt}t∈R is of absolutely continuous, or singular continuous, or discrete
type (as discussed in [6, §2.4]). This fact plays an important role to classify invariant
subspaces in this special context. It is easy to observe that A(x, t) and A(x, t) have
the same spectral type, so the following is an immediate consequence of Lemma 9.
Proposition 2. There is a simply invariant subspace of L2(σ) of either absolutely
continuous or singular continuous type which has no single generator.
Let w be a nonnegative function in L2(σ) satisfying (3), while log w does not lie in
L1(σ). We know that a cocycle is trivial if and only if it is of discrete type (see [6, §2.4,
Theorem 15 ]). It follows from Corollary 1 that the type of M[w] has to be continuous.
However, we have no idea to decide what kind of continuous spectrum M[w] may have.
(b) Using a suitable cocycle, we may construct a skew product on which the
H 2
0 −space is singly generated.
Indeed, let w be a bounded function as above and
let A(x, t) be the cocycle of M[w]. By Lemma 1 we see that almost every t →
A(x, t)w(x + et) is outer in H 2(dt/π(1 + t2)). Denote by (K × T, {St}t∈R) the skew
If A(x, t)n, n ≥ 1, is a coboundary q(x)q(x + et) with
product induced by A(x, t).
unitary function q on K, then qwn is a single generator of H 2
0 (σ). It then follows from
Theorem that A(x, t)n is a coboundary only for n = 0. Hence dµ = dσ × dθ/2π is an
ergodic measure on (K × T, {St}t∈R). If we set
φ(x, eiθ) = w(x)eiθ ,
(x, eiθ) ∈ K × T ,
then φ is a single generator of H 2
0 (µ), since log φ does not lie in L1(µ) and almost
every t → φ(St(x, eiθ)) is outer in H 2(dt/π(1 + t2)) (see [18] for another construction).
(c) We have a bit of information on the distribution of zeros of functions in H 2(σ)
which are connected with Dirichlet series (refer to [19] for related topics). Let {λn} be
a sequence in Γ such that
0 ≤ λ1 < λ2 < · · · < λn −→ λ ,
n → ∞ ,
for some λ in Γ. Define a function ψ in H 2(σ) by
with P∞
function.
n=1 an2 < ∞. Observe that almost every t → ψ(x + et) extends to an entire
ψ =
∞Xn=1
anχλn
18
Proposition 3. Let ψ be as above and let δ > 0. Then there is a decreasing sequence
{mn} with mn → −∞ such that the number of zeros of z → ψ(x + ez) in the strip
Sn = { z = t + iu ; mn > u > mn − δ}
is infinite, for σ − a.e. x in K.
Proof. Putting νn = λ − λn, we let φ = P∞
n=1 an χνn. Since z → eiλz has no zero,
z → ψ(x + ez) has zero at z if and only if so does z → φ(x + ez) at ¯z. For each r > 0,
t → φ ∗ Pir(x + et) cannot be an outer function in H 2(dt/π(1 + t2)), even if log φ does
not lie in L1(σ). Since φ has no weight at infinity, the inner part of t → φ ∗ Pir(x + et)
derives a Blaschke cocycle being not constant. From this fact, we may choose easily a
desired decreasing sequence {mn}.
(cid:3)
[Note on September 9, 2014: Made some changes in wording]
[Note on April 13, 2015: Accepted in ANNALES DE L'INSTITUTE FOURIER.]
References
[1] K. de Leeuw and I. Glicksberg, Quasi-invariance and analyticity of measures on compact groups,
Acta Math. 109 (1963), 179 -- 205.
[2] F. Forelli, Analytic and quasi-invariant measures, Acta Math. 118 (1967), 33 -- 59.
[3] T. Gamelin, H p spaces and extremal functions in H 1, Trans. Amer. Math. Soc. 124 (1966),
158 -- 167.
[4] T. Gamelin, Uniform algebras, Prentice-Hall, Englewood Cliffs, N.J., 1969.
[5] H. Helson, Cocycles in harmonic analysis, Actes Cong`es Intern. Math. 2 (1970), 477 -- 481.
[6] H. Helson, Analyticity on compact abelian groups, in Algebras in Analysis, Academic Press, Lon-
don, 1975, 1 -- 62.
[7] H. Helson, Compact groups with ordered duals VI, Bull. London Math. Soc. 8 (1976), 278 -- 281.
[8] H. Helson, Compact groups with ordered duals VII, J. London Math. Soc. (2) 20 (1979), 509 -- 515.
[9] H. Helson and D. Lowdenslager, Prediction theory and Fourier series of several complex variables
II, Acta Math. 106 (1961), 175 -- 213.
[10] H. Helson and W. Parry, Cocycles and spectra, Arkiv for Mat. 16 (1978), 195 -- 206.
[11] H. Helson and J. Tanaka, Singular cocycles and the generator problem, Proc. Internat. 17th
Conference on Operator Theory, Theta, 2000, 173 -- 186.
[12] C. Moore and K. Schmidt, Coboundaries and homomorphisms for non singular actions and a
problem of H. Helson, Proc. London Math. Soc. (3) 40 (1980), 443-475.
[13] P. Muhly, Function algebras and flows, Acta Sci. Math. (Szeged) 35 (1973), 111 -- 121.
[14] P. Muhly, Function algebras and flows III, Math. Z. 136 (1974), 253 -- 260.
[15] K. Petersen, Ergodic Theory, Cambridge University Press, Cambridge, 1983.
[16] W. Rudin, Fourier Analysis on Groups, Interscience, New York, 1962.
[17] J. Tanaka, Blaschke cocycles and generators, Pacific J. Math. 142 (1990), 357 -- 378.
[18] J. Tanaka, Single generator problem, Trans. Amer. Math. Soc. 348 (1996), 4113 -- 4129.
[19] J. Tanaka, Dirichlet series induced by the Riemann zeta-function, Studia Math. 187 (2008),
157 -- 184
Department of Mathematics, School of Education, Waseda University, Shinjuku,
Tokyo 169-8050, Japan
E-mail address: [email protected]
19
|
1502.01078 | 2 | 1502 | 2016-02-03T08:07:32 | Some applications of almost analytic extensions to operator bounds in trace ideals | [
"math.FA",
"math-ph",
"math-ph",
"math.SP"
] | Using the Davies-Helffer-Sj\"ostrand functional calculus based on almost analytic extensions, we address the following problem: Given self-adjoint operators $S_j$, $j=1,2$, in $\mathcal{H}$, and functions $f$ in an appropriate class, for instance, $f \in C_0^{\infty}(\mathbb{R})$, how to control the norm $\|f(S_2) - f(S_1)\|_{\mathcal{B}(\mathcal{H})}$ in terms of the norm of the difference of resolvents, $\|(S_2 - z_0 I_{\mathcal{H}})^{-1} - (S_2 - z_0 I_{\mathcal{H}})^{-1}\|_{\mathcal{B}(\mathcal{H})}$, for some $z_0 \in \mathbb{C}\backslash\mathbb{R}$. We are particularly interested in the case where $\mathcal{B}(\mathcal{H})$ is replaced by a trace ideal, $\mathcal{B}_p(\mathcal{H})$, $p \in [1,\infty)$. | math.FA | math |
SOME APPLICATIONS OF ALMOST ANALYTIC EXTENSIONS
TO OPERATOR BOUNDS IN TRACE IDEALS
FRITZ GESZTESY AND ROGER NICHOLS
Dedicated with deep admiration to Yurij Makarovich Berezansky
on the occasion of his 90th birthday.
Abstract. Using the Davies -- Helffer -- Sjostrand functional calculus based on
almost analytic extensions, we address the following problem: Given self-
adjoint operators Sj, j = 1, 2, in H, and functions f in an appropriate class, for
instance, f ∈ C∞
0 (R), how to control the norm kf (S2)−f (S1)kB(H) in terms of
the norm of the difference of resolvents, (cid:13)
(cid:13)B(H),
for some z0 ∈ C\R. We are particularly interested in the case where B(H) is
replaced by a trace ideal, Bp(H), p ∈ [1, ∞).
(cid:13)(S2 −z0IH)−1 −(S1 −z0IH)−1 (cid:13)
1. Introduction
Yurij M. Berezansky's contributions to analysis in general, and areas such as
functional analysis, operator theory, spectral and inverse spectral theory, harmonic
analysis, analysis in spaces of functions of an infinite number of variables, stochas-
tic calculus, mathematical physics, quantum field theory, integration of nonlinear
evolution equations, in particular, are legendary and of a lasting nature. The list of
fields his ground breaking work changed in dramatic fashion can easily be continued
in many directions as is demonstrated by the extraordinary breadth revealed in his
highly influential monographs [2] -- [8]. Since operator theoretic methods frequently
play a role in his research interests, we hope our modest contribution to operator
bounds in trace ideals will create some joy for him.
This paper has its origins in the following question: Given self-adjoint operators
Sj, j = 1, 2, in H, and functions f in an appropriate class, for instance, f ∈
C∞
0 (R), how to control the norm kf (S2) − f (S1)kB(H) in terms of the norm of the
difference of resolvents,(cid:13)(cid:13)(S2 − z0IH)−1 − S1 − z0IH)−1(cid:13)(cid:13)B(H), for some z0 ∈ C\R?
In particular, the question is just as natural with B(H) replaced by a trace ideal,
Bp(H), p ∈ [1, ∞).
In fact, our interest in these questions stems from computations of the Witten
index (a suitable extension of the Fredholm index) for a class of non-Fredholm
model operators
D
A =
d
dt
+ A,
dom(D
A) = W 1,2(R; H) ∩ dom(A),
(1.1)
Date: March 30, 2019.
2010 Mathematics Subject Classification. Primary: 47A30, 47A53, 47A60, 47B10; Secondary:
47B25, 30D99, 35Q40.
Key words and phrases. Almost analytic extensions, trace ideals, operator bounds.
R.N. gratefully acknowledges support from an AMS -- Simons Travel Grant.
Appeared in Methods Funct. Anal. Topology, 21, 151 -- 169 (2015).
1
2
F. GESZTESY AND R. NICHOLS
in the Hilbert space L2(R; H), where
(Af )(t) = A(t)f (t) for a.e. t ∈ R,
f ∈ dom(A) =(cid:26)g ∈ L2(R; H)(cid:12)(cid:12)(cid:12)(cid:12) g(t) ∈ dom(A(t)) for a.e. t ∈ R;
H < ∞(cid:27),
t 7→ A(t)g(t) is (weakly) measurable; R
dt kA(t)g(t)k2
(1.2)
with A(t), t ∈ R, a family of self-adjoint operators in H with asymptotes A± (in
norm resolvent sense).
Interesting concrete examles for A± are given by mass-
less Dirac-type operators in H = L2(Rd), d ∈ N (the latter are known to be
non-Fredholm), see, [10] -- [15]. More precisely, given the sequence of self-adjoint
operators H j,n, H j in L2(R; dt; H), j = 1, 2, n ∈ N, and self-adjoint operators
A+,n, A+, A− in H, n ∈ N, and a Pushnitski-type relation between the spectral
shift functions ξ( · ; H 2,n, H 1,n) and ξ( · ; A+,n, A−) for the pairs, (H 2,n, H 1,n) and
(A+,n, A−) of the form
ξ(λ; H 2,n, H 1,n) =( 1
π ´ λ1/2
−λ1/2
0, λ < 0,
ξ(ν;A+,n,A−) dν
(λ−ν 2)1/2
for a.e. λ > 0,
n ∈ N,
(1.3)
we were interested in performing the limit n → ∞ in (1.3) to obtain the analogous
relation for the limiting spectral shift functions ξ( · ; H 2, H 1) and ξ( · ; A+, A−) cor-
responding to the limiting pairs (H 2, H 1) and (A+, A−), respectively. The latter
is instrumental in computing the Witten index for DA. The task of performing the
limit n → ∞ in (1.3) is considerably complicated since due to the nature of the ap-
proximations involved, no suitable bounds on ξ( · ; H 2,n, H 1,n) and ξ( · ; A+,n, A−)
(independent of n ∈ N) are readily available. To circumvent this difficulty one can
resort to a distributional approach considering
R
dλ ξ(λ; H 2,n, H 1,n)f ′(λ) =
1
π R
dν ξ(ν; A+,n, A−)F ′(ν), n ∈ N,
(1.4)
where f ∈ C∞
0 (R) is arbitrary, and F ′ ∈ C∞
0 (R) is given by
F ′(ν) = ∞
ν 2
dλ f ′(λ)(λ − ν2)−1/2,
ν ∈ R.
(1.5)
Focusing now on the left-hand side of (1.4), one recalls Krein's trace formula,
dλ ξ(λ; H 2,n, H 1,n)f ′(λ),
f ∈ C∞
0 (R).
lim
n→∞
k[f (H2,n) − f (H 1,n)] − [f (H 2) − f (H 1)]kB1(L2(R;H)) = 0,
and hence obtain
lim
n→∞[0,∞)
dλ ξ(λ; H 2,n, H 1,n)f ′(λ) = lim
n→∞
trB1(L2(R;H))(f (H 2,n) − f (H 1,n))
Thus, given control of resolvents in the form
trB1(L2(R;H))(f (H 2,n) − f (H 1,n)) = [0,∞)
n→∞(cid:13)(cid:13)(cid:2)(H 2,n − z I)−m2 − (H 1,n − z I)−m2(cid:3)
lim
− [(H 2 − z I)−m2 − (H 1 − z I)−m2(cid:3)(cid:13)(cid:13)B1(L2(R;H)) = 0,
one can hope to control
z ∈ C\R.
(1.6)
(1.7)
(1.8)
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
3
= trB1(L2(R;H))(f (H 2) − f (H 1)) = [0,∞)
dλ ξ(λ; H 2, H 1)f ′(λ),
f ∈ C∞
0 (R).
(1.9)
Together with controlling the limit n → ∞ on the right-hand side of (1.4), this
leads to
dλ ξ(λ; H 2, H 1)f ′(λ) =
dν ξ(ν; A+, A−)F ′(ν).
(1.10)
R
1
π R
Without going into further details we note that (1.10) in turn can be used to
prove the limiting relation in (1.3) and the latter leads to a compution of the
semigroup regularized Witten index, Ws(DA), of DA: Assuming that 0 is a left
and a right Lebesgue point of ξ( · ; A+, A−), denoting the corresponding values by
ξ(0±; A+, A−), the semigroup regularized Witten index is found to be
Ws(DA) := lim
t↑∞
trL2(R;H)(cid:0)e−tH 1 − e−tH 2(cid:1)
= [ξ(0−; A+, A−) + ξ(0+; A+, A−)]/2,
(1.11)
see, for instance, [10] -- [15]. (We here use the semigroup regularized Witten index
rather than the resolvent regularised one as the former is applicable in the case of d-
dimensional Dirac-type operators A±, d ∈ N.) We trust this sufficiently illustrates
our interest in using control of resolvents of self-adjoint operators to gain control
over their C∞
0 -functions.
We also note a further complication lies in the fact that when studying multi-
dimensional Dirac-type operators A±, resolvents alone are not sufficient in the trace
class context and hence sufficiently high powers (depending on the space dimension
involved) of resolvents have to be employed.
Our principal tool to gain control over C∞
0 -functions of S in terms of (powers of)
resolvents of S is furnished by a suitable application of almost analytic extensions
efℓ,σ of f in the form of a Davies -- Helffer -- Sjostrand functional calculus [18], [20,
Ch. 2], [40, Proposition 7.2], of the form
(z)(S − zIH)−1,
(1.12)
f (S) = π−1C
dxdy
∂efℓ,σ
∂z
and a refinement due to Khochman [43] of the type
f (S) =
1
π C
dxdy
∂efℓ,σ
∂z
(z)(z − z0)m(S − z0IH)−m(S − zIH)−1, m ∈ N ∪ {0}.
(1.13)
to be discussed in some detail in Section 2. Section 3 contains our principal results
and some applications. Finally, Appendix A recalls various useful facts concerning
(powers of) resolvents.
We conclude with some comments on the notation employed in this paper: Let
H be a separable complex Hilbert space, (·, ·)H the scalar product in H (linear in
the second argument), and IH the identity operator in H.
Next, if T is a linear operator mapping (a subspace of) a Hilbert space into
another, then dom(T ) and ker(T ) denote the domain and kernel (i.e., null space)
of T . The spectrum and resolvent set of a closed linear operator in a Hilbert space
will be denoted by σ(·) and ρ(·), respectively. t
The convergence of bounded operators in the strong operator topology (i.e.,
pointwise limits) will be denoted by s-lim.
4
F. GESZTESY AND R. NICHOLS
The Banach spaces of bounded and compact linear operators on a separable
complex Hilbert space H are denoted by B(H) and B∞(H), respectively; the cor-
responding ℓp-based trace ideals will be denoted by Bp(H), their norms are abbre-
viated by k · kBp(H), p > 1. Moreover, trH(A) denotes the corresponding trace of a
trace class operator A ∈ B1(H).
The symbol C∞
0 (R) represents C∞-functions of compact support on R; contin-
uous functions on R vanishing at infinity are denoted by C∞(R).
2. Basic Facts on Almost Analytic Extensions And The Functional
Calculus For Self-Adjoint Operators
In this preparatory section we briefly recall the basics of almost analytic ex-
tensions and the ensuing functional calculus for self-adjoint operators, following
Davies' detailed treatment in [18], [20, Ch. 2].
One introduces the class Sβ(R), β ∈ R, consisting of all functions f ∈ C∞(R)
such that
f (m)(x) =
x→∞
O((cid:10)xiβ−m(cid:1), m ∈ N0,
where hzi = (cid:0)z2 + 1(cid:1)1/2
pointwise multiplication, Sβ(R) · Sγ(R) ⊆ Sβ+γ(R), β, γ ∈ R, and the space
, z ∈ C. Then in obvious notation, with "·" denoting
is an algebra under pointwise multiplication with
A(R) = [β<0
Sβ(R)
C∞
0 (R) ⊂ A(R).
(2.1)
(2.2)
(2.3)
In particular, f ∈ A(R) implies f ∈ C∞(R) (the continuous functions vanishing at
±∞) and f (m) ∈ L1(R), m ∈ N.
Given f ∈ A(R), one defines an almost analytic extension efℓ,σ, of f to C by
f (k)(x)(iy)k
z = x + iy ∈ C, ℓ ∈ N,
,
(2.4)
efℓ,σ(z) = σ(x, y)
ℓXk=0
k!
where
σ(x, y) = τ (y/hxi), x, y ∈ R,
τ ∈ C∞
0 (R), τ (s) =(1,
0,
s 6 1,
s > 2.
(2.5)
possible (cf., [18]).
We note the formula
The precise structure of efℓ,σ will not be important and other expresssions for it are
∂efℓ,σ
2(cid:18) ∂efℓ,σ
(z) + i
∂efℓ,σ
∂y
(z) =
∂x
∂z
1
=
[σx(x, y) + σy(x, y)]
1
2
f (k)(x)(iy)k
f ℓ+1(x)(iy)ℓ
+
σ(x, y)
1
2
ℓ!
,
(2.6)
implying the crucial fact,
∂efℓ,σ
∂z
(cid:12)(cid:12)(cid:12)(cid:12)
z ∈ C,
(2.7)
O(cid:0)yℓ(cid:1),
k!
(z)(cid:19)
ℓXk=0
(x + iy)(cid:12)(cid:12)(cid:12)(cid:12) =
y↓0
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
5
in particular,
∂efℓ,σ
∂z
(x) = 0,
x ∈ R.
(2.8)
Following Helffer and Sjostrand [40, Proposition 7.2], particularly, in the form
presented by Davies [18], [20, Ch. 2], one then establishes a functional calculus for
self-adjoint operators S in a complex, separable Hilbert space H via the formula
f (S) = π−1C
dxdy
∂efℓ,σ
∂z
(z)(S − zIH)−1.
(2.9)
Since the integrand is norm continuous, the integral in (2.9) is norm convergent,
in particular, one notes that (2.7) and (2.8), together with the standard estimate
6 Im(z)−1, overcome the apparent singularity of the integrand
(cid:13)(cid:13)(S −zIH)−1(cid:13)(cid:13)B(H)
in (2.9) for z ∈ σ(S) ⊆ R (cf. also (2.8)).
The justification for calling this a functional calculus follows upon proving the
following facts:
• The left-hand side of (2.8) is independent of the choice of ℓ ∈ N and the precise
form of σ in (2.5).
0 (R) with supp(f ) ∩ σ(S) = ∅, then f (S) = 0.
• If f ∈ C∞
• If f, g ∈ A(R), then (f g)(S) = f (S)g(S), f (S)∗ = f (S), kf (S)kB(H) = kf kL∞(R).
• Let z ∈ C\R and fz(x) = (x − z)−1, then fz ∈ A(R) and fz(S) = (S − zIH)−1.
In addition, we note that Khochman [43] proved the following extension of (2.9):
Lemma 2.1 ([43]). Let m ∈ N, f ∈ C∞
H. Then,
0 (R), and suppose that S is self-adjoint in
f (S) =
1
π C
dxdy
∂efℓ,σ
∂z
(z)(z − z0)m(S − z0IH)−m(S − zIH)−1.
(2.10)
We will employ (in fact, rederive) (2.10) in the proof of Theorem 3.8. Next, we
discuss another extension focusing on semigroups rather than powers of resolvents.
Lemma 2.2. Let t > 0, f ∈ C∞
from below in H. Then,
0 (R), and suppose that S is self-adjoint and bounded
f (S) =
1
π C
dxdy
∂efℓ,σ
∂z
(z)etze−tS(S − zIH)−1.
(2.11)
Proof. We start by noting that if f, g ∈ C∞
0 (R), it is proved in [20, p. 28] that
C
dxdy
∂g(f g)ℓ′′,σ′′
∂z
(z)(S − zIH)−1 = C
dxdy
∂z
0 (R) and let Et ∈ C∞
∂( fℓ′,σ′ gℓ,σ)
Next, suppose that f ∈ C∞
coincides with etx on an open interval I with supp(f ) ⊂ I. Then
(z)(S − zIH)−1. (2.12)
0 (R) denote a function which
eEt,ℓ,σ(z) = σ(x, y)etx
(ity)k
k!
ℓXk=0
,
z = x + iy, x ∈ I.
(2.13)
almost analytic extension of g, in light of the identity f (S) = g(S)e−tS, one infers
Let fℓ′,σ′ denote an almost analytic extension of f . Setting g = f Et, withegℓ′′,σ′′ an
6
F. GESZTESY AND R. NICHOLS
from the Davies -- Helffer -- Sjostrand functional calculus (2.9) applied to g,
f (S) =
=
=
=
1
1
1
1
∂z
∂z
∂z
∂z
dxdy
dxdy
dxdy
(z)e−tS(S − zIH)−1
(z)e−tS(S − zIH)−1
(z)e−tS(S − zIH)−1
π C
π C
π C
π C
π C
∂egℓ′′,σ′′
∂(efℓ′,σ′eEt,ℓ,σ)
∂efℓ′,σ′
(z)eEt,ℓ,σ(z)e−tS(S − zIH)−1
∂eEt,ℓ,σ
dxdy efℓ′,σ′ (z)
(z)σ(x, y)etx(cid:18) ℓXk=0
∂efℓ′′,σ′′
dxdy efℓ′,σ′ (z)( 1
2(cid:20)σx(x, y)etx(cid:18) ℓXk=0
k! (cid:19)
+ σ(x, y)etx(cid:18) ℓXk=0
k! (cid:19)(cid:21)
k! (cid:19) + σy(x, y)etx(cid:18) ℓXk=0
2(cid:20)σ(x, y)etxit(cid:18) ℓ−1Xk=0
π C
(ity)k
(ity)k
(ity)k
(ity)k
dxdy
∂z
1
i
+
1
+
+
k! (cid:19)e−tS(S − zIH)−1
(ity)k
k! (cid:19)(cid:21))
× e−tS(S − zIH)−1.
(2.14)
Exploiting the fact that f (S) is independent of ℓ, we now take the limit ℓ → ∞
in (2.14). Since efℓ′,σ′ has compact support and takes care of the singularity of the
resolvent, e−tS is bounded and z-independent, and the exponential series converges
uniformly on compact sets, one may pass the limit under the integral to obtain
dxdy
f (S) =
=
=
+
1
1
π C
π C
π C
π C
1
1
dxdy
dxdy
∂
∂
∂z
∂z
e−tS(S − zIH)−1
∂(σ(x, y)etz)
σ(x, y)etze−tS(S − zIH)−1
∂efℓ′,σ′
dxdy efℓ′,σ′ (z)
∂z(cid:0)efℓ′,σ′ (z)σ(x, y)etz(cid:1)e−tS(S − zIH)−1
∂z(cid:0)efℓ′,bσ(z)etz(cid:1)e−tS(S − zIH)−1,
∂z(cid:0)efℓ′,bσ(z)etz(cid:1) =
∂efℓ′,bσ
π C
etze−tS(S − zIH)−1
∂efℓ,bσ
dxdy
etz,
∂z
∂z
∂
1
which then shows
f (S) =
and hence (2.11) (renaming ℓ′ andbσ).
verify that
where bσ = σ′σ (which corresponds to choosing bτ = τ ′τ ). It is a simple matter to
(2.15)
(2.16)
(2.17)
(cid:3)
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
7
Historically, the idea of almost analytic (resp., pseudo-analytic) extensions ap-
peared in Hormander [39] and Dynkin [28], [29], Melin and Sjostrand [46] (see also
[26, Ch. 8], [45, Sect. III.6] for expositions and [42] for an alternative approach).
The functional calculus was used by Helffer and Sjostrand in their seminal 1989
paper on Schrodinger operators with magnetic fields [40], which in turn was the
basis for the systematic treatment by Davies [18], [20, Ch. 2]. Since these early
developments, there has been a large body of literature in connection with spectral
theory for Schrodinger and Dirac-type operators applying this functional calculus.
While a complete list of references in this context is clearly beyond the scope of
this paper, we want to illustrate the great variety of applications that rely on this
functional calculus a bit and hence refer to [16], [21], [22], [23], [24], [25], [27], [30],
[31], [34], [35], [37], [41], [43], [47], [48], [52], [53], and the references cited therein.
While we here exclusively focus on linear operators in a Hilbert space, this func-
tional calculus applies to operators in Banach spaces with real spectrum, see, for
instance, [1], [17], [18], [19], [32], [33]. Extensions to the case where the spectrum
is contained in the unit circle or contained in finitely-many smooth arcs were also
treated in [28].
3. Some Applications
In this section we apply the almost analytic extension method and its ensuing
functional calculus for self-adjoint operators to derive various norm bounds and
convergence properties of operators in trace ideals.
We start with the following estimates established in the proof of [20, Theo-
rem 2.6.2] (more precisely, (3.1) is proved in [20], but then the rest of Lemma 3.1
is obvious):
Lemma 3.1. Let z0 ∈ C\R, f ∈ A(R) and suppose that Sj, j = 1, 2, are self-
adjoint in H. Then
kf (S2) − f (S1)kB(H) 6
8
∂ef2,σ
(cid:2)z02 + z2(cid:3)
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
π C
×(cid:13)(cid:13)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:13)(cid:13)B(H).
(z)(cid:12)(cid:12)(cid:12)(cid:12)
Im(z)2
∂z
In addition, if for some p ∈ [1, ∞), (cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3) ∈ Bp(H) for
some (and hence for all ) z0 ∈ C\R, then
[f (S2) − f (S1)] ∈ Bp(H)
(3.1)
(3.2)
(3.3)
8
∂z
∂ef2,σ
(cid:2)z02 + z2(cid:3)
and
kf (S2) − f (S1)kBp(H) 6
If(cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3) ∈ B∞(H), the inclusion (3.2) extends to p = ∞.
f (S2) − f (S1) = π−1 C
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
π C
×(cid:13)(cid:13)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:13)(cid:13)Bp(H).
∂ef2,σ
(z)(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)
Proof. Combining (2.9), (A.1), and (A.2) one obtains,
(z)(cid:12)(cid:12)(cid:12)(cid:12)
Im(z)2
dxdy
(3.4)
∂z
and hence
kf (S2) − f (S1)kB(H)
8
F. GESZTESY AND R. NICHOLS
∂z
∂z
∂ef2,σ
∂ef2,σ
∂ef2,σ
∂ef2,σ
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
6 π−1C
6 π−1C
×(cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3)(S1 − z0IH)(S1 − zIH)−1(cid:13)(cid:13)B(H)
6 π−1C
×(cid:13)(cid:13)(S1 − z0IH)(S1 − zIH)−1(cid:13)(cid:13)B(H)(cid:13)(cid:13)(cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3)(cid:13)(cid:13)B(H)
6(cid:18) 8
π C
×(cid:13)(cid:13)(cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3)(cid:13)(cid:13)B(H).
(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(S2 − zIH)−1 − (S1 − zIH)−1(cid:13)(cid:13)B(H)
(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(S2 − z0IH)(S2 − zIH)−1
(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(S2 − z0IH)(S2 − zIH)−1(cid:13)(cid:13)B(H)
(cid:2)z02 + z2(cid:3)
(z)(cid:12)(cid:12)(cid:12)(cid:12)
Im(z)2 (cid:19)
Precisely the same chain of estimates applies to B(H) replaced by Bp(H), relying on
the ideal properties of Bp(H), p ∈ [1, ∞). The case p = ∞ in (3.2) is a consequence
of the norm convergent integral on the right-hand side of (3.4).
(cid:3)
(3.5)
∂z
∂z
Combined with a Stone -- Weierstrass approximation argument and the fact that
kf (S)kB(H) = kf kL∞(R), f ∈ A(R), Lemma 3.1 yields the following well-known
fact, recorded, for instance, in [20, Theorem 2.62], [50, Theorem V.III.20(a)]:
Lemma 3.2. Let Sn, n ∈ N, and S be self-adjoint in H, and suppose that Sn
converges to S in norm resolvent sense as n → ∞. Then,
lim
n→∞
kf (Sn) − f (S)kB(H) = 0
(3.6)
for all f ∈ C∞(R).
Remark 3.3. We note that the functional calculus based on almost analytic ex-
tensions is not the only possible approach to address estimates such as (3.1) and
(3.3). As a powerful alternative we mention the theory of double operator integrals
(DOI), which can prove stronger inequalities of the following type (cf. [9], [12],
[56]): Given m ∈ N odd and p ∈ [1, ∞), there exist constants a1, a2 ∈ R\{0} and
C = C(f, m, a1, a2) ∈ (0, ∞) such that
(cid:13)(cid:13)f (A) − f (B)kBp(H) 6 C(cid:0)(cid:13)(cid:13)(A − a1iIH)−m − (B − a1iIH)−m(cid:13)(cid:13)Bp(H)
+(cid:13)(cid:13)(A − a2iIH)−m − (B − a2iIH)−m(cid:13)(cid:13)Bp(H)(cid:1),
which permits the use of differences of higher powers m ∈ N of resolvents to control
the k · kBp(H)-norm of the left-hand side [f (A) − f (B)] for f ∈ C∞
0 (R). In fact, this
extends to a much larger class of functions f , see [13] for details.
f ∈ C∞
0 (R),
(3.7)
Moreover, repeatedly differentiating
(z)(λ − z)−1,
λ ∈ R,
(3.8)
with respect to λ yields
dxdy
f (λ) = π−1C
∂efℓ,σ
f (m−1)(S) = π−1(−1)m−1(m − 1)!C
∂z
This leads to estimates of the type (3.7) but with f replaced by f (m−1).
⋄
dxdy
(z)(S − z)−m,
λ ∈ R.
(3.9)
∂efℓ,σ
∂z
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
9
We recall a useful result:
Lemma 3.4. Let p ∈ [1, ∞) and assume that R, Rn, T, Tn ∈ B(H), n ∈ N, satisfy
s-limn→∞ Rn = R and s-limn→∞ Tn = T and that S, Sn ∈ Bp(H), n ∈ N, satisfy
limn→∞ kSn − SkBp(H) = 0. Then limn→∞ kRnSnT ∗
n − RST ∗kBp(H) = 0.
This follows, for instance, from [38, Theorem 1], [51, p. 28 -- 29], or [55, Lemma
6.1.3] with a minor additional effort (taking adjoints, etc.).
Next, we describe a typical convergence result:
Theorem 3.5. Let Sj,n, n ∈ N, and Sj, j = 1, 2, be self-adjoint in H, and assume
that Sj,n converges in strong resolvent sense as n → ∞ to Sj, j = 1, 2, respectively.
(i) Suppose that for some (and hence for all ) z0 ∈ C\R,
lim
n→∞(cid:13)(cid:13)(cid:2)(S2,n − z0IH)−1 − (S2 − z0IH)−1(cid:3)
−(cid:2)(S1,n − z0IH)−1 − (S1 − z0IH)−1(cid:3)(cid:13)(cid:13)B(H) = 0.
k[f (S2,n) − f (S1,n)] − [f (S2) − f (S1)]kB(H) = 0,
f ∈ C∞
0 (R).
(3.10)
(3.11)
Then,
lim
n→∞
(ii) Let p ∈ [1, ∞) and suppose that for some (and hence for all ) z0 ∈ C\R,
(cid:2)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:3),(cid:2)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:3) ∈ Bp(H),
n ∈ N,
(3.12)
and
Then,
lim
n→∞
lim
n→∞(cid:13)(cid:13)(cid:2)(S2,n − z0IH)−1 − (S2 − z0IH)−1(cid:3)
−(cid:2)(S1,n − z0IH)−1 − (S1 − z0IH)−1(cid:3)(cid:13)(cid:13)Bp(H) = 0.
k[f (S2,n) − f (S1,n)] − [f (S2) − f (S1)]kBp(H) = 0,
f ∈ C∞
0 (R).
(3.13)
(3.14)
Proof. As usual, a combination of identity (A.1), Lemma 3.4, and the assumed
strong resolvent convergence of Sj,n to Sj as n → ∞, j = 1, 2, proves sufficiency
of the conditions (3.10) and (3.12) for just one z0 ∈ C\R. Thus, assumption (3.10)
actually implies
lim
n→∞(cid:13)(cid:13)(cid:2)(S2,n − zIH)−1 − (S2 − zIH)−1(cid:3)
−(cid:2)(S1,n − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:13)(cid:13)B(H) = 0,
Next, mimicking (3.4), one obtains,
[f (S2,n) − f (S1,n)] − [f (S2) − f (S1)]
z ∈ C\R.
(3.15)
= π−1C
dxdy
∂ef2,σ
∂z
and hence,
6 π−1 lim
n→∞C
(z)h(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1(cid:3)
−(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)i,
(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1(cid:3)
∂ef2,σ
∂z
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
(3.16)
f ∈ C∞
0 (R),
(3.17)
lim
n→∞
k[f (S2,n) − f (S1,n)] − [f (S2) − f (S1)]kB(H)
10
F. GESZTESY AND R. NICHOLS
−(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:13)(cid:13)(cid:13)B(H)
0 (R) implies ef2,σ ∈ C∞
,
In this context one observes that f ∈ C∞
0 (C). Since the
exceptional set R in (3.15) has dxdy-measure zero (in addition to the fact that by
(2.8), ∂ef2,σ/∂z vanishes on R), an application of the Lebesgue dominated conver-
gence theorem to interchange the limit n → ∞ with the integral on the right-hand
side of (3.17) requires establishing an n-independent integrable majorant of the
integrand in (3.17). Employing identity (A.1) and estimate (A.2), this majorant
can be obtained as follows:
f ∈ C∞
0 (R).
(3.18)
(3.19)
(cid:13)(cid:13)(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1(cid:3) −(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:13)(cid:13)B(H)
6(cid:13)(cid:13)(S2,n − zIH)−1 − (S1,n − zIH)−1(cid:13)(cid:13)B(H)
+(cid:13)(cid:13)(S2 − zIH)−1 − (S1 − zIH)−1(cid:13)(cid:13)B(H)
6(cid:13)(cid:13)(S2,n − z0IH)(S2,n − zIH)−1(cid:13)(cid:13)B(H)
×(cid:13)(cid:13)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:13)(cid:13)B(H)
×(cid:13)(cid:13)(S1,n − z0IH)(S1,n − zIH)−1(cid:13)(cid:13)B(H)
+(cid:13)(cid:13)(S2 − z0IH)(S2 − zIH)−1(cid:13)(cid:13)B(H)
×(cid:13)(cid:13)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:13)(cid:13)B(H)
×(cid:13)(cid:13)(S1 − z0IH)(S1 − zIH)−1(cid:13)(cid:13)B(H)
6 8(cid:2)z02 + z2(cid:3)Im(z)−2h(cid:13)(cid:13)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:13)(cid:13)B(H)
+(cid:13)(cid:13)(S2 − z0IH)−1 − (S1 − z0IH)−1(cid:13)(cid:13)B(H)i
6 8(cid:2)z02 + z2(cid:3)Im(z)−2C(z0),
for some 0 < C(z0) < ∞, independent of n ∈ N, since by assumption (3.10),
(cid:13)(cid:13)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:13)(cid:13)B(H)
6 eC(z0)
for some 0 < eC(z0) < ∞, independent of n ∈ N. Together with the properties (2.7),
(2.8) of ∂ef2,σ/∂z, this establishes the sought integrable majorant, independent of
n ∈ N, and thus permits the interchange of the limit n → ∞ with the integral on
the right-hand side of (3.17). This completes the proof of (3.11).
The proof of (3.14) proceeds exactly along the same lines employing once more
(cid:3)
the ideal properties of Bp(H), p ∈ [1, ∞).
We remark in passing that double operator integral techniques permit one to
enlarge the class of functions f to which Theorem 3.5 applies (cf. also Remark 3.9).
Remark 3.6. The proof to Theorem 3.5 uses dominated convergence and given
(3.18), the crucial observation is that
which is obvious if f ∈ C∞
it is possible to once again prove (3.20) if f ∈ Sβ(R) for some β < −1. Indeed,
z02 + z2
Im(z)2 < ∞,
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
C
0 (R), since then efℓ,σ is compactly supported. However,
(z)(cid:12)(cid:12)(cid:12)(cid:12)
∂efℓ,σ
∂z
(3.20)
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
11
following [20, p. 25], and setting
U = {(x, y) hxi < y < 2hxi},
V = {(x, y) 0 < y < 2hxi},
(3.21)
one infers
σx + iσy 6 chxi−1χU (x, y),
for some constant c > 0. Then for f ∈ Sβ(R),
z = x + iy ∈ C,
(3.22)
hxiβ−k−1ykχU (x, y)) + Chxiβ−ℓ−1yℓχV (x, y),
where C > 0 is an appropriate constant. Therefore,
z = x + iy ∈ C,
z02 + x2 + y2
(3.23)
)
∂z
(z)(cid:12)(cid:12)(cid:12)(cid:12) 6 C( ℓXk=0
∂efℓ,σ
(z)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
∂efℓ,σ
z02 + z2
Im(z)2
∂z
y2
y2
z02 + 5hxi2
hxiβ−1χU (x, y)
z02 + x2 + y2
hxiβ−k−1ykχU (x, y)
+ Chxiβ−ℓ−1yℓχV (x, y)
6 C( ℓXk=0
)
6 C2ℓ( ℓXk=0
+ Chxiβ−ℓ−1yℓ−2χV (x, y)(cid:8)z02 + 5hxi2(cid:9)
6 C2ℓ( ℓXk=0
hxiβ−3χU (x, y)(cid:8)z02 + 5hxi2(cid:9))
+ C2ℓ−2hxiβ−3χV (x, y)(cid:8)z02 + 5hxi2(cid:9)
6 bChxiβ−1(cid:8)χU (x, y) + χV (x, y)(cid:9),
hxi2
z = x + iy ∈ C\R,
(3.24)
where bC = bC(z0) > 0 is a constant. Here, we used z02 + 5hxi2 < eChxi2, for an
appropriate constant eC = eC(z0) > 0. Given (3.24), (3.20) holds if β < −1 since
C
−∞
dxdy hxiβ−1(cid:8)χU (x, y) + χV (x, y)(cid:9)
= ∞
= 2 ∞
= 6 ∞
dx hxiβ−1 ∞
dx hxiβ−1(cid:26) 2hxi
dx hxiβ < ∞.
−∞
−∞
0
−∞
dy(cid:8)χU (x, y) + χV (x, y)(cid:9)
dy + 2hxi
hxi
dy(cid:27)
(3.25)
Thus, the majorant (3.24) is integrable as long as f ∈ Sβ(R) for some β < −1 and
hence Theorem 3.5 extends from C∞
0 (R) to Sβ(R), β < −1.
Next, we briefly apply Theorem 3.5 to a concrete (1 + 1)-dimensional example
treated in great detail in [10] and [11] by alternative methods.
Example 3.7. Assuming the real-valued functions φ, θ satisfy
φ ∈ ACloc(R) ∩ L∞(R) ∩ L1(R), φ′ ∈ L∞(R),
(3.26)
12
F. GESZTESY AND R. NICHOLS
θ ∈ ACloc(R) ∩ L∞(R), θ′ ∈ L∞(R) ∩ L1(R),
lim
t→∞
θ(t) = 1,
θ(t) = 0,
t→−∞
lim
(3.27)
we introduce the family of self-adjoint operators A(t), t ∈ R, in L2(R),
A(t) = −i
d
dx
+ θ(t)φ,
dom(A(t)) = W 1,2(R), t ∈ R,
(3.28)
and its self-adjoint asymptotes as t → ±∞,
A+ = −i
+ φ, A− = −i
,
dom(A±) = W 1,2(R).
(3.29)
d
dx
d
dx
In addition, we introducing the operator d/dt in L2(cid:0)R; dt; L2(R; dx)(cid:1) by
(cid:18) d
f(cid:19)(t) = f ′(t) for a.e. t ∈ R,
f ∈ dom(d/dt) =(cid:8)g ∈ L2(cid:0)R; dt; L2(R)(cid:1)(cid:12)(cid:12) g ∈ ACloc(cid:0)R; L2(R)(cid:1),
g′ ∈ L2(cid:0)R; dt; L2(R)(cid:1)(cid:9)
dt
= W 1,2(cid:0)R; dt; L2(R; dx)(cid:1).
Next, we agree to identify L2(cid:0)R; dt; L2(R; dx)(cid:1) with L2(R2; dtdx) (denoting the lat-
ter by L2(R2) for brevity ) and introduce DA in L2(R2) by
D
A =
d
dt
+ A,
dom(D
A) = W 1,2(R2),
(3.32)
with A defined as in (1.2) and A(t), t ∈ R, given by (3.28). Moreover, we introduce
the nonnegative, self-adjoint operators H j, j = 1, 2, in L2(R2) by
H 1 = D∗
ADA, H 2 = DAD∗
A.
(3.33)
(3.30)
(3.31)
As shown in [10], the assumptions on φ and θ guarantee that
(cid:2)(A+ − zI)−1 − (A− − zI)−1(cid:3) ∈ B1(cid:0)L2(R)(cid:1),
(for simplicity, we adopt the abbreviation I = IL2(R) throughout this example ), and
thus, the spectral shift function ξ( · ; A+, A−) for the pair (A+, A−) exists and is
well-defined up to an arbitrary additive real constant, satisfying
z ∈ C\R
(3.34)
ξ( · ; A+, A−) ∈ L1(cid:0)R; (ν2 + 1)−1dν(cid:1).
− +n2I)−1/2 and A+,n = A− +χn(A−)φχn(A−), n ∈ N,
(3.35)
Introducing χn(A−) = n(A2
the fact
A+,n − A− = χn(A−)φχn(A−) ∈ B1(cid:0)L2(R)(cid:1), n ∈ N,
implies that also the spectral shift functions ξ( · ; A+,n, A−), n ∈ N, exist and are
uniquely determined by
(3.36)
ξ( · ; A+,n, A−) ∈ L1(R; dν),
n ∈ N.
(3.37)
In fact, as has been shown in [10], the open constant in ξ( · ; A+, A−) can naturally
be determined via the limiting procedure
lim
n→∞
ξ(ν; A+,n, A−) =
1
2π R
dx φ(x) = ξ(ν; A+, A−),
ν ∈ R.
(3.38)
In particular, ξ( · ; A+, A−) turns out to be constant in this example.
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
13
Replacing A(t) by An(t) = A− + χn(A−)θ(t)φχn(A−), n ∈ N, t ∈ R, and hence,
A by An, DA by DA
, H j by H j,n, j = 1, 2, n ∈ N, one verifies the facts,
n
z ∈ C\[0, ∞),
(cid:2)(H 2 − z I)−1 − (H 1 − z I)−1(cid:3) ∈ B1(cid:0)L2(R2)(cid:1),
(cid:2)(H 2,n − z I)−1 − (H 1,n − z I)−1(cid:3) ∈ B1(cid:0)L2(R2)(cid:1), n ∈ N, z ∈ C\[0, ∞), (3.40)
showing that the spectral shift functions ξ( · ; H 2, H 1) and ξ( · ; H 2,n, H 1,n) for the
pairs (H 2, H 1) and (H 2, H 1), n ∈ N, respectively, are well-defined and satisfy
ξ( · ; H 2, H 1), ξ( · ; H 2,n, H 1,n) ∈ L1(cid:0)R; (λ2 + 1)−1dλ(cid:1), n ∈ N.
Since H j > 0, H j,n > 0, n ∈ N, j = 1, 2, one uniquely introduces ξ( · ; H 2, H 1)
and ξ( · ; H 2,n, H 1,n), n ∈ N, by requiring that
(3.39)
(3.41)
ξ(λ; H 2, H 1) = 0,
ξ( · ; H 2,n, H 1,n) = 0,
λ < 0, n ∈ N.
(3.42)
As shown in [10], one can now prove the following intimate connection between
ξ( · ; A+,n, A−) and ξ( · ; H 2,n, H 1,n), n ∈ N, the Pushnitski-type formula [36], [49],
ξ(λ; H 2,n, H 1,n) =
1
π λ1/2
−λ1/2
ξ(ν; A+,n, A−)dν
(λ − ν2)1/2
for a.e. λ > 0, n ∈ N.
(3.43)
Moreover, as shown in [10] and [11], one indeed has the convergence property
lim
n→∞(cid:13)(cid:13)(cid:2)(H 2,n − z I)−1 − (H 1,n − z I)−1(cid:3)
− [(H 2 − z I)−1 − (H 1 − z I)−1(cid:3)(cid:13)(cid:13)B1(L2(R2)) = 0,
Thus, Theorem 3.5 applies and hence yields,
lim
n→∞
k[f (H 2,n) − f (H 1,n)] − [f (H 2) − f (H 1)]kB1(L2(R2)) = 0,
z ∈ C\R.
(3.44)
f ∈ C∞
0 (R).
(3.45)
This, in turn permits one to take the limit n → ∞ in (3.43), implying
ξ(λ; H 2, H 1) =
1
π λ1/2
−λ1/2
ξ(ν; A+, A−)dν
(λ − ν2)1/2
for a.e. λ > 0, n ∈ N.
(3.46)
Equation (3.46) combined with (3.38) yields
ξ(λ; H 2, H 1) = ξ(ν; A+, A−) =
1
2π R
dx φ(x)
(3.47)
for a.e. λ > 0 and a.e. ν ∈ R. As a consequence of (3.47), the semigroup regularized
Witten index Ws(DA) of the non-Fredholm operator DA exists and equals
Ws(D
A) = ξ(0+; H 2, H 1) = ξ(0; A+, A−) =
1
2π R
dx φ(x).
(3.48)
This yields an alternative proof of the principal Witten index results in [10] and
[15].
The following result provides an extension of Theorem 3.5 to higher powers of
resolvents (necessitated by applications to d-dimensional Dirac-type operators as
hinted at in the introduction).
Theorem 3.8. Let Sj,n, n ∈ N, and Sj, j = 1, 2, be self-adjoint in H, and assume
that Sj,n converges in strong resolvent sense as n → ∞ to Sj, j = 1, 2, respectively.
Suppose that for some m ∈ N and some p ∈ [1, ∞),
(cid:2)(S2,n − zIH)−m − (S1,n − zIH)−m(cid:3),(cid:2)(S2 − zIH)−m − (S1 − zIH)−m(cid:3) ∈Bp(H),
14
If
F. GESZTESY AND R. NICHOLS
z ∈ C\R, n ∈ N.
(3.49)
and for some z0 ∈ C\R,
lim
z ∈ C\R,
n→∞(cid:13)(cid:13)(cid:2)(S2,n − zIH)−m − (S1,n − zIH)−m(cid:3)
−(cid:2)(S2 − zIH)−m − (S1 − zIH)−m(cid:3)(cid:13)(cid:13)Bp(H) = 0,
(S1,n − z0IH)−m(cid:2)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:3),
(S1 − z0IH)−m(cid:2)(S2 − z0IH)−1 − (S1 − z0)−1(cid:3) ∈ Bp(H),
n→∞(cid:13)(cid:13)(S1,n − z0IH)−m(cid:2)(S2,n − z0IH)−1 − (S1,n − z0IH)−1(cid:3)
− (S1 − z0IH)−m(cid:2)(S2 − z0IH)−1 − (S1 − z0)−1(cid:3)(cid:13)(cid:13)Bp(H) = 0,
lim
k[f (S2,n) − f (S1,n)] − [f (S2) − f (S1)]kBp(H) = 0,
f ∈ C∞
0 (R).
with
then
lim
n→∞
(3.50)
(3.52)
(3.53)
n ∈ N,
(3.51)
In addition, these results hold upon systematically replacing Bp(H) by B(H) in
(3.49) -- (3.53).
Proof. Let f ∈ C∞
analytic extension. Following a device due to Khochman [43], one introduces
0 (R2) a compactly supported almost
0 (R) be fixed and efℓ,σ ∈ C∞
g(x) = f (x)(x − z0)m,
(3.54)
concluding g ∈ C∞
applied to g, one obtains for any self-adjoint operator S in H,
0 (R). By the Davies -- Helffer -- Sjostrand functional calculus (2.9)
g(S) =
1
π C
dxdy
∂efℓ,σ
∂z
and hence,
f (S) = (S − z0IH)−mg(S)
(z)(z − z0)m(S − zIH)−1,
(3.55)
∂z
∂efℓ,σ
(z)(z − z0)m(cid:12)(cid:12)(cid:12)(cid:12)
=
1
π C
dxdy
(z)(z − z0)m(S − z0IH)−m(S − zIH)−1.
(3.56)
Applying (3.56) multiple times choosing H ∈ {S2, S1, S2,n, S1,n}, one infers
(cid:13)(cid:13)(cid:2)f (S2,n) − f (S1,n)(cid:3) −(cid:2)f (S2) − f (S1)(cid:3)(cid:13)(cid:13)Bp(H)
1
6
∂z
∂efℓ,σ
dxdy(cid:12)(cid:12)(cid:12)(cid:12)
π C
×(cid:13)(cid:13)(cid:2)(S2,n − z0IH)−m(S2,n − zIH)−1 − (S1,n − z0IH)−m(S1,n − zIH)−1(cid:3)
−(cid:2)(S2 − z0IH)−m(S2 − zIH)−1 − (S1 − z0IH)−m(S1 − zIH)−1(cid:3)(cid:13)(cid:13)Bp(H),
n ∈ N.
(3.57)
In order to prove the convergence claim in (3.53), the idea is to take the limit
n → ∞ and apply dominated convergence in (3.57). However, doing so requires
one to obtain an n-independent integrable majorant for the expression under the
integral in (3.57) and then to show that the integrand converges to zero pointwise
+ (S1,n − z0IH)−m(S2,n − zIH)−1 − (S1,n − z0IH)−m(S1,n − zIH)−1
= (S2,n − z0IH)−m(S2,n − zIH)−1 − (S1,n − z0IH)−m(S2,n − zIH)−1
(cid:2)(S2,n − z0IH)−m(S2,n − zIH)−1 − (S1,n − z0IH)−m(S1,n − zIH)−1(cid:3)
−(cid:2)(S2 − z0IH)−m(S2 − zIH)−1 − (S1 − z0IH)−m(S1 − zIH)−1(cid:3)
−(cid:2)(S2 − z0IH)−m(S2 − zIH)−1 − (S1 − z0IH)−m(S2 − zIH)−1
+ (S1 − z0IH)−m(S2 − zIH)−1 − (S1 − z0IH)−m(S1 − zIH)−1(cid:3)
=(cid:2)(S2,n − z0IH)−m − (S1,n − z0IH)−m(cid:3)(S2,n − zIH)−1
+ (S1,n − z0IH)−m(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1(cid:3)
−(cid:8)(cid:2)(S2 − z0IH)−m − (S1 − z0IH)−m(cid:3)(S2 − zIH)−1
+ (S1 − z0IH)−m(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:9)
=(cid:8)(cid:2)(S2,n − z0IH)−m − (S1,n − z0IH)−m(cid:3)(S2,n − zIH)−1
−(cid:2)(S2 − z0IH)−m − (S1 − z0IH)−m(cid:3)(S2 − zIH)−1(cid:9)
+(cid:8)(S1,n − z0IH)−m(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1
− (S1 − z0IH)−m(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:9),
(cid:13)(cid:13)(cid:2)(S2,n − z0IH)−m − (S1,n − z0IH)−m(cid:3)(S2,n − zIH)−1
−(cid:2)(S2 − z0IH)−m − (S1 − z0IH)−m(cid:3)(S2 − zIH)−1(cid:13)(cid:13)Bp(H)
6 eC(z0)Im(z)−1,
z ∈ C\R, n ∈ N,
For the first term in braces after the final equality in (3.58), one has a bound of
the type
(3.58)
z ∈ C\R, n ∈ N.
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
15
with respect to z as n → ∞. In order to carry this out, one expresses the difference
in the k · kBp(H)-norm in (3.57) as follows:
(3.59)
for a constant eC(z0) > 0 which does not depend on n ∈ N or z ∈ C\R. The
estimate in (3.59) follows at once from the triangle inequality, basic properties of
the Schatten -- von Neumann trace ideals, the standard resolvent estimate
k(S − zIH)−1kB(H) 6 Im(z)−1,
z ∈ C\R,
(3.60)
for an arbitrary self-adjoint operator S in H, and assumption (3.50). Moreover, by
Lemma 3.2, one also infers
lim
n→∞(cid:13)(cid:13)(cid:2)(S2,n − z0IH)−m − (S1,n − z0IH)−m(cid:3)(S2,n − zIH)−1
−(cid:2)(S2 − z0IH)−m − (S1 − z0IH)−m(cid:3)(S2 − zIH)−1(cid:13)(cid:13)Bp(H) = 0,
For the second term in braces after the final equality in (3.58),
z ∈ C\R.
(3.61)
(S1,n − z0IH)−m(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1
− (S1 − z0IH)−m(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)
= (S1 − z0IH)−m(cid:2)(S1 − zIH)−1 − (S2 − zIH)−1(cid:3)
− (S1,n − z0IH)−m(cid:2)(S1,n − zIH)−1 − (S2,n − zIH)−1(cid:3)
= {IH + (z − z0)(S1 − zIH)−1}(S1 − z0IH)−m(cid:2)(S1 − z0IH)−1 − (S2 − z0IH)−1(cid:3)
16
F. GESZTESY AND R. NICHOLS
× {IH + (z − z0)(S2 − zIH)−1} − {IH + (z − z0)(S1,n − zIH)−1}
× (S1,n − z0IH)−m(cid:2)(S1,n − z0IH)−1 − (S2,n − z0IH)−1(cid:3)
× {IH + (z − z0)(S2,n − zIH)−1},
z ∈ C\R, n ∈ N.
(3.62)
Using (3.60), one finds for any self-adjoint operator H,
kIH + (z0 − z)(H − zIH)−1kB(H) 6 1 + (z0 + z)Im(z)−1
6 2(z0 + z)Im(z)−1,
z ∈ C\R.
(3.63)
As a result of (3.52), (3.62) and (3.63),
(cid:13)(cid:13)(S1,n − z0IH)−m(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1
− (S1 − z0IH)−m(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:13)(cid:13)Bp(H)
6 bC(z0)
z02 + z2
Im(z)2 ,
z ∈ C\R, n ∈ N,
(3.52) and another application of Lemma 3.4 immediately imply
for a constant bC(z0) > 0 which does not depend on n ∈ N or z ∈ C\R. Moreover,
n→∞(cid:13)(cid:13)(S1,n − z0IH)−m(cid:2)(S2,n − zIH)−1 − (S1,n − zIH)−1
z ∈ C\R.
lim
− (S1 − z0IH)−m(cid:2)(S2 − zIH)−1 − (S1 − zIH)−1(cid:3)(cid:13)(cid:13)Bp(H) = 0,
(cid:12)(cid:12)z − z0(cid:12)(cid:12)m(cid:12)(cid:12)(cid:12)(cid:12)
(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) eC(z0)
Im(z)2 (cid:19),
+ bC(z0)
∂efℓ,σ
[z02 + z2]
Im(z)
∂z
The estimates in (3.60) and (3.64) show that away from R, which has dxdy-measure
equal to zero, the integrand in (3.57) is bounded above by
(3.64)
(3.65)
z ∈ C\R,
(3.66)
which is integrable with ℓ = 2 in light of (2.7) and the fact that efℓ,σ is compactly
supported. Therefore, taking the limit n → ∞ on both sides of (3.57) and then
applying dominated convergence in combination with (3.59), (3.61), and (3.65)
yields (3.53).
Clearly, the proof remains valid with Bp(H) replaced by B(H).
(cid:3)
Remark 3.9. Although applicable to the Witten index computation described in
the introduction, Theorem 3.8 is far from optimal.
Indeed, upon communicat-
ing Theorem 3.8 to G. Levitina, D. Potapov, and F. Sukochev, they subsequently
pointed out to us [44] that an application of the double operator integral method
permits one to extend the classes of functions f to the one employed in [56], and
more importantly, the DOI approach permits one to dispense with the conditions
(3.51) and (3.52) altogether. (Conditions (3.51) and (3.52) are clearly an artifact
of the resolvent term (S − zIH)−1 in formula (3.56)). This will be further pursued
elsewhere [13].
Remark 3.10. While we exclusively focused on applications to self-adjoint operators
S, as long as σ(T ) ⊂ R and the singularity of the resolvent (T − zIH)−1 of T as z
approaches the spectrum is uniformly bounded by Im(z)−N for some fixed N ∈ N,
choosing ℓ ∈ N sufficiently large in efℓ,σ, one can handle such classes of non-self-
adjoint operators T , particularly, operators in Banach spaces with real spectrum.
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
17
In fact, a functional calculus for the case of a non-self-adjoint operator T with
σ(T ) ⊂ R and a resolvent that satisfies an estimate of the type
k(T − zIH)−1kB(H) 6 cIm(z)−1(cid:18) hzi
Im(z)(cid:19)α
,
z ∈ C\R,
(3.67)
for some c > 0 and α > 0, was discussed in [18], [19], and in subsequent develop-
ments in [1], [17], [32], [33]. Moreover, the case where the spectrum is contained in
⋄
the unit circle or contained in finitely-many smooth arcs was discussed in [28].
Appendix A. Some Useful Resolvent Identities
In this appendix we recall some well-known, yet useful relations for (powers of)
resolvents.
We start by recalling the well-known identity (see, e.g., [54, p. 178]),
(T2 − zIH)−1 − (T1 − zIH)−1 = (T2 − z0IH)(T2 − zIH)−1
×(cid:2)(T2 − z0IH)−1 − (T1 − z0IH)−1(cid:3)(T1 − z0IH)(T1 − zIH)−1,
z, z0 ∈ ρ(T1) ∩ ρ(T2),
(A.1)
where Tj, j = 1, 2, are linear operators in H with ρ(T1) ∩ ρ(T2) 6= ∅. In addition, if
S is self-adjoint in H, we recall the elementary estimate,
Im(z)−1,
z, z0 ∈ C\R.
In addition, for m ∈ N, we note (cf. [55, p. 315]),
(T2 − zIH)−(m+1) − (T1 − zIH)−(m+1)
(cid:13)(cid:13)(S − z0IH)(S − zIH)−1(cid:13)(cid:13)B(H) =(cid:13)(cid:13)IH + (z − z0)(S − zIH)−1(cid:13)(cid:13)B(H)
6 81/2(cid:2)z02 + z2(cid:3)1/2
=(cid:2)(T2 − zIH)−m − (T1 − zIH)−m(cid:3)(T1 − zIH)−1
+ (T2 − zIH)−m(cid:2)(T2 − zIH)−1 − (T1 − zIH)−1(cid:3),
= (T2 − zIH)−1(cid:2)(T2 − zIH)−m − (T1 − zIH)−m(cid:3)
+(cid:2)(T2 − zIH)−1 − (T1 − zIH)−1(cid:3)(T2 − zIH)−m
−(cid:2)(T2 − zIH)−1 − (T1 − zIH)−1(cid:3)(cid:2)(T2 − zIH)−m − (T1 − zIH)−m(cid:3),
(T2 − zIH)−(m+1) − (T1 − zIH)−(m+1)
z ∈ ρ(T1) ∩ ρ(T2).
z ∈ ρ(T1) ∩ ρ(T2),
(A.2)
(A.3)
(A.4)
and
Next, by applying Cauchy's integral formula,
f (k)(z) = −
k!
2πi ‰Γ
dζ
f (ζ)
(ζ − z)k+1 ,
z ∈ Ω,
(A.5)
where f is an analytic function in the open set Ω ⊂ C, and Γ is a counterclockwise-
oriented contour encompassing the point z ∈ Ω, to a densely defined, closed linear
operator T in H with nonempty resolvent set, one obtains a formula for higher
powers of the resolvent of T in terms of a fixed lower power as follows,
(T − zIH)−k = −
(k − m)!(m − 1)!
2πi[(k − 1)!]
‰Γz
dζ (ζ − z)m−k−1(T − ζIH)−m,
(A.6)
z ∈ ρ(H0),
18
F. GESZTESY AND R. NICHOLS
where for each z ∈ ρ(T ), Γz is any counterclockwise-oriented circular contour cen-
tered at z which does not intersect or encompass points of σ(T ).
The following lemma (cf. [55, p. 210]) states an elementary, yet useful, fact:
z ∈ C\R,
(A.7)
z ∈ C\R, n > m.
(A.8)
Lemma A.1. Let Sj, j ∈ {1, 2} be self-adjoint operators in H. If
for some p ∈ [1, ∞) ∪ {∞} and some m ∈ N, then
(cid:2)(S2 − zIH)−m − (S1 − zIH)−m(cid:3) ∈ Bp(H),
(cid:2)(S2 − zIH)−n − (S1 − zIH)−n(cid:3) ∈ Bp(H),
Proof. It suffices to apply the Cauchy-type formula (A.6) and note that
(S2 − zIH)−n − (S1 − zIH)−n
= −
(n − m)!(m − 1)!
2πi[(n − 1)!]
‰Γz
dζ (ζ − z)m−n−1(cid:2)(S2 − ζIH)−m − (S1 − zIH)−m(cid:3),
z ∈ C\R, n > m,
(A.9)
where Γz is a counterclockwise-oriented circular contour centered at z that does
not intersect R.
(cid:3)
Acknowledgments. We are indebted to Alan Carey, Galina Levitina, Denis
Potapov, Fedor Sukochev, Yuri Tomilov, and Dmitriy Zanin for helpful discus-
sions, and particularly to Yuri Tomilov for pointing out to us a number of key
references in connection with almost analytic extensions.
References
[1] A. B´atkai, and E. Facanga, The spectral mapping theorem for Davies' functional calculus.
Rev. Roumaine Math. Pures Appl. 48, 365 -- 372 (2003).
[2] Ju. M. Berezanskii, Expansions in Eigenfunctions of Selfadjoint Operators, Translations of
Mathematical Monographs, Vol. 17, Amer. Math. Soc., Providence, R.I. 1968.
[3] Yu. M. Berezanskii, Selfadjoint Operators in Spaces of Functions of Infinitely Many Vari-
ables, Translations of Mathematical Monographs, Vol. 63, Amer. Math. Soc., Providence, R.I.
1986.
[4] Yu. M. Berezansky and A. A. Kalyuzhnyi, Harmonic Analysis in Hypercomplex Systems,
Mathematics and its Applications, 434, Kluwer, Dordrecht, 1998.
[5] Y. M. Berezansky and Y. G. Kondtratiev, Spectral methods in infinite-dimensional analysis.
Vol. 1, Mathematical Physics and Applied Mathematics, 12/1, Kluwer Academic, Dordrecht,
1995.
[6] Y. M. Berezansky and Y. G. Kondtratiev, Spectral methods in infinite-dimensional analysis.
Vol. 2, Mathematical Physics and Applied Mathematics, 12/2, Kluwer Academic, Dordrecht,
1995.
[7] Y. M. Berezansky, Z. G. Sheftel, and G. F. Us, Functional Analysis. Vol. I, Operator Theory:
Advances and Applications, Vol. 85, Birkhauser, Basel, 1996.
[8] Y. M. Berezansky, Z. G. Sheftel, and G. F. Us, Functional Analysis. Vol. II, Operator Theory:
Advances and Applications, Vol. 86, Birkhauser, Basel, 1996.
[9] M. Sh. Birman and M. Solomyak, Double operator integrals in a Hilbert space, Integral Eq.
Oper. Theory 47, 131 -- 168 (2003).
[10] A. Carey, F. Gesztesy, G. Levitina, D. Potapov, F. Sukochev, and D. Zanin, On index theory
for non-Fredholm operators: a (1 + 1)-dimensional example, arXiv:arXiv:1509.01356, Math.
Nachrichten, to appear.
[11] A. Carey, F. Gesztesy, G. Levitina, D. Potapov, F. Sukochev, and D. Zanin, Trace formulas
for a (1 + 1)-dimensional model operator, preprint, 2014.
[12] A. Carey, F. Gesztesy, G. Levitina, D. Potapov, F. Sukochev, and D. Zanin, On index theory
for non-Fredholm operators: a (2 + 1)-dimensional example, in preparation.
ALMOST ANALYTIC EXTENSIONS TO OPERATOR BOUNDS
19
[13] A. Carey, F. Gesztesy, G. Levitina, R. Nichols, D. Potapov, and F. Sukochev, Double operator
integral methods applied to continuity of spectral shift functions, arXiv:1511.07998.
[14] A. Carey, F. Gesztesy, G. Levitina, and F. Sukochev, On the index of a non-Fredholm model
operator, arXiv:1509.01580, Operators and Matrices, to appear.
[15] A. Carey, F. Gesztesy, D. Potapov, F. Sukochev, and Y. Tomilov, On the Witten index in
terms of spectral shift functions, arXiv:1404.0740, J. Analyse Math., to appear.
[16] G. Carron, T. Coulhon, and E. M. Ouhabaz, Gaussian estimates and Lp-boundedness of
Riesz means. J. Evol. Equ. 2, 299 -- 317 (2002).
[17] N. S. Claire, Spectral mapping theorem for the Davies -- Helffer -- Sjostrand functional calculus,
J. Math. Phys. Anal. Geom. 8, 221 -- 239 (2012).
[18] E. B. Davies, The functional calculus, J. London Math. Soc. (2) 52, 166 -- 176 (1995).
[19] E. B. Davies, Lp spectral independence and L1 analyticity, J. London Math. Soc. (2) 52,
177 -- 184 (1995).
[20] E. B. Davies, Spectral Theory and Differential Operators, Cambridge University Press, Cam-
bridge, 1995.
[21] M. Dimassi, Spectral shift function in the large coupling constant limit, Ann. H. Poincar´e 7,
513 -- 525 (2006).
[22] M. Dimassi and A. T. Duong, Trace asymptotics formula for the Schrodinger operators with
constant magnetic fields, J. Math. Anal. Appl. 416, 427 -- 448, (2014).
[23] M. Dimassi and V. Petkov, Spectral shift function and resonances for non-semi-bounded and
Stark Hamiltonians, J. Math. Pures Appl. 82, 1303 -- 1342 (2003).
[24] M. Dimassi and V. Petkov, Spectral shift function for operators with crossed magnetic and
electric fields, Rev. Math. Phys. 22, 355 -- 380 (2010).
[25] M. Dimassi and J. Sjostrand, Trace asymptotics via almost analytic extensions, in Partial
Differential Equations and Mathematical Physics, The Danish-Swedish Analysis Seminar,
1995, L. Hormander and A. Melin (eds.), Birkhauser, Basel, 1996, pp. 126 -- 142.
[26] M. Dimassi and J. Sjostrand, Spectral Asymptotics in the Semi-Classical Limit, London
Math. Soc. Lecture Note Series, Vol. 268, Cambridge University Press, Cambridge, 1999.
[27] M. Dimassi and M. Zerzeri, A tme-dependent approach for the study of spectral shift function,
C. R. Acad. Sci. Paris Ser. I 350, 375 -- 378 (2012).
[28] E. M. Dy'nkin, An operator calculus based upon the Cauchy -- Green formula, J. Sov. Math.
4, 329 -- 334 (1975).
[29] E. M. Dy'nkin, The pseudoanalytic extension, J. Analyse Math. 60, 45 -- 70 (1993).
[30] R. Froese, D. Hasler, and W. Spitzer, On the AC spectrum of one-dimensional random
Schrodinger operators with matrix-valued potentials, Math. Phys. Anal. Geom. 13, 219 -- 233
(2010).
[31] J. Frohlich, M. Griesemer, and I. M. Sigal, Spectral theory for the standard model of non-
relativistic QED, Comm. Math. Phys. 283, 613 -- 646 (2008).
[32] J. E. Gal´e, P. J. Miana, and T. Pytlik, Spectral properties and norm estimates associated to
(k)
the C
c
-functional calculus, J. Operator Theory 48, 385 -- 418 (2002).
[33] J. E. Gal´e and T. Pytlik, Functional calculus for infinitesimal generators of holomorphic
semigroups, J. Funct. Anal. 150, 307 -- 355 (1997).
[34] C. Gerard, Sharp propagation estimates for N -particle systems, Duke Math. J. 67, 483 -- 515
(1992).
[35] C. G´erard, A proof of the abstract limiting absorption principle by energy estimates, J. Funct.
Anal. 254, 2707 -- 2724 (2008).
[36] F. Gesztesy, Y. Latushkin, K. A. Makarov, F. Sukochev, and Y. Tomilov, The index formula
and the spectral shift function for relatively trace class perturbations, Adv. Math. 227, 319 --
420 (2011).
[37] M. Griesemer, Exponential decay and ionization thresholds in non-relativistic quantum elec-
trodynamics, J. Funct. Anal. 210, 321 -- 340 (2004).
[38] H. R. Grumm, Two theorems about Cp, Rep. Math. Phys. 4, 211 -- 215 (1973).
[39] L. Hormander, Fourier Integral Operators: Lectures at the Nordic Summer School of Math-
ematics, 1969.
[40] B. Helffer and J. Sjostrand, Equation de Schrodinger avec champ magn´etique et ´equaton de
Harper, in Schrodinger Operators, H. Holden and A. Jensen (eds.), Lecture Notes in Physics,
Vol. 345, Springer, Berlin, 1989, pp. 138 -- 197.
20
F. GESZTESY AND R. NICHOLS
[41] B. Helffer and J. Sjostrand, On diamagnetism and de Haas -- van Alphen effect, Ann. H.
Poincar´e A52, 303 -- 375 (1990).
[42] A. Jensen and S. Nakamura, Mapping properties of functions of Schrodinger operators be-
tween Lp-spaces and Besov spaces, in Spectral and Scattering Theory and Applications, K.
Yajima (ed.), Adv. Studies in Pure Math., Vol. 23, Math. Soc. Japan, Kinokuniya Company,
Tokyo, Japan, 1994, pp. 187 -- 209.
[43] A. Khochman, Resonances and spectral shift function for the semi-classical Dirac operator,
Rev. Math. Phys. 19, 1071 -- 1115 (2007).
[44] G. Levitina, D. Potapov, and F. Sukochev, private communication, January 2015.
[45] M. Martin and M. Putinar, Lectures on Hyponormal Operators, Operator Theory: Advances
and Applications, Vol. 39, Birkhauser, Basel, 1989.
[46] A. Melin and J. Sjostrand, Fourier integral operators with complex-valued phase functions,
in Fourier Integral Operators and Partial Differential Equations, J. Chazarain (ed.), Lecture
Notes in Math., Vol. 459, Springer, Berlin, 1975, pp. 120 -- 223.
[47] S. O'Rourke, D. Renfrew, and A. Soshnikov, On fluctuations of matrix entries of regular
functions of Wigner matrices with non-identically distributed entries, J. Theoret. Probab.
26, 750 -- 780 (2013).
[48] A. Pizzo, D. Renfrew, and A. Soshnikov, On finite rank deformations of Wigner matrices.
Ann. Inst. Henri Poincar´e Probab. Stat. 49, 64 -- 94 (2013).
[49] A. Pushnitski, The spectral flow, the Fredholm index, and the spectral shift function, in Spec-
tral Theory of Differential Operators: M. Sh. Birman 80th Anniversary Collection, T. Suslina
and D. Yafaev (eds.), AMS Translations, Ser. 2, Advances in the Mathematical Sciences, Vol.
225, Amer. Math. Soc., Providence, RI, 2008, pp. 141 -- 155.
[50] M. Reed and B. Simon, Methods of Modern Mathematical Physics. I: Functional Analysis,
revised and enlarged edition, Academic Press, New York, 1980.
[51] B. Simon, Trace Ideals and Their Applications, Mathematical Surveys and Monographs, Vol.
120, 2nd ed., Amer. Math. Soc., Providence, RI, 2005.
[52] J. Sjostrand and M. Zworski, Complex scaling and the distribution of scattering poles, J.
Amer. Math. Soc. 4, 729 -- 769 (1991).
[53] E. Skibsted, Smoothness of N -body scattering amplitudes, Rev. Math. Phys. 4, 619 -- 658
(1992).
[54] J. Weidmann, Linear Operators in Hilbert Spaces, Graduate Texts in Mathematics, Vol. 68,
Springer, New York, 1980.
[55] D. R. Yafaev, Mathematical Scattering Theory. General Theory, Amer. Math. Soc., Provi-
dence, RI, 1992.
[56] D. R. Yafaev, A trace formula for the Dirac operator, Bull. London Math. Soc. 37, 908 -- 918
(2005).
Department of Mathematics, University of Missouri, Columbia, MO 65211, USA
E-mail address: [email protected]
URL: http://www.math.missouri.edu/personnel/faculty/gesztesyf.html
Mathematics Department, The University of Tennessee at Chattanooga, 415 EMCS
Building, Dept. 6956, 615 McCallie Ave, Chattanooga, TN 37403, USA
E-mail address: [email protected]
URL: http://www.utc.edu/faculty/roger-nichols/index.php
|
1103.2090 | 1 | 1103 | 2011-03-10T17:53:30 | Random Series of Trace Class Operators | [
"math.FA",
"math.OA",
"math.PR"
] | In this lecture, we present some results on Gaussian (or Rademacher) random series of trace class operators, mainly due jointly with F. Lust-Piquard. We will emphasize the probabilistic reformulation of these results, as well as the open problems suggested by them. We start by a brief survey of what is known about the problem of characterizing a.s. convergent (Gaussian or Rademacher) series of random vectors in a Banach space. The main result presented here is that for certain pairs of Banach spaces $E,F$ that include Hilbert spaces (and type 2 spaces with the analytic UMD property), we have $$ R(E\hat\otimes F) =R(E)\hat\otimes F + E\hat\otimes R(F) $$ where $R(E)$ denotes the space of convergent Rademacher series with coefficients in $E$ and $E\hat\otimes F$ denotes the projective tensor product. | math.FA | math | Random Series of Trace Class Operators∗
by
Gilles Pisier
Texas A&M University
College Station, TX 77843, U. S. A.
and
Universit´e Paris VI
Equipe d'Analyse, Case 186, 75252
Paris Cedex 05, France
1
1
0
2
r
a
M
0
1
]
.
A
F
h
t
a
m
[
1
v
0
9
0
2
.
3
0
1
1
:
v
i
X
r
a
(1)
(2)
In this lecture, we would like to present some recent results on Gaussian (or Rademacher)
random series of trace class operators, (cf. mainly [17]). We will emphasize the probabilistic refor-
mulation of the results of [17], as well as the open problems suggested by these results. We will
start by a survey of what is known about the following problems.
Let B be a Banach space. Let (xn) be a sequence of elements of B. Let (gn) be a sequence of
i.i.d. Gaussian random variables.
Problem 1. Find a necessary and sufficient condition on the sequence (xn) for the a.s. convergence
in norm of the series
gnxn.
∞Xn=1
A similar question can be raised for series of the form P ±xn, where the signs are chosen
at random. More precisely, let (εn) be a sequence of i.i.d. ±1-value random variables such that
P (εn = +1) = P (εn = −1) = 1/2. We can ask
Problem 2. Find a necessary and sufficient on condition (xn) for the a.s. convergence of the series
εnxn.
∞Xn=1
εkxk. Let 0 ≤ p ≤ ∞. We will say that the sequence of partial
Let Sn = Pk≤n
gkxk and Rn = Pk≤n
sums {Sn} converges in Lp(B) if there is a Bochner-measurable B-valued random variable S∞ such
that kSn − S∞k → 0 in Lp. For p = 0, this corresponds to the convergence in probability of the
series (1). (Recall that the definition of a Bochner-measurable random variable is equivalent to
∗In Proceedings IV CLAPEM (Mexico City, September 1990).
1
requiring that the variable essentially takes its values in a separable subspace of B. Equivalently,
the image probability measure on B is a Radon measure.)
The following result is well known. In the Gaussian case, it goes back to [5] and [13], while the
Rademacher case goes back to [10] and [11] (cf. also [9] for the case p = 0).
Theorem 1. The following assertions are equivalent:
(i) The sequence of partial sums {Sn} converges a.s. in the norm of B.
(ii) The sequence {Sn} converges in Lp(B) for some 0 ≤ p < ∞.
(iii) The sequence {Sn} converges in Lp(B) for all 0 ≤ p < ∞.
The same equivalence holds with the partial sums {Rn} of the series (2) instead of the series
(1). Moreover, when the series (1) or (2) is convergent a.s. to a limit denoted by S, necessarily
there is a δ > 0 such that
Z exp(δkSk2) < ∞.
Corollary. For all 0 < p < q < ∞, the norms induced by Lp(B) and Lq(B) on the set of all a.s.
convergent series of the form (1) (resp. (2)) are equivalent. (In particular, they are all equivalent
to the norm induced by L2(B).)
the convergence in L2(B)
Remark. Of course, the preceding result gives an answer to Problem 1:
for instance is necessary and sufficient for the a.s. convergence. However, this does not help us
(except in the few simple cases below) since we usually cannot compute the norm in L2(B) for a
general Banach space B. We are looking for a condition as simple as possible characterizing the
a.s. convergence of (1).
Remark. Recently, Talagrand [30] gave a remarkable necessary and sufficient condition for the a.s.
continuity of Gaussian processes. This can be rephrased as a solution of Problem 1 in full generality,
as follows. Let K be the closed unit ball of the dual B∗ equipped with the topology σ(B∗, B) for
which it is compact. The series (1) converges a.s. in B iff the "majorizing measure condition" holds.
The latter means that there is a positive finite measure m on K such that if we set ∀t, s ∈ K
and ∀ε > 0 B(t, ε) = {s ∈ K d(t, s) < ε} then the condition can be stated as
d(t, s) =(cid:16)X hxn, t − si2(cid:17)1/2
m(B(t, ε))(cid:19)1/2
t∈KZ δ
0 (cid:18)Log
1
sup
lim
δ→0
dε = 0.
This completely solves the problem when we do not have any extra information on the Banach
space B. However, as the discussion below will show, the spaces B which arise in analysis are often
given with an additional structure (for instance a function space, or an operator space . . .) which
allows to find a simple very explicit necessary and sufficient condition. For instance, it is not easy
at all (although Talagrand did do it) to deduce the condition in Example 1 below (the Hilbert space
2
case) from the majorizing measure condition. In this lecture we want to concentrate on Banach
spaces B which are as "concrete" as possible. This is somehow the other end of the spectrum from
Talagrand's result which completely answers (and quite remarkably so) the "abstract" case.
One interesting feature of Theorem 1 is that it allows us to define the Banach space of all a.s.
convergent series of the form (1) or (2).
Indeed, given a fixed sequence (gn) of i.i.d. Gaussian
normal variables on a probability space (Ω, A, P), we define the space G(B) as the subspace of
L2(Ω, A, P; B) formed by all the convergent series of the form (1), and we equip it with the norm
induced by L2(B). Similarly, we define the space R(B) as the subspace of all convergent series of
the form (2) in L2(B).
It is not hard to show that G(B) (resp. R(B)) is a Banach space which coincides with the
closure in L2(B) of all the finite sums of the form (1) (resp. (2)).
By Theorem 1, our Problems 1 and 2 are the same as finding a "simple" description of the
Banach spaces G(B) and R(B).
Let us first review several known cases to illustrate what we mean by a "simple" characterization.
Example 1. If B is a Hilbert space, it is well known that (1) [resp. (2)] converges a.s. iff
Moreover, G(B) (resp. R(B)) can be identified with ℓ2(B) with equivalent norms.
X kxnk2 < ∞.
Example 2. If B = Lp(S, Σ, µ) for some measure space (S, Σ, µ) and if 1 ≤ p < ∞ (we insist that
p < ∞) then the series (1) [resp. (2)] is a.s. convergent iff
Moreover, there is a constant C (depending only on p) such that
xn(s)2!p/2
dµ(s) < ∞.
Z Xn
dµ(s)(cid:19)1/p
1
C(cid:18)Z (cid:16)X xn(s)2(cid:17)p/2
and similarly with the series (2) instead of (1).
≤(cid:13)(cid:13)(cid:13)X gnxn(cid:13)(cid:13)(cid:13)L2(B)
≤(cid:18)Z (cid:16)X xn(s)2(cid:17)p/2
dµ(s)(cid:19)1/p
,
In particular, if B = Lp(µ), the spaces G(B) and R(B) can be identified with the space Lp(µ; ℓ2),
with equivalent norms. This has been known for a long time, as a consequence of classical inequal-
ities. The "modern" way to prove this is to use Fubini's theorem and the Corollary of Theorem 1
and to work with the Lp(B) norm on G(B) or R(B) instead of the L2(B)-norm.
The basic examples above have been one of the motivation for the theory of type and cotype
of Banach spaces, (cf. [19]). A Banach space B is called of type p (1 ≤ p ≤ 2) if the condition
P kxnkp < ∞ is sufficient for the a.s. convergence of the series (2). Similarly, the space B is called
of cotype q (2 ≤ q < ∞) if the conditionP kxnkq < ∞ is necessary for the a.s. convergence of the
series (2).
3
With this terminology, we find that (by Example 1) if B is isomorphic to a Hilbert space then
B is of type 2 and cotype 2. By a well known result of Kwapie´n [12] the converse is also true, so
that only in spaces isomorphic to Hilbert do we have as simple a characterization as in Example 1.
From Example 2, it is rather easy to deduce (by Holder -- Minkowski) that Lp is of cotype max(p, 2)
and, if p 6= ∞, of type min(p, 2).
In our discussion, it is natural to ask for which spaces B the series of the form (1) and (2) are
equivalent in the sense that (1) converges a.s. iff (2) also does. This was answered in [19] as follows.
Theorem 2. The following properties of a Banach space B are equivalent:
(i) The space B is of cotype q for some q < ∞.
(ii) The a.s. convergence of a series of the form (1) is equivalent to the a.s. convergence of the
corresponding series of the form (2).
(iii) The spaces G(B) and R(B) can be naturally identified.
This is also equivalent to the nonexistence in B of a constant λ for which there exists a sequence
of finite dimensional subspaces Bn ⊂ B with Bn λ-isomorphic to ℓn
(Note that the natural
inclusion G(B) ⊂ R(B) holds for an arbitrary B, it is the converse inclusion which only holds if B
has a finite cotype.)
∞.
Since all the spaces that we will consider below fall into that category, we will sometimes only
state our results for R(B), which is more natural in view of the methods of proof, but the reader
should recall that the same results holds for G(B) as well.
Example 3. Let B be a Banach lattice of cotype q for some q < ∞. Without loss of generality,
we can assume that B is a Banach lattice of functions on some measure space (S, Σ, m). Consider
again xn ∈ B. Then the series (2) converges a.s. if
(3)
(4)
Moreover, there is a constant C such that
∈ B.
xn(·)2!1/2
∞X1
C(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X xn2(cid:17)12/2(cid:13)(cid:13)(cid:13)(cid:13)B
≤(cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)L2(B)
1
≤ C(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X xn2(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)B
.
This useful result can be found (in a different but equivalent or more general formulation) in the
work of Maurey [18]. It gives us an identification of R(B) and G(B) with the space B(ℓ2) of all
sequences (xn) such that (3) holds equipped with the norm appearing on the left side of (4). For
a more thorough discussion of the many inequalities satisfied by a Banach lattice of finite cotype,
(cf. e.g. the first chapter in the book [14]). In another direction, we refer the reader to [25, 26]
for a class of function spaces which, although they are not Banach lattices, satisfy an inequality
analogous to (4).
4
Recently, F. Lust-Piquard has obtained a striking noncommutative version of Example 2 (and
of Example 3 but we will not discuss this here, see [16]). Although her results are valid in a
more general framework, we will state them only for the simplest examples of noncommutative
Lp-spaces, namely the Schatten p-classes Cp. For 1 ≤ p < ∞, we denote by Cp the space of all
compact operators x : ℓ2 → ℓ2 such that trxp < ∞, where x = (x∗x)1/2. We equip this space
with the norm
kxkCp = (trxp)1/p.
It is well known that trxp = tr(x∗x)p/2 = tr(xx∗)p/2 for each x in Cp.
We will now consider our main Problems 1 and 2 with B = Cp.
The first part (Theorem 3) seems perhaps less surprising at first glance than the second one
(Theorem 4), because it appears more similar to the commutative case, but see below for a clarifi-
cation of the duality between the two cases.
Theorem 3 ([15]). Assume 2 ≤ q < ∞. Let B = Cq and let xn ∈ B. Then the series (2) (or (1))
converges a.s. iff the series
xnx∗
n are both convergent in the strong operator topology
Moreover, the norm in the space R(B) (or G(B)) is equivalent to the expression
x∗
nxn and
∞P1
∞P1
n(cid:17)q/2
nxn(cid:17)q/2
< ∞ and tr(cid:16)X xnx∗
tr(cid:16)X x∗
max
n!1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Cq
nxn!1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Cq
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞X1
∞X1
,(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
xnx∗
x∗
< ∞.
.
and satisfy
(5)
(6)
∞P1
(7)
(8)
Remark. Here the two conditions appearing in (5) are not equivalent and the theorem does not
hold if we drop one condition in (5).
The second part comes from [15] for 1 < p < 2 and from [17] for p = 1. Surprisingly, the
condition is different for the interval [1,2[ than for the interval ]2, ∞[.
Theorem 4 ([15, 17]). Assume 1 ≤ p ≤ 2. Let B = Cp and let xn ∈ B. Then the series (2) (or
(1)) converges a.s. iff there is a decomposition xn = yn + zn with yn, zn ∈ Cp such that
y∗
nyn and
∞P1
znz∗
n converge in the strong operator topology and satisfy
Moreover, if we define
< ∞.
nyn(cid:17)p/2
tr(cid:16)X y∗
(xn)p = inf((cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X y∗
n(cid:17)p/2
< ∞ and tr(cid:16)X znz∗
n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)Cp) .
nyn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)Cp
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X znz∗
5
where the infimum runs over all possible decompositions xn = yn + zn, then there is a constant C
(depending only on p) such that
1
C
(xn)p ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞X1
εnxn(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R(Cp)
≤ C(xn)p.
Remark. To illustrate the preceding result, let us consider the following illuminating special case
given in [15]. Let (εij ) be a collection of i.i.d. symmetric ±1 valued random variables as before, but
this time indexed by N × N instead of N.
Let A = (aij) be an infinite matrix with complex entries. Then, if 2 ≤ q < ∞, the random
matrix (εij, aij) is a.s. in Cq iff we have both
q/2
< ∞ and Xj Xi
aji2!q/2
< ∞.
Xi
Xj
Xj
aij2
bij2
Moreover, if 1 ≤ p < 2, then (εij aij) is a.s. in Cp iff aij can be decomposed as aij = bij + cij
with
Xi
p/2
< ∞ and Xj Xi
cij 2!p/2
< ∞.
(Note: the case 0 < p < 1 does not seem to be known.)
p + 1
Remark. The norm appearing in (8) is equivalent to the dual norm to the norm appearing in (6)
when 1
q = 1. This is a special case of the general duality between the intersection of two
spaces and the sum of their duals. However, the reader should be warned that (6) and (8) are not
equivalent norms when p = q unless p = q = 2!
Remark. In the preceding examples, when B = Lp or Cp, we can observe that if 1 < p < ∞ the
dual of G(B) (resp. R(B)) can be identified with G(B∗) (resp. R(B∗)). This property has been
extensively studied under the name of K-convexity. It means equivalently that there is a natural
bounded linear projection from L2(B) onto G(B) or R(B). We refer the reader to [20, 24] for more
details on that property. In particular, we proved (cf. [24]) that B is K-convex iff B is of type
p for some p > 1 or iff B does not contain a sequence of finite dimensional subspaces uniformly
isomorphic to ℓn
1 .
In the case p = 1, the space C1 is the space of all trace class operators which can also be viewed
as a tensor product. Let us recall the definition of the projective tensor product of two Banach
spaces. Let E, F be Banach spaces and let u =
xi ⊗ yi ∈ E ⊗ F with xi ∈ E, yi ∈ F . We define
nPi=1
kukΛ = inf( nX1
kxikkyik)
where the infimum runs over all possible representations of u as a finite sum as above. The projective
tensor product Eb⊗F is defined as the completion of E ⊗ F for this norm. It is well known that
6
the spaces of all bounded operators either B(E, F ∗) (from E into F ∗) or B(F, E∗) (from F into
E∗). We will see below that for a certain class of Banach spaces E and F , there is a rather simple
its dual (Eb⊗F )∗ can be identified with the space of all bounded bilinear forms on E × F , or with
characterization of the sequences xn ∈ Eb⊗F such that the series (1) or (2) converges a.s. in Eb⊗F .
Quite surprisingly the only known proof (at the moment) of these results uses the factorization of
operator valued analytic functions. In the case of C1 this factorization goes back to Sarason [29],
following classical work in the matrix case by Wiener -- Masani and Helson -- Lowdenslager. Recently,
these classical results have been generalized in [22] to the setting of type 2 Banach spaces (cf. also
[8, 21, 23]). To explain more clearly the connection with series of independent random variables as
in (1) and (2), we present our results on the infinite dimensional torus ∆ = TTT N, equipped with the
probability measure m which is the infinite product of copies of the normalized Haar measure on
the one dimensional torus TTT . Let B be a Banach space. Consider first a function F : TTT → B which
is in L1(TTT , dt; B). We will say that F is analytic if its Fourier transform vanishes on the negative
integers.
Let 1 ≤ p ≤ ∞. Now consider a function f in Lp(∆, m; B). Let us denote by En the conditional
expectation operator on Lp(∆, m; B) with respect to the σ-algebra generated by the first (n + 1)
coordinates (t0, t1, . . . , tn) on the infinite dimensional torus ∆ = TTT N. We will say that f is a
Hardy function if for each integer n ≥ 0 the function Enf is analytic as a function of the last
variable tn, more precisely, the function t → E0f (t) is analytic and for all n > 0, for any fixed
(t0, . . . , tn−1) ∈ TTT n the function t → Enf (t0, . . . , tn−1, t) is analytic in the above sense.
We will denote by H p(∆, B) the closed subspace of Lp(∆, m; B) formed by all the Hardy func-
tions f . This is the same as the set of all Hardy martingales considered in [4] and [6]. These Hardy
martingales are a convenient discretization of a certain kind of stochastic integrals which includes
analytic functions of the complex Brownian motion, (cf. e.g. [3, 32, 33]).
We will need the following.
Definition 1. Let E be a Banach space. We will say that E has the analytic UMD property if
there is a constant K such that for all g in H 2(∆, E) and for all choices of signs εn = ±1 the series
εn(Eng − En−1g) converges in H 2(∆, E) and satisfies
(cid:18)Z (cid:13)(cid:13)(cid:13)X εn(Eng − En−1g)(cid:13)(cid:13)(cid:13)
2
dm(cid:19)1/2
≤ K(cid:18)Z kgk2dm(cid:19)1/2
.
We will denote by K(E) the smallest constant K such that this holds.
We observe for further use that if we denote by µ the normalized uniform probability measure
on all the choices of signs ε = (εn)n≥0 ∈ {−1, 1}N, then by averaging (9) over all choices of signs
we find (denoting dng = Eng − En−1g)
∞Pn=1
(9)
(10)
hence by Fubini and a simple invariance argument, we also have for all points t ∈ TTT N
2
ZZ (cid:13)(cid:13)(cid:13)X εndng(t)(cid:13)(cid:13)(cid:13)
ZZ (cid:13)(cid:13)(cid:13)X εne−itn dng(t)(cid:13)(cid:13)(cid:13)
dm(t)dµ(ε) ≤ K 2Z kgk2dm,
dm(t)dµ(ε) ≤ 4K 2Z kgk2dm.
2
7
This property is slightly weaker than the so-called UMD property for which we refer to Burkholder's
paper [2]. Here we only require that "analytic" martingales (meaning those associated to a Hardy
function g) are unconditional in L2(∆, m; B). The UMD property corresponds to the same as (9)
but for arbitrary E-valued martingales. We refer to [6] for a discussion of the analytic UMD prop-
erty. Note that the spaces Lp or Cp are UMD for 1 < p < ∞, and that L1 has the analytic UMD
property (see [6]), while C1 fails it (cf. [8]).
Remark. To illustrate the possible applications of Theorems 3 and 4, let us consider their significance
for martaingales with values in Cq. Let (Mn) be a martingale with values in a UMD Banach space
B. Let us denote as usual dMn = Mn − Mn−1. We can define the vector valued version of the
maximal function and of the square function, as follows
M ∗ = sup
n
kMnk and S(M ) =sup
n Z (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
nX1
2
ǫkdMk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dµ(ǫ)
1/2
.
Then, it is known (cf. [2]) that for any 1 ≤ r ≤ ∞, there is a constant Kr such that for all B-valued
martingales (Mn) in Lr(B), we have
(Kr)−1kS(M )kr ≤ kM ∗kr ≤ KrkS(M )kr.
It was observed in [1] (cf. also [27] for more information) that the spaces Cq are all UMD spaces
when 1 < q < ∞. Thus, we can reformulate Theorems 3 and 4 in a martingale setting as follows.
Given a martingale (Mn) with values in Cq, for any ω in our probability space, let xn = dMn(ω)
and let us denote if 2 ≤ q < ∞ (resp. 1 ≤ p ≤ 2) by Sq(M )(ω) (resp. Sp(M )(ω)) the expression
(6) (resp. (8)). Then, we can state for all 1 ≤ r, q < ∞, there is a constant Kr,q such that for all
Cq-valued martingales (Mn), we have
(Kr,q)−1kSq(M )kr ≤ kM ∗kr ≤ Kr,qkSq(M )kr.
(11)
xn =
n ∈ E, zm ∈ F , ym ∈ E, zm
n ∈ F such that
case p = 1 of Theorem 4 to more general projective tensor products. Indeed, we have
Theorem 5. Assume that E, F are both Banach spaces of type 2 with the analytic UMD property.
Since it is well known that C1 can be identified with ℓ2b⊗ℓ2 it is natural to try to extend the
Let B = Eb⊗F and let xn ∈ Eb⊗F . Then the series (1) or (2) converges a.s. iff there are positive
scalars λm withP λm < ∞ and elements ym
∞Xm=1
n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn
R(Eb⊗F ) ≈ R(E)b⊗F + Eb⊗R(F ),
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L2(F )
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L2(E)
sup
m (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn
Equivalently, we have the isomorphic identification
λm[ym
n ⊗ zm + ym ⊗ zm
n ]
sup
m
kymkE ≤ 1, sup
m
kzmkF ≤ 1.
≤ 1,
sup
such that
ǫnym
ǫnzm
≤ 1,
(12)
8
and the norm of R(Eb⊗F ) is equivalent to the natural norm in the space R(E)b⊗F + Eb⊗R(F ) which
is defined as
(xn) = infnk(yn)kR(E) b⊗F + k(zn)kE b⊗R(F )o
where the infimum runs over all decompositions xn = yn + zn. (Another trivially equivalent norm
can be defined as the infimum ofP λm over all possible representations as in (11) above.)
Remark. The sum in (12) is not a direct sum. The meaning of (12) is that the three spaces
appearing in (9) are all naturally continuously injected into the set of all sequences of elements of
Eb⊗F and a sequence (xn) comes from R(Eb⊗F ) iff it can be decomposed as the sum of a sequence
in R(E)b⊗F and a sequence in Eb⊗R(F ).
For example, if we take E = F = ℓ2, then R(E) ≈ ℓ2(E) and R(F ) ≈ ℓ2(F ), and one can check
rather easily that k(yn)kℓ2(E) b⊗F (resp. k(zn)kE b⊗ℓ2(F )) can be identified with
(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X y∗
nyn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)E b⊗F(cid:18)resp. (cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X znz∗
n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)E b⊗F(cid:19) .
Thus, Theorem 5 extends the case p = 1 of Theorem 4.
The proof of Theorem 5 is based on a factorization property of functions of H 1(∆, Eb⊗F ) as
a convex hull of tensor products of bounded sets of functions in H 2(E) and H 2(F ). From the
viewpoint of Harmonic Analysis the space H p(∆, B) is nothing but the B-valued case of the H p-
space associated to the compact group ∆ with its dual group ZN ordered by the lexicographical
order, as is explained for instance in the chapter devoted to compact groups with ordered duals
in [28]. It is easy to see that if g, h are complex valued functions in H 2(∆), then the pointwise
product gh is in H 1(∆) with kghk1 ≤ kgk2khk2. Since the methods described in Rudin's book,
(which are due to Helson -- Lowdenslager) extend to the matrix valued case (cf. [7]) they allow us to
trivially modify the proofs of the Appendix B in the paper [22] (cf. also [23]) to obtain the following
statement (which could alternatively be phrased as a factorization of H 1-functions with values in
the space of all nuclear operators from E∗ into F ).
Theorem 6. Let E, F be Banach spaces for type 2. Then the natural product mapping
defines canonically a surjective norm one mapping
H 2(∆) × H 2(∆) → H 1(∆)
QE,F : H 2(∆, E)b⊗H 2(∆, F ) → H 1(∆, Eb⊗F ).
More explicitly, there is a constant C such that for any f in H 1(∆, Eb⊗F ) there are functions
gm ∈ H 2(∆, E) and hm ∈ H 2(∆, F ) such that
∀z ∈ ∆
and
(13)
f (z) =
gm(z) ⊗ hm(z)
∞X1
X kgmkH 2(∆,E)khmkH 2(∆,F ) ≤ Ckf kH 1(∆,E b⊗F ).
9
Using this result, it is not hard to complete the
Proof of Theorem 5. We first note that there is obviously a norm one inclusion
R(E)b⊗F + Eb⊗R(F ) → R(Eb⊗F ).
converse assuming (as we clearly may) that B is a complex Banach space. We will work with the
complex version of the series (2). (The corresponding series are sometimes called Steinhaus series.)
Equivalently, for every sequence (xn) of the form (11), the series P εnxn is in R(Eb⊗F ) with
norm kP εnxnkR(E b⊗F ) ≤ 2P λn. This is immediate from the definitions. We will prove the
Consider xn ∈ Eb⊗F such that the series S = P εnxn converges a.s. in Eb⊗F and consider the
function f : ∆ → B defined by
eitnxn.
f (t0, t1, . . .) =Xn≥0
Z∆
kf kdm ≤ 2kSkL1(E b⊗F ).
Then, it is easy to check (working with finite sums) that f converges a.s. and that
(14)
Moreover (since f is a polynomial of degree one in each complex variable) f is clearly in H 1(∆, B).
By Theorem 6, we can find gm, hm such that (13) holds.
Now let us denote
A simple calculation shows that
dnf = Enf − En−1f.
(15)
(16)
Let us set
xn =Z e−itnf (t0, t1, . . .)dm(t) =Xm Z e−itn gm ⊗ hmdm(t)
=Xm Z e−itn [dngm ⊗ hm + gm ⊗ dnhm]dm(t).
ym
n (t) = e−itn dngm(t),
zm(t) = hm(t),
ym(t) = gm(t),
zm
n (t) = e−itndnhm(t).
n (t) ∈ L2(∆, dm; R(E)) with
≤ 2K(E)kgmkH 2(∆,E),
(17)
andP εnzm
(18)
Then by (10) (the analytic UMD property) we haveP εnym
2
n (t) ∈ L2(∆, dm; R(F )) with
R(E)
dm(t)(cid:19)1/2
(cid:18)Z (cid:13)(cid:13)(cid:13)X εnym
n (t)(cid:13)(cid:13)(cid:13)
dm(t)(cid:19)1/2
(cid:18)Z (cid:13)(cid:13)(cid:13)X εnzm
n (t)(cid:13)(cid:13)(cid:13)
h (t)! ⊗ zm(t) + ym(t) ⊗ Xn
εnxn =Xn Z " Xn
εnym
2
R(F )
Xn
therefore it follows from (16) that we have
10
≤ 2K(F )khmkH 2(∆,F ),
n (t)!# dm(t)
εnzm
so that recalling (13), (14), (17) and (18) we find
(cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)R(E) b⊗F +E b⊗R(F )
≤ (2CK(E) + 2CK(F ))kf kH 1(∆,E b⊗F )
≤ 4(CK(E) + CK(F ))kSkR(E b⊗F )
so that we obtain the announced converse inclusion
R(Eb⊗F ) ⊂ R(E)b⊗F + Eb⊗R(F ),
and this concludes the proof.
Final Remarks.
(i) Note that Theorem 5 applies in the case E = Lp, F = Lq and 2 ≤ p, q < ∞. We do not know
what happens for 1 < p, q < 2 or even in the case E = Lp, F = L2 for 1 < p < 2.
(ii) Similarly it is easy to deduce from Theorem 5 that the space Lpb⊗Lq is of cotype max(p, q) if
2 ≤ p, q < ∞. However it remains an open problem whether Lpb⊗Lq is of finite cotype (cotype
2?) for 1 < p < 2 and 1 < q ≤ 2. For the case p = q = 2, the result goes back to [31].
(iii) The preceding Theorem 5 raises many other natural questions. One fascinating point is that
(12) appears like a derivation formula. It seems that R operates on the tensor product exactly
as a derivation does on a product of noncommuting objects.
Therefore, it is natural to ask the following two questions:
Is there a similar result for R(Eb⊗Fb⊗G)?
The guess is that we should find the sum
The most interesting open case is the case E = F = G = ℓ2. It is a long standing
R(E)b⊗Fb⊗G + Eb⊗R(F )b⊗G + Eb⊗Fb⊗R(G).
We can also ask about "second derivatives":
open question whether the space Eb⊗Fb⊗G has finite cotype in this case.
Is there an analogous result for R[R(Eb⊗F )]?
Here the guess is that we should have a formula analogous to the second derivative.
We must distinguish the first R (associated to a first sequence (ǫ1
(associated to a second sequence (ǫ2
R1 and R2 respectively. Then the guess is that we should have
n)) and the second one
n), independent of the first one) so let us denote them by
R1(R2(Eb⊗F )) ≈ R1(R2(E))b⊗F + R2(E)b⊗R1(F ) + R1(E)b⊗R2(F ) + Eb⊗R1(R2(F )).
Even in the case E = F = ℓ2, we could not check this.
11
References
[1] J. Bourgain, Vector valued singular integrals and the H 1 -- BMO duality, Probability Theory
and Harmonic Analysis, Chao and Woyczynski (eds.), Marcel Dekker, (1986), pp. 1 -- 19.
[2] D. Burkholder, A geometric characterization of Banach spaces in which martingale difference
sequences are unconditional, Ann. Probab. 9 (1981), 997 -- 1011.
[3] K. Carne, The algebra of bounded holomorphic margingales, J. Funct. Anal. 45 (1982), 95 -- 108.
[4] G. Edgar, Analytic martingale convergence, J. Funct. Anal. 69 (1986), 268 -- 280.
[5] X. Fernique, Int´egrabilit´e des vecteurs gaussiens, C.R. Acad. Sci. Paris A 270 (1970), 1698 --
1699.
[6] D.J.H. Garling, On martingales with values in a complex Banach space, Proc. Cambridge Phil.
Soc. 104 (1988), 399 -- 406.
[7] H. Helson and D. Lowdenslager, Prediction theory and Fourier series in several variables, Acta
Math. 99 (1958), 165 -- 202.
[8] U. Haagerup and G. Pisier, Factorization of analytic functions with values in noncommutative
L1-spaces, Canadian J. Math. 41 (1989), 882 -- 906.
[9] K. Ito and M. Nisio, On the convergence of sums of independent Banach space valued random
variables, Osaka J. Math. 5 (1968), 35 -- 48.
[10] J.P. Kahane, Some random series of functions, (1968), Heath Mathematical Monographs, Sec-
ond Edition, Cambridge University Press, (1985).
[11] S. Kwapie´n, A theorem on Rademacher series with vector valued coefficients, in Probability in
Banach Spaces, Springer Lecture Notes, no. 526, (1976), 157 -- 158.
[12] S. Kwapie´n, Isomorphic characterizations of inner product spaces by orthogonal series with
vector coefficients, Studia Math. 44 (1972), 583 -- 595.
[13] H. Landau and L. Shepp, On the supremum of a Gaussian process, Sankhya A 32 (1970),
369 -- 378.
[14] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces II, Springer Verlag, 1979.
[15] F. Lust-Piquard, In´egalit´es de Khintchine dans Cp, (1 < p < ∞), C.R. Acad. Sci. Paris 303
(1986), 289 -- 292.
[16] F. Lust-Piquard, A Grothendieck factorization theorem on 22-convex Schatten spaces. Israel
J. Math. 79 (1992), 331365.
12
[17] F, Lust-Piquard and G. Pisier, Noncommutative Khintchine and Paley inequalities, Ark. Mat.
29 (1991), no. 2, 241 -- 260.
[18] B. Maurey, Type et cotype dans les espaces munis de structure locale inconditionnelle,
S´eminaire Maurey -- Schwartz 73-75, Expos´e no. 24 -- 25, Ecole Polytechnique, Paris.
[19] 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 (1976), 45 -- 90.
[20] G. Pisier, Probabilistic methods in the geometry of Banach spaces, CIME Summer School,
June 1985, Springer Lecture Notes, no. 1206, (1986), 167 -- 241.
[21] G. Pisier, Factorization de fonctions analytiques `a valeurs op´erateurs, C.R. Acad. Sci. Paris
307 (1988), 955 -- 960.
[22] G. Pisier, Factorization of operator valued analytic functions, Adv. Math. 93 (1992), 61 -- 125.
[23] G. Pisier, Factorization of operator valued analytic functions and complex interpolation,
Festschrift in honor of I. Piatetski-Shapiro, Part II S. Gilbert, R. Howe and P. Sarnak (eds.),
Weizmann Science Press of Israel (1990), pp. 197 -- 220.
[24] G. Pisier, Holomorphic semi-groups and the geometry of Banach spaces, Ann. Math. 115
(1982), 375 -- 392
[25] G. Pisier, The dual J ∗ of the James space has cotype 2 and the Gordon -- Lewis property, Proc.
Cambridge Phil. Soc. 103 (1988), 323 -- 331.
[26] G. Pisier and Q. Xu, Random series in the real interpolation spaces between the spaces Vp,
GAFA, Springer Lecture Notes, no. 1267, (1987), 185 -- 209.
[27] J.L. Rubio de Francia, Martingale and integral transforms of Banach space valued functions,
Springer Lecture Notes, no. 1221, (1985), 195 -- 222.
[28] W. Rudin, Fourier Analysis on Group, Interscience, New York, 1962.
[29] D. Sarason, Generalized interpolation in H ∞, Trans. Amer. Math. Soc. 127 (1967), 179 -- 203.
[30] M. Talagrand, Regularity of Gaussian processes, Acta Math. 159 (1987), 99 -- 149.
[31] N. Tomczak-Jaegermann, On the moduli of convexity and smoothness and the Rademacher
averages of the trace classes Sp (1 ≤ p < ∞), Studia Math. 50 (1974), 163 -- 182.
[32] N. Varopoulos, The Helson-Szego theorem and Ap-functions for Brownian motion and several
variables, J. Funct. Anal. 39 (1980), 85 -- 121.
[33] N. Varopoulos, Probabilistic approach to some problems in complex analysis, Bull. Sci. Math.
105 (1981), 181 -- 224.
13
|
1609.09648 | 1 | 1609 | 2016-09-30T09:35:04 | Separably injective $C_\sigma$-spaces | [
"math.FA"
] | We show that a (complex) $C_\sigma$-space is separably injective if and only if it is linearly isometric to the Banach space $C_0(\Omega)$ of complex continuous functions vanishing at infinity on a substonean locally compact Hausdorff space $\Omega$. | math.FA | math |
SEPARABLY INJECTIVE Cσ -SPACES
CHO-HO CHU AND LEI LI
Abstract. We show that a (complex) Cσ-space is separably injective if and
only if it is linearly isometric to the Banach space C0(Ω) of complex continuous
functions vanishing at infinity on a substonean locally compact Hausdorff space
Ω.
1. Introduction
Recently, separably injective Banach spaces have been studied in depth by Avil´es,
Cabello S´anchez, Castillo, Gonz´alez and Moreno in [2, 3, 4], where one can find a
number of interesting examples of these spaces despite the scarcity of examples of
injective Banach spaces. In contrast to the fact that 1-injective Banach spaces are
isometric to the Banach space C(Ω) of continuous functions on a compact Hausdorff
space Ω [12, 14], 1-separably injective Banach spaces need not be complemented
in C(Ω).
In view of this, a natural question arises: when is a 1-complemented
subspace of C(Ω) separably injective? We address this question in this paper.
A Banach space V is 1-separably injective if every continuous linear map
T : Y −→ V on a closed subspace Y of a separable Banach space Z admits a norm
preserving extension to Z. It is known that C(Ω) itself is separably injective if and
only if Ω is an F -space [2]. Abelian C*-algebras with identity are of the form C(Ω)
for some Ω. The ones without identity can be represented as the algebra C0(S) of
complex continuous functions vanishing at infinity on a locally compact Hausdorff
space S and it has been shown lately that C0(S) is separably injective if and only
if S is substonean [7]. Following [11], we call S substonean if any two disjoint open
σ-compact subsets of S have disjoint compact closures. The compact substonean
spaces are exactly the F-spaces defined in [8, 18]. However, infinite discrete spaces
are F-spaces without being substonean. We refer to [13, Example 5] for an example
of a substonean space which is not an F-space.
Noting that the class of 1-complemented subspaces of C(Ω) is identical to that of
Cσ-spaces [15, Theorem 3] (see also Remark 2.1), our question amounts to asking
for a characterisation of separably injective Cσ-spaces. We give a complete answer
by showing that a Cσ-space is separably injective if and only if it is linearly isometric
to the function space C0(S) on a substonean locally compact Hausdorff space S.
In what follows, all Banach spaces are over the complex field and we will denote
by C0(K) the C*-algebra of complex continuous functions vanishing at infinity on
a locally compact Hausdorff space K. If K is compact, we omit the subscript 0.
Given a function g ∈ C0(K), we denote by g the complex conjugate of g.
1991 Mathematics Subject Classification. Primary 46B20, 17C65, 54G05.
Key words and phrases. Separably injective Banach space. Cσ-space. Substonean space.
The second author is supported by NSF (China) research grants No.11301285.
1
2
CHO-HO CHU AND LEI LI
Let T = {α ∈ C : α = 1} be the circle group. By a T-space, we mean a locally
compact Hausdorff space K equipped with a continuous group action
σ : (α, ω) ∈ T × K 7→ α · ω ∈ K.
A complex Banach space is called a complex Cσ-space if it is linearly isometric to a
function space of the form Cσ(K) for some T-space K, defined by
Cσ(K) = {f ∈ C0(K) : f (σ(α, ω)) = αf (ω), ∀ω ∈ K}.
We note that the definition of a complex Cσ-space in [16] requires K to be compact.
To achieve our result, we make substantial use of the Jordan algebraic structure
of Cσ(K). Indeed, although Cσ(K) lacks a C*-algebraic structure, it is equipped
with a triple product
{f, g, h} = f gh
(f, g, h ∈ Cσ(K))
which turns it into a JB*-triple with many useful Jordan properties.
First, let us give a brief introduction to JB*-triples which generalise C*-algebras.
For further references and the geometric origin of JB*-triples, we refer to [6, 17, 19].
A complex Banach space V is a JB*-triple if it admits a continuous triple product
{·, ·, ·} : V 3 −→ V
which is symmetric and linear in the outer variables, but conjugate linear in the
middle variable, and satisfies
(i) {x, y, {a, b, c}} = {{x, y, a}, b, c} − {a, {y, x, b}, c} + {a, b, {x, y, c}};
(ii) the operator a a : x ∈ V 7→ {a, a, x} ∈ V has real numerical range and
non-negative spectrum;
(iv) ka ak = kak2
for a, b, c, x, y ∈ V . We always have k{a, b, c}k ≤ kakkbkkck and (i) is called the
Jordan triple identity. A C*-algebra A is a JB*-triple with the triple product
{a, b, c} =
1
2
(ab∗c + cb∗a)
(a, b, c ∈ A).
More generally, the range of a contractive projection on a C*-algebra is a JB*-triple
(cf. [6, Theorem 3.3.1]), but not always a C*-algebra. An element e in a JB*-triple
is called a tripotent if e = {e, e, e}. Tripotents in C*-algebras are exactly the partial
isometries.
A subspace W of a JB*-triple V is called a subtriple if a, b, c ∈ W implies
{a, b, c} ∈ W . Closed subtriples of a JB*-triple are JB*-triples in the inherited
norm and triple product. A triple ideal of V is a subspace J ⊂ V such that
{a, b, c} ∈ J whenever one of a, b and c belongs to J. Given a closed triple ideal
J ⊂ V , the quotient space V /J is a JB*-triple in the triple product
{a + J, b + J, c + J} := {a, b, c} + J.
Two elements a, b ∈ V are said to be orthogonal to each other if a b = 0, where
a b is the continuous linear map x ∈ V 7→ {a, b, x} ∈ V . Two subspaces I, J ⊂ V
are orthogonal if I J := {a b : a ∈ I, b ∈ J} = {0}.
The bidual V ∗∗ of a JB*-triple V carries a natural structure of a JB*-triple, with
a unique predual, in which the triple product is separately weak* continuous and
the natural embedding of V into V ∗∗ identifies V as a subtriple of V ∗∗. Given a
closed triple ideal I ⊂ V , the bidual I ∗∗ embeds as a weak* closed triple ideal in
SEPARABLY INJECTIVE Cσ-SPACES
3
V ∗∗, which can be decomposed into an ℓ∞-sum V ∗∗ = I ∗∗ ⊕∞ J for some weak*
closed triple ideal J ⊂ V ∗∗, orthogonal to I ∗∗ [6, lemma 3.3.16].
A linear map ϕ : V −→ W between two JB*-triples is called a triple homo-
morphism if {ϕ(a), ϕ(b), ϕ(c)} = ϕ{a, b, c} for a, b, c ∈ V . The triple isomorphisms
between V and W are exactly the surjective linear isometries (cf. [6, Theorem 3.1.7,
Theorem 3.1.20]).
A JB*-triple V is called abelian if its triple product satisfies
{{x, y, z}, u, v} = {x, {y, z, u}, v} = {x, y, {z, u, v}}
for all x, y, z, u, v ∈ V . An abelian C*-algebra is an abelian JB*-triple and so is
Cσ(K) in the triple product defined above. In fact, Cσ(K) is a closed subtriple of
C0(K).
By [5, Lemma 2.2, Theorem 3.7], an abelian closed subtriple V of a C*-algebra
admits a composition series (Jλ)0≤λ≤µ of closed triple ideals, indexed by ordinals
λ, such that the quotient Jλ+1/Jλ is linearly isometric to an abelian C*-algebra, for
λ < µ. We recall that (Jλ)0≤λ≤µ is called a composition series if J0 = {0}, Jµ = V
and for a limit ordinal λ ≤ µ, the ideal Jλ is the closure of Sλ′<λ Jλ′ .
2. Jordan structure in Cσ-spaces
We will make use of the abelian JB*-triple structure of Cσ-spaces to derive our
result. To pave the way, we first present some detailed analysis of this structure.
Let V be an abelian closed subtriple of a C*-algebra (e.g. a Cσ-space) in this
section. One can consider it as as a subtriple of its bidual V ∗∗ via the natural
embedding v ∈ V 7→ bv ∈ V ∗∗, where bv(ψ) = ψ(v) for ψ ∈ V ∗. By [5, 9], V ∗∗
is (isometric to and identified as) an abelian von Neumann algebra with identity
denoted by 1 and involution z ∈ V ∗∗ 7→ z∗ ∈ V ∗∗. The triple product in V ∗∗ is
given by {a, b, c} = ab∗c. Each ψ ∈ V ∗ can be viewed naturally as a functional of
V ∗∗. If ψ is a positive functional of V ∗∗, then ψ(z∗) = ψ(z) for each z ∈ V ∗∗. A
positive functional ψ ∈ V ∗ is called a normal state of V ∗∗ if ψ(1) = 1. It is called
pure if it is an extreme point of the norm closed convex set of normal states in V ∗.
Let S be the set of all pure normal states of V ∗∗, which are exactly the multiplica-
tive normal states of V ∗∗. Given a projection p ∈ V ∗∗ and s ∈ S, we have s(p) = 0 or
1 since s(p) = s(p2) = s(p)2. If u ∈ V ∗∗ is unitary, then 1 = s(1) = s(u∗u) = s(u)2
for all s ∈ S. We equip S with the weak* topology of V ∗ and call it the pure normal
state space of V ∗∗.
The nonzero triple homomorphisms from V to C are exactly the set K = ext V ∗
1
1 , where K ∪ {0} is weak*-compact [9,
of extreme points of the dual unit ball V ∗
Proposition 2.3, Corollary 2.4] and S ⊂ K.
For each ω ∈ K and tripotent c ∈ V ∗∗, we have
ω(c) = ω(cc∗c) = ω(c)ω(c)2
which implies ω(c) = 0 or ω(c) ∈ T. We note that K is a T-space with the natural
T-action
σ : (α, ω) ∈ T × K 7→ αω ∈ K
and we have K = {αs : α ∈ T, s ∈ S}.
In fact, each ω ∈ K has a unique
representation ω = αs for some α ∈ T and s ∈ S, where s = ω(1)ω. By [9,
Theorem 1], the map
(2.1)
v ∈ V 7→ bvK ∈ Cσ(K)
4
CHO-HO CHU AND LEI LI
is a surjective linear isometry, which enables us to identify V with the Cσ-space
Cσ(K).
Remark 2.1. Let π : C(Ω) −→ C(Ω) be a contractive projection. Then its
image π(C(Ω)) is an abelian closed subtriple in some C*-algebra [10] and hence the
previous discussion implies that it is a Cσ-space.
For each a ∈ V \{0}, let V (a) be the JB*-subtriple generated by a in V . Then
there is a surjective linear isometry and triple isomorphism
(2.2)
φ : C0(Sa) → V (a) ⊂ V
which identifies V (a) with the abelian JB*-triple C0(Sa) of continuous functions
vanishing at infinity on the triple spectrum Sa ⊂ (0, kak], where Sa ∪{0} is compact
[6, Theorem 3.1.12].
Let φ∗∗ : C0(Sa)∗∗ → V (a)∗∗ be the bidual map and let i be the identity in
the von Neumann algebra C0(Sa)∗∗. Then e = φ∗∗(i) is the identity in the von
Neumann algebra V (a)∗∗ with product and involution given by
x · y = {x, e, y},
x ∈ V (a)∗∗ 7→ {e, x, e} ∈ V (a)∗∗.
While this abelian von Neumann algebraic structure of V (a)∗∗ will be assumed
throughout, it should be noted that V (a)∗∗ need not be a subalgebra of V ∗∗ in its
natural embedding. Nevertheless, V (a)∗∗ can always be considered as a subtriple
of V ∗∗ and the identity e ∈ V (a)∗∗ is a tripotent in V ∗∗ satisfying
{e, a, e} = {φ∗∗(i), φ∗∗(ιa), φ∗∗(i)} = φ∗∗{i, ιa, i} = a.
(2.3)
For each ρ ∈ K, viewed as a complex-valued triple homomorphism on V (a)∗∗, we
have ρ(e) = 0 if and only if ρ(a) = ρ{e, e, a} = 0.
The norm closed triple ideal Ja generated by a in V contains V (a) and is the
norm closure of {a, V, a}. It has been shown in [5, Lemma 2.2] that Ja is linearly
isometric to the abelian C*-algebra C0(Xa) of continuous functions vanishing at
infinity on a locally compact Hausdorff space Xa. We need some detail here for
later application. In fact, Ja is an abelian C*-algebra with the same product and
involution of V (a)∗∗, and e ∈ V (a)∗∗ ⊂ J ∗∗
a , which can be
a
seen from the following computation using (2.3).
is the identity of J ∗∗
{{a, x, a}, e, {a, y, a}} = {a, {x, {a, e, a}, y}, a} ∈ {a, V, a};
{e, {a, x, a}, e} = ea∗xa∗e = ee∗ae∗xe∗ae∗e = ae∗xe∗a = {x, a, a} ∈ Ja.
Given ρ ∈ K = ext V ∗
1 with ρ(e) = 1, its restriction ρJa ∈ J ∗
a . Conversely, each pure normal state ϕ of J ∗∗
a
a and can be extended to an extreme point eϕ ∈ ext V ∗
1
a is a pure normal
is an extreme point of
state of J ∗∗
the closed unit ball of J ∗
satisfying eϕ(e) = 1. Let
denote the pure normal state space of J ∗∗
topology w(J ∗
a , Ja) of J ∗
a .
Xa = {ρJa : ρ ∈ K, ρ(e) = 1}
a which is locally compact in the weak*
We note that for each ρ ∈ K, we have ρ(a) = 0 if and only if ρ(V (a)) = {0},
which in turn is equivalent to ρ(Ja) = ρ({a, V, a}) = {0}.
Lemma 2.2. In the above notation, the set K(e) = {ρ ∈ K : ρ(e) = 1} with the
relative weak* topology of V ∗ is homeomorphic to Xa = {ρJa : ρ ∈ K, ρ(e) = 1} in
the topology w(J ∗
a , Ja). In particular, K(e) is weak* locally compact in K.
SEPARABLY INJECTIVE Cσ-SPACES
5
Proof. We show that the restriction map ρ ∈ K(e) 7→ ρJa ∈ Xa is a homeomor-
phism in these topologies. It is clearly continuous and surjective. Given ρ, ρ′ ∈ K
such that ρJa = ρ′Ja , then ρ(a) = ρ′(a) 6= 0 since ρ(e) = ρ′(e) = 1. For any
v ∈ V , we have {a, a, v} ∈ Ja and hence ρ(a)2ρ(v) = ρ{a, a, v} = ρ′{a, a, v} =
ρ′(a)2ρ′(v), giving ρ(v) = ρ′(v). This proves injectivity of the map.
Finally, to show that the inverse of the map is continuous, let (ργJa) be a net
converging to ρJa ∈ Xa. Then again, for each v ∈ V , we have ργ(a) → ρ(a) 6= 0 and
ργ(a)2ργ(v) = ργ{a, a, v} → ρ{a, a, v} = ρ(a)2ρ(v), which implies ργ(v) → ρ(v),
proving continuity.
(cid:3)
Remark 2.3. The above lemma enables us to identify the pure normal state space
Xa with K(e) and write Xa = {ρ ∈ K : ρ(e) = 1}.
We retain the above notation in the sequel.
3. Separably injective Cσ-spaces
We characterize separably injective Cσ-spaces in this section. Throughout, let
V be a Cσ-space. We will identify V , as in the previous section, with the Cσ-space
Cσ(K), where K = ext V ∗
1 is the set of nonzero triple homomorphisms from V to
C.
Lemma 3.1. Let V be separably injective. Given a ∈ V of unit norm and the
identity e ∈ V (a)∗∗, let K(e) = {ρ ∈ K : ρ(e) = 1}. Then there exists an element
va ∈ V such that K(e) ⊂ Ka where
and Ka is weak* compact in V ∗.
Ka = {ρ ∈ K : ρ(va) = 1}
Proof. Since Ka is weak* closed in K ∪ {0}, it is weak* compact. Let
φ : C0(Sa) → V (a) ⊂ V
be the embedding in (2.2), where the triple spectrum Sa ⊂ (0, 1] can be identified,
via the evaluation map as usual, with the pure normal state space of C0(Sa)∗∗.
Let χa be the constant function on Sa with value 1 and consider the separable
subspace C0(Sa) + Cχa of ℓ∞(Sa). By separable injectivity of V , the embedding
φ : C0(Sa) → V (a) ⊂ V admits a norm preserving extension Φ : C0(Sa)+Cχa → V .
Let
va = Φ(χa) ∈ V.
To complete the proof, we show that ρ(va) = 1 for each ρ ∈ K(e).
We first observe that the sequence (rn) of odd roots of the identity function ιa
in C0(Sa) converges pointwise to the function χa. Let un = φ(rn) ∈ V (a).
Let ρ ∈ K(e). Then the map ρ ◦ φ : C0(Sa) → C is a pure normal state of
C0(Sa)∗∗. Hence we have ρ(un) = ρ ◦ φ(rn) ∈ [0, 1] and limn ρ(un) = 1.
The norm preserving extension Φ satisfies kΦ(χa)k ≤ 1 and
kΦ(χa) − 2unk = kΦ(χa) − 2φ(rn)k = kΦ(χa) − 2Φ(rn)k ≤ kχa − 2rnk ≤ 1.
It follows that ρ(Φ(χa)) − 2ρ(un) ≤ 1 for all n, which implies
Now ρ(Φ(χa)) ≤ 1 gives ρ(va) = ρ(Φ(χa)) = 1.
(cid:3)
ρ(Φ(χa)) − 2 ≤ 1.
6
CHO-HO CHU AND LEI LI
Our next task is to show that a separably injective Cσ-space V is actually linearly
isometric to an abelian C*-algebra. We adopt the following strategy. Since V is
abelian, it has been noted in Section 1 that there is a composition series (Jλ)0≤λ≤µ
of closed triple ideals in V such that for each ordinal λ < µ, the quotient Jλ+1/Jλ
is linearly isometric to the C*-algebra C0(Xλ) of continuous functions vanishing at
infinity on a locally compact Hausdorff space Xλ, and V ∗∗ is linearly isometric to
λ<µ(Jλ+1/Jλ)∗∗ = Lℓ∞
the ℓ∞-sum Lℓ∞
λ<µ C0(Xλ)∗∗. By the uniqueness of predual,
V ∗ is linearly isometric to the ℓ1-sum
ℓ1
M
λ<µ
(Jλ+1/Jλ)∗ =
ℓ1
M
λ<µ
C0(Xλ)∗.
Given that V is separably injective, we will refine this construction to show that V
is isometric to the abelian C*-algebra
C0(Xλ).
c0M
λ<µ
Let V be separably injective and let a ∈ V be of unit norm. Consider the
closed subtriple V (a) generated by a in V as well as the closed triple ideal Ja,
the latter is linearly isometric to the C*-algebra C0(Xa) as shown before, where
Xa = {ρ ∈ K : ρ(ea) = 1} is weak* locally compact by Lemma 2.2 and Remark 2.3,
and ea is the identity of the von Neumann algebra J ∗∗
a . By separable injectivity
and Lemma 3.1, there exists va ∈ V such that ρ(va) = 1 for each ρ ∈ Xa.
If Ja = V , then we are done. Otherwise we have the ℓ∞-sum
V ∗∗ = J ∗∗
a ⊕∞ (J ∗∗
a )
where (J ∗∗
a )
is, J ∗∗
(J ∗∗
a
identifies (J ∗∗
is a nonzero weak* closed triple ideal in V ∗∗, orthogonal to J ∗∗
a , that
a ) = {0} (cf. [6, Lemma 3.3.16]). The quotient map V ∗∗ → V ∗∗/J ∗∗
a
a ) with the quotient V ∗∗/J ∗∗
a and we can write
(3.1)
V ∗∗ = J ∗∗
a ⊕∞ (V ∗∗/J ∗∗
a ) = C0(Xa)∗∗ ⊕∞ (V /Ja)∗∗.
We have the ℓ1-sum V ∗ = C0(Xa)∗ ⊕1 (V /Ja)∗ where Ja is an M-ideal in V .
Consider the quotient map
x ∈ V 7→ [x] := x + Ja ∈ [V ] := V /Ja
which maps the closed unit ball of V onto the closed unit ball of V /Ja (cf. [1,
Corollary 5.6]).
Pick [b] = b + Ja ∈ [V ] with unit norm. Let V ([b]) and J[b] be respectively the
closed subtriple and triple ideal generated by [b] in [V ]. We can repeat the previous
arguments in the setting V ([b]) ⊂ J[b] ⊂ [V ] to deduce that J[b] is linearly isometric
to some abelian C*-algebra C0(X[b]) with
X[b] = {θ ∈ ext [V ]∗
1 : θ(e[b]) = 1}
[b] ⊂ [V ]∗∗ = V ∗∗/J ∗∗
a ) with e[b] = eeb + J ∗∗
where X[b] is locally compact in the weak* topology of [V ]∗ by Lemma 2.2 and
e[b] is the identity of J ∗∗
a , which identifies with a tripotent
eeb ∈ (J ∗∗
a . Moreover, the quotient JB*-triple [V ] = V /Ja is
abelian and seprably injective by [2, Proposition 4.6], and hence Lemma 3.1 implies
that there exists v[b] = evb + Ja ∈ [V ] = V /Ja such that θ(v[b]) = 1 for each θ ∈ X[b].
1 consists of nonzero complex-valued triple homomorphisms on
[V ], which can be lifted to nonzero complex triple homomorphisms on V via the
The set ext [V ]∗
SEPARABLY INJECTIVE Cσ-SPACES
7
quotient map and we have
where ¯ρ(x + Ja) := ρ(x). Hence we have
ext [V ]∗
1 = { ¯ρ : ρ ∈ K, ρ(Ja) = {0}}
X[b] = { ¯ρ : ρ ∈ K, ρ(a) = 0, ρ(eeb) = 1}
and ¯ρ ∈ X[b] implies ρ(b) = ¯ρ([b]) 6= 0 and ρ(evb) = 1.
A weak* convergent net in [V ]∗ lifts to a weak* convergent net in V ∗ via the
quotient map. Considering X[b] ⊂ C0(X[b])∗ = J ∗
[b] and Lemma 2.2, we see that a
net (¯ργ) in X[b] converges to ¯ρ ∈ X[b] in the weak* topology of C0(X[b])∗ if and only
if (ργ) weak* converges to ρ in K. The homeomorphism
¯ρ ∈ X[b] 7→ ρ ∈ {ρ ∈ K : ρ(a) = 0, ρ(eeb) = 1}
enables us to identify these two spaces. We note that
TXa ∩ TX[b] ⊂ TXa ∩ ext [V ]∗
(3.2)
1 = ∅
where ρ(a) 6= 0 for all ρ ∈ Xa.
In the ℓ1-sum
V ∗ = J ∗
(3.3)
each ω ∈ V ∗ admits a decomposition ω = ω1 + ω2 in V ∗ with ω2(Ja) = {0} and
kωk = kω1Ja k + kω2k, which provides the identification of ω as an element (eω1, eω2)
a ⊕1 (V /Ja)∗ = C(Xa)∗ ⊕1 [V ]∗,
in the ℓ1-sum, defined by
eω1 = ω1Ja ∈ J ∗
a
Hence for an extreme point ω ∈ ext V ∗
and eω2([·]) = ω2(·) ∈ (V /Ja)∗.
1 = K, we have eω1 = 0 or eω2 = 0.
a , Ja)-topology and the net (eω2
γ) converges to eω1 in the w(J ∗
Given a net (ωγ) in V ∗ weak* converging to a limit ω ∈ V ∗, it can be seen that
γ) converges
the net (eω1
to eω2 in the weak* topology of (V /Ja)∗. In particular, if the net (ωγ) is in K and
γ) to eωj implies that
ω ∈ K with eωj 6= 0 (j ∈ {1, 2}), then the convergence of (eωj
γ 6= 0 eventually, and hence eωj ′
eωj
The closed unit ball of the ℓ1-sum in (3.3) has extreme points
γ = 0 for j′ 6= j eventually.
(TXa, 0) ∪ (0, ext [V ]∗
1) := {(ω, 0) : ω ∈ TXa} ∪ {(0, ω) : ω ∈ ext [V ]∗
1}
and in the identification of (3.2), we have the disjoint union K = TXa ∪ ext [V ]∗
1.
Given a net (ωγ) in K weak* converging to some ω ∈ K, and given either ω ∈ TXa
or ω ∈ ext [V ]∗
1, the above observation implies that ωγ belongs to the same set
eventually.
Observe that [Jb] = J[b] = (Jb + Ja)/Ja and if J[b] 6= [V ], we have the ℓ∞-sum
V ∗∗ = J ∗∗
a ⊕∞((Jb+Ja)/Ja)∗∗⊕∞([V ]/J[b])∗∗ = C0(Xa)∗∗⊕∞C0(X[b])∗∗⊕∞([V ]/J[b])∗∗
where the quotient JB*-triple [[V ]] := [V ]/J[b] is separably injective, ρ(va) = 1 for
ρ ∈ Xa and ρ(evb) = 1 for ρ ∈ X[b].
The closed unit ball of the ℓ1-sum
V ∗ = C0(Xa)∗ ⊕1 C0(X[b])∗ ⊕1 ([V ]/J[b])∗
has extreme points
(TXa, 0, 0) ∪ (0, TX[b], 0) ∪ (0, 0, ext [[V ]]∗
1)
and we have the disjoint union K = ext V ∗
1. Given a
net (ωγ) in K weak* converging to some ω ∈ K, if ω belongs to one of the three
1 = TXa ∪ TX[b] ∪ ext [[V ]]∗
8
CHO-HO CHU AND LEI LI
sets above, then repeating the arguments as before, ωγ belongs to the same set
eventually.
Now transfinite induction together with separable injectivity yields a composition
series (Jλ)0≤λ≤µ of closed triple ideals in V , with vλ ∈ V , such that V ∗ is linearly
isometric to, and identifies with, the ℓ1-sum
ℓ1
M
(Jλ+1/Jλ)∗ =
λ<µ
ℓ1
M
λ<µ
C0(Xλ)∗
where ρ(vλ) = 1 for each ρ ∈ Xλ, and the pure normal state space Xλ of (Jλ+1/Jλ)∗∗
identifies with the set
in which eeλ is the identity of (Jλ+1/Jλ)∗∗. In this identification, we have the disjoint
union
{ρ ∈ K : ρ(Jλ) = {0}, ρ(eeλ) = 1}
K = [
TXλ
λ<µ
and for a weak* convergent net (ωγ) in K with limit ω ∈ TXλ for some λ, we
have (ωγ) in TXλ eventually. As a consequence of Lemma 2.2, in the identification
Xλ ⊂ C0(Xλ)∗ and Xλ ⊂ K, the weak* convergence in C0(Xλ)∗ of a net (ργ) in
Xλ to ρ ∈ Xλ is the same as the weak* convergence in K.
Lemma 3.2. Given that V is separably injective and in the above notation, the
subset Xλ ∪ {0} of K ∪ {0} is weak* compact for all λ < µ and also, TXλ is
relatively weak* open in K ∪ {0}.
Proof. Let (ργ) be a net in Xλ weak* converging to a nonzero limit ω ∈ V ∗. Then
ω ∈ K and by the above remark, we must have ω ∈ TXλ, say ω = αρ with α ∈ T
and ρ ∈ Xλ. Since Xλ is contained in the weak* compact set {ρ′ ∈ K : ρ′(vλ) = 1},
it follows that α = αρ(vλ) = limγ ργ(vλ) = 1 and ω = ρ ∈ Xλ. This proves that
Xλ ∪ {0} is weak* closed in K ∪ {0} and hence weak* compact.
For the second assertion, let (ωγ) be a net in (K ∪ {0})\TXλ weak* converging
to some ω ∈ K. Then again ω /∈ TXλ for otherwise, the previous remark implies
that ωγ belongs to TXλ eventually which is impossible.
(cid:3)
The above construction enables us to show that a separably injective Cσ-space
is isometric to an abelian C*-algebra.
Theorem 3.3. Let V be a separably injective Cσ-space. Then V is linearly iso-
metric to an abelian C*-algebra.
Proof. Let K = ext V ∗
1 and as shown previously, we have the ℓ1-sum
V ∗ =
ℓ1
M
λ<µ
C0(Xλ)∗
with the disjoint union K = Sλ<µ
TXλ. For each λ < µ, there is an element vλ ∈ V
such that ρ(vλ) = 1 for all ρ ∈ Xλ. We show that V is linearly isometric to the
c0-sum
C0(Xλ) which would complete the proof.
c0M
λ<µ
We continue to identify V with Cσ(K) in (2.1). By Lemma 3.2, each f ∈ Cσ(K)
restricts to a continuous function f Xλ ∈ C0(Xλ).
SEPARABLY INJECTIVE Cσ-SPACES
9
We show that the map
f ∈ V ≈ Cσ(K) 7→ (f Xλ) ∈
c0M
λ<µ
C0(Xλ)
is a surjective linear isometry.
To see that (f Xλ) indeed belongs to the c0-sum, we need to show (kf Xλk) ∈
c0(Λ), where Λ = [0, µ). Let ε > 0. By Lemma 3.2, for each λ ∈ Λ, the set TXλ is
relatively weak* open in K ∪ {0}. Since f vanishes at infinity on K, the set
Uε = {ω ∈ K : f (ω) < ε} ∪ {0}
is a relatively weak* open neighbourhood of 0 in the one-point compactification
K ∪ {0} of K. We have
K ∪ {0} = [
TXλ ∪ Uε
and by weak* compactness, there are finitely many λ1, . . . , λn such that
λ∈Λ
K ∪ {0} ⊂ TXλ1 ∪ · · · ∪ TXλn ∪ Uε.
It follows that kf Xλk ≤ ε for λ /∈ {λ1, . . . , λn} which proves (kf Xλk) ∈ c0(Λ).
Since K = ∪λT Xλ and each f ∈ Cσ(K) satisfies f (αω) = αf (ω) for α ∈ T and
ω ∈ K, it is evident that the map is a linear isometry.
It remains to show that the map is surjective. Let (gλ) ∈
function f : K → C by
c0M
λ∈Λ
C0(Xλ). Define a
f (ω) = αgλ(ρλ)
for ω = αρλ ∈ TXλ.
The function f is well-defined since the sets {TXλ}λ are mutually disjoint and each
ω ∈ K has a unique representation ω = αρ ∈ TXλ for if αρ = βσ ∈ TXλ, we
have α = αρ(vλ) = βσ(vλ) = β, where Xλ is contained in the weak* compact set
{ρ′ ∈ K : ρ′(vλ) = 1}.
We complete the proof by showing f ∈ Cσ(K). We have readily f (αω) = αf (ω)
for α ∈ T and ω ∈ K.
For continuity, let (ωγ) be a net weak* converging to ω ∈ K and say, ω =
αρ ∈ TXλ for some λ. By a previous remark, the net (ωγ) is in TXλ eventually.
Therefore we have ωγ = αγργ with αγ ∈ T and ργ ∈ Xλ eventually. It follows that
eventually αγ = αγργ(vλ) → αρ(vλ) = α and ργ → ρ. Hence we have
lim
γ
f (ωγ) = lim
γ
αγgλ(ργ) = αgλ(ρ) = f (ω).
Finally, for any ε > 0, there are finitely many λ1, . . . , λn such that kgλk < ε
for λ /∈ {λ1, . . . , λn}. For each λj with j = 1, . . . , n, there is a weak* compact set
Ej ⊂ Xλj such that {ρ ∈ Xλj : gλj (ρ) ≥ ε} ⊂ Ej. This gives
T{ρ ∈ Xλj : gλj (ρ) ≥ ε} ⊂ ∪n
{ω ∈ K : f (ω) ≥ ε} = ∪n
j=1
TEj ⊂ K
j=1
where the finite union ∪j TEj is weak* compact and therefore f ∈ C0(K).
(cid:3)
Finally, by the characterisation of separably injective abelian C*-algebras in [7,
Theorem 3.5], together with Theorem 3.3, we conclude with the following main
result of the paper.
10
CHO-HO CHU AND LEI LI
Theorem 3.4. Let V be a Cσ-space. The following conditions are equivalent.
(i) V is separably injective.
(ii) V is linearly isometric to the Banach space C0(S) of complex continuous
functions vanishing at infinity on a substonean locally compact Hausdorff
space S.
References
[1] E.M. Alfsen and E.G. Effros, Structure in real Banach spaces, Ann. of Math. 96 (1972)
98-173.
[2] A. Avil´es, F. Cabello S´anchez, J. M.F. Castillo, M. Gonz´alez and Y. Moreno, On separably
injective Banach spaces, Adv. Math. 234 (2013) 192 -- 216.
[3] A. Avil´es, F. Cabello S´anchez, J. M.F. Castillo, M. Gonz´alez and Y. Moreno, ℵ-injective
Banach spaces and ℵ-projective compacta, Rev. Mat. Iberoamericana, 31 (2015), 575 -- 600.
[4] A. Avil´es, F. Cabello S´anchez, J. M.F. Castillo, M. Gonz´alez and Y. Moreno, Separably
injective Banach spaces, Lecture Notes in Math. 2132, Springer, 2016.
[5] L.J. Bunce and C-H. Chu, Real contractive projections on commutative C*-algebras, Math.
Z. 226 (1997) 85-101.
[6] C-H. Chu, Jordan structures in geometry and analysis, Cambridge Tracts in Math. 190,
Cambridge Univ. Press, Cambridge, 2012.
[7] C-H. Chu and L. Li, Separably injective C*-algebras, Arch. Math. 106(2016), 553 -- 559.
[8] L. Gillman and M. Jerison, Rings of continuous functions, Van Nostrand, Princeton, 1960.
[9] Y. Friedman and B. Russo, Function representation of commutative operator triple systems,
J. London Math. Soc. 27(1983), 513-524.
[10] Y. Friedman and B. Russo, Solution of the contractive projection problem, J. Funct. Anal.
60 (1985) 56-79.
[11] K. Grove and G.K. Pedersen, Sub-Stonean spaces and corona sets, J. Funct. Anal., 56 (1984),
124 -- 143.
[12] M. Hasumi, The extension property of complex Banach spaces, Tohoku Mathematical Jour-
nal, 10(2) (1958), 135 -- 142.
[13] M. Henriksen and R. G. Woods, F-spaces and substonean spaces: general topology as a tool in
functional analysis, Papers on general topology and related category theory and topological
algebra (New York, 1985/1987), 60 -- 68, Ann. New York Acad. Sci., 552, New York Acad. Sci.,
New York, 1989.
[14] J.L. Kelley, Banach spaces with the extension property, Trans. Amer. Math. Soc. 72 (1952)
323-326.
[15] J. Lindenstrauss and D.E. Wulbert, On the classification of the Banach spaces whose duals
are L1 spaces, J. Funct. Anal. 4 (1969) 332-349.
[16] G.H. Olsen, On the classification of complex Lindenstrauss spaces, Math. Scand. 35 (1974)
237 -- 258.
[17] B. Russo, Structure of JB*-triple, in Proc. Oberwolfach Conf. on Jordan algebras, 1992, eds.
W. Kaup, K. McCrimmon, H.P. Petersson (Walter de Gruyter, Berlin, 1994) 209-280.
[18] G.L. Seever, Measures on F-spaces, Trans. Amer. Math. Soc. 133 (1968), 267-280.
[19] H. Upmeier, Symmetric Banach manifolds and Jordan C*-algebras (North Holland Math.
Studies 104) North Holland, Amsterdam 1985.
[20] M. Zippin, The separable extension problem, Israel J. Math. 26 (1977) 372-387.
School of Mathematical Sciences, Queen Mary, University of London, London E1
4NS, UK
E-mail address: [email protected]
School of Mathematical Sciences and LPMC, Nankai University, Tianjin 300071,
China
E-mail address: [email protected]
|
1904.00174 | 1 | 1904 | 2019-03-30T08:48:38 | Barrier functions in the subdifferential theory | [
"math.FA"
] | We present a new method for proving Correa-Jofr\'e-Thibault theorem that monotonicity of subdifferential implies convexity of the function.
This new method is based on barrier functions. Barrier functions help overcome some of the main technical difficulties when working with lower semicontinuous functions. | math.FA | math |
Barrier functions in the subdifferential theory
Milen Ivanov∗
Nadia Zlateva†
Radiant Life Technologies Ltd.
Faculty of Mathematics and Informatics
Nicosia, Cyprus
Sofia University
[email protected]
5, James Bourchier Blvd.
1164 Sofia, Bulgaria
[email protected]
Abstract
We present a new method for proving Correa-Jofr´e-Thibault theo-
rem that monotonicity of subdifferential implies convexity of the func-
tion.
This new method is based on barrier functions. Barrier functions
help overcome some of the main technical difficulties when working
with lower semicontinuous functions.
2010 Mathematics Subject Classification: 49J52, 47H05, 52A41.
Keywords: subdifferential, monotonicity, convex function, barrier function
1
Introduction
In 1990's Correa, Jofr´e and Thibault in series of papers proved that a convex
lower semicontinuous function can be characterized by monotonicity property
of its subdifferential -- in reflexive Banach space for the Clarke subdifferential
in [2] and in any Banach space for axiomatically introduced subdifferential
in [3] and for more general axiomatic presubdifferential in [4].
∗Supported by Bulgarian National Scientific Fund under grant KP-06-22/4.
†Supported by Scientific Fund of Sofia University under grant for 2019.
1
The main tool for proving this characterization is the Mean Value The-
orem of Zagrodny [13] which holds for a lower semicontinuous function in a
Banach space for any presubdifferential, see [11].
Jules and Lassonde prove a subdifferential test for optimality, see [7] of
Minty type [8] involving a subdifferential satisfying certain axioms.
In setting up the axiomatic framework we follow [12], but we pick the
apparently minimal set of axioms under which proofs can work. In this way
our results are slightly more general. As mentioned below, adding another
natural axiom can significantly simplify the presentation.
We work in a real Banach space X with dual X ∗.
Definition 1 (axioms for subdifferential). Multi-valued operator ∂ which
associates to any function f : X → R ∪ {+∞} and any x ∈ X a (possibly
empty) subset ∂f (x) ⊂ X ∗ is feasible subdifferential if
(P1) ∂f (x) = ∂cf (x) whenever f is a convex and continuous function in a
neighbourhood U of x, where ∂c stands for the Fenchel subdifferential,
i.e.
∂cf (x) := {x∗ ∈ X ∗ : hx∗, y − xi ≤ f (y) − f (x), ∀y ∈ U};
(P2) For f lower semicontinuous and g convex and continuous in a neigh-
bourhood of x ∈ dom f ,
0 ∈ ∂g(x) + lim sup
∂f (y)
y→x, f (y)→f (x)
(1)
whenever x is a local minimum point of f + g.
In more details (1) means that there are y ∗
n ∈ ∂f (yn) such that yn → x,
n converges in the w∗ topology to some y ∗
f (yn) → f (x) and the sequence y ∗
such that −y ∗ ∈ ∂g(x).
Note that all presubdifferentials considered by Correa, Jofr´e and Thibault
in [2, 3, 4], as well as subdifferentials considered by Jules and Lassonde in [7]
are feasible subdifferentials in the sense of Definition 1.
In terms of Ioffe's extensive classification, see [6], (P1) is called contiguity,
while (P2) is weak-star form of trustworthness.
We prove the above mentioned results for feasible subdifferentials and in
a different and unified way -- by using barrier functions instead of (a variant
of) Zagrodny Theorem.
2
For convenience of notation we often identify the map ∂f : X ⇒ X ∗ with
its graph, that is, (x, x∗) ∈ ∂f is a shorthand for x∗ ∈ ∂f (x).
The main contribution of this work is a new method, based on [5], see also
[10, p.569], for proving the following result of Correa, Jofr´e and Thibault.
Theorem 2 (Correa-Jofr´e-Thibault). Let X be a Banach space and let ∂ be
a feasible subdifferential.
Let f : X → R ∪ {+∞} be a proper lower semicontinuous function.
If ∂f is monotone, then f is convex.
Recall that monotonicity of ∂f means
(xi, x∗
i ) ∈ ∂f, i = 1, 2 ⇒ hx∗
2 − x∗
1, x2 − x1i ≥ 0.
The routine way of demonstrating the above result can be sketched like
this: examining the proof in [11] it is clear that Zagrodny Mean Value The-
orem holds for any feasible subdifferential. Using it, one can prove a Minty
test for optimality [8] of the following form
hx∗, x − x0i ≥ 0, ∀(x, x∗) ∈ ∂f ⇒ 0 ∈ ∂cf (x0),
where f is proper and lower semicontinuous and ∂ is feasible subdifferential.
If -- on top of (P1) and (P2) -- ∂ also satisfies the natural axiom
(P3) ∂(f + x∗) = ∂f + x∗,
∀x∗ ∈ X ∗
(called in [7] stability property which is rather limited form of calculability
axiom in [6]), then from Minty test immediately follows that ∂ ∪ ∂c has the
somewhat surprising property (first noted by Jules and Lasonde) to be max-
0) is monotonous
imal with respect to monotonicity relation. That is, if (x0, x∗
related to ∂f :
hx∗ − x∗
0, x − x0i ≥ 0,
∀(x, x∗) ∈ ∂f,
0) ∈ ∂cf . From the latter Moreau-Rockafellar Theorem about
then (x0, x∗
maximal monotonicity of ∂cf for convex, proper and lower semicontinuous f
follows immediately, but the surprising fact is that the above is true even if
∂f is not itself monotone. (For precise statement see Theorem 6.)
So, in particular if ∂ is feasible and satisfies in addition (P3), then
∂f ⊂ ∂cf,
3
whenever ∂f is monotone. Further the proof can be completed as we do in
here presented proof of Theorem 2.
Note that the additional axiom (P3) is not really necessary.
Our approach is based on a different technique involving barrier func-
tions instead of Zagrodny Theorem. We will also make the effort to obtain
Theorem 2 for general feasible subdifferental.
The paper is organized as follows.
In Section 2 we construct and consider the class of barrier functions we
use. In Section 3 we show some additional properties of the feasible subdif-
ferential linked to (P2) axiom. Finally, in Section 4 we present our proof of
Correa-Jofr´e-Thibault theorem.
2 Barrier functions
Let U ⊂ X be a open, convex and bounded neighbourhood of 0, i.e. 0 ∈ U.
Let µ be the Minkowski functional of U, that is,
µ(x) = inf{λ : λ > 0, x ∈ λU} for x 6= 0, µ(0) = 0.
It is clear that µ(x) < 1 if x ∈ U, and µ(x) ≥ 1 if x 6∈ U.
Let us first list few properties of Minkowski functional, see e.g. [1]:
(i) µ has values in [0, +∞);
(ii) µ is positively homogeneous, i.e. µ(tx) = tµ(x) for all x ∈ X and t ≥ 0;
(iii) µ(x + y) ≤ µ(x) + µ(y) for all x, y ∈ X and hence µ is convex;
(iv) {x : µ(x) ≤ 1} = cl U ⇒ {x : µ(x) = 1} = ∂U := cl U \ U, where cl U
denotes the topological closure of U and ∂U denotes its boundary.
Moreover,
(v) There exists c > 0 such that µ(x) ≤ ckxk for all x ∈ X;
(vi) There exists b > 0 such that µ(x) ≥ bkxk for all x ∈ X;
(vii) µ is Lipschtz continuous.
4
Indeed, let δ > 0 be such that δBX ⊂ U, where BX := {x ∈ X : kxk ≤ 1}.
Then
δx
kxk
∈ δBX ⊂ U, so µ(cid:18) δx
kxk(cid:19) ≤ 1, and µ(x) ≤ δ−1kxk, which is (v)
with c := δ−1.
Further, let U ⊂ sBX , then
giving (vi) with b := s−1.
sx
kxk
6∈ U, so µ(cid:18) sx
kxk(cid:19) ≥ 1, and µ(x) ≥
1
s
kxk
Finally, µ(x) = µ((x − y) + y) ≤ µ(x − y) + µ(y) by (ii). Hence, µ(x) −
µ(y) ≤ µ(x − y) and using (vi) we get µ(x) − µ(y) ≤ bkx − yk which yields
µ(x) − µ(y) ≤ bkx − yk for all x, y ∈ X and (vii) holds.
For x ∈ U, define the function
k(x) :=
µ(x)
1 − µ(x)
=
1
1 − µ(x)
− 1.
(2)
The next lemma shows that k is a barrier function for U, i.e. a continuous
function whose values tend to infinity while the arguments tend to ∂U (see
e.g. [9]).
Lemma 3. The function k defined by (2) has the following properties:
(a) k(0) = 0 and for some b > 0, k(x) ≥ µ(x) ≥ bkxk.
(b)
lim
x→∂U
k(x) = +∞;
(c) k is convex and Lipschitz on each level set {x : k(x) ≤ λ};
(d) the function k(x) := (cid:26) k(x), x ∈ U
+∞, x ∈ X \ U
convex.
Proof. The properties (a) and (b) are clear.
is lower semicontinuous and
In order to prove that k is convex, it is enough to show that the function
g(x) := k(x) + 1 =
is convex.
1
1 − µ(x)
To this end, fix x, y ∈ U and λ ∈ [0, 1]. From (i)
µ(λx + (1 − λ)y) ≤ λµ(x) + (1 − λ)µ(y) ⇔
1 − µ(λx + (1 − λ)y) ≥ λ(1 − µ(x)) + (1 − λ)(1 − µ(y)).
5
Set α := 1 − µ(x), β := 1 − µ(y), so α, β ∈ (0, 1). Then the latter says
g(λx + (1 − λ)y) ≤
1
λα + (1 − λ)β
.
We claim that
1
λα + (1 − λ)β
≤
λ
α
+
1 − λ
β
.
Of course, (4) is equivalent to
(3)
(4)
0 ≤ λβ(λα + (1 − λ)β) + (1 − λ)α(λα + (1 − λ)β) − αβ
= λ(1 − λ)(α2 + β2) + (λ2 + (1 − λ)2 − 1)αβ
= λ(1 − λ)(α2 + β2) + (2λ2 − 2λ)αβ
= λ(1 − λ)(α − β)2.
From (3) and (4) it follows that
g(λx + (1 − λ)y) ≤ λg(x) + (1 − λ)g(y)
and k is convex. Lipschitz continuity of k on each level set is inherited by
the Lipschitz continuity of µ.
(d) follows from (b) and (c).
3 Properties of feasible subdifferential
For the sake of clarity, we will take two technical parts out of the proof of
the main result. Namely, we will show some additional properties of feasible
subdifferential linked to (P2) axiom.
Lemma 4. Let ∂ be a feasible subdifferential. Let f : X → R ∪ {+∞} be
lower semicontinuous and bounded below on the open, convex and bounded
set U ⊂ X. Let, moreover, dom f ∩ U 6= ∅. Let g be convex continuous and
bounded below barrier function for U. Let ¯x ∈ Xbe fixed.
n ∈ ∂f (xn), y ∗
n ∈
Then there exist sequences (xn)∞
n=1, (yn)∞
n=1 ⊂ U and x∗
∂cg(yn) such that
kxn − ynk → 0,
f (xn) + g(yn) → inf
U
(f + g),
hx∗
n, xn − ¯xi + hy ∗
n, yn − ¯xi → 0.
6
(5)
(6)
(7)
Proof. Define
M := sup{ky − ¯xk : y ∈ U}.
(8)
Since U is bounded, M ∈ R.
Fix a sequence (εn)∞
Pick zn ∈ U such that
n=1 such that εn > 0 and εn → 0 as n → ∞.
f (zn) + g(zn) < inf
U
(f + g) + εn.
By Ekeland Variational Principle there are yn ∈ U such that
f (yn) + g(yn) < inf
U
(f + g) + εn,
(9)
and the function
x → f (x) + g(x) + εkx − ynk
attains its minimum at yn. By (P2) there are uk → yn, as k → ∞, and
k ∈ ∂f (uk) such that f (uk) → f (yn) and
u∗
w∗− lim u∗
k = u∗ ∈ −∂(g + εnk · −ynk)(yn).
By Sum Theorem for the Fenchel subdifferential and the fact that ∂εnk ·
−ynk = εn∂k · −ynk it follows that there are y ∗
n ∈ ∂g(yn) such that
ku∗ + y ∗
nk ≤ εn.
(10)
Since (uk, f (uk)) → (yn, f (yn)) as k → ∞, there is k1 ∈ N such that
kuk − ynk < εn, f (uk) − f (yn) < εn,
∀k > k1.
(11)
By (9) we have
f (uk) + g(yn) − inf
U
(f + g) ≤ 2εn,
∀k > k1.
(12)
Note that
hu∗
k, uk − ¯xi + hy ∗
n, yn − ¯xi = hu∗ + y ∗
n, yn − ¯xi + hu∗
k − u∗, yn − ¯xi + hu∗
k, uk − yni.
n, yn − ¯xi ≤ ku∗ + y ∗
Now, hu∗ + y ∗
k weak-star converges to u∗, we have hu∗
u∗
since by Banach-Steinhaus Theorem the sequence (u∗
nkkyn − ¯xk ≤ Mεn by (8) and (10). Since
k − u∗, yn − ¯xi → 0 as k → ∞. Also,
k=1 is bounded, and
k)∞
7
uk → yn as k → ∞, we have hu∗
Therefore, there is k2 ∈ N such that
k, uk − yni ≤ ku∗
kkkuk − ynk → 0 as k → ∞.
hu∗
k, uk − ¯xi + hy ∗
n, yn − ¯xi ≤ (M + 1)εn,
∀k > k2.
(13)
Set
¯k = max{k1, k2} + 1 and (xn, x∗
n) := (u¯k, u∗
¯k).
From (11), (12) and (13) it follows that so constructed sequences (xn, x∗
and (yn, y ∗
n=1 satisfy (5), (6) and (7).
n)∞
n)∞
n=1
Proposition 5. Let ∂ be a feasible subdifferential. Let f : X → R ∪ {+∞}
be a proper and lower semicontinuous function. Then dom ∂f is nonempty
and
f (x) ≤ sup{f (x) + hx∗, x − xi : (x, x∗) ∈ ∂f }, for all x ∈ X.
(14)
Proof. Fix arbitrary ¯x ∈ X and
r < f (¯x).
Fix some y ∈ dom f .
Since f is lower semicontinuous and the segment [¯x, y] is compact, there
is δ > 0 such that f is bounded below on the set
C := [¯x, y] + δB ◦
X.
Let k be the barrier function defined by (2) for the set U := C − {¯x}.
Let ξ > 0 be such that f > r on ¯x + ξBX. Fix a > 0 and such that
a > (r − inf
U
f )/(ξb),
where b > 0 and k(x) ≥ µ(x) ≥ bkxk (see Lemma 3 (a)).
Then it is immediate that
f (x) + ak(x − ¯x) > r,
∀x ∈ U.
(15)
Apply Lemma 4 to f and the barrier function g = ak(·−x) to get x∗
and y ∗
n ∈ ∂ck(· − ¯x)(yn) such that (5) and (6) are fulfilled and, moreover,
n ∈ ∂f (xn)
hx∗
n, xn − ¯xi + ahy ∗
n, yn − ¯xi = αn, where lim
n→∞
αn = 0.
(16)
8
Since y ∗
n ∈ ∂ck(yn − ¯x), we have
k(0) ≥ k(yn − ¯x) + hy ∗
n, ¯x − yni ⇐⇒ hy ∗
n, yn − ¯xi ≥ k(yn − ¯x),
From the boundedness below of f and (6) it follows that the sequence
n=1 is bounded. Since k is Lipschitz on its level sets (see Lemma 3
(k(yn − ¯x))∞
(c)), from (5) it follows that
k(yn − ¯x) = k(xn − ¯x) + βn, where lim
n→∞
βn = 0.
These and (16) give hx∗
−ak(xn − ¯x) + αn − aβn. So,
n, xn − ¯xi = αn − ahy ∗
n, yn − ¯xi ≤ αn − ak(yn − ¯x) =
hx∗
n, xn − ¯xi ≤ −ak(xn − ¯x) + γn, where lim
n→∞
γn = 0.
From (15) it follows that hx∗
n, xn − ¯xi < f (xn) − r + γn, or, equivalently,
r < f (xn) + hx∗
n, ¯x − xni + γn.
Therefore,
r ≤ sup{f (xn) + hx∗
n, ¯x − xni : n ∈ N}
≤ sup{f (x) + hx∗, ¯x − xi : (x, x∗) ∈ ∂f }.
Since r < f (¯x) were arbitrary, we are done.
4 Monotonicity and convexity
We start with the following extension to the case of feasible subdifferential
of a result of Jules and Lassonde [7].
Theorem 6. Let ∂ be a feasible subdifferential. Let f : X → R ∪ {+∞} be a
0) ∈ X × X ∗ be in monotone
proper lower semicontinuous function. Let (x0, x∗
relation to ∂f , that is,
hx∗ − x∗
0, x − x0i ≥ 0,
∀(x, x∗) ∈ ∂f.
(17)
Then (x0, x∗
0) ∈ ∂cf .
9
Proof. Let r ∈ R be such that
r < f (x0).
(18)
Let y ∈ dom f be arbitrary.
Since f is lower semicontinuous and the segment [x0, y] is compact, there
is δ > 0 such that f is bounded below on
Let k be the barrier function defined by (2) for the set U := C − {x0}.
C := [x0, y] + δB ◦
X.
Let a > 0 be arbitrary.
Obviously,
g(x) := −hx∗
0, x − x0i + ak(x − x0)
is a convex and continuous barrier for U. So, we can apply Lemma 4 to f
and g. Since
∂cg = −x∗
0 + a∂ck(· − x0),
there are x∗
fulfilled and, moreover,
n ∈ ∂f (xn) and y ∗
n ∈ ∂ck(· − x0)(yn) such that (5) and (6) are
hx∗
n, xn − x0i + h−x∗
0 + ay ∗
n, yn − x0i → 0.
Clearly, hx∗
0, xn − yni → 0, so (19) is equivalent to
hx∗
n − x∗
0, xn − x0i + ahy ∗
n, yn − x0i → 0.
This and (17) give
Since y ∗
n ∈ ∂ck(· − x0)(yn), we have
lim sup
n→∞
hy ∗
n, yn − x0i ≤ 0.
k(0) ≥ k(yn − x0) + hy ∗
n, x0 − yni ⇐⇒ hy ∗
n, yn − x0i ≥ k(yn − x0),
(19)
(20)
(21)
because k(0) = 0. But k(yn − x0) ≥ bkyn − x0k (cf. Lemma 3 (a)), so (20)
and (21) imply that
lim sup
n→∞
kyn − x0k ≤ 0 ⇐⇒ lim
n→∞
yn = x0.
From (5) it follows that xn → x0 as well; and from the lower semicontinuity
of f and (6) it follows that
f (x0) + g(x0) = inf
U
(f + g).
10
In particular,
f (y) − hx∗
0, y − x0i + ak(y − x0) ≥ f (x0).
Since a > 0 was arbitrary,
f (y) ≥ f (x0) + hx∗
0, y − x0i.
So, x0 ∈ dom f and, since y ∈ dom f was arbitrary, x∗
0 ∈ ∂cf (x0).
After all the above development, the proof of Correa-Jofr´e-Thibault is
now almost immediate.
Theorem 2 (Correa-Jofr´e-Thibault). Let X be a Banach space and ∂ be a
feasible subdifferential.
Let f : X → R ∪ {+∞} be a proper lower semicontinuous function.
If ∂f is monotone, then f is convex.
Proof. Consider g : X → R ∪ {+∞} defined as
g(x) := sup{f (x) + hx∗, x − xi : (x, x∗) ∈ ∂f }.
As a supremum of linear functions, g is convex and lower semicontinuous.
By (14) we have that
f ≤ g.
From Theorem 6 and monotonicity of ∂f we have that ∂f ⊂ ∂cf , which
implies f ≥ g. Therefore,
f = g.
References
[1] F. Clarke, Functional Analysis, Calculus of Variations and Optimal Con-
trol, Graduate Texts in Mathematics 264 (2013) Springer-Verlag, Lon-
don.
[2] R. Correa, A. Jofr´e, L. Thibault, Characterization of lower semicontin-
uous convex functions, PAMS 116 (1992) 67 -- 72.
11
[3] R. Correa, A. Jofr´e, L. Thibault, Subdifferential monotonicity as char-
acterization of convex functions, Numer. funct. anal. and Optimiz. 15
(1994) 531 -- 535.
[4] R. Correa, A. Jofr´e, L. Thibault, Subdifferential characterization of con-
vexity, In: Recent advances in non-smooth optimization, Eds. D.-Z. Du,
L. Qi and R. S. Womersly (1995) World Scentific publishing, 18 -- 23.
[5] M. Ivanov and N. Zlateva, Maximal Monotonicity of the Subdifferential
of a Convex Function: a Direct Proof, Journal of Convex Analysis 24
(2017) 1307 -- 1311.
[6] A. Ioffe, On the theory of subdifferentials, Adv. Nonlinear Anal. 1 (2012)
47-120.
[7] F. Jules, M. Lassonde, Subdifferential test for optimality, J. Global Op-
tim. 59 (2014) 101 -- 106.
[8] G. J. Minty: On the monotonicity of the gradient of a convex function,
Pacific J. Math. 14 (1964) 243 -- 247.
[9] J. Nocedal, S. Wright, Numerical Optimization (1999) Springer, New
York.
[10] J.-P. Penot, Analysis: From Concepts to Applications (2016) Springer.
[11] L. Thibault, A note on the Zagrodny mean value theorem, Optimization
35 (1995) 127 -- 130.
[12] L. Thibault, N. Zlateva, Integrability of subdifferentials of directionally
Lipschitz functions, PAMS 133 (10) (2005) 2939 -- 2948.
[13] D. Zagrodny, Approximate mean value theorem for upper subderiva-
tives, Nonlinear Anal. 12 (1988) 1413 -- 1428.
12
|
1111.2435 | 1 | 1111 | 2011-11-10T10:01:48 | A Simple Algorithm for Constructing all Real Hessenberg Unitary Matrices | [
"math.FA"
] | Unitary matrices which are zero below the secondary diagonal (Hessenberg unitary matrices) have many uses in analysis. Given a set of needed conditions on a unitary matrix, this algorithm will give the sparsest unitary matrix. We give an algorithm for constructing all real Hessenberg unitary matrices. The $n\times n$ unitary matrices given by the algorithm have $n-1$ variables which can be chosen to give additional properties needed for a particular application. | math.FA | math |
A SIMPLE ALGORITHM FOR CONSTRUCTING ALL REAL
HESSENBERG UNITARY MATRICES
JANET C. TREMAIN
Abstract. Unitary matrices which are zero below the secondary diagonal
(Hessenberg unitary matrices) have many uses in analysis. Given a set of
needed conditions on a unitary matrix, this algorithm will give the sparsest
unitary matrix. We give an algorithm for constructing all real Hessenberg
unitary matrices. The n × n unitary matrices given by the algorithm have
n − 1 variables which can be chosen to give additional properties needed for a
particular application.
1. Introduction
Hessenberg matrices have may uses in both pure an applied mathematics.
It
is a well used fact that every matrix is unitarily equivalent to such a matrix and
there are polynomial time algorithms for doing this. In numerical linear algebra,
they are used to speed-up eigenvalue computations. Also, real Hessenberg unitary
matrices have various uses in analysis. They also have the advantage of being quite
sparse. Here we give an algorithm which constructs all such real unitary matrices.
The algorithm starts with 2 × 2 matrices and for each new dimension, we drop the
top row of the previous dimension, introduce a new variable and then construct
two new rows with the new variable added and finally use the rest of the rows of
the previous case by adding the first column of zeroes. The n × n unitary matrices
have n − 1 variables which can be chosen to give additional properties needed for a
particular application.
This algorithm was developed at the request of the people working in the Frame
Research Center (www.framerc.org) for their work in frame theory where they
needed the sparsest unitary matrices.
Remark 1.1. Since we can multiply rows and columns of a unitary by −1 and
still have a unitary, and if we permute rows or columns while maintaining the
Hessenberg property we still have a unitary, we will construct all real triangular
unitary matrices up to such changes in sign and permutations of rows or columns.
2. The Construction
We will give an algorithm which constructs all real Hessenberg unitary matrices.
So we can better see how the algorithm works, we will start by giving some small
examples.
The author was supported by DTRA/NSF 1042701.
1
2
J. TREMAIN
2.1. 2 × 2 Matrices. This is a unique class and is given by:
(cid:20)√1 − z1
√z1
−√z1 √1 − z1(cid:21)
Here we must have 0 ≤ z1 ≤ 1.
2.2. 3 × 3 Matrices. This class is given by:
p(1 − z1)z2
√1 − z2
√z1z2
−√z2 p(1 − z2)(1 − z1) p(1 − z2)z1
√1 − z1
Here, 0 ≤ z1, z2 ≤ 1.
Proposition 2.1. This class of matrices consists of unitary matrices.
0
−√z1
Proof. The proof is done by cases:
Case 1: The rows have norm 1.
The square sum of the elements of row 1 is
The square sum of the elements of row 2 is
(1 − z2) + (1 − z1)z2 + z1z2 = 1.
z2 + (1 − z2)(1 − z1) + (1 − z2)z1 = z2 + (1 − z2) = 1.
It is clear that row 3 is norm 1.
Because of the symmetry of the matrix (I.e. switching z1 and z2 in the calcula-
tion), it is clear that the row sums and column sums are equal.
Case 2: The rows are orthogonal.
The inner product of rows 1 and 2 is
−p(1 − z2)z2 + (1 − z1)p(1 − z2)z2 + z1p(1 − z2)z2 =
p(1 − z2)z2[−1 + (1 − z1) + z1] = 0.
−p(1 − z1)z1z2 +p(1 − z1)z1z2 = 0.
The inner product of rows 1 and 3 is
The inner product of rows 2 and 3 is
−p(1 − z2)(1 − z1)z1 +p(1 − z2)(1 − z1)z1 = 0.
By symmetry again, the columns are orthogonal.
(cid:3)
We give some examples of how to use this:
Case 1: If z1 = z2 = 0 we get
1 0
0 1
0 0
0
0
1
HESSENBERG UNITARY MATRICES
3
Case 2: If we let z1 = z2 = 1 we get
0
0
0
1
−1
0
0 −1 0
Case 3: If we let z1 = 1 and z2 = 0 we get
0
0
1
0
0
1
0 −1 0
Case 4: If we let z1 = 1
2 and z2 = 1 we get
2
0 q 1
−1
0 −q 1
2 q 1
2 q 1
0
0
2
Case 5: If z1 = 1
2 and z2 = 2
3 we get
3
3 q 1
q 1
3 q 1
−q 2
−q 1
3 q 1
6 q 1
2 q 1
0
6
2
Theorem 2.2. This algorithm gives all 3 × 3 real Hessenberg unitary matrices.
Proof. We will just outline this. Given a unitary
a11 a12 a13
a21 a22 a23
0
a32 a33
If one of a32 or a33 is zero, it is easily checked that the matrix is a permutation of
the rows or columns of the identity multiplied by −1 if necessary and this matrix
easily arises from our algorithm. Otherwise, we may assume a32 is negative and
2
32 to get the last row correct given that it square sums to 1. Now,
let our z1 = a
the last row must be orthogonal to the rows above it so (a22, a23) = c(a12, a13) are
the unique vectors (up to length) in R2 which are orthogonal to row 3. Since z1 is
given, we can find the unique z2 which makes the last two terms of our first and
second rows equal to the given one. The first column is now uniquely determined
and must be of our form.
(cid:3)
2.3. 4 × 4 Matrices. We construct a unitary matrix by:
√1 − z3
√z1z2z3
p(1 − z1)z2z3
−√z3 p(1 − z3)(1 − z2) p(1 − z3)(1 − z1)z2 p(1 − z3)z1z2
p(1 − z2)(1 − z1) p(1 − z2)z1
√1 − z1
p(1 − z2)z3
−√z1
−√z2
0
0
0
We leave it to the reader to check that this constructs all 4 × 4 real Hessenberg
unitary matrices.
4
J. TREMAIN
2.4. 5 × 5 Matrices: The Algorithm. Given our general n × n real Hessenberg
unitary matrix with n − 1 variables in it, we introduce a new multiplier √zn+1
on the second row of the n × n case and the first entry in the second row of the
(n + 1) × (n + 1) matrix is −√zn+1. We also need the sum of the squares of the
entries of the two new rows we are adding at the top to equal the squares of the
entries of the top row of the n × n matrix. After that, we add the remaining rows
of the previous case with a column of zeroes in front.
p(1 − z3)z4
p(1 − z2)z3z4
Combining this with the previous case we get:
So, to pass from the 4 × 4 case to the 5 × 5 case we create two new rows:
(cid:20)√1 − z4
−√z4 p(1 − z4)(1 − z3) p(1 − z4)(1 − z2)z3 p(1 − z4)(1 − z1)z2z3 p(1 − z4)z1z2z3(cid:21)
√1 − z4
√z1z2z3z4
p(1 − z2)z3z4
−√z4 p(1 − z4)(1 − z3) p(1 − z4)(1 − z2)z3 p(1 − z4)(1 − z1)z2z3 p(1 − z4)z1z2z3
p(1 − z3)(1 − z2) p(1 − z3)(1 − z1)z2 p(1 − z3)z1z2
p(1 − z2)z1
√1 − z1
p(1 − z2)(1 − z1)
p(1 − z1)z2z3z4
p(1 − z1)z2z3z4
p(1 − z3)z4
0
0
0
−√z3
0
0
√z1z2z3z4
−√z2
0
−√z1
Theorem 2.3. The matrix given above is a unitary matrix.
Proof. We will outline the proof. Since the last three rows come from the previous
unitary matrix, we just need to check that the first two rows are orthogonal to all
the others, that they are norm 1 and the column vectors have norm 1.
The two new rows are norm 1:
The norm of row 1 is:
1 − z4 + (1 − z3)z4 + (1 − z2)z3z4 + (1 − z1)z2z3z4 + z1z2z3z4 =
1 − z3z4 + (1 − z2)z3z4 + z2z3z4 = 1 − z3z4 + z3z4 = 1.
The norm of row 2 is:
z4 + (1 − z4)(1 − z3) + (1 − z4)(1 − z2)z3 + (1 − z4)(1 − z1)z2z3 + (1 − z4)z1z2z3 =
z4 + (1 − z4)[(1 − z3) + (1 − z2)z3 + (1 − z1)z2z3 + z1z2z3] =
z4 + (1 − z4)[1 − z2z3 + z2z3] = 1.
The column vectors are norm 1:
Column 1 square sums to (1 − z4) + z4 = 1. Column 2 square sums to
(1 − z3)z4 + (1 − z4)(1 − z3) + z3 = 1 − z3 + z3 = 1.
Column 3 square sums to
(1 − z2)z3z4 + (1 − z4)(1 − z2)z3 + (1 − z3)(1 − z2) + z2 =
(1 − z2)[z3z4 + (1 − z4)z3 + (1 − z3)] + z2 = (1 − z2)[1] + z2 = 1.
Column 4 square sums to
(1 − z1)z2z3z4 + (1 − z4)(1 − z1)z2z3 + (1 − z3)(1 − z1)z2 + (1 − z2)(1 − z1) + z1 =
(1 − z1)[z2z3z4 + (1 − z4)z2z3 + (1 − z3)z2 + (1 − z2)] + z1 =
(1 − z1)[z2z3 + 1 − z2z3] + z1 = 1.
HESSENBERG UNITARY MATRICES
5
Column 5 square sums to
z1z2z3z4 + (1 − z4)z1z2z3 + (1 − z3)z1z2 + (1 − z2)z1 + (1 − z1) =
z1z2z3 + z1z2 − z1z2z3 + x1 − z1z2 + 1 − z1 = 1.
The rows are orthogonal: The inner product of rows 1 and 2 is:
−p(z4(1 − z4) +p(1 − z3)2
z4(1 − z4) +q(1 − z2)2
z
2
3 z4(1 − z4)+
q(1 − z1)2
2
2 z
z
2
3 z4(1 − z4) +qz
2
1 z
2
2 z
2
3 z4(1 − z4) =
−pz4(1 − z4) + (1 − z3)pz4(1 − z4) + (1 − z2)z3pz4(1 − z4)+
(1 − z1)z2z3pz4(1 − z4) + z1z2z3pz4(1 − z4) =
pz4(1 − z4)[−1 + (1 − z3) + (1 − z2)z3 + (1 − z1)z2z3 + z1z2z3] =
pz4(1 − z4)[−1 + (1 − z3) + z3 − z2z3 + z2z3 − z1z2z3 + z1z2z3] = 0.
(cid:3)
The inner product of rows 1 and 3 and taking our squares as above at the same
time is
−pz3(1 − z3)z4+(1−z2)pz3(1 − z3)z4+(1−z1)z2px3(1 − z3)z4+z1z2pz3(1 − z3)z4 =
pz3(1 − z3)z4[−1 + (1 − z2) + (1 − z1)z2 + z1z2] = 0
The inner product of rows 1 and 4 is
−pz2(1 − z2)z3z4 + (1 − z1)p(1 − z2)z2z3z4 + z1p(1 − z2)z2z3z4 =
The inner product of rows 1 and 5 is
pz2(1 − z2)z3z4[−1 + (1 − z1) + z1] = 0.
−pz1(1 − z1)z2z3z4 +pz1(1 − z1)z2z3z4] = 0.
The inner product of rows 2 and 3 is
−p(1 − z4)(1 − z3)z3 + (1 − z2)p(1 − z4)(1 − z3)z3+
(1 − z1)z2p(1 − z4)(1 − z3)z3 + z1z2p(1 − z4)(1 − z3)z3 =
p(1 − z4)(1 − z3)z3)[−1 + (1 − z2) + (1 − z1)z2 + z1z2] = 0.
The inner product of rows 2 and 4 is
−p(1 − z4)z2(1 − z2)z3+(1−z1)p(1 − z4)(1 − z2)z2z3+z1p(1 − z4)(1 − z2)z2z3 = 0.
The inner product of rows 2 and 5 is
−p(1 − z4)z1(1 − z1)z2z3 +p(1 − z4)z1(1 − z1)z2z3 = 0.
6
J. TREMAIN
3. The General Case
For the general case, we take the n × n case and delete the first row and add
a first column of zeroes. Then we add two new rows at the top of this matrix by
choosing a new variable zn and making row 1 as:
√1 − zn p(1 − zn−1)zn p(1 − zn−2)zn−1zn p(1 − zn−3)zn−1zn−2
· · ·
and making row 2 as
p(1 − z1)z2z3 · · · zn . . . √z1z2 · · · zn
−√zn p(1 − zn)(1 − zn−1) p(1 − zn)(1 − zn−2)zn−1 p(1 − zn)(1 − zn−3)zn−1zn−2 · · ·
p(1 − zn)(1 − z1)z2z3 . . . zn−1 p(1 − zn)z1z2z3 . . . zn−1.
Remark 3.1. It takes a significant amount of effort to show that we can produce
all the required Hessenberg unitary matrices and it does not seem to be important
enough to justify the effort since our intention is not to publish this but just post
it on the arXiv so it is available to researchers. So we will not address this here.
Basically, it can be done by induction on n and a case analysis of the placement of
zeroes.
References
Department of Mathematics, University of Missouri, Columbia, MO 65211-4100
E-mail address: [email protected]
|
1803.07468 | 1 | 1803 | 2018-03-20T14:56:04 | Equiangular tight frames from group divisible designs | [
"math.FA",
"math.CO"
] | An equiangular tight frame (ETF) is a type of optimal packing of lines in a real or complex Hilbert space. In the complex case, the existence of an ETF of a given size remains an open problem in many cases. In this paper, we observe that many of the known constructions of ETFs are of one of two types. We further provide a new method for combining a given ETF of one of these two types with an appropriate group divisible design (GDD) in order to produce a larger ETF of the same type. By applying this method to known families of ETFs and GDDs, we obtain several new infinite families of ETFs. The real instances of these ETFs correspond to several new infinite families of strongly regular graphs. Our approach was inspired by a seminal paper of Davis and Jedwab which both unified and generalized McFarland and Spence difference sets. We provide combinatorial analogs of their algebraic results, unifying Steiner ETFs with hyperoval ETFs and Tremain ETFs. | math.FA | math |
Equiangular tight frames from group divisible designs
Matthew Fickusa, John Jasperb
aDepartment of Mathematics and Statistics, Air Force Institute of Technology, Wright-Patterson AFB, OH 45433
bDepartment of Mathematics and Statistics, South Dakota State University, Brookings, SD 57007
Abstract
An equiangular tight frame (ETF) is a type of optimal packing of lines in a real or complex Hilbert
space. In the complex case, the existence of an ETF of a given size remains an open problem in
many cases. In this paper, we observe that many of the known constructions of ETFs are of one
of two types. We further provide a new method for combining a given ETF of one of these two
types with an appropriate group divisible design (GDD) in order to produce a larger ETF of the
same type. By applying this method to known families of ETFs and GDDs, we obtain several
new infinite families of ETFs. The real instances of these ETFs correspond to several new infinite
families of strongly regular graphs. Our approach was inspired by a seminal paper of Davis and
Jedwab which both unified and generalized McFarland and Spence difference sets. We provide
combinatorial analogs of their algebraic results, unifying Steiner ETFs with hyperoval ETFs and
Tremain ETFs.
Keywords:
2010 MSC: 42C15
equiangular tight frames, group divisible designs
1. Introduction
Let N ≥ D be positive integers, let F be either R or C, and let hx1, x2i = x∗
product on FD. The Welch bound [56] states that any N nonzero vectors {ϕn}N
1x2 be the dot
n=1 in FD satisfy
max
n6=n′
hϕn,ϕn′ i
kϕnkkϕn′ k ≥(cid:2) N −D
D(N −1)(cid:3)
1
2 .
(1)
A > 0 such that Akxk2 = PN
n=1 in FD achieve equality in (1) if and
It is well-known [49] that nonzero equal-norm vectors {ϕn}N
only if they form an equiangular tight frame for FD, denoted an ETF(D, N ), namely if there exists
n=1 hϕn, xi2 for all x ∈ FD (tightness), and the value of hϕn, ϕn′i
is constant over all n 6= n′ (equiangularity). In particular, an ETF is a type of optimal packing in
projective space, corresponding to a collection of lines whose minimum pairwise angle is as large as
possible. ETFs arise in several applications, including waveform design for communications [49],
compressed sensing [5, 6], quantum information theory [59, 45] and algebraic coding theory [34].
In the general (possibly-complex) setting, the existence of an ETF(D, N ) remains an open
problem for many choices of (D, N ). See [23] for a recent survey. Beyond orthonormal bases and
regular simplices, all known infinite families of ETFs arise from combinatorial designs. Real ETFs in
particular are equivalent to a class of strongly regular graphs (SRGs) [36, 46, 32, 55], and such graphs
have been actively studied for decades [11, 12, 15]. This equivalence has been partially generalized
to the complex setting in various ways, including approaches that exploit properties of roots of
1
unity [9, 8], abelian distance-regular covers of complete graphs (DRACKNs) [16], and association
schemes [33]. Conference matrices, Hadamard matrices, Paley tournaments and quadratic residues
are related, and lead to infinite families of ETFs whose redundancy N
D is either nearly or exactly
two [49, 32, 44, 48]. Harmonic ETFs and Steiner ETFs offer more flexibility in choosing D and
N . Harmonic ETFs are equivalent to difference sets in finite abelian groups [54, 49, 58, 18], while
Steiner ETFs arise from balanced incomplete block designs (BIBDs) [29, 25]. Recent generalizations
of Steiner ETFs have led to new infinite families of ETFs arising from projective planes that contain
hyperovals [24] as well as from Steiner triple systems [19], dubbed hyperoval ETFs and Tremain
ETFs, respectively. Another new family arises by generalizing the SRG construction of [28] to the
complex setting, using generalized quadrangles to produce abelian DRACKNs [22].
Far less is known in terms of necessary conditions on the existence of complex ETF(D, N ).
The Gerzon bound implies that N ≤ min{D2, (N − D)2} whenever a complex ETF(D, N ) with
N > D > 1 exists [38, 32, 52]. Beyond this, the only known nonexistence result in the complex case
is that an ETF(3, 8) does not exist [51], a result proven using computational techniques in algebraic
geometry. In quantum information theory, ETF(D, D2) are known as symmetric informationally-
complete positive operator-valued measures (SIC-POVMs). It is famously conjectured that such
Gerzon-bound-equality ETFs exist for any D [59, 44, 26].
In this paper, we give a new method for constructing ETFs that yields several new infinite
families of them. Our main result is Theorem 3.2, which shows how to combine a given initial ETF
with a group divisible design (GDD) in order to produce another ETF. In that result, we require
the initial ETF to be of one of the following types:
Definition 1.1. Given integers D and N with 1 < D < N , we say (D, N ) is type (K, L, S) if
D = S
K [S(K − 1) + L] = S2 − S(S−L)
K , N = (S + L)[S(K − 1) + L],
(2)
where K and S are integers and L is either 1 or −1. For a given K, we say (D, N ) is K-positive
or K-negative when it is type (K, 1, S) or type (K, −1, S) for some S, respectively. We simply say
(D, N ) is positive or negative when it is K-positive or K-negative for some K, respectively. When
we say that an ETF is one of these types, we mean its (D, N ) parameters are of that type.
It turns out that every known ETF(D, N ) with N > 2D > 2 is either a harmonic ETF, a
SIC-POVM, or is positive or negative. In particular, every Steiner ETF is positive, while every
hyperoval ETF and Tremain ETF is negative. In this sense, the ideas and results of this paper are
an attempt to unify and generalize several constructions that have been regarded as disparate. This
is analogous to-and directly inspired by-a seminal paper of Davis and Jedwab [17], which unifies
McFarland [41] and Spence [47] difference sets under a single framework, and also generalizes them
so as to produce difference sets whose corresponding harmonic ETFs have parameters
D = 1
3 22J−1(22J+1 + 1), N = 1
3 22J+2(22J − 1),
(3)
for some J ≥ 1. It is quickly verified that such ETFs are type (4, −1, S) where S = 1
3 (22J+1 +1). As
we shall see, combining our main result (Theorem 3.2) with known ETFs and GDDs recovers the
existence of ETFs with these parameters, and also provides several new infinite families, including:
Theorem 1.2. An ETF(D, N ) of type (K, −1, S) exists whenever:
(a) K = 4 and either S ≡ 3 mod 8 or S ≡ 7 mod 60;
2
(b) K = 5 and either S ≡ 4 mod 15 or S ≡ 5, 309 mod 380 or S ≡ 9 mod 280.
This result extends the S for which an ETF of type (4, −1, S) is known to exist from a geometric
progression to a finite union of arithmetic progressions, with the smallest new ETF having S = 19,
namely (D, N ) = (266, 1008), cf. [23]. Meanwhile, the ETFs given by Theorem 1.2 in the K = 5
case seem to be completely new except when S = 4, 5, 9, with (D, N ) = (285, 1350) being the
smallest new example. Using similar techniques, we were also able to find new, explicit infinite
families of K-negative ETFs for K = 6, 7, 10, 12. The description of these families is technical, and
so is given in Theorem 4.4 as opposed to here. More generally, using asymptotic existence results
for GDDs, we show that an infinite number of K-negative ETFs also exist whenever K = Q + 2
where Q is a prime power, K = Q + 1 where Q is an even prime power, or K = 8, 20, 30, 42, 56, 342.
In certain cases, the new ETFs constructed by these methods can be chosen to be real:
Theorem 1.3.
(a) There are an infinite number of real Hadamard matrices of size H ≡ 1 mod 35, and a real
ETF of type (5, −1, 8H + 1) exists for all such H.
(b) There are an infinite number of real Hadamard matrices of size H ≡ 1, 8 mod 21, and a real
ETF of type (6, −1, 2H + 1) exists for all sufficiently large such H.
(c) There are an infinite number of real Hadamard matrices of size H ≡ 1, 12 mod 55, and a real
ETF of type (10, −1, 4H + 1) exists for all sufficiently large such H.
(d) There are an infinite number of real Hadamard matrices of size H ≡ 1, 277 mod 345, and a
real ETF of type (15, −1, 4H + 1) exists for all sufficiently large such H.
These correspond to four new infinite families of SRGs, with the smallest new example being a
real ETF(66759, 332640), which is obtained by letting H = 36 in (a).
In the next section, we introduce known concepts from frame theory and combinatorial design
that we need later on. In Section 3, we provide an alternative characterization of when an ETF is
positive or negative (Theorem 3.1), which we then use to help prove our main result (Theorem 3.2).
In the fourth section, we discuss how many known ETFs are either positive or negative, and then
apply Theorem 3.2 to them along with known GDDs to obtain the new infinite families of negative
ETFs described in Theorems 1.2 and 4.4. We conclude in Section 5, using these facts as the basis
for new conjectures on the existence of real and complex ETFs.
2. Preliminaries
2.1. Equiangular tight frames
For any positive integers N and D, and any sequence {ϕn}N
sponding synthesis operator is Φ : FN → FD, Φy := PN
n=1 of vectors in FD, the corre-
n=1 y(n)ϕn, namely the D × N matrix
whose nth column is ϕn. Its adjoint (conjugate transpose) is the analysis operator Φ∗ : FD → FN ,
which has (Φ∗x)(n) = hϕn, xi for all n = 1, . . . , N . That is, Φ∗ is the D × N matrix whose nth
n. Composing these two operators gives the N × N Gram matrix Φ∗Φ whose (n, n′)th
row is ϕ∗
We say {ϕn}N
entry is (Φ∗Φ)(n, n′) = hϕn, ϕn′i, as well as the D × D frame operator ΦΦ∗ =PN
n=1 is a tight frame for FD if there exists A > 0 such that ΦΦ∗ = AI, namely if
the rows of Φ are orthogonal and have an equal nontrivial norm. We say {ϕn}N
n=1 is equal norm if
there exists some C such that kϕnk2 = C for all n. The parameters of an equal norm tight frame
n=1 ϕnϕ∗
n.
3
are related according to DA = Tr(AI) = Tr(ΦΦ∗) = Tr(Φ∗Φ) = PN
n=1 kϕnk2 = N C. We say
n=1 is equiangular if it is equal norm and the value of hϕn, ϕn′i is constant over all n 6= n′.
{ϕn}N
For any equal norm vectors {ϕn}N
n=1 in FD, a direct calculation reveals
0 ≤ Tr[( 1
C ΦΦ∗ − N
D I)2] =
N
N
Xn=1
Xn′=1
n′6=n
hϕn,ϕn′ i2
C 2
− N (N −D)
D
≤ N (N − 1) max
n6=n′
hϕn,ϕn′ i2
C 2
− N (N −D)
D
.
Rearranging this inequality gives the Welch bound (1). Moreover, we see that achieving equality
in (1) is equivalent to having equality above throughout, which happens precisely when {ϕn}N
is a tight frame for FD that is also equiangular, namely when it is an ETF for FD.
n=1
If N > D and {ϕn}N
n=1 is a tight frame for FD then completing the D rows of Φ to an
equal-norm orthogonal basis for FN is equivalent to taking a (N − D) × N matrix Ψ such that
Ψ∗Ψ = AI, ΦΨ∗ = 0 and Φ∗Φ + Ψ∗Ψ = AI. The sequence {ψn}N
n=1 of columns of any such Ψ is
n=1. Since ΨΨ∗ = AI and Ψ∗Ψ = AI − Φ∗Φ, any Naimark
called a Naimark complement of {ϕn}N
complement of an ETF(D, N ) is an ETF(N − D, N ). Since any nontrivial scalar multiple of an
ETF is another ETF, we will often assume without loss of generality that a given ETF(D, N ) and
its Naimark complements satisfy
1
2 ,
∀n = 1, . . . , N,
(4)
A = N(cid:2) N −1
D(N −D)(cid:3)
1
2 ,
kϕnk2 =(cid:2) D(N −1)
N −D (cid:3)
1
2 ,
kψnk2 =(cid:2) (N −D)(N −1)
D
(cid:3)
which equates to having hϕn, ϕn′i = 1 = hψn, ψn′i for all n 6= n′. For positive and negative
ETFs in particular (Definition 1.1), we shall see that all of these quantities happen to be integers.
Any ETF(D, N ) with N = D + 1 is known as a regular simplex, and such ETFs are Naimark
complements of ETFs for F1, namely sequences of scalars that have the same nontrivial modulus.
n=1 in FN −1 is a Naimark complement of the all-ones
In particular, a sequence of vectors {fn}N
sequence in F1 if and only if
FF∗ = N I,
N
Xn=1
fn = F1 = 0, F∗F = N I − J,
(5)
where 1 and J denote an all-ones column vector and matrix, respectively. Equivalently, the vectors
n=1 in FN are equal norm and orthogonal. In particular, for any N > 1, we can always
{1 ⊕ fn}N
take {1 ⊕ fn}N
n=1 to be the columns of a possibly-complex Hadamard matrix of size N . In this
case, F satisfies (5) and is also flat, meaning every one of its entries has modulus one. As detailed
below, flat regular simplices can be used to construct several families of ETFs, including Steiner
ETFs as well as those we introduce in Theorem 3.2.
Harmonic ETFs are the best-known class of ETFs [48, 58, 18]. A harmonic ETF(D, N ) is
obtained by restricting the characters of an abelian group G of order N to a difference set of
cardinality D, namely a D-element subset D of G with the property that the cardinality of
{(d, d′) ∈ D × D : g = d − d′} is constant over all g ∈ G, g 6= 0. The set complement G\D of
any difference set in G is another difference set, and the two corresponding harmonic ETFs are
Naimark complements.
In particular, for any abelian group G of order N , the harmonic ETF
arising from G\{0} is a flat regular simplex that satisfies (5).
4
2.2. Group divisible designs
For a given integer K ≥ 2, a K-GDD is a set V of V > K vertices, along with collections G
and B of subsets of V, called groups and blocks, respectively, with the property that the groups
partition V, every block has cardinality K, and any two vertices are either contained in a common
group or a common block, but not both. A K-GDD is uniform if its groups all have the same
cardinality M , denoted in exponential notation as a "K-GDD of type M U " where V = U M .
Letting B be the number of blocks, a {0, 1}-valued B × U M incidence matrix X of a K-GDD
of type M U has the property that each row of X contains exactly K ones. Moreover, for any
v = 1, . . . , V = U M , the vth column of X is orthogonal to M − 1 other columns of X, and has a
dot product of 1 with each of the remaining (U − 1)M columns. This implies that the replication
number Rv of blocks that contain the vth vertex satisfies
(U −1)M =
V
Xv′=1
v′6=v
(X∗X)(v, v′) =
X(b, v)
B
Xb=1
V
Xv′=1
v′6=v
X(b, v′) =
B
Xb=1(cid:26) K − 1, X(b, v) = 1
0, X(b, v) = 0(cid:27) = Rv(K −1).
As such, this number Rv = R is independent of v. At this point, summing all entries of X gives
BK = V R and so B is also uniquely determined by K, M and U . Because of this, the existence
of a K-GDD of type M U is equivalent to that of a {0, 1}-valued B × U M matrix X with
R = M (U −1)
K−1 , B = M U R
K = M 2U (U −1)
K(K−1) , X1 = K1, X∗X = R I + (JU − IU ) ⊗ JM .
(6)
In the special case where M = 1, a K-GDD of type 1U is called a BIBD(U, K, 1). In the special case
where U = K, a K-GDD of type M K is called a transversal design TD(K, M ), which is equivalent
to a collection of K − 2 mutually orthogonal Latin squares (MOLS) of size M .
In order for a K-GDD of type M U to exist, the expressions for R and B given in (6) are
necessarily integers. Beyond this, we necessarily have U ≥ K since we can partition any given
block into its intersections with the groups, and the cardinality of these intersections is at most
one. Altogether, the parameters of a K-GDD of type M U necessarily satisfy
U ≥ K, M (U −1)
K−1 ∈ Z, M 2U (U −1)
K(K−1) ∈ Z.
(7)
Though these necessary conditions are not sufficient [27], they are asymptotically sufficient in two
distinct ways: for any fixed K ≥ 2 and M ≥ 1, there exists U0 = U0(K, M ) such that a K-GDD
of type M U exists for all U ≥ U0 such that (7) is satisfied [13, 37]; for any fixed U ≥ K ≥ 2, there
exists M0 = M0(K, U ) such that a K-GDD of type M U exists for all M ≥ M0 such that (7) is
satisfied [43]. In the M = 1 and U = K cases, these facts reduce to more classical asymptotic
existence results for BIBDs and MOLS, respectively.
Many specific examples of GDDs are formed by combining smaller designs in clever ways. We
in particular will make use of the following result, which is a special case of Wilson's approach [57]:
Lemma 2.1. If a K-GDD of type M U and a U -GDD of type N V exist, then a K-GDD of type
(M N )V exists.
Proof. Let X and Y be incidence matrices of the form (6) for the given K-GDD of type M U
and U -GDD of type N V , respectively.
In particular, taking R and B as in (6), we can write
uXu′ = J for any
X = (cid:2) X1 · · · XU (cid:3) where each Xu is a B × M matrix with X∗
uXu = RI, and X∗
5
u 6= u′. We now construct the incidence matrix Z of a K-GDD of type (M N )V in the following
manner: in each row of Y, replace each of the U nonzero entries with a distinct matrix Xu, and
replace each of the zero entries with a B × M matrix of zeros.
This result generalizes MacNeish's classical method for combining MOLS [40]: if a TD(K, M )
and a TD(K, N ) exist, then applying Lemma 2.1 to them produces a TD(K, M N ). We will also
use one GDD to "fill the holes" of another:
Lemma 2.2. If K-GDDs of type M U and (M U )V exist, then a K-GDD of type M U V exists.
Proof. Letting X and Y be incidence matrices of the form (6) for the given K-GDDs of type M U
and (M U )V , respectively, it is straightforward to verify that
is the incidence matrix of a K-GDD of type M U V .
Z =(cid:20) IV ⊗ X
Y (cid:21)
2.3. Previously known constructions of ETFs involving BIBDs and MOLS
In the next section, we introduce a method for constructing ETFs that uses GDDs. This
method makes use of a concept from [19], which we now generalize from BIBDs to GDDs:
Definition 2.3. Take a K-GDD of type M U where M ≥ 1 and U ≥ K ≥ 2, and define R, B
and an incidence matrix X according to (6). Without loss of generality, write the columns of X as
{xu,m}U
m=1 have disjoint support. Then, for any u
and m, a corresponding embedding operator Eu,m is any {0, 1}-valued B × R matrix whose columns
are standard basis elements that sum to xu,m.
m=1 where, for each u, the vectors {xu,m}M
M
u=1,
letting {Ev}V
v=1 = {Eu,1}U
In the special case where M = 1, this concept leads to an elegant formulation of Steiner
u=1 be the embedding operators of a BIBD(V, K, 1), and
ETFs [19]:
letting {1 ⊕ fi}R
K−1 + 1,
the V (R + 1) vectors {Evfi}V
i=0 form an ETF for FB. In [19], this fact is proven using several
properties of embedding operators. We now show those properties generalize to the GDD setting;
later on, we use these facts to prove our main result:
i=0 be the columns of a possibly-complex Hadamard matrix of size R + 1 = V −1
v=1,
R
Lemma 2.4. If {Eu,m}U
u=1,, M
m=1 are the embedding operators arising from a K-GDD of type M U ,
E∗
u,mEu′,m′ =
u = u′, m = m′,
u = u′, m 6= m′,
I,
0,
δrδ∗
r′, u 6= u′.
Here, for any u 6= u′ and m, m′, δr and δr′ are standard basis elements in FR whose indices r, r′
depend on u, u′, m, m′.
Proof. Each Eu,m is a matrix whose columns are standard basis elements that sum to xu,m, and
so is an isometry, that is, E∗
u,mEu′,m′ is a matrix
whose entries are nonnegative integers that sum to:
u,mEu,m = I. Moreover, for any u, u′, m, m′, E∗
R
R
Xr=1
Xr′=1
(E∗
u,mEu′,m′)(r, r′) = 1∗E∗
u,mEu′,m′ 1 = hxu,m, xu′,m′i =
6
R, u = u′, m = m′,
0, u = u′, m 6= m′,
1, u′ 6= u.
u,mEu′,m′ = 0. If instead u 6= u′ then this implies that
u,mEu′,m′ has a single nonzero entry, and that this entry has value 1. This means there exists
When u = u′ and m 6= m′, this implies E∗
E∗
some r, r′ = 1, . . . , R, r = r(u, m, u′, m′), r′ = r′(u, m, u′, m′) such that E∗
u,mEu,m′ = δrδ∗
r′.
Other Steiner-like constructions of ETFs include hyperoval ETFs [24] and Tremain ETFs [22].
Beyond Steiner and Steiner-like techniques, there are at least two other methods for constructing
ETFs that make direct use of the incidence matrix of some kind of GDD. One method leads to
the phased BIBD ETFs of [22]: if X is the B × V incidence matrix of a BIBD(V, K, 1), and Φ is
any matrix obtained by replacing each 1-valued entry of X with any unimodular scalar, then the
columns of Φ are immediately equiangular, and the challenge is to design them so that they form
a tight frame for their span. Another method constructs ETFs with (D, N ) = ( 1
2 M (M ± 1), M 2)
from MOLS. To elaborate, a TD(K, M ) is a K-GDD of type M K , meaning by (6) that it has an
M 2 × KM incidence matrix X that satisfies
X1 = K1, X∗X = M I + (JK − IK) ⊗ JM .
(8)
Here, the columns of X have support M , and are arranged as K groups of M columns apiece,
where columns in a common group have disjoint support. Together, these facts imply, in turn, that
X(IK ⊗ 1M ) = 1M 21∗
K ,
(XX∗)2 = M XX∗ + K(K − 1)J.
(9)
At this point, the traditional approach is to let A = XX∗ − KI be the adjacency matrix of the
TD's block graph, and use (8) and (9) to show that this graph is strongly regular with parameters
(M 2, K(M − 1), M + K(K − 3), K(K − 1)). In the M = 2K case, applying Theorem 4.4 of [21]
to this graph then produces a real ETF with (D, N ) = ( 1
2 M (M − 1), M 2) whose vectors sum
to zero, while applying this same result in the M = 2(K − 1) case produces a real ETF with
(D, N ) = ( 1
2 M (M + 1), M 2) whose synthesis operator's row space contains the all-ones vector.
That said, a careful read of the literature reveals that this construction can be made more
explicit, and that doing so has repercussions for coding theory. To elaborate, in [10], MOLS
are used to produce quasi-symmetric designs (QSDs) which, via the techniques of [42], yield self-
complementary binary codes that achieve equality in the Grey-Rankin bound. In [34], such codes
are shown to be equivalent to flat real ETFs. A method for directly converting the incidence matri-
ces of certain QSDs into synthesis operators of ETFs was also recently introduced [20]. Distilling
these ideas leads to the following streamlined construction:
let X be the incidence matrix of a
TD(K, M ), let {1 ⊕ fm}M
m=1 be the columns of a possibly-complex Hadamard matrix of size M , let
m=1, and consider the K(M − 1) × M 2 matrix
F be the (M − 1) × M synthesis operator of {fm}M
Φ = (IK ⊗ F)X∗.
(10)
Using (5), (8), and (9) along with the fact that F is flat, it is straightforward to show that Φ
is flat and satisfies ΦΦ∗ = M 2I and Φ∗Φ = M XX∗ − KJ. As such, the columns of Φ form
a flat two-distance tight frame (TDTF) for FK(M −1) [7]. Moreover, this TDTF is an ETF when
M = 2K.
In particular, if there exists a TD(K, 2K) and a real Hadamard matrix of size 2K,
then there exists a flat real ETF(K(2K − 1), 4K 2). Using the equivalence between flat real ETFs
and Grey-Rankin-bound-equality codes given in [34], or alternatively the equivalence between such
ETFs and certain QSDs given in [20], this recovers Theorem 1 of [10]. In the K = 6 case, that
result gives the only known proof to date of the existence of a flat real ETF(66, 144).
7
For any TD and corresponding flat regular simplex, it is quickly verified that the corresponding
TDTF (10) is centered [21] in the sense that Φ1 = 0, namely that the all-ones vector is orthogonal
to the row space of Φ. This fact leads to an analogous reinterpretation of the second main result
of [10]: in lieu of (10), we instead consider the [K(M − 1) + 1] × M 2 flat matrix
Ψ =(cid:20) 1∗
Φ (cid:21) =(cid:20)
1∗
(IK ⊗ F)X∗(cid:21) .
(11)
Here, the properties of Φ immediately imply ΨΨ∗ = M 2I and Ψ∗Ψ = M XX∗ −(K −1)J, meaning
the columns of Ψ form a flat TDTF. However, unlike (10), the columns of (11) are equiangular
precisely when M = 2(K − 1). Replacing K with K + 1, this implies in particular that if there
exists a TD(K + 1, 2K) and a real Hadamard matrix of size 2K, then there exists a flat real
ETF(K(2K + 1), 4K 2). This recovers Theorem 2 of [10] via the equivalences of [34, 20], and gives
the only known proof of the existence of a flat real ETF(78, 144).
Simply put, if certain TDs exist, then certain ETFs exist. In the next section, we introduce a
new method of constructing ETFs from TDs.
3. Constructing equiangular tight frames with group divisible designs
In [17], Davis and Jedwab unify McFarland [41] and Spence [47] difference sets under a single
framework, and also generalize them so as to produce difference sets with parameters (3). McFar-
land's construction relies on nice algebro-combinatorial properties of the set of all hyperplanes in
a finite-dimensional vector space over a finite field. Davis and Jedwab exploit these properties to
form various types of building sets, which in some cases lead to difference sets.
In [34], it is shown that every harmonic ETF arising from a McFarland ETF is unitarily-
equivalent to a Steiner ETF arising from an affine geometry. When we applied a similar analysis to
the building sets of [17], we discovered that they have an underlying TD-like incidence structure.
(We do not provide this analysis here since it is nontrivial and does not help us prove our results
in their full generality.) This eventually led us to the ETF construction technique of Theorem 3.2
below. In short, our approach here is directly inspired by that of [17], though this is not apparent
from our proof techniques. In particular, the fact that the L parameter in Definition 1.1 is either
1 or −1 is a generalization of Davis and Jedwab's notion of extended building sets with "+" and
"−" parameters, respectively. To facilitate our arguments later on, we now consider these types of
parameters in greater detail:
Theorem 3.1. If 1 < D < N and (D, N ) is type (K, L, S), see Definition 1.1, then
where S ≥ 2. Conversely, given (D, N ) such that 1 < D < N , and letting L be either 1 or −1, if
the above expressions for S and K are integers then (D, N ) is type (K, L, S).
S =(cid:2) D(N −1)
N −D (cid:3)
1
2 , K = N S
D(S+L) ,
(12)
Moreover, in the case that the equivalent conditions above hold, scaling an ETF {ϕn}N
FD so that hϕn, ϕn′i = 1 for all n 6= n′ gives that it and its Naimark complements {ψn}N
tight frame constant A = K(S + L) and
n=1 for
n=1 have
kϕnk2 =(cid:2) D(N −1)
N −D (cid:3)
1
2 = S,
kψnk2 =(cid:2) (N −D)(N −1)
D
8
1
2 = S(K − 1) + KL,
∀n = 1, . . . , N.
(13)
(cid:3)
Proof. Whenever (D, N ) is type (K, L, S) we have that L is either 1 or −1 by assumption, at which
point the fact that L2 = 1 coupled with (2) gives
N − 1 = (S + L)[S(K − 1) + L] − 1 = S2(K − 1) + SKL = S[S(K − 1) + KL],
D − 1 = K(S+L)[S(K−1)+L]
S[S(K−1)+L] − 1 = K(S+L)
S [S(K − 1) + KL].
− 1 = 1
N
S
Multiplying and dividing these expressions immediately implies that
1
2 = S,
(cid:2) D(N −1)
N −D (cid:3)
(cid:2) (N −D)(N −1)
D
(cid:3)
1
2 = S(K − 1) + KL.
(14)
Here, S is an integer by assumption, and is clearly positive. Moreover, if S = 1 then (14) implies
D = 1. Since D > 1 by assumption, we thus have S ≥ 2. Continuing, (2) further implies
N S
D(S+L) = K(S+L)[S(K−1)+L]
S[S(K−1)+L]
S
S+L = K.
Conversely, now assume that S and K are defined by (12), where L is either 1 or −1, and that
S and K are integers. As before, the fact that D > 1 implies that S ≥ 2 and so K > 0. We solve
for N in terms of D, K, and L. Here, (12) gives N
S . Since L2 = 1, this implies
S = 1 + L
DK = S+L
Squaring this equation and multiplying the result by N − 1 gives
1
2 = 1
(cid:2) N −D
D(N −1)(cid:3)
D = (N − 1)(cid:0) N
S = L(cid:0) N
DK − 1(cid:1)2 = N(cid:0) N
DK − 1(cid:1).
DK − 1(cid:1)2 − N
N
D − 1 = N −D
Adding 1 to this equation and multiplying by (DK)2
N
then leads to a quadratic in N :
DK(cid:0) N
DK − 2(cid:1) − 1.
DK 2 = (N − DK)2 − (N − 2DK) = N 2 − (2DK + 1)N + DK(DK + 2).
Applying the quadratic formula then gives
(15)
(16)
(17)
(18)
N = DK + 1
1
2(cid:9).
2(cid:8)1 ±(cid:2)4DK(K − 1) + 1(cid:3)
2DK(cid:8)1 ±(cid:2)4DK(K − 1) + 1(cid:3)
As such, (15) becomes 1
here correspond to L = 1 and L = −1 respectively, that is,
S = L(cid:0) N
DK − 1(cid:1) = L
2DK(cid:8)1 + L(cid:2)4DK(K − 1) + 1(cid:3)
1
1
S = L
2(cid:9) = 1
2DK(cid:8)L +(cid:2)4DK(K − 1) + 1(cid:3)
2(cid:9).
1
2(cid:9), implying "+" and "−"
1
Moreover, since N is an integer, (16) implies that 4DK(K − 1) + 1 is the square of an odd integer,
that is, that 4DK(K −1)+1 = (2J −1)2 = 4J(J −1)+1 or equivalently that DK(K −1) = J(J −1)
for some positive integer J. Writing D = J(J−1)
K(K−1), (16) and (17) then become
N = DK + 1
2(cid:8)1 + L(cid:2)4DK(K − 1) + 1(cid:3)
1
S = 1
1
2DK(cid:8)L +(cid:2)4DK(K − 1) + 1(cid:3)
2(cid:9) = K−1
1
2 [1 + L(2J − 1)],
2(cid:9) = DK + 1
2J(J−1) [L + (2J − 1)] =( K−1
J−1 , L = 1
K−1
J , L = −1) = 2(K−1)
2J−L−1 .
9
That is, J = S(K − 1) + 1
the fact that L2 = 1 then gives the expressions for D and N given in Definition 1.1:
2 (L + 1). Substituting this into D = J(J−1)
K(K−1) and (18) and again using
[S(K−1)+ 1
2 (L+1)][S(K−1)+ 1
2 (L−1)]
D =
N = S[S(K − 1) + L] + 1
K(K−1))
= S 2(K−1)2+S(K−1)L
K [S(K − 1) + L],
2 {1 + L[2S(K − 1) + L]} = (S + L)[S(K − 1) + L].
K(K−1))
= S
Finally, in the case where (D, N ) is type (K, L, S), if {ϕn}N
n=1 is an ETF for FD, and is without
loss of generality scaled so that hϕn, ϕn′i = 1 for all n 6= n′, then (4) and (14) immediately imply
that {ϕn}N
n=1 satisfy (13), and that both are
tight frames with tight frame constant A = N S
n=1 and any one of its Naimark complements {ψn}N
D = K(S + L).
Theorem 3.1 implies that the (D, N ) parameters of an ETF with N > D > 1 are type (K, L, S)
with K = 1 if and only if that ETF is a regular simplex, and moreover that this only occurs
when L = 1 and S = D.
Indeed, for any D > 1, the pair (D, N ) = (D, D + 1) satisfies (2)
when (K, L, S) = (1, 1, D). Conversely, in light of (13), an ETF of type (1, L, S) has a Naimark
n=1 with the property that kψnk2 = L = 1 and hψn, ψn′i = 1 for all n 6= n′,
complement {ψn}N
namely a Naimark complement that is an ETF(1, N ).
We also emphasize that it is sometimes possible for the parameters of a single ETF to be
In particular, if D > 1 and
simultaneously positive and negative for different choices of K.
(D, D + 1) is type (K, −1, S), then (12) gives S = D and K = N S
D−1 .
Since K is an integer, this implies either D = 2 or D = 3. And, letting (K, L, S) be (3, −1, 2) and
(2, −1, 3) in (2) indeed gives that (D, N ) is (2, 3) and (3, 4), respectively.
D(S+L) = (D+1)D
D(D−1) = D+1
D−1 = 1 + 2
In the next section, we provide a much more thorough discussion of positive and negative ETFs,
including some other examples of ETFs that are both. For now, we turn to our main result, which
shows how to combine a given ETF(D, N ) whose parameters are type (K, L, S) with a certain
K-GDD to produce a new ETF whose parameters are type (K, L, S′) for some S′ > S. Here,
as with any GDD, we require K ≥ 2. In light of the above discussion, this is not a significant
restriction since (D, N ) has type (1, L, S) if and only if L = 1 and S = D, and we already know
that ETFs of type (1, 1, D) exist for all D > 1, being regular simplices.
Theorem 3.2. Assume an ETF of type (K, L, S) exists where K ≥ 2, and let M = S(K − 1) + L.
The necessary conditions (7) on the existence of a K-GDD of type M U reduce to having
Moreover, if such a GDD exists, and U has the additional property that
U ≥ K,
U −1
K−1 ∈ Z,
(S−L)U (U −1)
K(K−1)
∈ Z.
(K−2)(U −1)
(S+L)(K−1) ∈ Z,
(19)
(20)
then there exists an ETF of type (K, L, S′) where S′ = S + R = M U −L
K−1 , where R = M (U −1)
K−1 .
In particular, under these hypotheses, W := R
given ETF as {ϕm,i}M
embedding operators of the GDD (Definition 2.3), letting {δu}U
and letting {ei}S+L
S + L and W + 1, respectively, then the following vectors form an ETF of type (K, L, S′):
S+L ∈ Z, and without loss of generality writing the
M
m=1 be the
u=1 be the standard basis for FU ,
j=0 be the columns of possibly-complex Hadamard matrices of size
S+L
i=1 where kϕm,ik2 = S for all m and i, letting {Eu,m}U
i=1 and {1 ⊕ fj}W
m=1,
u=1,
{ψu,m,i,j}U
u=1,
M
m=1,
S+L
i=1,
W
As special cases of this fact, an ETF of type (K, L, S′) exists whenever either:
j=0, ψu,m,i,j := (δu ⊗ ϕm,i) ⊕(cid:0)Eu,m(ei ⊗ fj)(cid:1).
(21)
10
(a) U is sufficiently large and satisfies (19) and (20);
(b) there exists a K-GDD of type M U , provided we also have that S + L divides K − 2.
Proof. Since M = S(K − 1) + L,
M
K−1 = S + L
K−1 = (S + L) − L(cid:0) K−2
K−1(cid:1), M
K = S − S−L
K .
(22)
As such, the replication number of any K-GDD of type M U is R = M (U −1)
K−1). In
particular, such a GDD can only exist when K − 1 necessarily divides U − 1. Moreover, multiplying
the expressions in (22) gives that the number of blocks in any such GDD is
K−1 = S(U −1)+L( U −1
B = M 2U (U −1)
K(K−1)
=(cid:0)S − S−L
K (cid:1)(cid:0)S + L
= S2U (U − 1) + LSU U −1
K−1(cid:1)U (U − 1)
K−1 − S(S−L)
K U (U − 1) − L (S−L)U (U −1)
K(K−1)
.
Here, since our initial ETF(D, N ) is type (K, L, S), S(S−L)
K = S2 −D is an integer, and so the above
expression for B is an integer precisely when K(K − 1) divides (S − L)U (U − 1). To summarize,
since M = S(K − 1) + L where K divides S(S − L), the necessary conditions (7) on the existence
of a K-GDD of type M U reduce to having (19). These necessary conditions are known to be
asymptotically sufficient [13, 37]: for this fixed K and M , there exists U0 such that a K-GDD of
type M U exists for any U ≥ U0 that satisfies (19). Regardless, to apply our construction below
with any given K-GDD of type M U , we only need U to satisfy the additional property that
W = R
S+L = M (U −1)
(S+L)(K−1) = U −1
S+L(cid:2)(S + L) − L(cid:0) K−2
K−1(cid:1)(cid:3) = (U − 1) − L (K−2)(U −1)
(S+L)(K−1)
is an integer, namely to satisfy (20). Since K − 1 necessarily divides U − 1, this is automatically
satisfied whenever S + L happens to divide K − 2, and some of the ETFs we will identify in the
next section will have this nice property. Regardless, there are always an infinite number of values
of U which satisfy (19) and (20), including, for example, all U ≡ 1 mod (S + L)K(K − 1).
Turning to the construction itself, the fact that the given ETF(D, N ) is type (K, L, S) implies
D = S
K [S(K − 1) + L] = SM
K , N = (S + L)[S(K − 1) + L] = (S + L)M.
(23)
In particular, since N = (S + L)M , the vectors in our initial ETF(D, N ) can indeed be indexed as
{ϕm,i}M
S+L
i=1 . Moreover,
m=1,
M U = M (U − 1) + M = (K − 1)R + [S(K − 1) + L] = (S + R)(K − 1) + L.
(24)
As such, the number of vectors in the collection (21) is
N ′ = U M (S + L)(W + 1) = (S + L)( R
S+L + 1)M U = [(S + R) + L][(S + R)(K − 1) + L].
(25)
Also, for each i and j, ei ⊗ fj lies in a space of dimension F(S+L)W = FR. And, for each u and m,
Eu,m is a B × R matrix. As such, for any u, m, i and j, ψu,m,i,j is a well-defined vector in FD′
where, by combining (6), (23) and (24), we have
D′ = U D + B = U SM
K + M U
K R = S+R
K M U = S+R
11
K [(S + R)(K − 1) + L].
(26)
Comparing (25) and (26) against (2), we see that (D′, N ′) is indeed type (K, L, S′) where S′ = S+R.
Here, (24) further implies S′ = S + R = M U −L
K−1 .
Continuing, since (D′, N ′) is type (K, L, S′), Theorem 3.1 gives that the Welch bound for N ′ vec-
tors in FD′
, it suffices to prove that kψu,m,i,jk2 = S′
for all u, m, i, j, and that hψu,m,i,j, ψu′,m′,i′,j ′i is unimodular whenever (u, m, i, j) 6= (u′, m′, i′, j′).
Here, for any u, u′ = 1, . . . , U , m, m′ = 1, . . . , M , i, i′ = 1, . . . , S + L, j, j′ = 0, . . . , W ,
S ′ . As such, to show (21) is an ETF for FD′
is 1
hψu,m,i,j, ψu′,m′,i′,j ′i = h(δu ⊗ ϕm,i) ⊕(cid:0)Eu,m(ei ⊗ fj)(cid:1), (δu′ ⊗ ϕm′,i′) ⊕(cid:0)Eu′,m′(ei′ ⊗ fj ′)(cid:1)i
= hδu, δu′ihϕm,i, ϕm′,i′i + hEu,m(ei ⊗ fj), Eu′,m′(ei′ ⊗ fj ′)i
= hδu, δu′ihϕm,i, ϕm′,i′i + hei ⊗ fj, E∗
u,mEu′,m′(ei′ ⊗ fj ′)i.
When u = u′, m = m′, i = i′ and j = j′, (27) indeed becomes
(27)
kψu,m,i,jk2 = kδuk2kϕm,ik2 + keik2kfjk2 = 1S + (S + L)W = S + R = S′.
As such, all that remains is to show that (27) is unimodular in all other cases. For instance, if u 6= u′
u,mEu′,m′ = δrδ∗
then Lemma 2.4 gives that for any m, m′ there exists r, r′ = 1, . . . , R such that E∗
r′
meaning in this case (27) becomes
hψu,m,i,j, ψu′,m′,i′,j ′i = 0hϕm,i, ϕm′,i′i + hei ⊗ fj , δrδ∗
r′(ei′ ⊗ fj ′)i = (ei ⊗ fj)(r)(ei′ ⊗ fj ′)(r′),
which is unimodular, being a product of unimodular numbers.
m 6= m′ then Lemma 2.4 gives E∗
u,mEu,m′ = 0 and so (27) becomes
If we instead have u = u′ and
hψu,m,i,j, ψu,m′,i′,j ′i = kδuk2hϕm,i, ϕm′,i′i + hei ⊗ fj, 0(ei′ ⊗ fj ′)i = hϕm,i, ϕm′,i′i,
m=1,
S+L
which is unimodular since {ϕm,i}M
i=1 is an ETF of type (K, L, S), and has been scaled so that
kϕm,ik2 = S for all m and i. Next, if we instead have u = u′ and m = m′ then Lemma 2.4 gives
E∗
u,mEu,m = I and so (27) becomes
hψu,m,i,j, ψu,m,i′,j ′i = kδuk2hϕm,i, ϕm,i′i + hei ⊗ fj, ei′ ⊗ fj ′i = hϕm,i, ϕm,i′i + hei, ei′ihfj , fj ′i. (28)
In particular, when u = u′, m = m′ and i 6= i′, the fact that {ei}S+L
is orthogonal implies
i=1
that (28) reduces to hϕm,i, ϕm,i′i, which is unimodular for the same reason as the previous case.
The final remaining case is the most interesting: when u = u′, m = m′, i = i′ but j 6= j′, we
have 0 = h1 ⊕ fj, 1 ⊕ fj ′i = 1 + hfj, fj ′i and so hfj , fj ′i = −1; when combined with the fact that
kϕm,ik2 = S and keik2 = S + L, this implies that in this case (28) becomes
hψu,m,i,j, ψu,m,i,j ′i = kϕm,ik2 + keik2hfj , fj ′i = S + (S + L)(−1) = −L,
where, in Definition 1.1, we have assumed that L is either 1 or −1.
The construction of Theorem 3.2 leads to the concept of ETFs of type (K, L, S). To clarify, the
construction (21) and the above proof of its equiangularity is valid for any initial ETF(D, N ) and
S+L for some L ∈ {−1, 1}, and S +L divides R = M (U −1)
any K-GDD of type M U , provided M = N
K−1 .
However, it turns out that the first U D rows of the corresponding synthesis operator have squared-
norm (W + 1) N S
D , whereas the last B rows have squared-norm (W + 1)K(S + L). As such, the
equiangular vectors (21) are only a tight frame when K = N S
D(S+L) . Applying the techniques of the
proof of Theorem 3.1 then leads to the expressions for (D, N ) in terms of (K, L, S) given in (2).
These facts are not explicitly discussed in our proof above since any equal-norm vectors that attain
the Welch bound are automatically tight.
12
Remark 3.3. There is no apparent value to recursively applying Theorem 3.2. To elaborate,
given an ETF(D, N ) of type (K, L, S) and a K-GDD of type M U where M = S(K − 1) + L and U
satisfies (20), Theorem 3.2 yields an ETF(D′, N ′) of type (K, L, S′) where S′ = M U −L
K−1 . The "M "
parameter of this new ETF is thus M ′ = S′(K − 1) + L = (M U − L) + L = M U , and we can
apply Theorem 3.2 a second time provided we have a K-GDD of type (M U )U ′
where U ′ satisfies
the appropriate analog of (20), namely
(K−2)(U ′−1)
(S ′+L)(K−1) ∈ Z.
(29)
Doing so yields an ETF of type (K, −1, S′′) where S′′ = M ′U ′−L
. However, under
these hypotheses, there is a simpler way to construct an ETF of this same type. Indeed, using the
first GDD to fill the holes of the second GDD via Lemma 2.2 produces a K-GDD of type M U U ′
.
Moreover, U U ′ is a value of "U " that satisfies (20): since S′ = S + R,
K−1 = M U U ′−L
K−1
S+L = S ′−S
S ′+L
S+L + 1 = R
S+L + 1 = W + 1 ∈ Z,
and this together with (20) and (29) imply
(K−2)(U U ′−1)
(S+L)(K−1) = (K−2)[U (U ′−1)+(U −1)]
(S+L)(K−1)
= U (S ′+L)
(S+L)
(K−2)(U ′−1)
(S ′+L)(K−1) + (K−2)(U −1)
(S+L)(K−1) ∈ Z.
As such, we can combine our original ETF of type (K, L, S) with our K-GDD of type M U U ′
Theorem 3.2 to directly produce an ETF of type (K, L, M U U ′−L
).
K−1
via
We also point out that, in a manner analogous to how every McFarland difference set can
be viewed as a degenerate instance of a Davis-Jedwab difference set [17], every Steiner ETF can
be regarded as a degenerate case of the construction of Theorem 3.2. Here, in a manner con-
sistent with (2), we regard (D, N ) = (0, 1) as being type (K, L, S) = (K, 1, 0) where K ≥ 1 is
arbitrary. Under this convention, M = S(K − 1) + L = 1, R = U −1
K−1 and W = R, meaning
we need a K-GDD of type 1U , namely a BIBD(V, K, 1) where V = U . When such a BIBD ex-
ists, Theorem 3.2 suggests we let {ϕ1,1} be some fictitious ETF for the nonexistent space F0,
1
scaled so that kϕ1,1k2 = 0. It also suggests we let {Ev}V
m=1 be the embedding
operators of the BIBD, let e1 = 1, and let {1 ⊕ fj}R
j=0 be the columns of a possibly-complex
Hadamard matrix of size R. Under these conventions, (21) reduces to a collection of V (R + 1)
vectors {ψv,j}V
R
j=0, where for any v = u and j, the fact that ϕ1,1
lies in a "zero-dimensional space" makes it reasonable to regard
j=0 = {ψu,m,i,j}U
v=1 = {Eu,m}U
1
m=1,
1
i=1,
u=1,
u=1,
v=1,
R
ψv,j = ψu,1,1,j = (δv ⊗ ϕ1,1) ⊕(cid:0)Eu,1(e1 ⊗ fj)(cid:1) = Eu,1(1 ⊗ fj) = Eufj .
Here, Theorem 3.2 leads us to expect that {Evfj}V
v=1,
This is indeed the case: as discussed in the previous section, {Evfj}V
R
j=0 is an ETF of type (K, 1, S+R) = (K, 1, R).
R
j=0 is by definition a
v=1,
Steiner ETF for FB where B = V R
K , and moreover letting (K, L, S) = (K, 1, R) in (2) gives:
S
K [S(K − 1) + L] = R
K [R(K − 1) + 1] = R
K V = B,
(S + L)[S(K − 1) + L] = (R + 1)[R(K − 1) + 1] = V (R + 1).
In particular, every Steiner ETF is a positive ETF.
As a degenerate case of Remark 3.3, we further have that when a Steiner ETF is regarded
as being positive, applying Theorem 3.2 to it yields an ETF whose parameters match those of
13
another Steiner ETF. Indeed, since a Steiner ETF arising from a BIBD(V, K, 1) is type (K, 1, R)
where R = V −1
K−1, we can only apply Theorem 3.2 to it whenever there exists a K-GDD of type
M U where M = R(K − 1) + 1 = V and U satisfies (20). In this case, the resulting ETF is type
(K, 1, V U −1
K−1 ). However, under these same hypotheses, we can more simply use the BIBD(V, K, 1)
to fill the holes of the K-GDD of type V U via Lemma 2.2 to obtain a BIBD(U V, K, 1), and its
corresponding Steiner ETF is type (K, 1, V U −1
K−1 ).
4. Families of positive and negative equiangular tight frames
In this section, we use Theorems 3.1 and 3.2 to better our understanding of positive and negative
ETFs, in particular proving the existence of the new ETFs given in Theorems 1.2 and 4.4. Recall
that by (2), any ETF(D, N ) of type (K, L, S) has D = S
K , and so K
necessarily divides S(S − L). As we shall see, it is reasonable to conjecture that such ETFs exist
whenever this necessary condition is satisfied.
K [S(K − 1) + L] = S2 − S(S−L)
4.1. Positive equiangular tight frames
In light of Definition 1.1 and Theorem 3.1, an ETF(D, N ) with 1 < D < N is positive if and
only if there exists integers K ≥ 1 and S ≥ 2 such that
D = S
K [S(K − 1) + 1] = S2 − S(S−1)
K , N = (S + 1)[S(K − 1) + 1],
(30)
2 and K = N S
D(S+1) are integers.
or equivalently that S =(cid:2) D(N −1)
N −D (cid:3)
1
As discussed in the previous section, every regular simplex is a 1-positive ETF and vice versa,
and moreover, every Steiner ETF arising from a BIBD(V, K, 1) is type (K, 1, R) where R = V −1
K−1 .
Here, the fact that K necessarily divides S(S − 1) = R(R − 1) is equivalent to having that D = B
is necessarily an integer. Fisher's inequality also states that a BIBD(V, K, 1) can only exist when
K ≤ R. These necessary conditions on the existence of a BIBD(V, K, 1) are known to also be
sufficient when K = 2, 3, 4, 5, and also asymptotically sufficient in general: for any K ≥ 2, there
exists V0 such that for all V ≥ V0 with the property that R = V −1
K(K−1) , a
BIBD(V, K, 1) exists [2]. Moreover, explicit infinite families of such BIBDs are known, including
affine geometries, projective geometries, unitals and Denniston designs [25], and each thus gives
rise to a corresponding infinite family of (positive) Steiner ETFs.
K−1 and B = V R
K = V (V −1)
That said, not every positive ETF is a Steiner ETF. In particular, for any prime power Q, there
is a phased BIBD ETF [22] whose Naimark complements are ETF(D, N ) where
D = Q3+1
Q+1 , N = Q3 + 1, S =(cid:0) N −1
D −1(cid:1)
N
1
2 =(cid:0) Q3
Q (cid:1)
1
2 = Q, K = N S
D(S+1) = (Q+1)Q
Q+1 = Q,
namely an ETF of type (Q, 1, Q). These parameters match those of a Steiner ETF arising from
a projective plane of order Q − 1. However, such an ETF can exist even when no such projective
plane exists. For example, such an ETF exists when Q = 7 despite the fact that no projective
plane of order Q − 1 = 6 exists.
Other ETFs of type (K, 1, S) have K > S, and are thus not Steiner ETFs since the underlying
BIBD would necessarily violate Fisher's inequality. To elaborate, when K is a prime power, the
requirement that K divides S(S − 1) where S and S − 1 are relatively prime implies that either
K divides S or S − 1, and in either case K ≤ S. However, when K is not a prime power, we can
sometimes choose it to be a divisor of S(S − 1) that is larger than S.
14
For example, taking K = S(S − 1) in (30) gives D = S2 − 1 and N = (S2 − 1)2 = D2. Such an
ETF thus corresponds to a SIC-POVM in a space whose dimension is one less than a perfect square.
Such SIC-POVMs are known to exist when S = 2, . . . , 7, 18, and are conjectured to exist for all
S [26, 31]. Similarly, taking K =(cid:0)S
2(cid:1) in (30) gives D = S2 −2 and N = 1
We refer to such (D, N ) as being of real maximal type since real-valued examples of such ETFs meet
the real-variable version of the Gerzon bound, and are known to exist when S = 3, 5. Remarkably,
it is known that a real ETF of this type does not exist when S = 7 [23]. We also caution that there
2 (S2 − 2)(S2 − 1) =(cid:0)D+1
2 (cid:1).
is a single pair (D, N ) with N =(cid:0)D+1
ETF(D, N ) exists, namely (D, N ) = (3, 6).
2 (cid:1) that is neither positive or negative despite the fact that an
We summarize these facts as follows:
Theorem 4.1. An ETF of type (K, 1, S) exists whenever:
(a) K = 1 and S ≥ 2, (regular simplices);
(b) K ≥ 2 and there exists a BIBD(S(K − 1) + 1, K, 1) (Steiner ETFs [25]), including:
Q−1 where Q is a prime power and J ≥ 2 (affine geometries);
(i) K = 2, 3, 4, 5 and S ≥ K has the property that K divides S(S − 1);
(ii) K = Q and S = QJ −1
(iii) K = Q + 1 and S = QJ −1
(iv) K = Q + 1 and S = Q2 where Q is a prime power (unitals);
(v) K = 2J1 and S = 2J2 + 1 where 2 ≤ J1 < J2 (Denniston designs);
(vi) K ≥ 2 and S is sufficiently large and has the property that K divides S(S − 1);
Q−1 where Q is a prime power and J ≥ 2 (projective geometries);
(c) K = Q and S = Q whenever Q is a prime power [22];
(d) K = S(S − 1) where S = 2, . . . , 7, 18 (SIC-POVMs [26, 31]);
(e) K =(cid:0)S
2(cid:1) where S = 3, 5 (real maximal type [23]).
Because so much is already known regarding the existence of positive ETFs, we could not find
any examples where Theorem 3.2 makes a verifiable contribution. As we now discuss, much less is
known about negative ETFs, and this gives Theorem 3.2 an opportunity to be useful.
4.2. Negative equiangular tight frames
By Definition 1.1 and Theorem 3.1, an ETF(D, N ) with 1 < D < N is negative if and only if
there exists integers K ≥ 1 and S ≥ 2 such that
D = S
K [S(K − 1) − 1] = S2 − S(S+1)
K , N = (S − 1)[S(K − 1) − 1],
(31)
2 and K = N S
D(S−1) are integers. Here, since N
D > 1 and S
S−1 > 1,
2 S(S − 1), (S − 1)2) and such ETFs exist for any S ≥ 3,
or equivalently that S =(cid:2) D(N −1)
N −D (cid:3)
we actually necessarily have that K ≥ 2.
When K = 2, (31) becomes (D, N ) = ( 1
1
being the Naimark complements of ETFs of type (2, 1, S − 1).
In the K = 3 case, (31) becomes
D = S(2S−1)
3
= S2 − S(S+1)
3
, N = (S − 1)(2S − 1).
15
For D to be an integer, we necessarily have S ≡ 0, 2 mod 3. Moreover, for any S ≥ 2 with
S ≡ 0, 2 mod 3, an ETF of type (3, −1, S) exists. Indeed, the recent paper [19] gives a way to
modify the Steiner ETF arising from a BIBD(V, 3, 1) to yield a Tremain ETF(D, N ) with
D = 1
6 (V + 2)(V + 3), N = 1
2 (V + 1)(V + 2),
for any V ≥ 3 with V ≡ 1, 3 mod 6. Such an ETF is type (3, −1, S) with S = 1
2 (V + 3). We
also note that for every J ≥ 1, there is a harmonic ETF(D, N ) arising from a Spence difference
set [47] with (D, N ) = ( 1
2 3J+1(3J+1 − 1)), and such an ETF is type (3, −1, S) where
S = 1
2 (3J+1 + 1). It remains unclear whether any Spence ETFs are unitarily equivalent to special
instances of Tremain ETFs.
2 3J (3J+1 + 1), 1
In the K = 4 case, for any positive integer J, Davis and Jedwab [17] give a difference set whose
harmonic ETF has parameters (3) and so is a type (4, −1, S) ETF where S = 1
3 (22J+1 +1). Beyond
these examples, a few other infinite families of negative ETFs are known to exist. In particular, in
order for the expression for D given in (31) to be an integer, K necessarily divides S(S + 1), and
so it is natural to consider the special cases where S = K and S = K − 1.
When S = K − 1, (31) gives that ETFs of type (K, −1, K − 1) have
D = (K − 1)(K − 2), N = K(K − 2)2.
Remarkably, these are the same (D, N ) parameters as those of an ETF of type (K − 2, 1, K − 1).
In particular, every Steiner ETF arising from an affine plane of order Q is both of (positive) type
(Q, 1, Q + 1) and (negative) type (Q + 2, −1, Q + 1). More generally, a Steiner ETF arising from a
BIBD(V, K, 1) has (D, N ) = (B, V (R + 1)) where R = V −1
K , and so is only negative
when
K−1 and B = V R
D(S−1) = V (R+1)R
B(R−1) = K(R+1)
R−1
N S
(32)
is an integer. When R is even, R − 1 and R + 1 are relatively prime, and this can only occur when
R − 1 divides K. Here, since Fisher's inequality gives K ≤ R, this happens precisely when either
R = K = 2 or R = K + 1, namely when the underlying BIBD is a BIBD(3, 2, 1) or is an affine
plane of odd order K, respectively. Meanwhile, when R is odd, R − 1 and R + 1 have exactly
one prime factor in common, namely 2, and (32) is an integer precisely when 1
2 (R − 1) divides K.
Since K ≤ R, this happens precisely when either R = K = 3, R = K + 1 or R = 2K + 1, namely
when the underlying BIBD is the projective plane of order 2, an affine plane of even order, or is a
BIBD(V, K, 1) where V = (2K + 1)(K − 1) + 1 = K(2K − 1) for some K ≥ 2, respectively. With
regard to the latter, it seems to be an open question whether a BIBD(K(2K − 1), K, 1) exists for
every K ≥ 2, though a Denniston design provides one whenever K = 2J for some J ≥ 1, and they
are also known to exist when K = 3, 5, 6, 7 [39]. For such ETFs, (32) becomes K + 1, meaning
they are type (K ′, −1, 2K ′ − 1) where K ′ = K + 1.
Meanwhile, when S = K, (31) gives that ETFs of type (K, −1, K) have
D = K 2 − K − 1, N = (K − 1)(K 2 − K − 1).
(33)
The recently-discovered hyperoval ETFs of [24] are instances of such ETFs whenever K = 2J + 1
for some J ≥ 1. In the K = 4 case, (33) becomes (D, N ) = (11, 33), and this seems to be the
smallest set of positive or negative parameters for which the existence of a corresponding ETF
remains an open problem.
16
Apart from these examples, it seems that only a finite number of other negative ETFs are known
to exist. For example, since K necessarily divides S(S + 1), it is natural to also consider the cases
where K = S(S+1) and K =(cid:0)S+1
In particular, every positive SIC-POVM is also negative. Similarly, when K = (cid:0)S+1
N =(cid:0)D+1
2 (cid:1). When K = S(S+1), (31) becomes (D, N ) = (S2−1, (S2−1)2).
2 (cid:1), (31) gives
2 (cid:1) where D = S2 − 2, meaning every positive (D, N ) of real maximal type is also negative.
The only other example of a negative ETF that we found in the literature was an ETF(22, 176),
which has type (10, −1, 5), and arises from a particular SRG. When searching tables of known ETFs
such as [23], it is helpful to note that most positive and negative ETFs have redundancy N
D > 2,
with the only exceptions being 1-positive ETFs (regular simplices), 2-negative ETFs (Naimark
complements of 2-positive ETFs), ETF(2, 3) when regarded as type (3, −1, 2), and ETF(5, 10),
which are type (3, −1, 3). Indeed, when K ≥ 2, any K-positive ETF has N
> K ≥ 2.
Meanwhile, when K ≥ 3, any ETF of type (K, −1, S) only has K(S−1)
= N
K−2 .
Since K necessarily divides S(S + 1) where S ≥ 2, such an ETF can only exist when K = 3 and
S = 2, 3. We summarize these previously-known constructions of negative ETFs as follows:
D = K(S+1)
D ≤ 2 when S ≤ K
S
S
Theorem 4.2. An ETF of type (K, −1, S) exists whenever:
(a) K = 2 and S ≥ 3 (Naimark complements of 2-positive ETFs);
(b) K = 3 and S ≥ 2 with S ≡ 0, 2 mod 3 (Tremain ETFs [19] and Spence harmonic ETFs [47]);
(c) K = 4 and S = 1
3 (22J+1 + 1) for some J ≥ 1 (Davis-Jedwab harmonic ETFs [17]);
(d) K = Q + 2 and S = K − 1 where Q is a prime power (Steiner ETFs from affine planes [25]);
(e) K = 2J + 1 and S = 2K − 1 where J ≥ 1 (Steiner ETFs from Denniston designs [25]);
(f) K = 2J + 1 and S = K where J ≥ 1 (hyperoval ETFs [24]);
(g) K = S(S + 1) where S = 2, . . . , 7, 18 (SIC-POVMs [26]);
(h) K =(cid:0)S+1
2 (cid:1) where S = 3, 5 (real maximal type [23]);
(i) (K, S) = (4, 7), (6, 11), (7, 13), (8, 15), (10, 5) (various other ETFs [23]).
From this list, we see that for any K ≥ 5, the existing literature provides at most a finite
number of K-negative ETFs. Theorem 3.2(a) implies that many more negative ETFs exist: if an
ETF of type (K, −1, S) exists, then an ETF of type (K, −1, M U +1
K−1 ) exists for all sufficiently large
U that satisfy (19) and (20). Combining this fact with Theorem 4.2 immediately gives:
Corollary 4.3. There exists an infinite number of K-negative ETFs whenever:
(a) K = Q + 2 where Q is a prime power;
(b) K = Q + 1 where Q is an even prime power;
(c) K = 2, 8, 12, 20, 30, 42, 56, 342.
In particular, we now know that there are an infinite number of values of K for which an
infinite number of K-negative ETFs exist, with K = 14 being the smallest open case. With these
asymptotic existence results in hand, we now focus on applying Theorem 3.2 with explicit GDDs.
For example, the "Mercedes-Benz" regular simplex ETF(2, 3) is type (K, L, S) = (3, −1, 2) and
so has M = S(K − 1) + L = 3. Since S + L = 1 divides K − 2 = 1, Theorem 3.2 can be applied
with any 3-GDD of type 3U so as to produce an ETF of type (K, L, M U −L
2 (3U + 1)),
and moreover that the necessary conditions (7) on the existence of such a GDD reduce to (19),
K−1 ) = (3, −1, 1
17
2 (U − 1) ∈ Z and 1
namely to having U ≥ 3, 1
2 U (U − 1) ∈ Z. In fact, such GDDs are known to
exist whenever these necessary conditions are satisfied [27], namely when U ≥ 3 is odd. (This
also follows from the fact that such GDDs are equivalent to the incidence structures obtained by
removing a parallel class from a resolvable Steiner triple system.) That is, writing U = 2J + 1 for
some J ≥ 1, we can apply Theorem 3.2 with an ETF(2, 3) and a known 3-GDD of type 32J+1 to
produce an ETF of type (3, −1, 1
2 (3U + 1)) = (3, −1, 3J + 2).
In summary, applying Theorem 3.2 to an ETF of type (3, −1, 2) produces ETFs of type
(3, −1, S) for any S ≡ 2 mod 3, and so recovers the parameters of "half" of all possible 3-negative
ETFs, cf. Theorem 4.2 and [19], including the parameters of all harmonic ETFs arising from Spence
difference sets. To instead recover some of the ETFs of type (3, −1, S) with S ≡ 0 mod 3, one may,
for example, apply Theorem 3.2 to the well-known ETF(5, 10), which is type (3, −1, 3).
In order to obtain ETFs with verifiably new parameters, we turn our attention to applying
Theorem 3.2 to known K-negative ETFs with K ≥ 4. Here, the limiting factor seems to be a lack of
knowledge regarding uniform K-GDDs: while the literature has much to say when K = 3, 4, 5 [27],
we are relegated to well-known simple constructions involving Lemmas 2.1 and 2.2 whenever K > 5.
As such, we consider the K = 4, 5 cases separately from those with K > 5:
Proof of Theorem 1.2. The ETF(6, 16) is type (4, −1, 3), and so we can apply Theorem 3.2 when-
ever there exists a 4-GDD of type 8U where U satisfies (20). By Theorem 3.2, the known necessary
conditions (7) on the existence of such GDDs reduce to (19):
U ≥ 4,
U −1
3 ∈ Z,
U (U −1)
3
= 4U (U −1)
4(3) ∈ Z,
3 = 2(U −1)
namely to having U ≥ 4 and U ≡ 1 mod 3. These necessary conditions on the existence of 4-GDDs
of type 8U are known to be sufficient [27]. Moreover, for any such U , we have (20) is automatically
satisfied since U −1
2(3) ∈ Z. Altogether, for any U ≥ 4 with U ≡ 1 mod 3, we can apply
Theorem 3.2 to the ETF of type (4, −1, 3) with a 4-GDD of type 8U , and doing so produces an
ETF of type (4, −1, 1
3 (8U + 1)). Here, letting U = 1 recovers the parameters of the original ETF.
Overall, writing U = 3J + 1 for some J ≥ 0, this means that an ETF of type (4, −1, S) exists
whenever S = 1
3 [8(3J + 1) + 1] = 8J + 3 for any J ≥ 0, namely whenever S ≡ 3 mod 8.
In particular, for any J ≥ 1 we can take U = 4J−1 to obtain an ETF of type (4, −1, S) where
S = 1
3 (22J−1 +1). This means that applying Theorem 3.2 to the ETF of type (4, −1, 3)
recovers the parameters (3) of harmonic ETFs corresponding to Davis-Jedwab difference sets.
3 (8U + 1) = 1
3 (8U +1) = 1
In light of Remark 3.3, applying Theorem 3.2 to any ETF of type (4, −1, S) with S ≡ 3 mod 8
simply recovers a subset of the ETF types obtained by applying Theorem 3.2 to the ETF of type
(4, −1, 3). As such, to obtain more 4-negative ETFs via Theorem 3.2, we need to apply it to initial
ETFs that lie outside of this family. By Theorem 4.2, the existing literature gives one such set of
parameters, namely ETFs of type (4, −1, 7), which have (D, N ) = (35, 120). Here, M = 20, and a
4-GDD of type 20U can only exist if U satisfies (19):
U ≥ 4,
U −1
3 ∈ Z,
U (U −1)
3
= 4U (U −1)
4(3) ∈ Z.
Moreover, these necessary conditions are known to be sufficient [27], meaning Theorem 3.2 can
be applied whenever U also satisfies (20), namely U −1
6(3) ∈ Z. Thus, for any U ≡ 1 mod 9,
Theorem 3.2 yields an ETF of type (4, −1, 1
3 (20U + 1)). In particular, an ETF of type (4, −1, S)
exists for any S ≡ 7 mod 60. ETFs of type (4, −1, S) with S ≡ 67 mod 120 thus arise from both
constructions; in the statement of Theorem 1.2, we elect to not remove the overlapping values of
9 = 2(U −1)
18
S from either family so as to not emphasize one family over the other, and possibly make it easier
for future researchers to identify potential patterns.
In a similar manner, as summarized in Theorem 4.2, the existing literature provides ETFs of
type (5, −1, S) for exactly three values of S, namely ETF(12, 45), ETF(19, 76) and ETF(63, 280)
which have S = 4, 5, 9 and so M = 15, 19, 35, respectively. For these particular values of M , the
corresponding necessary conditions (7) on the existence of 5-GDDs of type M U are known to be
sufficient [27], meaning we only need U to satisfy (19) and (20) for the corresponding value of S,
namely to satisfy U ≥ 5, 1
4 (U − 1) ∈ Z and that
5U (U −1)
5(4) ∈ Z,
6U (U −1)
5(4) ∈ Z,
10U (U −1)
5(4)
∈ Z,
3(U −1)
3(4) ∈ Z, when S = 4,
3(U −1)
4(4) ∈ Z, when S = 5,
3(U −1)
8(4) ∈ Z, when S = 9.
An ETF of type (5, −1, S) thus exists when S = 1
with U ≡ 1, 65 mod 80, or S = 1
exists when S ≡ 4 mod 15, or S ≡ 5, 309 mod 380, or S ≡ 9 mod 280.
4 (19U + 1)
4 (35U + 1) with U ≡ 1 mod 32. That is, an ETF of type (5, −1, S)
4 (15U + 1) with U ≡ 1 mod 4, or S = 1
In certain special cases, these techniques yield real ETFs:
Proof of Theorem 1.3. The ETF (21) constructed in Theorem 3.2 is clearly real when the initial
ETF(D, N ) {ϕn}N
n=1 and the Hadamard matrices of size S + L and W + 1 are real.
In particular, a real ETF(63, 280) exists [23], and such ETFs are type (5, −1, 9). Since a real
Hadamard matrix of size S + L = 8 exists, then letting M = S(K − 1) + L = 35, Theorem 3.2
yields a real ETF of type (5, −1, 1
4 (35U + 1)) whenever there exists a 5-GDD of type 35U where U
satisfies (20) and there exists a real Hadamard matrix of size
H = W + 1 = R
S+L + 1 = M (U −1)
(S+L)(K−1) + 1 = 35(U −1)
8(4) + 1 = 35U −3
.
32
Here, we recall from the proof of Theorem 1.2 that such GDDs exist whenever U ≡ 1 mod 32,
namely whenever H ≡ 1 mod 35. Altogether, since 1
) + 1] = 8H + 1,
Theorem 3.2 yields a real ETF of type (5, −1, 8H + 1) whenever there exists a Hadamard matrix
of size H when H ≡ 1 mod 35. An infinite number of such Hadamard matrices exist: since 17 is
relatively prime to 140, Dirichlet's theorem implies an infinite number of primes Q ≡ 17 mod 140
exist, and each has the property that Q ≡ 1 mod 4, meaning that Paley's construction yields a real
Hadamard matrix of size 2(Q + 1) ≡ 36 mod 280, in particular of size 2(Q + 1) ≡ 1 mod 35.
4 (35U + 1) = 1
4 [35( 32H+3
35
Similarly, a real ETF(7, 28) exists [23], is type (6, −1, 3), and there exists a real Hadamard
matrix of size S + L = 2. Letting M = 14, Theorem 3.2 thus yields a real ETF of type
(6, −1, 1
5 (14U + 1)) whenever there exists a 6-GDD of type 14U where U satisfies (20) and there
exists a real Hadamard matrix of size H = W + 1 = 1
5 (7U − 2). Here, the necessary condi-
tions (19) and (20) on the existence of such a GDD reduce to having U ≡ 1, 6 mod 15, namely to
having H ≡ 1, 8 mod 21. Since K = 6, these necessary conditions are not known to be sufficient.
Nevertheless, they are asymptotically sufficient: there exists U0 such that for all U ≥ U0 with
U ≡ 1, 6 mod 15, there exists a 6-GDD of type 14U where U satisfies (20). As such, there exists
H0 such that for all H ≥ H0 with H ≡ 1, 8 mod 21, if there exists a real Hadamard matrix of
size H, then there exists a real ETF of type (6, −1, 1
5 (14U + 1)) = (6, −1, 2H + 1). As above,
an infinite number of such Hadamard matrices exist: since gcd(73, 84) = 1, there are an infinite
19
number of primes Q ≡ 73 mod 84, and each has the property that Q ≡ 1 mod 4, meaning Paley's
construction yields a real Hadamard matrix of size 2(Q + 1) ≡ 148 mod 168, in particular of size
2(Q + 1) ≡ 1 mod 21.
Applying these same techniques to real ETF(22, 176) and ETF(23, 276) yields the infinite fam-
ilies stated in (c) and (d) of the result.
We have now seen that the construction of Theorem 3.2 recovers all Steiner ETFs as a degen-
erate case, recovers the parameters of "half" of all Tremain ETFs including those of all harmonic
ETFs arising from Spence difference sets, and also recovers the parameters (3) of harmonic ETFs
arising from the Davis-Jedwab difference sets. This is analogous to how the approach of [17] unifies
McFarland, Spence and Davis-Jedwab difference sets with those with parameters (3). From this
perspective, the value of the generalization of [17] given in Theorem 3.2 is that it permits weaker
conclusions to be drawn from weaker assumptions: while [17] forms new difference sets (i.e., new
harmonic ETFs) by combining given difference sets with building sets formed from a collection of
hyperplanes, Theorem 3.2 forms new ETFs by combining given ETFs with GDDs.
In fact, a careful read of [17] indicates that the building sets used there to produce difference sets
with parameters (3) are related to 4-GDDs of type 84J −1
obtained by recursively using TD(4, 22j+1)
to fill the holes of TD(4, 22j+3) for every j = 1, . . . , J − 1 via Lemma 2.2. Alternatively, such GDDs
can be constructed by using Lemma 2.1 to combine a TD(4, 8) with a BIBD(4J−1, 4, 1) arising from
an affine geometry. We now generalize these approaches, using Lemmas 2.1 and 2.2 to produce the
GDDs needed to apply Theorem 3.2 to several known K-negative ETFs with K > 5:
Theorem 4.4. If an ETF of type (K, −1, S) exists where K−2
M = S(K − 1) − 1, then an ETF of type (K, −1, M U +1
K−1 ) exists when either:
S−1 ∈ Z, and a TD(K, M ) exists where
(i) U = 1 or U = K;
(ii) a BIBD(U, K, 1) exists;
(iii) U = K J for some J ≥ 2, provided a TD(K, M K j ) exists for all j = 1, . . . , J − 1.
As a consequence, an ETF of type (K, −1, S) exists when either:
(a) K = 6, S = 1
(b) K = 6, S = 1
(c) K = 7, S = 1
(d) K = 10, S = 1
(e) K = 12, S = 1
5 (9U + 1) where either U = 6J for some J ≥ 0 or a BIBD(U, 6, 1) exists;
5 (24U + 1) where either U = 6J for some J ≥ 0 or a BIBD(U, 6, 1) exists;
6 (35U + 1) where either U = 7J for some J ≥ 0 or a BIBD(U, 7, 1) exists;
9 (80U + 1) where either U = 10J for some J ≥ 0 or a BIBD(U, 10, 1) exists;
11 (32U + 1) when either U = 12J for some J ≥ 0 or a BIBD(U, 12, 1) exists.
Proof. Since K−2
S−1 ∈ Z and M = S(K − 1) − 1, any K-GDD of type M U can be combined with the
given ETF of type (K, −1, S) via Theorem 3.2 in order to construct an ETF of type (K, −1, M U +1
K−1 ).
(Since M +1
K−1 = S, such an ETF also exists when U = 1, namely the given initial ETF.) For
instance, the given TD(K, M ) is a K-GDD of type M K , and so such an ETF exists when U = K.
Other examples of such GDDs can be constructed by combining the given TD(K, M ) with any
BIBD(U, K, 1)-a K-GDD of type 1U -via Lemma 2.1. In particular, when K is a prime power,
we can construct these BIBDs from affine geometries of order K to produce examples of such
GDDs with U = K J for any J ≥ 2. GDDs with these parameters also sometimes exist even when
K is not a prime power: if a TD(K, M K j) exists for all j = 1, . . . , J − 1, then recursively using
20
TD(K, M K j−1) to fill the holes of TD(K, M K j ) for all j = 1, . . . , J − 1 via Lemma 2.1 gives a
K-GDD of type M K J
.
We now apply these ideas to some known K-negative ETFs with K > 5, organized according
to the families of Theorem 4.2.
In particular, ETFs of type (Q + 2, −1, Q + 1) exist whenever
Q is a prime power, and S − 1 = Q divides K − 2 = Q. For such ETFs, M = Q(Q + 2) and
a TD(Q + 2, Q(Q + 2)) equates to Q MOLS of size Q(Q + 2), which is known to occur when
Q = 2, 3, 4, 5, 8 [1]. (When Q and Q + 2 are both prime powers, the standard method [40] only
produces Q − 1 MOLS of size Q(Q + 2).) Taking Q = 2, 3 recovers a subset of the ETFs produced
in Theorem 1.2, and so we focus on Q = 4, 5, 8. In particular, for these values of Q, the above
1
Q+1 [Q(Q + 2)U + 1]) when either U = 1, U = Q + 2, or
methods yield an ETF of type (Q + 2, −1,
a BIBD(U, Q + 2, 1) exists. Moreover, even when Q + 2 is not a prime power, such an ETF with
U = (Q+2)J exists, provided a TD(Q+2, Q(Q+2)j+1) exists for all j = 1, . . . , J −1. When Q = 4, 8,
this occurs for any J ≥ 2, since for any j ≥ 1, the number of MOLS of size (4)6j+1 = 2j+33j+1 is
at least min{2j+3, 3j+1} − 1 ≥ 7, while the number of MOLS of size (8)10j+1 = 2j+45j+1 is at least
min{2j+4, 5j+1} − 1 ≥ 23.
Other new explicit infinite families of negative ETFs arise from SIC-POVMs, which by Theo-
rem 4.2 are ETFs of type (K, −1, S) where K = S(S + 1). Such ETFs always have the property
that S − 1 divides K − 2 = (S − 1)(S + 2), meaning we can apply the above ideas whenever there
exists a TD(K, M ) where M = (S − 1)(S + 1)2. Such TDs exist when S = 2, 3, being TD(6, 9)
and TD(12, 32), respectively. When S = 2 in particular, the above methods yield an ETF of type
(6, −1, 1
5 (9U + 1)) when either U = 1, U = 6, a BIBD(U, 6, 1) exists, or U = 6J for some J ≥ 2.
This final family arises from the fact that a TD(6, (9)6j ) exists for all j ≥ 1: there are at least 5
MOLS of size 54 = (9)6, at least 8 MOLS of size 324 = 9(62) (since there are 8 MOLS of size 9, and
at least 8 MOLS of size 36) [1], while for any j ≥ 3, the number of MOLS of size (9)6j = 2j3j+2
is at least 2j − 1 ≥ 7. Similarly, taking S = 3 yields an ETF of type (12, −1, 1
11 (32U + 1)) when
either U = 1, U = 12, a BIBD(U, 12, 1) exists, or U = 12J for some J ≥ 2: the number of MOLS
of size 32(12) = 384 and 32(12)2 = 4608 is at least 15 [1], while for j ≥ 3, the number of MOLS of
size 32(12)j = 22j+53j is at least min{22j+5, 3j } − 1 ≥ 26.
To be clear, we view the ETFs produced in Theorem 4.4 as a "proof of concept," and believe
that GDD experts will be able to find many more examples of new ETFs using Theorems 3.2
and 4.2. For this reason, we have omitted technical cases where an ETF of type (K, −1, S) and
a TD(K, M ) exist where M = S(K − 1) − 1, but S − 1 does not divide K − 2, such as when
S = K = 9, and when S = 2K − 1 where K = 6, 7, 8, 9, 17, 65537.
In such cases, one can still
combine the TD with a BIBD(U, K, 1) via Lemma 2.1 to produce a K-GDD of type M U , but (20)
is not automatically satisfied.
We also point out that though Theorem 3.2 is a generalization of [17], the new ETFs we have
found here are disjoint from those produced by another known generalization of this same work.
In particular, the parameters (3) of Davis-Jedwab difference sets can be regarded as the Q = 2
case of the more general family:
D = Q2J−1( 2Q2J +Q−1
Q+1
), N = 4Q2J ( Q2J −1
Q2−1 ),
(34)
where J ≥ 1 and Q ≥ 2 are integers. In particular, in [14], Chen generalizes the theory of [17] in
a way that produces difference sets with parameters (34) for any J ≥ 1 and any Q that is either
a power of 3 or any even power of an odd prime. For this reason, difference sets with parameters
21
of the form (34) are said to be Davis-Jedwab-Chen difference sets [35]. The inverse Welch bounds
for the corresponding harmonic ETFs and their Naimark complements are always integers:
However, such ETFs are seldom positive or negative. Indeed, a direct computation reveals
(cid:2) (N −D)(N −1)
D
(cid:3)
1
2 = 2Q2J −Q−1
.
Q−1
1
,
Q+1
2 = 2Q2J +Q−1
S =(cid:2) D(N −1)
N −D (cid:3)
D(S+1) = 2 + 2Q(Q2J −2−1)
N S
(Q−1)(Q2J −1+1)
,
N S
D(S−1) = 2 + 2
Q−1 .
2 (9J + 1)) or (4, −1, 1
Thus, by Theorem 3.1, the only such ETFs that are positive or negative are of type (2, 1, 2Q − 1),
(3, −1, 1
3 (22J+1 + 1)), corresponding to the special cases of (34) where J = 1,
Q = 3 and Q = 2, respectively. That is, the only overlap between the K-negative ETFs constructed
in Theorems 1.2 and 4.4 and the ETFs constructed in [14] are those with parameters (34).
5. Conclusions
Theorems 1.2 and 4.4 make some incremental progress towards resolving the ETF existence
problem. By comparing the existing ETF literature, as summarized in [23] for example, against
Theorems 4.1 and 4.2, we find that every known ETF is either an orthonormal basis, or has the
property that either it or its Naimark complement is a regular simplex, has N = 2D or N = 2D±1,
is a SIC-POVM, arises from a difference set, or is either positive or negative. In particular, every
known ETF(D, N ) with N > 2D > 2 is either a SIC-POVM (N = D2), or is a harmonic ETF, or
has N = 2D + 1 where D is odd, or is either positive or negative. This fact, along with the known
nonexistence of the ETF(3, 8) [51], and the available numerical evidence [53], leads us to make the
following conjecture:
Conjecture 5.1. If N > 2D > 2, an ETF(D, N ) exists if and only if N = D2 or D(D−1)
(D, N ) is positive or negative (Definition 1.1).
N −1 ∈ Z or
This is a substantial strengthening of an earlier conjecture made by one of the authors, namely
that if N > D > 1 and an ETF(D, N ) exists, then one of the three numbers D, N − D and
N − 1 necessarily divides the product of the other two. To help resolve Conjecture 5.1, it would in
particular be good to know whether an ETF(9, 25) exists: this is the smallest value of D for which
there exists an N such that N − 1 divides D(D − 1) but no known ETF exists. For context, we
note that there are numerous pairs (D, N ) for which N − 1 divides D(D − 1) and an ETF(D, N )
is known to exist, despite the fact that a difference set of that size does not exist [30], including
(35, 120), (40, 105), (45, 100), (63, 280), (70, 231), (77, 210), (91, 196), (99, 540), (130, 560),
(143, 924), (176, 561), (187, 528), (208, 1105), (231, 484), (247, 780), (260, 741).
It would also be good to know whether an ETF(11, 33) exists, since such an ETF would be
type (4, −1, 4): as detailed in the previous section, an ETF of type (K, L, S) can only exist when
K divides S(S − L), and moreover this necessary condition for existence is known to be sufficient
when L = 1 and K = 1, 2, 3, 4, 5, as well as when L = −1 and K = 2, 3.
The evidence also supports an analogous conjecture in the real case. In fact, comparing the
relevant literature [12, 23] against Theorems 4.1 and 4.2, we find that every known real ETF(D, N )
with N > 2D > 2 is either positive or negative. Moreover, when 1 < D < N − 1, N 6= 2D and
22
a real ETF(D, N ) exists, then it and its Naimark complements' Welch bounds are necessarily
the reciprocals of odd integers [50].
In particular, any such ETF automatically satisfies one of
the two integrality conditions given in Theorem 3.1 that characterize positive and negative ETFs.
Conversely, since S(K − 1) + KL is odd whenever S is odd, Theorem 3.1 implies that any ETF of
type (K, L, S) satisfies the necessary conditions of [50] when S is odd. That said, real ETF(19, 76),
ETF(20, 96) and ETF(47, 1128) do not exist [3, 4, 23], despite the fact that (19, 76) is type (5, −1, 5),
(20, 96) is both type (4, 1, 5) and type (6, −1, 5), and (47, 1128) is both type (21, 1, 7) and type
(28, −1, 7). These facts suggest the following analog of Conjecture 5.1:
Conjecture 5.2. If N > 2D > 2 and a real ETF(D, N ) exists, then (D, N ) is positive or negative.
Acknowledgments
The views expressed in this article are those of the authors and do not reflect the official
policy or position of the United States Air Force, Department of Defense, or the United States
Government. This work was partially supported by the Summer Faculty Fellowship Program of
the United States Air Force Research Laboratory.
References
[1] R. J. R. Abel, C. J. Colbourn, J. H. Dinitz, Mutually Orthogonal Latin Squares (MOLS), in: C.J. Colbourn,
J.H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 160–193.
[2] R. J. R. Abel, M. Greig, BIBDs with small block size, in: C.J. Colbourn, J.H. Dinitz (Eds.), Handbook of
Combinatorial Designs, Second Edition (2007) 72–79.
[3] J. Azarija, T. Marc, There is no (75,32,10,16) strongly regular graph, arXiv:1509.05933.
[4] J. Azarija, T. Marc, There is no (95,40,12,20) strongly regular graph, arXiv:1603.02032.
[5] W. U. Bajwa, R. Calderbank, D. G. Mixon, Two are better than one:
fundamental parameters of frame
coherence, Appl. Comput. Harmon. Anal. 33 (2012) 58-78.
[6] A. S. Bandeira, M. Fickus, D. G. Mixon, P. Wong, The road to deterministic matrices with the Restricted
Isometry Property, J. Fourier Anal. Appl. 19 (2013) 1123–1149.
[7] A. Barg, A. Glazyrin, K. A. Okoudjou, W.-H. Yu, Finite two-distance tight frames, Linear Algebra Appl. 475
(2015) 163–175.
[8] B. G. Bodmann, H. J. Elwood, Complex equiangular Parseval frames and Seidel matrices containing pth roots
of unity, Proc. Amer. Math. Soc. 138 (2010) 4387–4404.
[9] B. G. Bodmann, V. I. Paulsen, M. Tomforde, Equiangular tight frames from complex Seidel matrices con-
taining cube roots of unity, Linear Algebra Appl. 430 (2009) 396–417.
[10] C. Bracken, G. McGuire, H. Ward, New quasi-symmetric designs constructed using mutually orthogonal Latin
squares and Hadamard matrices, Des. Codes Cryptogr. 41 (2006) 195–198.
[11] A. E. Brouwer, Strongly regular graphs, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial
Designs, Second Edition (2007) 852-868.
[12] A. E. Brouwer, Parameters of Strongly Regular Graphs, http://www.win.tue.nl/∼aeb/graphs/srg/
[13] K. I. Chang, An existence theory for group divisible designs, Ph.D. Thesis, The Ohio State University, 1976.
[14] Y. Q. Chen, On the existence of abelian Hadamard difference sets and a new family of difference sets, Finite
Fields Appl. 3 (1997) 234–256.
[15] D. Corneil, R. Mathon, eds., Geometry and combinatorics: Selected works of J. J. Seidel, Academic Press,
1991.
[16] G. Coutinho, C. Godsil, H. Shirazi, H. Zhan, Equiangular lines and covers of the complete graph, Linear
Algebra Appl. 488 (2016) 264–283.
[17] J. A. Davis, J. Jedwab, A unifying construction for difference sets, J. Combin. Theory Ser. A 80 (1997) 13–78.
[18] C. Ding, T. Feng, A generic construction of complex codebooks meeting the Welch bound, IEEE Trans.
Inform. Theory 53 (2007) 4245–4250.
[19] M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, Tremain equiangular tight frames, J. Combin. Theory
Ser. A 153 (2018) 54-66.
23
[20] M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, Hadamard equiangular tight frames, submitted,
arXiv:1703.05353.
[21] M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, C. E. Watson, Equiangular tight frames with centroidal
symmetry, to appear in Appl. Comput. Harmon. Anal.
[22] M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, C. E. Watson, Polyphase equiangular tight frames and
abelian generalized quadrangles, to appear in Appl. Comput. Harmon. Anal.
[23] M. Fickus, D. G. Mixon, Tables of the existence of equiangular tight frames, arXiv:1504.00253 (2016).
[24] M. Fickus, D. G. Mixon, J. Jasper, Equiangular tight frames from hyperovals, IEEE Trans. Inform. Theory
62 (2016) 5225–5236.
[25] M. Fickus, D. G. Mixon, J. C. Tremain, Steiner equiangular tight frames, Linear Algebra Appl. 436 (2012)
1014–1027.
[26] C. A. Fuchs, M. C. Hoang, B. C. Stacey, The SIC question: history and state of play, Axioms 6 (2017) 21.
[27] G. Ge, Group divisible designs, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial Designs,
Second Edition (2007) 255–260.
[28] C. D. Godsil, Krein covers of complete graphs, Australas. J. Combin. 6 (1992) 245–255.
[29] J. M. Goethals, J. J. Seidel, Strongly regular graphs derived from combinatorial designs, Can. J. Math. 22
(1970) 597–614.
[30] D. Gordon, La Jolla Covering Repository, https://www.ccrwest.org/diffsets.html.
[31] M. Grassl, A. J. Scott, Fibonacci-Lucas SIC-POVMs, J. Math. Phys. 58 (2017) 122201.
[32] R. B. Holmes, V. I. Paulsen, Optimal frames for erasures, Linear Algebra Appl. 377 (2004) 31–51.
[33] J. W. Iverson, J. Jasper, D. G. Mixon, Optimal
line packings from nonabelian groups, submitted,
arXiv:1609.09836.
[34] J. Jasper, D. G. Mixon, M. Fickus, Kirkman equiangular tight frames and codes, IEEE Trans. Inform. Theory
60 (2014) 170-181.
[35] D. Jungnickel, A. Pott, K. W. Smith, Difference sets, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of
Combinatorial Designs, Second Edition (2007) 419–435.
[36] J. H. van Lint, J. J. Seidel, Equilateral point sets in elliptic geometry, Indag. Math. 28 (1966) 335–348.
[37] E. R. Lamken, R. M. Wilson, Decompositions of edge-colored complete graphs, J. Combin. Theory Ser. A 89
(2000) 149–200.
[38] P. W. H. Lemmens, J. J. Seidel, Equiangular lines, J. Algebra 24 (1973) 494–512.
[39] R. Mathon, A. Rosa, 2 − (v, k, λ) designs of small order, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook
of Combinatorial Designs, Second Edition (2007) 25–58.
[40] H. F. MacNeish, Euler Squares, Ann. of Math. 23 (1922) 221–227.
[41] R. L. McFarland, A family of difference sets in non-cyclic groups, J. Combin. Theory Ser. A 15 (1973) 1–10.
[42] G. McGuire, Quasi-symmetric designs and codes meeting the Grey-Rankin bound, J. Combin. Theory Ser. A
78 (1997) 280–291.
[43] Hedvig Moh´acsy, The asymptotic existence of group divisible designs of large order with index one, J. Combin.
Theory Ser. A 118 (2011) 1915–1924.
[44] J. M. Renes, Equiangular tight frames from Paley tournaments, Linear Algebra Appl. 426 (2007) 497–501.
[45] J. M. Renes, R. Blume-Kohout, A. J. Scott, C. M. Caves, Symmetric informationally complete quantum
measurements, J. Math. Phys. 45 (2004) 2171–2180.
[46] J. J. Seidel, A survey of two-graphs, Coll. Int. Teorie Combin., Atti dei Convegni Lincei 17, Roma (1976)
481–511.
[47] E. Spence, A family of difference sets, J. Combin. Theory Ser. A 22 (1977) 103–106.
[48] T. Strohmer, A note on equiangular tight frames, Linear Algebra Appl. 429 (2008) 326-330.
[49] T. Strohmer, R. W. Heath, Grassmannian frames with applications to coding and communication, Appl.
Comput. Harmon. Anal. 14 (2003) 257–275.
[50] M. A. Sustik, J. A. Tropp, I. S. Dhillon, R. W. Heath, On the existence of equiangular tight frames, Linear
Algebra Appl. 426 (2007) 619–635.
[51] F. Szollosi, All complex equiangular tight frames in dimension 3, arXiv:1402.6429.
[52] J. A. Tropp, Complex equiangular tight frames, Proc. SPIE 5914 (2005) 591401/1–11.
[53] J. A. Tropp, I. S. Dhillon, R. W. Heath, Jr., T. Strohmer, Designing structured tight frames via an alternating
projection method, IEEE Trans. Inform. Theory 51 (2005) 188–209.
[54] R. J. Turyn, Character sums and difference sets, Pacific J. Math. 15 (1965) 319–346.
[55] S. Waldron, On the construction of equiangular frames from graphs, Linear Algebra Appl. 431 (2009) 2228–
2242.
24
[56] L. R. Welch, Lower bounds on the maximum cross correlation of signals, IEEE Trans. Inform. Theory 20
(1974) 397-399.
[57] R. M. Wilson, An existence theory for pairwise balanced designs I. Composition theorems and morphisms, J.
Combin. Theory Ser. A 13 (1972) 220–245.
[58] P. Xia, S. Zhou, G. B. Giannakis, Achieving the Welch bound with difference sets, IEEE Trans. Inform.
Theory 51 (2005) 1900–1907.
[59] G. Zauner, Quantum designs: Foundations of a noncommutative design theory, Ph.D. Thesis, University of
Vienna, 1999.
25
|
1601.02122 | 1 | 1601 | 2016-01-09T15:29:08 | Tensor products and joint spectra for solvable Lie algebras of operators | [
"math.FA"
] | Given two complex Hilbert spaces, $H_1$ and $H_2$, and two complex solvable finite dimensional Lie algebras of operators, $L_1$ and $L_2$, such that $L_i$ acts on $H_i$ (i= 1,2), the joint spectrum of the Lie algebra $L_1\times L_2$, which acts on $H_1\overline\otimes H_2$, is expressed by the cartesian product of $Sp(L_1,H_1)$ and $Sp(L_2,H_2)$. | math.FA | math |
TENSOR PRODUCTS AND JOINT SPECTRA
FOR SOLVABLE LIE ALGEBRAS OF OPERATORS
ENRICO BOASSO
Abstract. Given two complex Hilbert spaces, H1 and H2, and two complex
solvable finite dimensional Lie algebras of operators, L1 and L2, such that Li
acts on Hi (i= 1,2), the joint spectrum of the Lie algebra L1 ×L2, which acts on
H1⊗H2, is expressed by the cartesian product of Sp(L1, H1) and Sp(L2, H2).
1.
Introduction
J. L. Taylor developed in [6] a notion of joint spectrum for an n-tuple a, a =
(a1, ..., an), of mutually commuting operators acting on a Banach space E, i.e.,
ai ∈ L(E), the algebra of all bounded linear operators on E, and [ai, aj] = 0,
1 ≤ i, j ≤ n. This interesting notion, which extends in a natural way the spectrum
of a single operator, has many important properties, among then, the projection
property and the fact that Sp(a, E) is a compact non empty subset of Cn, where
Sp(a, E) denotes the joint spectrum of a in E.
One of the most remarkable results of the Taylor joint spectrum is the one
related with tensor products of tuples of operators. For example, in [2], Z. Ceaus-
escu and F. H. Vasilescu proved the following result. Let Hi, 1 ≤ i ≤ n, be com-
plex Hilbert spaces, and ai, 1 ≤ i ≤ n, be bounded linear operators defined on Hi,
respectively, 1 ≤ i ≤ n. If we denote by H the completion of the tensor product
H1 ⊗ . . . ⊗ Hn with respect to the canonical scalar product, we may consider the
n-tuple of operators a, a = (a1, . . . , an), where ai = 1 ⊗ . . . ⊗ 1 ⊗ ai ⊗ 1 ⊗ . . . ⊗ 1,
1 ≤ i ≤ n, and 1 denotes the identity of the corresponding spaces. Then the
following identity holds,
Sp(a, E) = Sp(a1) × . . . × Sp(an).
Furthermore, in [3], Z. Ceausescu and F. H. Vasilescu showed that if H1 (resp.
H2) is a complex Hilbert space and a = (a1, . . . , an), (resp. b = (b1, . . . , bm)),
is a mutually commuting tuple of operators acting on H1, (resp. H2), then, the
commuting tuple (a, b) = (a1 ⊗ 1, . . . , an ⊗ 1, 1 ⊗ b1, . . . , 1 ⊗ bm) in L(H1⊗H2),
satisfies the relation,
Sp((a, b), H1⊗H2) = Sp(a, H1) × Sp(b, H2),
where 1 denotes the identity map of the corresponding Hilbert spaces, and H1⊗H2
is the completion of the tensor product H1 ⊗ H2 with respect to the canonical
scalar product, see [3,Theorem 2.2].
In [1] we defined a joint spectum for complex solvable finite dimensional Lie
algebras of operators L, acting on a Banach space E, and we denoted it by
Sp(L, E). We proved that Sp(L, E) is a compact non empty subset of L∗ and that
1
2
ENRICO BOASSO
the projection property for ideals still holds. Besides, when L is a commutative
algebra, our spectrum reduces to Taylor joint spectrum in the following sense. If
dim L = n, {ai}(1≤i≤n) is a basis of L and we consider the n-tuple a = (a1, . . . , an),
then {(f (a1), . . . , f (an)) : f ∈ Sp(L, E)} = Sp(a, E), i.e., Sp(L, E) in terms of
the basis of L∗ dual of {ai}(1≤i≤n) coincides with the Taylor joint spectrum of
the n-tuple a. Then, the following question arises naturally. If Hi, i = 1, 2, are
two complex Hilbert spaces, and Li, i = 1, 2, are two complex solvable finite
dimensional Lie algebras of operators such that Li acts on Hi, respectively, i =
1, 2, is there any relation between Sp(L1×L2, H1⊗H2) and Sp(L1, H1)×S(L2, H2).
In this paper we answer this question in the affirmative. Moreover, by a re-
finement of the argument of Z. Ceausescu and F. H. Vasilescu in [3], we extend
the main results of [2] and [3] for complex solvable finite dimensional Lie algebras
and its joint spectrum. In order to describe in more detail our main theorem we
need to introduce a definition. If Hi and Li are as above, i = 1, 2, we consider
the direct product of L1 and L2, i.e., the complex solvable finite dimensional Lie
algebra L1 × L2 defined by,
L1 × L2 = {x1 ⊗ 1 + 1 ⊗ x2 : xi ∈ Li, i = 1, 2},
where 1 is as above. Then, its is clear that L1 × L2 is a Lie algebra of operators
which acts on H1⊗H2, and our main theorem may be stated as follows,
Sp(L1 × L2, H1⊗H2) = Sp(L1, H1) × Sp(L2, H2),
where the above sets are considered as subsets of (L1 × L2)∗ under the natural
identification (L1 × L2)∗ ∼= L∗
1 × L∗
2.
The paper is organized as follows. In Section 2 we review several definitions
and results of [1] and we also prove a proposition which is an important steps to
our main result. Finally, in Section 3, we prove our main theorem.
2. Preliminaries
We briefly recall several definitions and results related to the spectrum of a
complex solvable Lie algebra of operators, see [1]. From now on L denotes a
complex solvable finite dimensional Lie algebra and H a complex Hilbert space
on which L acts as right continous operators, i.e., L is a Lie subalgebra of L(H)
with the opposite product. If dim L = n and f is a character of L, i.e., f ∈ L∗
and f (L2) = 0, where L2 = {[x, y] : x, y ∈ L}, then consider the following chain
complex, (H ⊗ ∧L, d(f )), where ∧L denotes the exterior algebra of L, and dp(f )
is as follows,
dp(f ) : H ⊗ ∧pL → H ⊗ ∧p−1L,
LIE ALGEBRAS OF OPERATORS
3
dp(f )ehx1 ∧ · · · ∧ xpi =
k=p
X
k=1
(−1)k+1e(xk − f (xk))hx1 ∧ . . . ∧ xk ∧ · · · ∧ xpi
+ X
(−1)k+leh[xk, xl] ∧ x1 ∧ . . . ∧ xk ∧ . . . ∧ xl ∧ . . . ∧ xpi,
1≤k<l≤p
wheremeans deletion. If p ≤ 0 or p ≥ n + 1, we define dp(f ) = 0.
If we denote by H∗((H⊗∧L, d(f ))) the homology of the complex (H⊗∧L, d(f )),
we may state our first definition.
Definition 2.1. With L and f as above, the set {f ∈ L∗ : f (L2) = 0, H∗((H ⊗
∧L, d(f ))) 6= 0} is the joint spectrum of L acting on H, and it is denoted by
Sp(L, H).
As we have said, in [1] we proved that Sp(L, E) is a compact non empty subset
of L∗ which reduces to Taylor joint spectrum when L is a commutative algebra,
in the sense explained in the introduction. Besides, if I is an ideal of L and π
denotes the projection map from L∗ to I ∗, then,
Sp(I, H) = π(Sp(L, H)),
i.e., the projection property for ideals still holds. With regard to this property, we
ought to mention the paper of C. Ott, see [5], who pointed out a gap in the proof
of this result, and give another proof of it. In any case, the projection property
remains true.
We observe that the set H ⊗ ∧L has a natural structure of Hilbert space, so
that the sets H ⊗ hxi1 ∧ · · · ∧ xipi, 1 ≤ i1 < . . . < ip ≤ n, 0 ≤ p ≤ n, are
orthogonal subspaces of H ⊗ ∧L, and if <, > denotes the inner product of H,
< ahxi1 ∧ . . . ∧ xipi, bhxi1 ∧ . . . ∧ xipi >=< a, b >.
We shall have occasion to use the direct product of two complex solvable finite
dimensional Lie algebras and its action on the tensor product of two complex
Hilbert spaces. We recall here the main facts which we need for our work. If Hi,
i =1,2, are two complex Hilbert spaces, then H1⊗H2 denotes the completion of
the tensor product H1 ⊗ H2 with respect to the canonical scalar product. Now, if
Li, i=1,2, are two complex solvable finite dimensional Lie algebras of operators,
such that Li acts on Hi, respectively, i=1,2, we consider the algebra L1 × L2, the
direct product of L1 and L2, which acts in a natural way on H1⊗H2, and it is
defined by,
L1 × L2 = {x ⊗ 1 + 1 ⊗ y : x ∈ L1 + y ∈ L2},
where 1 denotes the identity of the corresponding space.
It is clear that L1×L2, defined as above, is a complex solvable finite dimensional
Lie subalgebra of L(H1⊗H2). Moreover, by the structure of the Lie bracket in
L1 × L2, we have two distinguished ideals, L′
2, which we define as follows,
1 and L′
′
L
1 = {x ⊗ 1 : x ∈ L1},
′
L
2 = {1 ⊗ y : y ∈ L2}.
4
ENRICO BOASSO
In addition, if we consider the natural identification K : (L1 × L2)∗ ∼= L∗
1 × L∗
2,
K(f ) = (f ◦ i1, f ◦ i2), where f ∈ (L1 × L2)∗ and ij : Lj → L1 × L2, j = 1, 2,
are the canonical inclusions, as (L1 × L2)2 = L2
2, we have that the set of
characters of L1 × L2 is the cartesian product of the sets of characters of L1 and
L2.
1 × L2
The following proposition is an important step to our main theorem.
Proposition 2.2. Let Hi, i = 1, 2, be two complex Hilbert spaces, and L a
complex solvable finite dimensional Lie algebra of operators acting on H1. Let L′
1,
(resp. L′
2), be the Lie algebra of operators {x⊗1 : x ∈ L}, (resp. {1⊗x : x ∈ L}),
which acts on H1⊗H2, (resp. H2⊗H1), where 1 denotes the identity map of H2,
(resp. H1). Then,
(i) Sp(L
(ii) Sp(L
(iii) Sp(L
1, H1⊗H2) = Sp(L
1, H1⊗H2) ⊆ Sp(L, H1),
2, H2⊗H1) ⊆ Sp(L, H1).
2, H2⊗H1),
′
′
′
′
Proof. It is clear that (iii) is a consequence of (i) and (ii). Let us prove (i).
If f is a character of L, we consider C1 (resp. C2), the Koszul complex associ-
ated to the Lie algebra L
2) and f , then,
′
′
1 (resp. L
1, d1(f )),
′
C1 = (H1⊗H2 ⊗ ∧L
C2 = (H2⊗H1 ⊗ ∧L
′
2, d2(f )),
where the maps di(f ), i= 1,2, are as above.
In addition, a elementary calculation shows that the map µ,
µp : H1⊗H2 ⊗ ∧pL
′
1 → H2⊗H1 ⊗ ∧pL
′
2,
µp(e1 ⊗ e2hx1, . . . , xpi) = e2 ⊗ e1hx1, . . . , xpi,
defines an isomorphism which commutes with di(f ), i=1,2, i.e., µ is an isomor-
phism of chain complexes. Then,
′
′
Sp(L
1, H1⊗H2) = Sp(L
2, H2⊗H1).
In order to verify (ii), let us consider the Koszul complex associated to L and
f , C = (H1 ⊗ ∧pL, d(f )), and C the following chain complex,
C = ((H1 ⊗ ∧L)⊗H2, d(f ) ⊗ 1),
where 1 denotes the identity map of H2.
We observe that if η is the map defined by,
η : (H1⊗H2) ⊗ ∧pL
′
1 → (H1 ⊗ ∧pL)⊗H2,
ηp(e1 ⊗ e2hx1, . . . , xpi) = e1hx1, . . . , xpi ⊗ e2,
then an easy calculation shows that η defines an isomorphism of chain complexes
between C1 and C.
Now, if f does not belong to Sp(L, H1), by [8, Lemma 2.4], the complex C is
exact. As η is an isomorphism of chain complexes, if C1 is exact, then f does not
belong to Sp(L
′
1, H1⊗H2).
(cid:3)
LIE ALGEBRAS OF OPERATORS
5
We now state our main result.
3. The Main Result
Theorem 3.1. Let H1 and H2 be two complex Hilbert spaces and L1 and L2 two
complex solvable finite dimensional Lie algebras of operators such that Li acts on
Hi, respectively, i= 1,2. Let us consider the complex solvable finite dimensional
Lie algebra L1 × L2, which acts on H1⊗H2 and it is defined by,
L1 × L2 = {x1 ⊗ 1 + 1 ⊗ x2 : xi ∈ Li, i = 1, 2},
where 1 denotes the identity of the corresponding spaces.
Then,
Sp(L1 × L2, H1⊗H2) = Sp(L1, H1) × Sp(L2, H2),
1 × L∗
2 under the natural identification K : (L1 × L2)∗ ∼= L∗
where, in the above equality, the set Sp(L1 × L2, H1⊗H2) is considered as a subset
of L∗
2 of Section 2.
Proof. In order to prove that Sp(L1 × L2, H1⊗H2) is contained in the cartesian
product of Sp(L1, H1) and Sp(L2, H2), let L
2) be the ideal of L1 × L2
defined by {x ⊗ 1 : x ∈ L1} (resp. {1 ⊗ y : y ∈ L2}), where 1 is as above, see
Section 2. Then, by the projection property of the joint spectrum, Proposition 1
and the identification K,
1 (resp. L
1 × L∗
′
′
Sp(L1 × L2, H1⊗H2) ⊆ Sp(L
1, H1⊗H2) × Sp(L
2, H1⊗H2)
′
′
⊆ Sp(L1, H1) × Sp(L2, H2).
Let us prove the converse inclusion. We consider the following preliminary
facts first.
Let fi be a character of Li and Ki the Koszul complex of Li acting on Hi
associated to fi, for i = 1, 2,
K1 = (H1 ⊗ ∧L1, d1(f1)),
K2 = (H2 ⊗ ∧L2, d2(f2)).
We observe that there is a natural identification between the spaces (H1⊗H2)⊗
∧(L1 × L2) and (H1 ⊗ ∧L1) ⊗ (H2 ⊗ ∧L2). If we denote by ψ this identification,
ψ is the following map,
ψ : (H1 ⊗ H2) ⊗ ∧(L1 × L2) → (H1 ⊗ ∧L1) ⊗ (H2 ⊗ ∧L2),
ψ(e1 ⊗ e2hx1 ∧ · · · ∧ xp ∧ y1 ∧ . . . ∧ yqi) = e1hx1 ∧ · · · ∧ xpi ⊗ e2hy1 ∧ · · · ∧ yqi,
where e1 ∈ H1, e2 ∈ H2, p ∈ [[1, n]], q ∈ [[1, m]], n = dim(L1), and m = dim(L2).
As
(H1⊗H2) ⊗ ∧k(L1 × L2) = ⊕p+q=k(H1 ⊗ ∧pL1)⊗(H2 ⊗ ∧qL2),
if we consider Hi ⊗∧Li, i=1,2, (H1⊗H2)⊗∧(L1 ×L2) and (H1 ⊗∧L1)⊗(H2 ⊗∧L2)
with their natural structure of Hilbert spaces, a straightforward calculation shows
that the map ψ may be extended to an isometric isomorphism from (H1⊗H2) ⊗
∧(L1 × L2) onto (H1 ⊗ ∧L1)⊗(H2 ⊗ ∧L2).
Besides, if f is a character of L1 × L2 and if we define fj = f ◦ ij ∈ L∗
j , where
ij : Lj → L1 × L2 are the canonical inclusions and j = 1, 2, i.e., if we consider f
6
ENRICO BOASSO
decomposed under the natural identification K : (L1 × L2)∗ ∼= L∗
the Koszul complex of L1 × L2 acting on H1⊗H2 associated to f ,
1 × L∗
2, if K is
K = ((H1⊗H2) ⊗ ∧(L1 × L2), dk(f )),
then an easy calculation shows that,
ψdk(f ) = (d1(f1) ⊗ 1 + ξ ⊗ d2(f2))ψ,
where ξ is the map,
ξ : ⊕n
p=0 (H1 ⊗ ∧pL1) → ⊕n
p=0(−1)p,
ξ = ⊕n
p=0(H1 ⊗ ∧pL1),
1 is the identity map of H1 and dk(f ) is the boundary map of the Koszul complex
K, equivalently, if we consider the algebraic tensor product of the complexes K1
and K2, K1 ⊗ K2, and its natural completion K1⊗K2, the map ψ provides an
isometric isomorphism of chain complexes, from K onto K1⊗K2.
Moreover, if Ti, i = 1, 2, and Tk are the maps,
Ti = di(fi) + di(fi)
∗
,
Tk = dk(f ) + dk(f )
∗
,
as ξ is a selfadjoint map, an easy calculation shows that,
ψTk = (T1 ⊗ 1 + ξ ⊗ T2)ψ.
Let us return to the proof. If f does not belong to Sp(L1 × L2, H1⊗H2), by
[7, Lemma 3.1], the operator Tk is an invertible map. On the other hand, if fi,
i = 1, 2, belongs to Sp(Li, Hi), i = 1, 2, as Ti is a selfadjoint operator, i= 1,2,
there exist by [7, Lemma 3.1] and by [4, Chapter II, Section 31, Theorem 2] two
sequences of unit vectors, (ai
n ∈ Hi ⊗ ∧Li,
n) → 0 (n → ∞), for i = 1, 2. However, as ψ and ψ−1 are isometric
and Ti(ai
isomorphisms, by elementary properties of the tensor product in Hilbert spaces,
we have that: k a1
n)) → 0
(n → ∞), equivalently, by [4, Chapter II, Section 31, Theorem 2], 0 ∈ Sp(Tk),
which is impossible by our assumption.
n)n∈N, i = 1, 2, such that k ai
n) k= 1 and Tk(ψ−1(a1
n k= 1, ai
n ⊗ a2
n ⊗ a2
n k= 1, k ψ−1(a1
n ⊗ a2
(cid:3)
References
1. E. Boasso and A. Larotonda, A spectral theory for solvable Lie algebras of operators, Pacific
J. Math. 158 (1993), 15-22.
2. Z. Ceausescu and F. H. Vasilescu, Tensor products and Taylor's joint spectrum, Studia
Math. 62 (1978), 305-311.
3. Z. Ceausescu and F. H. Vasilescu, Tensor products and the joint spectrum in Hilbert spaces,
Proc. Amer. Math. Soc. 72 (1978), 505-508.
4. P. Halmos, Introduction to Hilbert spaces and the theory of spectral multiplicity, Chelsea
Publishing Company, 1972.
5. C. Ott, A note on a paper of E. Boasso and A. Larotonda, Pacific J. Math., 173 (1996),
173-179.
6. J. L. Taylor, A joint spectrum for several commuting operators, J. Funct. Anal. 6 (1970),
172-191.
7. F. H. Vasilescu, Analytical perturbations of the δ-operator and integral representation
formulas in Hilbert spaces, J. Operator Theory 1 (1979), 187-205.
LIE ALGEBRAS OF OPERATORS
7
8. C. Grosu and F. H. Vasilescu, The Kunneth formula for Hilbert complexes, Integral Equa-
tions Operator Theory 5 (1982), 1-17.
Enrico Boasso
E-mail address: enrico [email protected]
|
1808.01189 | 1 | 1808 | 2018-08-03T13:46:27 | On the exponential ultradistribution semigroups in Banach spaces | [
"math.FA"
] | In this paper, we continue our previous research studies of exponential ultradistribution semigroups in Banach spaces. The existence and uniqueness of analytical solutions of abstract fractional relaxation equations associated with the generators of exponential ultradistribution semigroups have been considered. Some other results are also proved. | math.FA | math |
ON THE EXPONENTIAL ULTRADISTRIBUTION SEMIGROUPS
IN BANACH SPACES
MARKO KOSTI ´C, STEVAN PILIPOVI ´C, AND DANIEL VELINOV
Abstract. In this paper, we continue our previous research studies of ex-
ponential ultradistribution semigroups in Banach spaces. The existence and
uniqueness of analytical solutions of abstract fractional relaxation equations as-
sociated with the generators of exponential ultradistribution semigroups have
been considered. Some other results are also proved.
1. Introduction and preliminaries
The class of (exponential) distribution semigroups in Banach spaces was intro-
duced by J. L. Lions in [24]. Along with the papers [3] by W. Arendt on integrated
semigroups and vector-valued Widder representation theorem, and [6] by G. Lumer
and I. Cioranescu on K-convoluted semigroups, the paper [24] has been a motivation
for a great number of other authors to study various classes of (ultra-)distribution
semigroups in the setting of Banach spaces or, more generally, locally convex spaces
(cf.
[1], [2], [10]-[12], [17], [22], [25]-[26], [29]-[30] and references cited therein for
further information). Following the approaches employed in [23]-[24], the authors
of [20] have defined the classses of L-ultradistribution semigroups and ultradistri-
bution semigroups with non-densely defined generators. In [21], some structural
characterizations of ultradistribution semigroups and exponential ultradistribution
semigroups have been proved. Motivated by the ideas from D. Fujiwara's research
of exponential distribution semigroups [9], we shall prove some new characteriza-
tions of spaces of vector-valued tempered ultradistributions in Proposition 2.1. In
Theorem 2.3, we consider the existence and uniqueness of analytical solutions of
abstract fractional relaxation equations associated with the generators of exponen-
tial ultradistribution semigroups. Some new qualitative properties of solutions of
the first order abstract Cauchy problem (ACP ) clarified below have been proved in
the case that the operator A generates an exponential ultradistribution semigroup
of (Mp)-class and that the initial value x belongs to the abstract Beurling space
associated with the operator A.
We use the standard terminology throughout the paper. We impose the following
conditions (cf. [14]) on the sequence (Mp)p∈N0 of positive real numbers satisfying
M0 = 1:
(M.1) M 2
(M.2) For some A, H > 0, Mp ≤ AH p min0≤q≤p Mp−qMq,
(M.3)′ :
p ≤ Mp−1Mp+1 for p ∈ N;
p, q ∈ N;
Every employment of the condition
Xp=1
∞
Mp−1
Mp
< ∞.
1
2
MARKO KOSTI ´C, STEVAN PILIPOVI ´C, AND DANIEL VELINOV
(M.3) :
∞
sup
p∈N
Xq=p+1
Mq−1Mp+1
pMpMq
< ∞,
K
:=(cid:8)ϕ ∈ C∞(Rn) : supp(ϕ) ⊆ K and there exists C > 0 such that kDαϕkC(K) ≤ Ch−αMα, α = 0, 1, 2,···(cid:9).
becomes a Ba-
:= supx∈K,α∈N0
K
Mp
which is a slightly stronger than (M.3)′, will be explicitly emphasized. The Gevrey
sequence Mp = p!s, s > 1 satisfies all the above conditions. The associated function
M (ρ) of sequence (Mp) on (0, +∞) is defined by M (ρ) = supp∈N ln ρp
, ρ > 0. We
know that he function t 7→ M (t), t ≥ 0 is increasing as well as that limλ→∞ M (λ) =
∞ and that the function M (·) vanishes in some open neighborhood of zero. Put
mp := Mp/Mp−1, p ∈ N. Then (M.1) implies that the sequence (mp)p∈N is increas-
ing.
The spaces of ultradifferentiable functions of ∗-class and ultradistributions of
∗-class are defined in the sequel. Let ∅ 6= Ω ⊆ R, and let ∅ 6= K ⋐ Ω denote a
compact set in Ω. Set, for every h > 0,
D{Mp},h
Equipped with the norm k·kD
nach space. Define also, as locally convex spaces, D(Mp)
D(Mp)(Ω) := lim indK⋐ΩD(Mp)
lim indK⋐ΩD{Mp}
E {Mp}(Ω) will be understood in the sense of [13].
type, are defined in [27] as duals of the following test spaces
, D{Mp},h
:= lim projh→+∞D{Mp},h
,
and D{Mp}(Ω) :=
. The notions of spaces E (Mp)(K), E (Mp)(Ω), E {Mp}(K) and
The spaces of tempered ultradistributions of the Beurling, resp., the Roumieu
:= lim indh→0D{Mp},h
K , D{Mp}
Dα·(x)
h−αMα
{Mp},h
K
K
K
K
K
K
S(Mp)(R) := lim projh→∞SMp,h(R), resp., S{Mp}(R) := lim indh→0SMp,h(R),
where for each h > 0,
and
SMp,h(R) :=(cid:8)φ ∈ C∞(R) : kφkMp,h < ∞(cid:9)
kφkMp,h := sup( hα+β
MαMβ(cid:0)1 + t2(cid:1)β/2(cid:12)(cid:12)φ(α)(t)(cid:12)(cid:12) : t ∈ R, α, β ∈ N0).
The common notation for symbols (Mp) and {Mp} will be ∗.
It is said that a function P (ξ) = Pα∈N0
the class (Mp), resp. {Mp}, if the coefficients aα satisfy the estimate
(1.1)
aαξα, ξ ∈ R is an ultrapolynomial of
aα ≤
for some L > 0 and C > 0, resp.
, α ∈ N0,
CLα
Mα
for every L > 0 and some CL > 0. The
aαDα is said to be an ultradifferential
corresponding operator P (D) = Pα∈N0
operator of the class (Mp), resp. {Mp} (see [13]).
Let X and Y be two Hausdorff sequentially complete locally convex spaces over
the field of complex numbers. By L(X, Y ) we denote the space consisting of all
continuous linear mappings from X into Y ; L(X) ≡ L(X, X). A C0 semigroup
(T (t))t≥0 on X (cf. [16] and [28] for the notion) is called locally equicontinuous if
the family {T (t) : t ∈ [0, t0]} is equicontinuous for all t0 > 0; (T (t))t≥0 is called
(globally) equicontinuous if the family {T (t) : t ≥ 0} is equicontinuous. Recall
that any C0-semigroup in a barreled space X is automatically locally equicontinuous
ON THE EXPONENTIAL ULTRADISTRIBUTION SEMIGROUPS IN BANACH SPACES
3
([16]). If for some α ∈ R the C0-semigroup (Tα(t) ≡ e−αtT (t))t≥0 is equicontinuous,
then (T (t))t≥0 is called α quasi -- equicontinuous.
Henceforth we shall always assume that (E,k·k) is a complex Banach space. Let
A be a closed linear operator with domain D(A) and range R(A) contained in E.
Since no confusion seems likely, we will identify A with its graph; [D(A)] denotes
the Banach space D(A) equipped with the graph norm kxk[D(A)] := kxk + kAxk,
x ∈ D(A). As is usually the case, ρ(A) denotes the resolvent set of A and R(λ : A)
denotes the operator (λI−A)−1 (λ ∈ ρ(A)), where I stands for the identity operator
on E. If F is a linear subspace of E, then we define the part of A in F, AF
for short, by AF := {(x, y) ∈ A : x, y ∈ F}. Set D∞(A) := T∞
n=1 D(An) and
Σα := {z ∈ C \ {0} : arg(z) < α} (α ∈ (0, π]). We denote by Aα(E) the vector
space consisting of all analytic mappings from Σα into E (α ∈ (0, π]). The following
system of seminorms k·kn ≡ k·k +··· +kAn·k, n ∈ N0 turns D∞(A) into a Fr´echet
space. We define the abstract Beurling space of (Mp) class associated to a closed
linear operator A as in [5]. Put E(Mp)(A) := lim projh→+∞E{Mp}
(A), where
h
E{Mp}
h
(A) :=(x ∈ D∞(A) : kxk{Mp}
(A), k · k{Mp}
h
h
≡ sup
p∈N0
hpkApxk
Mp
< ∞).
h
h
h′
(A) ⊆ E{Mp}
) is a Banach space, E{Mp}
Then (E{Mp}
(A) if 0 < h <
h′ < ∞, E(Mp)(A) is a Fr´echet space, and E(Mp)(A) is a dense subspace of E
whenever A is the generator of a regular (Mp)-ultradistribution semigroup ([5]).
By D′∗
+ (R, L(E)) we denote the space consisting of those L(E)-valued ultradis-
tributions G of ∗-class for which supp(G) ⊆ [0,∞) (cf. [17] for the notion) and by
D∗
0 we denote the space consisting of those ultradifferentiable functions ϕ of ∗-class
such that ϕ(t) = 0 for t ≤ 0. We refer the reader to [13]-[14] for further informa-
tion concerning the convolution of (vector-valued) ultradifferentiable functions and
ultradistributions of ∗-class.
Definition 1.1. ([20]) Let G ∈ D′∗
ultradistribution semigroup of ∗-class if:
φ, ψ ∈ D∗
(U.1) G(φ ∗ ψ) = G(φ)G(ψ),
0;
(U.2) N (G) :=Tφ∈D∗
N (G(φ)) = {0};
(U.3) R(G) :=Sφ∈D∗
R(G(φ)) is dense in E;
(U.4) For every x ∈ R(G) there exists a function u ∈ C([0,∞), E) satisfying
+ (R, L(E)). Then it is said that G is an L-
0
0
u(0) = x and G(φ)x =
∞
Z0
φ(t)u(t) dt, φ ∈ D∗
0.
If G ∈ D′∗
+ (R, L(E)) satisfies
(U.5) G(φ ∗0 ψ) = G(φ)G(ψ),
0 , where
t
φ, ψ ∈ D∗
Z0
(f ∗0 g)(t) :=
f (t − s)g(s) ds, t ∈ R,
then it is said that G is a pre-(UDSG) of ∗-class. If (U.5) and (U.2) are fulfilled for
G, then G is called an ultradistribution semigroup of ∗-class, in short, (UDSG). A
pre-(UDSG) G is said to be dense if it additionally satisfies (U.3).
4
MARKO KOSTI ´C, STEVAN PILIPOVI ´C, AND DANIEL VELINOV
Definition 1.2. An L-ultradistribution semigroup G of ∗-class is said to be ex-
ponential, EL-ultradistribution semigroup of ∗-class in short,
if in addition to
(U.1) − (U.4), G fulfills:
(U.6) (∃a ≥ 0)(e−a·G ∈ S′∗(R, L(E))).
Then we say that G is of order a. Conditions (U.6), (U.5), resp., (U.6), (U.5) and
(U.2), define an exponential pre-(UDSG), resp., exponential (UDSG), and they are
denoted by pre-(EUDSG), resp., (EUDSG).
The definition of (infinitesimal) generator A of an ultradistribution semigroup G
of ∗-class is defined by A := {(x, y) ∈ E × E : G(−ϕ′)x = G(ϕ)y for all ϕ ∈ D∗
0}.
Since every L-ultradistribution semigroup G of ∗-class is also an ultradistribution
semigroup of ∗-class ([20]), the definition of (infinitesimal) generator of such a
semigroup is clear. We refer the reader to [20] for the notion of an (exponential)
ultradistribution fundamental solution for a closed linear operator A acting on E.
For the sake of convenience, we shall remind the reader of the following important
result.
Theorem 1.3. ([19]) Suppose A is a closed linear operator on E. Then there exists
an exponential ultradistribution fundamental solution of ∗-class for A iff there exist
a ≥ 0, k > 0 and L > 0, in the Beurling case, resp., there exists a ≥ 0 such that,
for every k > 0 there exists Lk > 0, in the Roumieu case, such that:
{λ ∈ C : ℜλ > a} ⊆ ρ(A), and
kR(λ : A)k ≤ LeM(kλ), λ ∈ C, ℜλ > a, resp.,
kR(λ : A)k ≤ LkeM(kλ)
for all k > 0 and λ ∈ C,ℜλ > a.
Before proceeding further, we would like to write down the following facts about
(exponential) ultradistribution semigroups of ∗-class. If the operator A generates an
(exponential) ultradistribution semigroup G of ∗-class, then there exists an injective
operator C ∈ L(E) such that the operator A generates a (an exponentially bounded)
[17, Section 3.6]). This implies that
C-regularized semigroup (T (t))t≥0 on E (cf.
there exist two Fr´echet spaces Y and W such that Y ֒→ X ֒→ W (֒→ means the
continuous embedding), as well as that AY generates a C0-semigroup on Y and that
A = BX, where B generates a C0-semigroup on W ; without going into full details, it
should be only noticed that in (Mp)-case Y may be chosen to be exactly E(Mp)(A),
the abstract Beurling space of the operator A (then AE(Mp )(A) ∈ L(E(Mp)(A));
cf.
[17, Theorem 3.6.6] for more details). Furthermore, if the ultradistribution
semigroup G generated by A is exponential, then Y and W may be chosen to be
Banach spaces ([8]).
derivative Dα
Let α > 0, β ∈ R, γ ∈ (0, 1), and let m = ⌈α⌉. Then the Caputo fractional
t u(t) is defined for those functions u ∈ Cm−1([0,∞) : E) for which
gm−α ∗ (u −Pm−1
k=0 uk
m−1
·k
k! ) ∈ Cm([0,∞) : E), by
dtm"gm−α ∗ u −
t u(t) =
dm
Dα
k!!#.
uk ·k
Xk=0
The Mittag-Leffler function Eα,β(z) is defined by Eα,β(z) := P∞
n=0(zn/Γ(αn +
β)), z ∈ C. Here we assume that 1/Γ(αn + β) = 0 if αn + β ∈ −N0. Set, for
short, Eα(z) := Eα,1(z), z ∈ C. The Wright function Φγ(·) is defined by Φγ(t) :=
ON THE EXPONENTIAL ULTRADISTRIBUTION SEMIGROUPS IN BANACH SPACES
5
L−1(Eγ(−λ))(t), t ≥ 0, where L−1 denotes the inverse Laplace transform. We refer
the reader to [4] and [18, Section 1.3] for more details about Mittag-Leffler and
Wright functions.
2. Characterization of exponential ultradistribution semigroups
The following proposition is motivated by D. Fujiwara's research of exponential
distribution semigroups (cf. [9, Lemma 1]).
Proposition 2.1.
(i) Let a0 ≥ 0, and let e−a0·G ∈ L(S(Mp)(R), L(E)). Then
there exists h′ > 0 such that, for every a > a0 and for every non-empty
compact set K ⊆ [0,∞), there exists C > 0 such that for any ϕ, ψ ∈ D(Mp)
with supp(ϕ)∪ supp(ψ) ⊆ K, the following estimate holds:
0
(2.1)
(cid:13)(cid:13)e−atG(ϕ ∗ ψ)(cid:13)(cid:13)L(E) ≤ CkϕkD
{Mp},h′
K
kψkL1.
K
(ii) Let a0 ≥ 0, and let e−a0·G ∈ L(S{Mp}(R), L(E)). Then, for every h′ > 0,
for every a > a0, and for every non-empty compact set K ⊆ [0,∞), there
exists C > 0 such that for any ϕ, ψ ∈ D{Mp},h′
, the estimate (2.1) holds.
Proof. Let ζ ∈ D{Mp}, and let ζ(t) = 1 for all t ∈ [−1, 1]. Put γ1 := Hζ (H is the
1 = δ + Hζ′ in the sense of Roumieu ultradistributions.
Heaviside function). Then γ′
So, with ω1 = Hζ′ ∈ D{Mp}
, we have
δ′ ∗ γ1 = δ + ω1.
(2.2)
Let the numbers h′ > 0 and a > a0 be fixed, let K be a non-empty compact subset of
[0,∞), and let ϕ, ψ ∈ D{Mp},h′
. We will prove that the function e(a0−a)·(ϕ∗ ψ)(·) ∈
SMp,h(R) for all h ∈ (0, h′H/4(1 + a − a0)); here H denotes the constant from the
condition (M.2). Set K ′ := (K+supp(γ1)) ∪ (K+supp(ω1)) and K ′′ := K + K ′.
Using the product rule, Young's inequality and the condition (M.2), we have that
K
0
sup
sup
=
sup
≤
+
MαMβ
MαMβ
: t ∈ R, α, β ∈ N0)
(cid:13)(cid:13)e(a0−a)·(ϕ ∗ ψ)(·)(cid:13)(cid:13)Mp,h = sup( hα+β
t∈R, α, β∈N0( hα+β(cid:0)1 + t2(cid:1)β/2
t∈R, α, β∈N0( hα+β(cid:0)1 + t2(cid:1)β/2
t∈R, α, β∈N0( hα+β(cid:0)1 + t2(cid:1)β/2
t∈K ′′, α, β∈N0( hα+β(cid:0)1 + t2(cid:1)β/2
t∈K ′′, α, β∈N0( hα+β(cid:0)1 + t2(cid:1)β/2
≤ eM(cid:0)h√1+(sup K ′′)2(cid:1)
t∈K ′′, α∈N0( hα
MαMβ(cid:0)1 + t2(cid:1)β/2(cid:12)(cid:12)(cid:12)(cid:10)δ(α) ∗ e(a0−a)·(ϕ ∗ ψ)(·)(cid:11)(t)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:10)δ(α+1) ∗ γ1 ∗ (e(a0−a)·ϕ ∗ e(a0−a)·ψ)(cid:11)(t) −(cid:10)δ(α) ∗ ω1 ∗ (e(a0−a)·ϕ ∗ e(a0−a)·ψ)(cid:11)(t)(cid:12)(cid:12)(cid:12)
ZR(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α+1)
ZR(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α)
ZK ′(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α+1)
ZK ′(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α)
Mα ZK ′(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α+1)
ds)
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ γ1(cid:17)(s)(cid:12)(cid:12)(cid:12)
ds)
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ ω1(cid:17)(s)(cid:12)(cid:12)(cid:12)
ds)
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ γ1(cid:17)(s)(cid:12)(cid:12)(cid:12)
ds)
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ ω1(cid:17)(s)(cid:12)(cid:12)(cid:12)
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ γ1(cid:17)(s)(cid:12)(cid:12)(cid:12)
ds)
MαMβ
MαMβ
MαMβ
+
sup
=
sup
sup
6
MARKO KOSTI ´C, STEVAN PILIPOVI ´C, AND DANIEL VELINOV
ds)
α+1
(t − s)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:0)e(a0−a)·ψ ∗ ω1(cid:17)(s)(cid:12)(cid:12)(cid:12)
Xj=0(cid:13)(cid:13)ϕ(j)kL∞(K) (cid:13)(cid:13)e(a0−a)·ψ ∗ γ1(cid:13)(cid:13)L1(R))
Xj=0(cid:13)(cid:13)ϕ(j)kL∞(K) (cid:13)(cid:13)e(a0−a)·ψ ∗ ω1(cid:13)(cid:13)L1(R))
α
sup
Mα
Mα
4
A
≤
Mα ZK ′(cid:12)(cid:12)(cid:12)(cid:16)e(a0−a)·ϕ(·)(cid:17)(α)
α∈N0( (2(1 + a − a0)h)α
α∈N0( (2(1 + a − a0)h)α
+ eM(cid:0)h√1+(sup K ′′)2(cid:1)
t∈K ′′, α∈N0( hα
≤ 2(1 + a − a0)eM(cid:0)h√1+(sup K ′′)2(cid:1) sup
+ 2(1 + a − a0)eM(cid:0)h√1+(sup K ′′)2(cid:1) sup
(1 + a − a0)eM(cid:0)h√1+(sup K ′′)2(cid:1)(cid:13)(cid:13)ψ(cid:13)(cid:13)L1(R)h(cid:13)(cid:13)γ1(cid:13)(cid:13)L1(R) +(cid:13)(cid:13)ω1(cid:13)(cid:13)L1(R)i
)
(cid:17)j (2(1 + a − a0)h)α−j Mj
(1 + a − a0)eM(cid:0)h√1+(sup K ′′)2(cid:1)(cid:13)(cid:13)ψ(cid:13)(cid:13)L1(R)h(cid:13)(cid:13)γ1(cid:13)(cid:13)L1(R) +(cid:13)(cid:13)ω1(cid:13)(cid:13)L1(R)i
(1 + a − a0)eM(cid:0)h√1+(sup K ′′)2(cid:1)(cid:13)(cid:13)ψ(cid:13)(cid:13)L1(R)h(cid:13)(cid:13)γ1(cid:13)(cid:13)L1(R) +(cid:13)(cid:13)ω1(cid:13)(cid:13)L1(R)ikϕkD
Xj=0(cid:16) 2(1 + a − a0)h
Xj=0(cid:16) 2(1 + a − a0)h
eM(2(1+a−a0)h)
α∈N0(α+1
sup
H jMjMα−j
× kϕkD
× kϕkD
{Mp},h′
K
4
A
≤
{Mp},h′
K
(cid:17)j
h′
∞
h′H
4
A
≤
The final conclusion follows from the decomposition he−a·G, ϕ∗ψi = he−a0·G, e(a0−a)·(ϕ∗
ψ)i and the continuity of mapping e−a0·G : SMp,h → L(E), with the number
h ∈ (0, h′H/4(1 + a − a0)) chosen arbitrarily.
Remark 2.2.
(i) Notice that we have not used the structural theorem for the
space of tempered vector-valued ultradistributions in the proof (cf.
[21,
Theorem 2.2]) and that we do not need the condition (M.3) here. Observe
also that the proof of Proposition 2.1 works in the case that E is a general
sequentially complete locally convex space ([22]).
{Mp},h′
K
(cid:3)
eM(h′H/2).
(ii) As mentioned above, Proposition 2.1 is inspired by Lemma 1 of [9]. Re-
grettably, it seems that the assertions of [9, Theorem 2-Theorem 3] can-
not be so simply reconsidered for exponential ultradistribution semigroups.
Speaking-matter-of-factly, in ultradistribution case we cannot prove the ex-
istence of number N such that an analog of the estimate [9, (9)] holds.
In the remaining part of this section, we shall always assume that the condition
(M.3) holds. The following entire function of exponential type zero ([5], [13]) plays
an important role in our further work:
∞
ω(Mp)(z) =:
Yi=1(cid:0)1 +
iz
mp(cid:1),
z ∈ C.
Then ω(Mp)(z) ≥ eM(z), z ∈ C and the properties (P.1)-(P.5) stated in [17,
Subsection 3.6.2] hold.
Suppose now that A generates an exponential ultradistribution semigroup of
(Mp)-class. By Theorem 1.3, it readily follows that there exist numbers a ≥ 0,
k > 0 and L > 0 such that {λ ∈ C : ℜλ > a} ⊆ ρ(A) and kR(λ : A)k ≤ LeM(kλ),
λ ∈ C, ℜλ > a. Let ¯a > a. Then there exists a sufficiently large number n0 ∈ N such
ON THE EXPONENTIAL ULTRADISTRIBUTION SEMIGROUPS IN BANACH SPACES
7
that, for every n ∈ N with n ≥ n0, the operator A generates a global exponentially
bounded Cn-regularized semigroup (Sn(t))t≥0 on E, where
Sn(t)x :=
1
2πi
¯a+i∞
Z¯a−i∞
eλt R(λ : A)x
ωn(iλ)
dλ,
t ≥ 0, n ∈ N, x ∈ E and Cn := Sn(0).
cf. the proof of [17, Theorem 3.6.4] (in this result, we have clarified the result on
generation of local C-regularized semigroups; observe, however, that in the case of
exponential ultradistribution semigroups the resulting Cn-regularized semigroup is
global and exponentially bounded because the integration is taken along the line
connecting the points ¯a− i∞ and ¯a + i∞, not along the upwards oriented boundary
of a suitable ultra-logarithmic region of (Mp)-class). An elementary application of
Cauchy's formula, taken together with the consideration in the paragraph contain-
ing the equation [17, (316)] and [17, Lemma 3.6.5, Theorem 3.6.6], implies that
E(Mp)(A) ⊆ Cn(D∞(A)), n ≥ n0 and that for each x ∈ E(Mp)(A) a unique solution
of the abstract Cauchy problem
u ∈ C∞([0,∞) : E) ∩ C([0,∞) : [D(A)]),
u′(t) = Au(t), t ≥ 0,
u(0) = x,
(ACP ) :
is given by u(t) = Sn(t)C−1
n x, t ≥ 0, n ≥ n0. Using the same argumentation as
in the proof of [17, Theorem 3.6.4], we get that for each x ∈ E(Mp)(A) and h > 0
there exist n ≥ n0 and c > 0 such that
(2.3)
sup
p∈N0
hp(cid:13)(cid:13)
dp
dtp u(t)(cid:13)(cid:13)Mp
≤ ce¯at(cid:13)(cid:13)C−1
n x(cid:13)(cid:13),
t ≥ 0.
Applying [18, Theorem 2.4.2] (cf. also [4, Theorem 3.1, Theorem 3.3]) and the
Ljubich uniqueness theorem for abstract time-fractional equations [18, Theorem
2.1.34], it readily follows that for each x ∈ E(Mp)(A) and α ∈ (0, 1) there exists a
unique solution of the following abstract fractional Cauchy problem
v ∈ C([0,∞) : E) ∩ Amin(( 1
Dα
v(0) = x,
t v(t) = Av(t), t ≥ 0,
α −1) π
2 ,π)(E),
(ACP )α :
the following theorem.
0 Φα(s)Sn(stα)C−1
n x ds, t ≥ 0 (n ≥ n0). Thus, we have proved
given by v(t) := R ∞
Theorem 2.3. Suppose that (M.3) holds for (Mp) and A generates an exponential
ultradistribution semigroup of (Mp)-class. Then for each x ∈ E(Mp)(A) there exists
a unique solution u(t) of the abstract Cauchy problem (ACP ) satisfying additionally
that for each h > 0 there exist n ≥ n0 and c > 0 such that (2.3) holds. Furthermore,
for each x ∈ E(Mp)(A) and α ∈ (0, 1) there exists a unique solution of the abstract
fractional Cauchy problem (ACP )α.
It is worth noting that the assertion of Theorem 2.3 continues to hold, with
suitable modifications, in the setting of Fr´echet spaces and that possible applications
can be made to differential operators considered in [17, Example 3.5.15] and [22,
Example 9.3].
8
MARKO KOSTI ´C, STEVAN PILIPOVI ´C, AND DANIEL VELINOV
References
[1] W. Arendt, C. J. K. Batty, M. Hieber and F. Neubrander, Vector-valued Laplace Transforms
and Cauchy Problems. Birkhauser Verlag, Basel, 2001.
[2] W. Arendt, O. El-Mennaoui and V. Keyantuo, Local integrated semigroups: evolution with
jumps of regularity. J. Math. Anal. Appl. 186 (1994), 572 -- 595.
[3] W. Arendt, Vector-valued Laplace transforms and Cauchy problems. Israel J. Math. 59
(1987), 327 -- 352.
[4] E. Bazhlekova, Fractional Evolution Equations in Banach Spaces. PhD Thesis, Eindhoven
University of Technology, Eindhoven, 2001.
[5] I. Cioranescu, Beurling spaces of class (M
Math. 102 (1978), 167 -- 192.
p) and ultradistribution semi-groups. Bull. Sci.
[6] I. Cioranescu and G. Lumer, Probl`emes d'´evolution r´egularis´es par un noyan g´en´eral K(t),
Formule de Duhamel, prolongements, th´eor`emes de g´en´eration. C. R. Acad. Sci. Paris S´er. I
Math. 319 (1995), 1273 -- 1278.
[7] G. Da Prato and U. Mosco, Semigruppi distribuzioni analitici. Ann. Scuola Norm. Sup. Pisa
19 (1965), 367-396.
[8] R. deLaubenfels, C-semigroups and strongly continuous semigroups. Israel J. Math. 81
(1993), 227 -- 255.
[9] D. Fujiwara, A characterisation of exponential distribution semi-groups. J. Math. Soc. Japan.
18 (1966), 267 -- 274.
[10] M. Hieber, Integrated semigroups and differential operators on Lp spaces. Math. Ann. 29
(1991), 1- 16.
[11] M. Li, F. Huang and Q. Zheng, Local integrated C-semigroups. Studia Math. 145 (2001),
265 -- 280.
[12] V. Keyantuo, Integrated semigroups and related partial differential equations. J. Math. Anal.
Appl. 212 (1997), 135 -- 153.
[13] H. Komatsu, Ultradistributions, I. Structure theorems and a characterization. J. Fac. Sci.
Univ. Tokyo Sect. IA Math. 20 (1973), 25 -- 105.
[14] H. Komatsu, Ultradistributions, III. Vector valued ultradistributions the theory of kernels. J.
Fac. Sci. Univ. Tokyo Sect. IA Math. 29 (1982), 653 -- 718.
[15] H. Komatsu, Microlocal Analysis in Gevrey Class and in Complex Domains, Microlocal
analysis and applications. (Montecatini Terme, 1989), 161 -- 236, Lecutre Notes in Math.,
1494, Springer, Berlin, 1991.
[16] T. K¯omura, Semigroups of operators in locally convex spaces, J. Funct. Anal. 2 (1968), 258 --
296.
[17] M. Kosti´c, Generalized Semigroups and Cosine Functions. Mathematical Institute SANU,
Belgrade, 2011.
[18] M. Kosti´c, Abstract Volterra Integro-Differential Equations. Taylor and Francis Group/CRC
Press/Science Publishers, New York, in press.
[19] M. Kosti´c, Differential and analytical properties of semigroups of operators. Integral Equa-
tions Operator Theory 67 (2010), 499 -- 557.
[20] M. Kosti´c and S. Pilipovi´c, Ultradistribution semigroups. Siberian Math. J. 53 (2012), 232 --
242.
[21] M. Kosti´c, S. Pilipovi´c and D. Velinov, Structural theorems for ultradistribution semigroups.
To appear in Siberian Math. J.
[22] M. Kosti´c, S. Pilipovi´c and D. Velinov, C-distribution semigroups and C-ultradistribution
semigroups in locally convex spaces. Preprint.
[23] P. C. Kunstmann, Distribution semigroups and abstract Cauchy problems. Trans. Amer.
Math. Soc. 351 (1999), 837 -- 856.
[24] J. L. Lions, Semi-groupes distributions. Portugal. Math. 19 (1960), 141 -- 164.
[25] I. V. Melnikova, A. I. Filinkov, Abstract Cauchy Problems: Three Approaches. Chapman &
Hall/CRC, Boca Raton, 2001.
[26] F. Neubrander, Integrated semigroups and their applications to the abstract Cauchy problem.
Pacific J. Math. 135 (1988), 111 -- 155.
[27] S. Pilipovi´c, Tempered ultradistributions, Boll. Un. Mat. Ital. 7 (1988), 235-251.
[28] K. Yosida, Functional Analysis. Springer-Verlag, Berlin, 1965.
ON THE EXPONENTIAL ULTRADISTRIBUTION SEMIGROUPS IN BANACH SPACES
9
[29] K. Yoshinaga, Values of vector-valued distributions and smoothness of semigroup distribu-
tions. Bull. Kyushu Inst. Tech. Math. Nat. Sci. 12 (1965), 1 -- 27.
[30] S. Wang, Quasi-distribution semigroups and integrated semigroups. J. Funct. Anal. 146
(1997), 352 -- 381.
Faculty of Technical Sciences, University of Novi Sad, Trg D. Obradovi´ca 6, 21125
Novi Sad, Serbia
E-mail address: [email protected]
Department for Mathematics and Informatics, University of Novi Sad, Trg D. Obradovi´ca
4, 21000 Novi Sad, Serbia
E-mail address: [email protected]
Department for Mathematics, Faculty of Civil Engineering, Ss. Cyril and Methodius
University, Skopje, Partizanski Odredi 24, P.O. box 560, 1000 Skopje, Macedonia
E-mail address: [email protected]
|
1906.10349 | 1 | 1906 | 2019-06-25T06:58:09 | Optimal extension of the Fourier transform and convolution operator on compact groups | [
"math.FA"
] | Let $G$ be a compact group (not necessarily abelian) and let $\Phi$ be a Young function satisfying the $\Delta_2$-condition. We determine the optimal domain and the associated extended operator for both Fourier transform and the convolution operator defined on the Orlicz spaces $L^\Phi(G).$ | math.FA | math |
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND
CONVOLUTION OPERATOR ON COMPACT GROUPS
MANOJ KUMAR AND N. SHRAVAN KUMAR
Abstract. Let G be a compact group (not necessarily abelian) and let Φ be a
Young function satisfying the ∆2-condition. We determine the optimal domain and
the associated extended operator for both Fourier transform and the convolution
operator defined on the Orlicz spaces LΦ(G).
1. Introduction
Let X be a σ-order continuous Banach function space [13] defined over a finite
positive measure space (Ω,A, µ). If Y is any Banach space and T : X → Y is a
bounded linear operator, define νT : A → Y as νT (A) = T (χA). It is well known
that νT is a σ-additive vector measure. Further, by [13, Pg. 194, Theorem 4.14], X
can be embedded continuously inside L1(νT ) and the mapping IνT : L1(νT ) → Y
as the integration operator. Also, the space L1(νT ) is optimal in the sense that if
given by IνT (f ) =RΩ f dνT is a continuous extension of T. The mapping IνT is called
T extends continuously to a space eX with values in Y, then eX can be embedded
continuously inside L1(νT ).
The determination of the optimal domain and the extension for operators arising
from analysis is one of the classical problems of functional analysis. This optimal
extension process is known for kernel operators [1] and Sobolev embeddings [2, 5, 8].
In 2007, S. Okada and W. J. Ricker considered the optimal extension problem for
the convolution operators arising from measures defined on compact abelian groups.
Further, in [13, Chapter 7], the authors have also studied the Fourier transform
operator on compact abelian groups from the optimal domain view point. This
paper aims to study the optimal domain, extension and their properties for both
convolution operator and Fourier transform defined on a compact group. The initial
domain for these operators are the Orlicz spaces on a compact group. In this way,
these results are new even when the underlying group is abelian.
In section 3, we study the Fourier transform operators while in section 4, we study
the convolution operator and the compactness properties of the extended operator.
2010 Mathematics Subject Classification. Primary 43A15, 43A30, 46G10; Secondary 46E30,
43A77.
Key words and phrases. Compact group, Vector measure, Fourier transform, Convolution, Orlicz
space.
1
2
MANOJ KUMAR AND N. SHRAVAN KUMAR
2. Preliminaries
2.1. Fourier analysis on compact groups. Let G be a compact Hausdorff group
and let mG denote the normalized positive Haar measure on G. For 1 ≤ p ≤ ∞,
Lp(G) will denote the usual pth-Lebesgue space with respect to the measure mG.
It is well known that an irreducible unitary representation of a compact group G
Let {(Xα,k.kα)}α∈∧ be a collection of Banach spaces. We shall denote by ℓ1- ⊕α∈∧
is always finite-dimensional. Let bG be the set of all unitary equivalence classes of
irreducible unitary representations of G. The set bG is called the unitary dual of G
and bG is given the discrete topology.
Xα : Pα∈∧kxαkα < ∞(cid:27) equipped with the norm
Xα, the Banach space (cid:26)(xα) ∈ Π
k(xα)k∞ := Pα∈∧kxαkα. Similarly, we shall denote by ℓ∞- ⊕α∈∧
(cid:26)(xα) ∈ Π
α∈∧kxαkα < ∞(cid:27) equipped with the norm k(xα)k∞ := sup
α∈∧kxαkα.
Xα, the space consisting of those
Xα, the Banach space
Xα : sup
α∈∧
α∈∧
In the same lines, we shall also denote by c0- ⊕α∈∧
vectors (xα) from ℓ∞- ⊕α∈∧
is a closed subspace of ℓ∞- ⊕α∈∧
Xα.
Xα which goes to 0 as α → ∞. It is clear that c0- ⊕α∈∧
Xα
Let π be an irreducible unitary representation of G on the Hilbert space Hπ of
dimension dπ. For f ∈ L1(G), the Fourier transform of f at π denoted bf (π), is
defined as
bf (π) =ZG
f (t)π(t)∗ dmG(t), [π] ∈ bG.
Note that the Fourier transform operator f 7→ bf maps L1(G) into ℓ∞- ⊕[π]∈ bG B2(Hπ).
This operator is injective and bounded.
For more details on compact groups, we refer to [6].
2.2. Orlicz Spaces. Let Φ : [0,∞] → [0,∞] be a convex function. Then Φ is called
a Young function if it satisfies Φ(0) = 0 and lim
Φ(x) = +∞. If Φ is any Young
x→∞
function, then define Ψ as
Ψ(y) := sup {xy − Φ(x) : x ≥ 0}, y ≥ 0.
Then Ψ is also a Young function and is termed as the complementary function to Φ.
Further, the pair (Φ, Ψ) is called a complementary pair of Young functions. A pair
(Φ, Ψ) of complementary Young functions is said to be normalized if Φ(1)+Ψ(1) = 1.
We say that a Young function Φ satisfies the ∆2-condition if there exists a constant
K > 0 and x0 > 0 such that Φ(2x) ≤ KΦ(x) whenever x ≥ x0.
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR 3
The Orlicz space, denoted LΦ(G), is a Banach space consisting of complex mea-
surable functions on G for which
kfkΦ := inf(cid:26)k > 0 :ZG
Φ(cid:18)f
k (cid:19) dmG ≤ 1(cid:27) < ∞.
The above norm is called as the Luxemburg norm or Gauge norm. If (Φ, Ψ) is
a complementary Young pair, then there is a norm on LΦ(G), equivalent to the
Luxemburg norm, given by,
Φ = sup(ZG f gdmG :ZG
kfko
Ψ(g)dmG ≤ 1).
This norm is called as the Orlicz norm. Note that the Orlicz space is a natural
generalisation of the classical Lp spaces.
Let Φ be a Young function satisfying the ∆2-condition. Then the space of all
continuous functions on G is dense in LΦ(G). Similarly, the space of simple functions
is also dense in LΦ(G). Further, the dual of (LΦ(G),k·kΦ) is isometrically isomorphic
to (LΨ(G),k · ko
Ψ). In particular, if both Φ and Ψ satisfies the ∆2-condition, then
LΦ(G) is reflexive [14, Pg. 112, Theorem 10]. Further, by [16, Pg. 595, Theorem
133.3], LΦ(G) is a σ-order continuous Banach function space.
The following is an analogue of [7, Pg. 380, Lemma 35.11]. The proof of this
follows as in the case of Lp spaces.
Lemma 2.1. Let Φ be a Young function satisfying the ∆2-condition and let µ ∈
M(G) such that
sup{kµ ∗ fkΦ : kfk1 ≤ 1} < ∞.
Then there exists g ∈ LΦ(G) such that dµ = gdmG.
For more on Orlicz spaces see [14].
2.3. Vector measure. Let G be a compact group. Let B(G) denote the σ-algebra
consisting of all Borel subsets of G. Further, let M(G) denote the space of all
bounded complex Radon measures on G. Let X be a complex Banach space and let
ν be a σ-additive X-valued measure called vector measure on G. Let X ′ be the dual
of X and let BX ′ be the closed unit ball in X ′. For each x′ ∈ X ′, we shall denote
by hν, x′i, the corresponding scalar valued measure for the vector measure ν, which
is defined as hν, x′i(A) = hν(A), x′i, A ∈ B(G). A set A ∈ B(G) is said to be ν-null
if ν(B) = 0 for every Borel set B ⊂ A. The variation of ν, denoted ν, is a positive
measure defined as follows: For a set A ∈ B(G),
ν(A) = sup(XE∈ρ
kν(E)k : ρ the finite partition of A) .
4
MANOJ KUMAR AND N. SHRAVAN KUMAR
The vector measure ν is said to be measure of bounded variation if ν(G) < ∞.
x′∈BX ′hν, x′i(A),
The semivariation of ν on a set A ∈ B(G) is given by kνk(A) = sup
where hν, x′i is the total variation of the scalar measure hν, x′i. Let kνk denote
the quantity kνk(G). The vector measure ν is said to be absolutely continuous with
respect to a non-negative scalar measure µ if
ν(A) = 0, A ∈ B(G). A Banach
space X is said to have the Radon-Nikodym Property with respect to (G, B(G), mG)
if for each X-valued absolutely continuos vector measure ν of bounded variation
there exists a Bochner integrable function f : G → X such that dν = f dmG.
A complex valued function f on G is said to be ν-weakly integrable if f ∈
w(ν) the Banach space of all
L1(hν, x′i), for all x′ ∈ X ′. We shall denote by L1
ν-weakly integrable functions equipped with the norm
lim
µ(A)→0
kfkν = sup
x′∈BX ′ZG f dhν, x′i.
A ν-weakly integrable function f is said to be ν-integrable if for each A ∈ B(G) there
exists a unique xA ∈ X such thatRA f dhν, x′i = hxA, x′i, x′ ∈ X ′. The vector xA is
denoted byRA f dν. We shall denote by L1(ν) the space of all ν-integrable functions
the continuous linear operator Iν : L1(ν) → X given by Iν(f ) =RG f dν, f ∈ L1(ν).
and it is also a Banach space when equipped with the k · kν norm. We denote by Iν
See [13, Pg. 152]. Note that the space of simple functions on G is dense in L1(ν).
Let T : LΦ(G) → X be a bounded linear operator. Define νT : B(G) → X as
νT (A) = T (χA). Then νT is a vector measure on G. The measure νT will be called
as the vector measure associated with the operator T. See [13, Pg. 185].
For more details on vector measures and integration with respect to vector mea-
sures, we refer to [3, 4, 13].
Throughout this paper G will be a compact group and mG will denote the unique
normalized Haar measure on G. Further, Φ will always denote a Young function sat-
isfying the ∆2-condition and Ψ will be the Young function which is complementary
to Φ. Also, LΦ(G) will be considered only with the Luxemburg norm while its dual
LΨ(G) will be considered only with the Orlicz norm.
3. Optimal extension of the Fourier transform
In this section, we study the Fourier transform from two different view points.
Firstly we restrict the Fourier transform operator on L1(G) to LΦ(G) by treating
LΦ(G) as a subspace of L1(G). On the other hand, there is also a Fourier transform on
the LΦ(G) spaces arising from the Hausdorff-Young inequality. The optimal domain
for these operators is studied in this section. The results obtained in this section are
motivated by the results presented in [13, Chapter 7].
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR 5
3.1. Optimal extension of c0-valued Fourier transform. Define FΦ,0 : LΦ(G) →
c0- ⊕[π]∈ bG B2(Hπ) as FΦ,0(f ) = bf . Here bf denotes the Fourier transform of f as an
element of L1(G), since LΦ(G) is contained in L1(G). Further, it is well known that
FΦ,0 is a bounded linear operator. We shall denote by νΦ,0 the associated vector
measure.
Lemma 3.1. The vector measure νΦ,0 has finite variation and the variation of νΦ,0
coincides with the Haar measure mG.
Proof. Let A ∈ B(G). Then,
Thus νΦ,0 is finite. Let π0 denote the one dimensional representation given by
π0(t) = 1 for all t ∈ G. Let
kνΦ,0(A)k∞ =kcχAk∞ ≤ kχAk1 = mG(A).
χ{π0}(π) =( 1
[π] = [π0]
[π] 6= [π0]
if
if
0dπ
,
for [π] ∈ bG. Here 0dπ denotes the dπ × dπ zero matrix. Then, it is clear that χ{π0} ∈
B2(Hπ ). Further,
bG
Bℓ1- ⊕
[π]∈
hνΦ,0(A), χ{π0}i =hcχA, χ{π0}i
=X[π]∈ bG
tr(cχA(π)χ{π0}(π))
=tr(cχA(π0)) = cχA(π0)
=ZA
π0(t−1) dmG(t) = mG(A).
Therefore, mG(A) = hνΦ,0(A), χ{π0}i ≤ kνΦ,0(A)k∞. Hence kνΦ,0(A)k∞ = mG(A).
Since A ∈ B(G) is arbitrary, it follows that νΦ,0 = mG.
(cid:3)
As an immediate corollary we have the following.
Corollary 3.2. The measures νΦ,0 and mG have same null sets.
Remark 3.3. By Corollary 3.2 and [13, Pg. 194, Theorem 4.14], the space L1(νΦ,0)
is the largest Banach function space with σ-order continuous norm such that LΦ(G)
is continuously embedded into L1(νΦ,0) and the operator FΦ,0 has a continuous ex-
tension with values in c0- ⊕[π]∈ bGB2(Hπ). Further, this extension is unique and is the
integration operator IνΦ,0.
Corollary 3.4. The spaces L1(νΦ,0) and L1(G) coincide. Further, the extended op-
erator IνΦ,0 coincides with the Fourier transform on L1(G).
6
MANOJ KUMAR AND N. SHRAVAN KUMAR
Proof. By Lemma 3.1, it follows that the spaces L1(G) and L1(νΦ,0) coincide. Fur-
ther, the inclusions
L1(νΦ,0) ⊆ L1(νΦ,0) ⊆ L1(hνΦ,0, χ{π0}i)
are clear, where the first inclusion follows from [13, Pg. 116, Lemma 3.14]. Here π0
and χ{π0} are defined as in proof of Lemma 3.1. Also, it follows from the proof of
Lemma 3.1 that L1(hνΦ,0, χ{π0}i) coincides with L1(G). Thus the spaces L1(G) and
L1(νΦ,0) coincide. Hence, by the density of the space continuous functions on G in
both L1(G) and LΦ(G), it follows that the extended operator is just the Fourier
transform on L1(G).
(cid:3)
Corollary 3.5. Let (Φ, Ψ) be a pair of complementary Young functions satisfying
the ∆2-condition. If G is an infinite compact group then the domain of the optimal
extension of FΦ,0 contains LΦ(G) as a proper subspace.
Proof. Since Φ and Ψ both satisfy the ∆2-condition, the space LΦ(G) is reflexive
and hence the proof follows from Corollary 3.4.
(cid:3)
3.2. Optimal extension of the Hausdorff-Young inequality. Let (Φ, Ψ) de-
note a pair of complementary Young functions which are normalized. Further, we
shall denote by ℓΦ- ⊕[π]∈ bGB2(Hπ), the Banach space consisting of (x[π])[π]∈ bG ∈ Π
[π]∈ bGB2(Hπ)
for which
k(x[π])kΦ := inf
k > 0X[π]∈ bG
d2
πΦ(cid:18)kx[π]kB2(Hπ )
k√dπ (cid:19) ≤ Φ(1)
< ∞.
The following is the Hausdorff-Young inequality for Orlicz spaces. See [15, Pg. 211,
Theorem 2]. Also, see [15] for any undefined notations regarding this theorem.
Theorem 3.6. Let (Φ, Ψ) be a complementary pair of Young functions which are
continuous and normalized. Suppose that
(i) Φ ≺′ Φ0, where Φ0(x) = c1x2 and
(ii) Ψ′(x) ≤ c2xr,
for some r ≥ 1, c1, c2 > 0. Then the Fourier transform is a bounded linear operator
from LΦ(G) into ℓΨ- ⊕[π]∈ bG B2(Hπ).
Throughout this subsection, the complementary Young pair (Φ, Ψ) will satisfy the
conditions of the above theorem. We shall denote by FΦ,Ψ the operator from LΦ(G)
to ℓΨ- ⊕[π]∈ bG B2(Hπ) given by the above theorem. Further, we shall also denote by
νΦ,Ψ the associated vector measure.
Lemma 3.7. The measures νΦ,Ψ and mG have same null sets.
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR 7
Proof. Let A ∈ B(G). Then mG(A) = 0 if and only ifcχA = 0 if and only if νΦ,Ψ(A) =
0. Hence the proof.
Remark 3.8. By Lemma 3.7 and by [13, Pg. 194, Theorem 4.14], the space L1(νΦ,Ψ)
is the largest Banach function space with σ-order continuous norm such that LΦ(G)
is continuously embedded into L1(νΦ,Ψ) and the operator FΦ,Ψ has a continuous
extension with values in ℓΨ- ⊕[π]∈ bGB2(Hπ). Further, this extension is unique and is the
integration operator IνΦ,Ψ.
(cid:3)
Theorem 3.9. Let (Φ, Ψ) be a complementary pair of Young functions satisfying
the ∆2-condition.
(i) The space LΦ(G) is a dense subspace of L1(νΦ,Ψ) and kfkνΦ,Ψ ≤ kFΦ,ΨkkfkΦ,
(ii) The space L1(νΦ,Ψ) is a dense subspace of L1(G) and kfk1 ≤ kfkνΦ,Ψ, f ∈
f ∈ LΦ(G).
L1(νΦ,Ψ).
(iii) The extended operator IνΦ,Ψ : L1(νΦ,Ψ) → ℓΨ- ⊕[π]∈ bGB2(Hπ) is just the Fourier
transform map, i.e., IνΦ,Ψ(f ) = bf , for all f ∈ L1(νΦ,Ψ).
Proof. (i) Since L1(νΦ,Ψ) is the optimal domain, it follows that LΦ(G) is contained
in L1(νΦ,Ψ). Further, by [13, Pg. 185, Theorem 4.4], it follows that the inclusion of
LΦ(G) inside L1(νΦ,Ψ) is continuous with its operator norm equal to kFΦ,Ψk. The
conclusion follows from the density of simple functions in both the spaces.
(ii) Consider the element χ{π0} from ℓΦ- ⊕[π]∈ bGB2(Hπ). Here π0 and χ{π0} are defined
as in the proof of Lemma 3.1. Here, note that χ{π0} ∈ BℓΦ- ⊕
B2(Hπ ). Again, as in the
proof of Lemma 3.1, one can show that mG = hνΦ,Ψ, χ{π0}i. Thus, if f ∈ L1(νΦ,Ψ),
then
[π]∈
bG
ZG fdmG =ZG fdhνΦ,Ψ, χ{π0}i ≤ kfkνΦ,Ψ.
Thus f ∈ L1(G). Further, the density of L1(νΦ,Ψ) in L1(G), again, follows from the
density of the simple functions.
(iii) This is again a consequence of the density of simple functions and (ii). (cid:3)
4. Optimal extension of the convolution operator
In this section, we consider the convolution operator and determine its optimal
domain. Further, we also characterize when the extended operator will be compact.
The results of this section are analogues of the results obtained in [11]. Throughout
this section, (Φ, Ψ) will denote a pair of complementary Young functions satisfying
the ∆2-condition.
8
MANOJ KUMAR AND N. SHRAVAN KUMAR
Let µ ∈ M(G). Define TΦ,µ : LΦ(G) → LΦ(G) as TΦ,µ(f ) = µ ∗ f, where
µ ∗ f (t) =ZG
f (s−1t)dµ(s).
Then TΦ,µ is a bounded linear operator and kTΦ,µk ≤ kµk. Let νΦ,µ denote the
associated vector measure. If dµ = gdmG, then νΦ,g will denote the measure νΦ,µ.
4.1. Optimal extension.
Proposition 4.1. Let g ∈ LΦ(G). Let FΦ,g(t) = τtg, t ∈ G, where τtg(s) = g(st).
(i) If f ∈ L1(G), then the function t 7→ f FΦ,g(t) from G to LΦ(G) is integrable
(in the sense of Bochner).
(ii) If f ∈ LΦ(G), then g ∗ f =RG f FΦ,g dmG.
Proof. (i) Since G is compact, it follows that the range of the continuous mapping
t 7→ f FΦ,g(t) is compact and hence separable. Thus, by Pettis measurability theorem
[3, Pg. 42, Theorem 2], it follows that the function f FΦ,g is strongly measurable.
By the fact that LΦ(G) is translation invariant and that the translations are norm
(ii) By (i), Fubini's theorem and the fact that the dual of LΦ(G) is LΨ(G), we
get that
preserving, it follows thatRG kf FΦ,gkΦ dmG < ∞, thus proving (i).
f FΦ,g dmG, h(cid:29) , ∀ h ∈ LΨ(G).
hg ∗ f, hi =(cid:28)ZG
(cid:3)
Corollary 4.2. Let g ∈ LΦ(G). Then the LΦ(G)-valued measure νΦ,g can be written
as
νΦ,g(A) =ZA
FΦ,g dmG,
where FΦ,g is defined as in Proposition 4.1.
Proof. Follows from (ii) of Proposition 4.1 by letting f = χA.
(cid:3)
Corollary 4.3. Let µ ∈ M(G). If f ∈ LΦ(G) such that dµ = f dmG a.e., then there
exists an integrable function (in the sense of Bochner) F : G → LΦ(G) such that for
every Borel set A of G,
νΦ,µ(A) =ZA
F dmG.
Proof. By our assumption that the function f is the Radon-Nikodym derivative of
µ w.r.t. the Haar measure, it follows that νΦ,µ = νΦ,f . Now the proof follows from
Corollary 4.2.
(cid:3)
The following theorem is the converse to the previous corollary.
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR 9
exists f ∈ LΦ(G) such that dµ = f dmG a.e..
Theorem 4.4. Let µ ∈ M(G). Suppose that an integrable function (in the sense
of Bochner) F : G → LΦ(G) satisfies νΦ,µ(A) =RA F dmG, A ∈ B(G). Then there
Proof. Let F : G → LΦ(G) such that νΦ,µ(A) =RA F dmG. By [3, Pg. 42, Theorem
2], there exists a Borel set E of G such that mG(E) = 1 and {F (x) : x ∈ E}
is separable in LΦ(G). So, without loss of generality, we can assume that F (G) is
a separable subset of LΦ(G). Let FΦ,0 denote the Fourier transform defined as in
section 3. As F (G) is separable and the map FΦ,0 is continuous, (FΦ,0 ◦ F )(G) is
a separable subset of c0- ⊕[π]∈ bGB2(Hπ). Thus, by definition of the c0 direct sum, there
(1)
i.e.,
bf (π) = 0dπ ∀ f ∈ F (G) and ∀ [π] ∈ bG \ Γ.
exists a countable subset Γ of bG such that if f ∈ F (G) then bf vanishes outside Γ,
Let [π] ∈ bG. Now, consider the map eFπ : G → B2(Hπ) given by eFπ(t) = dF (t)(π).
Since G is compact, LΦ(G) is continuously embedded in L1(G) and hence the above
map is well-defined. Note that the linear maps FΦ,0 : LΦ(G) → c0- ⊕[π]∈ bGB2(Hπ)
and Pπ : c0- ⊕[σ]∈ bGB2(Hσ) → B2(Hπ) are continuous, where Pπ is the projection onto
B2(Hπ) and hence, by [13, Pg. 149], the composition Pπ ◦ FΦ,0 ◦ F is Bochner
integrable. Further, as
ZA eFπ(t) dmG(t) =ZA
F (t) dmG(t)(cid:19) (Pπ and FΦ,0 are linear maps)
it follows that eFπ is Bochner integrable.
We now claim that, for every Borel subset A of G,
eFπ = Pπ ◦ FΦ,0 ◦ F,
(2)
(3)
where τtµ(A) = µ(At−1). Let A be a Borel subset of G. Then
ZA eFπ(t) dmG(t) =ZA cτtµ(π) dmG(t),
(Pπ ◦ FΦ,0 ◦ F )(t) dmG(t) (by (2))
=Pπ ◦ FΦ,0(cid:18)ZA
=Pπ ◦ FΦ,0 (νΦ,µ(A)) = \νΦ,µ(A)(π)
=\µ ∗ χA(π) = cχA(π)bµ(π)
=(cid:18)ZA
π(t−1) dmG(t)(cid:19)bµ(π)
=ZA
π(t−1)bµ(π) dmG(t)
10
MANOJ KUMAR AND N. SHRAVAN KUMAR
π(st)∗ dµ(s) dmG(t)
=ZAZG
=ZA cτtµ(π) dmG(t).
ZA
Let [π] ∈ bG \ Γ. Then, by (1) and (3), we have, for every Borel set A of G,
π(t−1)bµ(π) dmG(t) =ZAcτtµ(π) dmG(t) =ZA eFπ(t) dmG(t) = 0dπ .
This implies that π(t−1)bµ(π) = 0dπ a.e. t ∈ G. As π is a unitary representation, it
follows thatbµ(π) = 0dπ and hence cτtµ(π) = 0dπ , for every t ∈ G. Thus,
(4)
On the other hand, using the fact that Γ is countable and (3), there exists a Borel
set A of G such that mG(A) = 1 and
cτtµ(π) = eFπ(t) = 0dπ ∀ t ∈ G.
cτtµ(π) = eFπ(t) ∀ t ∈ A and ∀ [π] ∈ Γ.
By (4) and (5) we have
(5)
cτtµ(π) = eFπ(t) = dF (t)(π) ∀ t ∈ A and ∀ [π] ∈ bG
and therefore by the uniqueness of the Fourier transform it follows that τtµ =
F (t) ∀ t ∈ A. Now the conclusion follows by applying τt−1 on both sides.
(cid:3)
Lemma 4.5. If µ ∈ M(G), then the semivariation of νΦ,µ satisfies the following:
for any Borel set A,
µ(G)
Ψ−1(1)
mG(A) ≤ kνΦ,µk(A) = sup
g∈BLΨ(G)ZA eµ ∗ g dmG ≤(cid:20)Φ−1(cid:18) 1
mG(A)(cid:19)(cid:21)−1
kµk.
hνΦ,µ, gi(A) = hTΦ,µ(χA), gi = hχA, T ∗
Proof. Note that, for any g ∈ LΨ(G), T ∗
For any g ∈ LΨ(G), we have
Thus, by [3, Pg. 46, Theorem 4] and from the definition of the semivariation of a
vector measure, it follows that
Φ,µ(g) =eµ ∗ g. Let A be a Borel subset of G.
Φ,µ(g)i = hχA,eµ ∗ gi =ZAeµ ∗ g dmG.
g∈BLΨ(G)ZA eµ ∗ g dmG.
Ψ−1(1) χG(cid:13)(cid:13)(cid:13)
Note that, by [14, Pg. 78, Corollary 7], it follows that(cid:13)(cid:13)(cid:13) 1
Ψ−1(1)ZA eµ ∗ χG dmG.
the above equality, it follows that
kνΦ,µk(A) = sup
kνΦ,µk(A) ≥
= 1. Thus, from
o
Ψ
1
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR11
Now the inequality on the left follows from the fact that eµ ∗ χG = µ(G). Further,
the right inequality follows from the Holder's inequality for Orlicz spaces [14, Pg.
62, Remark 1] and [14, Pg. 78, Corollary 7].
(cid:3)
Corollary 4.6. If µ ∈ M(G), then the vector measure νΦ,µ is absolutely continuous
w.r.t. mG. Further, if µ(G) is non-zero, then mG is absolutely continuous w.r.t. νΦ,µ.
Proof. The proof of this follows from Lemma 4.5.
(cid:3)
Remark 4.7. If µ ∈ M(G) is such that µ(G) 6= 0 then, by Corollary 4.6, the
measures νΦ,µ and mG have same null sets. Hence, by [13, Pg. 194, Theorem 4.14],
the space L1(νΦ,µ) is the largest Banach function space with σ-order continuous norm
such that LΦ(G) is continuously embedded into L1(νΦ,µ) and the operator TΦ,µ has a
continuous extension with values in LΦ(G). Further, this extension is unique and is
the integration operator IνΦ,µ.
Corollary 4.8. Let µ ∈ M(G). Then, there exists f ∈ LΦ(G) such that dµ = f dmG
a.e. if and only if νΦ,µ < ∞.
Proof. Forward part follows from Corollary 4.3 and [3, Pg. 46, Theorem 4(iv)]. For
the converse, note that the space LΦ(G) is reflexive and hence, by [3, Pg. 76, Corol-
lary 13], LΦ(G) has the Radon-Nikodym property. Thus the conclusion follows from
Corollary 4.6 and Theorem 4.4.
(cid:3)
4.2. Properties of the optimal domain. From now onwards µ will denote a
complex Radon measure such that µ(G) 6= 0.
Proposition 4.9. Let f be a complex-valued Borel measurable function on G. Then,
f ∈ L1(νΦ,µ) if and only ifRG feµ ∗ g dmG < ∞ for every g ∈ LΨ(G). Further, if
f ∈ L1(νΦ,µ), then
(6)
kfkνΦ,µ = sup
g∈BLΨ(G)ZG feµ ∗ g dmG.
(7)
Proof. Since Φ satisfies the ∆2-condition, by [14, Pg. 139, Theorem 5], LΦ(G) does
not contain a copy of c0 and hence, by [13, Pg. 138], the spaces L1(νΦ,µ) and L1
w(νΦ,µ)
coincide. Let g ∈ LΨ(G). Then, by following the steps as in Lemma 4.5 and by [3,
Pg. 46, Theorem 4], it follows that
hνΦ,µ, gi(A) =ZA eµ ∗ g dmG.
Thus, f ∈ L1(νΦ,µ) if and only if f ∈ L1
w(νΦ,µ), i.e.,RG f dhνΦ,µ, gi < ∞ for every
g ∈ LΨ(G), i.e.,RG feµ ∗ g dmG < ∞ for every g ∈ LΨ(G). Now the formula for
the norm of f ∈ L1(νΦ,µ) follows from the definition and from (7).
Theorem 4.10.
(cid:3)
12
MANOJ KUMAR AND N. SHRAVAN KUMAR
(i) If f ∈ LΦ(G), then
kfkνΦ,µ ≤ kfkΦkµk.
(ii) The inclusion ı : LΦ(G) → L1(νΦ,µ) is continuous with the operator norm
satisfying
µ(G) ≤ kık ≤ kµk.
Proof. (i) The proof follows from (6) and from the Holder's inequality for Orlicz
spaces [14, Pg. 62, Remark 1].
(ii) The continuity of the inclusion map ı and the inequality on the right follows
from (i). We now prove the inequality on the left side. By (6), it is clear that
kık = sup
g∈BLΨ(G)keµ ∗ gko
Ψ
Ψ for all g ∈ BLΨ(G). In particular, taking g = 1
Ψ−1(1) χG, we
(cid:3)
and hence kık ≥ keµ ∗ gko
get the required inequality.
Corollary 4.11. The space LΦ(G) is dense in L1(νΦ,µ).
Proof. This follows from the previous theorem and the fact that the space of simple
functions is dense in the both of the spaces w.r.t. their respective norms.
(cid:3)
Corollary 4.12. The extended map IνΦ,µ : L1(νΦ,µ) → LΦ(G) is given by IνΦ,µ(f ) =
µ ∗ f.
Proof. This follows from [13, Pg. 185, Theorem 4.4] and Corollary 4.11.
(cid:3)
Theorem 4.13.
(i) If f ∈ L1(νΦ,µ), then
kfk1 ≤
Ψ−1(1)
µ(G) kfkνΦ,µ.
(ii) The inclusion : L1(νΦ,µ) → L1(G) is continuous with the operator norm
satisfying
(iii) The range of is a dense and translation invariant subspace of L1(G).
1
kνΦ,µk ≤ kk ≤
Ψ−1(1)
µ(G)
.
particular, taking g = 1
Proof. (i) It follows from (6) that kfkνΦ,µ ≥RG feµ ∗ g dmG for all g ∈ BLΨ(G). In
(ii) The continuity of the inclusion map and the inequality on the right follows
from (i). We now prove the inequality on the left side. From the definition of the
operator norm, it follows that kk ≥ k(f )k1 for all f ∈ BL1(νΦ,µ). In particular,
taking f = 1
Ψ−1(1) χG, we get the required inequality.
kνΦ,µk χG, we get the required inequality.
Borel set A of G,
Borel set A of G,
νΦ,µ(A) =ZA
νΦ,µ(A) =ZA
F1 dmG.
F2 dmG.
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR13
(iii) Denseness of the range of follows from the fact that the space of simple
functions is dense in the both of the spaces w.r.t. their respective norms. Further,
translation invariance follows from (6) and the fact that translation on LΨ(G) is an
isometry.
(cid:3)
4.3. Compactness of the extended operator. A function f : G → X is said
to be Pettis integrable if hf, x′i ∈ L1(G), x′ ∈ X ′ and if for each A ∈ B(G) there
exists a unique xA ∈ X such thatRAhf, x′i dmG = hxA, x′i, x′ ∈ X ′.
We now characterize the compactness of the extended operator in terms of the
measure νΦ,µ.
Theorem 4.14. Let µ ∈ M(G). Then the following are equivalent:
(i) The LΦ(G)-valued integral operator IνΦ,µ is compact.
(ii) The measure νΦ,µ has finite variation.
(iii) There exists a LΦ(G)-valued Bochner integrable function F1 such that for every
(iv) There exists a LΦ(G)-valued Pettis integrable function F2 such that for every
(v) There exists a Borel set A of G such that νΦ,µ(A) is non-zero and finite.
(vi) There exists f ∈ LΦ(G) such that dµ = f dmG.
Proof. (i)⇒(ii) is a consequence of [12, Theorem 4].
(ii)⇒(iii) follows from Corollary 4.8 and Corollary 4.3.
(iii)⇒(iv) follows from the fact that a Bochner integrable function is Pettis inte-
grable [13, Pg. 149].
(iv)⇒(v) is a consequence of the fact that a vector measure defined by using a
Pettis integrable function has σ-finite variation [9, Corollary 1].
(v)⇒(vi). Let A be a Borel subset subset of G such that 0 < νΦ,µ(A) < ∞.
Then, by [13, Pg. 106] and Lemma 4.5, mG(A) > 0. As LΦ(G) is reflexive, by [3,
Pg. 76, Corollary 13] applied to the measure νΦ,µ restricted to A, we get a function
F : A → LΦ(G) which is the Radon-Nikodym derivative of νΦ,µ restricted to A w.r.t.
the Haar measure mG. Now the argument given for Theorem 4.4 can be modified to
obtain the necessary conclusion.
(vi)⇒(i). By Corollary 4.8 the measure νΦ,µ has finite variation and by [3, Pg.
46, Theorem 4] and Corollary 4.2, it follows that the variation of νΦ,µ is kfkΦmG.
Further, the Radon-Nikodym derivative dνΦ,µ
with a compact range. Thus
the compactness of the extended operator IνΦ,µ follows from [12, Theorem 1].
(cid:3)
dνΦ,µ is FΦ,f
kf kΦ
14
MANOJ KUMAR AND N. SHRAVAN KUMAR
Our next characterization is in terms of the L1-spaces.
Corollary 4.15. Let µ ∈ M(G). Then the following are equivalent:
(i) The LΦ(G)-valued integral operator IνΦ,µ is compact.
(ii) The spaces L1(G) and L1(νΦ,µ) coincide.
(iii) The spaces L1(G) and L1(νΦ,µ) coincide.
(iv) The spaces L1(νΦ,µ) and L1(νΦ,µ) coincide.
Proof. (i)⇒(iv) follows from [12, Theorem 1]. Conversely, (iv) implies that νΦ,µ is
finite as the constant function 1 ∈ L1(νΦ,µ). Hence, by Theorem 4.14, (i) follows.
(i)⇒(ii). By Theorem 4.14 there exists f ∈ LΦ(G) such that dµ = f dmG. Further,
by [3, Pg. 46, Theorem 4] and Corollary 4.2, it follows that the variation of νΦ,µ is
kfkΦmG. Hence (ii).
(ii)⇒(iii) follows from [13, Pg. 116, Lemma 3.14] and Theorem 4.13. We now prove
(iii)⇒(i). Let f ∈ L1(G). By our assumption, f ∈ L1(νΦ,µ) and hence, by Corollary
4.12, µ ∗ f ∈ LΦ(G). Note that, by [13, Pg. 152],
sup
f ∈BL1(G)kµ ∗ fkΦ =
sup
f ∈BL1(νΦ,µ)kIνΦ,µ(f )k = kIνΦ,µk = 1.
Now the conclusion follows from Lemma 2.1 and Theorem 4.14.
(cid:3)
As our final result, we show that the optimal extension is genuine.
Corollary 4.16. Let G be an infinite compact group and let µ ∈ M(G) be such that
dµ = f dmG for some f ∈ LΦ(G). Then the optimal extension of TΦ,µ is proper, i.e.,
LΦ(G) is a proper subspace of L1(νΦ,µ).
Proof. By Theorem 4.10, LΦ(G) ⊆ L1(νΦ,µ). Further, our assumption implies, by
Theorem 4.14 and Corollary 4.15, that LΦ(G) ⊆ L1(G). As Φ and Ψ satisfy the ∆2
condition, LΦ(G) is reflexive and as G is infinite it follows that LΦ(G) is a proper
subspace of L1(G) and hence a proper subspace of L1(νΦ,µ).
(cid:3)
Acknowledgement
The first author would like to thank the University Grants Commission, India,
for providing the research grant.
References
[1] G. P. Curbera and W. J. Ricker, Optimal domains for kernel operators via interpolation,
Math. Nachr., 244 (2002) 4763.
[2] G. P. Curbera and W. J. Ricker, Optimal domains for the kernel operator associated
with Sobolevs inequality, Studia Math., 158 (2003) 131152.
[3] J. Diestel and J. J. Uhl, Vector Measures, Math. Surveys, vol. 15. Amer. Math. Soc.,
Providence (1977).
OPTIMAL EXTENSION OF THE FOURIER TRANSFORM AND CONVOLUTION OPERATOR15
[4] N. Dinculeanu, Vector measures, VEB Deutscher Verlag der Wissenschaften, Berlin
(1966).
[5] D. Edmunds, R. Kerman and L. Pick, Optimal Sobolev imbeddings
involving
rearrangement-invariant quasinorms, J. Funct. Anal., 170 (2000) 307355.
[6] G. B. Folland, A Course in Abstract Harmonic Analysis, CRC Press, Boca Raton, 1995.
[7] E. Hewitt and K. A. Ross, Abstract harmonic analysis. Vol II: Structure and Analy-
sis for Compact Groups Analysis on Locally Compact Abelian Groups, Grundlehren der
Math.Wissenschaften, 152, Springer, 1970.
[8] R. Kerman and L. Pick, Optimal Sobolev imbeddings, Forum Math., 18 (2006) 535570.
[9] K. Musial, The weak Radon-Nikodym property in Banach spaces, Studia Math., 64 (1979)
151-174.
[10] S. Okada and W. J. Ricker, Optimal domains and integral representations of convolution
operators in Lp(G), Integr. Equ. Oper. Theory, 48 (2004), 525-526.
[11] S. Okada and W. J. Ricker, Optimal domains and integral representations of Lp(G)-
valued convolution operators via measures, Math. Nachr., 280 (2007), 423-436.
[12] S. Okada, W. J. Ricker and L. Rodr´ıguez-Piazza, Compactness of the integration
operator associated with a vector measure, Studia Math., 150 (2002) 133-149.
[13] S. Okada, W. J. Ricker, and E. A. S´anchez P´erez, Optimal Domain and Integral
Extension of Operators acting in Function Spaces, Oper. Theory Adv. Appl., vol. 180.
Birkhauser, Basel (2008).
[14] M. M. Rao and Z. D. Ren, Theory of Orlicz spaces, Dekker, New York, 1991.
[15] M. M. Rao and Z. D. Ren, Applications of Orlicz spaces, Dekker, New York, 2002.
[16] A. C. Zaanen Riesz spaces, Volume II, North Holland, 1983.
Department of Mathematics
Indian Institute of Technology Delhi
Delhi - 110016
India
E-mail address: [email protected]
Department of Mathematics
Indian Institute of Technology Delhi
Delhi - 110016
India
E-mail address: [email protected]
|
1503.06112 | 1 | 1503 | 2015-03-20T15:31:34 | Norm-attaining functionals and proximinal subspaces | [
"math.FA"
] | G. Godefroy asked whether, on any Banach space, the set of norm-attaining functionals contains a 2-dimensional linear subspace. We prove that a recent construction due to C.J. Read provides an example of a space which does not have this property. This is done through a study of the relation between the following two sentences where X is a Banach space and Y is a closed subspace of finite codimension in X: (A) Y is proximinal in X. (B) The annihilator of Y consists of norm-attaining functionals. We prove that these are equivalent if X is the Read's space. Moreover, we prove that any non-reflexive Banach space X with any given closed subspace Y of finite codimension at least 2 admits an equivalent norm such that (B) is true and (A) is false. | math.FA | math |
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL
SUBSPACES
MARTIN RMOUTIL
Abstract. G. Godefroy asked whether, on any Banach space, the set of norm-
attaining functionals contains a 2-dimensional linear subspace. We prove that
a recent construction due to C.J. Read provides an example of a space which
does not have this property.
This is done through a study of the relation between the following two
sentences where X is a Banach space and Y is a closed subspace of finite
codimension in X:
(A) Y is proximinal in X.
(B) Y ⊥ consists of norm-attaining functionals.
We prove that these are equivalent if X is the Read's space.
Moreover, we prove that any non-reflexive Banach space X with any given
closed subspace Y of finite codimension at least 2 admits an equivalent norm
such that implication (B) =⇒ (A) fails.
1. Introduction
Let X = (X, k · k) be a normed linear space. Then by X ∗ we denote the topo-
logical dual of X and by NA(X) = NA(X, k · k) the set of all functionals in X ∗ for
which there exists x ∈ SX such that x∗(x) = kxk; its elements are norm-attaining
functionals.
A set F ⊂ X is said to be proximinal if for each x ∈ X there is y ∈ F with
kx − yk = dist(x, F ). Note that proximinal sets are closed.
In this paper we investigate the relation of the following two sentences in which
X is a Banach space and Y ⊂ X is a closed subspace:
(A) Y is proximinal in X.
(B) Y ⊥ := {x∗ ∈ X ∗; Y ⊂ Ker(x∗)} ⊂ NA(X).
In view of the easy observation that a functional x∗ ∈ X ∗ is norm-attaining if
and only if Ker(x∗) is proximinal (see e.g. Corollary 3.4), it is reasonable to ask
whether sentences (A) and (B) are also equivalent when the codimension of Y in
X is higher than 1. And, indeed, implication (A) =⇒ (B) is true whenever the
quotient space X/Y is reflexive (cf. Lemma 3.3). In particular, it holds when the
codimension of Y in X is finite (and therefore dim(X/Y ) < ∞) which is the case
we are interested in the most.
Date: March 6, 2018.
2010 Mathematics Subject Classification. Primary 46B10, 46B20; Secondary 46B03.
Key words and phrases. proximinal subspace, norm-attaining functional, lineability, Banach
space, non-reflexive space, equivalent norm.
The research leading to these results has received funding from the Grant GA CR P201/12/0290,
and the European Research Council under the European Union's Seventh Framework Programme
(FP/2007-2013) / ERC Grant Agreement n.291497.
1
2
MARTIN RMOUTIL
However, implication (B) =⇒ (A) is not necessarily true even if Y has finite
codimension. The validity of this implication (and related statements) for vari-
ous spaces has been studied by several authors (e.g. A.L. Garkavi, V. Indumathi
or G. Godefroy) and counterexamples are known: e.g. Indumathi [12, Section 4.]
shows that in any infinite-dimensional space L1(T, ν) (here (T, ν) is a positive mea-
sure space) there exists a closed non-proximinal subspace M of finite codimension
d ≥ 2 such that all subspaces N ) M are proximinal; in particular, L1(T, ν) fails
implication (B) =⇒ (A). In another work [13] of Indumathi, the author studies
the relation of (A) and (B) (and similar statements) in more detail and defines a
space to be R(1) if implication (B) =⇒ (A) is true for any closed subspace Y ⊂ X
of finite codimension (in general, a non-reflexive space is said to be R(n) space if
proximinality of its closed subspaces of finite codimension at least n is implied by
the proximinality of all n-codimensional subspaces containing it).
In Section 5 we show that, given any non-reflexive Banach space X and its closed
subspace Y of finite codimension at least 2, there is an equivalent norm · on X
such that Y is not proximinal in (X, ·), but Y ⊥ ⊂ NA(X, ·) (see Theorem 5.6).
In particular, any non-reflexive Banach space can be renormed so that it is not R(1).
The main motivation for the present work is the following Problem of G. Gode-
froy together with a recent theorem of C.J. Read below.
Problem 1.1 ([9, Problem III.], also e.g. [2] or [1]). Does the set NA(X) contain
a 2-dimensional subspace for any Banach space X?
The well-known theorem of E. Bishop and R.R. Phelps [4] states that for any
Banach space X, the set NA(X) is dense in X ∗; in particular, NA(X) is nonempty
and so it clearly contains 1-dimensional subspaces. Since then, a number of papers
studying the structure of NA(X) from various viewpoints have been written; the
ones which are interesting for us are concerned with the linear structure of NA(X),
in particular the question whether NA(X) is lineable or spaceable (for definition
see Section 2). For example, in [2] it is shown that if X is an Asplund space
enjoying the Dunford-Pettis property, then NA(X) is not spaceable (e.g. C(K)
with K scattered (Hausdorff) compactum; however, NA(C(K)) for K infinite is
always lineable-see [1]). On the other hand, in [1] the authors prove that if the
Banach space X possesses an infinite-dimensional complemented subspace with a
Schauder basis, then X can be equivalently renormed so that NA(X) is lineable.
Some articles also investigate the lineability of X ∗ \ NA(X), e.g. [1] or [7]. In [7]
it is observed that if J denotes the James space, then J ∗ \ NA(J ) is not even
2-lineable (i.e. (J ∗ \ NA(J )) ∪ {0} contains no 2-dimensional subspace).
A related problem (again [9, Problem III.]) was to decide whether every Banach
space contains a proximinal subspace of codimension 2. This problem was recently
solved by C.J. Read:
Theorem R (C.J. Read [20]). Let · be the renorming of c0 from Definition 1.
Then (c0, ·) contains no proximinal subspaces of finite codimension 2 or larger.
In Section 4 we show that this space also works as a counterexample for Prob-
lem 1.1. More precisely, our Theorem 4.2 states that NA(c0, ·) contains no
2-dimensional subspaces. This is done by showing that (c0, ·) is R(1) (i.e.
im-
plication (B) =⇒ (A) holds for it) and then applying Read's result. Among the
known sufficient conditions for a Banach space X to be R(1) are the following three:
• the dual space X ∗ is (Gateaux) smooth [13, Corollary 2];
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
3
• X is WLUR (for definition see Section 2.) [13, Corollary 3];
• X is a subspace of c0 [10].
None of these seems to be suitable in the case of (c0, ·), however; our proof has
to be more direct, although it is done similarly to our new proof of the sufficiency
of X being WLUR (see Corollary 3.6).
The essential (but simple) tool we use throughout the paper is a characterization
of conditions (A) and (B) in terms of the quotient space X/Y (see parts (i) and
(ii) of Lemma 3.3); for example, a closed subspace Y ⊂ X is proximinal if and only
if the quotient mapping q : X → X/Y maps BX onto the whole BX/Y (where BZ
denotes the closed unit ball in Z), and we have a similar condition characterizing
the fact that Y ⊥ ⊂ NA(X). It is not difficult to see that our characterization of
proximinality is equivalent to that of Garkavi [8] (see [13, Theorem A]), which is
not formulated using the quotient space. The characterization of Garkavi was used
in many subsequent articles on the topic, but it seems that our approach using
quotient spaces is, at least in some situations, more natural.
In Lemma 3.3 (iv) we use our characterizations of (A) and (B) to prove that
implication (B) =⇒ (A) holds whenever X/Y is strictly convex. Moreover, in
Proposition 3.5 we show that this is the case whenever the norm on X is WLUR
(thus obtaining a different proof of this known sufficient condition for R(1)). This
is not completely trivial since in Example 3.7 we have a strictly convex Banach
space X and its closed subspace Y of codimension 2 such that X/Y is not strictly
convex.
The proof of Theorem 4.2 (i.e. that Read's space solves Problem 1.1) is done
by contradiction: Setting X = (c0, ·) we assume that there is a closed subspace
Y ⊂ X of finite codimension at least 2 such that Y ⊥ ⊂ NA(X), and then we prove
that X/Y is strictly convex. It follows from Lemma 3.3 that Y is proximinal which
is a contradiction with Theorem R.
We also use quotient spaces to obtain our renorming of a non-reflexive space
X such that the resulting space contradicts implication (B) =⇒ (A). Indeed, in
Lemma 5.5 we take any closed subspace Y ⊂ X of finite codimension in X. Given
a set A ⊂ X/Y such that UX/Y ⊂ A ⊂ BX/Y satisfying some technical conditions,
we find a renorming of X such that q(BX ) = A. Then, using our Lemma 3.3 we
show in Theorem 5.6 that if the codimension of Y is at least 2, then A can be
chosen in such a way that the resulting renorming satisfies (B), but not (A).
2. Notation and Definitions
Let X = (X, k · k) be a normed linear space over the field of real numbers.
Then 0 is the origin in X (there is no risk of confusion), X ∗ is its (topological)
dual, BX its closed unit ball, UX is the open unit ball and SX denotes the unit
sphere. For a functional x∗ ∈ X ∗ we use the notation Ker(x∗) := (x∗)−1(0) and
att(x∗) := {x ∈ BX ; x∗(x) = kx∗k}. As explained in the Introduction, NA(X) =
{x∗ ∈ X ∗; att(x∗) 6= ∅}. If Y is a normed linear space, by X ∼= Y we mean that X
is linearly isomorphic to Y , and X ⊕∞Y is the direct sum of X and Y equipped with
the maximum norm, i.e. for (x, y) ∈ X ⊕∞ Y we have k(x, y)k = max{kxk, kyk}.
If Y is a subspace of X, we write Y ⊥ for its annihilator {x∗ ∈ X ∗; Y ⊂ Ker(x∗)}.
Recall that the quotient space X/Y consists of all cosets of the form x + Y and is
equipped with the norm kx + Y k = dist(0, x + Y ). If X is a Banach space and Y is
closed, then X/Y is a Banach space and its dimension is equal to the codimension of
4
MARTIN RMOUTIL
Y in X. We shall often work with the quotient mapping q : X → X/Y : x 7→ x + Y ,
which is a linear operator satisfying kqk = 1.
Next, if A ⊂ X is any set, then Int(A), A and ∂A stand for the interior, closure
w→ x we mean that xn converges to x
and boundary of A respectively. By xn
weakly in X, and similarly for w∗. If we do not specify which topology we mean,
it is usually the norm topology; e.g. xn → x means limn→∞ kxn − xk = 0.
We recall that the norm k · k on X is called strictly convex (or rotund ; SC for
short) if all the points of SX are extremal points of BX , that is, if there are no
nontrivial line segments contained in SX .
In that case X is also called strictly
convex.
xn
limn→∞ kx + xnk = 2kxk. Then the space X is called WLUR.
The norm k · k is said to be weakly locally uniformly rotund (WLUR for short) if
w→ x whenever x ∈ X and (xn)∞
n=1 ⊂ X are such that limn→∞ kxnk = kxk and
For any sequence a ∈ ℓ∞ we denote a = (cid:0)aj(cid:1)∞
Let A be a bounded set in a Banach space. We denote by convσ(A) the σ-convex
j=1 = (cid:0)a1, a2, . . .(cid:1).
hull of A, that is
convσ(A) = n
∞
X
i=1
λixi; (λi)∞
i=1 ⊂ [0, 1],
∞
X
i=1
λi = 1 and (xi)∞
i=1 ⊂ Ao.
A subset M of a Banach space is said to be:
• n-lineable if M ∪ {0} contains an n-dimensional subspace;
• lineable if M ∪ {0} contains an infinite-dimensional subspace;
• spaceable if M ∪ {0} contains an infinite-dimensional closed subspace.
A comprehensive survey on linear subspaces of various sets is [3].
3. Observations and lemmas
The following observations are obvious and will be applied without reference.
Observations 3.1. Let X be a normed linear space.
(i) Any proximinal set in X is closed.
(ii) Let F be a proximinal set in X and let x ∈ X. Then x + F is also proximinal
in X.
(iii) A subspace Y ⊂ X is proximinal if and only if for any x ∈ X there is a point
z with minimal norm in the coset x + Y .
The next facts are well-known; we include them for the sake of completeness.
Fact 3.2.
n ∈ X ∗.
1, . . . , x∗
i ) ⊂ Ker(y∗).
Then
(i) Let X be a normed linear space and y∗, x∗
i=1 Ker(x∗
y∗ ∈ span{x∗
n} if and only if Tn
1, . . . , x∗
(ii) Let Z be any normed linear space. Then Z is not strictly convex if and only
if there is a functional ϕ ∈ Z ∗ such that att(ϕ) contains more than one point.
(iii) Let X be a Banach space, Y ⊂ X be a closed subspace and q : X → X/Y the
quotient mapping. Then q∗ : (X/Y )∗ → Y ⊥ is an isometry onto Y ⊥.
Proof. (i) and (iii) are standard; for proof of (iii) see e.g. [6, Proposition 2.6].
(ii): Assume that Z is not strictly convex. Take two distinct points u, v ∈ SZ
such that for any λ ∈ (0, 1), λu + (1 − λ)v ∈ SZ , and find (using the Hahn–Banach
theorem) a functional ϕ ∈ SZ ∗ such that ϕ( u+v
2 ) = 1. Clearly, if ϕ(u) < 1, then
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
5
ϕ(v) > 1, and vice versa. But kϕk = 1, whence ϕ(u) = ϕ(v) = 1. The opposite
implication is trivial.
(cid:3)
The following lemma translates properties (A) and (B) to the language of quo-
tient spaces, and is essential in the sequel.
Lemma 3.3. Let X be a Banach space, Y ⊂ X be a closed subspace and q : X →
X/Y be the quotient mapping. Then:
(i) Y is proximinal in X ⇐⇒ q(BX ) = BX/Y .
(ii) Y ⊥ ⊂ NA(X) ⇐⇒ (∀ϕ ∈ (X/Y )∗) (∃u ∈ q(BX )) ϕ(u) = kϕk.
(iii) If X/Y is reflexive and Y is proximinal, then Y ⊥ ⊂ NA(X).
(iv) If X/Y is strictly convex and Y ⊥ ⊂ NA(X), then Y is proximinal.
Proof. (i): Assume Y is proximinal. Obviously q(BX ) ⊂ BX/Y ; to prove the
other inclusion, take any u ∈ BX/Y . Further, pick z ∈ q−1(u). Then dist(0, z +
Y ) = kuk ≤ 1, and by Observation 3.1 (iii) there is a point y ∈ z + Y such that
kyk = dist(0, z + Y ), so q(y) = u and y ∈ BX .
To prove the opposite implication, assume that q(BX ) = BX/Y and take an
arbitrary x ∈ X. By homogeneity we can clearly assume that dist(0, x + Y ) = 1,
so q(x) ∈ BX/Y , and our assumption now implies that there exists y ∈ BX such
that q(y) = q(x) (which is equivalent to y ∈ x + Y ). By Observation 3.1 (iii) we
are done.
(ii): Let Y ⊥ ⊂ NA(X) and take an arbitrary ϕ ∈ (X/Y )∗. Then y∗ := ϕ ◦ q ∈
Y ⊥, and so there exists x ∈ BX such that ϕ(q(x)) = y∗(x) = ky∗k = kϕk (by
Fact 3.2 (iii)). Hence q(x) ∈ q(BX ) ∩ att(ϕ).
Conversely, take any y∗ ∈ Y ⊥ and the corresponding ϕ ∈ (X/Y )∗ such that
y∗ = ϕ ◦ q. By the assumption there exists u ∈ q(BX ) ∩ att(ϕ). So there also exists
x ∈ BX ∩ q−1(u). Then y∗(x) = ϕ(q(x)) = ϕ(u) = kϕk = ky∗k, and y∗ is therefore
norm-attaining.
(iii): Trivial from (i), (ii) and the fact that on a reflexive space, all functionals
are norm-attaining.
(iv): Assume the condition in (ii) is satisfied. Clearly UX/Y ⊂ q(BX ) ⊂ BX/Y ,
so by (i) it suffices to show that SX/Y ⊂ q(BX ). To that end, take an arbitrary
u ∈ SX/Y and use the Hahn-Banach theorem to find ϕ ∈ S(X/Y )∗ such that u ∈
att(ϕ). Since X/Y is strictly convex, it follows by Fact 3.2 that att(ϕ) = {u}, and
the condition in (ii) now implies that u ∈ q(BX ).
(cid:3)
Corollary 3.4. Let X be a Banach space. Then:
(a) Let Y ⊂ X be a closed subspace of finite codimension in X. If Y is proximinal,
then Y ⊥ ⊂ NA(X).
(b) A functional x∗ ∈ X ∗ is norm-attaining if and only if Ker(x∗) is proximinal.
Proof. (a): The quotient space X/Y is finite-dimensional and in particular it is
reflexive.
(b): In addition to being reflexive, the quotient space X/ Ker(x∗) is (for trivial
(cid:3)
reasons) also strictly convex, and the statement follows.
Proposition 3.5. Let X be a Banach space with a WLUR norm, and Y ⊂ X be
its closed subspace such that Y ⊥ ⊂ NA(X). Then X/Y is strictly convex.
Proof. Aiming for a contradiction, suppose that X is WLUR and X/Y is not strictly
convex. Then we can find a functional ϕ ∈ (X/Y )∗ such that kϕk = 1 and att(ϕ)
6
MARTIN RMOUTIL
contains more than one point (Fact 3.2 (ii)). Now, set x∗ = ϕ◦q where q : X → X/Y
is the quotient mapping. Fact 3.2 (iii) gives us that kx∗k = kϕk = 1 and x∗ ∈ Y ⊥.
As Y ⊥ ⊂ NA(X), x∗ attains its norm at a point x ∈ SX and ϕ(q(x)) = x∗(x) = 1,
whence q(x) ∈ att(ϕ).
Take any u ∈ att(ϕ) \ {q(x)} and any sequence of elements xn ∈ u such that
limn→∞ kxnk = 1 = kxk. Then limn→∞ kx + xnk = 2 since
2 = ϕ(q(xn) + q(x)) = x∗(xn + x) ≤ kxn + xk ≤ kxnk + kxk → 2, n → ∞.
Our assumption that X is WLUR yields that xn converges to x weakly in X.
To obtain the desired contradiction, take any ψ ∈ (X/Y )∗ such that ψ(q(x)) 6=
ψ(u) and set y∗ := ψ ◦ q ∈ X ∗. Then for all n ∈ N we have y∗(x) = ψ(q(x)) 6=
ψ(u) = y∗(xn), so y∗(xn) does not converge to y∗(x) which is a contradiction. (cid:3)
Let us note that these observations provide a new proof for the following result
of Indumathi (see [13, Corollary 3]).
Corollary 3.6. Let X be a Banach space with a WLUR norm and Y ⊂ X be its
closed subspace of finite codimension. Then Y is proximinal if and only if Y ⊥ ⊂
NA(X).
Proof. The statement follows immediately from Proposition 3.5, Lemma 3.3 and
Corollary 3.4.
(cid:3)
The following is a simple example of a strictly convex space X with a finite-
codimensional closed subspace Y such that X/Y is not strictly convex.
Example 3.7. By [0, ω] we mean the compact interval of ordinals with the order
topology. Let us define the Banach space X as C([0, ω] × {0, 1}) with the norm k · k
defined by
∞
kxk2 = kxk2
∞ +
X
k=0
2−k(cid:0)x(k, 0)2 + x(k, 1)2(cid:1).
It is easy to check that k · k is indeed a norm on C([0, ω] × {0, 1}), which is moreover
equivalent to k · k∞. It is also clear that the dual to (C([0, ω] × {0, 1}), k · k∞) is
(ℓ1([0, ω] × {0, 1}), k · k1), so X ∗ is isomorphic to the latter space.
Set Y := {x ∈ X; x(ω, 0) = x(ω, 1) = 0}; then Y is a closed subspace of
codimension 2 and Y ⊥ = span{δ(ω,0), δ(ω,1)} and Y . We claim that X is strictly
convex, but X/Y is not.
Let x, y ∈ X be arbitrary points satisfying
2kxk2 + 2kyk2 − kx + yk2 = 0;
(1)
[6]) it
by a well-known (and easy) characterization of strict convexity (see e.g.
suffices to prove that x = y. One can rewrite equation (1) using the definition of
k · k, and observe that the left hand side is the sum of the term
2kxk2
∞ + 2kyk2
∞ − kx + yk2
∞,
and all terms of the form
2−k(cid:0)2x(k, i)2 + 2y(k, i)2 − (x(k, i) + y(k, i))2(cid:1),
(k, i) ∈ [0, ω) × {0, 1}
which are easily seen to be non-negative. Since the sum of all these terms is equal
to 0, it follows that so is each of the terms. In particular, x(k, i) = y(k, i) for all
(k, i) ∈ [0, ω) × {0, 1}, and the continuity of x and y now implies that x = y.
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
7
We shall now prove that X/Y is isometrically isomorphic to (R2, k · k∞); in
particular, X/Y is not strictly convex. Consider the quotient mapping q : X →
X/Y : x 7→ x + Y = q(x). Obviously z ∈ q(x) if and only if z(ω, i) = x(ω, i),
i = 0, 1, whence the mapping L : X/Y → R2 : x + Y 7→ (x(ω, 0), x(ω, 1)) is a linear
bijection. It remains to show that L : X/Y → (cid:0)R2, k · k∞(cid:1) is an isometry.
Let x ∈ X be given; by definition,
kx + Y k = dist(0, x + Y ) ≥ max{x(ω, 0), x(ω, 1)}.
To prove the opposite inequality, set for all N ∈ N,
xN := x · χ[N,ω]×{0,1},
where χA denotes the characteristic function of A ⊂ [0, ω] × {0, 1}; then xN ∈ x + Y
as xN (ω, i) = x(ω, i), i = 0, 1. Now,
kxN k2 = kxN k2
∞ +
∞
X
k=N
2−k(cid:0)x(k, 0)2 + x(k, 1)2(cid:1),
and it follows from the continuity of x that
lim
N→∞
kxN k = max{x(ω, 0), x(ω, 1)};
hence, dist(0, x + Y ) ≤ max{x(ω, 0), x(ω, 1)}.
In view of the last example and Proposition 3.5 it is natural to ask the following.
Problem 3.8. Can the assumption that X be WLUR in Proposition 3.5 be weak-
ened to SC? Equivalently, can Example 3.7 be done in such a way that we moreover
have Y ⊥ ⊂ NA(X)?
Remark 3.9. The following classical results are related to the problem above. Let
X be a Banach space.
(a) Let X be SC and Y be a reflexive subspace of X. Then X/Y is SC (V. Klee,
[17]).
(b) The following are equivalent (see e.g. M. Day, [5, Chapter VII, §2]):
(i) X ∗ is Gateaux smooth (equivalently, at each point of SX ∗ there is only
one supporting hyperplane of BX ∗);
(ii) every quotient of X is SC;
(iii) every 2-dimensional quotient of X is SC.
It is interesting to note that having a strictly convex norm on X/Y is not the
only known condition which ensures that the implication (B) =⇒ (A) holds. The
following theorem can be found in [10] or [9]:
Theorem GI (G. Godefroy, V. Indumathi). Let X be a closed subspace of c0 and
Y ⊂ X be a closed finite-codimensional subspace of X. Then Y is proximinal in X
if and only if Y ⊥ ⊂ NA(X).
Instead of showing that X/Y is strictly convex (which is not the case), the authors
essentially prove that X/Y is polyhedral (i.e. the unit ball is the convex hull of
finitely many points), and then show that each of the extremal points is contained
in q(BX ) where q : X → X/Y is the quotient mapping (in the same manner as
we show it for each point of the sphere SX/Y ); it then follows from the convexity
of q(BX ) that q(BX ) = BX/Y , and therefore Y is proximinal. (In their proof, our
8
MARTIN RMOUTIL
Lemma 3.3 (i) is replaced by an essentially equivalent theorem of Garkavi, which
is not formulated in terms of quotient spaces.)
The next proposition can be regarded as a generalization of Corollary 3.4 (a).
This can be seen if we reformulate the corollary as follows:
Let X be a Banach space and x∗ ∈ X ∗. Then the hyperplane Ker(x∗) is proxim-
inal whenever it contains a finite-codimensional proximinal subspace Y .
Proposition 3.10. Let X be a normed linear space and let Y ⊂ Z ⊂ X be linear
subspaces such that the codimension of Y in Z is finite. If Y is proximinal (in X),
then Z is also proximinal.
Proof. First assume that the codimension of Y in Z is 1. Fix a functional x∗ ∈ SX
such that Y = Z ∩ Ker(x∗) and any vector v ∈ Z \ Y ; then it is easy to see that
Z = [
c∈R
(cv + Y ).
Now, pick an arbitrary x ∈ X; we aim to find a point y ∈ x + Z with kyk =
dist(0, x + Z). To that end, let us consider the function ϕ : R → R defined at c ∈ R
by the formula
ϕ(c) := dist(0, x + cv + Y ).
Assume, for the moment, that we have found a global minimum of ϕ at a point
c; then we can use the proximinality of Y to find a point y ∈ x + cv + Y with
minimal norm. Clearly y has the minimal norm also in x + Z.
Since ϕ is easily seen to be continuous (even kvk-Lipschitz), it is enough to show
that there is a compact set I ⊂ R with ϕ(R \ I) ⊂ (ϕ(0), ∞). We claim that it
suffices to set
I := {c ∈ R; x∗(x + cv) ≤ kxk}.
Indeed, if c ∈ R \ I, then x∗(x + cv) > kxk. Since Y ⊂ Ker(x∗), the functional
x∗ ∈ SX ∗ is constant on any coset of the form z + Y , z ∈ X. It follows that
ϕ(c) = inf{kx + cv + yk; y ∈ Y }
≥ inf{x∗(x + cv + y); y ∈ Y } = x∗(x + cv) > kxk ≥ ϕ(0);
In particular, 0 ∈ I. The boundedness of I follows from the fact that v /∈ Ker(x∗).
Now assume that we have proved the assertion of the Proposition in case the
codimension of Y in Z is equal to n ∈ N, and suppose we have closed subspaces Y
and Z of X such that Y ⊂ Z with codimension n+1. This means that there are some
x∗
1, . . . , x∗
i ).
Then the codimension of Y in Z1 is 1, so Z1 is proximinal, and the codimension of
Z1 in Z is n, so Z is proximinal.
(cid:3)
n+1 ∈ X ∗ such that Y = Z ∩ Tn+1
i ). Set Z1 := Z ∩ Tn
i=1 Ker(x∗
i=1 Ker(x∗
Remark 3.11. Following W. Pollul [19], by Z ⊂p Y we mean that Z is a proxim-
inal subspace of the normed linear space Y ; it is then natural to ask under what
conditions is the relation ⊂p transitive (in which case the ambient space is said to be
a P -space). Indumathi [13, Corollary 5] gives a characterization: A normed linear
space X is a P -space if and only if it is an R(1) space and NA(X) is orthogonally
linear (for definition see the same article). In particular, X is a P -space if it is an
R(1) space and NA(X) is a vector space.
Another related question is whether proximinality of subspaces is preserved un-
der intersections. This is not always the case either; if a Banach space X has the
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
9
property that the intersection of arbitrary two finite-codimensional proximinal sub-
spaces is proximinal, X is said to be lattice-proximinal. A simple characterization
was given by Godefroy and Indumathi in [11]: X is lattice-proximinal if and only
if it is R(1) and NA(X) is a vector space. An example of such a space is c0; we
already know that it is R(1) (Theorem GI), and it is easy to see that norm-attaining
functionals on c0 are exactly finitely supported elements of ℓ1.
4. NA(X) need not be 2-lineable
The following renorming of c0 is due to C.J. Read [20] and it is the first Banach
space known to have no proximinal subspaces of finite codimension strictly larger
than 1 (Theorem R). In this section we will show that the set of norm-attaining
functionals on this space is not 2-lineable, solving Problem 1.1.
Definition 1 (C.J. Read). Let c00(Q) be the terminating sequences with rational
coefficients, and let (un)∞
n=1 be a sequence of elements of c00(Q) which lists every
element infinitely many times.
Further, let (an)∞
fying for each n ∈ N,
n=1 be a strictly increasing sequence of positive integers satis-
an > max supp un,
and an ≥ kunk1.
The norm · on c0 is defined as follows:
x = kxk∞ +
∞
X
n=1
2−a2
nhx, un − eani.
Here kxk∞ = supn xn is the usual norm on c0, ej's are the vectors from the usual
cannonical basis, and the duality h·, ·i is the hc0, l1i duality.
Remark 4.1. Proposition 3.10 implies that if a normed linear space X has no
proximinal subspace of codimension 2, then it has no proximinal subspace of any
finite codimension larger than 2 either. Hence, in the proof of C.J. Read it is enough
to show that the space (c0, ·) constructed there has no proximinal subspace of
codimension 2 (he proves it at once for all codimensions, slightly complicating the
proof).
The following theorem shows that the construction of Read does, in fact, answer
Problem 1.1.
Theorem 4.2. Let X be the space (c0, ·) from Definition 1. Then NA(X) con-
tains no linear subspaces of dimension larger than one.
Proof. We shall denote by · the norm on ℓ1 which is dual to · on c0. Assume
that Y $ X is a closed subspace of finite codimension and such that Y ⊥ ⊂ NA(X);
we aim to prove that the codimension of Y in X is equal to 1.
It follows from
Theorem R and Lemma 3.3 that it is enough to prove the strict convexity of X/Y ;
we shall do so by contradiction.
To that end, we can (by Fact 3.2) assume that we have a functional ϕ ∈ S(X/Y )∗
such that att(ϕ) contains more than one point; set x∗ := ϕ ◦ q. Then x∗ ∈ Y ⊥ ⊂
NA(X), and so there is x ∈ att(x∗). Fact 3.2 gives that x∗ = kϕk = 1 and
q(x) ∈ att(ϕ).
Let us take any point u ∈ att(ϕ)\{q(x)} and choose any sequence (xn)∞
X such that limn→∞ xn = 1 (recall that kuk = 1). Then (xn)∞
n=1 ⊂ u ⊂
n=1 is a bounded
10
MARTIN RMOUTIL
sequence and we can without loss of generality assume that there is x ∈ ℓ∞ such
that xn
Now,
w∗
→ x in ℓ∞.
2 = ϕ(q(x) + q(xn)) = x∗(x + xn) ≤ x + xn ≤ x + xn n→∞−→ 2,
whence
x + xn − x + xn −→ 0.
Since the norm · is defined by a series of pseudonorms (which satisfy the triangle
inequality), we in turn obtain for each k ∈ N:
hx, uk − eak i + hxn, uk − eak i − hx + xn, uk − eak i n→∞−→ 0,
so, since (xn) converges to x in weak∗ topology, we have that
hx, uk − eak i + hx, uk − eak i − hx + x, uk − eak i = 0.
(2)
Claim: There is λ ≥ 0 such that x = λx. In particular, x ∈ X and xn
Assume this is not the case. Then there is z∗ ∈ c00(Q) ⊂ (ℓ1, ·) = X ∗ such
w→ x.
that z∗(x) < 0 < x(z∗), z∗(x) > kxk∞ and x(z∗) > kxk∞.
Indeed, by the Hahn-Banach theorem we can find z∗
x and the weak∗-closed convex set [0, ∞) · x, and z∗
weak∗-closed convex set [0, ∞) · x with
1 ∈ ℓ1 separating the point
2 ∈ ℓ1 separating x and the
z∗
1 (x) < 0 ≤ x(z∗
1 ),
and z∗
2(x) ≤ 0 < x(z∗
2).
Setting z∗
3 := z∗
1 + z∗
2, one can readily see that
z∗
3(x) < 0 < x(z∗
3 ).
Using the density of c00(Q) in ℓ1 we find a functional z∗ ∈ c00(Q) ⊂ ℓ1 so close to z∗
3
that, again, z∗(x) < 0 < x(z∗). Without loss of generality we can now assume that
z∗(x) > kxk∞ and x(z∗) > kxk∞ because in the opposite case we would simply
multiply z∗ by a sufficiently large rational constant.
Now, take k ∈ N such that uk = z∗ (see Definition 1); Then
hx, uk − eak i = z∗(x) − xak ≤ z∗(x) + kxk∞ < 0
hx, uk − eak i = x(z∗) − xak ≥ x(z∗) − kxk∞ > 0.
and
But on the other hand it follows from (2) that for each natural k we have
hx, uk − eak i · hx, uk − eak i ≥ 0,
a contradiction which proves the Claim.
Finally, find a functional ψ ∈ (X/Y )∗ such that ψ(q(x)) < 0 < ψ(u) (we can
3 above in this proof). But then, setting
do so e.g. using the same idea as with z∗
y∗ := ψ ◦ q, we have that
y∗(x) = λy∗(x) = λψ(q(x)) ≤ 0 < ψ(u) = y∗(xn)
which contradicts the fact that xn
w→ x, concluding the proof.
(cid:3)
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
11
5. Non-R(1) renormings
The main result of the present section is Theorem 5.6, which shows that the
implication (B) =⇒ (A) (see the Introduction) can fail in all non-reflexive spaces
if we take a suitable renorming; the purpose of all the preceding lemmas is to prove
this theorem. At the end of this section we also prove a slightly stronger version of
the key Lemma 5.5 which holds for the space ℓ1.
Lemma 5.1. Let x1, x2, . . . , xn ∈ Rd be such that C := conv{x1, x2, . . . , xn} has
nonempty interior. Let x = Pn
i=1 λi = 1 and λi > 0 for each i.
Then x ∈ Int(C).
i=1 λixi with Pn
Proof. Assume that x ∈ ∂C and take any non-zero functional ϕ ∈ (Rd)∗ such
that supC ϕ = ϕ(x) (it exists due to the Hahn-Banach Theorem). The hyperplane
H := ϕ−1(ϕ(x)) is nowhere dense, so there is a j ∈ {1, . . . , n} such that xj ∈ C \ H
(otherwise C ⊂ H which is impossible). But λj > 0, so λj ϕ(xj ) < λjϕ(x), whence
ϕ(x) = ϕ(cid:16)
n
X
i=1
λixi(cid:17) = X
i6=j
λiϕ(xi) + λj ϕ(xj ) < X
i6=j
λiϕ(x) + λj ϕ(x) = ϕ(x),
which is a contradiction.
(cid:3)
Lemma 5.2. If C is a convex set in a normed linear space X, y ∈ Int(C) and
z ∈ C, then λy + (1 − λ)z ∈ Int(C) for any λ ∈ (0, 1).
Proof. If C, y, z and λ are as above, we can find r > 0 such that U (y, r) ⊂ C and
a point x ∈ C ∩ U (z, λr
1−λ ). Then the ball U := λU (y, r) + (1 − λ)x = U (λy +
(1 − λ)x, λr) is contained in C as it consists of convex combinations of points of
U (y, r) and x. But kλy + (1 − λ)z − (λy + (1 − λ)x)k = (1 − λ)kz − xk < λr, and
so λy + (1 − λ)z ∈ U .
(cid:3)
Proposition 5.3. Let A ⊂ Rd be bounded. Then convσ(A) = conv(A).
Proof. Take any σ-convex combination x = P∞
i=1 λixi of elements of A. If there
are only finitely many non-zero λi's, then x ∈ conv(A) trivially. So we can clearly
assume that λi 6= 0 for every i, and we aim to prove that x ∈ conv{xi; i ∈ N} ⊂
conv(A).
Now, any convex set in Rd is either contained in a hyperplane, or has a nonempty
interior. Thus, setting C := conv{xi; i ∈ N}, we can assume that Int(C) 6= ∅. Take
N ∈ N so large that the interior of CN := conv{xi; i = 1, . . . , N } is nonempty and
set λ := PN
i=1 λi; then by Lemma 5.1, y := 1
i=1 λixi ∈ Int(CN ) ⊂ Int(C).
λ PN
The point z := 1
1−λ P∞
λ)z, the result follows from Lemma 5.2.
i=N +1 λixi is clearly an element of C. Since x = λy + (1 −
(cid:3)
Lemma 5.4. Let X be a Banach space and let for each j, k ∈ N, f j
k=1 ⊂ ℓ1 and (bk)∞
Let (ak)∞
limk→∞ kak − bkk1 = 0. Then
k ∈ BX .
k=1 ⊂ ℓ1 be two sequences of elements of ℓ1 such that
lim
k→∞
∞
X
j=1
aj
kf j
k = lim
k→∞
∞
X
j=1
bj
kf j
k
whenever one of the limits exists.
12
MARTIN RMOUTIL
Proof. Since X is a Banach space, all the series in the statement, as well as the
estimate below, converge because they converge absolutely. Hence, the following
suffices to conclude the proof:
(cid:13)(cid:13)(cid:13)
X
∞
j=1
≤
∞
X
j=1
aj
kf j
k −
∞
X
j=1
k − bj
(cid:12)(cid:12)aj
k(cid:12)(cid:12) · (cid:13)(cid:13)f j
∞
j=1
(cid:0)aj
X
bj
kf j
= (cid:13)(cid:13)(cid:13)
k(cid:13)(cid:13)(cid:13)
k − bj
(cid:12)(cid:12)aj
X
k(cid:13)(cid:13) ≤
j=1
∞
k − bj
k(cid:13)(cid:13)(cid:13)
k(cid:1)f j
k(cid:12)(cid:12) = kak − bkk1 −→ 0.
(cid:3)
The following lemma is the core result of the present section showing how to
construct certain renormings of non-reflexive Banach spaces. The proof is rela-
tively complicated from the technical point of view, but except for its use of the
James's theorem [14] (stating that on a non-reflexive Banach space there always is
a functional which does not attain its norm on the closed unit ball) it is elementary.
The proof is divided into 20 paragraphs to make it easier to read and also to relate
it to the proof of Proposition 5.7 which follows the same scheme. The main idea of
the proof appears already in the second paragraph where we define the renorming
by choosing a new unit ball for the space. The rest of the proof is to show that
the image of this new unit ball via the quotient mapping is exactly what we want
it to be (i.e. the set A). The difficult part is to show that the image contains no
extra points, and that follows from the choice of renorming which uses the non-
norm-attaining functional and the subsequent existence of a "very non-convergent"
bounded sequence of points (a norming sequence for the functional contained in the
unit sphere).
Lemma 5.5. Let X be a non-reflexive Banach space, d ∈ N, Y ⊂ X be a closed
subspace of codimension d in X, and q : X → X/Y be the quotient mapping. Let
k · k be an equivalent norm on X/Y and let A ⊂ X/Y be a symmetric convex Fσ-set
satisfying U(X/Y,k·k) ⊂ A ⊂ B(X/Y,k·k) for which there exists v ∈ ∂A and r ∈ (0, 1)
such that A \ (cid:0)Uk·k(v, r) ∪ Uk·k(−v, r)(cid:1) is closed.
Then X admits an equivalent norm · such that q(B(X,·)) = A.
Proof. §1: Since Y is a closed subspace of finite codimension d in X, it is comple-
mented in X, its (topological) complement is isomorphic to Rd as well as to X/Y ,
and X ∼= Y ⊕∞ X/Y (see e.g.
[6, Section 4.1]). It follows that X ∼= Y ⊕∞ Rd.
Note also that Y is non-reflexive as in the opposite case X would be reflexive, be-
ing isomorphic to the direct sum of two reflexive spaces, which is easily seen to be
reflexive.
Our goal is to find a certain equivalent norm on X; we can therefore assume
that X = Y ⊕∞ Rd and consider the projection to the second coordinate q : X →
Rd instead of the quotient mapping X → X/Y . This is, indeed, correct as the
topological and linear structure remains the same, so we can also assume without
loss of generality that A ⊂ Rd and that k · k is a norm on Rd (and it is the one
satisfying B(Rd,k·k) = A). To avoid any confusion, let us emphasize that we shall
use no other norm on Rd in this proof.
§2: Now, denote F0 := conv(cid:0)A \ (U (v, r) ∪ U (−v, r))(cid:1); then F0 is a compact
subset of A. Clearly, the set A ∩ U (v, r) is of the type Fσ, so there are closed sets
Ej, j ≥ 1, such that A ∩ U (v, r) = S∞
j=1 Ej. From the convexity of A ∩ U (v, r) it
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
13
follows that A ∩ U (v, r) = S∞
each j.
j=1 Fj where Fj := conv(Ej ); then Fj is also closed for
To define an equivalent norm with the desired properties, we need to use the
fact that Y is non-reflexive. The James's theorem implies that there exists y∗ ∈
Y ∗ \ NA(Y ), and we can fix a sequence (cid:0)zj(cid:1)∞
j=0 ⊂ SY such that y∗(cid:0)zj(cid:1) ր ky∗k as
j → ∞. The following set B clearly is the closed unit ball for an equivalent norm
· on X.
B := conv(cid:16)BY ×
1
2
A ∪
∞
[
j=0
(cid:0)(cid:8)zj(cid:9) × Fj ∪ (cid:8) − zj(cid:9) × (−Fj)(cid:1)(cid:17)
Indeed, B is closed absolutely convex with BY × 1
2 A ⊂ B ⊂ BY × A.
§3: We will be done when we prove that q(B) = A. The inclusion q(B) ⊃ A
j=0(Fj ∪ (−Fj )), and obviously
is easy to see; indeed, A is symmetric so A = S∞
S∞
j=0(Fj ∪ (−Fj)) ⊂ q(B). It remains to prove that q(B) ⊂ A.
Note that the sequence (zj)∞
j=0 does not converge; our task would therefore be
easy if the definition of B only involved closure instead of the closed convex hull.
If that were the case, our argument would be based on the simple observation that
a sequence of the form (zjk , fk)∞
k=1 with fk ∈ Fjk converges only if the sequence of
indices (jk)∞
k=1 is eventually constant with limk→∞ jk = j∞, whence limk→∞ fk ∈
Fj∞ ⊂ A.
The rest of the proof deals with the technical difficulties arising from the presence
of convex combinations.
§4: To that end, let us fix an arbitrary point u ∈ q(B) and an x ∈ Y such
that (x, u) ∈ B. Then (x, u) can be written as the limit of a sequence of convex
combinations:
(x, u) = lim
k→∞(cid:16)tk(cid:16)xk,
1
2
vk(cid:17) +
nk
X
j=0
(cid:0)αj
k(cid:0)zj, f j
k(cid:1) − βj
k(cid:0)zj, gj
k(cid:1)(cid:1)(cid:17)
= lim
k→∞(cid:16)tkxk +
nk
X
j=0
kzj − βj
(cid:0)αj
kzj(cid:1),
1
2
tkvk +
nk
X
j=0
kf j
k − βj
(cid:0)αj
k(cid:1)(cid:17)
kgj
(3)
k + βj
j=0 (cid:0)αj
where nk ∈ N, tk + Pnk
k(cid:1) = 1 with all the coefficients non-negative,
vk ∈ A and f j
k ∈ Fj for each k ∈ N and j ∈ N ∪ {0}. Note that all the sets Fj , as
well as A, are convex, and so in the convex combinations it is enough to take only
one element from each of them instead of a convex combination of finitely many
elements.
k, gj
In paragraphs §5 to §14 of this proof we essentially make a series of simplifying
assumptions which will be shown to cause no loss of generality; in this manner we
shall reduce the matter enough to be able to formulate the final argument in the
last six paragraphs.
§5: First, let us formally make all the sums in (3) infinite series by defining for
each k ∈ N and each j > nk,
αj
k := 0,
βj
k := 0,
f j
k := f 1
k
and gj
k := g1
k.
Thus we now have
(x, u) = lim
k→∞(cid:16)tkxk +
∞
X
j=0
kzj − βj
(cid:0)αj
kzj(cid:1),
1
2
tkvk +
∞
X
j=0
kf j
k − βj
(cid:0)αj
k(cid:1)(cid:17)
kgj
(4)
(5)
14
MARTIN RMOUTIL
where each of the infinite series has only finitely many non-zero summands (for
each k ∈ N). In the following we shall often use this fact without recalling it.
§6: By passing to a subsequence (using a diagonal argument) we can assume
that all the following limits exist for all j ∈ N (recall that all the sequences (f j
and (gj
k=1 are bounded in Rd):
k)∞
k )∞
k=1
αj := lim
k→∞
αj
k,
βj := lim
k→∞
f j
k
βj
k,
t := lim
k→∞
and gj := lim
k→∞
f j := lim
k→∞
tk,
gj
k.
(6)
§7: We now distinguish two cases: Either t = 0 or t > 0. Assume first t > 0;
then we shall conclude that u ∈ A as follows. We can find k0 ∈ N such that tk > t
2
for each k ≥ k0. Hence, for each k ≥ k0 we have
tkvk +
∞
X
j=0
kf j
(cid:0)αj
tk
2
+
∞
X
j=0
k + βj
(cid:0)αj
k − βj
kgj
k(cid:1)(cid:13)(cid:13)(cid:13)
k(cid:1) = tk +
tk
2
vk(cid:13)(cid:13)(cid:13)
≤ (cid:13)(cid:13)(cid:13)
(cid:0)αj
X
∞
j=0
+
∞
X
j=0
kf j
(cid:13)(cid:13)αj
k(cid:13)(cid:13) +
∞
X
j=0
kgj
(cid:13)(cid:13)βj
k(cid:13)(cid:13)
k + βj
k(cid:1) −
tk
2
= 1 −
tk
2
< 1 −
t
4
.
1
2
(cid:13)(cid:13)(cid:13)
≤
Hence,
u = lim
k→∞(cid:16) 1
2
tkvk +
∞
X
j=0
kf j
k − βj
(cid:0)αj
k(cid:1)(cid:17) ∈ (cid:16)1 −
kgj
t
4(cid:17)B(Rd,k·k) ⊂ U(Rd,k·k) ⊂ A.
§8: Till the end of the proof, we shall therefore investigate the case when t =
limk→∞ tk = 0. It follows from (5) that
u = lim
k→∞
(1 − tk)
∞
X
j=1
(cid:16) αj
k
1 − tk
f j
k −
βj
k
1 − tk
k(cid:17) = lim
gj
k→∞
∞
X
j=1
(cid:16) αj
k
1 − tk
f j
k −
βj
k
1 − tk
k(cid:17),
gj
and similarly for the first coordinate x; the sum on the right is a convex combination
for each k ∈ N. Hence, we can assume without loss of generality that tk = 0 for
each k, so we have
∞
X
j=0
k + βj
(cid:0)αj
k(cid:1) = 1 for each k, and
(x, u) = lim
k→∞(cid:16)
∞
X
j=0
kzj − βj
(cid:0)αj
kzj(cid:1),
∞
X
j=0
kf j
k − βj
(cid:0)αj
k(cid:1)(cid:17).
kgj
§9: Next assumption we want to make is α0
k = β0
k = 0 for all k ∈ N. In the
present paragraph we show that this causes no loss of generality:
Let us first consider the case when α0 + β0 = 1; then
lim
k→∞
∞
X
j=1
k + βj
(cid:0)αj
k(cid:1) = lim
k→∞(cid:0)1 − (cid:0)α0
k + β0
k(cid:1)(cid:1) = 1 − (cid:0)α0 + β0(cid:1) = 0.
It is now easy to see (e.g. using Lemma 5.4) that limk→∞ P∞
whence (by (5) and (6)) u = α0f 0 + β0g0 ∈ F0 ⊂ A.
j=1 (cid:0)αj
kf j
k − βj
kgj
k(cid:1) = 0,
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
15
Assume now that α0 + β0 < 1. Clearly we can also assume that for each k ∈ N
we have α0
k + β0
k < 1, and consider the limit of convex combinations
u := lim
k→∞
∞
X
j=1
(cid:16)
αj
k
k + β0
1 − (cid:0)α0
k(cid:1)
f j
k −
βj
k
k + β0
1 − (cid:0)α0
k(cid:1)
k(cid:17);
gj
(7)
it is easy to see that the limit exists. Then we have
u = (cid:0)α0 + β0(cid:1)(cid:16) α0
α0 + β0 f 0 −
+ (cid:0)1 − (cid:0)α0 + β0(cid:1)(cid:1) lim
= (cid:0)α0 + β0(cid:1)(cid:16) α0
α0 + β0 f 0 −
k→∞
β0
α0 + β0 g0(cid:17)
αj
(cid:16)
X
k
∞
j=1
1 − (cid:0)α0 + β0(cid:1)
β0
α0 + β0 g0(cid:17) + (cid:0)1 − (cid:0)α0 + β0(cid:1)(cid:1)u,
f j
k −
βj
k
1 − (cid:0)α0 + β0(cid:1)
k(cid:17)
gj
(8)
where the last equality follows from Lemma 5.4 whose assumption is verified as
follows:
j=1
αj
k
(cid:17)∞
(cid:13)(cid:13)(cid:13)
(cid:16)
= (cid:12)(cid:12)(cid:12)
1 − (cid:0)α0 + β0(cid:1)
1 − (cid:0)α0 + β0(cid:1)
−
αj
(cid:17)∞
j=1(cid:13)(cid:13)(cid:13)1
− (cid:16)
k
k + β0
1 − (cid:0)α0
k(cid:1)
1
(cid:12)(cid:12)(cid:12)
k(cid:1)∞
· (cid:13)(cid:13)(cid:0)αj
j=1(cid:13)(cid:13)1 −→ 0,
k + β0
1 − (cid:0)α0
k(cid:1)
k instead of αj
k.
1
and similarly with βj
k → ∞,
In (8) we expressed u as a convex combination of an element of F0 ⊂ A and u;
therefore, to see that u ∈ A, it is sufficient to prove u ∈ A. Of course, u is in q(B)
as it is the projection of (x, u) ∈ B where
x := lim
k→∞
∞
X
j=1
(cid:16)
αj
k
k + β0
1 − (cid:0)α0
k(cid:1)
zj −
βj
k
k + β0
1 − (cid:0)α0
k(cid:1)
zj(cid:17).
To avoid introducing more notation, we henceforth assume that α0
k = β0
k = 0 for
all k ∈ N, so we now have
(x, u) = lim
k→∞(cid:16)
∞
X
j=1
kzj − βj
(cid:0)αj
kzj(cid:1),
∞
X
j=1
kf j
k − βj
(cid:0)αj
k(cid:1)(cid:17).
kgj
§10: Next we distinguish two subcases depending on the value of
ν := minn lim
k→∞
∞
X
j=1
αj
k ,
lim
k→∞
∞
X
j=1
ko;
βj
by passing to a subsequence we can guarantee that both limits exist.
§11: Let us first consider the case when ν > 0; we claim that then u ∈ A. To
show that, set for k ∈ N,
αk :=
∞
X
j=1
αj
k,
βk :=
∞
X
j=1
βj
k,
fk :=
1
αk
∞
X
j=1
αj
kf j
k
and gk :=
1
βk
∞
X
j=1
kgj
βj
k,
and find k0 ∈ N such that for each k ≥ k0, αk, βk ∈ (cid:0) ν
k ≥ k0, let us assume for example αk ≥ βk; the opposite case is analogous.
2(cid:1). Now, for a given
2 , 1 − ν
16
MARTIN RMOUTIL
Recall that for j ≥ 1, f j
k , gj
k ∈ Fj ⊂ U (v, r) which implies that fk, gk ∈ U (v, r),
and thus kfk − gkk < 2r. We now have
∞
X
j=1
(cid:13)(cid:13)(cid:13)
kf j
k − βj
kgj
(cid:0)αj
k(cid:1)(cid:13)(cid:13)(cid:13)
= kαkfk − βkgkk = k(αk − βk)fk + βk(fk − gk)k
≤ αk − βk + 2rβk = 1 − 2βk + 2rβk = 1 − 2βk(1 − r) < 1 − ν(1 − r) < 1.
This is true for any k ≥ k0, and consequently
u = lim
k→∞
∞
X
j=1
kf j
k − βj
kgj
(cid:0)αj
k(cid:1) ∈ (1 − ν(1 − r)) · B(Rd,k·k) ⊂ U(Rd,k·k) ⊂ A.
§12: We now turn to the case ν = 0. The first step is to observe that we can
assume without loss of generality that either αk = 0 for all k, or βk = 0 for all k.
Setting
ck := (cid:0)α1
k, α2
k, . . .(cid:1) and dk := (cid:0)β1
k, β2
k, . . .(cid:1),
we see that either kckk1 = αk → 0 or kdkk1 = βk → 0. Let us examine only the
second case as the first one is analogous. Then we have (e.g. by Lemma 5.4)
lim
k→∞
∞
X
j=1
βj
kgj
k = 0,
and therefore (recall that αk + βk = 1, whence αk → 1)
lim
k→∞
∞
X
j=1
αj
k
αk
f j
k = lim
k→∞
∞
X
j=1
αj
kf j
k = lim
k→∞
∞
X
j=1
kf j
k − βj
kgj
(cid:0)αj
k(cid:1) = u,
from where the observation follows.
From now on, we assume that βj
k = 0 for all j, k ∈ N, so
(x, u) = lim
k→∞(cid:16)
∞
X
j=1
αj
kzj ,
∞
X
j=1
αj
kf j
k(cid:17).
(9)
§13: Again we need to distinguish two further subcases: This time depending
on the value of
λ :=
∞
X
j=1
αj .
§14: It is easy to see that λ ∈ [0, 1]; the easier situation is when λ = 1: In this
case we need to observe that the limits in the definition of u commute, that is:
u = lim
k→∞
∞
X
j=1
kf j
αj
k =
∞
X
j=1
αjf j.
(10)
Indeed, let ε > 0 be given and let us find j0 ∈ N such that P∞
and k0 ∈ N such that for each k ≥ k0 we have (cid:13)(cid:13)Pj0
(cid:12)(cid:12)Pj0
j=j0+1 αj < ε
4 ,
4 and
4 . Then for each k ≥ k0 we also obtain (using the assumption
kf j
k − αj f j(cid:1)(cid:13)(cid:13) < ε
k − αj(cid:1)(cid:12)(cid:12) < ε
j=1 (cid:0)αj
j=1 (cid:0)αj
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
17
λ = 1) that (cid:12)(cid:12)P∞
j=j0+1 (cid:0)αj
αj
kf j
X
k −
X
∞
∞
j=1
j=1
(cid:13)(cid:13)(cid:13)
4 , and so
k − αj(cid:1)(cid:12)(cid:12) < ε
αjf j(cid:13)(cid:13)(cid:13)
k − αjf j(cid:1)(cid:13)(cid:13)(cid:13)
X
k +
αj
∞
j0
X
j=1
≤ (cid:13)(cid:13)(cid:13)
<
ε
4
+
kf j
(cid:0)αj
∞
X
j=j0+1
j=j0+1
∞
X
j=j0+1
+ (cid:13)(cid:13)(cid:13)
αj <
ε
4
+ (cid:16)
kf j
αj
k(cid:13)(cid:13)(cid:13)
+ (cid:13)(cid:13)(cid:13)
X
∞
j=j0+1
∞
X
j=j0+1
αjf j(cid:13)(cid:13)(cid:13)
4(cid:17) +
ε
4
ε
αj +
< ε.
This proves the second equality in (10), which implies that u ∈ convσ(A) (note
that f j, gj ∈ Fj ⊂ A for every j ∈ N) and it follows by Proposition 5.3 that
u ∈ conv(A) = A.
§15: Finally, assume that λ ∈ [0, 1); we will show that this assumption leads to
a contradiction. Set
xλ :=
∞
X
j=1
αjzj
and x(1−λ) := x − xλ = lim
k→∞
∞
X
j=1
(cid:0)αj
k − αj(cid:1)zj.
Now we aim to prove that x(1−λ)
of zj. But it need not suffice to simply divide all the coefficients αj
as some of them could be negative. Hence, for any a ∈ R we define
1−λ is the limit of a sequence of convex combinations
k − αj by 1 − λ
a+ := max{a, 0},
a− := (−a)+,
and γj
k := (cid:0)αj
k − αj(cid:1)+
,
and we claim that
x(1−λ) = lim
k→∞
∞
X
j=1
γj
kzj.
§16: By Lemma 5.4 it is enough to prove that
j=1 − (cid:0)γj
k − αj(cid:1)∞
lim
k→∞(cid:13)(cid:13)(cid:0)αj
k(cid:1)∞
j=1(cid:13)(cid:13)1 = 0.
≤ αj as αj
j=N +1 αj < ε
2
2 . Note
k ≥ 0. Then for each k ≥ k0 we have
k − αj(cid:1)−
k − αj(cid:1)∞
Let an arbitrary ε > 0 be given. We can find N ∈ N such that P∞
and k0 ∈ N such that for each natural k ≥ k0 we have PN
that (cid:0)αj
k(cid:1)∞
(cid:13)(cid:13)(cid:0)αj
j=1(cid:13)(cid:13)1 =
§17: Now, set σk := P∞
j=1 γj
j=1 (cid:0)αj
we obtain
k −αj(cid:1) = 1−λ also implies that limk→∞ σk = 1−λ. Thus, setting γj
k; then the last part together with the fact that
k := γ j
(cid:12)(cid:12)αj
k − αj(cid:12)(cid:12) +
k − αj(cid:12)(cid:12) < ε
k − αj(cid:1)−
j=1 − (cid:0)γj
j=1 (cid:12)(cid:12)αj
P∞
αj < ε.
(cid:0)αj
X
j=1
X
j=1
j=N +1
∞
X
k
σk
≤
∞
N
x(1−λ) = lim
k→∞
1 − λ
σk
∞
X
j=1
γj
kzj = (1 − λ) lim
k→∞
γj
kzj,
X
j=1
∞
and from the definition it immediately follows that
∞
X
j=1
γj
k = 1 for each k, and lim
k→∞
γj
k = 0 for each j.
18
MARTIN RMOUTIL
§18: From these last facts we can readily see that
x(1−λ)
1 − λ
= lim
k→∞
∞
X
j=1
γj
kzj ∈
∞
\
N =1
conv(cid:8)zj; j > N(cid:9).
(11)
Indeed, pick arbitrary N ∈ N and ε > 0, and find k0 such that for each natural
k ≥ k0,
γk :=
N
X
j=1
γj
k <
ε
2
.
Then for each k ≥ k0,
dist(cid:16)
kzj, conv(cid:8)zj; j > N(cid:9)(cid:17)
γj
∞
∞
j=1
j=1
X
X
≤ (cid:13)(cid:13)(cid:13)
≤ (cid:13)(cid:13)(cid:13)
X
≤ γk + (cid:12)(cid:12)(cid:12)
j=1
N
γj
kzj −
1
1 − γk
∞
X
j=N +1
γj
kzj(cid:13)(cid:13)(cid:13)
+ (cid:13)(cid:13)(cid:13)
1
1 −
1 − γk
(cid:16)1 −
(cid:12)(cid:12)(cid:12)
γj
kzj(cid:13)(cid:13)(cid:13)
(cid:17)
X
∞
j=N +1
1
1 − γk
γj
kzj(cid:13)(cid:13)(cid:13)
(1 − γk) = γk + γk < ε.
This estimate yields that x(1−λ)
1−λ ∈ conv(cid:8)zj; j > N(cid:9).
§19: But this is impossible since the intersection in (11) is empty as we will
show below. On the other hand, the limit in the definition (9) of x converges by
the assumption and since xλ is a well-defined element of the Banach space Y , the
limit defining x(1−λ) necessarily converges as well. This is a contradiction showing
that the case λ ∈ [0, 1) does not occur; since in all the other cases we have already
shown that u ∈ A, the proof will be complete, once we make the observation that
C :=
∞
\
N =1
conv(cid:8)zj; j > N(cid:9) = ∅.
§20: Recall that (cid:0)zj(cid:1)∞
j=1 ⊂ SY was chosen to be a norming sequence of the
non-norm-attaining functional y∗. If there existed a point z ∈ C, then z ∈ BY and
y∗(z) > y∗(cid:0)zN(cid:1) for each N ∈ N. Hence, y∗(z) = ky∗k which is a contradiction, and
the proof is complete.
(cid:3)
As a corollary we obtain the following result.
Theorem 5.6. Let X be a non-reflexive Banach space and Y ⊂ X a closed subspace
of finite codimension d ≥ 2. Then X admits an equivalent norm such that Y ⊥ ⊂
NA(X) and Y is not proximinal.
Proof. As explained in §1 of the proof of Lemma 5.5, we can assume that X/Y is
isomorphic to Rd where d is the codimension of Y in X. Let k · k2 be the Euclidean
norm on Rd and let ei, i = 1, . . . , d, be the canonical basis in Rd. Put
A := conv(cid:0)Bk·k2(e1, 1) ∪ Bk·k2(−e1, 1)(cid:1) \ {e1 + e2, −e1 − e2}.
Clearly A is a symmetric convex Fσ set with nonempty interior satisfying the condi-
tion from Lemma 5.5, so there is a renorming · of X such that q(BX ) = A where
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
19
q : X → Rd is the quotient mapping; then A is the closed unit ball in Rd ∼= X/Y
with the corresponding norm. In the sequel, by X we mean (X, ·).
Since q(BX ) = A 6= BX/Y , Lemma 3.3 yields that Y is not proximinal in X.
To prove that Y ⊥ ⊂ NA(X) we will use Lemma 3.3 again. Take any ϕ ∈ (X/Y )∗;
we are to show that it attains its norm at a point of q(BX ). Since dim(X/Y ) < ∞,
there is a point u ∈ att(ϕ).
If u /∈ {e1 + e2, −e1 − e2}, we are done. Suppose
u = e1 + e2; of course, the case u = −e1 − e2 is symmetrical. Further, we can
clearly assume kϕk = 1, so ϕ(u) = 1.
Consider the hyperplane H := {(x1, . . . , xn) ∈ Rd; x2 = 1}; obviously H is a
tangent hyperplane to A at the point e1+e2 (i.e. H∩Int(A) = ∅ and e1+e2 ∈ H∩A).
But there is only one such hyperplane because A ⊃ Bk·k2(e1, 1) and e1 + e2 lies on
the boundary of Bk·k2(e1, 1). It follows that H = ϕ−1(1). But e2 ∈ H ∩ q(BX )
which concludes the proof.
(cid:3)
Proposition 5.7. Let Y ⊂ ℓ1 be a closed subspace of codimension d ∈ N, and
q : ℓ1 → ℓ1/Y be the quotient mapping. Let k · k be an equivalent norm on ℓ1/Y and
let A ⊂ ℓ1/Y be a symmetric convex Fσ-set satisfying U(ℓ1/Y,k·k) ⊂ A ⊂ B(ℓ1/Y,k·k).
Then there is an equivalent norm · on ℓ1 such that q(B(ℓ1,·)) = A.
Proof. The proof is very similar to that of Lemma 5.5, and so we will do it more
briefly, emphasizing the necessary changes in each paragraph of the proof.
§1: Denote X := ℓ1. The subspace Y ⊂ X is isomorphic to Z := (cid:8)(cid:0)xj(cid:1)∞
j=1 ∈
ℓ1; x1 = x2 = · · · = xd = 0(cid:9) as all subspaces of the same finite codimension in
X are isomorphic. But Z is obviously (isometrically) isomorphic to ℓ1, and so we
can (similarly as in the proof of Lemma 5.5) assume that X = Y ⊕∞ Rd where
Y = ℓ1. Throughout the proof we shall only consider the norm k · k on Rd for which
B(Rd,k·k) = A.
§2: In a similar (but simpler) way as in 5.5 we can find closed convex sets Fj
(j ∈ N) of diameter less than 1 and such that A = S∞
of Y = ℓ1 and we define
j=1 Fj.
Now, instead of using the James's theorem, we take the canonical basis (ej)∞
j=1
B := conv(cid:16)BY ×
1
2
A ∪
∞
[
j=1
(cid:0)(cid:8)ej(cid:9) × Fj ∪ (cid:8) − ej(cid:9) × (−Fj)(cid:1)(cid:17).
Again, B is the closed unit ball for an equivalent norm · on X.
§3: We want to prove that q(B) = A; the inclusion q(B) ⊃ A follows immedi-
ately from the definition of B and the fact that A = S∞
such that (x, u) ∈ B. As in 5.5,
j=1 Fj.
§4: To prove the converse inclusion, we again take arbitrary u ∈ q(B) and x ∈ Y
(x, u) = lim
k→∞(cid:16)tkxk +
nk
X
j=1
kej − βj
(cid:0)αj
kej(cid:1),
1
2
nk
tkvk +
kf j
X
(cid:0)αj
k − βj
k(cid:1)(cid:17)
kgj
k , gj
k ≥ 0, vk ∈ A and f j
j=1
(12)
§5, §6: Setting αj
where nk ∈ N, tk + Pnk
j=1 (cid:0)αj
k, βj
k, f j
k(cid:1) = 1 with αj
k for k ∈ N and j > nk as in (4), we can assume
again that all the limits in (6) exist. As in 5.5, we now formally consider all the
sums in (12) as infinite series.
k + βj
k and gj
k ∈ Fj.
k, βj
§7, §8: In the same way as in 5.5 we prove that we can assume all tk's to be
zero (otherwise u ∈ Int(A)).
20
MARTIN RMOUTIL
§9: This step of the proof we skip entirely because in this case we do not work
with the set F0.
§10: Here we make one of the important changes by defining
µ := lim
k→∞
∞
X
j=1
min(cid:8)αj
k, βj
k(cid:9);
of course, we can assume that this limit exists, and we distinguish two cases: µ > 0
and µ = 0.
§11: Suppose µ > 0. For k ∈ N set
∞
uk :=
kgj
k(cid:1);
then u = limk→∞ uk. Further, for all j, k ∈ N set µj
k := min(cid:8)αj
and for a given j ∈ N assume that αj
obtain the same estimate:
k ≥ βj
k − βj
(cid:0)αj
X
kf j
j=1
k(cid:9). Fix k ∈ N
k; the opposite case is analogous and we
k, βj
k − βj
kf j
(cid:13)(cid:13)αj
k(cid:1)f j
k(cid:1) + µj
k − µj
k;
here the inequality follows from the fact that f j
k , gj
Now, by taking the sum over all j ∈ N we obtain
kgj
k(cid:13)(cid:13) = (cid:13)(cid:13)(cid:0)αj
≤ (cid:0)αj
= αj
k − µj
k − µj
k + βj
k
k + µj
k(cid:0)f j
k − gj
k(cid:1)(cid:13)(cid:13)
k ∈ Fj ⊂ BRd and diam(Fj ) < 1.
kukk ≤
∞
X
j=1
k − βj
kf j
(cid:13)(cid:13)αj
kgj
k(cid:13)(cid:13) ≤
∞
X
j=1
k + βj
k − µj
(cid:0)αj
k(cid:1) = 1 −
∞
X
j=1
µj
k.
It follows after passage to limit as k → ∞ that kuk ≤ 1−µ, and so u ∈ (1−µ)BRd ⊂
URd ⊂ A.
§12: For the rest of the proof we shall assume that µ = 0. Similarly as in the
proof of Lemma 5.5 we prove that there is no loss of generality in assuming that
µk = 0 for each k. In other words, we can assume that for all j, k ∈ N we have
αj
k = 0 or βj
§13: Recalling the notation in (4), we finally distinguish two cases depending
k = 0.
on the value of
λ :=
∞
X
j=1
(cid:0)αj + βj(cid:1).
§14: Again, λ ∈ [0, 1] and we first resolve the case λ = 1: We shall observe that
the limit and the series in the definition of u commute, that is:
u = lim
k→∞
∞
X
j=1
kf j
k − βj
kgj
(cid:0)αj
k(cid:1) =
∞
X
j=1
(cid:0)αj f j − βjgj(cid:1).
(13)
Indeed, let ε > 0 be given and let us find j0 ∈ N such that P∞
j=j0+1 (cid:0)αj + βj(cid:1) < ε
4 ,
and k0 ∈ N such that for each k ≥ k0 we have (cid:13)(cid:13)Pj0
kgj
k − αj f j − (cid:0)βj
j=1 (cid:0)αj
kf j
k −
βjgj(cid:1)(cid:1)(cid:13)(cid:13) < ε
j=1 (cid:0)αj + βj(cid:1)(cid:12)(cid:12) < ε
4 . By the same calculation
(cid:0)αj f j − βjgj(cid:1)(cid:13)(cid:13)(cid:13)
X
4 and (cid:12)(cid:12)Pj0
(cid:13)(cid:13)(cid:13)
X
as in the proof of 5.5 we obtain that for each k ≥ k0,
k(cid:1) −Pj0
j=1 (cid:0)αj
k − βj
k + βj
k(cid:1) −
(cid:0)αj
kf j
kgj
< ε.
j=1
j=1
∞
∞
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
21
This proves (13), so u ∈ convσ(A) = conv(A) = A by Proposition 5.3 and the
convexity of A.
§15: Aiming for a contradiction, suppose that λ ∈ [0, 1) and set
xλ :=
∞
X
j=1
(cid:0)αj − βj(cid:1)ej
and
x(1−λ) := x − xλ = lim
k→∞
∞
X
j=1
(cid:0)αj
k − αj − (cid:0)βj
k − βj(cid:1)(cid:1)ej.
Note that the last limit must exist as it is the difference of an existing limit and
xλ. Define
γj
k := (cid:0)αj
and we now claim that
k − αj(cid:1)+
and δj
k := (cid:0)βj
k − βj(cid:1)+
,
x(1−λ) = lim
k→∞
∞
X
j=1
k − δj
(cid:0)γj
k(cid:1)ej.
§16: By Lemma 5.4 it suffices to prove that
This can be done in the same way as in the proof of 5.5.
lim
k − αj(cid:1)∞
j=1 − (cid:0)γj
k→∞(cid:13)(cid:13)(cid:0)αj
§17: Now, set σk := P∞
Further, define
k(cid:1)∞
j=1(cid:13)(cid:13)1 = 0
j=1 (cid:0)γj
γj
k
σk
γj
k :=
and
lim
j=1 − (cid:0)δj
k − βj(cid:1)∞
k→∞(cid:13)(cid:13)(cid:0)βj
k(cid:1)∞
j=1(cid:13)(cid:13)1 = 0.
k(cid:1) and observe that limk→∞ σk = 1 − λ.
k + δj
and δj
k :=
δj
k
σk
;
then we have:
(i) x(1−λ) = (1 − λ) limk→∞ P∞
j=1 (cid:0)γj
(ii) for each k ∈ N, P∞
k + δj
j=1 (cid:0)γj
k(cid:1) = 1;
k = 0 or δj
(iii) for all j, k ∈ N, either γj
k = 0 and limk→∞ δj
(iv) and for each j ∈ N, limk→∞ γj
§18, §19, §20: Now, on one hand we can readily see from (i), (ii) and (iii) that
k(cid:1)ej;
k − δj
k = 0;
k = 0.
kx(1−λ)k = (1 − λ) lim
k→∞
∞
X
j=1
k + δj
(cid:0)γj
k(cid:1) = (1 − λ) lim
k→∞
1 = 1 − λ > 0.
On the other hand, (i) and (iv) yield that x(1−λ) = 0. Indeed, (i) says that x(1−λ)
is the norm-limit of a certain sequence which by (iv) converges pointwise to 0. (cid:3)
Remark 5.8.
(a) Let X be a Banach space, Y ⊂ X be a closed subspace and
q : X → X/Y be the quotient mapping. Since any functional y∗ ∈ Y ⊥ can
be expressed as y∗ = ϕ ◦ q for some ϕ := (X/Y )∗, one can notice that in the
proof of Lemma 3.3 (ii) we actually show the following equivalence for each
y∗ ∈ Y ⊥:
y∗ ∈ NA(X) ⇐⇒ the corresponding ϕ ∈ (X/Y )∗ attains its norm on q(BX ).
Consider the situation when X = ℓ1 and Y has codimension d in X; then
X/Y ∼= Rd. The set A := U(Rd,k·k2) satisfies the assumptions of Proposi-
tion 5.7, so there is an equivalent norm · on ℓ1 such that q(B(ℓ1,·)) =
22
MARTIN RMOUTIL
A. From the above equivalence we immediately obtain that Y ⊥ ⊂ (X ∗ \
NA(ℓ1, ·)) ∪ {0}.
In other words, given any finite-dimensional subspace Z ⊂ (ℓ1)∗, we can find
a renorming such that Z \ {0} is contained in non-norm-attaining functionals.
Note that a result of F.J. Garc´ıa-Pacheco [7, Theorem 1.2] implies that ℓ∗
1 \
NA(ℓ1) is, in fact, spaceable.
It is also easy to see that, in the new norm, Y is antiproximinal (i.e. no
point x /∈ Y has in Y a closest point). Indeed, the proof of Lemma 3.3 (i) also
shows that
X is antiproximinal ⇐⇒ q(BX ) = UX/Y .
Likewise, given a closed subspace Y of finite codimension in a non-reflexive
Banach space X, there exists an equivalent norm on X such that Y is prox-
iminal in X. This follows from Lemma 5.5 as the set A := BX/Y satisfies the
assumptions.
(b) The assumption that Y is finite-codimensional is very important for our ap-
proach to both Lemma 5.5 and Proposition 5.7. In the proof of the former,
we use this assumption to see that Y is complemented, and therefore non-
reflexive as its complement is finite-dimensional; this argument is not neces-
sary in the latter case as ℓ1 contains no infinite-dimensional complemented
reflexive subspaces-see [6, Section 13.7]. The fact that Y is non-reflexive is
needed as any reflexive space is proximinal in any superspace (see [21]; the
converse is also true-[19] or [22]). In both cases we rely on Proposition 5.3.
(c) It is natural to ask whether the assumption that A is Fσ in Lemma 5.5 and
Proposition 5.7 can be relaxed to higher Borel classes or even analytic sets.
Related questions have already been studied: R. Kaufman proves in [15] that
any non-reflexive Banach space admits an equivalent norm such that the set of
norm-attaining functionals is not Borel; in [16] he improves this result showing
that such a renorming can be made even Fr´echet smooth. Following Kaufman
and using his deep abstract lemmas, O. Kurka proves in his paper [18] that
given any separable non-reflexive Banach space X and an ordinal α < ω1, there
exists an equivalent strictly convex norm on X such that the corresponding
set of norm-attaining functionals is not of the (additive) Borel class α (it is
observed in [15] that when X is strictly convex, N A(X) is Borel).
It seems likely that Kurka's method could be (at least in some cases) used
to prove a version of Lemma 5.5 with weaker assumptions on the descriptive
quality of A.
Acknowledgment. I would like to thank Bernardo Cascales (who suggested the
topic of this article during my visit in Murcia), Richard Aron and Mat´ıas Raja for
their helpful comments, moral support and interest in my work. Most of all, I am
very grateful to Ondrej Kalenda for many fruitful discussions and careful reading
of the manuscript which resulted in further helpful suggestions.
References
[1] Mar´ıa D. Acosta, Antonio Aizpuru, Richard M. Aron, and Francisco J. Garc´ıa-Pacheco.
Functionals that do not attain their norm. Bull. Belg. Math. Soc. Simon Stevin, 14(3):407–
418, 2007.
[2] Pradipta Bandyopadhyay and Gilles Godefroy. Linear structures in the set of norm-attaining
functionals on a Banach space. J. Convex Anal., 13(3-4):489–497, 2006.
NORM-ATTAINING FUNCTIONALS AND PROXIMINAL SUBSPACES
23
[3] Luis Bernal-Gonz´alez, Daniel Pellegrino, and Juan B. Seoane-Sep´ulveda. Linear subsets of
nonlinear sets in topological vector spaces. Bull. Amer. Math. Soc. (N.S.), 51(1):71–130,
2014.
[4] Errett Bishop and Robert R. Phelps. A proof that every Banach space is subreflexive. Bull.
Amer. Math. Soc., 67:97–98, 1961.
[5] Mahlon M. Day. Normed linear spaces. Springer-Verlag, New York-Heidelberg, third edition,
1973. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 21.
[6] Mari´an Fabian, Petr Habala, Petr H´ajek, Vicente Montesinos, and V´aclav 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.
[7] Francisco J. Garc´ıa-Pacheco. Vector subspaces of the set of non-norm-attaining functionals.
Bull. Aust. Math. Soc., 77(3):425–432, 2008.
[8] Aleksandr L. Garkavi. On best approximation by the elements of infinite-dimensional sub-
spaces of a certain class. Mat. Sb. (N.S.), 62 (104):104–120, 1963.
[9] Gilles Godefroy. The Banach space c0. Extracta Math., 16(1):1–25, 2001.
[10] Gilles Godefroy and V. Indumathi. Proximinality in subspaces of c0. J. Approx. Theory,
101(2):175–181, 1999.
[11] Gilles Godefroy and V. Indumathi. Proximinality and renormings: some new examples. J.
Approx. Theory, 176:118–128, 2013.
[12] V. Indumathi. Proximinal subspaces of finite codimension in general normed linear spaces.
Proc. London Math. Soc. (3), 45(3):435–455, 1982.
[13] V. Indumathi. On transitivity of proximinality. J. Approx. Theory, 49(2):130–143, 1987.
[14] Robert C. James. Characterizations of reflexivity. Studia Math., 23:205–216, 1963/1964.
[15] Robert P. Kaufman. Topics on analytic sets. Fund. Math., 139(3):215–229, 1991.
[16] Robert P. Kaufman. On smooth norms and analytic sets. Israel J. Math., 116:21–27, 2000.
[17] Victor Klee. Some new results on smoothness and rotundity in normed linear spaces. Math.
Ann., 139:51–63 (1959), 1959.
[18] Ondrej Kurka. Structure of the set of norm-attaining functionals on strictly convex spaces.
Canad. Math. Bull., 54(2):302–310, 2011.
[19] Walter Pollul. Reflexivitat und Existenz-Teilraume in der linearen Approximationstheorie.
Gesellschaft fur Mathematik und Datenverarbeitung, Bonn, 1972. Gesellschaft fur Mathe-
matik und Datenverarbeitung, Bonn, Ber. No. 53.
[20] Charles J. Read. Banach spaces with no proximinal
subspaces of codimension 2.
arXiv:1307.7958.
[21] Ivan Singer. Best approximation in normed linear spaces by elements of linear subspaces.
Translated from the Romanian by Radu Georgescu. Die Grundlehren der mathematischen
Wissenschaften, Band 171. Publishing House of the Academy of the Socialist Republic of
Romania, Bucharest; Springer-Verlag, New York-Berlin, 1970.
[22] Ivan Singer. On normed linear spaces which are proximinal in every superspace. J. Approxi-
mation Theory, 7:399–402, 1973.
E-mail address, M. Rmoutil: [email protected]
University of Warwick, Mathematics Institute, Zeeman Building, Coventry CV4 7AL,
United Kingdom
Charles University, Faculty of Mathematics and Physics, Department of
Mathematical Analysis, Sokolovsk´a 83, 186 75 Prague 8, Czech Republic
|
1004.2386 | 1 | 1004 | 2010-04-14T12:52:32 | The Topological Centers Of Module Actions | [
"math.FA"
] | In this article, for Banach left and right module actions, we will extend some propositions from Lau and $\ddot{U}lger$ into general situations and we establish the relationships between topological centers of module actions. We also introduce the new concepts as $Lw^*w$-property and $Rw^*w$-property for Banach $A-bimodule$ $B$ and we investigate the relations between them and topological center of module actions. We have some applications in dual groups. | math.FA | math |
THE TOPOLOGICAL CENTERS OF MODULE ACTIONS
KAZEM HAGHNEJAD AZAR
Abstract. In this article, for Banach left and right module actions, we will
extend some propositions from Lau and U lger into general situations and we
establish the relationships between topological centers of module actions. We also
introduce the new concepts as Lw∗w-property and Rw∗w-property for Banach
A − bimodule B and we investigate the relations between them and topological
center of module actions. We have some applications in dual groups.
1.Introduction and Preliminaries
As is well-known [1], the second dual A∗∗ of A endowed with the either Arens mul-
tiplications is a Banach algebra. The constructions of the two Arens multiplications
in A∗∗ lead us to definition of topological centers for A∗∗ with respect both Arens
multiplications. The topological centers of Banach algebras, module actions and ap-
plications of them were introduced and discussed in [6, 8, 13, 14, 15, 16, 17, 21, 22],
and they have attracted by some attentions.
Now we introduce some notations and definitions that we used throughout this paper.
Let A be a Banach algebra. We say that a net (eα)α∈I in A is a left approximate
identity (= LAI) [resp.
right approximate identity (= RAI)] if, for each a ∈ A,
eαa −→ a [resp. aeα −→ a]. For a ∈ A and a′ ∈ A∗, we denote by a′a and aa′
respectively, the functionals on A∗ defined by < a′a, b >=< a′, ab >= a′(ab) and
< aa′, b >=< a′, ba >= a′(ba) for all b ∈ A. The Banach algebra A is embedded in
its second dual via the identification < a, a′ > - < a′, a > for every a ∈ A and a′ ∈ A∗.
We denote the set {a′a : a ∈ A and a′ ∈ A∗} and {aa′ : a ∈ A and a′ ∈ A∗} by A∗A
and AA∗, respectively, clearly these two sets are subsets of A∗. Let A has a BAI.
If the equality A∗A = A∗, (AA∗ = A∗) holds, then we say that A∗ factors on the
left (right). If both equalities A∗A = AA∗ = A∗ hold, then we say that A∗ factors
on both sides. Let X, Y, Z be normed spaces and m : X × Y → Z be a bounded
bilinear mapping. Arens in [1] offers two natural extensions m∗∗∗ and mt∗∗∗t of m
from X ∗∗ × Y ∗∗ into Z ∗∗ as following:
1. m∗ : Z ∗ × X → Y ∗, given by < m∗(z ′, x), y >=< z ′, m(x, y) > where x ∈ X,
y ∈ Y , z ′ ∈ Z ∗,
2. m∗∗ : Y ∗∗ × Z ∗ → X ∗, given by < m∗∗(y′′, z ′), x >=< y′′, m∗(z ′, x) > where
x ∈ X, y′′ ∈ Y ∗∗, z ′ ∈ Z ∗,
3. m∗∗∗ : X ∗∗ × Y ∗∗ → Z ∗∗, given by < m∗∗∗(x′′, y′′), z ′ > =< x′′, m∗∗(y′′, z ′) >
where x′′ ∈ X ∗∗, y′′ ∈ Y ∗∗, z ′ ∈ Z ∗.
The mapping m∗∗∗ is the unique extension of m such that x′′ → m∗∗∗(x′′, y′′) from
X ∗∗ into Z ∗∗ is weak∗ − to − weak∗ continuous for every y′′ ∈ Y ∗∗, but the mapping
2000 Mathematics Subject Classification. 46L06; 46L07; 46L10; 47L25.
Key words and phrases. Arens regularity, bilinear mappings, Topological center, Second dual,
Module action.
1
2
y′′ → m∗∗∗(x′′, y′′) is not in general weak∗ − to − weak∗ continuous from Y ∗∗ into Z ∗∗
unless x′′ ∈ X. Hence the first topological center of m may be defined as following
Z1(m) = {x′′ ∈ X ∗∗ : y′′ → m∗∗∗(x′′, y′′) is weak∗ − to − weak∗ − continuous}.
Let now mt : Y × X → Z be the transpose of m defined by mt(y, x) = m(x, y) for
every x ∈ X and y ∈ Y . Then mt is a continuous bilinear map from Y × X to Z, and
so it may be extended as above to mt∗∗∗ : Y ∗∗ × X ∗∗ → Z ∗∗. The mapping mt∗∗∗t :
X ∗∗ × Y ∗∗ → Z ∗∗ in general is not equal to m∗∗∗, see [1], if m∗∗∗ = mt∗∗∗t, then
m is called Arens regular. The mapping y′′ → mt∗∗∗t(x′′, y′′) is weak∗ − to − weak∗
continuous for every y′′ ∈ Y ∗∗, but the mapping x′′ → mt∗∗∗t(x′′, y′′) from X ∗∗ into
Z ∗∗ is not in general weak∗ − to − weak∗ continuous for every y′′ ∈ Y ∗∗. So we define
the second topological center of m as
Z2(m) = {y′′ ∈ Y ∗∗ : x′′ → mt∗∗∗t(x′′, y′′) is weak∗ − to − weak∗ − continuous}.
It is clear that m is Arens regular if and only if Z1(m) = X ∗∗ or Z2(m) = Y ∗∗. Arens
regularity of m is equivalent to the following
lim
i
lim
j
< z ′, m(xi, yj) >= lim
j
lim
i
< z ′, m(xi, yj) >,
whenever both limits exist for all bounded sequences (xi)i ⊆ X , (yi)i ⊆ Y and
z ′ ∈ Z ∗, see [6, 18].
The regularity of a normed algebra A is defined to be the regularity of its algebra
multiplication when considered as a bilinear mapping. Let a′′ and b′′ be elements
of A∗∗, the second dual of A. By Goldstin,s Theorem [6, P.424-425], there are nets
(aα)α and (bβ)β in A such that a′′ = weak∗ − limα aα and b′′ = weak∗ − limβ bβ. So
it is easy to see that for all a′ ∈ A∗,
and
lim
α
lim
β
< a′, m(aα, bβ) >=< a′′b′′, a′ >
lim
β
lim
α
< a′, m(aα, bβ) >=< a′′ob′′, a′ >,
where a′′b′′ and a′′ob′′ are the first and second Arens products of A∗∗, respectively,
see [6, 14, 18].
The mapping m is left strongly Arens irregular if Z1(m) = X and m is right strongly
Arens irregular if Z2(m) = Y .
This paper is organized as follows.
a) In section two, for a Banach A − bimodule, we have
(1) a′′ ∈ ZB∗∗(A∗∗) if and only if π∗∗∗∗
(2) F ∈ ZB∗∗((A∗A)∗) if and only if π∗∗∗∗
(3) G ∈ Z(A∗A)∗(B ∗∗) if and only if π∗∗∗∗
(4) Let B has a BAI (eα)α ⊆ A such that eα
r
ℓ
ℓ
(b′, a′′) ∈ B ∗ for all b′ ∈ B ∗.
(g, F ) ∈ B ∗ for all g ∈ B ∗.
(g, G) ∈ A∗A for all g ∈ B ∗.
e∗∗(B ∗∗) = B ∗∗ [
resp. Ze∗∗(B ∗∗) = B ∗∗] and B ∗ factors on the left [resp. right], but not on
the right [resp. left], then ZB∗∗(A∗∗) 6= Z t
B∗∗(A∗∗).
(5) B ∗A ⊆ wapℓ(B) if and only if AA∗∗ ⊆ ZB∗∗(A∗∗).
(6) Let b′ ∈ B ∗. Then b′ ∈ wapℓ(B) if and only if the adjoint of the mapping
w∗
→ e′′. Then if Z t
π∗
ℓ (b′, ) : A → B ∗ is weak∗ − to − weak continuous.
3
b) In section three, for a Banach A − bimodule B, we define Lef t − weak∗ − to − weak
property [=Rw∗w− property] and Right − weak∗ − to − weak property [=Rw∗w−
property] for Banach algebra A and we show that
(1) If A∗∗ = a0A∗∗ [resp. A∗∗ = A∗∗a0] for some a0 ∈ A and a0 has Rw∗w−
property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗.
(2) If B ∗∗ = a0B ∗∗ [resp. B ∗∗ = B ∗∗a0] for some a0 ∈ A and a0 has Rw∗w−
property [resp. Lw∗w− property] with respect to B, then ZA∗∗(B ∗∗) = B ∗∗.
(3) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w−
property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗.
(4) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w−
property [resp. Lw∗w− property] with respect B, then ZA∗∗(B ∗∗) = B ∗∗.
(5) If a0 ∈ A has Rw∗w− property with respect to B, then a0A∗∗ ⊆ ZB∗∗(A∗∗)
and a0B ∗ ⊆ wapℓ(B).
(6) Assume that AB ∗ ⊆ wapℓB. If B ∗ strong factors on the left [resp. right],
then A has Lw∗w− property [resp. Rw∗w− property ] with respect to B.
(7) Assume that AB ∗ ⊆ wapℓB. If B ∗ strong factors on the left [resp. right],
then A has Lw∗w− property [resp. Rw∗w− property ] with respect to B.
2. The topological centers of module actions
Let B be a Banach A − bimodule, and let
πℓ : A × B → B and πr : B × A → B.
be the left and right module actions of A on B. Then B ∗∗ is a Banach A∗∗ − bimodule
with module actions
π∗∗∗
ℓ
: A∗∗ × B ∗∗ → B ∗∗ and π∗∗∗
r
: B ∗∗ × A∗∗ → B ∗∗.
Similarly, B ∗∗ is a Banach A∗∗ − bimodule with module actions
πt∗∗∗t
ℓ
: A∗∗ × B ∗∗ → B ∗∗ and πt∗∗∗t
r
: B ∗∗ × A∗∗ → B ∗∗.
We may therefore define the topological centers of the right and left module actions
of A on B as follows:
ZA∗∗(B ∗∗) = Z(πr) = {b′′ ∈ B ∗∗ : the map a′′ → π∗∗∗
r
(b′′, a′′) : A∗∗ → B ∗∗
is weak∗ − to − weak∗ continuous}
ZB∗∗(A∗∗) = Z(πℓ) = {a′′ ∈ A∗∗ : the map b′′ → π∗∗∗
ℓ
(a′′, b′′) : B ∗∗ → B ∗∗
is weak∗ − to − weak∗ continuous}
Z t
A∗∗(B ∗∗) = Z(πt
ℓ) = {b′′ ∈ B ∗∗ : the map a′′ → πt∗∗∗
ℓ
(b′′, a′′) : A∗∗ → B ∗∗
is weak∗ − to − weak∗ continuous}
Z t
B∗∗(A∗∗) = Z(πt
r) = {a′′ ∈ A∗∗ : the map b′′ → πt∗∗∗
r
(a′′, b′′) : B ∗∗ → B ∗∗
is weak∗ − to − weak∗ continuous}
We note also that if B is a left(resp. right) Banach A − module and πℓ : A × B →
B (resp. πr : B × A → B) is left (resp. right) module action of A on B, then B ∗ is
a right (resp. left) Banach A − module.
4
We write ab = πℓ(a, b), ba = πr(b, a), πℓ(a1a2, b) = πℓ(a1, a2b), πr(b, a1a2) = πr(ba1, a2),
π∗
ℓ (a1b′, a2) = π∗
b′ ∈ B ∗ when there is no confusion.
r (b′, ab), for all a1, a2, a ∈ A, b ∈ B and
ℓ (b′, a2a1), π∗
r (b′a, b) = π∗
Theorem 2-1. We have the following assertions.
(1) Assume that B is a Left Banach A − module. Then, a′′ ∈ ZB∗∗(A∗∗) if and
only if π∗∗∗∗
ℓ
(b′, a′′) ∈ B ∗ for all b′ ∈ B ∗.
(2) Assume that B is a right Banach A − module. Then, b′′ ∈ ZA∗∗(B ∗∗) if and
only if π∗∗∗∗
r
(b′, b′′) ∈ A∗ for all b′ ∈ B ∗.
Proof.
(1) Let b′′ ∈ B ∗∗. Then, for every a′′ ∈ ZB∗∗ (A∗∗), we have
< π∗∗∗∗
ℓ
=< πt∗∗∗t
ℓ
(b′, a′′), b′′ >=< b′, π∗∗∗
ℓ
(a′′, b′′), b′ >=< πt∗∗∗
(b′, a′′) = πt∗∗
ℓ
It follow that π∗∗∗∗
Conversely, let a′′ ∈ A∗∗ and let π∗∗∗∗
all b′′ ∈ B ∗∗, we have
ℓ
ℓ
ℓ
(a′′, b′′) >=< π∗∗∗
ℓ
(a′′, b′′), b′ >
(b′′, a′′), b′ >=< b′′, πt∗∗
(a′′, b′) ∈ B ∗.
ℓ
(a′′, b′) > .
(a′′, b′) ∈ B ∗ for all b′ ∈ B ∗. Then for
< π∗∗∗
ℓ
(a′′, b′′), b′ >=< b′, π∗∗∗
ℓ
(a′′, b′′) >=< π∗∗∗∗
ℓ
(b′, a′′), b′′ >
=< πt∗∗
ℓ
(a′′, b′), b′′ >=< b′′, πt∗∗
(a′′, b′) >=< πt∗∗∗
ℓ
ℓ
=< πt∗∗∗t
(a′′, b′′), b′ > .
ℓ
Consequently a′′ ∈ ZB∗∗(A∗∗).
(2) Prof is similar to (1).
(b′′, a′′), b′ >
(cid:3)
Theorem 2-2. Assume that B is a Banach A−bimodule. Then we have the following
assertions.
(1) F ∈ ZB∗∗((A∗A)∗) if and only if π∗∗∗∗
(2) G ∈ Z(A∗A)∗(B ∗∗) if and only if π∗∗∗∗
ℓ
r
(g, F ) ∈ B ∗ for all g ∈ B ∗.
(g, G) ∈ A∗A for all g ∈ B ∗.
Proof.
(1) Let F ∈ ZB∗∗((A∗A)∗) and (b′′
α)α ⊆ B ∗∗ such that b′′
α
w∗
→ b′′. Then for
all g ∈ B ∗, we have
< π∗∗∗∗
ℓ
(g, F ), b′′
α >=< g, π∗∗∗
ℓ
(F, b′′
α) >=< π∗∗∗
ℓ
(F, b′′
α), g >
→< π∗∗∗
ℓ
(F, b′′), g >=< π∗∗∗∗
ℓ
(g, F ), b′′ > .
(g, F ) ∈ (B ∗∗, weak∗)∗ = B ∗.
(g, F ) ∈ B ∗ for F ∈ (A∗A)∗ and g ∈ B ∗. Assume that
Thus, we conclude that π∗∗∗∗
Conversely, let π∗∗∗∗
ℓ
ℓ
α)α ⊆ B ∗∗ such that b′′
b′′ ∈ B ∗∗ and (b′′
α
α) >=< π∗∗∗∗
α), g >=< g, π∗∗∗
(F, b′′
(g, F ) >=< π∗∗∗∗
(g, F ) >→< b′′, π∗∗∗∗
w∗
→ b′′. Then
ℓ
α, π∗∗∗∗
< π∗∗∗
=< b′′
(F, b′′
ℓ
ℓ
ℓ
ℓ
=< π∗∗∗
It follow that F ∈ ZB∗∗((A∗A)∗).
ℓ
(2) Proof is similar to (1).
ℓ
(F, b′′), g > .
(g, F ), b′′
α >
(g, F ), b′′ >
(cid:3)
5
In the proceeding theorems, if we take B = A, we obtain some parts of Lemma 3.1
from [14].
An element e′′ of A∗∗ is said to be a mixed unit if e′′ is a right unit for the first Arens
multiplication and a left unit for the second Arens multiplication. That is, e′′ is a
mixed unit if and only if, for each a′′ ∈ A∗∗, a′′e′′ = e′′oa′′ = a′′. By [4, p.146], an
element e′′ of A∗∗ is mixed unit if and only if it is a weak∗ cluster point of some BAI
(eα)α∈I in A.
Let B be a Banach A − bimodule and a′′ ∈ A∗∗. We define the locally topological
center of the left and right module actions of a′′ on B, respectively, as follows
a′′ (B ∗∗) = Z t
Z t
a′′(πt
Za′′ (B ∗∗) = Za′′ (πt
ℓ) = {b′′ ∈ B ∗∗ : πt∗∗∗t
r) = {b′′ ∈ B ∗∗ : πt∗∗∗t
r
ℓ
(a′′, b′′) = π∗∗∗
(b′′, a′′) = π∗∗∗
ℓ
r
(a′′, b′′)},
(b′′, a′′)}.
Thus we have
\
a′′ ∈A∗∗
\
a′′ ∈A∗∗
Z t
a′′(B ∗∗) = Z t
A(B ∗∗) = Z(πt
r),
Za′′(B ∗∗) = ZA(B ∗∗) = Z(πr).
Definition 2-3. Let B be a left Banach A − module and e′′ ∈ A∗∗ be a mixed
unit for A∗∗. We say that e′′ is a left mixed unit for B ∗∗, if
π∗∗∗
ℓ
(e′′, b′′) = πt∗∗∗t
ℓ
(e′′, b′′) = b′′,
for all b′′ ∈ B ∗∗.
The definition of right mixed unit for B ∗∗ is similar. B ∗∗ has a mixed unit if it has
left and right mixed unit that are equal.
It is clear that if e′′ ∈ A∗∗ is a left (resp. right) unit for B ∗∗ and Ze′′ (B ∗∗) = B ∗∗,
then e′′ is left (resp. right) mixed unit for B ∗∗.
w∗
→ e′′.
e∗∗ (B ∗∗) = B ∗∗ [ resp. Ze∗∗ (B ∗∗) = B ∗∗] and B ∗ factors on the left [resp.
Theorem 2-4. Let B be a Banach A−bimodule with a BAI (eα)α such that eα
Then if Z t
right], but not on the right [resp. left], then ZB∗∗(A∗∗) 6= Z t
B∗∗(A∗∗).
Proof. Suppose that B ∗ factors on the left with respect to A, but not on the right.
w∗
→ e′′. Thus for all b′ ∈ B ∗ there are
Let (eα)α ⊆ A be a BAI for A such that eα
a ∈ A and x′ ∈ B ∗ such that x′a = b′. Then for all b′′ ∈ B ∗∗ we have
< π∗∗∗
ℓ
(e′′, b′′), b′ >=< e′′, π∗∗
ℓ (b′′, b′) >= lim
α
< π∗∗
ℓ (b′′, b′), eα >
= lim
α
= lim
α
ℓ (b′, eα) >= lim
α
ℓ (x′, aeα) >= lim
α
< b′′, π∗
ℓ (x′a, eα) >
< π∗∗
ℓ (b′′, x′), aeα >
< b′′, π∗
< b′′, π∗
=< π∗∗
ℓ (b′′, x′), a >=< b′′, b′ > .
Thus π∗∗∗
e′′ ∈ ZB∗∗(A∗∗). If we take ZB∗∗(A∗∗) = Z t
(e′′, b′′) = b′′ consequently B ∗∗ has left unit A∗∗ − module. It follows that
B∗∗(A∗∗). Then the
B∗∗(A∗∗), then e′′ ∈ Z t
ℓ
6
mapping b′′ → πt∗∗∗t
w∗
→ e′′, πt∗∗∗t
Since eα
r
(b′′, e′′) is weak∗ − to − weak∗ continuous from B ∗∗ into B ∗∗.
(b′′, eα) w∗
(b′′, e′′). Let b′ ∈ B ∗ and (bβ)β ⊆ B such that
→ πt∗∗∗t
r
w∗
→ b′′. Since Z t
e∗∗(B ∗∗) = B ∗∗, we have the following quality
r
bβ
< πt∗∗∗t
r
(b′′, e′′), b′ >= lim
α
< πt∗∗∗t
r
(b′′, eα), b′ >= lim
α
< πt∗∗∗
r
(eα, b′′), b′ >
= lim
α
lim
β
< πt∗∗∗
r
(eα, bβ), b′ >= lim
α
lim
β
< πr(bβ, eα), b′ >
= lim
α
lim
β
< b′, πr(bβ, eα) >= lim
β
lim
α
< b′, πr(bβ, eα) >
= lim
β
< b′, bβ >=< b′′, b′ > .
Thus πt∗∗∗t
that b′′ ∈ B ∗∗ and (bβ)β ⊆ B such that bβ
(b′′, e′′) = π∗∗∗
r
r
(b′′, e′′) = b′′. It follows that B ′′ has a right unit. Suppose
w∗
→ b′′. Then for all b′ ∈ B ∗ we have
< b′′, b′ >=< π∗∗∗
r
(b′′, e′′), b′ >=< b′′, π∗∗
r (e′′, b′) >= lim
β
< π∗∗
r (e′′, b′), bβ >
= lim
β
< e′′, π∗
r (b′, bβ) >= lim
β
lim
α
< π∗
r (b′, bβ), eα >
= lim
β
lim
α
< π∗
r (b′, bβ), eα >= lim
β
lim
α
< b′, πr(bβ, eα) >
= lim
α
lim
β
< π∗∗∗
r
(bβ, eα), b′ >= lim
α
lim
β
< bβ, π∗∗
r (eα, b′) >
= lim
α
< b′′, π∗∗
r (eα, b′) > .
It follows that weak − limα π∗∗
factors on the right that is contradiction.
r (eα, b′) = b′. So by Cohen Factorization Theorem, B ∗
(cid:3)
Corollary 2-5. Let B be a Banach A − bimodule and e′′ ∈ A∗∗ be a left mixed unit
for B ∗∗. If B ∗ factors on the left, but not on the right, then ZB∗∗(A∗∗) 6= Z t
B∗∗(A∗∗).
In the proceeding corollary, if we take B = A, then it is clear Z t
so we obtain Proposition 2.10 from [14].
e∗∗ (A∗∗) = A∗∗, and
Theorem 2-6. Suppose that B is a weakly complete Banach space. Then we have
the following assertions.
(1) Let B be a Left Banach A − module and e′′ be a left mixed unit for B ∗∗. If
AB ∗∗ ⊆ B, then B is reflexive.
(2) Let B be a right Banach A − module and e′′ be a right mixed unit for B ∗∗.
If ZA∗∗(B ∗∗)A ⊆ B, then ZA∗∗(B ∗∗) = B.
Proof.
(1) Assume that b′′ ∈ B ∗∗. Since e′′ is also mixed unit for A∗∗, there is a
BAI (eα)α ⊆ A for A such that eα
b′′ in B ∗∗. Since AB ∗∗ ⊆ B, we have π∗∗∗
π∗∗∗
and so B is reflexive.
(e′′, b′′) =
(eα, b′′) ∈ B. Consequently
(e′′, b′′) = b′′ in B. Since B is a weakly complete, b′′ ∈ B,
(eα, b′′) w→ π∗∗∗
→ π∗∗∗
ℓ
ℓ
ℓ
w∗
→ e′′. Then π∗∗∗
(eα, b′′) w∗
ℓ
ℓ
7
(2) Since b′′ ∈ ZA∗∗(B ∗∗), we have π∗∗∗
ZA∗∗(B ∗∗)A ⊆ B, π∗∗∗
π∗∗∗
(b′′, e′′) = b′′ in B ∗∗. Since
(b′′, eα) w→
(b′′, e′′) = b′′ in B. It follows that b′′ ∈ B, since B is a weakly complete.
(cid:3)
(b′′, eα) ∈ B. Consequently we have π∗∗∗
→ π∗∗∗
r
r
r
r
r
(b′′, eα) w∗
A functional a′ in A∗ is said to be wap (weakly almost periodic) on A if the mapping
a → a′a from A into A∗ is weakly compact. The procceding definition to the equiva-
lent following condition, see [6, 14, 18].
For any two net (aα)α and (bβ)β in {a ∈ A : k a k≤ 1}, we have
limαlimβ < a′, aαbβ >= limβlimα < a′, aαbβ >,
whenever both iterated limits exist. The collection of all wap functionals on A is de-
noted by wap(A). Also we have a′ ∈ wap(A) if and only if < a′′b′′, a′ >=< a′′ob′′, a′ >
for every a′′, b′′ ∈ A∗∗.
Definition 2-7. Let B be a left Banach A − module. Then, b′ ∈ B ∗ is said to be left
weakly almost periodic functional if the set {πℓ(b′, a) : a ∈ A, k a k≤ 1} is relatively
weakly compact. We denote by wapℓ(B) the closed subspace of B ∗ consisting of all
the left weakly almost periodic functionals in B ∗.
The definition of the right weakly almost periodic functional (= wapr(B)) is the same.
By [18], the definition of wapℓ(B) is equivalent to the following
< π∗∗∗
ℓ
(a′′, b′′), b′ >=< πt∗∗∗t
ℓ
(a′′, b′′), b′ >
for all a′′ ∈ A∗∗ and b′′ ∈ B ∗∗. Thus, we can write
wapℓ(B) = {b′ ∈ B ∗ : < π∗∗∗
ℓ
(a′′, b′′), b′ >=< πt∗∗∗t
ℓ
(a′′, b′′), b′ >
f or all a′′ ∈ A∗∗, b′′ ∈ B ∗∗}.
Theorem 2-8. Suppose that B is a left Banach A − module. Consider the following
statements.
(1) B ∗A ⊆ wapℓ(B).
(2) AA∗∗ ⊆ ZB∗∗(A∗∗).
(3) AA∗∗ ⊆ AZB∗∗((A∗A)∗).
Then, we have (1) ⇔ (2) ⇐ (3).
Proof. (1) ⇒ (2)
Let (b′′
α)α ⊆ B ∗∗ such that b′′
α
w∗
→ b′′. Then for all a ∈ A and a′′ ∈ A∗∗, we have
< π∗∗∗
ℓ
(aa′′, b′′
=< a′′, π∗∗
ℓ (b′′
α, b′)a >
ℓ (b′′
(a′′, b′′), b′a) >
ℓ (b′′
α), b′ >=< aa′′, π∗∗
(a′′, b′′
(aa′′, b′′), b′) > .
α, b′a) >=< π∗∗∗
=< π∗∗∗
ℓ
ℓ
α, b′) >=< a′′, π∗∗
α), b′a) >→< π∗∗∗
ℓ
Hence aa′′ ∈ ZB∗∗ (A∗∗).
(2) ⇒ (1)
Let a ∈ A and b′ ∈ B ∗. Then
< π∗∗∗
ℓ
(a′′, b′′
α), b′a >=< aπ∗∗∗
ℓ
(a′′, b′′
α), b′ >=< π∗∗∗
ℓ
(aa′′, b′′
α), b′ >
8
=< πt∗∗∗t
It follow that b′a ∈ wapℓ(B).
(3) ⇒ (2)
Since AZB∗∗((A∗A)∗) ⊆ ZB∗∗(A∗∗), proof is hold.
α), b′ >=< πt∗∗∗t
(aa′′, b′′
ℓ
ℓ
(a′′, b′′
α), b′a > .
(cid:3)
In the proceeding theorem, if we take B = A, then we obtain Theorem 3.6 from [14]
and the same as proceeding theorem, we can claim the following assertions:
If B is a right Banach A − module, then for the following statements we have
(1) ⇔ (2) ⇐ (3).
(1) AB ∗ ⊆ wapr(B).
(2) A∗∗A ⊆ ZB∗∗(A∗∗).
(3) A∗∗A ⊆ ZB∗∗((A∗A)∗)A.
The proof of the this assertion is similar to proof of Theorem 2-8.
Corollary 2-9. Suppose that B is a Banach A − bimodule. Then if A is a left
[resp. right] ideal in A∗∗, then B ∗A ⊆ wapℓ(B) [resp. AB ∗ ⊆ wapr(B)].
Example 2-10. Suppose that 1 ≤ p ≤ ∞ and q is conjugate of p. We know that if
G is compact, then L1(G) is a two-sided ideal in its second dual of it. By proceeding
Theorem we have Lq(G) ∗ L1(G) ⊆ wapℓ(Lp(G)) and L1(G) ∗ Lq(G) ⊆ wapr(Lp(G)).
Also if G is finite, then Lq(G) ⊆ wapℓ(Lp(G))∩wapr(Lp(G)). Hence we conclude that
ZL1(G)∗∗(Lp(G)∗∗) = Lp(G) and ZLp(G)∗∗(L1(G)∗∗) = L1(G).
Theorem 2-11. We have the following assertions.
(1) Suppose that B is a left Banach A − module and b′ ∈ B ∗. Then b′ ∈ wapℓ(B)
ℓ (b′, ) : A → B ∗ is weak∗−to−weak
if and only if the adjoint of the mapping π∗
continuous.
(2) Suppose that B is a right Banach A−module and b′ ∈ B ∗. Then b′ ∈ wapr(B)
r (b′, ) : B → A∗ is weak∗−to−weak
if and only if the adjoint of the mapping π∗
continuous.
Proof.
(1) Assume that b′ ∈ wapℓ(B) and π∗
ℓ (b′, )∗ : B ∗∗ → A∗ is the adjoint of
ℓ (b′, ). Then for every b′′ ∈ B ∗∗ and a ∈ A, we have
π∗
ℓ (b′, a) > .
ℓ (b′, )∗b′′, a >=< b′′, π∗
< π∗
Suppose (b′′
α)α ⊆ B ∗∗ such that b′′
α
w∗
→ b′′ and a′′ ∈ A∗∗ and (aβ)β ⊆ A such
that aβ
w∗
→ a′′. By easy calculation, for all y′′ ∈ B ∗∗ and y′ ∈ B ∗, we have
< π∗
ℓ (y′, )∗, y′′ >= π∗∗
ℓ (y′′, y′).
Since b′ ∈ wapℓ(B),
< π∗∗∗
ℓ
(a′′, b′′
α), b′ >→< π∗∗∗
ℓ
(a′′, b′′), b′ > .
Then we have the following statements
lim
α
< a′′, π∗
ℓ (b′, )∗b′′
α >= lim
α
< a′′, π∗∗
ℓ (b′′
α, b′) >
= lim
α
< π∗∗∗
ℓ
(a′′, b′′
α), b′ >=< π∗∗∗
ℓ
9
(a′′, b′′), b′ >
=< a′′, π∗
ℓ (b′, )∗b′′ > .
It follow that the adjoint of the mapping π∗
weak continuous.
Conversely, let the adjoint of the mapping π∗
weak continuous. Suppose (b′′
for every a′′ ∈ A∗∗, we have
ℓ (b′, ) : A → B ∗ is weak∗ − to −
α)α ⊆ B ∗∗ such that b′′
α
ℓ (b′, ) : A → B ∗ is weak∗ − to −
w∗
→ b′′ and b′ ∈ B ∗. Then
lim
α
< π∗∗∗
ℓ
(a′′, b′′
α), b′ >= lim
α
< a′′, π∗∗
ℓ (b′′
α, b′) >
= lim
α
< a′′, π∗
ℓ (b′, )∗b′′
α >=< a′′, π∗
ℓ (b′, )∗b′′ >=< π∗∗∗
ℓ
It follow that b′ ∈ wapℓ(B).
(2) proof is similar to (1).
(a′′, b′′), b′ > .
(cid:3)
Corollary 2-12. Let A be a Banach algebra. Assume that a′ ∈ A∗ and Ta′ is the
linear operator from A into A∗ defined by Ta′ a = a′a. Then, a′ ∈ wap(A) if and only
if the adjoint of Ta′ is weak∗ − to − weak continuous. So A is Arens regular if and
only if the adjoint of the mapping Ta′ a = a′a is weak∗ − to − weak continuous for
every a′ ∈ A∗.
3. Lw∗w-property and Rw∗w-property
In this section, we introduce the new definition as Lef t − weak∗ − to − weak property
and Right−weak∗−to−weak property for Banach algebra A and make some relations
between these concepts and topological centers of module actions. As some conclu-
sion, we have ZL1(G)∗∗(M (G)∗∗) 6= M (G)∗∗ where G is a locally compact group. If G
is finite, we have ZM(G)∗∗ (L1(G)∗∗) = L1(G)∗∗ and ZL1(G)∗∗(M (G)∗∗) = M (G)∗∗.
Definition 3-1. Let B be a left Banach A − module. We say that a ∈ A has
Lef t − weak∗ − to − weak property (= Lw∗w− property) with respect to B, if for
w∗
→ 0 implies ab′
α
w→ 0.
If every a ∈ A has Lw∗w− property
all (bα)α ⊆ B ∗, ab′
α
with respect to B, then we say that A has Lw∗w− property with respect to B. The
definition of the Right − weak∗ − to − weak property (= Rw∗w− property) is the
same.
We say that a ∈ A has weak∗ − to − weak property (= w∗w− property) with respect
to B if it has Lw∗w− property and Rw∗w− property with respect to B.
If a ∈ A has Lw∗w− property with respect to itself, then we say that a ∈ A has
Lw∗w− property.
For proceeding definition, we have some examples and remarks as follows.
a) If B is Banach A-bimodule and reflexive, then A has w∗w−property with respect
to B. Then
i) L1(G), M (G) and A(G) have w∗w−property when G is finite.
ii) Let G be locally compact group. L1(G) [resp. M (G)] has w∗w−property [resp.
10
Lw∗w− property ]with respect to Lp(G) whenever p > 1.
b) Suppose that B is a left Banach A − module and e is left unit element of A such
that eb = b for all b ∈ B. If e has Lw∗w− property, then B is reflexive.
c) If S is a compact semigroup, then C+(S) = {f ∈ C(S) : f > 0} has w∗w−property.
Theorem 3-2. Suppose that B is a Banach A − bimodule. Then we have the
following assertions.
(1) If A∗∗ = a0A∗∗ [resp. A∗∗ = A∗∗a0] for some a0 ∈ A and a0 has Rw∗w−
property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗.
(2) If B ∗∗ = a0B ∗∗ [resp. B ∗∗ = B ∗∗a0] for some a0 ∈ A and a0 has Rw∗w−
property [resp. Lw∗w− property] with respect to B, then ZA∗∗(B ∗∗) = B ∗∗.
Proof.
(1) Suppose that A∗∗ = a0A∗∗ for some a0 ∈ A and a0 has Rw∗w−
w∗
→ b′′. Then for all a ∈ A and
α)α ⊆ B ∗∗ such that b′′
α
property. Let (b′′
b′ ∈ B ∗, we have
α, π∗
α, b′) w∗
< π∗∗
ℓ (b′′
α, b′), a >=< b′′
ℓ (b′, a) >→< b′′, π∗
ℓ (b′, a) >=< π∗∗
ℓ (b′′, b′), a >,
ℓ (b′′
w∗
α, b′)a0
it follow that π∗∗
→
ℓ (b′′, b′)a0.
ℓ (b′′, b′)a0. Since a0 has Rw∗w− property, π∗∗
π∗∗
Now let a′′ ∈ A∗∗. Then there is x′′ ∈ A∗∗ such that a′′ = a0x′′ consequently
we have
ℓ (b′′, b′). Also we can write π∗∗
ℓ (b′′
w→ π∗∗
α, b′)a0
→ π∗∗
ℓ (b′′
< π∗∗∗
ℓ
(a′′, b′′
α), b′ >=< a′′, π∗∗
ℓ (b′′
α, b′) >=< x′′, π∗∗
α, b′)a0 >
→< x′′, π∗∗
ℓ (b′′, b′)a0 >=< π∗∗∗
ℓ
(a′′, b′′
ℓ (b′′
α), b′ > .
We conclude that a′′ ∈ ZB∗∗(A∗∗). Proof of the next part is the same as the
proceeding proof.
(2) Let B ∗∗ = a0B ∗∗ for some a0 ∈ A and a0 has Rw∗w− property with respect
w∗
→ a′′. Then for all b ∈ B, we
α)α ⊆ A∗∗ such that a′′
α
to B. Assume that (a′′
have
< π∗∗
r (a′′
α, b′), b >=< a′′
α, π∗∗
r (b′, b) >→< a′′, π∗∗
r (b′, b) >=< π∗∗
r (a′′, b′), b > .
α, b′) w∗
r (a′′
→ π∗∗
We conclude that π∗∗
π∗∗
r (a′′, b′)a0. Since a0 has Rw∗w− property with respect to B, π∗∗
π∗∗
r (a′′, b′)a0.
Now let b′′ ∈ B ∗∗. Then there is x′′ ∈ B ∗∗ such that b′′ = a0x′′. Hence, we
have
r (a′′, b′) then we have π∗∗
α, b′)a0
α, b′)a0
r (a′′
r (a′′
w∗
→
w→
< π∗∗∗
r
(b′′, a′′
=< x′′, π∗∗
r (a′′
r (a′′
α), b′ >=< b′′, π∗∗
α, b′)a0 >→< x′′, π∗∗
α, b′) >=< a0x′′, π∗∗
r (a′′
r (a′′, b′)a0 >=< b′′, π∗∗
α, b′) >
r (a′′, b′) >
=< π∗∗∗
r
(b′′, a′′), b′ > .
It follow that b′′ ∈ ZA∗∗(B ∗∗). The next part is similar to the proceeding
proof.
(cid:3)
11
i) Let G be a locally compact group. Since M (G) is a Banach
Example 3-3.
L1(G)-bimodule and the unit element of M (G) has not Lw∗w− property or Rw∗w−
property, by Theorem 2-3, ZL1(G)∗∗(M (G)∗∗) 6= M (G)∗∗.
ii) If G is finite, then by Theorem 2-3, we have ZM(G)∗∗(L1(G)∗∗) = L1(G)∗∗ and
ZL1(G)∗∗(M (G)∗∗) = M (G)∗∗.
Assume that B is a Banach A − bimodule. We say that B factors on the left (right)
with respect to A if B = BA (B = AB). We say that B factors on both sides, if
B = BA = AB.
Theorem 3-4. Suppose that B is a Banach A − bimodule and A has a BAI. Then
we have the following assertions.
(1) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w−
property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗.
(2) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w−
property [resp. Lw∗w− property] with respect B, then ZA∗∗(B ∗∗) = B ∗∗.
Proof.
(1) Assume that B ∗ factors on the left and A has Rw∗w− property. Let
α)α ⊆ B ∗∗ such that b′′
α
w∗
→ b′′. Since B ∗A = B ∗, for all b′ ∈ B ∗ there are
(b′′
x ∈ A and y′ ∈ B ∗ such that b′ = y′x. Then for all a ∈ A, we have
< π∗∗
=< b′′
ℓ (b′′
α, π∗
α, y′)x, a >=< b′′
α, π∗
ℓ (b′, a) >→< b′′, π∗
ℓ (y′, a)x >=< π∗∗
ℓ (b′, a) >=< π∗∗
α, b′), a >
ℓ (b′′, y′)x, a > .
ℓ (b′′
Thus, we conclude that π∗∗
property, π∗∗
ℓ (b′′
w→< π∗∗
ℓ (b′′
(a′′, b′′
α, y′)x
α), b′ >=< a′′, π∗∗
ℓ (b′′, y′)x. Now let b′′ ∈ A∗∗. Then
α, y′)x >
α, b′) >=< a′′, π∗∗
ℓ (b′′
< π∗∗∗
ℓ
α, y′)x
w∗
→< π∗∗
ℓ (b′′, y′)x. Since A has Rw∗w−
→< a′′, π∗∗
ℓ (b′′, y′)x >=< π∗∗∗
ℓ
ℓ (b′′
(a′′, b′′), b′ > .
It follow that a′′ ∈ ZB∗∗(A∗∗) = A∗∗.
If B ∗ factors on the right and A has Lw∗w− property, then proof is the same
as preceding proof.
(2) Let B ∗ factors on the left with respect to A and A has Rw∗w− property with
w∗
respect to B. Assume that (a′′
→ a′′. Since B ∗A = B,
for all b′ ∈ B ∗ there are x ∈ A and y′ ∈ B ∗ such that b′ = y′x. Then for all
b ∈ B, we have
α)α ⊆ A∗∗ such that a′′
α
< π∗∗
r (a′′
α, y′)x, b >=< π∗∗
=< a′′, π∗
r (a′′
r (b′, b) >=< π∗∗
α, b′), b >=< a′′
α, π∗
r (a′′, y′)x, b > .
r (b′, b) >
Consequently π∗∗
respect to B, π∗∗
have
r (a′′
r (a′′
α, y′)x
α, y′)x
w∗
→ π∗∗
w→ π∗∗
r (a′′, y′)x. Since A has Rw∗w− property with
r (a′′, y′)x. It follow that for all b′′ ∈ B ∗∗, we
< π∗∗∗
r
(b′′, a′′
α), b′ >=< b′′, π∗∗
=< π∗∗∗
r (a′′
(b′′, a′′), b′ > .
r
α, y′)x >→< b′′, π∗∗
r (a′′, y′)x >
12
Thus we conclude that b′′ ∈ ZA∗∗(B ∗∗).
The proof of the next assertions is the same as proceeding proof.
(cid:3)
Theorem 3-5. Suppose that B is a Banach A−bimodule. Then we have the following
assertions.
(1) If a0 ∈ A has Rw∗w− property with respect to B, then a0A∗∗ ⊆ ZB∗∗(A∗∗)
and a0B ∗ ⊆ wapℓ(B).
(2) If a0 ∈ A has Lw∗w− property with respect to B, then A∗∗a0 ⊆ ZB∗∗(A∗∗)
and B ∗a0 ⊆ wapℓ(B).
(3) If a0 ∈ A has Rw∗w− property with respect to B, then a0B ∗∗ ⊆ ZA∗∗(B ∗∗)
and B ∗a0 ⊆ wapr(B).
(4) If a0 ∈ A has Lw∗w− property with respect to B, then B ∗∗a0 ⊆ ZA∗∗(B ∗∗)
and a0B ∗ ⊆ wapr(B).
Proof.
(1) Let (b′′
α)α ⊆ B ∗∗ such that b′′
α
w∗
→ b′′. Then for all a ∈ A and b′ ∈ B ∗,
we have
< π∗∗
ℓ (b′′
α, b′)a0, a >=< π∗∗
ℓ (b′′
α, b′), a0a >=< b′′
α, π∗
ℓ (b′, a0a) >
→< b′′, π∗
ℓ (b′, a0a) >=< π∗∗
ℓ (b′′, b′)a0, a > .
w∗
→ π∗∗
ℓ (b′′
It follow that π∗∗
with respect to B, π∗∗
We conclude that a0a′′ ∈ ZB∗∗(A∗∗) so that a0A∗∗ ∈ ZB∗∗(A∗∗). Since
ℓ (b′′, b′)a0 = π∗∗
π∗∗
ℓ (b′′, b′)a0. Since a0 has Rw∗w− property
w→ π∗∗
ℓ (b′′, b′a0), a0B ∗ ⊆ wapℓ(B).
α, b′)a0
ℓ (b′′
ℓ (b′′, b′)a0.
α, b′)a0
(2) proof is similar to (1).
(3) Assume that (a′′
α)α ⊆ A∗∗ such that a′′
α
w∗
→ a′′. Let b ∈ B and b′ ∈ B ∗. Then
we have
< π∗∗
r (a′′
α, b′)a0, b >=< π∗∗
r (a′′
α, b′), a0b >=< a′′
α, π∗
r (b′, a0b) >
→< a′′, π∗
r (b′, a0b) >=< π∗∗
r (a′′, b′)a0, b > .
Thus we conclude π∗∗
erty with respect to B, π∗∗
have
r (a′′
α, b′)a0
r (a′′
r (a′′, b′)a0. Since a0 has Rw∗w− prop-
r (a′′, b′)a0. If b′′ ∈ B ∗∗, then we
w→ π∗∗
w∗
→ π∗∗
α, b′)a0
< π∗∗∗
r
(a0b′′, a′′
α), b′ >=< a0b′′, π∗∗∗
r
(a′′
α, b′) >=< b′′, π∗∗
r (a′′
α, b′)a0 >
=< b′′, π∗∗
r (a′′
α, b′)a0 >=< π∗∗∗
r
(a0b′′, a′′), b′ > .
It follow that a0b′′ ∈ ZA∗∗(B ∗∗). Consequently we have a0B ∗∗ ∈ ZA∗∗(B ∗∗).
The proof of the next assertion is clear.
(4) Proof is similar to (3).
(cid:3)
13
Theorem 3-6. Let B be a Banach A − bimodule. Then we have the following
assertions.
(1) Suppose
lim
α
lim
β
< b′
β, bα >= lim
β
lim
α
< b′
β, bα >,
for every (bα)α ⊆ B and (b′
Rw∗w− property with respect to B.
β)β ⊆ B ∗. Then A has Lw∗w− property and
(2) If for some a ∈ A,
lim
α
lim
β
< ab′
β, bα >= lim
β
lim
α
< ab′
β, bα >,
for every (bα)α ⊆ B and (b′
respect to B. Also if for some a ∈ A,
β)β ⊆ B ∗, then a has Rw∗w− property with
lim
α
lim
β
< b′
βa, bα >= lim
β
lim
α
< b′
βa, bα >,
for every (bα)α ⊆ B and (b′
respect to B.
β)β ⊆ B ∗, then a has Lw∗w− property with
Proof.
(1) Assume that a ∈ A such that ab′
β
w∗
→ 0 where (b′
β)β ⊆ B ∗. Let b′′ ∈ B ∗∗
and (bα)α ⊆ B such that bα
w∗
→ b′′. Then
lim
β
< b′′, ab′
β >= lim
β
lim
α
< bα, ab′
β >= lim
β
lim
α
< ab′
β, bα >
= lim
α
lim
β
< ab′
β, bα >= 0.
We conclude that ab′
β
has Rw∗w− property.
w→ 0, so A has Lw∗w− property. It also easy that A
(2) Proof is easy and is the same as (1).
(cid:3)
α = b′aα [resp. b′
Definition 3-7. Let B be a left Banach A − module. We say that B ∗ strong factors
on the left [resp. right] if for all (b′
α)α ⊆ B ∗ there are (aα)α ⊆ A and b′ ∈ B ∗ such
that b′
If B ∗ strong factors on the left and right, then we say that B ∗ strong factors on the
both side.
It is clear that if B ∗ strong factors on the left [resp. right], then B ∗ factors on the
left [resp. right].
α = aαb′] where (aα)α has limit the weak∗ topology in A∗∗.
Theorem 3-8. Suppose that B is a Banach A − bimodule. Assume that AB ∗ ⊆
wapℓB. If B ∗ strong factors on the left [resp. right], then A has Lw∗w− property
[resp. Rw∗w− property ] with respect to B.
Proof. Let (b′
are (aα)α ⊆ A and b′ ∈ B ∗ such that b′
α)α ⊆ B ∗ such that ab′
α
w∗
→ 0. Since B ∗ strong factors on the left, there
α = b′aα. Let b′′ ∈ B ∗∗ and (bβ)β ⊆ B such
14
that bβ
w∗
→ b′′. Then we have
lim
α
< b′′, ab′
α >= lim
α
lim
β
< bβ, ab′
α >= lim
α
lim
β
< ab′
α, bβ >
= lim
α
lim
β
< ab′aα, bβ >= lim
α
lim
β
< ab′, aαbβ >
= lim
β
lim
α
< ab′, aαbβ >= lim
β
lim
α
< ab′
α, bβ >= 0
It follow that ab′
α
w→ 0.
Problems .
(cid:3)
(1) Suppose that B is a Banach A − bimodule. If B is left or right factors with
respect to A, dose A has Lw∗w−property or Rw∗w−property, respectively?
(2) Suppose that B is a Banach A − bimodule. Let A has Lw∗w−property with
respect to B. Dose ZB∗∗(A∗∗) = A∗∗?
References
1. R. E. Arens, The adjoint of a bilinear operation, Proc. Amer. Math. Soc. 2 (1951), 839-848.
2. N. Arikan, A simple condition ensuring the Arens regularity of bilinear mappings, Proc. Amer.
Math. Soc. 84 (4) (1982), 525-532.
3. J. Baker, A.T. Lau, J.S. Pym Module homomorphism and topological centers associated with
weakly sequentially compact Banach algebras, Journal of Functional Analysis. 158 (1998), 186-
208.
4. F. F. Bonsall, J. Duncan, Complete normed algebras, Springer-Verlag, Berlin 1973.
5. H. G. Dales, A. Rodrigues-Palacios, M.V. Velasco, The second transpose of a derivation, J.
London. Math. Soc. 2 64 (2001) 707-721.
6. H. G. Dales, Banach algebra and automatic continuity, Oxford 2000.
7. N. Dunford, J. T. Schwartz, Linear operators.I, Wiley, New york 1958.
8. M. Eshaghi Gordji, M. Filali, Arens regularity of module actions, Studia Math. 181 3 (2007),
237-254.
9. M. Eshaghi Gordji, M. Filali, Weak amenability of the second dual of a Banach algebra, Studia
Math. 182 3 (2007), 205-213.
10. K. Haghnejad Azar, A. Riazi, Arens regularity of bilinear forms and unital Banach module space,
(Submitted).
11. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol I 1963.
12. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol II 1970.
13. A. T. Lau, V. Losert, On the second Conjugate Algebra of locally compact groups, J. London
Math. Soc. 37 (2)(1988), 464-480.
14. A. T. Lau, A. U lger, Topological center of certain dual algebras, Trans. Amer. Math. Soc. 348
(1996), 1191-1212.
15. S. Mohamadzadih, H. R. E. Vishki, Arens regularity of module actions and the second adjoint
of a derivation, Bulletin of the Australian Mathematical Society 77 (2008), 465-476.
16. M. Neufang, Solution to a conjecture by Hofmeier-Wittstock, Journal of Functional Analysis.
217 (2004), 171-180.
17. M. Neufang, On a conjecture by Ghahramani-Lau and related problem concerning topological
center, Journal of Functional Analysis. 224 (2005), 217-229.
18. J. S. Pym, The convolution of functionals on spaces of bounded functions, Proc. London Math
Soc. 15 (1965), 84-104.
19. A. U lger, Arens regularity sometimes implies the RNP, Pacific Journal of Math. 143 (1990),
377-399.
20. A. U lger, Some stability properties of Arens regular bilinear operators, Proc. Amer. Math. Soc.
(1991) 34, 443-454.
15
21. A. U lger, Arens regularity of weakly sequentialy compact Banach algebras, Proc. Amer. Math.
Soc. 127 (11) (1999), 3221-3227.
22. A. U lger, Arens regularity of the algebra A ⊗B, Trans. Amer. Math. Soc. 305 (2) (1988) 623-639.
23. P. K. Wong, The second conjugate algebras of Banach algebras, J. Math. Sci. 17 (1) (1994),
15-18.
24. N. J. Young Theirregularity of multiplication in group algebra Quart. J. Math. Soc., 24 (2)
(1973), 59-62.
25. Y. Zhing, Weak amenability of module extentions of Banach algebras Trans. Amer. Math. Soc.
354 (10) (2002), 4131-4151.
Department of Mathematics, Amirkabir University of Technology, Tehran, Iran
Email address: [email protected]
|
1906.09599 | 1 | 1906 | 2019-06-23T15:55:58 | $L_p$ functional Busemann-Petty centroid inequality | [
"math.FA",
"math.MG"
] | If $K\subset\mathbb{R}^n$ is a convex body and $\Gamma_pK$ is the $p$-centroid body of $K$, the $L_p$ Busemann-Petty centroid inequality states that $\vol(\Gamma_pK) \geq \vol(K)$, with equality if and only if $K$ is an ellipsoid centered at the origin. In this work, we prove inequalities for a type of functional $r$-mixed volume for $1 \leq r < n$, and establish as a consequence, a functional version of the $L_p$ Busemann-Petty centroid inequality. \keywords{Convex body, Moment body, Busemann-Petty centroid} } | math.FA | math |
Lp functional Busemann-Petty centroid inequality
J. E. Haddad∗, C. H. Jim´enez†,
L. A. Silva‡
Abstract
If K ⊂ Rn is a convex body and ΓpK is the p-centroid body of K, the Lp Busemann-
Petty centroid inequality states that vol(ΓpK) ≥ vol(K), with equality if and only if K is an
ellipsoid centered at the origin. In this work, we prove inequalities for a type of functional
r-mixed volume for 1 ≤ r < n, and establish as a consequence, a functional version of the Lp
Busemann-Petty centroid inequality.
Keywords. Convex body, Moment body, Busemann-Petty centroid
1
Introduction
The study of affine isoperimetric inequalities on one side and affine Sobolev inequalities for
functions on Rn on the other is connected to a great extent. The equivalence of the classical
isoperimetric inequality and the classical L1 Sobolev inequality has been known for quite some
time (see for example[2, 33, 10, 6, 29, 24, 9]). Following this path Zhang in [34] established the
equivalence of an affine L1 Sobolev inequality with the Petty Projection inequality for convex
bodies. Some time after, along with Lutwak and Yang continued in this direction obtaining
Lp versions of the mentioned equivalence. These authors developed around the same time a
rich theory of geometrical inequalities for centroid bodies and established Lp extensions of
many other fundamental parameters in Convex Geometry, such as mixed volume and surface
area.
On top of the strong connections mentioned above, other geometrical inequalities of
isoperimetric flavour like the Busemann-Petty centroid inequality or Blaschke-Santal´o, among
others, have been fundamental in the study of several inequalities of Sobolev type,
like
Lp log-Sobolev, Gagliardo-Nirenberg, Sobolev trace or weighted Sobolev inequalities (e.g
[15, 8, 13, 14, 11, 12]). It is important to notice that in many of the works mentioned above,
where the Busemann-Petty centroid inequality was used to recover some known results for
Sobolev type inequalities, this inequality provided a more direct approach. This approach
often went around the use (in their original proofs) of other well known tools in the area of
convex geometric analysis like the Minkowski problem or the theory of mixed or dual mixed
volumes.
∗Departamento de Matem´atica, ICEx, Universidade Federal de Minas Gerais, 30123-970, Belo Horizonte, Brasil;
e-mail: [email protected]
†Departamento de Matem´atica, Pontif´ıcia Universidade Cat´olica do Rio de Janeiro, 22451-900 Rio de Janeiro,
Brasil [corresponding author]; e-mail: [email protected]
‡Pontif´ıcia Universidade Cat´olica do Rio de Janeiro, Departamento de Matem´atica, 22451-900, Rio de
leti-
Janeiro, Brasil. Professora do Ensino B´asico, T´ecnico e Tecnol´ogico no IFMG - Campus Bambu´ı; e-mail:
[email protected]
1
In this work we continue with this line of research. We obtain a family of inequalities for
functions on Rn, inequalities of Sobolev type, and that in particular recover the Lp Busemann-
Petty centroid inequality for convex bodies in Rn. Our main inequality is presented in the
form of a functional mixed volume inequality.
2
Theorem 1.1. Let f be a C 1 function and g a continuous non-negative function, both with
compact support in Rn, then for 1 ≤ r < n, q = nr
n−r and λ ∈(cid:16) n
n+p , 1(cid:17) ∪ (1, ∞),
[(n+p)(λ−1)+p]r
ZRn(cid:18)ZRn
g(y)h∇f (x), yipdy(cid:19)r/p
dx ≥ Cn,p,λg
1
np(λ−1)
− λr
g
λ
(λ−1)n
f r
q.
(1)
The sharp constant Cn,p,λ is computed in Section 3 and equality is attained if and only if f
and g have the following forms
g(x) = aGp,λ(Ax2)
f (x) = bFr(Ax2)
for positive constants a, b, A ∈ GLn(R), Gp,λ : R+ → R defined by
Gp,λ(t) =
and
(1 + tp)
(1 − tp)
1
λ−1
1
λ−1
+
Fr(t) = (1 + t
if λ ∈(cid:16) n
n+p , 1(cid:17)
if λ > 1,
r
r−1 )1− r
n .
2 Some notations and tools from Convex Geometry
In order to show the intrinsic geometric nature of inequality (1), and in particular, its relation
to the Lp Busemann-Petty centroid inequality, let us first recall some basic definitions. A
convex body is a convex set K ⊂ Rn which is compact and has non-empty interior. For a
convex body K, its support function hK , which uniquely characterizes it, is defined as
hK (x) = max{hx, yi : y ∈ K}.
If K contains the origin in the interior, then we also have the gauge k · kK and radial rK (·)
functions of K defined respectively as
kykK := inf{λ > 0 : y ∈ λK} ,
y ∈ Rn \ {0} ,
rK(y) := max{λ > 0 : λy ∈ K} ,
y ∈ Rn \ {0} .
Clearly, kykK = 1
rK (y) .
For a convex body K ⊂ Rn and p ≥ 1, its Lp-moment and Lp-centroid bodies, denoted
by MpK and ΓpK, are defined by their support functions
hMpK(x)p =ZK
hx, yipdy,
and
hΓpK(x)p =
1
vol(K)cn,p ZK
hx, yipdy,
(2)
respectively, where cn,p = ωn+p
of Rm. The Lp Busemann-Petty centroid inequality states that
ω2ωnωp−1
and ωm is the m-dimensional volume of the unit ball B
vol(ΓpK) ≥ vol(K)
or
vol(MpK) ≥ cn/p
n,p vol(K)
n+p
p ,
(3)
3
in terms of the moment body MpK. Equality holds in (3) if and only if K is a 0-symmetric
ellipsoid.
Centroid bodies for p = 1 can be found for the first time in a work of Blaschke [3]
whereas the respective Busemann-Petty centroid inequality for p = 1 is due to Petty [31].
The Lp version of centroid bodies above was introduced by Lutwak and Zhang [23], while (3)
was obtained by Lutwak, Yang and Zhang in [19]. For the history of the Busemann-Petty
centroid inequality and a comprehensive introduction on centroid and moment bodies we refer
to Chapter 10 in [32].
The theory of mixed volumes, first developed by Minkowski [28, 27], is one of the pillars
of the Brunn-Minkowski theory, it provides us with a unified approach to the study of several
of the most important parameters in Convex Geometry, such as volume, mean width, surface
area, among others. At the same time, it has been fundamental in many other problems
ranging from characterization of special families of convex bodies to establish new isoperi-
metric inequalities, we refer to [32, 4] for a comprehensive introduction to the theory of mixed
volumes. There are several extensions of the concept of mixed volume, in this work we will
focus mainly in the dual mixed volume and the Lp extension of the mixed volume, concepts
belonging to the dual and Lp Brunn-Minkowski theory respectively. Regarding the latter we
have the following Lp extension of mixed volume, for some background on this we refer to
[18] and to [22] and the references therein.
For r ≥ 1, the Lr-mixed volume Vr(K, L) of convex bodies K and L is defined by
Vr(K, L) =
r
n
lim
ε→0
vol(K +r ε ·r L) − vol(K)
ε
,
where K +r ε ·r L is the convex body defined by:
hK+rε·rL(x)r = hK (x)r + εhL(x)r,
∀x ∈ Rn.
One of the main aspects of the mixed volume is that it has an integral representation. As
in the classical case for the Lr version it is known (see [18]) that there exists a unique finite
positive Borel measure Sr(K, .) on Sn−1 such that
for each convex body L.
Vr(K, L) =
1
nZSn−1
hL(u)rdSr(K, u),
(4)
If 1 ≤ r < ∞ and K, L are convex bodies in Rn containing the origin as interior point, we
can find also in [18] that
Vr(K, L) ≥ vol(K)
n−r
n vol(L)
r
n ,
(5)
with equality if and only if K and L are dilates of each other. Combining inequalities (5) and
(3), we obtain:
Vr(L, MpK) ≥ cr/p
n,p vol(L)
n−r
n vol(K)
(n+p)r
np
.
(6)
Taking L = MpK in (6), we recover (3), hence (6) is an equivalent formulation for the Lp
Busemann-Petty centroid inequality. This and similar geometric inequalities for mixed vol-
umes involving centroid and projection bodies were already considered in [17]. The main
result, Theorem 1.1 is a functional version of inequality (6), replacing the sets L, K by func-
tions f, g.
In order to establish a functional version of (6) and considering the integral representation
of the geometric Lr mixed volume (4), let us recall the following result obtained by Lutwak,
Yang and Zhang, where they introduced the concept of surface area measure of a Sobolev
function.
The Lr surface area measure of a function f : Rn → R with Lr weak derivative is given
by:
4
Lemma 2.1 (Lemma 4.1 of [22]). Given 1 ≤ r < ∞ and a function f : Rn → R with Lr
weak derivative, there exists a unique finite Borel measure Sr(f, .) on Sn−1 such that
ZRn
φ(−∇f (x))rdx =ZSn−1
φ(u)rdSr(f, u),
(7)
for every non-negative continuous function φ : Rn → R homogeneous of degree 1. If f is
not equal to a constant function almost everywhere, then the support of Sr(f, .) cannot be
contained in any n − 1 dimensional linear subspace.
Conversely, for a convex body L the function fL(x) = F (kxkL) satisfies Sr(f, .) = Sr(L, .)
if F is any function F : R+ → R+ satisfying
(see [22]). By the Sobolev inequality we have
tn−1F ′(t)rdt = 1
Z ∞
0
fL(x)
ZRn
nr
n−r
n−r dx ≤ c
s
nr
(nωn)
vol(L)
n
n−r
ωn
where cs is the sharp constant in the Sobolev inequality on Rn, and there is equality when
F (t) = aFr(t) with a, b > 0, where
Fr(t) = (1 + t
r
r−1 )1− r
n .
The function F (x2) is an extremal function of the euclidean Lr Sobolev inequality on Rn.
In view of identity (7), for any f and L such that Sr(f, .) = Sr(L, .), we have
Vr(L, K) =
1
nZRn
hK(−∇f (x))rdx.
This motivates the following definition.
Definition 2.2. Given 1 ≤ r < ∞ and a function f : Rn → R with Lr weak derivative, we
define
Vr(f, K) =
hK (−∇f (x))rdx
1
nZRn
The Lp Sobolev inequality for general norms was proved in [7] and [1] and can be stated
as a mixed volume inequality for functions as follows:
Theorem 2.3. If f is a C 1 function with compact support in Rn and K is an origin-
symmetric convex body, then for 1 < r < n and q = nr
n−r
Vr(f, K) ≥ cr
1kf kr
q vol(K)
r
n ,
(8)
where c1 is the optimal constant and equality holds in (8) if and only if f (x) = aFr(bkxkK )
for some a, b > 0. Taking f (x) = Fr(kxkL) we recover inequality (5).
Theorem 2.3 was originally proved using an innovative approach based on optimal trans-
portation of mass in [7] and in [1] using Convex Symmetrization.
In Section 4 we give an alternative, simpler and elementary proof of this inequality using
the tools developed in [20]. Some of the tools we are using here, specially those contained in
[22], have been used in the study of Sobolev type inequalities. Their approach is often based
on a functional extension of the so-called LYZ body and other known geometric inequalities
for projection and polar projection bodies (see Subsection 10.15 in [32] and references therein
for more on this).
Let us go back to the definition of the moment body (2), it has been noticed that hMpK is
a convex function regardless of the set K (see e.g. Chapter 5 in [5]). This observation allows
us to make the following definition:
5
Definition 2.4. If g is a non-negative measurable function with compact support, we define
the convex body Mpg by
hMpg(ξ)p =ZRn
g(x)hx, ξipdx.
The left-hand side of (1) has then a geometric meaning:
Vr(f, Mpg) =
1
nZRn(cid:18)ZRn
g(y)h∇f (x), yipdy(cid:19)r/p
dx.
If K is a convex body and g(x) = G(kxkK ) for any non-negative continous function
G : R+ → R with compact support, it is not hard to verify using polar coordinates that
Mpg =(cid:18)(n + p)Z ∞
0
tn+p−1G(t)dt(cid:19)1/p
MpK.
Our main result (Theorem 1.1) is a consequence of Theorem 2.3, and Theorem 2.5 below:
Theorem 2.5. If g is a non-negative function with compact support in Rn, then, for each
λ ∈(cid:16) n
n+p , 1(cid:17) ∪ (1, ∞), we have that
p
vol(Mpg)
n ≥ cn,pan,p,λg
1
(n+p)(λ−1)+p
(λ−1)n
− λp
g
λ
(λ−1)n
,
(9)
where an,p,λ is given by the Lemma (3.4).
Let Gp,λ : R+ → R be defined by
Gp,λ(t) =( (1 + tp)
(1 − tp)
1
λ−1
1
λ−1
+
if λ < 1
if λ > 1,
then taking g(x) = Gp,λ(kxkK ) in (9) we recover (3).
Equality holds in (9) if and only if g(x) = aGp,λ(A.x2) for any a > 0 and A ∈ Gln(Rn).
Even though Theorem 2.5 contains the geometric core of the main Theorem 1.1, the term
vol(Mpg) cannot be expressed in terms of g in an elementary way, as Vr(f, Mpg) does. This
is the reason why we need to combine it with Theorem 2.3 to obtain a functional inequality.
Let us note that Theorem 1.1 cannot be regarded as a functional mixed volume inequality
in full generality since it can only be applied to a function f and the centroid/moment body
of another function g. We refer the interested reader to review the works of Milman and
Rotem [26, 25] where they have defined a functional extension of mixed volumes and have
extended some of their main properties to a functional setting.
We should finally also mention other related extension of the Busemann-Petty centroid
inequality obtained by Paouris and Pivovarov in [30] where the authors obtained randomized
versions of this and other important isoperimetric inequalities.
The rest of the paper is organized as follows: In Section 3 we shall prove some preliminary
results, including an extension of the Lp Busemann-Petty centroid inequality, to compact
domains. Then in Section 4 we prove Theorems 2.3 and 2.5.
We hope this work shed some more light into the deep connection between isoperimetric
and functional inequalities.
6
3 Preliminary results
In order to prove our main result, Theorem 1.1, we consider two cases: r = 1 and 1 < r < n.
For r = 1, inequality (5) holds for more general sets. As in [34], a compact domain is the
closure of a bounded open set.
Lemma 3.1 (Lemma 3.2 of [34]). If M is a compact domain with piecewise C 1 boundary and
K a convex body in Rn, then,
with equality if and only if M and K are homothetic.
V (M, K)n ≥ vol(M )n−1 vol(K),
In the same spirit, the next lemma shows that the Lp-Busemann-Petty Centroid inequality
remains valid for a compact domain:
Lemma 3.2. If M is a compact domain, then
vol(ΓpM ) ≥ vol(M ).
(10)
Equality holds in (10) if and only if M is a 0-symmetric ellipsoid.
Proof. For a compact domain M and ξ ∈ Sn−1, we define the set
Consider δ(t) = tn
n , for t ≥ 0, and the star set SM defined by its radial function
Lξ = {t ∈ [0, ∞) : tξ ∈ M }.
ρSM (ξ) = δ−1(µ(δ(Lξ))),
where µ denotes the one dimensional Lebesgue measure of Lξ. It is easy to see that vol(SM ) =
vol(M ). Also, let s = δ(t) = tn
n , then ds = tn−1dt. For x ∈ Rn, we have:
ZM
hx, yipdy =ZSn−1ZLξ
=ZSn−1ZLξ
=ZSn−1Zδ(Lξ)
n ZSn−1
= n
p
hx, tξiptn−1dtdξ
hx, ξiptptn−1dtdξ
hx, ξip(ns)
p
n dsdξ
hx, ξipZδ(Lξ)
p
n dsdξ
s
On the other hand, we have
ZSM
0
hx, yipdy =ZSn−1Z ρSM (ξ)
=ZSn−1Z ρSM (ξ)
=ZSn−1Z δ(ρSM (ξ))
0
0
hx, tξiptn−1dtdξ
hx, ξiptptn−1dtdξ
hx, ξip(ns)
p
n dsdξ
= n
p
n ZSn−1
hx, ξipZ µ(δ(Lξ ))
0
p
n dsdξ.
s
By the Bathtub principle (see Theorem 1.14, pag. 28 of [16]) we have
7
therefore,
p
n ds
s
p
0
s
n ds ≥Z µ(δ(Lξ ))
Zδ(Lξ)
hx, yipdy ≥ZSM
ZM
Since vol(SM ) = vol(M ), we obtain hΓpM (x)p ≥ hΓpSM (x)p, whence ΓpM ⊃ ΓpSM and
vol(ΓpM ) ≥ vol(ΓpSM ). We conclude,
hx, yipdy.
(11)
vol(ΓpM ) ≥ vol(ΓpSM ) ≥ vol(SM ) = vol(M ).
If M is a compact domain attaining equality in (10), then equality in (11) implies
µ(δ(Lξ)) = δ(Lξ) for a.e ξ, meaning that M is a star body. We conclude the proof recalling
the equality case of (3).
Let f be a C 1 function with compact support in Rn. For t > 0, consider the level sets of
f in Rn:
and
Nf,t = {x ∈ Rn : f (x) ≥ t}
Sf,t = {x ∈ Rn : f (x) = t}.
Since f is of class C 1, by Sard's Theorem, Sf,t is a C 1 submanifold which has non-zero
normal vector ∇f , for almost all t. Denote by dSt the surface area element of Sf,t. Then the
co-area formula relates the area elements dx = ∇f −1dStdt.
We present a lemma, whose proof is inside of the proof of Theorem 4.1 of [34]. It will be
useful to prove Theorem 1.1 for the case r = 1.
Lemma 3.3. If f is a C 1 function with compact support in Rn, then:
Z ∞
0
vol(Nf,t)
n−1
n dt ≥ f n
n−1
.
We observe that the proof of Lemma 3.3 carries over replacing n−1
not for η > 1. We prove an analogous result for η = n+p
p > 1.
n by any η ∈ (0, 1), but
Lemma 3.4. If g is a C 1 function with compact support in Rn and λ ∈(cid:16) n
n+p , 1(cid:17) ∪ (1, ∞)
n+p
vol(Ng,t)
n dt ≥ an,p,λg
1
(n+p)(λ−1)+p
(λ−1)n
− λp
g
λ
(λ−1)n
,
Z ∞
0
where
with
an,p,λ =
A
− (n+p)(λ−1)+p
n,p,λ
p
(λ−1)n
B
(λ−1)n
n,p,λ
if λ > 1
if λ ∈(cid:16) n
n+p , 1(cid:17)
An,p,λ = ((λ − 1)n + λp)
λ−1(cid:17) (λp)
Γ(cid:16) λ
1
1−λ ((λ − 1)(n + p))− n+p
Γ(cid:16) n
p + 1
λ−1 + 2(cid:17)
p Γ(cid:16) n
p + 2(cid:17)
(λ−1)p
(λ−1)n+λp
8
and
Bn,p,λ = λ
p
n + p(cid:18)λ −
n
n + p(cid:19)
(1−λ)(n+p)
p
−1
Proof. For λ > 1 and t > 0, let pλ(t) = (1 − tλ−1)
(1 − tλ−1s1−λ)
n
p
+ and
(1 − λ)− n
1−λ
p(cid:17)
1−λ − n
p −2Γ(cid:16) n
p + 2(cid:17) Γ(cid:16) λ
Γ(cid:16) λ−2
λ−1(cid:17)
+ and l(t) = vol(Ng,t). Then pλ(cid:0) t
n
p
s(cid:1) =
≥ 1 − tλ−1s1−λ.
(12)
p
n
pλ(cid:18) t
s(cid:19)
Multiplying (12) by l(t) and integrating, we obtain:
p
n
0
Z ∞
l(t)pλ(cid:18) t
s(cid:19)
g1 ≤Z ∞
dt ≥Z ∞
l(t)pλ(cid:18) t
s(cid:19)
0
0
p
n
0
l(t)dt − s1−λZ ∞
dt + s1−λZ ∞
0
l(t)tλ−1dt,
l(t)tλ−1dt.
whence
By Holder, observe that:
p
n
Z ∞
0
l(t)pλ(cid:18) t
s(cid:19)
dt ≤(cid:18)Z ∞
0
l(t)
n+p
n dt(cid:19)
Write u = t/s and dt = sdu. Then:
n
n+p Z ∞
0
pλ(cid:18) t
s(cid:19)
n+p
n
p
n+p
.
dt!
p
n
Z ∞
0
l(t)pλ(cid:18) t
s(cid:19)
dt ≤(cid:18)Z ∞
0
l(t)
n+p
n dt(cid:19)
Now, observe that:
n
n+p (cid:18)Z ∞
0
pλ(u)
n+p
n du(cid:19)
p
n+p
p
n+p .
s
Z ∞
0
l(t)tλ−1dt =Z ∞
0
vol(Ngλ,tλ)tλ−1dt.
Write v = tλ, dv = λtλ−1dt, then tλ−1dt = 1
λ dv and:
l(t)tλ−1dt =
Z ∞
0
1
λZ ∞
0
vol(Ngλ,t)dt =
1
λ
gλ
λ.
Hence,
g1 ≤
1
λ
gλ
where a = 1
λ gλ
0
λs1−λ +(cid:18)Z ∞
λ, b =(cid:16)R ∞
0
l(t)
n
0
n+p
n+p (cid:18)Z ∞
n dt(cid:19)
n+p (cid:16)R ∞
n dt(cid:17)
n
0 pλ(t)
n+p
l(t)
p
n+p
pλ(t)
n+p
n dt(cid:19)
p
n+p = as−α + bsβ,
s
(13)
p
n+p , α = λ − 1 e β = p
n+p .
n+p
n dt(cid:17)
Notice that the right-hand side of (13) has a unique minimum for s ∈ (0, ∞), then mini-
mizing with respect to s ∈ (0, ∞), we obtain:
(n+p)(λ−1)+p
g1 ≤ An,p,λg
λ
λp
(cid:18)Z ∞
0
l(t)
n+p
n dt(cid:19)
(λ−1)n
(n+p)(λ−1)+p
,
where An,p,λ is given in the statement of the Lemma.
Hence,
9
(n+p)(λ−1)+p
g1 ≤ An,p,λg
λ
λp
(cid:18)Z ∞
0
vol(Ng,t)
n+p
n dt(cid:19)
(λ−1)n
(n+p)(λ−1)+p
,
that proves the statement of Lemma for the case λ > 1.
n+p , 1(cid:17), we define qλ(t) = (tλ−1 − 1)
n
p
+. Then, qλ(t)
p
n ≥ tλ−1 − 1 and
p
For the case λ ∈ (cid:16) n
qλ(cid:0) t
s(cid:1)
n ≥ tλ−1s1−λ − 1.
It follows that
l(t)tλ−1dt = 1
l(t)dt = g1, we obtain
Z ∞
0
p
n
0
λ gλ
dt ≥ s1−λZ ∞
s(cid:19)
l(t)qλ(cid:18) t
λ and R ∞
λ ≤ g1 +Z ∞
gλ
s1−λ
λ
0
0
l(t)tλ−1dt −Z ∞
0
l(t)dt
p
n
dt.
l(t)qλ(cid:18) t
s(cid:19)
Since R ∞
0
By Holder
p
n
Z ∞
0
l(t)qλ(cid:18) t
s(cid:19)
0
dt ≤(cid:18)Z ∞
=(cid:18)Z ∞
0
l(t)
l(t)
n+p
n dt(cid:19)
n dt(cid:19)
n+p
n
0
n+p Z ∞
n+p (cid:18)Z ∞
0
n
n+p
n
dt!
qλ(cid:18) t
s(cid:19)
n du(cid:19)
qλ(u)
n+p
p
n+p
p
n+p
s
p
n+p
Hence,
1
λ
gλ
For λ ∈ (cid:16) n
n
p
n+p
l(t)
λ ≤ sλ−1g1 +(cid:18)Z ∞
n+p , 1(cid:17), the right-hand side of (14) has a unique minimum s ∈ (0, ∞), then
n+p (cid:18)Z ∞
n dt(cid:19)
n dt(cid:19)
n+p +λ−1.
qλ(t)
(14)
n+p
n+p
s
0
0
p
minimizing with respect to s ∈ (0, ∞), we obtain:
gλ
λ ≤ Bn,p,λg
1
p
(n+p)(λ−1)+p
(cid:18)Z ∞
0
l(t)
n+p
n dt(cid:19)
(1−λ)n
p
,
where Bn,p,λ is given in the statement of the Lemma.
Therefore,
vol(Ng,t)
n+p
n dt ≥ B
− p
c,d,λ
(1−λ)n
− (n+p)(λ−1)+p
g
1
(1−λ)n
pλ
(1−λ)n
g
λ
.
Z ∞
0
Now, we present other tools for the case 1 < r < n of our main result, introduced by
Lutwak, Yang, and Zhang in [20]. Let H 1,r(Rn) denote the usual Sobolev space of real-
valued functions of Rn with Lr partial derivatives. If f ∈ H 1,r(Rn) ∩ C ∞(Rn) and Q is a
compact convex set that contains the origin in its relative interior, then they define
Vr(f, t, Q) =
1
nZSf,t
hQ(ν(x))r∇f (x)r−1dSt(x),
10
where ν(x) = ∇f (x)
convex body Kt such that, for each origin-symmetric convex body Q
∇f (x) . They prove that for almost every t > 0, there exists an origin-symmetric
Vr(Kt, Q) = Vr(f, t, Q).
(15)
The next Lemma can be deduced from [20], inequalities (6.3), (5.3), (5.4), and (5.1).
Lemma 3.5. If r ∈ (1, n), f ∈ H 1,r(Rn) and q = nr
n−r , then
vol(Kt)
n−r
n dt ≥ n
Z ∞
0
r−n
n cr
2f r
q,
where
c2 = n
1
q (cid:18) n − r
r − 1(cid:19)
r−1
r Γ(cid:0) n
r(cid:1) Γ(cid:0)n + 1 − n
r(cid:1)
Γ(n)
1
n
!
.
4 Proof of the main results
We present separate proofs for the cases 1 < r < n and r = 1.
4.1 Case 1 < r < n:
Proof of 2.3. By the co-area formula, (15), (5) and Lemma 3.5:
Vr(f, K) =
hK (−∇f (x))rdx
hK(n
Sf,t
x
)r∇f (x)r−1dSf,tdt
Vr(f, t, K)dt
Vr(Kt, K)dt
1
1
=
nZRn
nZ ∞
0 ZSf,t
=Z ∞
=Z ∞
≥Z ∞
=Z ∞
0
r−n
≥ n
n cr
0
0
0
vol(Kt)
vol(Kt)
2f r
q vol(K)
r
n
n−r
n vol(K)
r
n dt
n−r
n dt vol(K)
r
n
Proof of 2.5. We may observe that:
g(x)hx, ξipdx
hx, ξipdxdt
hMpg(ξ)p =ZRn
=Z ∞
0 Z{g≥t}
=Z ∞
0
hMpNg,t(ξ)pdt.
In this sense, we regard Mpg as a generalized p-sum of sets, where we replace finite p-sums
11
by a p-integral of sets
and clearly, for any convex body K,
Vp(cid:18)K,Zp
We compute:
MpNg,tdt
Mpg =Zp
MpNg,tdt(cid:19) =Z ∞
0
Vp(K, MpNg,t)dt.
vol(Mpg) = Vp(Mpg, Mpg)
MpNg,tdt(cid:19)
Vp(Mpg, MpNg,t)dt
= Vp(cid:18)Mpg,Zp
=Z ∞
0
≥ vol(Mpg)
n−p
n Z ∞
0
vol(Mpg)
vol(MpNg,t)
p
0
n ≥Z ∞
≥ cn,pZ ∞
0
vol(Ng,t)
n+p
n dt
(n+p)(λ−1)+p
vol(MpNg,t)p/ndt
p
n dt
− λp
g
λ
(λ−1)n
,
Then using Lemmas 3.2 and 3.4, it follows that
where an,p,λ is given by Lemma (3.4).
≥ cn,pan,p,λg
1
(λ−1)n
4.2 Proof of Theorem 1.1: Case r = 1
Proof. Let V1(f, Mpg) = 1
V1(f, Mpg) =
dStdt.
We denote ηSt
x = ∇f (x)
∇f (x) , as
p
1
(cid:28) ∇f (x)
∇f (x)
nRRn(cid:0)RRn g(y)h∇f (x), yipdy(cid:1)1/p dx. Then,
dy(cid:19)1/p
nZ ∞
, y(cid:29)(cid:12)(cid:12)(cid:12)(cid:12)
0 ZSf,t(cid:18)ZRn
dy(cid:19)1/p
dyds!1/p
g(y)(cid:12)(cid:12)(cid:12)(cid:12)
g(y)(cid:12)(cid:12)(cid:10)ηSt
x , y(cid:11)(cid:12)(cid:12)
0 ZNg,s(cid:12)(cid:12)(cid:10)ηSt
x , y(cid:11)(cid:12)(cid:12)
x ) =(cid:18)ZRn
= Z ∞
p
p
,
hMpg(ηSt
it follows that:
V1(f, Mpg) =
=
hMpg(ηSt
x )dStdt
1
nZ ∞
0 ZSf,t
0 ZSf,t Z ∞
nZ ∞
1
p
0 ZNg,s(cid:12)(cid:12)(cid:10)ηSt
x , y(cid:11)(cid:12)(cid:12)
dyds!1/p
dStdt
Write hMpNg,s(ηSt
p
dy, then:
x )p =RNg,s(cid:12)(cid:12)(cid:10)ηSt
x , y(cid:11)(cid:12)(cid:12)
0 ZSf,t(cid:18)Z ∞
nZ ∞
V1(f, Mpg) =
1
0
hMpNg,s(ηSt
x )pds(cid:19)1/p
dStdt
By the co-area formula, the Minkowski integral inequality and Lemmas 3.1, 3.2, 3.5 and
3.4:
12
V1(f, Mpg) ≥
0
1
1
≥
0 Z ∞
nZ ∞
0 Z ∞
n Z ∞
=(cid:18)Z ∞
0 (cid:18)Z ∞
≥(cid:18)Z ∞
0 (cid:18)Z ∞
=(cid:18)Z ∞
0 (cid:18)Z ∞
=(cid:18)Z ∞
n,p(cid:18)Z ∞
≥ c
1
p
0
0
0
0
vol(Nf,t)
≥ c1/p
n,p f n
n−1
C
hMpNg,s(ηSt
0 ZSf,t
0 ZSf,t
V1(Nf,t, MpNg,s)dt(cid:19)p
hMpNg,s(ηSt
vol(Nf,t)
n−1
n vol(MpNg,s)
vol(Nf,t)
n−1
n dt(cid:19)p
n dt(cid:19)(cid:18)Z ∞
0
n−1
1
p
x )dSt!p
ds!
x )dSt! dt!p
ds(cid:19)
n dt(cid:19)p
1
p
1
vol(MpNg,s)
vol(MpNg,s)
dt
1
p
ds!
1
p
p
1
p
ds(cid:19)
n ds(cid:19)
n ds(cid:19)
n ds(cid:19)
n+p
1
p
p
1
p
vol(Nf,t)
n−1
n dt(cid:19)(cid:18)Z ∞
vol(Ng,s)
0
(n+p)(λ−1)+p
− (n+p)(λ−1)+p
n,p,λ
(λ−1)np
(λ−1)np
g
1
− λ
g
λ
(λ−1)n
.
Remark 4.1. Let us point out that a simpler proof of Theorem 1.1 for the case r = p can
be deduced using the Lp Affine Sobolev inequality [20] and the equivalence between the Lp
Busemann-Petty centroid inequality and the Lp Petty projection inequality (see [19]). The
well known identity for sets
Vp(L, ΓpK) =
ωn
vol(K)
V−p(K, Π◦
pL),
where Vp(·, ·) denotes the Lp dual mixed volume and Π◦
can be extended to functions as
pL the Lp polar projection body of L,
Vp(f, Mpg) = V−p(g, Π◦
pf )
where we define
and
V−p(g, L) =ZRn
pf, ξ)p =ZRn
h(Π◦
xp
Lg(x)dx
h∇f (x), ξ)ipdx.
Then an application of the dual mixed volume inequality for functions (Lemma 4.1 in [21])
and the Lp Affine Sobolev inequality (which corresponds to the Lp Petty Projection inequality
for functions), gives the result.
13
Acknowledgements
The first author was partially supported by Fapemig, Project APQ-01542-18 and CNPQ
grant PQ-301203/2017-2. The second and third authors are partially supported by FAPERJ
grant JCNE 236508 and CNPQ grant 428076/2018-1. The second author was also partially
supported by CNPQ grant PQ 305650/2016-5 and PUC-Rio programa de incentivo a pro-
dutividade em pesquisa. The third author acknowledges the support of the IFMG campus
Bambui while conducting this work.
References
[1] Angelo Alvino, Vincenzo Ferone, Guido Trombetti, and Pierre-Louis Lions. Convex
symmetrization and applications. In Annales de l'Institut Henri Poincare (C) Non Linear
Analysis, volume 14, pages 275 -- 293. Elsevier, 1997.
[2] Thierry Aubin. Problemes isop´erim´etriques et espaces de Sobolev. Journal of differential
geometry, 11(4):573 -- 598, 1976.
[3] Wilhelm Blaschke. Affine geometrie ix: Verschiedene Bemerkungen und aufgaben. Ber.
Verh. Sachs. Akad. Wiss. Leipzig Math. -- Phys. Kl, 69:412 -- 420, 1917.
[4] Fenchel W. Bonnesen, T. Theorie der Konvexen Korper. Springer, Berlin, 1934. Reprint:
Chelsea Publ. Co., New York, 1948. English translation: BCS Associates, Moscow, Idaho,
1987.
[5] Silouanos Brazitikos, Apostolos Giannopoulos, Petros Valettas, and Beatrice-Helen Vrit-
siou. Geometry of isotropic convex bodies, volume 196. American Mathematical Soc.,
2014.
[6] Yu.D. Burago and V.A. Zalgaller. Geometric Inequalities. Grundlehren der Mathematis-
chen Wissenschaften, 285, Springer-Verlag, Berlin, 1988, MR 89b:52020, Zbl 0633.53002.
[7] Dario Cordero-Erausquin, Bruno Nazaret, and C´edric Villani. A mass-transportation
approach to sharp Sobolev and Gagliardo-Nirenberg inequalities. Advances in Mathe-
matics, 182(2):307 -- 332, 2004.
[8] PL De N´apoli, Juli´an Haddad, Carlos Hugo Jim´enez, and Marcos Montenegro. The
sharp affine L2 Sobolev trace inequality and variants. Mathematische Annalen, 370(1-
2):287 -- 308, 2018.
[9] H. Federer. Geometric Measure Theory. Springer-Verlag, Berlin, 1969.
[10] Herbert Federer and Wendell H Fleming. Normal and integral currents. Annals of
Mathematics, pages 458 -- 520, 1960.
[11] Christoph Haberl and Franz E Schuster. Asymmetric affine Lp Sobolev inequalities.
Journal of Functional Analysis, 257(3):641 -- 658, 2009.
[12] Christoph Haberl and Franz E Schuster. General Lp affine isoperimetric inequalities.
Journal of Differential Geometry, 83(1):1 -- 26, 2009.
[13] J Haddad, C Jim´enez, and M Montenegro. Sharp affine weighted Lp Sobolev type
inequalities. Transactions of the American Mathematical Society, 2018.
[14] Julian Haddad, C Hugo Jimenez, and Marcos Montenegro. Asymmetric Blaschke-Santal´o
functional inequalities. arXiv preprint arXiv:1810.02288, 2018.
[15] Julian Haddad, Carlos Hugo Jim´enez, and Marcos Montenegro. Sharp affine Sobolev
type inequalities via the Lp Busemann-Petty centroid inequality. Journal of Functional
Analysis, 271(2):454 -- 473, 2016.
14
[16] Elliott H Lieb and Michael Loss. Analysis, graduate studies in mathematics, vol. 14.
American Mathematical Society, Providence, 2001.
[17] Erwin Lutwak. Mixed projection inequalities. Transactions of the American Mathemat-
ical Society, 287(1):91 -- 106, 1985.
[18] Erwin Lutwak. The Brunn-Minkowski-Firey theory. i. Mixed volumes and the Minkowski
problem. J. Differential Geom, 38(1):131 -- 150, 1993.
[19] Erwin Lutwak, Deane Yang, and Gaoyong Zhang. Lp affine isoperimetric inequalities.
J. Differential Geom, 56(1):111 -- 132, 2000.
[20] Erwin Lutwak, Deane Yang, and Gaoyong Zhang. Sharp affine Lp Sobolev inequalities.
Journal of Differential Geometry, 62(1):17 -- 38, 2002.
[21] Erwin Lutwak, Deane Yang, and Gaoyong Zhang. Moment-entropy inequalities. The
Annals of Probability, 32(1B):757 -- 774, 2004.
[22] Erwin Lutwak, Deane Yang, and Gaoyong Zhang. Optimal Sobolev norms and the Lp
Minkowski problem. International Mathematics Research Notices, 2006, 2006.
[23] Erwin Lutwak and Gaoyong Zhang. Blaschke-Santal´o inequalities. J. Differential Geom,
47(1):1 -- 16, 1997.
[24] Vladimir Gilelevich Maz'ya. Classes of domains and imbedding theorems for function
spaces. In Doklady Akademii Nauk, volume 133, pages 527 -- 530. Russian Academy of
Sciences, 1960.
[25] Vitali Milman and Liran Rotem. α-concave functions and a functional extension of mixed
volumes. arXiv preprint arXiv:1302.0823, 2013.
[26] Vitali Milman and Liran Rotem. Mixed integrals and related inequalities. Journal of
Functional Analysis, 264(2):570 -- 604, 2013.
[27] Hermann Minkowski. Theorie der konvexen korper,
insbesondere begrundung ihres
oberflachenbegriffs. Gesammelte Abhandlungen, 2:131 -- 229, 1911.
[28] Hermann Minkowski. Volumen und oberflache. Math. Ann. 57 (1903), 447 -- 495. Gesam-
melte Abhandlungen, vol. II, pp. 230 -- 276, Teubner, Leipzig, 1911.
[29] Robert Osserman. The isoperimetric inequality. Bulletin of the American Mathematical
Society, 84(6):1182 -- 1238, 1978.
[30] Grigoris Paouris and Peter Pivovarov. Randomized isoperimetric inequalities. In Con-
vexity and Concentration, pages 391 -- 425. Springer, 2017.
[31] Clinton M Petty. Centroid surfaces. Pacific Journal of Mathematics, 11(4):1535 -- 1547,
1961.
[32] Rolf Schneider. Convex bodies: the Brunn-Minkowski theory. Number 151. Cambridge
university press, 2014.
[33] Giorgio Talenti. Best constant in Sobolev inequality. Annali di Matematica pura ed
Applicata, 110(1):353 -- 372, 1976.
[34] Gaoyong Zhang. The affine Sobolev inequality. J. Differential Geom, 53(1):183 -- 202,
1999.
|
1705.04635 | 2 | 1705 | 2017-06-19T11:05:20 | Local approach to order continuity in Ces\`aro function spaces | [
"math.FA"
] | The goal of this paper is to present a complete characterisation of points of order continuity in abstract Ces\`aro function spaces $CX$ for $X$ being a symmetric function space. Under some additional assumptions mentioned result takes the form $(CX)_a = C(X_a)$. We also find simple equivalent condition for this equality which in the case of $I=[0,1]$ comes to $X\neq L^\infty$. Furthermore, we prove that $X$ is order continuous if and only if $CX$ is, under assumption that the Ces\`aro operator is bounded on $X$. This result is applied to particular spaces, namely: Ces\`aro-Orlicz function spaces, Ces\`aro-Lorentz function spaces and Ces\`aro-Marcinkiewicz function spaces to get criteria for OC-points. | math.FA | math |
LOCAL APPROACH TO ORDER CONTINUITY IN CES `ARO FUNCTION
SPACES
TOMASZ KIWERSKI AND JAKUB TOMASZEWSKI
Abstract. The goal of this paper is to present a complete characterization of points of order
continuity in abstract Ces`aro function spaces CX for X being a symmetric function space.
Under some additional assumptions mentioned result takes the form (CX)a = C(Xa). We
also find simple equivalent condition for this equality which in the case of I = [0, 1] comes
to X 6= L∞. Furthermore, we prove that X is order continuous if and only if CX is, under
assumption that the Ces`aro operator is bounded on X. This result is applied to particular
spaces, namely: Ces`aro-Orlicz function spaces, Ces`aro-Lorentz function spaces and Ces`aro-
Marcinkiewicz function spaces to get criteria for OC-points.
1. Introduction
Much has been said about the Ces`aro spaces both from isomorphic and isometric point of
view, see for example the papers [1], [2] by Astashkin and Maligranda, [9] by Curbera and
Ricker, [10] by Delgado and Soria and [22], [23], [24] by Le´snik and Maligranda and references
given there.
For a Banach ideal space X the abstract Ces`aro space CX is the space of all functions f such
that Cf ∈ X, equipped with the norm kf kCX = kCf kX , where C denotes the continuous
Ces`aro operator
General considerations for this construction has been initiated in [22]. The abstract Ces`aro
spaces CX are neither symmetric nor reflexive. Surprisingly, the descriptions of Kothe duals of
CX spaces is different for the case of I = [0, 1] and I = [0, ∞), see [22].
Here we will focus on considering the abstract Ces`aro spaces CX for those function spaces X
which are symmetric. We study the local structure of this spaces in the terms of order continuity
property.
The paper is organized as follows.
In Section 2 we collect some necessary preliminaries on Banach ideal spaces, symmetric func-
tion spaces, Ces`aro function spaces and order continuity property. Here, we specify Theorems
A, B, C, D and E, Fact 1 and Lemma 2 because we will used them often in this article.
Section 3 contains the main results of this paper. Curbera and Ricker in [9] proved that
(CX)a = C(Xa) for symmetric spaces X 6= L∞[0, 1] on I = [0, 1]. Moreover, Kiwerski and
Kolwicz in [15] have shown analogous equality in the case of the Ces`aro-Orlicz function spaces
CesΦ := CLΦ, see also Remark 6 for a more accurate discussion. We extend these results to
the class of symmetric spaces and we get a full characterization of order continuous points in
abstract Ces`aro function spaces CX.
In the last Section 4, we show applications of our characterization for particular cases of
symmetric spaces. Some results from this section were proved earlier directly for X = Lp by
Hassard and Hussein in [12], Shiue in [31], for X = L∞ by Zaanen in [32] and for X = LΦ by
Kiwerski and Kolwicz in [15].
2010 Mathematics Subject Classification: 46B20, 46B42, 46E30.
Key words and phrases: Ces`aro function spaces; Ces`aro-Orlicz function spaces; Ces`aro-Lorentz function spaces;
Ces`aro-Marcinkiewicz function spaces; order continuity; local structure of a separated point.
1
C : f 7→ Cf (x) :=
f (t)dt.
1
xZ x
0
2. Notation and preliminaries
Denote by m the Lebesgue measure on I and by L0 = L0(I) the space of all classes of real-
valued Lebesgue measurable functions defined on I, where I = [0, 1] or I = [0, ∞). Through all
the paper when we pick a subset A ⊂ I we assume that A is a Lebesgue measurable set. For a
subset A ⊂ I we define the essential infimum infess(A) of A as follows
infess(A) := inf{ǫ ≥ 0 : m([0, ǫ) ∩ A) = 0}.
A Banach space E = (E, k · k) is said to be a Banach ideal space on I (we write E[0, 1] or
E[0, ∞)) if E is a linear subspace of L0(I) and satisfies the condition that if g ∈ E, f ∈ L0 and
f ≤ g a.e. on I then f ∈ E and kf k ≤ kgk. Sometimes we write k·kE to be sure in which
space the norm has been taken. We say that E is non-trivial if E 6= {0}.
For f ∈ L0 we define support as supp(f ) := {t ∈ I : f (t) 6= 0}. Recall that the support
supp(X) of Banach ideal space X is defined as measurable subset of I such that m(supp(f ) \
supp(X)) = 0 for every f ∈ X, and for every U ⊂ supp(X) of finite positive measure we have
χU ∈ X.
For two Banach spaces E and F on I the symbol E ֒→ F means that the embedding E ⊂ F
is continuous, i.e. there exists a constant A > 0 (we call it the embedding constant) such that
kf kF ≤ A kf kE for all f ∈ E. Recall that for two Banach ideal spaces E and F the embedding
E ⊂ F is always continuous. Moreover, E = F (resp. E ≡ F ) means that the spaces are the
same as the sets and the norms are equivalent (resp. equal). By E ≃ F we denote the fact that
the Banach spaces E and F are isomorphic.
For function f ∈ L0 we define distribution functions as df (λ) := m({t ∈ I : f (t) > λ})
for λ ≥ 0. We say that two functions f, g ∈ L0(I) are equimeasurable when they have the
same distribution functions, i.e. df ≡ dy. By a symmetric function space (symmetric Banach
function space or rearrangement invariant Banach function space) on I we mean a Banach ideal
space E = (E, k·kE) with the additional property that for any two equimeasurable functions
f, g ∈ L0(I) if f ∈ E then g ∈ E and kf kE = kgkE (we also accept the convention to write
"symmetric space" within the meaning of "symmetric function space" because in this paper
we focus only on the consideration of this case). In particular, kf kE = kf ∗kE, where f ∗(t) :=
inf{λ > 0 : df (λ) ≤ t} for t ≥ 0. Note that, if a symmetric function space E on I is non-trivial
then supp(E) = I. For the theory of symmetric spaces the reader is referred to [4] and [19].
For a symmetric function space E on I its fundamental function ϕE is defined by the formula
for 0 < t ∈ I. Writing ϕE(0+) or ϕE(∞) we understand lim
t→0+
ϕE(t) or lim
t→∞
ϕE(t), respectively.
ϕE(t) :=(cid:13)(cid:13)χ[0,t](cid:13)(cid:13)E
,
A point f ∈ X is said to have an order continuous norm (f is an OC-point) if for any sequence
(fn) ⊂ X with 0 ≤ fn ≤ f and fn → 0 a.e. on I, we have kfnkX → 0. By Xa we denote
the subspace of all order continuous elements of X. A Banach ideal space X is called order
continuous (we write X ∈ (OC) for short) if every element of X has an order continuous norm,
if Xa = X. It is worth to notice that in case of Banach ideal spaces on I, f ∈ Xa if and
i.e.
only if kf χAnkX → 0 for any sequence (An) satisfying An ↓ ∅ (that is, decreasing sequence of
Lebesgue measurable sets An ⊂ I with intersection of measure zero, see [4, Proposition 3.5,
p. 15]). The subspace Xa is always closed, see [4, Th. 3.8, p. 16]. Characterization of order
continuity given in Theorem A (iii) and (iv) is well known.
Theorem A.
(i) ([8, Lemma 2.6]) Let E be symmetric space. Then f ∈ E is a point of order continuity
if and only if f ∗ is also.
2
(ii) ([3, Lemma 2.5], [8, Lemma 2.5], cf. [19]) Let E be symmetric space. Then f ∈ Ea if
and only if
for n → ∞.
In particular if f is a point of order continuity then f ∗(∞) = 0.
(cid:13)(cid:13)(cid:13)
f ∗χ[0, 1
→ 0
n )(cid:13)(cid:13)(cid:13)E
and (cid:13)(cid:13)f ∗χ[n,∞)(cid:13)(cid:13)E → 0
(iii) ([27]) A Banach ideal space E is order continuous if and only if E contains no isomorphic
copy of l∞.
(iv) ([4, Th. 5.5, p. 27]) A Banach ideal space E is order continuous if and only if is
separable.
(cid:3)
Let X be a Banach ideal space. The closure in X of the set of simple functions is denoted by
Xb. It is well known, that the subspace Xb is the closure in X of the set of bounded functions
supported in sets of finite measure, cf. [4, Prop. 3.10. p. 17]. Of course, Xb is always non-trivial
for non-trivial space X, moreover the subspace Xb is an order ideal of X. We always have
Xa ⊂ Xb, see [4, Th. 3.11, p. 18], and the inclusion {0} 6= Xa ⊂ Xb may be proper, cf.
[4,
Ex. 3, p. 30]. The fact, that this example is based on non-symmetric construction is essential.
In the symmetric case we have only two possibilities, more precisely Xa = {0} or Xa = Xb.
The subspaces Xa and Xb coincide if and only if the characteriztic functions of the sets of finite
measure all have absolutely continuous norms, cf.
[4, Th. 3.13, p. 19], [20, Prop. 2.2] and
Theorem B. Next theorem describes the opposite extreme, when Xa is trivial.
Theorem B. Let E be a symmetric space. The following conditions are equivalent:
(i) Ea = {0},
(ii) E ֒→ L∞,
(iii) Ea 6= Eb,
(iv) ϕE(0+) > 0.
In particular, if I = [0, 1] then condition (ii) is equivalent to the statement E = L∞.
Proof. Obviously, E ֒→ L∞ is equivalent to E = L∞ if I = [0, 1] because the embedding L∞ ֒→ E
holds for each symmetric space E on I = [0, 1], see [4, Corollary 6.7, p. 78]. Equivalence of
conditions (i), (iii) and (iv) follows immediately from [4, Th. 5.5, p. 67] and our discussion
preceding this theorem. For the implication (ii) ⇒ (iv) it is enough to observe, that
ϕE(0+) = lim
t→0+(cid:13)(cid:13)χ[0,t](cid:13)(cid:13)E ≥ A lim
t→0+(cid:13)(cid:13)χ[0,t](cid:13)(cid:13)L∞ = A > 0,
where A is the embedding constant. Finally, suppose ϕE(0+) > 0 and E * L∞. This means,
that there is an unbounded element f ∈ E. Then the f ∗ is unbounded in zero and f ∗ ∈ E
by the symmetry of E. Therefore we can find a sequence (tn) ⊂ I, tn → 0+ as n → ∞, with
f ∗(tn) > n. We have
1
n
as n → ∞ contrary to the assumption that ϕE(0+) > 0.
(cid:13)(cid:13)χ[0,tn](cid:13)(cid:13)E ≤(cid:13)(cid:13)(cid:13)(cid:13)
f ∗χ[0,tn](cid:13)(cid:13)(cid:13)(cid:13)E
1
n
≤
kf ∗kE → 0,
(cid:3)
For s > 0 the dilation operator Ds is defined, on L0(I), by
Dsf (t) := f (
t
s
)χI (
t
s
) = f (
t
s
)χ[0,min{1,s})(t),
for t ∈ I, is bounded in any symmetric space E on I and kDskE→E ≤ max{1, s} (see [4, p.
148]). The lower and upper Boyd indices of E are defined by
p(E) := lim
s→∞
lns
,
ln kDskE→E
3
q(E) := lim
s→0+
lns
ln kDskE→E
.
They satisfy the inequalities 1 ≤ p(E) ≤ q(E) ≤ ∞. For more details see [6], [5] and [28].
Next result will be used in Section 3.
Theorem C. ([26, Prop. 2.b.3, p. 132]) Let E be a symmetric function space. Then for every
1 ≤ p < p(E) and q(E) < q ≤ ∞ we have
Lp ∩ Lq ֒→ E ֒→ Lp + Lq.
In particular, if we take q = ∞ then for every 1 ≤ p < p(E) we get
Lp ∩ L∞ ֒→ E ֒→ Lp + L∞.
The continuous Ces`aro operator C : L0(I) → L0(I) is defined by
(cid:3)
Cf (x) :=
f (t)dt,
1
xZ x
0
for 0 < x ∈ I. For a Banach ideal space X on I we define an abstract Ces`aro space CX = CX(I)
by
CX := {f ∈ L0(I) : Cf ∈ X},
with the norm kf kCX = kCf kX (see [9], [10], [22], [23], [24]). Let us note that for non-
symmetric space X the space CX need not have a weak unit even if X has it (see [22, Example
2]), so in general supp(CX) ⊂ supp(X). Of course, if we assume that C : X → X then
supp(CX) = supp(X), cf. [22, Remark 1]. Moreover, if CX is non-trivial and X is symmetric
space on I then supp(CX) = I, see Lemma 2.
Let us mention the important result about boundedness of the Ces`aro operator.
Theorem D. [17, p. 127] For any symmetric space E on I the operator C : E → E is bounded
if and only if the lower Boyd index satisfies p(E) > 1.
(cid:3)
Note that if (X, k·kX) is a Banach ideal space then the assumption C : X → X is equivalent
to the statement C is bounded. In fact, if C : X → X, then X ֒→ CX. This means that there
is M > 0 with kf kCX = kCf kX ≤ M kf kX for all f ∈ X, i.e. C is bounded. On the other
hand, if C is bounded then for all f ∈ X we have kf kCX = kCf kX ≤ kCk kf kX = kCk kf kX ,
which means that X ⊂ CX and C : X → X. However, it may happen that X 6֒→ CX (see
[10, Proposition 2.1], [16, Example 14]). Moreover, C : CX → X is always bounded (from the
definition of CX) and CX is so-called optimal domain of C for X (cf. [10] and [23]).
The immediate consequence of Theorem D and the above discussion about boundedness of
the Ces`aro operator is a next
Theorem E. For any symmetric space E the embedding E ֒→ CE holds if and only if p(E) > 1.
In particular, if p(E) > 1 then the space CE is non-trivial.
(cid:3)
It is worth to notice the following useful observation.
Fact 1. (cf. [4, Prop. 3.2, p. 52]) For every function f : I → R we have the following inequalities
Cf ≤ Cf ≤ C f ≤ C(f ∗) := f ∗∗.
Moreover, (Cf )∗ ≤ C(f ∗).
(cid:3)
Let us note some basic fact about the non-triviality of the space CX when X is symmetric
function space.
Lemma 2. Let X be symmetric space on I. If I = [0, 1] then CX 6= {0}. If I = [0, ∞) then
the following conditions are equivalent
4
(i) CX 6= {0},
(ii) 1
(iii) 1
x χ(t0,∞) ∈ X for some t0 > 0,
x χ(t,∞) ∈ X for all t > 0.
Proof. We give an easy proof for the convenience of the reader. Assume that X is symmetric
space on I. Because L∞[0, 1] ֒→ X[0, 1], so we have that
{0} 6= CL∞[0, 1] := Ces∞[0, 1] ֒→ CX[0, 1],
see [4, Corollary 6.7, p. 78] and [15, Remark]. This means that CX[0, 1] is always non-trivial.
Equivalence of conditions (i) and (ii) follows from [22, Theorem 1 (a)]. Of course, (iii) implies
(ii). Therefore, we will only show that (ii) implies (iii). In fact, take 0 < t < t0 (if t0 < t there
is nothing to prove). Then
χ(t,t0] +
x χ(t0,∞) ∈ X from the assumption and 1
χ(t,∞) =
because 1
1
x
1
x
1
x
χ(t0,∞) ∈ X,
x χ(t,t0] ∈ X since X is symmetric.
(cid:3)
3. On the OC-points in Ces`aro function spaces CX
Theorem 3. Let X be symmetric space such that Ces`aro operator is bounded on X. Then
X ∈ (OC) if and only if CX ∈ (OC).
The proof of the above theorem will be given at the end of this section as a consequence of
characterization of subspace (CX)a. We will begin with some observations and examples.
Remark 4. The implication: X is order continuous then CX is also, follows easily (in fact
from the definition) without any assumption, see [24, Lemma 1 (a)]. Moreover, in the case of
I = [0, 1] Theorem 3 has been proved already, see [15, Proposition 2].
Leaving the assumption of symmetry Theorem 3 ceases to be true, as the following example
shows.
Example 5. (see [24, Example 1]) If X = L2[0, 1
but CX = Ces2[0, 1] ∈ (OC), see Proposition 20.
4 ]⊕L∞[ 1
4 , 1
2 ]⊕L2[ 1
2 , 1], then of course X /∈ (OC)
Remark 6. Curbera and Ricker showed in [9, Prop. 3.1 (c)], using the methods of vector
measures and integral representation, that (adapting into our notation) C(Xa) = CXa whenever
X is a symmetric space on [0, 1] and X 6= L∞. The last condition can be expressed equivalently
in several ways, cf. Theorem B. Similar results are also known in the case of I = [0, ∞) but
only for a certain class of the Orlicz spaces. Namely, suppose that Φ is an Orlicz function with
Φ < ∞, i.e. bΦ = ∞ and LΦ is the Orlicz space generated by function Φ (cf. the definitions
in the subsection 4.2 and references therein). From [15, Theorem 5 (i)] we have the following
equalities
C(LΦ
a ) = {f : C f ∈ Lϕ
a }
= {f : IΦ(λC f ) < ∞ for all λ(f ) := λ > 0}
= {f : ρΦ(λf ) < ∞ for all λ(f ) := λ > 0} = CΦ,
where CΦ is the space of all order continuous elements of space CLΦ := CesΦ, so
(CLΦ)a = (CesΦ)a = CΦ = C(LΦ
a ).
Example 7. The above observation leads naturally to the question - is it true, that C(Xa) =
(CX)a for any symmetric space X? Moment of thought show us that in general that is not the
case. Taking L∞ we observe that
C(L∞
a ) = C{0} = {0} (Ces∞)a := (CL∞)a 6= {0},
5
since elementary computation shows that, for example χ[ 1
2
,1) ∈ (CL∞)a.
However, in Theorem 16 we will give an answer when this representation is possible and
describe the space (CX)a in the other cases. We start with the following lemma.
Lemma 8. Let X be symmetric space such that CX 6= {0} and let A ⊂ I. If infess(A) > 0 and
m(A) < ∞ then χA ∈ (CX)a.
Proof. At the beginning we will show that χ(a,b) ∈ (CX)a, when 0 < a < b < ∞. Let (An) ⊂ I
be sequence of measurable sets with An ↓ ∅. Put Bn = An ∩ (a, b). Simply m(Bn) → 0 when
n → ∞. We have
χBn(t)dt
CχBn(x) =
=
≤
1
xZ x
0
χ(a,b)(x)
x
m(Bn)
b
Z(a,min{x,b})∩An
dt +
m(Bn)
x
χ(b,∞)(x)
χ(a,b)(x) +
χ(b,∞)(x).
m(Bn)
x
< ∞. Therefore
m(Bn)
x
) → 0
χ(b,∞)(cid:13)(cid:13)(cid:13)(cid:13)X
as n → ∞.
From Lemma 2, assumption CX 6= {0} implies (cid:13)(cid:13)
(cid:13)(cid:13)χ(a,b)χAn(cid:13)(cid:13)CX = kχBnkCX
= kCχBnkX ≤(cid:13)(cid:13)(cid:13)(cid:13)
= m(Bn)(
1
b
m(Bn)
1
x χ(b,∞)(cid:13)(cid:13)X
χ(a,b)(cid:13)(cid:13)(cid:13)(cid:13)X
+(cid:13)(cid:13)(cid:13)(cid:13)
χ(b,∞)(cid:13)(cid:13)(cid:13)(cid:13)X
1
x
b (cid:13)(cid:13)χ(a,b)(cid:13)(cid:13)X +(cid:13)(cid:13)(cid:13)(cid:13)
Since An was chosen arbitrary, we get χ(a,b) ∈ (CX)a.
Let A ⊂ I be a subset with infess(A) = a > 0 and m(A) < ∞. Take a sequence of measurable
sets (An) ⊂ I with An ↓ ∅. Denote Bn = An ∩ A. Then, like above, m(Bn) → 0 when n → ∞.
Note that
Since χ[a,m(Bn)+a] ≤ χ[a,m(A)+a] and χ[a,m(Bn)+a](t) ↓ 0 as n → ∞ for a.e. t ∈ I thus, from a
previous case, we obtain
C(χBn) ≤ C(χ[a,m(Bn)+a]).
kχAχAnkCX = kC(χBn)kX
≤(cid:13)(cid:13)Cχ[a,m(Bn)+a](cid:13)(cid:13)X
=(cid:13)(cid:13)χ[a,m(Bn)+a](cid:13)(cid:13)CX
→ 0 as n → ∞.
Thus χA ∈ (CX)a.
(cid:3)
Corollary 9. Let X be symmetric space such that CX 6= {0}. If f ∈ (CX)b and infess(A) > 0,
where A ⊂ I then f χA ∈ (CX)a.
Proof. Let f ∈ (CX)b, (fn) be sequence of simple functions with lim
fn = f and A ⊂ I be
n→∞
like in the assumptions above. Of course, fnχA → f χA. Moreover, m(supp(fn) ∩ A) < ∞ thus
fnχA ∈ (CX)a from Lemma 8. Since C : CX → X is always bounded and (CX)a is closed
hence f χA ∈ (CX)a.
(cid:3)
We will use the notation ∆0 = ∆0(E) := {f ∈ E : lim
t→0
space.
Cf (t) = 0}, where E is a Banach ideal
Lemma 10. Let X be symmetric space. Then (CX)b ∩ ∆0 ⊂ (CX)a.
6
Proof. If CX = {0} there is nothing to prove. Assume that CX 6= {0} and take 0 ≤ f ∈
(CX)b ∩ ∆0. Let (An) ⊂ I be arbitrary sequence of measurable subsets with An ↓ ∅. Take ǫ > 0.
Since Cf (0+) = 0 there exist δ > 0 such that supt∈[0,δ) Cf (t) < ǫ. From Corollary 9 we get
f χ[δ,∞) ∈ (CX)a. Therefore there exist nδ ∈ N with
for every n ≥ nδ. Without loss of generality suppose that δ ≤ 1. Then
C(f χAn∩[0,δ)) ≤ sup{Cf (s) : s ∈ [0, δ)}(χ[0,1) +
χ[1,∞)) ≤ ǫ(χ[0,1) +
1
t
χ[1,∞)),
(cid:13)(cid:13)f χ[δ,∞)χAn(cid:13)(cid:13)CX ≤ ǫ,
1
t
for n ∈ N. Moreover, (cid:13)(cid:13)
have
1
t χ[1,∞)(cid:13)(cid:13)X < ∞ in view of our assumption CX 6= {0} and Lemma 2. We
kf χAnkCX ≤(cid:13)(cid:13)(cid:13)
f χ[δ,∞)χAn(cid:13)(cid:13)(cid:13)CX
≤ ǫ + ǫ(cid:13)(cid:13)(cid:13)(cid:13)
≤ ǫ(1 + ϕX (1) +(cid:13)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)
f χ[0,δ)χAn(cid:13)(cid:13)(cid:13)CX
χ[1,∞)(cid:13)(cid:13)(cid:13)(cid:13)X
χ[1,∞)(cid:13)(cid:13)(cid:13)(cid:13)X
χ[0,1) +
1
t
1
t
),
for n ≥ nδ. Since ǫ > 0 was arbitrary, we conclude that
which ends the proof.
(cid:3)
kf χAnkCX → 0 as n → ∞,
Lemma 11. Let X be Banach ideal function space, then
C(Xa) ⊂ (CX)a.
Proof. Suppose 0 ≤ f ∈ C(Xa). Let (An) be arbitrary sequence of measurable subsets of I with
An ↓ ∅. First, note that f (t)χAn(t) ↓ 0 as n → ∞ for a.e. t ∈ I thus, from Lebesgue dominated
convergence theorem, we obtain that C(f χAn)(t) ↓ 0 as n → ∞ for a.e. t ∈ I. Since Cf ∈ Xa
and C(f χAn) ≤ Cf for every n ∈ N we have
kf χAnkCX = kC(f χAn)kX → 0 as n → ∞,
i.e. f ∈ (CX)a.
(cid:3)
Lemma 12. Let X be symmetric function space such that Ces`aro operator is bounded on X
and Xa = Xb. Then (CX)b ⊂ C(Xa).
Proof. We divide the proof into three steps. First, we will show that χ[0,a] ∈ C(Xa) for every
0 < a < ∞. Fix a > 0 and note that
C(χ[0,a])(t) = χ[0,a](t) +
a
t
χ[a,∞](t),
for 0 < t ∈ I. Since Xa = Xb thus χ[0,a] ∈ Xa and we only need to show that f (t) := 1
t χ[a,∞](t) ∈
Xa. Let (An) be sequence of measurable subsets of I with An ↓ ∅. Without loss of generality
we can assume that An ⊂ supp(f ) = [a, ∞) for every n ∈ N. Let δ > 0 be arbitrary. From
δ for
all n ≥ nδ. From Theorem C and our assumption that Ces`aro operator is bounded there exist
p > 1 with Lp ∩ L∞ ֒→ X. Since Lp ∈ (OC), we can find n′
Lemma 8 we have χ[a,a+δ] ∈ Xa therefore there exist nδ ∈ N such that (cid:13)(cid:13)
a χ[a,a+δ]χAn(cid:13)(cid:13)X ≤ 1
δ ∈ N satisfying
1
(cid:13)(cid:13)(cid:13)(cid:13)
1
t
χ[δ,∞]χAn(cid:13)(cid:13)(cid:13)(cid:13)Lp
7
≤
1
δ
,
for every n ≥ n′
δ. Let n ≥ max{nδ, n′
δ} we have
1
a
kf χAnkX ≤(cid:13)(cid:13)(cid:13)(cid:13)
≤
1
δ
1
δ
1
δ
=
≤
1
t
1
t
χ[δ,∞]χAn(cid:13)(cid:13)(cid:13)(cid:13)X
χ[a,a+δ]χAn(cid:13)(cid:13)(cid:13)(cid:13)X
+(cid:13)(cid:13)(cid:13)(cid:13)
χ[δ,∞]χAn(cid:13)(cid:13)(cid:13)(cid:13)L∞∩Lp
+ D(cid:13)(cid:13)(cid:13)(cid:13)
χ[δ,∞]χAn(cid:13)(cid:13)(cid:13)(cid:13)Lp
,(cid:13)(cid:13)(cid:13)(cid:13)
1
t
(p − 1)− 1
+ D max{
+ D max{
},
}
p
1
δ
1
δ
,
δ1− 1
p
where D > 0 is inclusion constant. Since δ > 0 was arbitrary, we obtain
kf χAnkX → 0
as n → ∞.
Consequently, f ∈ Xa and hence χ[0,a] ∈ C(Xa).
Now let A ⊂ I be arbitrary set of finite measure. From Fact 1 we have
C(χA) ≤ C(χ[0,m(A)]).
From a previous step of proof we also know that C(χ[0,m(A)]) ∈ Xa thus from ideal property of
Xa we get C(χA) ∈ Xa, i.e. χA ∈ C(Xa).
Finally, let f ∈ (CX)b and (fn) be a sequence of simple functions with lim
n→∞
fn = f pointwise.
We already know that Cfn ∈ Xa for every n ∈ N. From continuity of Ces`aro operator C :
CX → X we have that Cfn → Cf . Therefore Cf ∈ Xa because Xa is closed and the proof is
complete.
(cid:3)
Remark 13. We can prove more, namely: if X is a symmetric function space with Xa = Xb
then (CX)b ⊂ C(Xa) if and only if 1
t χ(λ0,∞)(t) ∈ Xa for some λ0 > 0. In above proof we in fact
proved the sufficiency. The necessity is even simpler, if 1
t χ(1,∞)(t) /∈ Xa then C(χ(0,1))(t) /∈ Xa,
i.e. χ(0,1)(t) /∈ C(Xa), but of course χ(0,1)(t) ∈ (CX)b. It follows from the proof of remaining
lemma that if C : X → X and Xa = Xb then 1
t χ[1,∞) ∈ Xa. Is also worth noting that in the
case of I = [0, 1] the assumption X 6= L∞ is equivalent to 1
t χ(λ,∞)(t) ∈ Xa for all 0 < λ < 1, see
Theorem B.
Lemma 14. Let X be symmetric space with Xa = {0}. Then (CX)a ⊂ ∆0.
Proof. If CX = {0} then inclusion is trivial, so we can assume that CX 6= {0}. Take 0 ≤ f /∈ ∆0.
Then
lim sup
Cf (t) = δ > 0.
t→0+
We can find a sequence (an) ⊂ (0, 1) such that an → 0 and Cf (an) > δ
is continuous function, therefore there exist open neighborhood Un of an with inf
t∈Un
for n ∈ N. We can chose a subsequence nk ∈ N such that Unk ∩ [0, 1
2 for every n ∈ N. Since Cf
{Cf (t)} ≥ δ
4
k ) 6= ∅ for k ∈ N. Now we
8
have
f χ[0, 1
C(f χ[0, 1
(cid:13)(cid:13)(cid:13)
k )(cid:13)(cid:13)(cid:13)CX
=(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≥(cid:13)(cid:13)(cid:13)
≥(cid:13)(cid:13)(cid:13)(cid:13)
Cf (t)χ[0, 1
Cf (t)χ[0, 1
δ
4
χUnk ∩[0, 1
k ) +
k ))(cid:13)(cid:13)(cid:13)X
k )(cid:13)(cid:13)(cid:13)X
k )(cid:13)(cid:13)(cid:13)(cid:13)X
Cf ( 1
k )
t
χ[ 1
k ,∞)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X
≥
δ
4
ϕX (0+),
(cid:3)
which shows that f /∈ (CX)a.
Lemma 15. Let X be symmetric space with Xa 6= {0}. Then
(CX)a = (CX)b.
Proof. Take δ > 0. Let (fn) ⊂ (CX)b be a sequence chosen so that fn → f almost everywhere.
Using Corollary 9 we conclude that f χ[δ,∞) ∈ (CX)a. Since C(Xa) ⊂ (CX)a and Xa 6= {0} we
have
Observe that
χ[0,δ) ∈ C(Xa) ⊂ (CX)a.
fnχ[0,δ) ≤ f ∗
n(0)χ[0,δ),
thus fnχ[0,δ) ∈ (CX)a using the ideal property of (CX)a. From continuity of C : CX → X we
have fnχ[0,δ) → f χ[0,δ) in CX. This means that f χ[0,δ) ∈ (CX)a because (CX)a is closed. (cid:3)
Theorem 16. Let X be symmetric function space. Then one of following holds:
(i) (CX)a = (CX)b if Xa 6= {0},
(ii) (CX)a = (CX)b ∩ ∆0 if Xa = {0}.
In particular, if Ces`aro operator is bounded on X and Xa 6= {0} then
(CX)a = C(Xa).
Proof. Suppose Xa 6= {0}. Then Xa = Xb from Theorem B. This is the case of Lemma 15 so
we have (CX)a = (CX)b. If we additionally assume that the Ces`aro operator C is bounded on
X, using Lemma 11 and Lemma 12 we have
C(Xa) ⊂ (CX)a = (CX)b ⊂ C(Xa).
Therefore (CX)a = C(Xa).
Assume now that Xa = {0}. From Lemma 14 we have (CX)a ⊂ ∆0. Since we always have
that (CX)a ⊂ (CX)b hence (CX)a ⊂ (CX)b ∩ ∆0. Combining this with Lemma 10 we obtain
(CX)b ∩ ∆0 ⊂ (CX)a ⊂ (CX)b ∩ ∆0,
thus (CX)a = (CX)b ∩ ∆0.
(cid:3)
Remark 17. The previous theorem can be formulated in a more concise form if we assume that
the Ces`aro operator is bounded on symmetric space X. In this case
(CX)a = C(Xa) + (CX)b ∩ ∆0.
Moreover, it follows from Remark 13 that if X is symmetric space and Xa is non-trivial then
(CX)a = C(Xa) if and only if 1
t χ(λ0,∞)(t) ∈ Xa for some λ0 > 0. Additionally, as consequence
of Remark 13, we have (CX)a = C(Xa) if and only if X 6= L∞ in the case of I = [0, 1].
Lemma 18. Let X be a symmetric space such that the Ces`aro operator is bounded on X. If
f ∈ X \ Xa then f ∗ ∈ CX \ (CX)a.
9
Proof. Take an element f ∈ X \ Xa. From a symmetry of X we get f ∗ ∈ X and C(f ∗) ∈ X
since the operator C is bounded on X. Now we have to consider two cases:
(i) Xa 6= {0}. Then, from Theorem 16, (CX)a = C(Xa). From Theorem A (i) f ∗ /∈ Xa.
From ideal property of Xa and since C(f ∗) ≥ f ∗, we have that C(f ∗) /∈ Xa, i.e.
f ∗ /∈ C(Xa) = (CX)a.
(ii) Xa = {0}. Observe that 0 6= f , since 0 ∈ Xa thus
lim sup
t→0+
C(f ∗)(t) ≥ lim sup
t→0+
f ∗(t) > 0.
From Theorem 16 we obtain f ∗ /∈ ∆0 ∩ (CX)b = (CX)a.
(cid:3)
It is time to give proof of the Theorem 3 announced at the beginning of this section.
Proof of Theorem 3. Necessity. Let X be a symmetric space such that Ces`aro operator is
bounded on X and suppose X ∈ (OC). Then X = Xa and from Theorem 16 we get
(CX)a = C(Xa) = CX,
which means that CX ∈ (OC).
Sufficiency. If X /∈ (OC) then there exist an element f /∈ Xa. From Lemma 18 f ∗ /∈ (CX)a
and consequently CX /∈ (OC) which completes the proof.
(cid:3)
4. Applications
Although each of the following spaces belongs to the class of symmetric spaces, in this special
cases, the criteria for OC-points become more specified.
4.1. The Ces`aro function spaces Cesp.
Remark 19. We have the following characterization of the closure of the set of simple functions
in the space L∞
and
(L∞[0, 1])b = L∞[0, 1],
(L∞[0, ∞))b = {f ∈ L∞[0, ∞) : lim
t→∞
f (t) = 0}.
Of course, (L∞)a = {0} and p(L∞) > 1. Let us define a set ∆∞ = ∆∞(E) := {f ∈ E :
Cf (t) = 0} for E being a Banach ideal space on [0, ∞). With this notation, we get the
lim
t→∞
following reformulation of Theorem 16 (ii):
(Ces∞)a =
∆0 = {f ∈ Ces∞ :
lim
t→0+
∆0 ∩ ∆∞ = {f ∈ Ces∞ :
Cf (t) = 0}
if I = [0, 1]
lim
t→0+,∞
Cf (t) = 0} if I = [0, ∞).
The space Ces∞[0, 1] is known as Korenblyum-Krein-Levin space K and it is also known that
Ka = {f ∈ K :
lim
h→0+
f dm = 0} = ∆0(K),
1
hZ h
0
see [32, pp. 469-471]. Thus, the above characterization is a generalization of this classical result.
It is worth to mention that Ces1[0, 1] ≡ L1(ln(1/t)). From Lebesgue dominated convergence
theorem we obtain that L1(ln(1/t) is order continuous. Moreover, Ces1[0, ∞) = {0}, see [2,
Th. 3.1 (a)]. Since the spaces Lp are order continuous for 1 ≤ p < ∞ and L∞
a = {0}, so as a
consequence of Theorem 16 and Remark 19 we obtain well known result (cf. [2, Th. 3.1 (b)]).
10
Proposition 20. We have the following characterization
(Cesp)a =( Cesp
{f ∈ Ces∞ :
lim
t→0+,∞
when 1 ≤ p < ∞
Cf (t) = 0} when p = ∞.
In particular, the Ces`aro function space Cesp is order continuous if and only if 1 ≤ p < ∞.
4.2. The Ces`aro-Orlicz function spaces CesΦ. A function Φ : [0, ∞) → [0, ∞] is called an
Orlicz function if:
(i) Φ is convex,
(ii) Φ(0) = 0, Φ(∞) = ∞,
(iii) Φ is neither identically equal to zero nor infinity on (0, ∞).
For more information about Orlicz functions see [7] and [18].
Let us denote by
and
aΦ = sup{u ≥ 0 : Φ(u) = 0},
bΦ = sup{u > 0 : Φ(u) < ∞}.
We have, aΦ < ∞ and bΦ > 0, since an Orlicz function is neither identically equal to zero nor
infinity on [0, ∞). The function ϕ is continuous and non-decreasing on [0, bΦ) and is strictly
increasing on [aΦ, bΦ).
We say an Orlicz function Φ satisfies the condition ∆2 for large arguments (Φ ∈ ∆2(∞) for
short) if there exists K > 0 and u0 > 0 such that Φ(u0) < ∞ and
Φ(2u) ≤ KΦ(u)
for all u ∈ [u0, ∞). Similarly, we can define the condition ∆2 for small, with Φ(u0) > 0
(Φ ∈ ∆2(0)) or for all arguments (Φ ∈ ∆2(R+)). These conditions play a crucial role in the
theory of Orlicz spaces, see [7], [18], [29] and [30].
The Orlicz function space LΦ = LΦ(I) generated by an Orlicz function Φ is defined by
LΦ := {f ∈ L0(I) : IΦ(f /λ) < ∞ for some λ = λ(f ) > 0},
where IΦ(f ) = RI Φ(f (t))dt is a convex modular (for the theory of Orlicz spaces and modular
spaces see [29] and [30]). The space LΦ is a Banach ideal function space with the Luxemburg-
Nakano norm
kf kΦ = inf{λ > 0 : IΦ(f /λ) ≤ 1}.
Let us recall that
LΦ
b = {f ∈ CesΦ : ∀λ>0∃sλ∈S0 such that IΦ(λ(f − s)) < ∞},
where S0 is the ideal of bounded functions with support of finite measure. It is also easy to see
that
b = {f ∈ LΦ : IΦ(λf ) < ∞ for all λ > 0} = LΦ
LΦ
a ,
when bΦ = ∞, cf. [11], [29, Th. 3.3 (a), p. 17] and Theorem B.
The Ces`aro-Orlicz function space CesΦ = CesΦ(I) is defined by CesΦ(I) := CLΦ(I). Conse-
quently, the norm in the space CesΦ is given by the formula
kf kCesΦ
= inf{λ > 0 : ρΦ(f /λ) ≤ 1},
where ρΦ(f ) = IΦ(Cf ) is a convex modular. Note that CesΦ[0, 1] 6= {0} for any Orlicz function
Φ, see Lemma 2, [15, Remark 4]. Moreover, for the equivalent conditions to the non-triviality
of CesΦ[0, ∞) see [15, Proposition 3].
Lemma 21. Let Φ be an Orlicz function. Then
(CesΦ)b = {f ∈ CesΦ : ∀λ>0∃sλ∈S0 ρΦ(
11
f − sλ
λ
) < ∞}.
Proof. The idea of this proof comes from [11, Th. 3.1]. Without loss of generality, we can
consider only the case when the Ces`aro-Orlicz function space is non-trivial. Like in mentioned
article we use the notation
H(S0) = {f ∈ CesΦ : ∀λ>0∃sλ∈S0 ρΦ(
f − sλ
λ
) < ∞}.
We claim that:
(i) H(S0) is closed,
(ii) H(S0) ⊂ clS0, where by clS0 we mean the closure of S0 in CesΦ.
Take f ∈ clH(S0) and fix λ > 0. There exist a sequence (fn) ⊂ H(S0) with fn → f . From
[29, Th. 1.3, p. 8] there exist n0 ∈ N such that
Let s ∈ S0 be such that
Since modular ρΦ is convex, we have
ρΦ(
f − fn0
λ/2
) < ∞.
ρΦ(
fn0 − s
λ/2
) < ∞.
ρΦ(
f − s
λ
) = ρΦ(2
(f − fn0 + fn0 − s)
2λ
)
≤
1
2
(ρΦ(
f − fn0
λ/2
) + ρΦ(
fn0 − s
λ/2
)) < ∞,
and (i) follows.
Pick f ∈ H(S0)+ and λ > 0. From the definition of H(S0) there exist 0 ≤ sλ ≤ f such that
ρΦ(
f − sλ
λ
) < ∞.
Take a sequence (yn) ⊂ S0 with yn ↑ f − sλ a.e. on I. Since 0 ↓ f − sλ − yn ≤ f − sλ a.e. on
I thus using Lebesgue dominated convergence theorem we get C(f − sλ − yn) ↓ 0 a.e. on I.
Because L1 ∈ (OC) and
for all n ∈ N, so
ρΦ(
f − sλ − yn
) ≤ ρΦ(
f − sλ
λ
) < ∞,
λ
ρΦ(
f − sλ − yn
λ
) → 0.
Therefore there exist n0 ∈ N with
ρΦ(
f − sλ − yn0
λ
) < λ,
i.e. dist(f, S0) ≤ λ, see [11]. Since λ > 0 was arbitrary so f ∈ clS0. This completes the proof of
part (ii).
It is clear that clS0 = (CesΦ)b and S0 ⊂ H(S0). Combining this observation with (i) and (ii)
we conclude that
and
i.e. (CesΦ)b = H(S0).
H(S0) ⊂ clS0 = (CesΦ)b,
(CesΦ)b = clS0 ⊂ clH(S0) = H(S0),
(cid:3)
Next proposition is a generalisation of Theorem 5, Theorem 7 and Corollary 8 from [15].
12
Proposition 22. Let Φ be an Orlicz function. Then
(CesΦ)a =
{f ∈ CesΦ : ρΦ(λf ) < ∞ for all λ > 0}
{f ∈ CesΦ : ∀λ>0∃sλ∈S0 ρΦ(λ(f − s)) < ∞ and lim
t→0+
{f ∈ CesΦ :
Cf (t) = 0}
lim
t→0+,∞
if bΦ = ∞
Cf (t) = 0} if aΦ = 0, bΦ < ∞
if aΦ > 0, bΦ < ∞
Moreover, if p(LΦ) > 1 then the Ces`aro-Orlicz function space CesΦ is order continuous if and
only if ϕ ∈ ∆2 (i.e. if p(LΦ) > 1 then CesΦ ∈ (OC) if and only if LΦ ∈ (OC)).
Proof. Suppose CesΦ is non-trivial and
(i) bΦ = ∞. Then there exist λ0 > 0 with IΦ( η0
that this inequality holds for all η > 0. There exist γ > 0 such that IΦ( 1
Indeed, taking γ = λ0
η0
we have
x χ(λ0,∞)) < ∞ for some η0 > 0. We will show
x χ(γ,∞)) < ∞.
IΦ(
1
x
χ(γ,∞)) =Z ∞
λ0
η0
Φ(
1
x
)dx = η−1
0 Z ∞
λ0
Φ(
η0
u
)du
Φ(
η0
u
χ(λ0,∞)(u))du = η−1
0 IΦ(
η0
x
χ(λ0,∞)) < ∞.
Take η > 0. Then
IΦ(
η
x
0
= η−1
0 Z ∞
χ(λ0,∞)) = ηZ ∞
≤ ηZ γ
≤ ηZ γ
0
0
0
Φ(
Φ(
Φ(
1
u
η
λ0
η
λ0
χ
( λ0
η ,∞)
(u))du
γ
)du + ηZ ∞
)du + ηZ ∞
γ
Φ(
Φ(
1
u
1
u
χ( λ0
η ,∞)(x))du
χ(γ,∞)(u))du < ∞,
which proves the claim. Therefore, 1
x χ(λ0,∞)(x) ∈ LΦ
a . Using Remark 13 we get
(CesΦ)a = C(LΦ
a ) = {f ∈ CesΦ : ρΦ(λf ) < ∞ for all λ > 0}.
(ii) aΦ = 0 and bΦ < ∞. In this case LΦ
a = {0}, thus from Theorem 16 (ii) and Lemma 21
we get
(CesΦ)a = (CesΦ)b ∩ ∆0
= {f ∈ CesΦ : ∀λ>0∃sλ∈S0 ρΦ(λ(f − s)) < ∞ and lim
t→0+
Cf (t) = 0}.
(iii) aΦ > 0 and bΦ < ∞. Then LΦ = L∞, cf. [29, Ex. 1, p. 98] thus CesΦ = Ces∞ and it
is enough to apply Remark 19.
Moreover, the Orlicz space LΦ is order continuous if and only if Φ ∈ ∆2, see [29, Remark 1, p.
22]. Therefore, if p(LΦ) > 1 then the Ces`aro-Orlicz function space CesΦ is order continuous if
and only if ϕ ∈ ∆2 from Theorem 3.
(cid:3)
Remark 23. Note that p(LΦ) = αΦ and q(LΦ) = βΦ, where αΦ and βΦ are the so-called lower
and upper Orlicz-Matuszewska indices, see [26, Prop. 2.b.5 and Remark 2, p. 140]) and [29],
[30] for more information.
4.3. The Ces`aro-Lorentz function spaces CΛϕ. The fundamental function ϕE of every
symmetric function space E on I is quasi-concave on I, that is, ϕE(t) = 0 if and only if t = 0,
ϕE(t) is increasing on I and ϕE(t)/t is non-increasing for t ∈ (0, m(I)). Moreover, for any
quasi-concave function ϕ there is a symmetric function space whose fundamental function is
13
ϕ. The smallest symmetric function space with fundamental function ϕ is called the Lorentz
function space Λϕ = Λϕ(I) and is defined as
Λϕ(I) := {f ∈ L0(I) :Z m(I)
0
f ∗dϕ < ∞},
with a norm given as kf kΛϕ(I) =R m(I)
f ∗dϕ. The fundamental function of a symmetric function
space E is not necessary concave but E can be equivalently renormed in such a way that the
resulting fundamental function is concave.
In this case the Riemann-Stieltjes integral in the
definition of the Lorentz function space ΛϕE my be rewritten in the form
0
kf kΛϕE (I) = kf kL∞(I) ϕE(0+) +Z m(I)
0
f ∗(t)φE(t)dt,
0 φE(s)ds. Then we have embedding ΛϕE ֒→ E with embedding constant equal
where ϕE(t) =R t
1.
Proposition 24. Let ϕ be quasi-concave function. Then
(CΛϕ)a =
CΛϕ
{f ∈ CΛϕ : lim
t→∞
{f ∈ (CΛϕ)b :
if ϕ(0+) = 0, ϕ(∞) = ∞
if ϕ(0+) = 0, ϕ(∞) < ∞
Cf (t) = 0} if ϕ(0+) > 0, ϕ(∞) = ∞
if ϕ(0+) > 0, ϕ(∞) < ∞
(Cf )∗(t) = 0}
lim
t→0+
lim
Cf (t) = 0}
{f ∈ CΛϕ :
t→0+,∞
In particular, the Ces`aro-Lorentz function space CΛϕ is order continuous if and only if ϕ(0+) = 0
and ϕ(∞) = ∞ (i.e. CΛϕ ∈ (OC) if and only if Λϕ ∈ (OC)).
Proof. We can assume that CΛϕ is non-trivial because if CΛϕ = {0} there is noting to prove.
To prove the first part we have to consider the following cases. Suppose
(i) ϕ(0+) = 0 and ϕ(∞) = ∞. It follows from [19, Lemma 5.1] and Theorem A (iv) that
x χ(1,∞)(x) ∈ (Λϕ)a so using
Λϕ ∈ (OC). Moreover, from [21, Corollary 4.13] we have 1
Remark 17 we obtain (CΛϕ)a = C(Λϕ)a = CΛϕ.
(ii) ϕ(0+) = 0 and ϕ(∞) < ∞. In this situation Λϕ /∈ (OC) but (Λϕ)a is non-trivial. More
precisely,
(Λϕ)a = {f ∈ Λϕ : lim
t→∞
f ∗(t) = 0},
see [21, Corollary 4.13]. Since C is bounded, using Theorem 16 we get
(CΛϕ)a = C(Λϕ)a = {f ∈ CΛϕ : lim
t→∞
(Cf )∗(t) = 0}.
(iii) ϕ(0+) > 0 and ϕ(∞) = ∞. From Theorem B we get (Λϕ)a = {0} and it is enough to
use Theorem 16 (ii).
(iv) ϕ(0+) > 0 and ϕ(∞) < ∞. In this case Λϕ = L∞, see [19, p. 108] and we can use
Remark 19.
The second part is a direct consequence of the foregoing considerations.
(cid:3)
4.4. The Ces`aro-Marcinkiewicz function spaces CMϕ. For any quasi-concave function ϕ
on I the Marcinkiewicz function space (called also weak Lorentz space) Mϕ = Mϕ(I) is defined
as
Mϕ(I) := {f ∈ L0(I) : sup
t∈I
ϕ(t)f ∗∗(t) < ∞},
with a norm kf kMϕ(I) = supt∈I ϕ(t)f ∗∗(t), where f ∗∗(t) := 1
0 f ∗(s)ds is a maximal function.
The Marcinkiewicz function space Mϕ is the largest symmetric function space with fundamental
function ϕ, i.e. for any symmetric function space E we have E ֒→ MϕE with embedding constant
equal 1.
t R t
14
Proposition 25. Let ϕ be quasi-concave function. Then
{f ∈ CMϕ :
lim
t→0+,∞
(CMϕ)b
{f ∈ (CMϕ)b :
{f ∈ CMϕ :
lim
t→0+
lim
t→0+,∞
ϕ(t)(Cf )∗∗(t) = 0} if ϕ(0+) = 0, p(Mϕ) > 1
if ϕ(0+) = 0, p(Mϕ) = 1
if ϕ(0+) > 0, ϕ(∞) = ∞
if ϕ(0+) > 0, ϕ(∞) < ∞
Cf (t) = 0}
Cf (t) = 0}
(CMϕ)a =
In particular, if we assume that p(Mϕ) > 1 then the non-trivial Ces`aro-Marcinkiewicz function
space CMϕ is never order continuous.
Proof. Let us consider the following cases:
(i) ϕ(0+) = 0 and p(Mϕ) > 1. Operator C is bounded from Theorem D and
(Mϕ)a = {f ∈ Mϕ :
lim
t→0+,∞
ϕ(t)f ∗∗(t) = 0},
see [13, Def. 1.3 and Th. 1.3] so applying Theorem 16 we get
(CMϕ)a = C(Mϕ)a = {f ∈ CMϕ :
lim
t→0+,∞
ϕ(t)(Cf )∗∗(t) = 0}.
(ii) ϕ(0+) = 0 and p(Λϕ) = 1. Because (Mϕ)a 6= {0} and the Ces`aro operator C is not
bounded on Mϕ we can use Theorem 16 (i).
(iii) ϕ(0+) > 0 and ϕ(∞) = ∞. In this situation (Mϕ)a is trivial in view of Theorem B, so
we simply use Theorem 16 (ii).
(iv) ϕ(0+) > 0 and ϕ(∞) < ∞. It is easy to see that Mϕ = L∞. Indeed, inclusion Mϕ ֒→ L∞
is a consequence of Theorem B. Now, if f ∈ L∞ then f ∗∗ also is a bounded function
and
kf kMϕ = sup
t∈I
ϕ(t)f ∗∗(t) ≤ kf ∗∗kL∞(I) kϕkL∞(I) < ∞,
since ϕ(∞) < ∞. Therefore also L∞ ֒→ Mϕ. From Remark 19 we get
(CMϕ)a = (Ces∞)a = {f ∈ CMϕ :
lim
t→0+,∞
Cf (t) = 0},
and the proof is finished.
Question 26. It seems that non-trivial Ces`aro-Marcinkiewicz function space CMϕ is never
order continuous but this is only our conjecture.
4.5. The spaces C(L1 ∩ L∞) and C(L1 + L∞). The spaces L1 ∩ L∞ and L1 + L∞ occupy
a special place in the theory of symmetric spaces because they are respectively the largest
and the smallest of all symmetric function spaces, i.e. L1 ∩ L∞ ֒→ E[0, ∞) ֒→ L1 + L∞ and
L∞ ֒→ E[0, 1] ֒→ L1, cf. [4, Th. 6.6 and Th. 6.7, pp. 77-78], kf kL1∩L∞ = max{kf kL1 , kf kL∞},
and
(cid:3)
kf kL1+L∞ = inf{kgkL1 + khkL∞ : f = g + h, f ∈ L1 and h ∈ L∞} =Z 1
0
cf. [4, Th. 6.2 and Th. 6.4, p. 74 and 76].
f ∗(t)dt,
The case of L1 ∩ L∞ is not very interesting because if I = [0, ∞) then C(L1 ∩ L∞) = {0}
and if I = [0, 1] then L∞ ֒→ L1 so L1 ∩ L∞ = L∞ and L1 + L∞ = L1. Therefore for I = [0, 1]
we have C(L1 ∩ L∞) = C(L∞) = Ces∞ and C(L1 + L∞) = C(L1) = Ces1 - these cases were
already discussed in the Section 4.1.
Proposition 27. (C(L1 + L∞))a = {f ∈ C(L1 + L∞) : (Cf )∗(∞) = 0}.
15
Proof. We claim that
(4.1)
(L1 + L∞)a = {f ∈ L1 + L∞ : f ∗(∞) = 0}.
This characterization is well known (see e.g. note in Section 2 in [13] and references therein)
but we will give a short proof for the sake of completeness. Inclusion ⊂ is a direct consequence
of Theorem A (ii). For the reverse inclusion ⊃ take f ∈ L1 + L∞ and observe that
(cid:13)(cid:13)f ∗χ[0,1)(cid:13)(cid:13)L1 =Z 1
=Z 1
0
0
(f ∗χ[0,1))∗(t)dt
f ∗(t)dt = kf kL1+L∞ < ∞.
Since f ∗χ[0, 1
Lebesgue dominated convergence theorem
n ] → 0 almost everywhere and is dominated by integrable function thus, from
as n → ∞. Therefore, if f /∈ (L1 + L∞)a then according to Theorem A (ii) we get
→ 0,
f ∗χ[0, 1
(cid:13)(cid:13)(cid:13)
f ∗χ[0, 1
=(cid:13)(cid:13)(cid:13)
n ](cid:13)(cid:13)(cid:13)L1
n ](cid:13)(cid:13)(cid:13)L1+L∞
(cid:13)(cid:13)f ∗χ[n,∞)(cid:13)(cid:13)L1+L∞ 9 0.
f ∗(t)dt =Z 1
0
=(cid:13)(cid:13)f ∗χ[n,∞)(cid:13)(cid:13)L1+L∞ ≥ δ > 0,
χ[1,∞)(t) ∈ (L1 + L∞)a.
1
t
Passing to subsequence if necessary we can assume that (cid:13)(cid:13)f ∗χ[n,∞)(cid:13)(cid:13)L1+L∞ ≥ δ for some δ > 0.
We have
f ∗(n) ≥Z n+1
n
(f ∗χ[n,∞))∗(t)dt
so f ∗(∞) > 0 and the claim follows. Note that from (4.1)
Using Remark 13 we obtain (C(L1 + L∞))a = C((L1 + L∞)a) and the proof is complete.
(cid:3)
16
References
[1] S. V. Astashkin, L. Maligranda, Structure of Ces`aro function spaces, Indag. Math. (N.S.) 20 (3) (2009)
329-379.
[2] S. V. Astashkin, L. Maligranda, Structure of Ces`aro function spaces: a survey, Banach Center Publ. 102
(2014) 13-40.
[3] W. Arendt, B. de Pagter, Spectrum and asymptotics of the Black-Scholes partial differential equation in
(L1, L∞)-interpolation spaces, Pacific J. Math. 202 (1) (2002) 1-36.
[4] C. Bennet, R. Sharpley, Interpolation of Operators, Pure Appl. Math., vol. 129, Academic Press, Inc., Boston,
San Diego, New York, Berkeley, London, Sydney, Tokyo, Toronto, 1988.
[5] D. W. Boyd, Indices for the Orlicz spaces, Pacific J. Math. 38 (2) (1971) 315-323.
[6] D. W. Boyd, Indices of function spaces and their relationship to interpolation, Canad. J. Math. 21 (1969)
1245-1254.
[7] S. Chen, Geometry of Orlicz spaces, Disserationes Math. (Rozprawy Mat.) CCCLVI (1996).
[8] M. Ciesielski, P. Kolwicz, A. Panfil, Local monotonicity structure of symmetric spaces with application, J.
Math. Anal. Appl. 409 (2014) 649-662.
[9] G. P. Curbera, W. J. Ricker, Abstract Ces`aro spaces: integral representation, J. Math. Anal. Appl. 441 (2016)
(1) 25-44.
[10] O. Delgado, J. Soria, Optimal domain for the Hardy operator, J. Funct. Anal. 244 (1) (2007) 119-133.
[11] A. S. Granero, H. Hudzik, On some proximinal subspaces of modular spaces, Acta Math. Hungar. 85 (1-2)
(1999) 59-79.
[12] B. D. Hassard, D. A. Hussein, On Ces`aro function spaces, Tamkang J. Math. 4 (1973) 19-25.
[13] A. Kami´nska, M. J. Lee, M-ideal properties in Marcinkiewicz spaces, Comment. Math. (2004), 123-144, Tomus
specialist in Honorem Juliani Musielak.
[14] A. Kami´nska, M. Maligranda, L.-E. Persson, Indices and regularization of measurable functions, in: Function
Spaces, The 5th Conference: Proceedings of the Conference at Pozna´n, Poland 2000, 231-246
[15] T. Kiwerski, P. Kolwicz, Isomorphic copies of l∞ in Ces`aro-Orlicz function spaces, Positivity, published
online on 3 November 2016, doi: 10.1007/s11117-016-0449-6.
[16] T. Kiwerski, P. Kolwicz, Isometric copies of l∞ in Ces`aro-Orlicz function spaces, preprint of 16 pages, 30 Jan-
uary 2016, \protect\vrule width0pt\protect\href{http://arxiv.org/abs/1701.08794}{arXiv:1701.08794}v1
at http://arxiv.org/abs/1701.08794.
[17] A. Kufner, L. Maligranda, L. E. Persson, The Hardy inequality. About its history and some related results,
Vydavatelsky Servis Publishing House, Pilsen 2007.
[18] M. A. Krasnosel'skiı, Ya. B. Rutickiı, Convex functions and Orlicz Spaces, P. Noorddhoff Ltd., Groningen,
1961 (translation).
[19] S. G. Krein, Yu. I. Petunin, E. M. Semenov, Interpolation of linear operators, Nauka, Moscow, Russia, 1978.
[20] P. Kolwicz, Rotundity properties in Calder`on-Lozanovskiı, Houston J. Math., 31 (3) (2005) 883-912.
[21] P. Kolwicz, Local structure of symmetrizations E(∗) with applications, J. Math. Anal. Appl. 440 (2) (2016)
doi: 10.1016/j.jmaa.2016.03.075
[22] K. Le´snik, M. Maligranda, On abstract Ces`aro spaces. I. Duality, J. Math. Anal. Appl. 424 (2) (2015) 932-951,
doi: 10.1016/j.jmaa.2014.11.023.
[23] K. Le´snik, M. Maligranda, On abstract Ces`aro spaces. II. Optimal Range, Integral Equations Operator Theory
81 (2) (2015) 227-235.
[24] K. Le´snik, M. Maligranda, Interpolation of abstract Ces`aro, Copson and Tandori spaces, Indag. Math. (N.S.)
27 (3) (2016), 764 -- 785, doi:10.1016/j.indag.2016.01.009.
[25] J. Lindenstrauss, L. Tzafriri, Classical Banach Spaces I. Sequence Spaces, Springer-Verlag, Berlin, Heidelberg,
1977.
[26] J. Lindenstrauss, L. Tzafriri, Classical Banach Spaces II. Function Spaces, Springer-Verlag, Berlin, Heidel-
berg, 1979.
[27] G. Ya. Lozanovskiı, On isomorphic Banach structures, Sibirsk. Math. J. 10 (1969) 93-98.
[28] L. Maligranda, Indices and interpolation, Disserationes Math. (Rozprawy Mat.) CCXXXIV (1985).
[29] L. Maligranda, Orlicz Spaces and Interpolation, Sem. Mat., vol. 5, Univ. of Campinas, Campinas SP, Brazil,
1989.
[30] J. Musielak, Orlicz Spaces and Modular Spaces, Lecture Notes in Math., vol. 1034, Springer-Verlag, 1983.
[31] J. S. Shiue, A note on Ces`aro function space, Tamkang J. Math. 1 (1970) 91-95.
[32] A. C. Zaanen, Riesz spaces II, North-Holland Math., Library 30, North-Holland, Amsterdam, 1983.
(Tomasz Kiwerski) Faculty of Mathematics, Computer Science and Econometrics, University
of Zielona G`ora, prof. Z. Szafrana 4a, 65-516 Zielona G`ora, Poland
17
e-mail address: [email protected]
(Jakub Tomaszewski) Institute of Mathematics, Faculty of Electrical Engineering, Pozna´n
University of Technology, Piotrowo 3A, 60-965 Pozna´n, Poland
e-mail address: [email protected]
18
|
1901.02796 | 3 | 1901 | 2019-03-26T17:05:31 | The pseudo-differential calculus in a Bargmann setting | [
"math.FA"
] | We give a fundament for Berezin's analytic $\Psi$do considered in \cite{Berezin71} in terms of Bargmann images of Pilipovi{\'c} spaces. We deduce basic continuity results for such $\Psi$do, especially when the operator kernels are in suitable mixed weighted Lebesgue spaces and act on certain weighted Lebesgue spaces of entire functions. In particular, we show how these results imply well-known continuity results for real $\Psi$do with symbols in modulation spaces, when acting on other modulation spaces. | math.FA | math |
PSEUDO-DIFFERENTIAL CALCULUS IN A
BARGMANN SETTING
NENAD TEOFANOV AND JOACHIM TOFT
Abstract. We give a fundament for Berezin's analytic Ψdo con-
sidered in [4] in terms of Bargmann images of Pilipovi´c spaces. We
deduce basic continuity results for such Ψdo, especially when the
operator kernels are in suitable mixed weighted Lebesgue spaces
and act on certain weighted Lebesgue spaces of entire functions. In
particular, we show how these results imply well-known continu-
ity results for real Ψdo with symbols in modulation spaces, when
acting on other modulation spaces.
0. Introduction
The aim of the paper is to put a fundament for the theory of an-
alytic pseudo-differential operators, considered in [4] by F. Berezin.
This is essentially done through a detailed analysis of Bargmann im-
ages of the so-called Pilipovi´c spaces of functions and distributions,
given in [11, 26]. More precisely, we consider kernels related to integral
representations of analytic pseudo-differential operators to deduce their
continuity properties. When the corresponding symbols belong to suit-
able (weighted) Lebesgue spaces of semi-conjugate analytic functions,
we prove the continuity of the analytic pseudo-differential operators
when acting between (weighted) Lebesgue spaces of analytic functions.
Moreover, by using the relationship between the Bargmann transform
and the short-time Fourier transform we show that our results can
be used to recover well-known (sharp) continuity properties of (real)
pseudo-differential operators with symbols in modulation spaces which
act between other modulation spaces, see [23, 25, 28]. We emphasize
that our approach here is more general, because we have relaxed the
assumptions on the involved weight functions, compared to earlier con-
tributions.
Analytic pseudo-differential operators, considered in [4] by Berezin
are well-designed when considering several problems in analysis and its
applications, e. g.
in quantum mechanics. In the context of abstract
harmonic analysis it follows that any linear and continuous operator be-
tween Fourier invariant function and (ultra-)distribution spaces may, in
2010 Mathematics Subject Classification. Primary:
32W25, 35S05, 32A17,
46F05, 42B35
Secondary: 32A25, 32A05.
Key words and phrases. Analytic kernels, Berezin operators, Pilipovi´c spaces,
modulation spaces, Gelfand-Shilov spaces.
1
a unique way, be transformed into an analytic pseudo-differential oper-
ator by the Bargmann transform (see Section 2). An advantage of such
reformulations is that all of the involved objects are essentially entire
functions and thereby possess several strong and convenient properties.
The definition of analytic pseudo-differential operators resembles the
definition of real pseudo-differential operators. In fact, let apx, ξq be a
suitable function or (ultra-)distribution on the phase space R2d. Then
the (real) pseudo-differential operator Oppaq acting on suitable sets of
functions or (ultra-)distributions on the configuration space Rd is given
by
2żRd
fpxq ÞÑ pOppaqfqpxq " p2πq´ d
apx, ξqpfpξqeixx,ξy dξ.
(0.1)
Here the integral in (0.1) should be interpreted in a distributional
(weak) sense, if necessary, and we refer to [15] or Section 1 for the
notation.
Suppose instead that a is a suitable semi-conjugate entire (analytic)
function on Cd Cd -- C2d, i. e. pz, wq ÞÑ apz, wq is an entire (analytic)
function. Then the analytic pseudo-differential operator OpVpaq acting
on suitable entire functions F on Cd is given by
Fpzq ÞÑ pOpVpaqFqpzq "żCd
apz, wqFpwqepz,wq dµpwq.
(0.2)
Here dµpwq is the Gauss measure π´de´w2
dλpwq, where dλpwq is
the Lebesgue measure on Cd, and pz, wq " řd
j"1 zj wj, when z "
pz1, . . . , zdq P Cd, and w " pw1, . . . , wdq P Cd. This means that the
operator kernel (with respect to dµ) is given by
Kpz, wq " Kapz, wq " apz, wqepz,wq.
pTKFqpzq "żCd
Evidently, pOpVpaqFqpzq is equal to the integral operator
Kpz, wqFpwq dµpwq
with respect to dµ, when K is given by (0.3). By the analyticity prop-
erties of the symbol a it follows that pz, wq ÞÑ Kpz, wq is an entire
function on C2d.
In [4, 25] several facts of analytic pseudo-differential operators are
(0.3)
(0.4)
deduced. For example, if a and F are chosen such that
z ÞÑ apz, qF epz, q
is locally uniformly bounded and analytic from Cd to L1pdµq, then
OpVpaqF in (0.2) is a well-defined entire function on Cd. In [4, 25] it is
also observed that
pOpVpzjqFqpzq " zjFpzq
and pOpVpwjqFqpzq " pBjFqpzq
when F P L1pdµ1q X ApCdq, and dµ1pwq " p1 ` wq dµpwq.
2
(0.5)
In such setting we study the mapping properties for complex integral
operators and pseudo-differential operators when respectively K " Ka
and a above belong to suitable classes of semi-conjugate entire func-
tions. In fact, we permit more generally that K and a belong to suitable
classes of formal semi-conjugate analytic power series expansions. That
is, Kpz, wq and apz, wq are of the forms
cKpα, βqeαpzqeβpwq
capα, βqeαpzqeβpwq,
eαpzq "
zα
?α!
,
and ÿα,β
ÿα,β
respectively.
To set the stage for our study we collect the background material
in Section 1. It contains a brief account on weight functions, Gelfand-
Shilov spaces, spaces of Hermite functions and power series expansions,
modulation spaces, and Bargmann transform and spaces of analytic
functions. Especially, we recall basic facts for the spaces
and their Bargmann duals
A5σpCdq`A0,5σpCdq, AspCdq`A0,spCdq,
0,spCdq,
5σpCdq`A1
spCdq`A1
A1
0,5σpCdq, A1
Fpzq "ÿα
cpF, αq À hαα!´ 1
2σ ,
cpF, αqÀ e´rα
cpF, αqeαpzq,
(0.6)
(0.7)
(0.8)
1
2s ,
when s, σ ą 0. The spaces in (0.6) consist of all formal power series
with coefficients satisfying
respectively, for some (for every) h, r ą 0, and the spaces in (0.7)
consist of all formal power series in (0.8) such that
cpF, αq À hαα!` 1
2σ ,
cpF, αqÀ e`rα
1
2s ,
respectively, for every (for some) h, r ą 0.
to belong to any of the spaces
In Section 2 we extend the definition of (0.4) to allow the kernels K
and their duals
where
uA5σpC2dq,
uA1
5σpC2dq,
uA0,spC2dq,
uA1
0,spC2dq,
uAspC2dq,
uA0,5σpC2dq,
uA1
uA1
spC2dq,
0,5σpC2dq,
uA5σpC2dq " t K ; pz, wq ÞÑ Kpz, wq P A5σpC2dqu,
TK : AspCdq ÞÑ A1
spC2dq,
spCdq when K P uA1
3
and similarly for the other spaces in (0.9) and (0.10). In the end we
prove that if s ą 0 or s " 5σ, then the integral operators in (0.4),
(0.9)
(0.10)
(0.11)
and
(0.12)
TK : A0,spCdq ÞÑA1
0,spCdq when K P uA1
0,spC2dq,
are uniquely defined and continuous, and similarly when the roles of
the non-duals in (0.6) and (0.9), and their duals in (0.7) and (0.10)
are swapped. We also prove the opposite direction, that any linear
and continuous operators between such spaces are given by such kernel
operators. These kernel results are given in Propositions 2.2 and 2.3.
Due to the Bargmann transform homeomorphisms, these results are
also equivalent to Theorems 3.3 and 3.4 in [7] on kernel theorems for
Pilipovi´c spaces. (See Subsection 1.5.)
σ ą 0,
σ ą 1,
2σ
2σ
Note that, if s ě 1
1
2 z2´rz
1
2 z2`rz
1
1
σ`1 for some r ą 0 u,
2s for some r ą 0 u
2s for every r ą 0 u
2, then the spaces of power series expansions above
can be identified with certain spaces of analytic and semi-conjugate
analytic functions. For example we have
A5σpCdq " t F P ApCdq ; Fpzq À erz
AspCdq " t F P ApCdq ; Fpzq À e
A1
spCdq " t F P ApCdq ; Fpzq À e
5σpCdq " t F P ApCdq ; Fpzq À erz
A1
A1
51pCdq " ApCdq
and similarly for uAspC2dq and uA1
Kpz, wqetpz,wq is homeomorphic on
uA1
0,5σpC2dq,
spC2dq. In particular, the mappings
(0.11) and (0.12) can be formulated in terms of those function spaces.
2s and σ ą 0, and t P C, then Kpz, wq ÞÑ
see Theorem 2.6. In particular, (0.3) implies that the mappings (0.11)
and (0.12) still hold true with OpVpaq in place of TK. (Cf. Theorems
2.7 and 2.8.)
2, the conditions on a and its kernel Ka of OpVpaq
are slightly different. More precisely, these conditions are of the form
In the case s ě 1
uA1
spC2dq and
σ´1 for every r ą 0 u,
and A1
0,51pCdq " Adpt0uq,
If instead s P p0, 1
uA1
0,spC2dq,
uA1
5σpC2dq,
and
apz, wq À e
1
2 z´w2`rpz
1
2s `w
1
2s q
Kpz, wq À e
1
2 pz2`w2q`rpz
1
2s `w
1
2s q
in order for the mappings (0.11) and (0.12) should hold. (Cf. Theorems
2.9 and 2.10.)
In Section 3 we consider operators (0.4), where certain linear pull-
backs of their kernels obey suitable mixed and weighted Lebesgue norm
4
estimates. We prove that such operators are continuous between ap-
propriate (weighted) Lebesgue spaces of entire functions. For example,
let ω be a weight on Cd Cd and ω1, ω2 be weights on Cd such that
ω2pzq
ω1pwq À ωpz, wq
and let
where
Kωpz, wq " e´ 1
If p, q, pj, qj P r1,8s satisfy
1
q1 ´
1
p2 "
1
p1 ´
GK,ωpz, wq " Kωpz, z ` wq,
2 pz2`w2qKpz, wqωp
?2z,?2wq.
1
q2 " 1 ´
1
p ´
1
q
and q ď p,
E,pω1qpCdq to Ap2
and GK,ω P Lp,qpCd Cdq, then it follows from Theorem 3.3 that TK is
E,pω2qpCdq. By slightly modifying the
continuous from Ap1
definition of GK,ω we also deduce another similar but different conti-
nuity result where the condition q ď p above is removed (cf. Theorem
3.5).
We also present some consequences of these results. Theorem 3.4
can be considered as a special case of Theorem 3.3 formulated by ana-
lytic pseudo-differential operators instead of integral operators. Theo-
rems 3.8 and 3.9 are obtained by imposing conditions on moderateness
on ω, ω1 and ω2 above and translating Theorem 3.3 and 3.5 to real
pseudo-differential operators via the Bargmann transform and its in-
verse. These approaches show that obtained continuity results on an-
alytic pseudo-differential or integral operators might be suitable when
investigating real pseudo-differential operators. In fact, Theorems 3.8
and 3.9 agree with the sharp results [24, Theorem 3.3], [27, Theorem
3.1] and [28, Theorem 2.2] in the Banach space case. Remark 3.10 in
the end of Section 3 shows that our approach can be used to extend the
latter results on real pseudo-differential operators to include situations
with non-moderate weights. We note that the moderate condition on
weights may in some situations be significantly restrictive (cf. Remark
1.14 in Section 1).
1. Preliminaries
In this section we recall some facts on involved function and distri-
bution spaces as well as on pseudo-differential operators. In Subsection
1.1 we introduce suitable weight classes. Thereafter we recall in Sub-
sections 1.2 -- 1.4 the definitions and basic properties for Gelfand-Shilov,
Pilipovi´c and modulation spaces. Then we discuss in Subsection 1.5
the Bargmann transform and recall some topological spaces of entire
functions or power series expansions on Cd. The section is concluded
with a review of some facts on pseudo-differential operators.
5
1.1. Weight functions. A weight on Rd is a positive function ω P
L8
locpRdq such that 1{ω P L8
locpRdq. The weight ω on Rd is called
moderate if there is a positive locally bounded function v on Rd such
that
x, y P Rd,
ωpx ` yq ď Cωpxqvpyq,
(1.1)
for some constant C ě 1. If ω and v are weights on Rd such that (1.1)
holds, then ω is also called v-moderate. The set of all moderate weights
on Rd is denoted by PEpRdq.
The weight v on Rd is called submultiplicative, if it is even and (1.1)
holds for ω " v. From now on, v always denotes a submultiplicative
In particular, if (1.1) holds and v
weight if nothing else is stated.
is submultiplicative, then it follows by straight-forward computations
that
ωpxq
vpyq À ωpx ` yq À ωpxqvpyq,
vpx ` yq À vpxqvpyq and vpxq " vp´xq,
(1.2)
x, y P Rd.
Here and in what follows we write Apθq À Bpθq, θ P Ω, if there is a
constant c ą 0 such that Apθq ď cBpθq for all θ P Ω.
If ω is a moderate weight on Rd, then by [25] and above, there is a
submultiplicative weight v on Rd such that (1.1) and (1.2) hold (see
also [13, 25]). Moreover if v is submultiplicative on Rd, then
(1.3)
for some constant r ą 0 (cf. [13]). In particular, if ω is moderate, then
(1.4)
1 À vpxq À erx
and e´rx ď ωpxq À erx,
ωpx ` yq À ωpxqery
x, y P Rd
for some r ą 0.
1.2. Gelfand-Shilov spaces. Let 0 ă s P R be fixed. Then the
(Fourier invariant) Gelfand-Shilov space SspRdq (ΣspRdq) of Roumieu
type (Beurling type) consists of all f P C 8pRdq such that
}f}S σ
s,h " sup xαBβfpxq
hα`βpα! β!qs
(1.5)
is finite for some h ą 0 (for every h ą 0). Here the supremum should
be taken over all α, β P Nd and x P Rd. The semi-norms } }S σ
induce an inductive limit topology for the space SspRdq and projective
limit topology for ΣspRdq, and the latter space becomes a Fr´echet space
under this topology.
2 (s ą 1
2 ).
spRdq are the
The space SspRdq ‰ t0u (ΣspRdq ‰ t0u), if and only if s ě 1
The Gelfand-Shilov distribution spaces S 1
spRdq and Σ1
dual spaces of SspRdq and ΣspRdq, respectively.
s,h
6
We have
S1{2pRdq ãÑ ΣspRdq ãÑ SspRdq ãÑ ΣtpRdq
tpRdq
1{2pRdq,
ãÑ S pRdq ãÑ S 1pRdq ãÑ Σ1
spRdq ãÑ S 1
ãÑ S 1
spRdq ãÑ Σ1
(1.6)
1
2 ă s ă t.
Here and in what follows we use the notation A ãÑ B when the topo-
logical spaces A and B satisfy A Ď B with continuous embeddings.
A convenient family of functions concerns the Hermite functions
hαpxq " π´ d
4p´1qαp2αα!q´ 1
2 e
2 pBαe´x2
q, α P Nd.
2
x
The set of Hermite functions on Rd is an orthonormal basis for L2pRdq.
It is also a basis for the Schwartz space and its distribution space, and
for any Σs when s ą 1
2 and their distribution spaces.
They are also eigenfunctions to the Harmonic oscillator H " Hd "
x2 ´ ∆ and to the Fourier transform F , given by
fpxqe´ixx,ξy dx,
2, Ss when s ě 1
2żRd
pF fqpξq " pfpξq " p2πq´ d
when f P L1pRdq. Here x , y denotes the usual scalar product on Rd.
In fact, we have
ξ P Rd,
spRdq and on Σ1
Hdhα " p2α ` dqhα.
The Fourier transform F extends uniquely to homeomorphisms on
S 1pRdq, S 1
spRdq. Furthermore, F restricts to homeo-
morphisms on S pRdq, SspRdq and on ΣspRdq, and to a unitary oper-
ator on L2pRdq. Similar facts hold true when the Fourier transform is
replaced by a partial Fourier transform.
Gelfand-Shilov spaces and their distribution spaces can also be char-
acterized by estimates of short-time Fourier transform, (see e. g. [14,
21, 26]). More precisely, let φ P S pRdq be fixed. Then the short-
time Fourier transform Vφf of f P S 1pRdq with respect to the window
function φ is the Gelfand-Shilov distribution on R2d, defined by
Vφfpx, ξq " Fpf φp ´ xqqpξq,
x, ξ P Rd.
If f, φ P S pRdq, then it follows that
2żRd
Vφfpx, ξq " p2πq´ d
fpyqφpy ´ xqe´ixy,ξy dy,
x, ξ P Rd.
By [25, Theorem 2.3] it follows that the definition of the map pf, φq ÞÑ
Vφf from S pRdq S pRdq to S pR2dq is uniquely extendable to a
continuous map from S 1
spR2dq, and restricts to a
continuous map from SspRdqSspRdq to SspR2dq. The same conclusion
holds with Σs in place of Ss, at each place.
In the following propositions we give characterizations of Gelfand-
Shilov spaces and their distribution spaces in terms of estimates of
spRdq to S 1
spRdq S 1
7
the short-time Fourier transform. We omit the proof since the first
part follows from [14, Theorem 2.7]) and the second part from [26,
Proposition 2.2]. See also [8] for related results.
Proposition 1.1. Let s ě 1
2), φ P SspRdqz0 (φ P ΣspRdqz0)
and let f be a Gelfand-Shilov distribution on Rd. Then the following
is true:
2 (s ą 1
(1) f P SspRdq (f P ΣspRdq), if and only if
Vφfpx, ξq À e´rpx
s q,
for some r ą 0 (for every r ą 0).
(2) f P S 1
spRdq (f P Σ1
spRdq), if and only if
1
1
s `ξ
Vφfpx, ξq À erpx
σ q,
for every r ą 0 (for some r ą 0).
1
1
s `ξ
x, ξ P Rd,
x, ξ P Rd,
(1.7)
(1.8)
1.3. Spaces of Hermite series and power series expansions.
Next we recall the definitions of topological vector spaces of Hermite se-
ries expansions, given in [26]. As in [26], it is convenient to use suitable
extensions of R` when indexing our spaces.
Definition 1.2. The sets R5 and R5 are given by
R5 " R`Ťσą0t5σu
and R5 " R5Ťt0u.
Moreover, beside the usual ordering in R, the elements 5σ in R5 and
R5 are ordered by the relations x1 ă 5σ1 ă 5σ2 ă x2, when σ1, σ2, x1
and x2 are positive real numbers such that x1 ă 1
Definition 1.3. Let p P r1,8s, s P R5, r P R, ϑ be a weight on Nd,
and let
2 and x2 ě 1
2.
ϑr,spαq "$&%
Then,
1
erα
2s ,
rαpα!q
when s P R`,
2σ , when s " 5σ,
1
α P Nd.
0pNdq is the set of all sequences tcαuαPNd Ď C on Nd;
(1) ℓ1
(2) ℓ0,0pNdq " t0u, and ℓ0pNdq is the set of all sequences tcαuαPNd Ď
(3) ℓp
C such that cα ‰ 0 for at most finite numbers of α;
rϑspNdq is the Banach space which consists of all sequences
tcαuαPNd Ď C such that
}tcαuαPNd}ℓp
rϑs " }tcαϑpαquαPNd}ℓp ă 8;
(4) ℓ0,spNdq " Şrą0
ℓp
rϑr,sspNdq and ℓspNdq " Ťrą0
jective respective inductive limit topologies of ℓp
respect to r ą 0;
ℓp
rϑr,sspNdq, with pro-
rϑr,sspNdq with
8
(5) ℓ1
0,spNdq " Ťrą0
ℓp
r1{ϑr,sspNdq and ℓ1
inductive respective projective limit topologies of ℓp
with respect to r ą 0.
spNdq " Şrą0
ℓp
r1{ϑr,sspNdq, with
r1{ϑr,sspNdq
Let p P r1,8s, and let ΩN be the set of all α P Nd such that α ď N.
Then the topology of ℓ0pNdq is defined by the inductive limit topology
of the sets
tcαuαPNd P ℓ1
0pNdq ; cα " 0 when α R ΩN(
with respect to N ě 0, and whose topology is given through the semi-
norms
(1.9)
tcαuαPNd ÞÑ }tcαuαďN}ℓppΩN q,
It is clear that these topologies are independent of p. Furthermore, the
topology of ℓ1
0pNdq is defined by the semi-norms (1.9). It follows that
ℓ1
0pNdq is a Fr´echet space, and that the topology is independent of p.
Next we introduce spaces of formal Hermite series expansions
f " ÿαPNd
F " ÿαPNd
cαhα,
tcαuαPNd P ℓ1
0pNdq.
cαeα,
tcαuαPNd P ℓ1
0pNdq.
(1.10)
(1.11)
(1.12)
(1.13)
cαeα (1.14)
and power series expansions
which correspond to
ℓ0,spNdq,
Here
eαpzq "
We consider the mappings
TH : tcαuαPNd ÞÑ ÿαPNd
cαhα
ℓspNdq,
zα
?α!
,
ℓ1
spNdq and ℓ1
0,spNdq.
z P Cd, α P Nd.
and TA : tcαuαPNd ÞÑ ÿαPNd
between sequences, and formal Hermite series and power series expan-
sions.
Definition 1.4. If s P R5, then
H0,spRdq, HspRdq, H1
0,spRdq,
spRdq
and H1
(1.15)
and
A0,spCdq, AspCdq, A1
spCdq
and A1
0,spCdq,
(1.16)
are the images of TH and TA respectively in (1.14) of corresponding
spaces in (1.12). The topologies of the spaces in (1.15) and (1.16) are
inherited from the corresponding spaces in (1.12).
9
Since locally absolutely convergent power series expansions can be
identified with entire functions, several of the spaces in (1.16) are iden-
tified with topological vector spaces contained in ApCdq (see Theo-
rem 1.9 below and the introduction). Here ApΩ0q is the set of all
(complex valued) functions which are analytic in Ω0. (For Ω0 Ď Cd,
contain Ω0.)
ApΩ0q "Ť ApΩq, where the union is taken over all open Ω Ď Cd which
We recall that f P S pRdq if and only if it can be written as (1.10)
such that
for every N ě 0 (cf.
definitions that the inclusions
e. g. [19]).
In particular it follows from the
cα À xαy´N ,
H0pRdq ãÑ H0,spRdq ãÑ HspRdq ãÑ H0,tpRdq
ãÑ S pRdq ãÑ S 1pRdq ãÑ H1
ãÑ H1
0,spRdq ãÑ H1
0,tpRdq ãÑ H1
spRdq
0pRdq, when s, t P R5, s ă t,
(1.17)
are dense.
Remark 1.5. By the definition it follows that TH in (1.14) is a homeo-
morphism between any of the spaces in (1.12) and corresponding space
in (1.15), and that TA in (1.14) is a homeomorphism between any of
the spaces in (1.12) and corresponding space in (1.16).
The next results give some characterizations of HspRdq and H0,spRdq
when s is a non-negative real number.
Proposition 1.6. Let 0 ď s P R and let f P H1
HspRdq (f P H0,spRdq), if and only if f P C 8pRdq and satisfies
0pRdq. Then f P
}H N
d f}L8 À hN N!2s,
(1.18)
for some h ą 0 (every h ą 0). Moreover, it holds
HspRdq " SspRdq ‰ t0u, H0,spRdq " ΣspRdq ‰ t0u when s P p 1
2 ,8q,
HspRdq " SspRdq ‰ t0u, H0,spRdq ‰ ΣspRdq " t0u when s " 1
2 ,
HspRdq ‰ SspRdq " t0u, H0,spRdq ‰ ΣspRdq " t0u when s P p0, 1
2q,
HspRdq ‰ SspRdq " t0u, H0,spRdq " ΣspRdq " t0u when s " 0.
We refer to [26] for the proof of Proposition 1.6.
Due to the pioneering investigations related to Proposition 1.6 by
Pilipovi´c in [17, 18], we call the spaces HspRdq and H0,spRdq Pilipovi´c
In fact, in the
spaces of Roumieu and Beurling types, respectively.
restricted case s ě 1
2, Proposition 1.6 was proved already in [17, 18].
10
Later on it will also be convenient for us to have the following defi-
nition. Here we let Fpz2, z1q and Fpz2, z1q be the formal power series
(1.19)
and ÿ cpα2, α1qeα2pz2qeα1pz1q,
ÿ cpα2, α1qeα2pz2qeα1pz1q
respectively, when Fpz2, z1q is the formal power series
ÿ cpα2, α1qeα2pz2qeα1pz1q.
(1.20)
Here zj P Cdj , j " 1, 2, and the sums should be taken over all pα2, α1q P
Nd2 Nd1.
Definition 1.7. Let d " d2`d1, s P R5, ΘC,1 and ΘC,2 be the operators
pΘC,1Fqpz2, z1q " Fpz2, z1q
and pΘC,2Fqpz2, z1q " Fpz2, z1q
between formal power series in (1.19) and (1.20), zj P Cdj , j " 1, 2.
Then
uAspCd2 Cd1q,
uA1
spCd2 Cd1q,
uA0,spCd2 Cd1q,
uA1
0,spCd2 Cd1q
(1.21)
are the images of (1.16) under ΘC,1, and uApCd2Cd1q and uAd2,d1pt0uq "
uAd2`d1pt0uq are the images of ApCdq and Ad2`d1pt0uq respectively un-
der ΘC,1. The topologies of the spaces in (1.21), uApCd2 Cd1q and
uAd2,d1pt0uq are inherited from the topologies in the spaces (1.16), ApCdq
the spaces in (1.16) can be considered as special cases of uApCd2 Cd1q
and Adpt0uq, respectively.
Remark 1.8. By letting d2 " d and d1 " 0, it follows that ApCdq and
and the spaces in (1.21).
51pCdq " ApCdq and A1
0,51pCdq " Adpt0uq, it follows that
Since A1
uA1
51pCd2 Cd1q " uApCd2 Cd1q,
uA1
0,51pCd2 Cd1q " uAd2,d1pt0uq.
(1.22)
The following results are now immediate consequences of Theorems
4.1, 4.2, 5.2 and 5.3 in [26] and Definition 1.7. Here let
and
κ1,r,spzq "
$''''&''''%
κ2,r,spzq "$&%
erplogxzyq
1
1´2s ,
erz
2σ
σ`1 ,
2
z
2 ´rz
1
s ,
e
erz
2σ
σ´1 ,
e
2
z
2 `rz
1
s ,
11
2
s ă 1
s " 5σ, σ ą 0,
s ě 1
2 ,
s " 5σ, σ ą 1,
s ě 1
2 ,
(1.23)
(1.24)
Theorem 1.9. Let s1, s2 P R5 be such that s2 ą 51, and let κ1,r,s and
κ2,r,s be given by (1.23) and (1.24) respectively, when r ą 0. Then the
following is true:
1
Cd1q such that K À κ1,r,s1 for some r ą 0 (for every 0 ă r ă
2 ).
(1) uAs1pCd2 Cd1q (uA0,s1pCd2 Cd1q) consists of all K P uApCd2
(2) uA1
0,s2pCd2 Cd1q) consists of all K P uApCd2
Cd1q such that K À κ2,r,s2 for every r ą 0 (for some r ą 0).
By Remark 1.8 it follows that Theorem 1.9 remains true after the
s2pCd2 Cd1q (uA1
spaces in (1.21) are replaced by corresponding spaces in (1.16).
1.4. Modulation spaces. Before giving the definition of a broad fam-
ily of modulation spaces, we make a review of mixed normed spaces of
Lebesgue types, adapted to suitable bases of the Euclidean space Rd.
Let E be the ordered basis te1, . . . , edu of Rd. Then the ordered basis
E1 " te1
1, . . . , e1
du (the dual basis of E) satisfies
xej, e1
ky " 2πδjk
for every j, k " 1, . . . , d.
The corresponding parallelepiped, lattice, dual parallelepiped and dual
lattice are given by
κpEq " t x1e1 ` ` xded ; px1, . . . , xdq P Rd, 0 ď xk ď 1, k " 1, . . . , d u,
ΛE " t j1e1 ` ` jded ; pj1, . . . , jdq P Zd u,
κpE1q " t ξ1e1
and
1 ` ` ξde1
d ; pξ1, . . . , ξdq P Rd, 0 ď ξk ď 1, k " 1, . . . , d u,
E " ΛE1 " t ι1e1
Λ1
1 ` ` ιde1
d ; pι1, . . . , ιdq P Zd u,
respectively. Note here that the Fourier analysis with respect to general
biorthogonal bases has recently been developed in [20].
We observe that there is a matrix TE such that e1, . . . , ed and e1
are the images of the standard basis under TE and TE1 " 2πpT ´1
spectively.
1, . . . , e1
E qt, re-
d
In the following we let
maxpqq " maxpq1, . . . , qdq and minpqq " minpq1, . . . , qdq
when q " pq1, . . . , qdq P r1,8sd.
Definition 1.10. Let E be an ordered basis of Rd and p " pp1, . . . , pdq P
r1,8sd. If f P L1
locpRdq, then }f}Lp
is defined by
E
where gkpzkq, zk P Rd´k, k " 0, . . . , d ´ 1, are inductively defined as
g0px1, . . . , xdq " fpx1e1 ` ` xdedq,
px1, . . . , xdq P Rd,
}f}Lp
E " }gd´1}Lpd pRq
12
and
gkpzkq " }gk´1p , zkq}Lpk pRq,
EpRdq consists of all f P L1
The space Lp
and is called E-split Lebesgue space (with respect to p).
zk P Rd´k, k " 1, . . . , d ´ 1.
locpRdq such that }f}Lp
is finite,
E
Next we discuss suitable conditions for bases in the phase space R2d.
We let σpX, Y q be the standard symplectic form on the phase space,
given by
σpX, Y q " xy, ξy ´ xx, ηy,
X " px, ξq P R2d, Y " py, ηq P R2d.
We notice that if
is the standard basis of R2d, then
te1, . . . , ed, ε1, . . . , εdu
(1.25)
σpej, ekq " 0, σpej, εkq " ´δj,k,
(1.26)
when j, k P t1, . . . , du. More generally, a basis in (1.25) for the phase
space R2d is called symplectic if (1.26) holds. A symplectic basis (1.25)
for R2d is called phase split if e1, . . . , ed and ε1, . . . , εd span
and σpεj, εkq " 0,
tpx, 0q P R2d ; x P Rd u
and tp0, ξq P R2d ; ξ P Rd u,
respectively.
Next we give the definition of our class of modulation spaces.
Definition 1.11. Let E be an ordered basis for R2d, p P r1,8s2d,
φpxq " π´ d
and let ω be a weight on R2d. Then the modulation
space M p
4 e´ 1
E,pωqpRdq consists of all f P H1
51pRdq such that
2 x2
}f}M p
E,pωq " }Vφf ω}Lp
E
(1.27)
is finite.
2 x2
4 e´ 1
and f P H1
We remark that if φpxq " π´ d
51pRdq, then px, ξq ÞÑ
Vφfpx, ξq is a smooth function (cf. [26]). Furthermore, by [26, Theorem
4.8] we get the following. The proof is omitted.
Proposition 1.12. Let E be an ordered basis for R2d, p P r1,8s2d
and let ω be a weight on R2d. Then M p
E,pωqpRdq is a Banach space with
norm given by (1.27).
If the weight ω in Definition 1.11 is a moderate weight, then we can
E,pωqpRdq. In what follows we let p1 P r1,8s be
say more concerning M p
the conjugate exponent of p P r1,8s, i. e. 1
Proposition 1.13. Let E be an ordered basis for R2d, p P r1,8s2d and
let ω, v P PEpR2dq be such that ω is v-moderate. Then the following is
true:
p1 " 1.
p ` 1
13
(1) Σ1pRdq ãÑ M p
E,pωqpRdq ãÑ Σ1
then Σ1pRdq is dense in M p
1pRdq. If in addition maxppq ă 8,
E,pωqpRdq;
(2) if φ P M 1
pvqpRdqzt0u and f P Σ1
E,pωqpRdq,
1pRdq, then f P M p
if and only if the right-hand side of (1.27) is finite. Further-
pvqpRdqzt0u in (1.27) give rise
more, different choices of φ P M 1
to equivalent norms;
(3) M p
(4) if p
1, . . . , p1
E,pωqpRdq increases with p1, . . . , p2d and decreases with ω;
1 " pp1
2dq, then the restriction of the L2pRdq scalar
product p , q to Σ1pRdq is uniquely extendable to a (semi-
E,p1{ωqpRdq. If in
conjugate) duality between M p
E,pωqpRdq can be iden-
addition maxppq ă 8, then the dual of M p
tified by M p
E,pωqpRdq and M p
E,p1{ωqpRdq through the form p , q.
1
1
Proposition 1.13 follows by similar arguments as in Chapters 11 and
12 in [12] (see also [25, 26]).
Remark 1.14. In some sense, the variable x at the weight ωpx, ξq in the
definition of modulation spaces quantify growth and decay properties
for the involved functions or distributions. In the same way the variable
ξ quantify regularity or lack of regularity for the involved functions or
distributions.
2 x2
4 e´ 1
By the analysis in [26] it follows that there are no bounds on how
fast Vφf may grow or decay at infinity when φpxq " π´ d
is fixed,
x P Rd, and f is taken in the class H1
51pRdq. Since weights in PEpR2dq
are bounded by exponential functions, the restrictions of the weights in
Proposition 1.13 are significantly stronger compared to what is the case
in Proposition 1.12. A question here concerns wether it is possible to
extend parts of Proposition 1.13 to larger weight classes than PEpR2dq
or not.
It seems that the invariance properties (2) in Proposition 1.13 con-
cerning the choice of weight function are not possible for weights that
are not moderate. On the other hand, (1) and (4) in Proposition 1.13
hold true for certain weights outside PEpR2dq. In fact, in [25], certain
weight classes which contain PEpR2dq as well as weights of the form
ωpx, ξq " erpx
1
s `ξ
1
s q,
x, ξ P Rd,
when r ą 0 and s ą 1
2 are introduced. For corresponding (broader)
families of modulation spaces it is then proved that Proposition 1.13
(1) and (4) hold true (with some modifications).
1.5. Bargmann transform and spaces of analytic functions.
The Bargmann transform Vd is the homeomorphism from the spaces
in (1.15) to respective spaces in (1.16), given by TA T ´1
H , where TH
and TA are given by (1.14).
14
We notice that if f P LppRdq for some p P r1,8s, then Vdf is the
entire function given by
pVdfqpzq " π´d{4żRd
which can also be formulated as
exp´´
pVdfqpzq "żRd
or
1
2pxz, zy`y2q`21{2xz, yy¯fpyq dy,
z P Cd,
Adpz, yqfpyq dy,
z P Cd,
(1.28)
z P Cd, y P Rd.
Adpz, yq " π´d{4 exp´ ´
(Cf. [1, 2].) Here
pVdfqpzq " xf, Adpz, qy,
where the Bargmann kernel Ad is given by
1
z P Cd,
2pxz, zy ` y2q ` 21{2xz, yy¯,
dÿj"1
z " pz1, . . . , zdq P Cd
xz, wy "
and pz, wq " xz, wy
zjwj
when
and w " pw1, . . . , wdq P Cd,
and otherwise x , y denotes the duality between test function spaces
and their corresponding duals which is clear form the context. We note
that the right-hand side in (1.28) makes sense when f P S 1
1{2pRdq and
defines an element in ApCdq, since y ÞÑ Adpz, yq can be interpreted as
an element in S1{2pRdq with values in ApCdq.
It was proved by Bargmann that f ÞÑ Vdf is a bijective and iso-
metric map from L2pRdq to the Hilbert space A2pCdq, the set of entire
functions F on Cd which fullfils
}F}A2 "´żCd Fpzq2dµpzq¯1{2
ă 8.
(1.29)
dλpzq, where dλpzq is the Lebesgue measure
Recall, dµpzq " π´de´z2
on Cd, and the scalar product on A2pCdq is given by
pF, GqA2 "żCd
pF, Gq ÞÑ xF, GyA2 " xF, GyA2pCdq "żCd
FpzqGpzq dµpzq, F, G P A2pCdq.
For future references we note that the latter scalar product induces the
bilinear form
FpzqGpzq dµpzq
(1.30)
(1.31)
on A2pCdq A2pCdq.
In [1] it was proved that the orthonormal basis thαuαPNd in L2pRdq
of Hermite functions is mapped to the orthonormal basis teαuαPNd in
A2pCdq (cf. (1.13)). Furthermore, there is a convenient reproducing
15
In fact, let ΠA be the operator from L2pdµq to
Fpwqepz,wq dµpwq,
z P Cd.
(1.32)
formula on A2pCdq.
ApCdq, given by
pΠAFqpzq "żCd
Then it is proved in [1] that ΠA is an orthonormal projection from
L2pdµq to A2pCdq.
Fourier transform is given by
From now on we assume that φ in the definition of the short-time
φpxq " π´d{4e´x2{2,
x P Rd,
(1.33)
if nothing else is stated. For such φ, it follows by straight-forward
computations that the relationship between the Bargmann transform
and the short-time Fourier transform is given by
and U ´1
V Vd " Vφ,
Vd " UV Vφ,
(1.34)
where UV is the linear, continuous and bijective operator on D 1pR2dq »
D 1pCdq, given by
pUVFqpx ` iξq " p2πqd{2epx2`ξ2q{2e´ixx,ξyFp21{2x,´21{2ξq,
cf. [25].
x, ξ P Rd,
(1.35)
Definition 1.15. Let E be an ordered basis for R2d, UV be the oper-
ator in (1.35), p P r1,8s2d, φpxq " π´ d
and let ω be a weight
on R2d.
4 e´ 1
2 x2
(1) The space B p
E,pωqpCdq consists of all F P L1
}F}Bp
E,pωq " }pU ´1
V Fq ω}Lp
E
locpCdq such that
is finite;
(2) The space Ap
with topology inherited from B p
E,pωqpCdq consists of all F P ApCdqŞ B p
E,pωqpCdq.
E,pωqpCdq
We note that the spaces in Definition 1.15 are normed spaces when
minppq ě 1.
For conveneincy we set }F}Bp
E,pωq " 8, when F R B p
E,pωqpCdq is mea-
E,pωq " 8, when F P ApCdqzB p
surable, and }F}Ap
Remark 1.16. In Definitions 1.11 and 1.15, important cases appear
when E is the standard basis for R2d and p1 " " pd " p P r1,8s
and pd`1 " " p2d " q P r1,8s. For such choices of E and p we set
Lp,q " Lp
E,
M p,q
E,pωqpCdq.
and
pωq " M p
E,pωq,
Ap,q
pωq " Ap
E,pωq
Bp,q
pωq " B p
E,pωq.
We also set
M p
pωq " M p,p
pωq,
Ap
pωq " Ap,p
pωq
16
and
Bp
pωq " Bp,p
pωq.
If in addition ω " 1, then we set
M p,q
M p
E,pωq " M p
E,
and similarly for Ap
E,pωq and B p
pωq " M p,q
E,pωq spaces.
and M p
pωq " M p,
If instead E " ted`1, . . . , e2d, e1, . . . , edu where e1, . . . , e2d is the stan-
dard basis for R2d and p1 " " pd " q P r1,8s and pd`1 " "
p2d " p P r1,8s, then we set Lp,q
W p,q
pωq " M p
E,pωq, Ap,q
and Bp,q
,pωq " B p
E,pωq.
" Lp
E,
,pωq " Ap
E,pωq
We notice that the space W p,q
pωq in Remark 1.16 is an example of a
(weighted) Wiener amalgam space (cf. [9, 10]).
For future references we observe that the Bp
pωq norm is given by
pωq " 2d{pp2πq´d{2żCd e´z2{2Fpzqωp21{2zqp dλpzq1{p
}F}Bp
" 2d{pp2πq´d{2ijR2de´px2`ξ2q{2Fpx ` iξqωp21{2x,´21{2ξqp dxdξ1{p
(1.36)
(with obvious modifications when p " 8). Especially it follows that
the norm and scalar product in B2
pωqpCdq take the forms
}F}B2
pF, GqB2
pωq "żCd Fpzqωp21{2zq2 dµpzq1{2
pωq "żCd
FpzqGpzqωp21{2zq2 dµpzq,
,
F P B2
pωqpCdq
F, G P B2
pωqpCdq
(cf. (1.29) and (1.30)).
By the definitions and (1.34) it follows that the Bargmann transform
E,pωqpCdq. In fact, we
is an isometric injection from M p
have the following refinement. We omit the proof since the result is a
special case of Theorem 4.8 in [26].
E,pωqpRdq to Ap
Proposition 1.17. Let E be an ordered basis for R2d, p P r1,8s2d, and
ω be a weight on R2d. Then the Bargmann transform is an isometric
bijection from M p
E,pωqpRdq to Ap
E,pωqpCdq.
Finally, the SCB transform (i. e. the Semi Conjugated Bargmann
transform), VΘ,d2,d1 is defined as ΘC,1 Vd2`d1. We also set VΘ,d "
VΘ,d,d. Evidently, all properties of the Bargmann transform carry over
to analogous properties for the SCB transform. Assume that E is a ba-
sis for R2d2 R2d1, p P r1,8s2d2`2d1, p, q P p0,8s and that ω is a weight
17
der the map ΘC,1 with the topology defined by the norm
E,pωqpCd2`d1q is the image of Ap
E,pωqpCd2`d1q un-
on R2d2 R2d1 Then uAp
,
E,pΘC,1ωq
E,pωqpCd2`d1q.
E,pωq " }ΘC,1a}Ap
}a}uAp
The spaces
uAp,q
uAppCd2Cd1q
pωqpCd2Cd1q,
and their norms, and the scalar product p , quA2 are defined analo-
uAp,qpCd2Cd1q and
uAp
pωqpCd2Cd1q,
a P uAp
gously.
1.6. Pseudo-differential operators. Next we recall some properties
in pseudo-differential calculus. Let Mpd, Ωq be the set of d d-matrices
with entries in the set Ω, a P Σ1pR2dq, and let A P Mpd, Rq be fixed.
Then the pseudo-differential operator OpApaq is the linear and contin-
uous operator on Σ1pRdq, given by
pOpApaqfqpxq " p2πq´dij apx´Apx´yq, ξqfpyqeixx´y,ξy dydξ,
x P Rd.
(1.37)
1pR2dq, the pseudo-differential operator OpApaq is
1pRdq with distri-
For general a P Σ1
defined as the continuous operator from Σ1pRdq to Σ1
bution kernel
Ka,Apx, yq " p2πq´d{2pF ´1
x, y P Rd. (1.38)
Here F2F is the partial Fourier transform of Fpx, yq P Σ1
1pR2dq with
respect to the y variable. This definition makes sense since the map-
pings
2 aqpx ´ Apx ´ yq, x´ yq,
F2
are homeomorphisms on Σ1
a homeomorphism on Σ1
and Fpx, yq ÞÑ Fpx ´ Apx ´ yq, x ´ yq
(1.39)
1pR2dq. In particular, the map a ÞÑ Ka,A is
The standard (Kohn-Nirenberg) representation, apx, Dq " Oppaq,
and the Weyl quantization Opwpaq of a are obtained by choosing A " 0
and A " 1
2 I, respectively, in (1.37) and (1.38), where I is the identity
matrix.
1pR2dq.
Remark 1.18. By Fourier's inversion formula, (1.38) and the kernel the-
orem [16, Theorem 2.2], [22, Theorem 2.5] for operators from Gelfand-
Shilov spaces to their duals, it follows that the map a ÞÑ OpApaq is
bijective from Σ1
1pR2dq to the set of all linear and continuous operators
from Σ1pRdq to Σ1
1pR2dq.
By Remark 1.18, it follows that for every a1 P Σ1
Mpd, Rq, there is a unique a2 P Σ1
By Section 18.5 in [15], the relation between a1 and a2 is given by
1pR2dq and A1, A2 P
1pR2dq such that OpA1pa1q " OpA2pa2q.
OpA1pa1q " OpA2pa2q ô a2 " eixpA1´A2qDξ,Dxya1.
18
(1.40)
Here we note that the operator eixADξ ,Dxy is homeomorphic on Σ1pR2dq
and its dual (cf. [5,6,29]). For modulation spaces we have the following
subresult of Proposition 2.8 in [28].
Proposition 1.19. Let s ě 1
Σ1pR2dq and let TA " eixADξ,Dxy. If ω P PEpR4dq and
ωApx, ξ, η, yq " ωpx ` Ay, ξ ` Aη, η, yq,
2, A P Mpd, Rq, p, q P p0,8s, φ, a P
then TA from Σ1pR2dq to Σ1pR2dq extends uniquely to a homeomor-
phism from M p,q
pωqpR2dq to M p,q
pωAqpR2dq, and
}TAa}M p,q
pωAq -- }a}M p,q
pωq
.
(1.41)
2. Kernel theorems and analytic pseudo-differential
operators
s) and mappings with kernels in uAs (uA1
In the first part of the section we show that there is a one to one
correspondence between linear and continuous mappings from A1
s to
As (As to A1
s) with respect to
the measure dµ (cf. Propositions 2.2 and 2.3). Thereafter we deduce
in Theorems 2.7 -- 2.10 analogous results for analytic pseudo-differential
operators based on Theorem 2.6 which deals with mapping properties
of the operator which takes apz, wq into epz,wqapz, wq.
Here and in what follows, any extension of the A2-form, p , qA2
from A0pCdq A0pCdq to C is still called A2-form and still denoted by
p , qA2. Similar approaches yield extensions of the forms x , yA2 and
p , quA2.
0,spNdq are the duals of ℓspNdq and
ℓ0,spNdq, respectively, through unique extensions of the ℓ2pNdq form on
ℓ0pNdq. Since the spaces in (1.16) are images of the spaces in (1.12)
under the map TA in (1.14), the following lemma is an immediate
consequence of these duality properties. The result is also implicitly
given in [7, 26].
By the definitions, ℓ1
spNdq and ℓ1
Lemma 2.1. Let s P R5. Then the following is true:
(1) the form pF, Gq ÞÑ pF, GqA2 from A0pCdq A0pCdq to C is
spCdq
0,spCdq to C. Furthermore, the
spCdq and
uniquely extendable to continuous forms from AspCdq A1
to C, and from A0,spCdq A1
duals of AspCdq and A0,spCdq can be identified by A1
A1
0,spCdq through the form p , qA2;
(2) the form pF, Gq ÞÑ xF, GyA2 from A0pCdq A0pCdq to C is
spCdq
0,spCdq to C. Furthermore, the
spCdq and
uniquely extendable to continuous forms from AspCdq A1
to C, and from A0,spCdq A1
duals of AspCdq and A0,spCdq can be identified by A1
A1
0,spCdq through the form x , yA2.
19
The following two propositions follow by applying VΘ,d2,d1 on Theo-
rem 3.3 and 3.4 in [7], and using Lemma 2.1. The details are left for
the reader.
Proposition 2.2. Let s P R5, and let T be a linear and continuous
map from A0pCd1q to A1
spCd1q to AspCd2q,
0pCd2q. Then the following is true:
(1) if T is a linear and continuous map from A1
then there is a unique K P uAspCd2 Cd1q such that
T F "`z2 ÞÑ xKpz2, q, FyA2pCd1 q
holds true;
(2.1)
(2) if T is a linear and continuous map from AspCd1q to A1
spCd2q,
spCd2 Cd1q such that (2.1) holds
A1
true.
then there is a unique K P uA1
The same holds true if As, uAs, A1
0,s and uA1
Proposition 2.3. Let K P uA1
0,s, respectively, at each occurrence.
s and uA1
s are replaced by A0,s, uA0,s,
0pCd2 Cd1q, s P R5 and let T be the
(2.2)
linear and continuous map from A0pCd1q to A1
0pCd2q, given by
F ÞÑ T F "`z2 ÞÑ xKpz2, q, FyA2pCd1 q.
Then the following is true:
continuous map from A1
spCd1q to AspCd2q;
continuous map from AspCd1q to A1
spCd2 Cd1q, then T extends uniquely to a linear and
(1) if K P uAspCd2 Cd1q, then T extends uniquely to a linear and
(2) if K P uA1
s are replaced by A0,s, uA0,s,
The same holds true if As, uAs, A1
0,s and uA1
F P A0pCd1q, G P A0pCd2q.
Next we recall the definition of analytic pseudo-differential operators.
The operator T in (2.2) should be interpreted as T in the formula
pT F, GqA2pCd2 q " pK, GbFqA2pCd2 Cd1 q,
0,s, respectively, at each occurrence.
s and uA1
spCd2q.
A1
(See [25, Definition 6.20] in the case t " 0, as well as [3, 4].)
differential operator OpVpaq (AΨDO) with symbol a is given by
51pCd Cdq. Then the analytic pseudo-
Definition 2.4. Let a P uA1
pOpVpaqFqpzq "żCd
apz, wqFpwqepz,wq dµpwq
By the definition it follows that the relation between the operator
" pF, apz, qep ,zqqA2pCdq,
z P Cd.
(2.3)
kernel K and the symbol a is given by
Kpz, wq " epz,wqapz, wq,
z, w P Cd,
provided the multiplication on the right-hand side makes sense.
20
This leads to the question about mapping properties of Tt defined
by
pTtaqpz, wq " etpz,wqapz, wq,
z, w P Cd, t P C
spC2dq.
(2.4)
the inverse of Tt is T´t. Hence Tt is well-defined and a homeomorphism
when a belongs to a suitable subspace of uA1
First we notice that a P uApC2dq, if and only if Tta P uApC2dq, and that
on uA1
If TH is the same as in (1.14) then we shall investigate the map T0,t
51pC2dq.
in the commutative diagram:
ℓ1
51pN2dq
51pN2dq
T0,tÝÝÝÑ ℓ1
TH§§đ
uA51pC2dq ÝÝÝÑTt
§§đTH
uA51pC2dq.
51pC2dq with the expansion
cpα, βq
cpα, βqeαpzqeβpwq " ÿα,βPNd
cpα, βq À rα`βaα!β!
tγzγwγ
γ!
,
z, w P Cd,
(2.5)
zαwβ
?α!β!
,
z, w P Cd,
(2.6)
Therefore, let a P uA1
apz, wq " ÿα,βPNd
where
for every r ą 0. Since
we have
etpz,wq " ÿγPNd
etpz,wqapz, wq " ÿγPNd ÿα,βPNd
where
ϕt,z,wpα, β, γq,
z, w P Cd, t P C,
(2.7)
ϕt,z,wpα, β, γq "
cpα, βqtγzα`γwβ`γ
γ!?α!β!
,
z, w P Cd, t P C,
(2.8)
We shall prove that the series in (2.7) is locally uniformly convergent
with respect to t, z and w. If t ă R, z ă R and w ă R for some
fixed R ą 0, then by (2.6) we get
ϕt,z,wpα, β, γq À
γ!
for all α, β, γ P Nd. Since the series
rα`βRα`β`2γ
ď prRqα`βpdR2qγ
γ!
ÿα,β,γPNd
prRqα`βp2R2qγ
γ!
is convergent when r is chosen strictly smaller than R´1, the asserted
uniform convergence follows from Weierstass' theorem.
21
In particular, we may change the order of summation in (2.7) to
obtain
pTtaqpz, wq " ÿα,βPNd ÿγPNd
" ÿα,βPNd ÿγPNd
cpα, βqtγzα`γ wβ`γ
γ!?α!β!
γ 1{2
γ β ` γ
cpα, βqtγα ` γ
" ÿα,βPNdpT0,tcqpα, βqeαpzqeβpwq,
γβ
cpα ´ γ, β ´ γqtγα
γ1{2
where
pT0,tcqpα, βq " ÿγďα,β
eα`γpzqeβ`γpwq
z, w P Cd, t P C,
,
t P C, (2.9)
and we have identified T0,t in the diagram (2.5).
We have now the following:
Proposition 2.5. Let t P C, s, s0 P R5 be such that s ă 1{2 and
0 ă s0 ď 1{2, and let T0,t be the map on ℓ1
0pN2dq given by (2.9). Then
T0,t is a continuous and bijective map on ℓ1
0pN2dq with the inverse T0,´t.
Furthermore, T0,t restricts to homeomorphism from ℓ1
spN2dq,
and from ℓ1
spN2dq to ℓ1
0,s0pN2dq to ℓ1
Proof. The topology on ℓ1
norms
0,s0pN2dq.
0pN2dq can be defined by the family of semi-
pNptcpα, βquα,βPNdq " sup
0pN2dq we have
αďN
Then, for a given c P ℓ1
pNpT0,tpcqq ď sup
α,βďN ÿγďα,β
βďN cpα, βq, N P N.
sup
γβ
γ1{2
cpα ´ γ, β ´ γqtγα
ď pNpcq ÿγďN
tγ2N
ď 2Np1 ` tqN dpNpcq,
0pN2dq follows.
t P C, N P N,
and the continuity of T0,t on ℓ1
By straight-forward computations it also follows that T0,´t is the
inverse of T0,t, which gives asserted homeomorphism properties of T0,t
on ℓ1
0pN2dq.
Next we consider the case when s, s0 P R`. Assume that
cpα, βq À e
1
h pα1{2s`β1{2sq, α, β P Nd,
22
1
e
where
for some constant h ą 0. Then
h pα´γ1{2s`β´γ1{2sqtγα
pT0,tcqpα, βq À ÿγďα,β
Ipαq "ÿγďαα
h pα´γ1{2sqtγ¸1{2
γe
γtγ¸1{2
h α1{2sÿγďαα
and similarly for Ipβq. Since
Ipαq ď e
" e
h α1{2s
,
1
1
2
we get
γβ
γ1{2
ď IpαqIpβq,
t P C, α P Nd,
p1`tqdα{2,
t P C, α P Nd,
pT0,tcqpα, βq À e
1
h pα1{2s`β1{2sqp1 ` tqdpα`βq{2 À e
2
h pα1{2s`β1{2sq,
t P C, α, β P Nd, where the last inequality follows from the fact that
s ă 1{2. This gives the continuity assertions for T0,t in the case when
s, s0 P R` and s, s0 ă 1{2.
For s0 " 1{2 we have
1
pT0,tcqpα, βq À e
for some other choice of h1 ą 0 which only depend on t, d and h and
the continuity of T0,t on ℓ1
h pα`βqp1`tqdpα`βq{2 " e
t P C, α, β P Nd,
pα`βq,
1
h1
0, 1
2pN2dq follows.
It remains to consider the case when s " s0 " 5σ for some σ ą 0.
Assume that
cpα, βq ď Crα`β pα!β!q
1
2σ , α, β P Nd,
1
2σ tγα
γβ
γ1{2
for some constants C, r ą 0. Then
rα`β´2γ ppα ´ γq!pβ ´ γq!q
pT0,tcqpα, βq ď C ÿγďα,β
ď Cr ÿγďα,β
ď Crp2rqα`β pα!β!q
1
rα`β pα!β!q
2σ tγ2α`β{2
2σ ÿγďα`β
ď Crp2rp1 ` tqdqα`β pα!β!q
tγ
1
1
t P C, α, β P Nd,
where Cr ą 0 only depends on C and r. This shows that T0,t is contin-
uous on ℓ1
2σ ,
(cid:3)
5σpN2dq and on ℓ1
0,5σpN2dq.
We have now the following:
23
0,s0pC2dq.
spC2dq to uA1
Proof. By the commutative diagram (2.5) we have
0,1{2pC2dq to uA1
Theorem 2.6. Let t P C, s, s0 P R5 be such that s ă 1
true:
2 and 0 ă s0 ď 1
2,
51pC2dq. Then the following is
0,1{2pC2dq;
0,1{2pC2dq extends uniquely to home-
0,s0pC2dq to
and let Tt be given by (2.4) when a P uA1
(1) Tt restricts to a homeomorphism from uA1
(2) Tt from uA1
omorphisms from uA1
uA1
0,1{2pC2dq to uA1
spC2dq and from uA1
a P uA51pC2dq,
1 qa for general a P uA1
and letting Tta " pT1 T0,t T ´1
nuity assertions follow from Proposition 2.5.
It remains to prove the uniqueness. Let a P uA1
respectively, and let cTtapα, βq be the coefficients of Tta P uA1
with the corresponding expansion coefficients capα, βq, and cbpα, βq,
0pC2dq,
α, β P Nd. Then
0pC2dq, b P uA0pC2dq,
Tta " pT1 T0,t T ´1
1 qa,
0pC2dq, the conti-
pa, bquA2 " ÿα`βďN
capα, βqcbpα, βq,
for some N P N depending on b. Now choose a sequence aj P uA0pC2dq
such that
(2.10)
lim
jÑ8paj, bquA2 " pa, bquA2,
If cajpα, βq denote the coefficients in the expansion of aj, j P N, then
it follows from (2.10)
for every b P uA0pC2dq.
for every α, β P Nd
lim
jÑ8
cajpα, βq " capα, βq,
(2.11)
by taking bpz, wq " eαpzqeβpwq. The uniqueness follows if we prove
that
lim
jÑ8ppTtajq, bquA2 " ppTtaq, bquA2,
Let the coefficients of Ttaj be denoted by ct,ajpα, βq, α, β P Nd. By
(2.9) and (2.11) we get ct,ajpα, βq Ñ ct,apα, βq as j Ñ 8, for every
pα, βq P N2d, and (2.12) follows since
for every b P uA0pC2dq.
(2.12)
ppTtajq, bquA2 " ÿα`βďN
ct,ajpα, βqcbpα, βq,
and
ppTtaq, bquA2 " ÿα`βďN
where N depends on b only.
ct,apα, βqcbpα, βq,
b P uA0pC2dq,
(cid:3)
The following two theorems now follows by combining Propositions
2.2 and 2.3 with Theorem 2.6. The details are left for the reader.
24
A0,s, A1
a P uA1
Theorem 2.8. Let a P uA1
If a P uA1
0,s and uA1
0,s and uA1
s and uA1
s and uA1
Theorem 2.7. Let s P R5 be such that s ă 1
and continuous map from AspCdq to A1
spCd Cdq such that T " OpVpaq.
2, As, A1
The same holds true if s ă 1
2 and let T be a linear
spCdq. Then there is a unique
s are replaced by s ď 1
2,
0,s, respectively, at each occurrence.
0pCd Cdq and s P R5 be such that s ă 1
2.
spCd Cdq, then OpVpaq extends uniquely to a linear and
s are replaced by s ď 1
2,
continuous map from AspCdq to A1
A0,s, A1
The same holds true if s ă 1
0,s, respectively, at each occurrence.
spCdq.
2, As, A1
The analogous results to Theorems 2.7 and 2.8 for larger s are equiv-
alent to kernel theorems for Fourier invariant Gelfand-Shilov spaces.
Theorem 2.9. Let s ě 1
2 ). Then the following is true:
2 (s ą 1
(1) If T is a linear and continuous map from A1
spCdq to AspCdq
(from A1
Cdq such that
0,spCdq to A0,spCdq), then there is a unique a P uApCd
1
apz, wq À e
1
2 z´w2´rpz
1
s `w
s q,
z, w P Cd,
(2.13)
for some (for every) r ą 0 and T " OpVpaq;
(2) If T is a linear and continuous map from AspCdq to A1
spCdq
(from A0,spCdq to A1
Cdq such that
0,spCdq), then there is a unique a P uApCd
1
apz, wq À e
1
2 z´w2`rpz
1
s `w
s q,
z, w P Cd,
(2.14)
for every (for some) r ą 0 and T " OpVpaq.
Theorem 2.10. Let s ě 1
2 (s ą 1
2 ). Then the following is true:
0, then OpVpaq from A0pCdq to A1
to a linear and continuous map from A1
A1
0,spCdq to A0,spCdq);
(1) If a P uApCd Cdq satisfies (2.13) for some (for every) r ą
(2) If a P uApCd Cdq satisfies (2.14) for every (for some) r ą
0, then OpVpaq from A0pCdq to A1
to a linear and continuous map from AspCdq to A1
A0,spCdq to A1
0pCdq is uniquely extendable
spCdq (from
0pCdq is uniquely extendable
spCdq to AspCdq (from
0,spCdq).
Proof. We only prove the results in the Roumieu case. The Beurling
case follows by similar arguments and is left for the reader.
If T is the same as in (2.2) for some K P uApCd Cdq, then T "
OpVpaq when apz, wq " e´pz,wqKpz, wq, pz, wq P Cd Cd. Since
1
e´pz,wqe
2 pz2`w2q " e´ 1
25
2 z´w2
,
z, w P Cd,
Theorem 1.9 gives
K P uAspCd Cdq ô Kpz, wq À e
ô apz, wq À e
for some r ą 0. In the same way,
1
2 pz2`w2q´rpz
1
s `w
1
s q
1
2 pz´w2q´rpz
1
s `w
1
s q,
z, w P Cd.
1
s `w
1
1
s q,
2 pz´w2q`rpz
spCd Cdq ô apz, wq À e
z, w P Cd,
for every r ą 0. The results now follows from these relations and
Propositions 2.2 and 2.3
K P uA1
Remark 2.11. For strict subspaces of uA1
0,1{2pCdq in Definition 1.7, the
estimates imposed on their elements are given by (1.23) or by (1.24)
for suitable assumptions on r ą 0. It is evident that in all such cases,
these conditions are violated under the action of Tt in Theorem 2.6
when t ‰ 0. Hence, Theorem 2.6 cannot be extended to other spaces
in Definition 1.7.
In particular, the conditions (2.13) and (2.14) in Theorems 2.9 and
2.10 can not be replaced by the convenient condition that a should
belong to e. g.
(cid:3)
2 and s2 ą 1
uAspCd Cdq,
uA0,spCd Cdq,
uA1
s1pCd Cdq or
uA1
0,s2pCd Cdq
(2.15)
when s P R5, s1 ě 1
2. On the other hand, the conditions on
a in Theorems 2.9 and 2.10 means exactly that pz, wq ÞÑ epz,wqapz, wq
belongs to the spaces in (2.15), depending on the choice between (2.13)
and (2.14), and the condition on r.
Remark 2.12. Let s P R5 be such that s ď 1
2 . By similar arguments as
in the proofs of Theorems 2.9 and 2.10, one may also characterize linear
and continuous operators from A1
0,spCdq
to A0,spCdq as operators of the form OpVpaq for suitable conditions on
a. The details are left for the reader.
spCdq to AspCdq, and from A1
3. Operators with kernels and symbols in mixed weighted
Lebesgue spaces
In this section we focus on operators in the previous section, whose
kernels should belong to uApCd Cdq and obey certain mixed norm
estimates of Lebesgue types. We deduce continuity properties of such
operators when acting between suitable Lebesgue spaces of analytic
functions. (See Theorems 3.3 -- 3.5.) Thereafter we show that our results
can be used to regain well-known and sharp continuity results in [27] for
pseudo-differential operators with symbols in modulation spaces when
acting on other modulation spaces. (See Theorems 3.8 and 3.9.) A key
step here is to deduce an explicit formula which relates the short-time
Fourier transform of the symbol to a real pseudo-differential operator
26
Oppaq with the Bargmann transform of the kernel to Oppaq.
Lemma 3.7.)
(See
We shall consider Lebesgue norm conditions of matrix pull-backs of
the involved kernels. Let
C l
C 3
j,k C 2
j,k
j,k C 4
jk P Mpd, Rq, Cjk "C 1
C21 C22¸ P Mp4d, Rq,
C "C11 C12
Udpx ` iξq " px, ξq P R2d,
j,k¸ P Mp2d, Rq,
j, k, l P Z`,
and let
Ud,dpx1 ` iξ1, x2 ` iξ2q " px1, ξ1, x2, ξ2q P R4d,
?2 z,?2 wq,
Kωpz, wq " e´ 1
2 pz2`w2qKpz, wq ωp
(3.2)
GK,C,ω " Kω U ´1
d,d C Ud,d,
z, w P Cd, j " 1, 2.
x, xj, ξ, ξj P Rd,
E,pω2qpCdq,
We will consider continuity of operators from Ap1
when GK,C,ω fullfils suitable Lp,qpC2dq estimates, where the weights full-
fil
E,pω1qpCdq to Ap2
(3.1)
(3.3)
(3.4)
(3.5)
(3.6)
pC2dq be the sets of
pC2dq " }GUd,d}Lp,q
pR4dq,
ω2pzq
ω1pwq À ωpz, wq,
z, w P Cd.
Here and in what follows we let Lp,qpC2dq and Lp,q
all G P L1
and }G}Lp,q
}G}Lp,qpC2dq " }GUd,d}Lp,qpR4dq
respectively, are finite. (See also Remark 1.16.)
locpC2dq such that
The involved Lebesgue exponents should satisfy
1
p1 ´
1
p2 " 1 ´
1
p ´
1
q
,
q ď p2 ď p,
p, q P r1,8s, p1, p2 P r1,8s2d.
We need that C and Cjk above should satisfy
detpCq detpC11C21q ‰ 0
or
In (3.4) and in what follows we use the convention
detpCq detpC12C22q ‰ 0.
1
p " 1
p1
, . . . ,
1
pd ,
p0 ď p, p ď p,
q0 ă q,
q ă q
and r " r,
27
when
p " pp1, . . . , pdq,
q " pq1, . . . , qdq,
r " pr1, . . . , rdq
belong to r1,8sd and p, q, r, p0, q0, r0 P r1,8s satisfy
k P t1, . . . , du.
and rk " r,
p0 ď pk,
Remark 3.1. We notice that (3.1) -- (3.6) implies that C is invertible and
that one of the following conditions hold true:
q0 ă qk,
pk ď p,
qk ă q
(1) both C11 and C21 are invertible;
(2) both C12 and C22 are invertible.
If (1) holds, then
11 C12
C21 C22¸ -- detI2d C ´1
detC11 C12
" detI2d
21 C22¸
11 C12¸ " detpC ´1
11 C12
21 C22 ´ C ´1
I2d C ´1
0 C ´1
C ´1
21 C22 ´ C ´1
11 C12q.
Here and in what follows, I " Id is the d d identity matrix. From
these computations it follows that
C ´1
11 C12 ´ C ´1
11 ´ C ´1
C12C ´1
21 C22,
21 C22,
C ´1
11 C12 ´ C22C ´1
11 ´ C22C ´1
C12C ´1
21
21
are invertible when (1) holds, and
22 C21,
12 C11 ´ C ´1
C ´1
12 ´ C ´1
C11C ´1
are invertible when (2) holds.
Remark 3.2. Let C0 P Mpd, Cq and let Ud be the same as in (3.2).
Then the matrix Ud C0 U ´1
in Mp2d, Rq which corresponds to C0
is given by
12 C11 ´ C21C ´1
C ´1
12 ´ C21C ´1
C11C ´1
22 C21,
22
22
d
Obviously, the map which takes C0 P Mpd, Cq into the matrix (3.7)
in Mp2d, Rq is injective, but not bijective.
In this way we identify
Mpd, Cq with the set of all matrices in Mp2d, Rq which are given by
(3.7) for some C0 P Mpd, Cq.
d Cjk Ud, then GK,C,ω in (3.2) is
given by
If Cjk P Mp2d, Rq and Tjk " U ´1
GK,C,ωpz, wq " KωpT11pzq ` T12pwq, T21pzq ` T22pwqq,
28
z, w P Cd.
RepC0q¸ .
RepC0q ´ ImpC0q
ImpC0q
(3.7)
If more restricted, Cjk can be identified as matrices in Mpd, Cq as
above, for j, k P t1, 2u, then GK,C,ω in (3.2) is given by
GK,C,ωpz, wq " KωpC11z ` C12w, C21z ` C22wq,
for such choices of C.
z, w P Cd,
Theorem 3.3. Let E be an ordered basis for R2d, ω1 and ω2 be weights
on Cd, ω be a weight on Cd Cd such that (3.3) holds, and let p1, p2,
p and q be as in (3.4). Also let C P Mp4d, Rq be such that (3.1) holds,
true:
K P uApCd Cdq, and let GK,C,ω be as in (3.2). Then the following is
(1) if (3.5) holds and GK,C,ωpz, wq P Lp,qpC2dq, then TK in (2.1)
from A51pCdq to ApCdq is uniquely extendable to a continuous
mapping from Ap1
E,pω1qpCdq to Ap2
}TKF}Ap2
E,pω2q À }GK,C,ω}Lp,q}F}Ap1
(2) if (3.6) holds and GK,C,ω P Lq,p
E,pω1q
E,pω2qpCdq, and
, F P Ap1
E,pω1qpCdq;
(3.8)
GK,C,ωpz, wqΦpz, wq dλpzqdλpwq
À }GK,C,ω}Lp,q}Φ}Lp1,q1 ,
29
pC2dq, then TK in (2.1) from
A51pCdq to ApCdq is uniquely extendable to a continuous map-
ping from Ap1
E,pω1qpCdq to Ap2
E,pω2qpCdq, and
}TKF}Apω2,Lp2 q À }GK,C,ω}Lq,p
2 }F}Ap1
E,pω1q
, F P Ap1
E,pω1qpCdq.
(3.9)
Proof. We only prove (1). The assertion (2) follows by similar argu-
ments and is left for the reader. Let
GK,C,ω,ppwq " }GK,C,ωp , wq}LppCdq, w P Cd.
Then }GK,C,ω}Lp,qpC2dq -- }GK,p,ω}LqpCdq. Also let Kω be as in (3.2),
F P Ap1
E,pω1qpCdq and H P B p2
E,p1{ω2qpCdq, and set
Fω1pwq " Fpwqe´w2{2ω1p
?2 wq, w P Cd
and
Hω2pzq " Hpzqe´z2{2{ω2p
?2 zq,
z P Cd.
By Holder's inequality we get
pT F, HqB2 -- żCdpT FqpzqHpzqe´z2
dλpzq
Kωpz, wqFω1pwqHω2pzq dλpzqdλpwq
ďij C2d
"ij C2d
where
Φpx ` iξ, y ` iηq
" Fω1pU ´1
d pC21px, ξq ` C22py, ηqqqHω2pU ´1
d pC11px, ξq ` C12py, ηqqq,
x, y, ξ, η P Rd.
Here we identify px, ξq P R2d by corresponding 2d 1-matrix` x
ξ, as
usual.
We need to estimate }Φ}Lp1,q1 , and start with reformulating }Φp , wq}Lp1 .
For }Φp , wq}Lp1 we take
px, ξq ÞÑ C21ppx, ξq ` C ´1
21 C22py, ηqq
as new variables of integration, and get
}Φp , wq}Lp1 -- }Fω1 Hω2pU ´1
where B1 and B2 are the matrices
B1 " C11C ´1
11 C12 ´ C22 P Mp2d, Rq,
which are invertible due to Remark 3.1 and the assumptions. Hence,
for F 0
21 P Mp2d, Rq and B2 " C21C ´1
d pB1p ´ B2py, ηqqq}Lp1 , w " y ` iη,
ω1 " Fω1pU ´1
}Φp , wq}Lp1 -- ´´F 0
ω1p1
If r1 " p1{p1 and r2 " p
ω2 " Hω2 U ´1
q and H 0
ω2p1¯pB2py, ηqq¯ 1
H 0
2{p1, then it follows from (3.4) that
d p´B1q we have
, w " y` iη. (3.10)
p1
d
1
1
r1 `
1
r2 " 1 `
p1
q1
,
and r1, r2,
q1
p1 ě 1.
Hence, by (3.10), that q1{p1 ě 1, the fact that B2 is invertible, and
Holder's and Young's inequalities we obtain
}Φ}Lp1,q1 À››››´´F 0
ω1p1
H 0
-- ›››F 0
ď´}F 0
ω1p1
p1››››Lq1
ω2p1¯ pB2 q¯ 1
E 1
ω2p1›››Lq1{p1
ω1p1
H 0
E¯ 1
E }H 0
}Lr1
ω2p1
}Lr2
p1
p1
-- }Fω1}Lp1
E }Hω2}L
p1
2
E
and the right-hand side of (3.8) follows by taking the supremum over
all such H with }H}B
E,p1{ω2q ď 1.
p1
2
The existence of extension now follows from Hahn-Banach's theorem.
By these estimates it also follows that
pz, wq ÞÑ Kωpz, wqFω1pwqHω2pzq
30
belongs to L1pCd Cdq, and the uniqueness is a straight-forward ap-
plication of Lebesgue's theorem.
(cid:3)
For corresponding pseudo-differential operator with symbol a, the
kernel is given by Kpz, wq " epz,wqapz, wq. By straight-forward compu-
tations it follows that GK,C,ω,p takes the form
GK,C,ω,ppwq " }aωp , wq}LppCdq, w P Cd,
when C11 " C12 " C22 " I2d and C21 " 0, or C11 " C12 " C21 " I2d
and C22 " 0, where
aωpz, wq " e´ 1
2 z2
apz ` w, wqωp
or
?2pz ` wq,?2 wq
aωpz, wq " e´ 1
apz ` w, zqωp
Hence, Theorem 3.3 gives the following.
2 w2
?2 pz ` wq,?2 zq.
(3.11)
(3.12)
Theorem 3.4. Let ω1 and ω2 be weights on Cd, ω be a weight on
Cd Cd such that (3.3) holds, p1, p2, p and q be as in (3.4). Also let
If aω P Lp,qpC2dq, then the operator OpVpaq in (2.4) from A51pCdq to
ApCdq is uniquely extendable to a continuous mapping from Ap1
E,pω1qpCdq
to Ap2
a P uApCd Cdq and let aω be given by (3.11) or by (3.12) for z, w P Cd.
E,pω2qpCdq.
We also have the following result related to Theorem 3.3. Here the
matrix C is given by (3.1) with
C11 " C21 " I2d, C12 "0
0 Id¸ and C22 "Id 0
0¸ (3.13)
0
0
which obviously satisfies (3.5). Also again recall Remark 1.16 for no-
tations.
Theorem 3.5. Let C be given by (3.1) with Cjk given by (3.13), ω1
and ω2 be weights on Cd, ω be a weight on Cd Cd such that (3.3)
holds, and let p, q P r1,8s. Also let K P uApCd Cdq and GK,C,ω be as
in (3.2). If GK,C,ω P Lp,q
pC2dq, then then TK in (2.1) from A51pCdq to
ApCdq is uniquely extendable to a continuous mapping from Ap1,q1
pω1qpCdq
to Aq,p
,pω2qpCdq, and
}TKF}Aq,p
,pω2q À }GK,C,ω}Lp,q
}F}Ap1,q1
pω1q
, F P Ap1,q1
pω1qpCdq.
(3.14)
31
Proof. Let Fω1, F 0
orem 3.3, and let Kω be as in (3.2). Then
ω1, Hω2 and H 0
ω2 be the same as in the proof of The-
pTKF, HqB2 ďżżżż Kωpx, ξ, y, ηqF 0
"żżżż GK,C,ωpx, ξ, y, ηqqF 0
ω1px ` y, ξqH 0
ω1py, ηqH 0
ω2px, ξq dxdξdydη
ω2px, ξ ` ηq dxdξdydη
ď }GK,C,ω}Lp,q
}Φ0}Lp1 ,
where
Φ0px, ξq "ij F 0
ω1px ` y, ξqH 0
ω2px, ξ ` ηqq1
ω1p , ξq}Lq1}H 0
" }F 0
q1
dydη 1
ω2px, q}Lq1 ,
x, ξ P Rd.
Hence,
}Φ0}Lp1 " }Fω1}Lq1,p1
1
}Hω2}Lp1,q1
2
.
The continuity assertion now follows from these estimates, an applica-
tion of Hahn-Banach's theorem and Lebesgue's theorem (cf. the end
of the proof of Theorem 3.3).
(cid:3)
Remark 3.6. In Theorem 3.5, the matrix C is chosen only as (3.1)
and (3.13), while Theorem 3.3 is valid for a whole family of matrices
with the only restriction (3.5) or (3.6). On the other hand, by similar
arguments, it follows that the conclusions in Theorem 3.5 are still true
when, more generally, C P Mp4d, Rq is of the form
Id
0 Id
Id
0
0 Id
Id
0
0
0 Id
0 Id
0
0
0
0
0
0 Id
0
0 Id
0
0 Id
0 Id
0
0 Id
,
‹‹‹‹‚
‹‹‹‹‚
or
Id
0
0 Id
Id
0 Id
0
0 Id
Id
0
0
0
0 Id
0
0 Id
0
0
0 Id
0
0 Id
0
0
Id
0
0 Id.
,
‹‹‹‹‚
‹‹‹‹‚
,
for any choice of at each place, provided the mixed Lebesgue condi-
tions on GK,C,ω are slightly modified.
In order to apply Theorem 3.3 to real pseudo-differential operators
we have the following.
32
2 px2`ξ2q, x, ξ P Rd, a P H1
51pR2dq
Lemma 3.7. Let φpx, ξq " π´ d
2 eixx,ξye´ 1
and let Ka be the kernel of Oppaq. Then
e´ 1
2 pz2`w2qVΘ,dKapz, wq
" 2
2 e´ipxx,ξ´2ηy`xy,ηyqpVφaqp
?2x,´
when z " x ` iξ P Cd and w " y ` iη P Cd.
Proof. Let φ0px, ξq " π´ d
putations and Fourier's inversion formula we get
2 e´ 1
d
?2η,?2pη ´ ξq,?2py ´ xqq
2 px2`ξ2q " e´ixx,ξyφpx, ξq. By formal com-
?2, w{
?2q
p4π3q
d
2 e
i
2 pxx,ξy`xy,ηyqe´ 1
4 pz2`w2qVΘ,dKapz{
d
d
d
" p2πq
" p2πq
"¡ apx1, ξ1qeixx1´y1,ξ1yφ0px1 ´ x, y1 ´ yqe´ipxy1,ηy´xx1,ξyq dx1dy1dξ1
2 e´ixy,ηyij apx1, ξ1qφ0px1´x, ξ1`ηqeixx1,ξ1ye´ipxy,ξ1y´xx1,ξyq dx1dξ1
2 eixx´y,ηyij apx1, ξ1qφpx1 ´ x, ξ1 ` ηqe´ipxx1,η´ξy`xy´x,ξ1yq dx1dξ1
2 eixx´y,ηypVφaqpx,´η, η ´ ξ, y ´ xq. (cid:3)
We can now use the previous lemma and theorems to obtain mapping
properties for pseudo-differential operators with symbols in modulation
spaces. For example, we may combine Lemma 3.7 and Theorem 3.3
to deduce the following result, which is the same as [28, Theorem 2.2].
Hence our kernel results on the Bargmann transform side can be used to
regain classical mapping properties pseudo-differential operators when
acting on modulation spaces.
Theorem 3.8. Let E be an ordered basis of R2d, A P MpR, dq, p, q P
r1,8s and p1, p2 P r1,8s2d be as in (3.4), ω0 P PEpR4dq and ω1, ω2 P
PEpR2dq be such that
" p4πq
ω2px ´ Ay, ξ ` pI ´ Aqηq
ω1px ` pI ´ Aqy, ξ ´ Aηq À ω0px, ξ, η, yq,
x, y, ξ, η P Rd,
(3.15)
and let a P M p,q
uniquely to continuous operator from M p1
pω0qpR2dq. Then OpApaq from Σ1pRdq to Σ1
E,pω1qpRdq to M p2
1pRdq extends
E,pω2qpRdq, and
} OpApaq}M p1
E,pω1qÑM p2
E,pω2q À }a}M p,q
pω0q
.
Proof. By (1.40) and Proposition 1.19 we may assume that A " 0. Let
and let ω be given by
C11 " C21 " C22 " I2d, C12 " 0
ωpx, ξ, y, ηq " ω0px,´η, ξ ` η, y ´ xq,
33
x, y, ξ, η P Rd,
which we identify with
ωpz, wq,
z " x ` iξ P Cd, w " y ` iη P Cd.
Then it follows by straight-forward computations that (3.15) is the
same as (3.3). Furthermore, let K0 " VΘ,dKa, where Ka is the kernel
of the operator Oppaq. Then it follows from Lemma 3.7 and straight-
forward computations that if
Ha,ω0px, ξ, η, yq " Vφapx, ξ, η, yq ω0px, ξ, η, yq, ,
x, y, ξ, η P Rd,
then
?2 x,´
?2pξ ` ηq,?2 η,?2 yq -- GK0,C,ωpz, wq,
Ha,ω0p
(3.16)
z " x ` iξ P Cd, w " y ` iη P Cd.
By first applying the Lp-norm on (3.16) with respect to x and ξ, and
thereafter applying the Lq-norm with respect to y and η, we get
}GK0,C,ω}Lp,qpC2dq -- }Ha,ω0}Lp,qpR4dq -- }a}M p,q
pω0q ă 8.
Hence, the assumptions in Theorem 3.3 are fullfiled, and we conclude
E,pω1qpCdq to
that the operator TK0 with kernel K0 is continuous from Ap1
E,pω2qpCdq. The asserted continuity for Oppaq is now a consequence
Ap2
of the commutative diagram
E,pω1qpRdq
ÝÝÝÑ M p2
E,pω2qpRdq
M p1
Oppaq
(3.17)
Vd§§đ
E,pω1qpCdq ÝÝÝÑTK0
Ap1
E,pω2qpCdq. (cid:3)
Ap2
§§đVd
The next result extends [24, Theorem 3.3] and follows by similar
arguments as in the previous proof, using Theorem 3.5 instead of The-
orem 3.3. The details are left for the reader.
Theorem 3.9. Let ω1 and ω2 P PEpR2dq, ω P PEpR4dq be such that
ω2px, ξ ` ηq
ω1px ` y, ξq À ω0px, ξ, η, yq, ,
x, y, ξ, η P Rd,
let p, q P r1,8s, and let a P W p,q
1pRdq is uniquely extendable to a continuous mapping from M q1,p1
Σ1
to W p,q
pω0qpR2dq. Then Op0paq from Σ1pRdq to
pω1qpRdqq
pω2qpRdqq, and
} Oppaqf}W p,q
pω2q À }a}W p,q
pω0q}f}M q1,p1
pω1q
,
f P M q1,p1
pω1qpRdqq.
Remark 3.10. Let E, p, q, pj, φ0 and φ be the same as in Lemma
3.7, Theorem 3.8 and their proofs. Also let A P MpR, dq and let
φA " eixADξ ,Dxyφ. Then the condition on a in Theorem 3.8 is that
}Vφ0a ω0}Lp,q ă 8. In view of [5, 6, 29] and Proposition 1.13 (2), the
34
previous condition is the same as }VφAa ω0}Lp,q ă 8 because ω0 is
moderate.
We observe that all weights in Theorem 3.8 are moderate, while there
are no such assumptions or other restrictions on the involved weight
functions in Theorem 3.3. Since the latter result is used to prove the
former one, a natural question is wether Theorem 3.8 can be extended
to broader classes of weight functions. In view of Remark 1.14, it is
evident that the imposing moderate conditions on weights might in
some context be considered as strong restrictions.
The answer on this question is affirmative in the sense that for suit-
able modifications, the moderate conditions on the weights in Theorem
3.8 can be removed.
In fact, let HA
51pR2dq be the modification of H51pR2dq, given by
HA
51pR2dq " t eixADξ ,Dxypeixx,ξyaq ; a P H51pR2dqu,
51q1pR2dq be the dual of HA
51pR2dq, ω1, ω2 be weights on R2d and let
pHA
ω0 be a weight on R4d such that (3.15) holds. Then it follows from the
proof of Theorem 3.8 that the following is true:
‚ if a P pHA
‚ if a P pHA
H51pRdq to H1
from M p1
51q1pR2dq, then VφAa makes sense as a smooth function;
51q1pR2dq satisfies }VφAaω0}Lp,q ă 8, then OpApaq from
51pRdq extends uniquely to a continuous operator
E,pω1qpRdq to M p2
E,pω2qpRdq.
In similar ways, Theorem 3.9 can be extended to permit more general
weight classes.
References
[1] V. Bargmann On a Hilbert space of analytic functions and an associated inte-
gral transform, Comm. Pure Appl. Math., 14 (1961), 187 -- 214.
[2] V. Bargmann On a Hilbert space of analytic functions and an associated in-
tegral transform. Part II. A family of related function spaces. Application to
distribution theory., Comm. Pure Appl. Math., 20 (1967), 1 -- 101.
[3] W. Bauer Berezin-Toeplitz quantization and composition formulas, J. Funct.
Anal., 256 (2007), 3107 -- 3142.
[4] F. A. Berezin Wick and anti-Wick symbols of operators, Mat. Sb. (N.S.), 86
(1971), 578 -- 610.
[5] M. Cappiello, J. Toft, Pseudo-differential operators in a Gelfand -- Shilov setting,
Math. Nachr. 290 (2017), 738 -- 755.
[6] E. Carypis, P. Wahlberg, Propagation of exponential phase space singularities
for Schrodinger equations with quadratic Hamiltonians, J. Fourier Anal. Appl.
23 (2017), 530 -- 571.
[7] Y. Chen, M. Signahl, J. Toft Factorizations and singular value estimates of
operators with Gelfand-Shilov and Pilipovi´c kernels, J. Fourier Anal. Appl. 24
(2018), 666 -- 698.
[8] 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.
35
[9] H. G. Feichtinger Banach spaces of distributions of Wiener's type and inter-
polation, in: Ed. P. Butzer, B. Sz. Nagy and E. Gorlich (Eds), Proc. Conf.
Oberwolfach, Functional Analysis and Approximation, August 1980, Int. Ser.
Num. Math. 69 Birkhauser Verlag, Basel, Boston, Stuttgart, 1981, pp. 153 --
165.
[10] H. G. Feichtinger Modulation spaces on locally compact abelian groups. Techni-
cal 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.
[11] C. Fernandez, A. Galbis, J. Toft The Bargmann transform and powers of har-
monic oscillator on Gelfand-Shilov subspaces, RACSAM 111 (2017), 1 -- 13.
[12] K. Grochenig Foundations of Time-Frequency Analysis, Birkhauser, Boston,
2001.
[13] 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.
[14] K. Grochenig, G. Zimmermann Spaces of test functions via the STFT J. Funct.
Spaces Appl. 2 (2004), 25 -- 53.
[15] L. Hormander The Analysis of Linear Partial Differential Operators, vol I -- III,
Springer-Verlag, Berlin Heidelberg NewYork Tokyo, 1983, 1985.
[16] Z. Lozanov Crvenkovi´c, D. Perisi´c Hermite expansions of elements of Gelfand
Shilov spaces in quasianalytic and non quasianalytic case, Novi Sad J. Math.
37 (2007), 129 -- 147.
[17] S. Pilipovi´c Generalization of Zemanian spaces of generalized functions which
have orthonormal series expansions, SIAM J. Math. Anal. 17 (1986), 477 -- 484.
[18] S. Pilipovi´c Tempered ultradistributions, Boll. U.M.I. 7 (1988), 235 -- 251.
[19] M. Reed and B. Simon Methods of modern mathematical physics, Academic
Press, London New York, 1979.
[20] M. Ruzhansky, N. Tokmagambetov Nonharmonic analysis of boundary value
problems, Int. Math. Res. Notices 12 (2016), 3548 -- 3615.
[21] N. Teofanov Ultradistributions and time-frequency analysis in: P. Boggiatto,
L. Rodino, J. Toft, M. W. Wong (eds) Pseudo-differential operators and related
topics, Operator Theory: Advances and Applications 164, Birkhauser, Basel,
2006, pp. 173 -- 192.
[22] N. Teofanov, Gelfand-Shilov spaces and localization operators, Funct. Anal.
Approx. Comput. 7 (2015), 135 -- 158.
[23] J. Toft Continuity properties for modulation spaces with applications to pseudo-
differential calculus, II, Ann. Global Anal. Geom. 26 (2004), 73 -- 106.
[24] J. Toft Pseudo-differential operators with symbols in modulation spaces, in:
B.-W. Schulze, M. W. Wong (Eds), Pseudo-Differential Operators: Complex
Analysis and Partial Differential Equations, Operator Theory Advances and
Applications 205, Birkhauser Verlag, Basel, 2010, pp. 223 -- 234.
[25] J. Toft The Bargmann transform on modulation and Gelfand-Shilov spaces,
with applications to Toeplitz and pseudo-differential operators, J. Pseudo-
Differ. Oper. Appl. 3 (2012), 145 -- 227.
[26] J. Toft Images of function and distribution spaces under the Bargmann trans-
form, J. Pseudo-Differ. Oper. Appl. 8 (2017), 83 -- 139.
[27] J. Toft Continuity and compactness for pseudo-differential operators with sym-
bols in quasi-Banach spaces or Hormander classes, Anal. Appl. 15 (2017),
353 -- 389.
36
[28] J. Toft Matrix parameterized pseudo-differential calculi on modulation spaces
in: M. Oberguggenberger, J. Toft, J. Vindas, P. Wahlberg (eds) Generalized
Functions and Fourier Analysis, Operator Theory: Advances and Applications
260, Birkhauser, Basel Heidelberg NewYork Dordrecht London, pp. 215 -- 235.
[29] G. Tranquilli Global normal forms and global properties in function spaces for
second order Shubin type operators PhD Thesis, 2013.
Department of Mathematics and Informatics, University of Novi
Sad, Novi Sad, Serbia
E-mail address: [email protected]
Department of Computer science, Mathematics and Physics, Linnaeus
University, Vaxjo, Sweden
E-mail address: [email protected]
37
|
1812.00204 | 1 | 1812 | 2018-12-01T13:24:18 | On compressions of self-adjoint extensions of a symmetric linear relation | [
"math.FA"
] | Let $A$ be a symmetric linear relation in the Hilbert space $\gH$ with equal deficiency indices $n_\pm (A)\leq\infty$. A self-adjoint linear relation $\wt A\supset A$ in some Hilbert space $\wt\gH\supset \gH$ is called an exit space extension of $A$; such an extension is called finite-codimensional if $\dim (\wt\gH\ominus\gH)< \infty$. We study the compressions $C (\wt A)=P_\gH\wt A\up\gH$ of exit space extensions $\wt A=\wt A^*$. For a certain class of extensions $\wt A$ we parameterize the compressions $C (\wt A)$ by means of abstract boundary conditions. This enables us to characterize various properties of $C (\wt A)$ (in particular, self-adjointness) in terms of the parameter for $\wt A$ in the Krein formula for resolvents. We describe also the compressions of a certain class of finite-codimensional extensions. These results develop the results by A. Dijksma and H. Langer obtained for a densely defined symmetric operator $A$ with finite deficiency indices. | math.FA | math |
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS OF A
SYMMETRIC LINEAR RELATION
V.I. MOGILEVSKII
Abstract. Let A be a symmetric linear relation in the Hilbert space H with equal
deficiency indices n±(A) ≤ ∞. A self-adjoint linear relation eA ⊃ A in some Hilbert
space eH ⊃ H is called an exit space extension of A; such an extension is called finite-
codimensional if dim(eH ⊖ H) < ∞. We study the compressions C(eA) = PHeA ↾ H of
exit space extensions eA = eA∗. For a certain class of extensions eA we parameterize
the compressions C(eA) by means of abstract boundary conditions. This enables us to
characterize various properties of C(eA) (in particular, self-adjointness) in terms of the
parameter for eA in the Krein formula for resolvents. We describe also the compressions
of a certain class of finite-codimensional extensions. These results develop the results
by A. Dijksma and H. Langer obtained for a densely defined symmetric operator A with
finite deficiency indices.
1. Introduction
Assume now that A is a not necessarily densely defined symmetric operator in a Hilbert
Self(A) = Self 0(A) if and only if A is densely defined.
problem in the extension theory of symmetric operators. In the paper by A.V. Shtraus
were recently studied in [2, 3, 4, 11, 12, 20]. In particular, it was shown in [3] that a
Let H be a subspace in a Hilbert space eH and let eA be a linear relation (in particular
operator) in eH. Recall that a linear relation C(eA) in H given by C(eA) = PHeA ↾ H is
called a compression of eA; moreover, the relation eA and its compression C(eA) are called
finite-codimensional if dim(eH ⊖ H) < ∞. Compressions of linear operators or relations
finite-codimensional compression C(eA) of a self-adjoint linear relation eA is self-adjoint
(for operators eA = eA∗ this fact was established earlier in [24]).
space H. A self-adjoint linear relation eA ⊃ A in a Hilbert space eH ⊃ H is called an exit
space extension of A. Denote by Self(A) the set of all exit space extensions eA = eA∗ of
A and by Self 0(A) the set of all eA ∈ Self(A) such that eA is an operator. As is known
If eA ∈ Self(A), then the compression C(eA) of eA is a symmetric extension of A. A
description of all extensions eA ∈ Self(A) and their compressions C(eA) is an important
[23] all extensions eA ∈ Self 0(A) of an operator A with arbitrary (equal or unequal)
eA ∈ Self(A).
In [22, 25] the compressions C(eA) of the extensions eA ∈ Self 0(A) we
a bijective correspondence eA = eAτ between all relation-valued Nevanlinna functions
deficiency indices n±(A) ≤ ∞ are parameterized by means of holomorphic operator-
functions F (λ), λ ∈ C+, whose values are contractions from Ni to N−i (here N±i =
ker (A∗∓i) are defect subspaces of A). In [13] the result of [23] was extended to extensions
In the case n+(A) = n−(A) another parametrization of the set Self(A) is given by the
Krein formula for generalized resolvents [15, 17]. This formula (see (2.36)) establishes
τ = τ (λ) (λ ∈ C \ R) (τ (λ) is a linear relation in an auxiliary Hilbert space H) and all
described in terms of the asymptotic behaviour of the corresponding parameter F (λ).
2010 Mathematics Subject Classification. 47A06, 47A20, 47A56, 47B25.
Key words and phrases. Symmetric and self-adjoint linear relation (operator), exit space self-adjoint
extension, compression, boundary triplet.
1
(1.1)
τ (λ) = A + λB +
Aj,
λ ∈ C \ R,
1
αj − λ
lXj=1
2
V.I. MOGILEVSKII
extensions eA ∈ Self(A). The canonical self-adjoint extension A0 of A in (2.36) is called
a basic extension. An operator-valued parameter τ = τ (λ) in (2.36) is called rational if
where αj ∈ R and A = A∗, B ≥ 0 and Aj ≥ 0 are bounded operators in H. In the case
parameter τ is rational.
n+(A) = n−(A) < ∞ an extension eAτ ∈ Self(A) is finite-codimensional if and only if a
In the paper by H. Langer and A. Dijksma [12] the compressions C(eAτ ) of extensions
eAτ are investigated in terms of the parameter τ from the Krein formula (2.36). The main
results of [12] can be formulated in the form of the following theorem.
Theorem 1.1. Assume that A is a densely defined, closed symmetric operator in H with
finite deficiency indices n+(A) = n−(A) < ∞ and A0 = A∗
0 is the basic extension in the
Krein formula (2.36). Let for simplicity a parameter τ = τ (λ), λ ∈ C \ R, in (2.36) be
an operator-valued function, let eAτ ∈ Self 0(A) be the corresponding extension of A and
let C(eAτ ) be the compression of eAτ . Assume also that Bτ ≥ 0 is an operator in H given
by Bτ = lim
y→∞
1
iy τ (iy). Then:
(i) If
(1.2)
lim
y→∞
yIm (τ (iy)h, h) = ∞,
h ∈ H, h 6= 0,
then C(eAτ ) ⊂ A0.
(ii) If (1.2) holds, then C(eAτ ) = A if and only if Bτ = 0.
(iii) If Bτ > 0, then C(eAτ ) = A0. If (1.2) holds and C(eAτ ) = A0, then Bτ > 0.
Moreover, if τ is a rational parameter (1.1), then C(eAτ ) = A0 ⇐⇒ Bτ > 0.
(iv) If τ is the rational parameter (1.1), then the extension eAτ is finite-codimensional
and the canonical self-adjoint extension C(eAτ ) of A corresponds in the Krein formula to
the self-adjoint linear relation τ∞ in H given by
τ∞ = {{h, Pker BAh + h′} : h ∈ ker B, h′ ∈ ran B}.
(1.3)
It is also shown in [12] that τ∞ admits the representation
(1.4)
τ∞ = {{h, h′} ∈ H2 : ∃h(λ) ∈ H : h = lim
λ→∞
h(λ), h′ = lim
λ→∞
τ (λ)h(λ)}.
In the present paper we study compressions of extensions eA ∈ Self(A) of a symmetric
linear relation (in particular, not necessarily densely defined symmetric operator) A
with possibly infinite deficiency indices n+(A) = n−(A) ≤ ∞. Our approach is based on
the theory of boundary triplets and their Weyl functions (see [9, 14, 18] and references
therein). Recall [14, 18] that a collection Π = {H, Γ0, Γ1} consisting of an auxiliary
Hilbert space H and linear mappings Γj : A∗ → H, j ∈ {0, 1}, is called a boundary
triplet for A∗ if the mapping (Γ0, Γ1)⊤ is surjective and the abstract Green identity
(2.15) is valid (here A∗ is the adjoint relation). If Π = {H, Γ0, Γ1} is a boundary triplet
for A∗, then the equality (the abstract boundary conditions)
(1.5)
θ → Aθ := { f ∈ A∗ : {Γ0
f , Γ1
f } ∈ θ}
gives a parametrization of all proper extensions eA of A (i.e., all linear relations eA in
H with A ⊂ eA ⊂ A∗) in terms of linear relations θ in H. Moreover, it was shown in
[9, 18] that each boundary triplet Π for A∗ gives rise to the Krein formula (2.36) with
coefficients A0, γ(λ) and M(λ) naturally defined in terms of Π.
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
3
Assume that A is a closed symmetric linear relation in H with equal deficiency indices
n+(A) = n−(A) ≤ ∞ and let Π = {H, Γ0, Γ1} be a boundary triplet for A∗. In the paper
linear relation θc, which in a certain sense is a limit value of τ (iy) when y → ∞. This
we consider extensions eAτ ∈ Self(A) corresponding to a certain subclass of Nevanlinna
parameters τ (λ) in (2.36). We show that the compression C(eAτ ) of eAτ being a symmetric
extension of A admits the representation C(eAτ ) = Aθc (see (1.5)) with a symmetric
result enables us to characterise various properties of C(eAτ ) (self-adjointness, inclusion
C(eAτ ) ⊂ A0, equalities C(eAτ ) = A and C(eAτ ) = A0) in terms of the parameter τ . In
the case n±(A) < ∞ the obtained results become valid for any Nevanlinna parameter
τ . This fact enables us to prove the following theorem which strengthens essentially
statements (i) -- (iii) of Theorem 1.1.
Theorem 1.2. Assume that A is a closed symmetric linear relation in H with finite
deficiency indices n+(A) = n−(A) < ∞ and let all other assumptions of Theorem 1.1 be
satisfied. Then:
ficiency indices n±(A).
2. Preliminaries
2.1. Notations. The following notations will be used throughout the paper: H, H de-
note separable Hilbert spaces; B(H1, H2) is the set of all bounded linear operators defined
on H1 with values in H2; A ↾ L is a restriction of the operator A ∈ B(H1, H2) onto
the linear manifold L ⊂ H1; PL is the orthoprojection in H onto the subspace L ⊂ H;
C+ (C−) is the open upper (lower) half-plane of the complex plane.
Recall that a linear manifold T in the Hilbert space H0 ⊕ H1 (H ⊕ H) is called a linear
in H). The set of all closed linear relations from H0 to
relation from H0 to H1 (resp.
H1 (in H) will be denoted by eC(H0, H1) (resp. eC(H)). Clearly for each linear operator
T : dom T → H1, dom T ⊂ H0, its grT = {{f, T f } : f ∈ dom T } is a linear relation
from H0 to H1. This fact enables one to consider an operator as a linear relation.
For a linear relation T from H0 to H1 we denote by
dom T := {h0 ∈ H0 : ∃h1 ∈ H1 {h0, h1} ∈ T },
ker T := {h0 ∈ H0 : {h0, 0} ∈ T }
ran T := {h1 ∈ H1 : ∃h0 ∈ H0 {h0, h1} ∈ T }, mul T := {h1 ∈ H1 : {0, h1} ∈ T }
yIm (τ (iy)h, h) < ∞ for all h ∈ ker Bτ .
y→∞
Theorem 1.3. Let the assumptions be the same as in Theorem 1.2. Then:
(i′) C(eAτ ) ⊂ A0 if and only if (1.2) holds.
(ii′) C(eAτ ) = A if and only Bτ = 0 and (1.2) holds.
(iii′) C(eAτ ) = A0 if and only if Bτ > 0.
We prove also the following theorem which characterizes extensions eA ∈ Self(A) with
self-adjoint compression C(eA).
(i) C(eAτ ) is self-adjoint if and only if lim
(i) C(eAτ ) is self-adjoint and transversal with A0 (that is, C(eAτ ) ∩ A0 = A and
C(eAτ )b+A0 = A∗) if and only if
the corresponding parameter for C(eAτ ) in the Krein formula is the operator N ′
τ (iy)h, h ∈ H. Moreover, in the sense of (1.5) C(eAτ ) = ANτ with
extensions eA ∈ Self(A) of a symmetric linear relation A with possibly infinite equal de-
finite-codimensional extension eA ∈ Self(A).
In the case n+(A) = n−(A) < ∞ this result holds for any
given by N ′
τ h = lim
y→∞
Nτ = −N ′
τ .
lim
y→∞
yIm (τ (iy)h, h) < ∞ for all h ∈ H. In this case
Finally, we extend statement (iv) of Theorem 1.1 to a certain class of finite-codimensional
τ in H
4
V.I. MOGILEVSKII
the domain, kernel, range and multivalued part of T respectively; moreover, we let
Denote also by T −1 and T ∗ the inverse and adjoint linear relations of T respectively.
For an operator T = T ∗ ∈ B(H) we write T ≥ 0 if (T f, f ) ≥ 0, f ∈ H, and T > 0 if
dmul T = {0} ⊕ mul T = {{0, h1} : h1 ∈ mul T }.
T − αI ≥ 0 with some α > 0.
For linear relations T1 and T2 in H we let T1b+T2 = {bf +bg : bf ∈ T1,bg ∈ T2}.
2.2. Nevanlinna functions. Recall that a holomorphic operator function M : C \ R →
B(H) is called a Nevanlinna function if Im λ·Im M(λ) ≥ 0 and M ∗(λ) = M(λ), λ ∈ C\R.
The class of all Nevanlinna B(H)-valued functions will be denoted by R[H].
Definition 2.1. The operator-function M ∈ R[H] is referred to the class:
(i) Rc[H], if ran Im M(λ) is closed for all λ ∈ C \ R;
(ii) Ru[H], if (Im M(λ))−1 ∈ B(H) for all λ ∈ C \ R or equivalently if Im λ · Im M(λ) >
0, λ ∈ C \ R.
Clearly Ru[H] ⊂ Rc[H] ⊂ R[H]; moreover in the case dim H < ∞ one has Rc[H] =
R[H].
The following proposition is well known (see e.g. [18]).
Proposition 2.2. If M ∈ R[H], then the equality
(2.1)
BM = s- lim
y→∞
1
iy M(iy)
defines the operator BM ∈ B(H) such that BM ≥ 0. Moreover, the equality
(2.2)
dom NM = {h ∈ H : lim
y→∞
yIm (M(iy)h, h) < ∞}
defines the (not necessarily closed) linear manifold dom NM ⊂ H and for each h ∈
dom NM there exists the limit
(2.3)
NM h := lim
y→∞
M(iy)h,
h ∈ dom NM .
Hence the equalities (2.2) and (2.3) define the linear operator NM : dom NM → H.
Proposition 2.3. Let τ ∈ R[H]. Then the subspace H′′ := ker Im τ (λ) ⊂ H does not
depend on λ ∈ C \ R and the block representation
holds with H′ = H ⊖ H′′, B2 = B∗
2 ∈ B(H′′), B1 ∈ B(H′′, H′) and the operator-function
τ1(·) ∈ R[H′] such that ker Im τ1(λ) = {0}, λ ∈ C \ R. Moreover, τ ∈ Rc[H] if and only
if τ1 ∈ Ru[H′].
Proof. As is known the Cayley transform K(λ) = (τ (λ) − i)(τ (λ) + i)−1, λ ∈ C+, is a
holomorphic contractive operator-function on C+ and for each λ ∈ C+ the operator τ (λ)
admits the representation
(2.5)
f = i(K(λ) − I)h,
τ (λ)f = (K(λ) + I)h,
h ∈ H.
It follows from (2.5) that
(2.6)
Im (τ (λ)f, f ) = h2 − K(λ)h2
for all λ ∈ C+, h ∈ H and f = i(K(λ) − I)h. According to [19] the subspace H0 = {h ∈
H : K(λ)h = h} and the operator V = K(λ) ↾ H0(∈ B(H0, H)) do not depend on
λ ∈ C+. Moreover, by (2.5) and (2.6) ker Im τ (λ) = (V − I)H0 and hence the subspace
(2.4)
τ (λ) =(cid:18)τ1(λ) B1
1 B2(cid:19) : H′ ⊕ H′′ → H′ ⊕ H′′,
B∗
λ ∈ C \ R
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
5
H′′ := ker Im τ (λ) does not depend on λ ∈ C+. Note also that by (2.5) the operator
τ (λ) ↾ H′′ is defined by
f = i(V − I)h0,
τ (λ)f = (V + I)h0,
h0 ∈ H0.
Therefore B = τ (λ) ↾ H′′(∈ B(H′′, H)) does not depend on λ ∈ C+ and, consequently,
τ (λ) admits the block representation
(2.7)
This implies that
τ (λ) =(cid:18)τ1(λ) B1
τ2(λ) B2(cid:19) : H′ ⊕ H′′ → H′ ⊕ H′′,
λ ∈ C+
Im τ (λ) =(cid:18) Im τ1(λ)
2i (τ2(λ) − B∗
1)
1
1
2i (B1 − τ ∗
Im B2
2 (λ)
(cid:19) : H′ ⊕ H′′ → H′ ⊕ H′′,
λ ∈ C+
1. Therefore by (2.7) the equality (2.4) holds with
and hence Im B2 = 0, τ1(λ) = B∗
B2 = B∗
2, which implies that
(2.8)
Im τ (λ) =(cid:18)Im τ1(λ) 0
0(cid:19) : H′ ⊕ H′′ → H′ ⊕ H′′,
0
λ ∈ C+.
Hence ker Im τ1(λ) = {0} and the required statements are proved for λ ∈ C+. The same
statements for λ ∈ C− are implied by τ (λ) = τ ∗(λ), λ ∈ C−.
(cid:3)
Nevanlinna relation valued functions if:
As is known (see e.g. [8]) a function τ : C \ R → eC(H) is referred to the class eR(H) of
(i) Im (f ′, f ) ≥ 0,
(ii) (τ (λ) + i)−1 ∈ B(H), λ ∈ C+, and (τ (λ) + i)−1 is a holomorphic operator-function
{f, f ′} ∈ τ (λ),
λ ∈ C+;
in C+;
(iii) τ ∗(λ) = τ (λ), λ ∈ C \ R.
(2.9)
λ ∈ C \ R
H = H0 ⊕ K,
the operator part of τ .
τ (λ) does not depend on λ ∈ C \ R and the decompositions
τ (λ) = τ0(λ) ⊕ bK,
mul τ (λ) = 0, λ ∈ C \ R. If τ (·) ∈ R(H), then dom τ (λ) = H, λ ∈ C \ R.
A function τ ∈ eR(H) is referred to the class R(H) if its values are operators, i.e., if
It is clear that R[H] ⊂ R(H) ⊂ eR(H).
According to [15] for each function τ ∈ eR(H) the multivalued part K := mul τ (λ) of
hold with τ0 ∈ R[H0] and bK = dmul τ (λ) = {0} ⊕ K. The operator function τ0 is called
Definition 2.4. A relation valued function τ ∈ eR(H) is referred to the class eRc(H)
(eRu(H)) if decompositions (2.9) hold with τ0 ∈ Rc[H0] (resp. with τ0 ∈ Ru[H0]).
Clearly eRu(H) ⊂ eRc(H) ⊂ eR(H) and in the case dim H < ∞ one has eRc(H) = eR(H).
Corollary 2.5. The relation valued function τ : C \ R → eC(H) belongs to the class
eRc(H) if and only if there exist a decomposition
1 −B2(cid:19) : H′ ⊕ H′′
{z
}
operators B1 ∈ B(H′′, H′), B2 = B∗
such that τ (λ) admits the representation (2.9) with the operator function
τ0(λ) =(cid:18)τ1(λ) −B1
2 ∈ B(H′′) and an operator-function τ1 ∈ Ru[H′]
The following corollary is immediate from Proposition 2.3.
H = H0 ⊕ K = H′ ⊕ H′′
→ H′ ⊕ H′′
λ ∈ C \ R.
H0
{z
,
}
H0
{z
}
H0
(2.10)
(2.11)
−B∗
⊕K,
6
V.I. MOGILEVSKII
This means that
(2.12)
τ (λ) = {{h ⊕ ϕ, (τ1(λ)h − B1ϕ) ⊕ (−B∗
1 h − B2ϕ) ⊕ k} : h ⊕ ϕ ∈ H′ ⊕ H′′, k ∈ K}.
Remark 2.6. In the following for the relation valued function τ ∈ eRc(H) represented
in accordance with Corollary 2.5 by (2.10) and (2.12) we will use the notation τ =
{H′ ⊕ H′′ ⊕ K, B1, B2, τ1}.
2.3. Boundary triplets and Weyl function. Recall that a linear relation T in H is
called: (i) symmetric if T ⊂ T ∗ or, equivalently, if
(2.13)
(f ′, g) − (f, g′) = 0,
{f, f ′}, {g, g′} ∈ T ;
(ii) self-adjoint if T = T ∗ (and hence T ∈ eC(H)).
Then
Assume that θ is a symmetric relation in H with a closed multivalued part mul θ.
(2.14)
H = H0 ⊕ K,
where K = mul θ, H0 = H ⊖ mul θ and B is a symmetric operator in H0 (the operator
part of θ) with dom B = dom θ ⊂ H0. Clearly, θ is closed (self-adjoint) if and only if B
is closed (self-adjoint).
θ = B ⊕ dmul θ = {{h, Bh + k} : h ∈ dom θ, k ∈ mul θ},
The following lemma will be useful in the sequel.
Lemma 2.7. Let θ be a symmetric linear relation in H. Then:
(i) dom θ ⊂ H ⊖ mul θ.
(ii) If mul θ = mul θ and H ⊖ mul θ ⊂ dom θ, then θ∗ = θ.
(iii) If dim H < ∞, then the condition H ⊖ mul θ ⊂ dom θ is equivalent to θ∗ = θ.
Proof. (i) Statement (i) directly follows from (2.13)
(ii) Assume that mul θ = mul θ and H ⊖ mul θ ⊂ dom θ. Then in the representation
(2.14) of θ one has dom B = dom θ = H0. Therefore B = B∗ ∈ B(H0) and, consequently,
θ∗ = θ.
(iii) Assume that dim H < ∞. Then mul θ is closed and hence decompositions (2.14)
hold. Moreover, dim H0 < ∞ and therefore θ∗ = θ if and only if dom B = H0 or,
equivalently, dom θ = H ⊖ mul θ. This yields statement (ii).
(cid:3)
In the following we denote by A a closed symmetric linear relation (in particular closed
not necessarily densely defined symmetric operator) in a Hilbert space H. Let Nλ(A) =
let N(A) = H ⊖ dom A(= mul A∗) and let n±(A) := dim Nλ(A) ≤ ∞, λ ∈ C±, be
deficiency indices of A. Denote by ext(A) the set of all proper extensions of A (i.e.,
ker (A∗ − λ) (λ ∈ C \ R) be a defect subspace of A, let bNλ(A) = {{f, λf } : f ∈ Nλ(A)},
the set of all relations eA in H such that A ⊂ eA ⊂ A∗) and by ext(A) the set of closed
extensions eA ∈ ext(A). Recall that two extensions eA1, eA2 ∈ ext(A) are called transversal
if eA1 ∩ eA2 = A and eA1b+eA2 = A∗.
Definition 2.8. [14, 5] A collection Π = {H, Γ0, Γ1} consisting of a Hilbert space H
and linear mappings Γj : A∗ → H, j ∈ {0, 1}, is called a boundary triplet for A∗, if
the mapping Γ = (Γ0, Γ1)⊤ from A∗ into H ⊕ H is surjective and the following Green's
identity holds:
(2.15)
(f ′, g) − (f, g′) = (Γ1bf , Γ0bg) − (Γ0bf , Γ1bg),
[14, 18] If Π = {H, Γ0, Γ1} is a boundary triplet for A∗, then n+(A) =
Proposition 2.9.
n−(A) = dim H. Conversely, for each symmetric relation A in H with equal deficiency
indices n+(A) = n−(A) there exists a boundary triplet for A∗.
bf = {f, f ′}, bg = {g, g′} ∈ A∗.
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
7
Proposition 2.10.
[14, 18] Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗. Then:
(i) ker Γ = A and Γ is a bounded operator from A∗ onto H ⊕ H.
(ii) The mapping
(2.16)
θ → Aθ := { f ∈ A∗ : {Γ0 f , Γ1 f } ∈ θ}
establishes a bijective correspondence between all linear relations θ in H and all extensions
(iii) (Aθ)∗ = Aθ∗
eA = Aθ ∈ ext(A). In the case θ ∈ B(H) one has Aθ = { f ∈ A∗ : Γ1 f = θΓ0 f }
(iv) Aθ = Aθ and hence Aθ ∈ ext(A) if and only if θ ∈ eC(H)
(v) Aθ is symmetric (self-adjoint) if and only if θ is symmetric (resp. self-adjoint)
(vi) The equality
(2.17)
A0 := ker Γ0 = { f ∈ A∗ : Γ0
f = 0}
defines a self-adjoint extension A0 of A.
(vii) The extensions A0 and Aθ ∈ ext(A) are transversal if and only if θ ∈ B(H).
The following lemma will be useful in the sequel.
Lemma 2.11. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗,
(2.18)
H = H′ ⊕ H′′ ⊕ K
Γ11
Γ12
Γ13
: A∗ → H′ ⊕ H′′ ⊕ K.
Moreover, let B ∈ B(H′′, H′ ⊕ H′′) be the operator with the block representation
(2.19)
(2.20)
and the block representations of Γ0 and Γ1 are
Γ01
Γ02
Γ03
Γ0 =
: A∗ → H′ ⊕ H′′ ⊕ K,
B =(cid:18)B1
Γ1 =
B2(cid:19) : H′′ → H′ ⊕ H′′
the domain dom B = H′′ )and let θ0 ∈ eC(H) be given by
with bK = dmul θ0 = {0} ⊕ K. Then:
such that B2 = B∗
(i)The equalities
(2.21)
θ0 = B ⊕ bK = {{h, B1h ⊕ B2h ⊕ k} : h ∈ H′′, k ∈ K}
(2.22)
S := Aθ0 = {bf ∈ A∗ : Γ01bf = 0, Γ03bf = 0, Γ11bf = B1Γ02bf , Γ12bf = B2Γ02bf }
(2.23)
define a symmetric extension S ∈ ext(A) and its adjoint S∗.
(ii) The collection Π′ = {H′, Γ′
j : S∗ → H′ defined by
2 (this means that B is a bounded symmetric operator in H′ ⊕ H′′ with
(2.24)
is a boundary triplet for S∗.
(iii) Let θ be a linear relation in H′ defined by
Γ′
0, Γ′
1} with operators Γ′
Γ′
1bf = Γ11bf − B1Γ02bf ,
S∗ = {bf ∈ A∗ : Γ03bf = 0, Γ12bf = B∗
0bf = Γ01bf ,
1Γ01bf + B2Γ02bf }
bf ∈ S∗
θ = θ′ ∔ dmul θ = {{h, θ′h + h′} : h ∈ dom θ′, h′ ∈ mul θ},
eθ′ =(cid:18)
1 ↾ dom θ′ B2(cid:19) : dom θ′ ⊕ H′′ → H′ ⊕ H′′
B∗
B1
θ′
with a linear operator θ′ : dom θ′ → H′ (dom θ′ ⊂ H′) and let the extension eA ∈ ext(S)
be given by eA = Sθ (in the triplet Π′). Moreover, let eθ′ : dom θ′ ⊕ H′′ → H′ ⊕ H′′ be a
linear operator, defined by
8
V.I. MOGILEVSKII
and let eθ be a linear relation in H given by
(2.25)
eθ :=eθ′ ∔ dmuleθ = {{h ⊕ ϕ, (θ′h + B1ϕ + h′) ⊕ (B∗
muleθ = mul θ ⊕ K
(2.26)
Then eA ∈ ext(A) and eA = Aeθ (in the triplet Π).
If in addition θ ∈ B(H′), then
h ⊕ ϕ ∈ dom θ′ ⊕ H′′, h′ ∈ mul θ, k ∈ K}.
1h + B2ϕ) ⊕ k} :
and (2.26) takes the form
eθ′ =(cid:18) θ B1
B∗
1 B2(cid:19) : H′ ⊕ H′′ → H′ ⊕ H′′.
(2.27)
eθ =eθ′ ⊕ bK = {h ⊕ ϕ, (θ + B1ϕ) ⊕ (B∗
1 h + B2ϕ) ⊕ k} : h ⊕ ϕ ∈ H′ ⊕ H′′, k ∈ K}
Proof. (i) Clearly the relation θ0 is symmetric and by Proposition 2.10, (v) S := Aθ0 ∈
ext(A) is symmetric. Moreover, by (2.19) the second equality in (2.22) holds. Since
obviously
θ∗
0 = {{h1 ⊕ h2, h′ ⊕ (B∗
1 h1 + B2h2) ⊕ k} : h1 ⊕ h2 ∈ H′ ⊕ H′′, h′ ∈ H′, k ∈ K}
and by Proposition 2.10, (iii) S∗ = Aθ∗
0 , the equality (2.23) is valid.
0 and Γ′
1.
1bf , Γ′
Next assume that h′
1, h′
Π) and (2.23) one obtains
0bf , Γ′
1bg).
This proves Green's identity (2.15) for operators Γ′
2 ∈ H′. Since the mapping Γ = (Γ0, Γ1)⊤ is surjective, there is
(ii) Let bf = {f, f ′},bg = {g, g′} ∈ S∗. By using Green identity (2.15) (for the triplet
(f ′, g) − (f, g′) = (Γ11bf , Γ01bg) − (Γ01bf , Γ11bg) + (Γ12bf , Γ02bg) − (Γ02bf , Γ12bg)+
(Γ13bf , Γ03bg) − (Γ03bf , Γ13bg) = (Γ11bf , Γ01bg) − (Γ01bf , Γ11bg) + (Γ01bf , B1Γ02bg)+
(B2Γ02bf , Γ02bg) − (B1Γ02bf , Γ01bg) − (Γ02bf , B2Γ02bg) = ((Γ11 − B1Γ02)bf , Γ01bg)−
(Γ01bf , (Γ11 − B1Γ02)bg) = (Γ′
bf ∈ A∗ such that
Γ01bf = h′
1, Γ02bf = 0, Γ03bf = 0, Γ11bf = h′
Hence Γ12bf = B∗
1Γ01bf + B2Γ02bf and by (2.23) bf ∈ S∗. Moreover, in view of (2.24) one
1bf = h′
0bf = h′
(iii) It follows from (2.23) and (2.24) that bf ∈ eA if and only if bf ∈ A∗ and
Γ01bf ∈ dom θ, Γ11bf − B1Γ02bf = θ′Γ01bf + h′, Γ12bf = B∗
1Γ01bf + B2Γ02bf , Γ03bf = 0
with some h′ ∈ mul θ. In turn, these conditions are equivalent to inclusions bf ∈ A∗ and
{Γ0bf , Γ1bf } ∈eθ. Hence eA = Aeθ (in the triplet Π).
Lemma 2.12. Let L1 and L2 be closed subspaces in a Hilbert space H such that L1∩L2 =
{0}. Then L1 ∔ L2 = L1 ∔ L2 if and only if there exists a closed subspace L′
1 ⊃ L1 in H
such that H = L′
0bg) − (Γ′
2, Γ12bf = B∗
2, which proves surjectivity of the mapping Γ′ = (Γ′
1, Γ13bf = 0.
has Γ′
1, Γ′
0, Γ′
1)⊤.
1h′
(cid:3)
1 ∔ L2.
1 ⊕ L2 with a closed subspace L′
Proof. Let L1 ∔ L2 = L1∔L2. Then H = (L1∔L2)⊕H′ with H′ = H⊖(L1∔L2) and hence
H = L′
1 ⊃ L1 be a closed
subspace in H such that H = L′
1) and π2 ∈ B(H, L2) be the
corresponding skew projections onto L′
1 and L2 respectively. Assume that fn ∈ L1 ∔ L2
and fn → f . Then π1fn ∈ L1, π1fn → π1f and hence π1f ∈ L1. Since π2f ∈ L2, it
follows that f (= π1f + π2f ) ∈ L1 ∔ L2. Therefore L1 ∔ L2 is closed.
(cid:3)
1 = L1 ⊕ H′. Conversely, let L′
1 ∔ L2 and let π1 ∈ B(H, L′
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
9
Proposition 2.13. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let A0 = ker Γ0
(2.28)
2.10, (i) that
and only if dom θ = dom θ).
linear operator B : dom θ → H0, dom θ ⊂ H0 (in particular, this assumption is satisfied
and let θ ∈ eC(H) be a linear relation such that decompositions (2.14) hold with a (closed)
for symmetric θ). Then Aθb+A0 = Aθb+A0 if and only if B ∈ B(dom θ, H0) (that is, if
Proof. Let bH = {0} ⊕ H and bH0 = {0} ⊕ H0. Since A0 = A bH, it follows from Proposition
Moreover, B ∩ bH0 = {0} and bH = bH0 ⊕ bK, which in view of (2.14) implies that θb+bH =
(B ∔ bH0) ⊕ bK. Therefore
Next, by Lemma 2.12 B ∔ bH0 = B ∔ bH0 if and only if there exists B′ ∈ eC(H0) such
that B ⊂ B′ and B′ ⊕ bH0 = H2
Aθb+A0 = Aθb+A0 ⇐⇒ θb+bH = θb+bH.
θb+bH = θb+bH ⇐⇒ B ∔ bH0 = B ∔ bH0.
0. Since the last equality is equivalent to the inclusion
B′ ∈ B(H0), the following equivalence holds:
(2.29)
(2.30)
Now combining (2.28), (2.29) and (2.30), one obtains the required statement.
Corollary 2.14. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let A0 = ker Γ0 and
(cid:3)
B ∔ bH0 = B ∔ bH0 ⇐⇒ B ∈ B(dom θ, H0).
let θ ∈ eC(H). Then Aθ ∩ A0 = A and Aθb+A0 = Aθb+A0 if and only if θ ∈ B(dom θ, H).
Proof. Since A0 = A{0}⊕H, it follows from Proposition 2.10, (i) that Aθ ∩ A0 = A if and
only if θ ∩ ({0} ⊕ H) = {0}, that is if and only if θ is an operator. Thus Aθ ∩ A0 = A
if and only if decompositions (2.9) hold with K = {0}. This and Proposition 2.13 yield
(cid:3)
the result.
In the rest of this subsection we recall some definitions and results from [9, 18, 10].
Proposition 2.15. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗ and let A0 = A∗
0 ∈
ext(A) be given by (2.17). Moreover, let π1 be the orthoprojection in H ⊕ H onto H ⊕ {0}.
Then the operator Γ0 ↾ bNλ(A), λ ∈ C \ R, isomorphically maps bNλ(A) onto H and hence
the equalities
γ(λ) = π1(Γ0 ↾ bNλ(A))−1,
Γ1 ↾ bNλ(A) = M(λ)Γ0 ↾ bNλ(A),
correctly define the operator functions γ(·) : C\ R → B(H, H) and M(·) : C\ R → B(H).
Moreover, γ(·) and M(·) satisfy the identities
λ ∈ C \ R
(2.31)
(2.32)
γ(λ) = γ(z) + (λ − z)(A0 − λ)−1γ(z)
M(z) − M ∗(λ) = (z − λ)γ∗(λ)γ(z),
z, λ ∈ C \ R
which imply that γ(·) and M(·) are holomorphic in C \ R and M(·) ∈ R[H].
Definition 2.16. The operator-functions γ(·) and M(·) defined in Proposition 2.15 are
called the γ-field and the Weyl function of the triplet Π respectively.
Remark 2.17. It follows from (2.31) and (2.32) that γ(·) and M(·) are the γ-field and
the Q-function of the pair (A, A0) respectively in the sense of [16, 17].
Definition 2.18. The (not necessarily closed) linear relation F in H defined by
is called a forbidden relation of the boundary triplet Π = {H, Γ0, Γ1} for A∗.
F := Γdmul A∗ = {{Γ0{0, n}, Γ1{0, n}} : n ∈ N(A)}
10
V.I. MOGILEVSKII
In the following theorem the forbidden relation F is characterized in terms of the
asymptotic behavior of the Weyl function.
Theorem 2.19. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let M(·) be the Weyl
function of Π and let BM and NM be the operators defined by (2.1) and (2.2), (2.3)
respectively. Moreover, let F be the forbidden relation of Π. Then
(2.33)
ran BM ⊂ mul F ,
ran BM = mul F ,
and F admits the representation
(2.34)
2.4. Exit space extensions and formula for generalized resolvents. As is known
F = NM ∔ dmul F = {{h, NM h + h′} : h ∈ dom NM , h′ ∈ mul F }.
(2.35)
λ ∈ C \ R
a linear relation eA = eA∗ in a Hilbert space eH ⊃ H is called an exit space extension of A
if A ⊂ eA and the minimality condition span{H, (eA − λ)−1H : λ ∈ C \ R} =eH is satisfied.
For an exit space extension eA ∈ eC(eH) of A the compressed resolvent
is called a generalized resolvent of A (here PH is the orthoprojection ineH onto H). If two
exit space extensions eA1 ∈ eC(eH1) and eA2 ∈ eC(eH2) of A generates the same generalized
resolvent R(λ), then eA1 and eA2 are equivalent. The latter means that there exists a
unitary operator V ∈ B(eH1 ⊖ H,eH2 ⊖ H) such that eA2 = eUeA1 with the unitary operator
1,eH2
eU = (IH ⊕ V ) ⊕ (IH ⊕ V ) ∈ B(eH2
2). Hence each exit space extension eA of A is defined
R(λ) = PH(eA − λ)−1 ↾ H,
by the generalized resolvent (2.35) uniquely up to the equivalence.
[9, 18] Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗,
Theorem 2.20.
A0 = ker Γ0 and γ(·) and M(·) are the γ-field and the Weyl function of Π respectively.
Then: (i) the equality (Krein formula for generalized resolvents)
(2.37)
(2.36)
λ ∈ C \ R
λ ∈ C \ R;
PH(eAτ − λ)−1 ↾ H = (A0 − λ)−1 − γ(λ)(τ (λ) + M(λ))−1γ∗(λ),
PH(eAτ − λ)−1 ↾ H = (A−τ (λ) − λ)−1,
establishes a bijective correspondence eA = eAτ between all relation valued functions τ =
τ (λ) ∈ eR(H) and all exit space self-adjoint extensions eA of A. Moreover, for each
τ ∈ eR(H) the following equality holds:
(ii) an extension eAτ is canonical (that is, eAτ ∈ eC(H)) if and only if τ (λ) ≡ θ(= θ∗), λ ∈
C \ R. In this case eAτ = A−θ (in the sense of Proposition 2.10, (ii)).
It follows from Theorem 2.20 that Krein formula (2.36) gives a parametrization eA = eAτ
of all exit space extensions eA = eA∗ of A in terms of functions τ (·) ∈ eR(H). Without
3.1. Parametrization of compressions. Assume thateH ⊃ H is a Hilbert space, Hr :=
eH⊖H, PH is the orthoprojection ineH onto H and eA = eA∗ ∈ eC(eH) is an exit space extension
connection with boundary triplets formula (2.36) was originally proved in [15, 17].
3. Compressions of exit space self-adjoint extensions
of A. In the following we let
(3.1)
(3.2)
(3.3)
S(eA) := eA ∩ H2 = {bf ∈ H2 : bf ∈ eA}
T (eA) := {{PHf, PHf ′} : {f, f ′} ∈ eA}
C(eA) := PHeA ↾ H = {{f, f ′} ∈ H2 : {f, f ′ ⊕ f ′
r} ∈ eA with some f ′
r ∈ Hr}
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
11
Proof. According to (2.37)
λ ∈ C \ R,
(3.4)
(3.5)
Π).
(3.6)
Π).
The following theorem directly follows from the results of [6, 8].
−τ (λ) = {h ⊕ ϕ, (−τ1(λ)h + B1ϕ) ⊕ (B∗
1 h + B2ϕ) ⊕ k} : h ⊕ ϕ ∈ H′ ⊕ H′′, k ∈ K}.
Lemma 3.3. Let the assumptions of Lemma 2.11 be satisfied. Moreover, let S ∈ ext(A)
be symmetric extension (2.22) of A, let Π′ = {H′, Γ′
1} be boundary triplet (2.24)
0, Γ′
Clearly, S(eA) is a closed symmetric relation in H, C(eA) is a symmetric relation in H,
T (eA) is a linear relation in H and
A ⊂ S(eA) ⊂ C(eA) ⊂ T (eA) ⊂ A∗.
Moreover, S(eA) ⊂ eA and according to [21, 7] (S(eA))∗ = T (eA).
Definition 3.1. The linear relation C(eA) is called the compression of eA.
Theorem 3.2. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let τ ∈ eR(H) and let
eAτ = eA∗
τ be the corresponding exit space extension of A. Then the equalities S(eA) = A
and T (eA) = T (eA) hold if and only if τ ∈ Ru[H].
for S∗, let τ1(·) ∈ Ru[H′] and let eA = eSτ1 be the corresponding exit space self-adjoint
extension of S. Assume also that τ = {H′ ⊕ H′′ ⊕ K, B1, B2, τ1} ∈ eRc(H) (see Remark
2.6). Then eA ⊃ A and eA = eAτ (in the triplet Π).
PH(eA − λ)−1 ↾ H = (eA(λ) − λ)−1,
where eA(λ) = S−τ1(λ) (in the triplet Π′). Moreover,
Comparing this equality with (2.27) we obtain from Lemma 2.11, (iii) that eA(λ) ∈ ext(A)
and eA(λ) = A−τ (λ) (in the triplet Π). Therefore by (3.5) and (2.37) eA = eAτ (in the triplet
Proposition 3.4. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let τ ∈ eRc(H) and
let eAτ = eA∗
(see Remark 2.6), Γ0 and Γ1 have the block representations (2.19) and θ0 ∈ eC(H) is given
by (2.21). Then the extension S = S(eAτ ) admits the representation (2.22) (in the triplet
1} be boundary triplet (2.24) for S∗. Assume that eA := eSτ1 (in the
triplet Π′). Since τ1 ∈ Ru[H′], it follows from Theorem 3.2 that S(eA) = S. Moreover,
according to Lemma 3.3 A ⊂ eA and eA = eAτ (in the triplet Π), which proves the required
let τ ∈ eRc(H). Then S(eAτ ) ∩ A0 = A if and only if τ ∈ Rc[H].
S(eAτ ) = Aθ0 with θ0 given by (2.21). Since obviously Aθ0 ∩ A0 = A bK, the equivalences
S(eAτ ) ∩ A0 = A ⇔ K = {0} ⇔ τ ∈ Rc[H] are valid.
eAτ corresponding to τ ∈ eRc(H).
Proof. Clearly the assumptions of Lemma 2.11 are satisfied with the subspaces H′, H′′, K
and the operator B = (B1, B2)⊤. Let S ∈ ext(A) be symmetric extension (2.22) and
let Π′ = {H′, Γ′
statement.
Corollary 3.5. Let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let A0 = ker Γ0 and
Proof. Assume that τ is represented as in (3.6). Then according to Proposition 3.4
In the following theorem we give a geometric characterization of self-adjoint extensions
τ be the corresponding exit space extension of A. Assume also that
(cid:3)
(cid:3)
(cid:3)
τ = {H′ ⊕ H′′ ⊕ K, B1, B2, τ1}
0, Γ′
12
V.I. MOGILEVSKII
Theorem 3.6. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗ and A0 =
0, Γ′
A. Then:
H = H0 ⊕ K,
decompositions
only if τ ∈ Rc[H].
T = T . Since S ∈ ext(A) and S ⊂ S∗, it follows from Proposition 2.10, (ii) and (v) that
ker Γ0. Let τ ∈ eR(H) and let eAτ be the corresponding exit space self-adjoint extension of
(i) The linear relations S(eAτ )b+A0 and T (eAτ ) are closed if and only if τ ∈ eRc(H).
(ii) S(eAτ ) ∩ A0 = A and the linear relations S(eAτ )b+A0 and T (eAτ ) are closed if and
Proof. (i) We put S = S(eAτ ) and T = T (eAτ ). Assume that Sb+A0 = Sb+A0 and
S = Aθ0 with some symmetric relation θ0 ∈ eC(H). Moreover, by Proposition 2.13 the
holds with K = mul θ0, bK = dmul θ0 and B ∈ B(dom θ0, H0), where dom θ0 ⊂ H0 is
closed. Let H′′ = dom θ0 and H′ = H0 ⊖ H′′, so that H0 = H′ ⊕ H′′. Then decomposition
(2.18) of H holds, the operator B admits the block representation (2.20) and θ0 can
be written in the form (2.21). Moreover, since B is symmetric, the equality B2 = B∗
2
holds. Therefore by Lemma 2.11, (ii) the equalities (2.24) define the boundary triplet
Π′ = {H′, Γ′
θ0 = B ⊕ bK
Π′ = {H′, Γ′
the triplet Π′) and in view of Theorem 3.2 one has T = T .
of H, an operator B = (B1, B2)⊤ ∈ B(H′′, H′ ⊕ H′′) and a function τ1 ∈ Ru[H′] such
that τ = {H′ ⊕ H′′ ⊕ K, B1, B2, τ1}. It follows from Proposition 3.4 that S = Aθ0 with
1} for S∗. Since eA = eAτ is an extension of S, it follows from Theorem
3.2 that eA = eAτ1 (in the triplet Π′) with some τ1 ∈ Ru[H′]. Therefore by Lemma 3.3
τ ∈ eRc(H).
Conversely, let τ ∈ eRc(H). Then by Corollary 2.5 there exist a decomposition (2.10)
θ0 given by (2.21). Therefore by Proposition 2.13 Sb+A0 = Sb+A0. Next assume that
1} is a boundary triplet (2.24) for S∗. Then by Lemma 3.3 eAτ = eSτ1 (in
τ is the corresponding exit space extension of A. Let eAτ ∈ eC(eH) with a
and eAτ = eA∗
Hilbert space eH ⊃ H and let Hr =eH ⊖ H. Then:
(i) There exist a symmetric relation Ar ∈ eC(Hr) and a boundary triplet Πr = {H, Γr
r : Γ0bf = Γr
(ii) C(eAτ ) = A−Fr (in the triplet Π), were Fr is the forbidden relation of Πr.
Proposition 3.7. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗, τ ∈ Ru[H]
eAτ = {bf ⊕ bfr ∈ A∗ ⊕ A∗
(ii) Combining statement (i) with Corollary 3.5 we arrive at statement (ii).
1bfr}.
0bfr, Γ1bf = −Γr
Proof. Statement (i) directly follows from the results of [6, 8].
r such that τ is the Weyl function of Πr and
(i) According to Proposition 2.10, (ii) statement (ii) is equivalent to the equality
for A∗
0, Γ′
(cid:3)
0, Γr
1}
(3.7)
(3.8)
(3.9)
and hence f ′
C(eAτ ) = {bf ∈ A∗ : {Γ0bf , −Γ1bf} ∈ Fr}.
r. Moreover, by (3.8) {Γ0bf , −Γ1bf} = {Γr
Let bf = {f, f ′} ∈ C(eAτ ). Then bf ∈ A∗ and by (3.3) there exists f ′
r} ∈ eAτ . Letting bfr = {0, f ′
r one getsbg = bf ⊕bfr. Therefore bfr ∈ A∗
bg := {f, f ′ ⊕ f ′
1bfr} ∈ Fr.
0bfr, Γr
Conversely, let bf = {f, f ′} ∈ A∗ and {Γ0bf , −Γ1bf } ∈ Fr. Then there exists f ′
r} = −Γ1bf . Letting bfr = {0, f ′
r} = Γ0bf and Γr
from (3.8) thatbg := bf ⊕bfr = {f, f ′ ⊕ f ′
r} ∈ eAτ and therefore bf = {f, f ′} ∈ C(eAτ ). This
r ∈ mul A∗
r
r we obtain
r ∈ Hr such that
such that Γr
r ∈ mul A∗
proves (3.9).
r} ∈ A∗
r} ∈ H2
1{0, f ′
0{0, f ′
(cid:3)
r
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
13
In the following theorem we parameterize in terms of τ the compressions C(eAτ ) of exit
space extensions eAτ with τ ∈ eRc(H).
Theorem 3.8. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗, τ ∈ eRc(H),
eAτ = eA∗
τ is the corresponding exit space extension of A and C(eAτ ) is the compression
of eAτ . Moreover, let τ0 ∈ Rc[H0] and K be the operator and multivalued parts of τ
be operators corresponding to τ0 in accordance with Proposition 2.2. Then C(eAτ ) = Aθc
respectively (see (2.9)) and let Bτ0 ∈ B(H0) and Nτ0 : dom Nτ0 → H0 (dom Nτ0 ⊂ H0)
(in the triplet Π) with the symmetric linear relation θc in H given by
θc = −Nτ0 ∔ dmul θc = {{h, −Nτ0h + h′} : h ∈ dom Nτ0, h′ ∈ mul θc}.
(3.10)
ran Bτ0 ⊕ K ⊂ mul θc,
mul θc = ran Bτ0 ⊕ K.
Moreover,
(3.11)
If in addition ran Bτ0 is closed, then mul θc = mul θc = ran Bτ0 ⊕ K and hence
θc = {{h, −Nτ0h + Bτ0ψ + k} : h ∈ dom Nτ0, ψ ∈ H0, k ∈ K}.
(3.12)
Proof. Assume that τ = {H′⊕H′′⊕K, B1, B2, τ1} with H′⊕H′′ = H0 and τ1 ∈ Ru[H′] (see
Corollary 2.5 and Remark 2.6). Moreover, let Γ0 and Γ1 have the block representations
admits the representation (2.22) and by Lemma 2.11, (ii) the equalities (2.24) define a
boundary triplet Π′ = {H′, Γ′
(in the triplet Π′) and by Proposition 3.7 there exist a Hilbert space Hr, a symmetric
r such that τ1 is the
(2.19) and let θ0 ∈ eC(H) be given by (2.21). Then according to Proposition 3.4 S = S(eAτ )
1} for S∗. Moreover, according to Lemma 3.3 eAτ = eSτ1
relation Sr ∈ eC(Hr) and a boundary triplet Πr = {H′, Γr
Weyl function of Πr and C(eAτ ) = S−Fr (in the triplet Π′), where Fr is the forbidden
relation of Πr. It follows from Theorem 2.19 that
1} for S∗
0, Γr
0, Γ′
ran Bτ1 ⊂ mul Fr,
mul Fr = ran Bτ1
(3.13)
(3.14)
where Nτ1 : dom Nτ1 → H′ is the operator given by
(3.15)
−Fr = −Nτ1 ∔ dmul Fr = {{h, −Nτ1h + h′} : h ∈ dom Nτ1, h′ ∈ mul Fr},
yIm (τ1(iy)h, h) < ∞}
dom Nτ1 = {h ∈ H′ : lim
y→∞
τ1(iy)h,
Nτ1h = lim
y→∞
c : dom θ′
h ∈ dom Nτ1.
Let dom θ′
c := dom Nτ1 ⊕ H′′ and θ′
c → H0 be the operator given by
θ′
c =(cid:18)
−Nτ1
1 ↾ dom Nτ1 B2(cid:19) : dom Nτ1 ⊕ H′′ → H′ ⊕ H′′.
B1
B∗
Then by Lemma 2.11 C(eAτ ) = Aθc (in the triplet Π) with the linear relation θc in H
given by
mul θc = mul Fr ⊕ K
(3.16)
(3.17)
(3.18)
(3.19)
It follows from (2.8) (with τ0(λ) in place of τ (λ)) that
θc = θ′
c ∔ dmul θc = {{h, θ′
ch + h′} : h ∈ dom θ′
c, h′ ∈ mul θc}
yIm (τ0(iy)h, h) = yIm (τ1(iy)PH′h, PH′h),
h ∈ H0.
Therefore by (3.15) dom θ′
for each h = h′ ⊕ h′′ ∈ dom θ′
c one obtains
c = dom Nτ0. Moreover, combining (3.17), (3.16) and (2.11)
θ′
ch = (−Nτ1h′ + B1h′′) ⊕ (B∗
1 h′ + B2h′′) =
− lim
y→∞
((τ1(iy)h′ − B1h′′) ⊕ (−B∗
1 h′ − B2h′′)) = − lim
y→∞
τ0(iy)h = −Nτ0h.
14
V.I. MOGILEVSKII
This and (3.19) yield (3.10).
Next, by (2.11) Bτ1 = Bτ0 PH′ and hence ran Bτ1 = ran Bτ0. Therefore by (3.13) and
(3.18) the relations (3.11) are valid. Finally, θc is symmetric in view Proposition 2.10,
(cid:3)
(v).
Theorem 3.9. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗, A0 = ker Γ0
(that is A0 is a fixed canonical self-adjoint extension of A in the Krein formula (2.36)),
statements are equivalent:
τ ∈ eR(H) and τ0 ∈ R[H0] is the operator part of τ (see (2.14)). Then the following
(i) T (eAτ ) is closed and C(eAτ ) ⊂ A0;
(ii) τ ∈ eRc(H) and τ0 satisfies
yIm (τ0(iy)h, h) = ∞,
h ∈ H0, h 6= 0.
lim
y→∞
(3.20)
If statement (i) (or equivalently (ii)) is valid then
(3.21)
(3.22)
C(eAτ ) = {bf ∈ A∗ : Γ0bf = 0, Γ1bf ∈ ran Bτ0 ⊕ K}.
If in addition ran Bτ0 is closed, then C(eAτ ) is closed and
C(eAτ ) = {bf ∈ A∗ : Γ0bf = 0, Γ1bf = Bτ0 h ⊕ k with some h ∈ H0 and k ∈ K}.
Proof. Assume statement (i). Then by (3.4) S(eAτ ) ⊂ A0 and hence S(eAτ )b+A0 = A0.
Therefore S(eAτ )b+A0 is closed and by Theorem 3.6, (i) τ ∈ eRc(H). This in view of
Theorem 3.8 implies that C(eAτ ) = Aθc with θc given by (3.10). It follows from (3.10)
that C(eAτ ) ⊂ A0 if and only if dom Nτ0 = {0}, which yields (3.20). Thus (ii) holds.
Conversely, let (ii) holds. Then by Theorem 3.6, (i) T (eAτ ) is closed. Moreover, by
Theorem 3.8 C(eAτ ) = Aθc with θc of the form (3.10) and (3.20) shows that dom Nτ0 =
{0}. Hence θc = {0} ⊕ mul θc and therefore C(eAτ ) ⊂ A0, which yields statement (i).
Moreover, by Proposition 2.10, (iv) C(eAτ ) = Aθc, where in view of (3.11) θc = {0} ⊕
mul θc = {0} ⊕ Hc with Hc = ran Bτ0 ⊕ K. This yields (3.21). Finally, the last statement
(cid:3)
of the theorem follows from the previous one.
Corollary 3.10. Let the assumptions be the same as in Theorem 3.9. Then the following
statements are equivalent:
H0 ⊕ K. Hence ran Bτ0 = H0 or, equivalently, ker Bτ0 = {0}. Thus (ii) is valid.
If in addition ran Bτ0 is closed, then statement (ii) is equivalent to the following one:
(i) T (eAτ ) is closed and C(eAτ ) = A0;
(ii) τ ∈ eRc(H) and ker Bτ0 = {0}.
(i ′) T (eAτ ) is closed and C(eAτ ) = A0.
Proof. Assume that (i) holds. Then by Theorem 3.9 τ ∈ eRc(H) and C(eAτ ) = A{0}⊕Hc
with Hc = ran Bτ0 ⊕ K. Since A0 = A{0}⊕H, it follows from C(eAτ ) = A0 that Hc = H =
Conversely, let (ii) holds. Then by Theorem 3.6, (i) T (eAτ ) is closed. Moreover, by
Theorem 3.8 C(eAτ ) = Aθc with θc given by (3.10) and satisfying (3.11). Since ran Bτ0 =
in view of Proposition 2.10, (iv) C(eAτ ) = A{0}⊕H = A0. This proves statement (i).
versely, let (ii) holds. Then (i) is valid and hence C(eAτ ) = Aθc with θc = {0} ⊕ mul θc.
C(eAτ ) = A{0}⊕H = A0, which yields (i ′).
Moreover, ran Bτ0 = H0 and by Theorem 3.8 mul θc = H0 ⊕ K = H. Therefore
(cid:3)
Assume now that ran Bτ0 is closed. Then (i ′) implies (i) and, consequently, (ii). Con-
H0, it follows from (3.11) that mul θc = H. Hence by Lemma 2.7, (i) θc = {0} ⊕ H and
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
15
Corollary 3.11. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗ and τ ∈
(3.23)
lim
y→∞
h ∈ H, h 6= 0.
yIm (τ (iy)h, h) = ∞,
operator-function τ0 ∈ Rc[H0] in decomposition (2.9) of τ satisfies (3.20) and ran Bτ0 ⊕
K = {0}. Hence K = {0} and, consequently, H0 = H and τ = τ0. This implies that
τ ∈ Rc[H] and (3.23) holds. Moreover, ran Bτ0 = {0} and therefore Bτ = Bτ0 = 0.
eR(H). Then T (eAτ ) is closed and C(eAτ ) = A if and only if τ ∈ Rc[H], Bτ = 0 and
Proof. Let T (eAτ ) be closed and C(eAτ ) = A. Then by Theorem 3.9 τ ∈ eRc(H), the
Conversely, let τ ∈ Rc[H], Bτ = 0 and (3.23) is satisfied. Then by Theorem 3.9 T (eAτ )
is closed and C(eAτ ) is given by (3.21) with ran Bτ0 = ran Bτ = {0} and K = {0}. Hence
C(eAτ ) = C(eAτ ) = ker Γ and by Proposition 2.10, (i) C(eAτ ) = A.
each exit space extension eA = eA∗ of A is a densely defined operator and according to
M.A. Naimark (see e.g. [1]) an extension eA of A is said to be of the second kind if dom eA∩
H = dom A or equivalently if C(eA) = A. Clearly, in the case dom A = H Corollary
3.11 gives a parametrisation of all extensions eA of the second kind satisfying T (eA) =
T (eA). Note that this result follows from the results of [8], where a parametrisation of
all extensions eA ⊃ A of the second kind of a densely defined A was obtained in terms
of the parameter τ from the Krein formula (2.36). Observe also that a somewhat other
parametrization of the second kind extensions can be found in [22, 25].
Remark 3.12. Assume that A is a closed densely defined symmetric operator in H. Then
(cid:3)
3.2. Exit space self-adjoint extensions with a self-adjoint compression. In the
following proposition we provide a sufficient condition on the parameter τ for self-
adjointness of the compression C(eAτ ).
Proposition 3.13. Assume that Π = {H, Γ0, Γ1} is a boundary triplet for A∗, τ ∈ eRc(H)
and τ0 ∈ Rc[H0] is the operator part of τ (see (2.9)). If ran Bτ0 is closed and
(3.24)
lim
y→∞
yIm (τ0(iy)h, h) < ∞,
h ∈ ker Bτ0,
Proof. Assume that ran Bτ0 is closed and (3.24) holds. Then according to Theorem 3.8
then C(eAτ ) is self-adjoint.
C(eAτ ) = Aθc with a symmetric linear relation θc in H given by (3.10) and satisfying
mul θc = mul θc = ran Bτ0 ⊕ K. Moreover, by (3.24) ker Bτ0 ⊂ dom Nτ0 = dom θc.
Therefore
H ⊖ mul θc = H0 ⊖ ran Bτ0 = ker Bτ0 ⊂ dom θc
and by Lemma 2.7, (ii) θc = θ∗
2.10, (v).
c . Now the required statement follows from Proposition
(cid:3)
Corollary 3.14. Let the assumptions be the same as in Proposition 3.13. If ran Bτ0
is closed and there is a compact interval [a, b] ⊂ R such that τ0 admits a holomorphic
continuation onto R \ [a, b], then C(eAτ ) is self-adjoint.
Proof. If the function τ0 is holomorphic in C\[a, b], then it admits the Laurent expansion
τ0(λ) = C0 + λBτ0 +
1
λk
∞Xk=1
Ck,
λ > R > 0
16
V.I. MOGILEVSKII
with operator-valued coefficients Ck = C ∗
k ∈ B(H0), k = 0, 1, 2, . . . . Therefore
Im (τ0(iy)h, h) =
and,consequently,
(−1)m
y2m−1 (C2m−1h, h),
∞Xm=1
y ∈ R, h ∈ ker Bτ0
lim
y→∞
yIm (τ0(iy)h, h) = −(C1h, h) < ∞,
h ∈ ker Bτ0.
Thus (3.24) holds and the required statement follows from Proposition 3.13.
(cid:3)
Theorem 3.15. Assume the conditions of Theorem 3.9. Then the following statements
are equivalent:
In the following theorem we describe exit space self-adjoint extensions eA of A such
that the compression C(eA) of eA is self-adjoint and transversal with A0.
(i) T (eAτ ) is closed, C(eAτ )∗ = C(eAτ ) and the extensions C(eAτ ) and A0 are transversal;
(ii) τ ∈ Rc[H] and
yIm (τ (iy)h, h) < ∞,
h ∈ H.
(3.25)
lim
y→∞
Moreover, if (i) (equivalently (ii)) is satisfied, then
(3.26)
where Nτ ∈ B(H) is the operator given by
C(eAτ ) = A−Nτ = {bf ∈ A∗ : Γ1bf = −Nτ Γ0bf},
(3.27)
Nτ = s- lim
y→∞
τ (iy).
Proof. Let statement (i) holds. Since S(eAτ ) ⊂ C(eAτ ), it follows from Lemma 2.12 that
the linear relation S(eAτ )b+A0 is closed. Moreover, by (3.4) S(eAτ ) ∩ A0 = A. Therefore
by Theorem 3.6, (ii) τ ∈ Rc[H]. Next, according to Theorem 3.8 C(eAτ ) = Aθc, where θc
is given by (3.10) with τ0 = τ , and by Proposition 2.10, (vii) θc ∈ B(H). Therefore by
(3.10) dom Nτ = dom θc = H, which in view of (2.2) yields (3.25). Thus statement (ii)
is valid. Moreover, since mul θc = {0}, it follows from (3.10) that θc = −Nτ and in view
of definition (2.3) of Nτ the equality (3.27) holds. This yields (3.26).
transversal.
c ∈ B(H). Therefore by
dom Nτ = H. Since θc is symmetric, this implies that θc = θ∗
Conversely, assume (ii). Then by Theorem 3.6 T (eAτ ) is closed. Moreover, by Theorem
3.8 C(eAτ ) = Aθc, where θc is given by (3.10) with τ0 = τ and (3.25) yields dom θc =
Proposition 2.10, (v) and (vii) C(eAτ )∗ = C(eAτ ) and the extensions C(eAτ ) and A0 are
3.3. Finite-codimensional exit space extensions. LeteH ⊃ H be a Hilbert space, let
Hr = eH ⊖ H and let eA = eA∗ ∈ eC(eH) be an exit space extension of A. Such an extension
Proposition 3.16. Let A ∈ eC(H) be a symmetric linear relation with finite deficiency
(hence dim H = d < ∞). Moreover, let τ ∈ eR(H) and let τ0 ∈ R[H0] be the operator
part of τ (see (2.9)). Then eAτ is finite-codimensional if and only if
indices n+(A) = n−(A) =: d < ∞ and let Π = {H, Γ0, Γ1} be a boundary triplet for A∗
The following proposition is well known (see e.g. [12]).
is called finite-codimensional, if nr := dim Hr < ∞.
(cid:3)
(3.28)
τ0(λ) = A + λB +
Aj,
λ ∈ C \ R
1
αj − λ
lXj=1
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
17
where
(3.29)
αj ∈ R, A, B, Aj ∈ B(H0), A = A∗, B ≥ 0, Aj ≥ 0, Aj 6= 0, j = 1, 2, . . . , l.
Moreover,
(3.30)
nr = dim ran B +
dim ran Aj.
lXj=1
In the following two theorems we extend Proposition 3.16 to a ceratin class of self-
∞, j = 1, 2, . . . , l.
satisfies (3.30).
given by
(ii) τ0 is the rational function (3.28), (3.29) such that dim ran B < ∞ and dim ran Aj <
indices n+(A) = n−(A) ≤ ∞, let Π = {H, Γ0, Γ1} be a boundary triplet for A∗, let
space extension of A (see Theorem 2.20). Assume also that τ0 ∈ R[H0] is the operator
part of τ (see (2.9)). Then the following statements are equivalent:
adjoint extensions eA of a symmetric relation A with possibly infinite deficiency indices
n+(A) = n−(A) ≤ ∞ and characterise compressions C(eA) of such extensions.
Theorem 3.17. Let A ∈ eC(H) be a symmetric linear relation in H with equal deficiency
τ ∈ eC(eH) (eH ⊃ H) be the corresponding exit
A0 = ker Γ0, let τ ∈ eR(H) and let eAτ = eA∗
(i) eAτ is finite-codimensional and the linear relation S(eAτ )b+A0 is closed;
Moreover, if statement (ii) (and hence (i)) is valid, then the dimension nr of Hr =eH⊖H
Proof. Assume statement (i) and let Sr = Sr(eAτ ) and Tr(eAτ ) be linear relations in Hr
Since dim Hr < ∞, it follows that Tr(eAτ ) is closed and according to [7, Proposition
2.15] T (eAτ ) is closed as well. Therefore by Theorem 3.6 τ ∈ eRc(H). Let in accordance
S = S(eAτ ) admits the representation (2.22). Let Π′ = {H′, Γ′
(2.24) for S∗. Then by Lemma 3.3 eAτ = eSτ1 (in the triplet Π′). Moreover, according to
with Corollary 2.5 τ0 be decomposed as in (2.11) with τ1 ∈ Ru[H′], that is τ = {H′ ⊕
H′′ ⊕ K, B1, B2, τ1} in the sense of Remark 2.6. Then by Proposition 3.4 the relation
1} be boundary triplet
[7, Lemma 2.14] n±(S) = n±(Sr) and, consequently, n+(S) = n−(S) < ∞. Therefore
dim H′ < ∞ and by Proposition 3.16
r,
Sr = eAτ ∩ H2
Tr(eAτ ) = {{PHrf, PHrf ′} : {f, f ′} ∈ eAτ }.
0, Γ′
(3.31)
τ1(λ) = A0 + λB0 +
1
αj − λ
lXj=1
A0j,
λ ∈ C \ R,
where αj ∈ R and A0, B0, A0j ∈ B(H′) are the operators satisfying A0 = A∗
0, B0 ≥ 0
and A0j ≥ 0, A0j 6= 0, j = 1, 2, . . . , l.This and the block representation (2.11) of τ0(λ)
imply that τ0(λ) is of the form (3.28) with
(3.32)
A =(cid:18) A0 −B1
1 −B2(cid:19) ,
−B∗
B =(cid:18)B0 0
0(cid:19) ,
0
Hence A = A∗, B ≥ 0, Aj ≥ 0, Aj 6= 0 and
Aj =(cid:18)A0j 0
0(cid:19) , j = 1, 2, . . . , l.
0
(3.33)
dim ran B = dim ran B0 < ∞,
dim ran Aj = dim ran A0j < ∞,
which yields statement (ii).
18
V.I. MOGILEVSKII
Conversely, let statement (ii) holds. Put ϕ0(λ) = Im λ and ϕj(λ) = Im 1
αj−λ. Then by
(3.28)
(3.34)
Im τ0(λ) = ϕ0(λ)B +
lXj=1
ϕj(λ)Aj,
λ ∈ C \ R.
Since Im λ · Im ϕj(λ) > 0, j = 0, 1, . . . , l, it follows from (3.34) that
(3.35)
H′′ := ker Im τ0(λ) = ker B ∩ l\j=1
ker Aj!
and, consequently, H′ := H0 ⊖ H′′ = ran B + ran A1 + · · ·+ ran Al. Therefore dim H′ < ∞
and in view of (3.35) block representations (3.32) hold (with respect to the decomposition
H0 = H′ ⊕H′′). Combining (3.32) with (3.28) one obtains the block representation (2.11)
of τ0(λ) with τ1 ∈ R[H′] given by (3.31). Since dim H′ < ∞, it follows that τ1 ∈ Ru[H′]
and according to Proposition 2.3 τ0 ∈ Rc[H0]. Therefore τ ∈ eRc(H) and by Theorem 3.6
S(eAτ )b+A0 is closed.
Now it remains to show that eAτ is finite-codimensional and (3.30) holds. Since τ =
{H′ ⊕ H′′ ⊕ K, B1, B2, τ1}, it follows from Proposition 3.4 that S = S(eAτ ) admits the
by Lemma 3.3 eAτ = eSτ1 (in the triplet Π′). Thus in accordance with Proposition 3.16
eAτ is finite-codimensional and
representation (2.22) and by Lemma 2.11 the equalities (2.24) define a boundary triplet
Π′ = {H′, Γ′
1} for S∗. Therefore by Proposition 2.9 n±(S) = dim H′ < ∞. Moreover,
0, Γ′
nr = dim ran B0 +
dim ran A0j.
lXj=1
Combining this equality with (3.33) we arrive at (3.30).
(cid:3)
Theorem 3.18. Let the assumptions of Theorem 3.17 be satisfied and let claim (ii)
(equivalently (i)) of this theorem be valid. Moreover, let A′ = Pker BA ↾ ker B ∈ B(ker B),
let K be the multivalued part of τ (see (2.9)) and let θc be a self-adjoint linear relation
in H given by
(3.36)
θc = {{h, −A′h ⊕ h′ ⊕ k} : h ∈ ker B, h′ ∈ ran B, k ∈ K}
(this means that −A′ is the operator part of θc and mul θc = ran B ⊕ K). Then the
Moreover, the following holds:
extension eAτ is finite-codimensional, the compression C(eAτ ) of eAτ is self-adjoint and
C(eAτ ) = Aθc (in the triplet Π).
(i) C(eAτ ) = A0 if and only if ker B = 0 (that is B > 0);
(ii) C(eAτ ) and A0 are transversal if and only if τ ∈ R[H] and B = 0, in which case
Proof. It follows from Theorem 3.17 that eAτ is finite-codimensional. Moreover, by The-
orem 3.8 C(eAτ ) = θc with symmetric relation θc in H given by (3.10). Since in view of
θc = −A.
(3.28)
yIm (τ0(iy)h, h) = y2(Bh, h) +
lXj=1
y2
j + y2 (Ajh, h),
α2
h ∈ H0,
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
19
it follows that lim
y→∞
yIm (τ0(iy)h, h) < ∞ if and only if h ∈ ker B. Therefore by definitions
(2.2) and (2.3) dom Nτ0 = ker B and
Nτ0h = lim
y→∞
τ0(iy)h = Ah,
h ∈ ker B(= dom Nτ0)
Moreover, by (2.1) Bτ0 = B and the finite-dimensionality of ran B yields ran Bτ0 =
ran Bτ0 = ran B. Therefore by Theorem 3.8 mul θc = ran B ⊕ K and (3.10) can be written
as
(3.37)
θc = {{h, −Ah + (h′ ⊕ k)} : h ∈ ker B, h′ ∈ ran B, k ∈ K}
Since H0 = ker B ⊕ ran B and h′ ∈ ran B in (3.37) is arbitrary, the equality (3.37) is
equivalent to (3.36). Finally, statements (i) and (ii) are implied by Corollary 3.10 and
(cid:3)
Theorem 3.15.
3.4. The case of finite deficiency indices. In this subsection we suppose that the
following assumption (a) is satisfied:
(a) A is a closed symmetric linear relation in H with finite deficiency indices n+(A) =
adjoint.
(3.40)
(3.38)
lim
y→∞
yIm (τ0(iy)h, h) = ∞,
h ∈ H0, h 6= 0.
τ is the corresponding exit space extension of A
Π) with the symmetric linear relation θc in H defined by (3.12).
Theorem 3.20. Let the assumption (a) be satisfied and let A0 = ker Γ0. Then:
and ran Bτ = ran Bτ for any subspace H′ ⊂ H and operator-function τ ∈ R[H′]. These
facts together with Theorems 3.8, 3.9 and Corollaries 3.10, 3.11 yield the following two
theorems.
n−(A) < ∞, Π = {H, Γ0, Γ1} is a boundary triplet for A∗, τ ∈ eR(H), τ0 ∈ R[H0] is the
operator part of τ (see (2.9)), eAτ = eA∗
and C(eAτ ) is the compression of eAτ .
Then ext(A) = ext(A) and hence for each exit space extension eA = eA∗ of A the linear
relations C(eA), T (eA) and S(eA)b+A0 are closed (here A0 = ker Γ0). Moreover, under
the assumption (a) dim H < ∞, which implies that eR(H) = eRc(H), R[H′] = Rc[H′]
Theorem 3.19. Let the assumption (a) be satisfied. Then C(eAτ ) = Aθc (in the triplet
(i) C(eAτ ) ⊂ A0 if and only if the operator function τ0 satisfies
In this case C(eAτ ) is defined by abstract boundary conditions (3.22).
(ii) C(eAτ ) = A0 if and only if ker Bτ0 = 0 (that is Bτ0 > 0 ).
(iii) C(eAτ ) = A if and only if τ ∈ R[H], Bτ = 0 and
In the following theorem we describe all exit space extensions eA = eA∗ of a symmetric
relation A with finite deficiency indices such that the compression C(eA) of eA is self-
(i) C(eAτ ) is self-adjoint if and only if
If in particular τ0 is holomorphic out of a compact interval [a, b] ⊂ R, then C(eAτ ) is
Theorem 3.21. Let the assumption (a) be satisfied and let A0 = ker Γ0. Then:
lim
y→∞
yIm (τ0(iy)h, h) < ∞,
h ∈ ker Bτ0.
yIm (τ (iy)h, h) = ∞,
h ∈ H, h 6= 0.
self-adjoint.
(3.39)
lim
y→∞
20
V.I. MOGILEVSKII
(ii) C(eAτ ) is self-adjoint and transversal with A0 if and only if τ ∈ R[H] and
yIm (τ (iy)h, h) < ∞,
h ∈ H.
lim
y→∞
In this case
where Nτ = N ∗
C(eAτ ) = A−Nτ = {bf ∈ A∗ : Γ1bf = −Nτ Γ0bf},
τ ∈ B(H) is the operator given by Nτ = lim
y→∞
τ (iy).
Proof. (i) According to Theorem 3.19 C(eAτ ) = Aθc with a symmetric linear relation θc
in H defined by (3.12). It follows from (3.12) that mul θc = ran Bτ0 ⊕ K and therefore
H ⊖ mul θc = H0 ⊖ ran Bτ0 = ker Bτ0. Moreover by (3.12) dom θc = dom Nτ0 and hence
the inclusion H ⊖ mul θc ⊂ dom θc is equivalent to (3.40). Therefore by Lemma 2.7, (iii)
(3.40) is equivalent to self-adjointness of θc, which by Proposition 2.10, (v) is equivalent
of the statement (i) is implied by Corollary 3.14.
to C(eAτ )∗ = C(eAτ ). Thus C(eAτ )∗ = C(eAτ ) if and only if (3.40) holds. The second part
(ii) Statement (ii) directly follows from Theorem 3.15 and arguments before Theorem
(cid:3)
3.19.
In the following theorem we characterize compressions of finite-codimensional exten-
relation θc = θ∗
(3.18) are valid.
this theorem follow from Theorem 3.18 and arguments before Theorem 3.19.
Theorem 3.22. Let under the assumption (a) τ0 be a rational function (3.28), (3.29)
(according to Proposition 3.16 this assumption is equivalent to the finite-codimensionality
sions eA = eA∗ of a symmetric relation A with finite deficiency indices. The statements of
of eAτ ). Moreover, let A′ = Pker BA ↾ ker B and let K be the multivalued part of τ (see
(2.9)). Then C(eAτ ) is a self-adjoint linear relation in H and C(eAτ ) = Aθc with the
c ∈ eC(H) given by (3.36). Moreover, statements (i) and (ii) of Theorem
Corollary 3.23. Let eH be a Hilbert space, let H be a subspace of a finite codimension
in eH (this means that dim(eH ⊖ H) < ∞) and let eA = eA∗ be a linear relation in eH. Then
the compression C(eA) = PHeA ↾ H of eA (see (3.3)) is a self-adjoint linear relation in H.
Proof. Clearly, eA is an exit space extension of the symmetric relation S = S(eA) (see
Therefore by Theorem 3.22 C(eA) is self-adjoint.
case of an operator eA = eA∗ see [24]).
(3.1)). Moreover, it was shown in the proof of Theorem 3.17 that n+(S) = n−(S) < ∞.
(cid:3)
Remark 3.24. Statement of Corollary 3.23 was proved by another method in [3] (for the
References
[1] N. I. Akhiezer, I. M.Glazman, Theory of linear operators in Hilbert space, vol. I and II, Pitman,
Boston-London-Melbourne, 1981.
[2] T.Ya. Azizov, A. Dijksma, Closedness and adjoints of products of operators, and compressions,
Int. Equ. Oper. Theory 74 (2012), 259-269.
[3] T.Ya. Azizov, A. Dijksma, and G. Wanjala, Compressions of maximal dissipative and self-adjoint
linear relations and of dilations, Linear Algeb. Appl. 439 (2013), 771- 792.
[4] T.Ya. Azizov, B. ´Curgus, and A. Dijksma, Finite-codimensional compressions of symmetric and
self-adjoint linear relations in Krein spaces, Int. Equ. Oper. Theory 86 (2016), 71-95.
[5] V. M. Bruk, Extensions of symmetric relations, Math. Notes 22 (1977), no.6, 953 -- 958.
[6] V.A. Derkach, S. Hassi, M.M. Malamud, and H.S.V. de Snoo, Generalized resolvents of symmetric
operators and admissibility, Methods of Functional Analysis and Topology 6 (2000), no. 3 , 24 -- 55.
[7] V. A. Derkach, S. Hassi, M. Malamud, and H. S.V. de Snoo,Boundary Relations and their Weyl
families, Trans. Amer. Math. Soc. 358 (2006), no. 12, 5351 -- 5400.
ON COMPRESSIONS OF SELF-ADJOINT EXTENSIONS
21
[8] V.A. Derkach, S. Hassi, M.M. Malamud, and H.S.V. de Snoo, Boundary relations and generalized
resolvents of symmetric operators, Russian J. Math. Ph. 16 (2009), no. 1, 17 -- 60.
[9] V.A. Derkach and M.M. Malamud, Generalized resolvents and the boundary value problems for
Hermitian operators with gaps, J. Funct. Anal. 95 (1991),1 -- 95.
[10] V.A. Derkach and M.M. Malamud, The extension theory of Hermitian operators and the moment
problem, J. Math. Sciences 73 (1995), no. 2, 141-242.
[11] A. Dijksma and H. Langer, Finite-dimensional self-adjoint extensions of a symmetric operator
with finite defect and their compressions, In: Advances in complex analysis and operator theory,
Festschrift in honor of Daniel Alpay, Birkhauser, Basel, (2017), 135-163.
[12] A. Dijksma and H. Langer,Compressions of self-adjoint extensions of a symmetric operator and
M.G. Kreins resolvent formula, Integr. Equ. Oper. Theory 90:41 (2018).
[13] A. Dijksma and H.S.V. de Snoo, Self-adjoint extensions of symmetric subspaces, Pasif. J. Math.
54 (1974), no. 1, 71 -- 100.
[14] V.I. Gorbachuk and M.L. Gorbachuk, Boundary problems for differential-operator equations,
Kluver Acad. Publ., Dordrecht-Boston-London, 1991. (Russian edition: Naukova Dumka, Kiev,
1984).
[15] M.G. Krein and H. Langer,On defect subspaces and generalized resolvents of a Hermitian operator
in the space Πκ, Funct. Anal. Appl. 5 (1971/1972), 136 -- 146, 217 -- 228
[16] M.G Krein and H. Langer, Uber die Q-functions eines π-hermiteschen operators in raume Πκ,
Acta. Sci. Math. (Szeged) 34 (1973), 191 -- 230.
[17] H. Langer and B. Textorious, On generalized resolvents and Q-functions of symmetric linear
relations (subspaces) in Hilbert space, Pacif. J. Math. 72(1977), no. 1 , 135 -- 165.
[18] M. M. Malamud, On the formula of generalized resolvents of a nondensely defined Hermitian
operator, Ukr. Math. Zh. 44(1992), no. 12, 1658 -- 1688.
[19] B. Sz.-Nagy, C. Foias, Harmonic Analysis of Operators in Hilbert Space, Paris and Akad..Kiado,
Budapest, 1967.
[20] M.A. Nudelman, A generalization of Stengers lemma to maximal dissipative operators. Integr.
Equ. Oper. Theory 70 (2011), 301-305.
[21] A.V.Shtraus, On selfadjoint operators in the orthogonal sum of Hilbert spaces, Dokl. Akad. Nauk
SSSR, 144 (1962), no.3, 512-515.
[22] A. V. Shtraus, On one-parameter families of extensions of a symmetric operator, Izv. Akad. Nauk
SSSR, Ser. Mat. 30 (1966), 1325-1352.
[23] A. V. Shtraus, Extensions and generalized resolvents of a symmetric operator which is not densely
defined, Izv. Akad. Nauk SSSR. Ser. Mat. 34 (1970),no.1, 175-202. (Russian); English translation:
Mathematics of the USSR-Izvestiya, 4 (1970), no. 1, 179 -- 208.
[24] W. Stenger, On the projection of a selfadjoint operator, Bull. Am. Math. Soc. 74(1968), 369-372.
the Shtraus
[25] S.M. Zagorodnyuk, Generalized resolvents of symmetric and isometric operators:
approach, Ann. Funct. Anal. 4 (2013), no. 1, 175 -- 285.
Department of Mathematical Analysis and Informatics, Poltava National V.G. Ko-
rolenko Pedagogical University, Ostrogradski Str. 2, 36000 Poltava, Ukraine
E-mail address: [email protected]
|
1504.05242 | 1 | 1504 | 2015-04-20T21:43:57 | Criterion for $\mathbb{Z}_d$--symmetry of a Spectrum of a Compact Operator | [
"math.FA",
"math.SP"
] | If $A$ is a compact operator in a Banach space and some power $A^q$ is nuclear we give a criterion of $\mathbb{Z}_{d}$ -- symmetry of its spectrum $\sigma{A}$ in terms of vanishing of the traces $\mathop{\mathit{Trace}} A^n$ for all $n$, $n \geq 0$, $n \neq 0 \mod d$, sufficiently large. | math.FA | math |
CRITERION FOR Zd -- SYMMETRY OF A SPECTRUM OF A COMPACT
OPERATOR
BORIS S. MITYAGIN
Abstract. If A is a compact operator in a Banach space and some power Aq is nuclear
we give a criterion of Zd -- symmetry of its spectrum σ(A) in terms of vanishing of the
traces Trace An for all n, n ≥ 0, n 6= 0 mod d, sufficiently large.
In the case of matrices, or linear operators T : X → X in a finite-dimensional space, one
can check (prove) that the following conditions are equivalent.
(a) The spectrum of T is symmetric, or Z2-symmetric, i.e., λ ∈ σ(T ) → −λ ∈ σ(T )
and their algebraic multiplicities m(λ), m(−λ) are equal;
(b) Trace T p = 0 for all odd p ∈ N.
M. Zelikin [Zel08] observed and proved that this claim could be extended to S1, the trace-
class operators in a Hilbert space. We will show that such claims could be made
(i) in general Banach spaces;
(ii) for Zd symmetry of a spectrum, d ≥ 2.
Of course, we need to make sure that Trace is well -- defined if we write conditions like
(b). Then, the formula for the trace Trace A = Pj λj(A) should be properly explained
if we use it. We now recall a few notions and facts about nuclear operators (see more in
[Kon86]).
An operator A : X → Y between two Banach spaces is called nuclear if it has represen-
tation
(1)
where
(2)
Ax =
qXk=1
akfk(x)yk,
q ≤ ∞
ak > 0,
kfkX ′k ≤ 1,
a∗ =X ak < ∞,
kykY k ≤ 1,
1
and
∀k.
2
BORIS S. MITYAGIN
A linear space of nuclear operators X → Y is a Banach space N (X; Y ) with the norm
(3)
kAk1 = inf{a∗ : (1), (2)}.
A linear functional Trace is well-defined on N (X; X) (for any Banach Space X) by
(4)
Of course,
(5)
and kTracek = 1.
Trace A =
akfk(yk).
qXk=1
kTrace Ak ≤ kAk1,
A. Grothendieck [Gro95] showed that for operators (1) [X = Y ],
(6)
(7)
(8)
if
qXk=1
a2/3
k < ∞
then Xλj(A) < ∞
Trace A =X λj(A).
where points of the spectrum σ(A) are enumerated with their multiplicity, and
The presentation (4) with (1) to (3) gives a factorization
(9)
where
(10)
(11)
with
(12)
A = J F, X
F−→ ℓ2(N)
J−→ X,
F x =
J ξ =
∞X1
∞X1
a1/2
k fk(x)ek,
and
a1/2
k ξkyk,
kF k ≤ (a∗)1/2,
kJk ≤ (a∗)1/2
Moreover, the product F J is a Hilbert-Schmidt operator, or of the Schatten class S2 in
a Hilbert space ℓ2(N); see more in [GK69], [Sim79]. Indeed,
(13)
hF J ek, emi = a1/2
k a1/2
m fm(yk)
CRITERION FOR Zd -- SYMMETRY
3
and
(14)
so kF Jk2 ≤ a∗.
hF J ek, emi2 =
∞Xk,m=1
∞Xk,m=1
akamfm(yk)2 ≤ (a∗)2
By Holder inequality for Schatten classes ([GK69] or [Sim79]),
(15)
kBCDk2/3 ≤ kBk2kCk2kDk2
so (F J)3 ∈ S2/3 and has a representation
(16)
(F J)3 =
where kfkk, khkk ≤ 1 and
(17)
Therefore,
∞Xk=1
∞Xk=1
ckh·, fkihk,
c > 0,
c2/3
k < ∞.
(18)
has
2
3
(19)
with
(20)
A4 = J(F J)3F =
ckhF (·), fkiJ hk
∞Xk=1
-property (6) and
λj(Aq) < ∞ for all q ≥ 4,
∞Xj=1
Trace Aq =
λj(Aq)
∞Xj=1
More careful geometric analysis, based on approximative characteristics of operators [MP66],
[Pie87] -- if we use [Kon80], or [Kon86, Theorem 4.a.6, p. 227] -- shows that we can lower
q in (19), (20) to 3. Indeed, (F J)2 is in S1(ℓ2(N)), so there are finite-dimensional operators
Gn, Rank Gn ≤ n, such that
(21)
Then
(22)
αn < ∞,
where αn := k(F J)2 − Gnk
Xn
kA3 − J GnF k ≤ a∗ · αn
4
BORIS S. MITYAGIN
and by [Kon86, Theorem 4.a.6]
(23)
(24)
and
(25)
A3 is nuclear,
λj(A3) ≤ 2a∗Xn
Xj
αn < ∞,
Trace A3 =Xj
λj(A3).
But this remark will not improve our Theorem 1 (below) in an essential way (just in (41)
we can say p ≥ p∗ ≥ 3q∗).
In a Hilbert space X = H by Lisdkiı Theorem [Lid59], for any trace-class operator
C ∈ S1,
(26)
and
(27)
λj(C) < ∞
∞Xj=1
Trace C =
λj(C).
∞Xj=1
Maybe, talking just about nuclear operators, M. Zelikin considered in [Zel08, Thm. 2] only
Hilbert spaces.
Before stating our main result let us recall [DS58, Chapter VII, Sections 3 and 4] elements
of Riesz theory of compact operators.
If T : X → X is compact its spectrum σ(T ) is discrete with 0 being the only accumulation
point, and it has the following properties
(i) for any ρ > 0, σ(T ) ∩ {z : z ≥ ρ} is a finite set;
(ii) if
(28)
(29)
δ(α) =
1
2
min{α − λ : λ ∈ σ(T ), λ 6= α}
[so δ(α) > 0 for any α ∈ C \ 0] and
P (α) =
1
2πi
Z
z−α=δ(α)
(z − T )−1 dz,
CRITERION FOR Zd -- SYMMETRY
5
then
with
(30)
(31)
m(α) = Rank P (α) < ∞, α ∈ C \ {0}
m(α) = 0
if and only if α 6∈ σ(T ).
For α ∈ σ(T ) \ 0, m(α) is an algebraic multiplicity of an eigenvalue α.
The operational calculus [DS58, Chapter VII, Sections 3 and 4] explains that for any ρ > 0
such that
(32)
we have
(33)
σ(T ) ∩ {z = ρ} = ∅
T = Xα>ρ
T (α) + S, where T (α) =
1
2πi
Z
z−α=δ(α)
z(z − T )−1 dz
is an operator of rank m(α) with
(34)
and
(35)
σ(T (α)) = {α},
S =
1
2πi Zz=ρ
z(z − T )−1 dz.
Moreover, for any entire function F (z), say, for polynomials,
(36)
F (T ) = Xα>ρ
F (T (α)) + F (S),
where by the Riesz-Cauchy formulae,
1
2πi
Z
z−α=δ(α)
(37) F (T (α)) =
It follows that
F (z)(z − T )−1 dz, F (S) =
1
2πi Zz=ρ
F (z)(z − T )−1 dz.
(38)
(39)
Trace F (T (α)) = F (α) · m(α).
F (T (α)) = 0
if F (j)(α) = 0,
0 ≤ j ≤ m(α).
Now we are ready to prove
6
BORIS S. MITYAGIN
Theorem 1. Let T be a compact operator in a Banach space X, and some power T q∗ is a
nuclear operator. Then σ(T ) is Zd -- symmetric, i.e., for any β ∈ C \ {0},
(40)
if and only if
(41)
m(βωk) = m(β) for all k = 0, 1, . . . d − 1, ω = exp(cid:18)i
2π
d (cid:19)
Trace T dp+r = 0,
1 ≤ r ≤ d − 1,
for all sufficiently large p, say p ≥ p∗ ≥ 4q∗.
Of course, if d = 2, this is an extension of [Zel08] , Thm. 2, to a Banach case.
Proof. Part 1: (40) ⇒ (41). This is an "algebraic" claim although first we notice: the
assumption p ≥ 4q∗ guarantees that all operators T n, n = dp+r, in (41) satisfy 2
3 -- condition
so by Grothendieck theorem
(42)
Trace T n =
λj(T n)
∞Xj=1
and the absolute convergence permits to rearrange the terms of the right sum as we wish
to write
(43)
With
(44)
Trace T n =X µ · m(µ; T n)
m(µ; T n) = 0 for µ 6∈ σ(T n)
we can "add" the terms with µ 6∈ σ(T n) and this does not change the right side in (43).
For
(45)
so
(46)
n = dp + r ∈ (41) define g = gcd{r, d}
r = ag,
d = bg,
(a, b) = 1
and with r ≤ d − 1 we have 1 ≤ a < b. For any µ ∈ C \ {0} take its Zb-orbit, i.e.,
(47)
eµ = {µ · τ j : 0 ≤ j < b},
τ = ωq = exp(cid:18)i
2π
b (cid:19).
CRITERION FOR Zd -- SYMMETRY
7
The sum in (43) could be written as
(48)
XZb−orbits
b−1Xj=0
µτ j · m(µτ j; T n)
where for certainty µ in the orbit (47) is chosen as µ = µeiϑ, 0 ≤ ϑ < 2π
b . Now we will
show that the sum in (48) over each orbit is equal to zero. With numbers as in (45) put
κ = exp(cid:18)i
2π
n(cid:19) so κn = 1 and notice that if µ = λn, we choose
λ = µ1/neiϑ′
, ϑ′ =
ϑ
n
,
(cid:16)λωk(cid:17)n
= µωk(dp+r) = µωkr = µτ ak
µτ j m(µτ j; T n) =
(1)
=
(49)
then
(50)
and
(51)
b−1Xj=0
m(λωkκs; T )
(2)
=
m(λκs · ωk; T )
(3)
=
=
=
1
g
1
g
µτ akm(µτ ak; T n)
d−1Xk=0
d−1Xk=0(cid:16)λωk(cid:17)n
n−1Xs=0
d−1Xk=0(cid:16)λωk(cid:17)n
n−1Xs=0
d−1Xk=0
n−1Xs=0
m(λκs; T )!
µ n−1Xs=0
b−1Xj=0
m(λκs; T )µ
1
g
1
g
=
τ αk (4)
=
τ j (5)
= 0
The steps in (51) are justified in the following way.
(1) comes from (50).
(2) is just
(3) uses in essential way the theorem's
the change of order of the double summation.
assumption (40) on m(βωk) being independent on k. (4) is bases on the properties of the
roots ω, τ , ωd = 1, τ = ωg under (46). Of course, in (5)
times over {τ j}b−1
j=0. Part (40) ⇒ (41) is proven.
b−1Xj=0
τ j = 0, and {τ ak}d−1
k=0 runs g
(cid:3)
8
BORIS S. MITYAGIN
Proof. Part 2: (41) ⇒ (40). Take λ 6= 0 and as before
(52)
n = dp∗ + dp + r,
1 ≤ r ≤ d − 1,
p ≥ 0
and 0 < ρ < λ is such that
σ(T ) ∩ {z ∈ C : z = ρ} = ∅,
(53)
with
(54)
(55)
where
(56)
(57)
Then by (39)
(58)
(59)
being the Zd-orbit of λ. Now we use (36) for the special choice F = Fpr with
eλ = {λωk : 0 ≤ k ≤ d − 1}
Fpr(z) =(cid:16) z
λ(cid:17)dp∗+dp+r
(cid:18) zd − αd
λd − αd(cid:19)m(α)
=
ϕ(z),
ϕ(z) = Yα≥ρ
α∈σ(T )
α6∈eλ
= ψ(zd),
and ψ is a polynomial.
ϕ(T (α)) = 0,
Fpr(T (α)) = 0,
α > ρ
∀α 6∈eλ,
Trace Fpr(T (β)) = m(β)Fpr(β) = m(λωk)ωkr.
but for β ∈eλ, i.e., β = λωk,
(60)
Therefore,
(61)
where
(62)
Put
(63)
ωkrm(λωk) + Trace Fpr(S)
Trace Fpr(T ) =
Fpr(S) =(cid:18) T
λ(cid:19)dp∗
·
d−1Xk=0
2πi Zz=ρ (cid:16) z
1
λ(cid:17)dp+r
ϕ(z)(z − T )−1 dz.
Φ = max{ϕ(z) : z ≤ ρ}
CRITERION FOR Zd -- SYMMETRY
9
and with (53)
(64)
Then
(65)
where
(66)
and
(67)
M = max{kR(z; T )k : z = ρ} < ∞
kFpr(S)k1 ≤ Ctp,
any r,
1 ≤ r ≤ d − 1,
C =
Φ · M · ρ · kT dp∗k1
λdp∗
λ(cid:19)d
t =(cid:18) ρ
< 1.
Now by (41) and (61)
(68)
0 =
d−1Xk=0
ωkrm(λωk) + ξpr
for any p ≥ 1 and r,
1 ≤ r ≤ d − 1.
d−1Xk=0
The sum
(69)
does not depend on p but the remainder by (65) to (67) have estimates
ξpr ≤ Ctp
so
ξpr → 0 (p → ∞)
This implies by (68)
(70)
or
(71)
ωkrm(λωk) = 0,
∀r,
1 ≤ r ≤ d − 1
d−1Xk=0
yk = m(λωk),
1 ≤ k ≤ d − 1.
is a solution of the system
(72)
d−1Xk=1
ωkryk = −y0,
1 ≤ r ≤ d − 1.
Its determinant is of Vandermonde type so
(73)
det{ωkr}d−1
k,r=1 6= 0,
10
BORIS S. MITYAGIN
and the identities
(74)
show that by (72)
(ωr)k = 0,
d−1Xk=0
∀r,
1 ≤ r ≤ d − 1
(75)
yk = y0, i.e., m(λωk) = m(ω), ∀k,
1 ≤ k ≤ d − 1.
This proves that the multiplicity function m is constant on Zd -- orbits in C \ {0}, and (40)
is proven.
(cid:3)
It is worth to notice that the proof of Part II does not use any form of Grothendick or
Lidskii thoerem but it uses only properties of a linear function Trace on N (X; X) and an
elementary formula for Trace K when K is an operator of finite rank.
Acknowledgements
I thank M. Zelikin who told me his results [Zel08] in September 2007 during the 18th
Crimean Autumn Mathematical School -- Symposium, in Laspi -- Batiliman, Crimea. At that
time I realized that his result could be extended to the Banach space case.
In the late
2007 I had useful discussions with P. Djakov, E. Gorin, H. Konig, M. Solomyak. Recently,
during the Aleksander Pelczynski Memorial Conference, July 2014, Bedlewo, Poland, O.
Reinov brought my attention to his papers [Rei12], [RL13]. I thank C. Baker and O. Reinov
for recent discussions.
References
[DS58] Nelson Dunford and Jacob T. Schwartz. Linear Operators. I. General Theory.
With the assistance of W. G. Bade and R. G. Bartle. Pure and Applied Mathe-
matics, Vol. 7. Interscience Publishers, Inc., New York; Interscience Publishers,
Ltd., London, 1958, pp. xiv+858.
[GK69]
I. C. Gohberg and M. G. Kreın. Introduction to the theory of linear nonselfadjoint
operators. Translated from the Russian by A. Feinstein. Translations of Mathe-
matical Monographs, Vol. 18. American Mathematical Society, Providence, R.I.,
1969, pp. xv+378.
[Gro95] A. Grothendieck. "Produits tensoriels topologiques et espaces nucl´eaires". In:
S´eminaire Bourbaki, Vol. 2. Soc. Math. France, Paris, 1995, Exp. No. 69, 193 --
200.
REFERENCES
11
[Kon80] Hermann Konig. "s-numbers, eigenvalues and the trace theorem in Banach spaces".
In: Studia Math. 67.2 (1980), pp. 157 -- 172. issn: 0039-3223.
[Kon86] Hermann Konig. Eigenvalue distribution of compact operators. Vol. 16. Operator
Theory: Advances and Applications. Birkhauser Verlag, Basel, 1986, p. 262. isbn:
3-7643-1755-8.
[Lid59] V. B. Lidskiı. "Non-selfadjoint operators with a trace". In: Dokl. Akad. Nauk
SSSR 125 (1959), pp. 485 -- 487. issn: 0002-3264.
[MP66] Boris S. Mityagin and A. Pelczy´nski. "Nuclear Operators and Approximative
Dimension". In: Proc. Inter. Congr. Math. Moscow, 1966, pp. 366 -- 372.
[Pie87] A. Pietsch. Eigenvalues and s-numbers. Vol. 43. Mathematik und ihre Anwen-
dungen in Physik und Technik [Mathematics and its Applications in Physics
and Technology]. Akademische Verlagsgesellschaft Geest & Portig K.-G., Leipzig,
1987, p. 360. isbn: 3-321-00012-1.
[Rei12] Oleg Reinov. "Some more remarks on Grothendieck-Lidskiı trace formulas". In:
J. Prime Res. Math. 8 (2012), pp. 5 -- 11. issn: 1817-2725.
[RL13] Oleg Reinov and Qaisar Latif. "Grothendieck-Lidskiı theorem for subspaces of
Lp-spaces". In: Math. Nachr. 286.2-3 (2013), pp. 279 -- 282. issn: 0025-584X. doi:
10.1002/mana.201100112. url: http://dx.doi.org/10.1002/mana.201100112.
[Sim79] Barry Simon. Trace ideals and their applications. Vol. 35. London Mathematical
Society Lecture Note Series. Cambridge University Press, Cambridge-New York,
1979, pp. viii+134. isbn: 0-521-22286-9.
[Zel08] Mikhail I. Zelikin. "A criterion for the symmetry of a spectrum". In: Dokl. Akad.
Nauk 418.6 (2008), pp. 737 -- 740. issn: 0869-5652. doi: 10.1134/S1064562408010328.
url: http://dx.doi.org/10.1134/S1064562408010328.
231 West 18th Avenue, The Ohio State University, Columbus, OH 43201
E-mail address: [email protected]
|
1112.4985 | 2 | 1112 | 2012-01-19T16:08:02 | Multipliers and Wiener-Hopf operators on weighted L^p spaces | [
"math.FA"
] | We study the operators T on the weighted space L^p commuting either with the right translations St or left translations P^+S_{-t} and we establish the existence of a symbol of T. We characterize completely the spectrum of St. We obtain a similar result for the spectrum of P^+S_{-t} and some spectral results for the bounded operators commuting with (St), t>0 or with (P^+St), t<0. | math.FA | math |
MULTIPLIERS AND WIENER-HOPF OPERATORS ON WEIGHTED
Lp SPACES
VIOLETA PETKOVA
Abstract. We study the multipliers M (bounded operators commuting with the trans-
lations) on weighted spaces Lp
ω(R). We establish the existence of a symbol µM for M
and some spectral results for the translations St and the multipliers. We also study the
operators T on the weighted space Lp
ω(R+) commuting either with the right translations
St, t ∈ R+, or left translations P +S−t, t ∈ R+, and we establish the existence of a sym-
bol µ of T . We characterize completely the spectrum σ(St) of the operator St proving
that
σ(St) = {z ∈ C : z ≤ etα0},
where α0 is the growth bound of (St)t≥0. We obtain a similar result for the spectrum
of (P +S−t), t ≥ 0. Moreover, for an operator T commuting with St, t ≥ 0, we establish
the inclusion µ(O) ⊂ σ(T ), where O = {z ∈ C : Im z < α0}.
1. Introduction
Let E be a Banach space of functions on R. For t ∈ R, define the translation by t on
E by
Stf (x) = f (x − t), a.e., ∀f ∈ E.
We call a multiplier on E, every bounded operator on E commuting with St for every
t ∈ R. For the multipliers on a Hilbert space we have the existence of a symbol and some
spectral results concerning the translations and the multipliers are obtained by using this
property of the multipliers (see [7], [8]). In the arguments exploited in [7], [8] the spectral
mapping theorem of Gearhart [3] for semigroups in Hilbert spaces plays an essential role.
The first purpose of this paper is to extend the main results in [8], [7] concerning the
existence of the symbol of a multiplier as well as the spectral results in the case where
E is a weighted Lp
ω(R) space. For general Banach spaces the characterization of the
spectrum of the semigroup V (t) = etG by the resolvent of its generator G is much more
complicated than for semigroups in Hilbert spaces (see for instance [4]). In particular,
the statements of Lemma 1, 2 and 3 (see Section 2) are rather difficult to prove and for
general Banach spaces this problem remains open. In this paper we restrict our attention
to Lp
ω(R), 1 ≤ p < ∞, weighted spaces. The advantage that we take account is that the
semigroup of the translations (St) preserves the positive functions. For semigroups having
this special property in the spaces Lp
ω(R) we have a spectral mapping theorem (see [1],
[12], [13]). We obtain Theorems 1-4 for multipliers on Lp
ω(R) and in this work we explain
only these parts of the proofs which are based on spectral mapping techniques and which
are different from the arguments used to establish Theorems 1-4 in the particular case
1
2
VIOLETA PETKOVA
p = 2 (see for more details [8], [7]).
For a Banach space E denote by E ′ the dual space of E. For f ∈ E, g ∈ E ′, denote
by < f, g > the duality. Let p ≥ 1, and let ω be a weight on R. More precisely, ω is a
positive, continuous function such that
ω(x + t)
ω(x)
sup
x∈R
< +∞, ∀t ∈ R.
Let Lp
ω(R) be the set of measurable functions on R such that
kf kp,ω =(cid:16)ZR
f (x)pω(x)pdx(cid:17)1/p
< +∞, 1 ≤ p < +∞.
Let Cc(R) (resp. Cc(R+) ) be the space of continuous functions on R (resp. R+) with
compact support in R (resp. R+). Notice that Cc(R) is dense in Lp
In the following we set E = Lp
for 1 ≤ p < +∞. In this case
ω(R) and we consider only Banach spaces having this form
ω(R).
hf, gi =ZR
f (x)¯g(x)ω2(x)dx
and
hf, gi ≤ kf kp,ωkgkq,ω, for 1 < p < +∞,
where 1
p + 1
q = 1. For p = 1, we have
E′ = L∞
ω (R) = {f is measurable : f (x)ω(x) < ∞, a.e.}
and
If M is a multiplier on E then, there exists a distribution µ such that
kgk∞,ω = esssup {f (x)ω(x), x ∈ R}.
For φ ∈ Cc(R+), the operator
Mf = µ ∗ f, ∀f ∈ Cc(R+).
Mφ : Lp
ω(R) ∋ f −→ φ ∗ f
is a multiplier on E. Introduce
α0 = lim
t→+∞
ln kStk
1
t , α1 = lim
t→+∞
ln kS−tk
1
t .
It is easy to see that α1 + α0 ≥ 0. Consider
U = {z ∈ C, Im z ∈ [−α1, α0]}.
For an operator T denote by ρ(T ) the spectral radius of T and by σ(T ) the spectrum of
T . It is well known that ρ(St) = eα0t, for t ≥ 0.
Given a function f and a ∈ C, denote by (f )a the function
R ∋ x −→ f (x)eax
WIENER-HOPF OPERATORS
3
and denote by M the algebra of the multipliers on E. We note by g the Fourier transform
of a function g ∈ L2(R). Our first result is a theorem saying that every multiplier on E
has a representation by a symbol.
Theorem 1. Let M be a multiplier on E. Then
1) For a ∈ [−α1, α0], we have (Mf )a ∈ L2(R), for every f ∈ E such that (f )a ∈ L2(R).
2) For a ∈ [−α1, α0], there exists a function νa ∈ L∞(R) such that
\(Mf )a(x) = νa(x)d(f )a(x), ∀f ∈ E, with (f )a ∈ L2(R), a.e.
Moreover, we have kνak∞ ≤ CkMk, ∀a ∈ [−α1, α0].
3) If
◦
U 6= ∅, there exists a function ν ∈ H∞(
◦
U ) such that
dMf (z) = ν(z) f (z), z ∈
◦
U, ∀f ∈ C ∞
c (R),
where dMf (ia + x) = \(Mf )a(x), for a ∈] − α1, α0[, f ∈ C ∞
c (R).
The function ν is called the symbol of M. The above result is similar to that
established in [8], [7] and the novelty is that we treat Banach spaces Lp
ω(R) and not only
Hilbert spaces. Define A as the closed Banach algebra generated by the operators Mφ,
for φ ∈ Cc(R). Notice that A is a commutative algebra. Our second result concerns the
spectra of St and M ∈ M.
Theorem 2. We have
i) σ(St) = {z ∈ C, e−α1t ≤ z ≤ eα0t}, ∀t ∈ R.
Let M ∈ M and let µM be the symbol of M.
ii) We have
iii) If M ∈ A, then we have
µM (U) ⊂ σ(M).
µM (U) = σ(M).
(1.1)
(1.2)
(1.3)
The equality (1.3) may be considered as a weak spectral mapping property (see [2])
for operators in the Banach algebra A. On the other hand, it is important to note that if
M ∈ M, but M /∈ A, in general we have µM (U) 6= σ(M). For the space E = L1(R), there
exists a counter-example (see section 2 and [2]). Thus the inclusion in (1.2) could be strict.
In section 3, we obtain similar results for Wiener-Hopf operators on weighted Lp
ω(R+)
spaces. In the analysis of Wiener-Hopf operators some new difficulties appear in com-
parison with the case of multipliers.
Let E be a Banach space of functions on R+. Let p ≥ 1 and let ω be a weight on R+.
It means that ω is a positive, continuous function such that
0 < inf
x≥0
ω(x + t)
ω(x)
≤ sup
x≥0
ω(x + t)
ω(x)
< +∞, ∀t ∈ R+.
4
VIOLETA PETKOVA
Let Lp
ω(R+) be the set of measurable functions on R+ such that
Z ∞
f (x)pω(x)pdx < +∞.
0
Notice that Cc(R+) is dense in Lp
ω(R+).
Let P + be the projection from L2(R−) ⊕ Lp
denote by Sa the restriction of Sa on Lp
denoted by S. Let I be the identity operator on Lp
ω(R+) into Lp
ω(R+). From now we will
ω(R+) for a ≥ 0 and, for simplicity, S1 will be
ω(R+).
Definition 1. A bounded operator T on Lp
ω(R+) is called a Wiener-Hopf operator if
P +S−aT Saf = T f, ∀a ∈ R+, f ∈ Lp
ω(R+).
As in [5] we can show that every Wiener-Hopf operator T has a representation by a
convolution. More precisely, there exists a distribution µ such that
T f = P +(µ ∗ f ), ∀f ∈ C ∞
c (R+).
If φ ∈ Cc(R), then the operator
Lp
ω(R+) ∋ f −→ P +(φ ∗ f )
is a Wiener-Hopf operator and we will denote it by Tφ. Moreover, we have
but it is obvious that
(P +S−aSa)f = f, ∀f ∈ Lp
ω(R+),
(SaP +S−a)f 6= f,
for all f ∈ Lp
ω(R+) with a support not included in ]a, +∞[. The fact that Sa is not
invertible leads to many difficulties in contrast to the case when we deal with the space
Lp
ω(R).
Let E be the space Lp
ω(R+). As above define
a0 = lim
t→+∞
ln kStk
1
t , a1 = lim
t→+∞
ln kS−tk
1
t
and set J = [−a1, a0]. The next theorem is similar to Theorem 1.
Theorem 3. Let a ∈ J and let T be a Wiener-Hopf operator. Then for every f ∈ Lp
such that (f )a ∈ L2(R+), we have
ω(R+)
with ha ∈ L∞(R) and
(T f )a = P +F −1(had(f )a)
khak∞ ≤ CkT k,
(1.4)
where C is a constant independent of a. Moreover, if a1 + a0 > 0, the function h defined
on U = {z ∈ C : Im z ∈ J} by h(z) = hIm z(Re z) is holomorphic on
◦
U.
Definition 2. The function h defined in Theorem 3 is called the symbol of T .
WIENER-HOPF OPERATORS
5
We are able to examine the spectrum of the operators in the space W of bounded
operators on E commuting with (St)t≥0 or (P +S−t)t≥0.
Let O = {z ∈ C, Im z < a0} and V = {z ∈ C, Im z < a1}.
Theorem 4. We have
i) σ(St) = {z ∈ C, z ≤ ea0t}, ∀t > 0.
ii) σ(P +S−t) = {z ∈ C, z ≤ ea1t}, ∀t > 0.
Let T ∈ W and let µT be the symbol of T .
iii) If T commutes with St, ∀t ≥ 0, then we have
iv) If T commutes with P +S−t, ∀t ≥ 0, then we have
µT (O) ⊂ σ(T ).
µT (V) ⊂ σ(T ).
(1.5)
(1.6)
(1.7)
(1.8)
The equalities (1.5),(1.6) generalize the well known results for the spectra of the right
and left shifts in the space of sequences l2 (see for instance, [10]). However, our proofs
are based heavily on the existence of symbols for Wiener-Hopf operators and having in
mind Theorem 3, we follow the arguments in [9].
In section 4, we obtain a sharp spectral result for Wiener-Hopf operators having the
ω(R+).
ω(R+) our result
form Tφ with φ ∈ Cc(R). This result is established here for operators in spaces Lp
It is important to note that even for p = 2 and for the Hilbert space L2
below is new.
Theorem 5. Let φ ∈ Cc(R). Then
i) if supp (φ) ⊂ R+, we have
ii) if supp (φ) ⊂ R−, we have
φ(O) = σ(Tφ).
φ(V) = σ(Tφ).
The above result yields a weak spectral mapping property and can be compared with
the equality (1.3) in Theorem 2, however the proof is more complicated.
Recall that we use the notation E = Lp
ω(R). We start with the following
2. Multipliers on Lp
ω(R)
Lemma 1. Let λ ∈ C be such that eλ ∈ σ(S) and let Re λ = α0. Then there exists a
sequence (fn)n∈N of functions of E and an integer k ∈ Z so that
lim
n→∞(cid:13)(cid:13)(cid:16)etA − e(λ+2πki)t(cid:17)fn(cid:13)(cid:13) = 0, ∀t ∈ R, kfnk = 1, ∀n ∈ N.
(2.1)
6
VIOLETA PETKOVA
Proof. Let A be the generator of the group (St)t∈R. It is clear that the group (St)t∈R
ω(R) the results of [12], [13] say that the spectral
preserves positive functions. Since E = Lp
mapping theorem holds and
σ(etA) \ {0} = etσ(A) = {etλ : λ ∈ σ(A)}.
In particular, for the spectral bound s(A) of A we get
s(A) := sup{Re z : z ∈ σ(A)} = α0.
Thus eλ ∈ σ(S) \ {0} = eσ(A) yields λ + 2πki = λ0 ∈ σ(A) for some k ∈ Z. On the other
hand, Re λ0 = α0, and we deduce that λ0 is on the boundary of the spectrum of A. By
a well known result, this implies that λ0 is in the approximative point spectrum of A.
Let µn be a sequence such that µn →n→∞ λ0, Re µn > λ0, ∀n ∈ N. Then
hence k(µnI − A)−1k → ∞. Applying the uniform boundedness principle and passing to
a subsequence of µn (for simplicity also denoted by µn), we may find f ∈ E such that
(cid:13)(cid:13)(µnI − A)−1(cid:13)(cid:13) ≥ (dist (µn, σ(A)))−1,
lim
n→∞(cid:13)(cid:13)(µnI − A)−1f(cid:13)(cid:13) → ∞.
Introduce fn ∈ D(A) defined by
fn =
(µnI − A)−1f
k(µnI − A)−1f k
.
The identity
(λ + 2πki − A)fn = (λ0 − µn)fn + (µn − A)fn
implies that (λ + 2πki − A)fn → 0 as n → ∞. Then the equality
(etA − et(λ+2πki))fn =(cid:16)Z t
0
e(λ+2πki)(t−s)eAsds(cid:17)(A − λ − 2πki)fn
yields (2.1). (cid:3)
Now we prove the following important lemma.
Lemma 2. For all φ ∈ C ∞
c (R) and λ such that eλ ∈ σ(S) with Re λ = α0 we have
φ(iλ + a) ≤ kMφk, ∀a ∈ R.
(2.2)
Proof. Let λ ∈ C be such that eλ ∈ σ(S) and Re λ = α0 and let (fn)n∈N be the
sequence constructed in Lemma 1. We have
1 = kfnk =
sup
g∈E ′ , kgk
E
′ ≤1
< fn, g > .
Then, there exists gn ∈ E ′ such that
< fn, gn > −1 ≤
1
n
WIENER-HOPF OPERATORS
7
and kgnkE ′ ≤ 1. Fix φ ∈ C ∞
c (R) and consider
φ(iλ + a) ≤ φ(iλ + a)hfn, gni +
φ(iλ + a)
1
n
The first two terms on the right side of the last inequality go to 0 as n → ∞ since by
Lemma 1 we have
On the other hand,
φ(iλa)
1
n
≤(cid:12)(cid:12)(cid:12)ZR(cid:10)φ(t)(cid:16)e(λ+2πki)t − St(cid:17)e−i(a+2πk)tfn, gn(cid:11)dt(cid:12)(cid:12)(cid:12) +
hφ(t)Ste−i(a+2πk)tfn, gnidt(cid:12)(cid:12)(cid:12).
+(cid:12)(cid:12)(cid:12)ZR
(cid:13)(cid:13)e−i(a+2πk)t(cid:16)e(λ+2πki)t − St(cid:17)fn(cid:13)(cid:13) −→n→∞ 0.
In =(cid:12)(cid:12)(cid:12)ZR
< φ(t)Ste−i(a+2πk)tfn, gn > dt(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)hhZR
φ(t)e−i(a+2πk)tfn(. − t)dti, gni(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)hZR
(φ(. − y)ei(a+2πk)yfn(y)dy, ei(a+2πk).gni(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)h(cid:16)Mφ(ei(a+2πk).fn)(cid:17), ei(a+2πk).gni(cid:12)(cid:12)(cid:12)
and In ≤ kMφkkfnkkgnkE ′ ≤ kMφk. Consequently, we deduce that
(cid:3)
Notice that the property (2.2) implies that
φ(iλ + a) ≤ kMφk.
Lemma 3. Let φ ∈ C ∞
Then we have
φ(λ) ≤ kMφk, ∀λ ∈ C, provided Im λ = α0.
c (R) and let λ be such that e−¯λ ∈ σ((S−1)∗) with Re λ = −α1.
φ(iλ + a) ≤ k(Mφ)k, ∀a ∈ R.
(2.3)
t∈R acting on E ′. Let λ ∈ C be such that e−¯λ ∈
Proof. Consider the group (S−t)∗
σ((S−1)∗) and
e−¯λ = ρ(S−1) = ρ((S−1)∗) = eα1.
The group (S−t)∗ preserves positive functions. To prove this, assume that g(x) ≥ 0, a.e.
is a positive function and let h ∈ E be such that h(x) ≥ 0, a.e. Then
hh, (S−t)∗gi = hS−th, gi ≥ 0.
If F (x) = ((S−t)∗g)(x) < 0 for x ∈ Λ ⊂ R and Λ has a positive measure, we choose
h(x) = 1Λ(x). Then h1Λ, F i ≤ 0 and we conclude that F (x) = 0 a.e.
in Λ which is
a contradiction. For the group (S−t)∗ the spectral mapping theorem holds and, by the
and we deduce that for every t ∈ R we have
≤ k(St)∗kE ′→E ′(cid:13)(cid:13)(cid:16)e(−¯λ+2πmi)t − (S−t)∗(cid:17)gk(cid:13)(cid:13)E ′
k→∞(cid:13)(cid:13)(cid:16)(St)∗ − e(¯λ−2πmi)t(cid:17)gk(cid:13)(cid:13)E ′ = 0.
lim
8
VIOLETA PETKOVA
same argument as in Lemma 1, we prove that there exists a sequence (gk)k∈N of functions
of E ′ and an integer m so that for all t ∈ R,
k(etB − e(−¯λ+2πmi)t))gkkE ′ = 0
lim
k→∞
and kgkkE ′ = 1.
Since S−tSt = I, we have (St)∗(S−t)∗ = I. This implies that
(cid:13)(cid:13)(St)∗gk − e(¯λ−2πmi)tgk(cid:13)(cid:13)E ′ =(cid:13)(cid:13)(cid:16)(St)∗ − e(¯λ−2πmi)t(St)∗(S−t)∗(cid:17)gk(cid:13)(cid:13)E ′
For 1 < p < +∞ the space E = Lp
with E. Consequently, since kgkkE ′ = 1, there exists fk ∈ E such that
ω(R) is reflexive and the dual to E′ can be identified
< fk, gk > −1 ≤
1
k
, kfkkE ≤ 1.
(2.4)
For p = 1 the space L1
In this case the dual to L1
the set
ω(R) is not reflexive and to arrange (2.4), we use another argument.
ω (R) = 1. Fix 0 < ǫ < 1 and consider
ω (R). Let kgkL∞
ω(R) is L∞
Mǫ,m = {x ∈ R : g(x)ω(x) ≥ 1 − ǫ, m ≤ x < m + 1}, m ∈ Z.
If µ(Mǫ,m) (the Lebesgue measure of Mǫ,m) is zero for all m ∈ Z, we obtain a contra-
ω (R) = 1. Thus there exists r ∈ Z such that µ(Mǫ,r) > 0. Now we
diction with kgkL∞
take
f (x) =
1Mǫ,r(x)ei arg(g(x))
µ(Mǫ,r)ω2(x)
.
Then
1 ≥ hf, gi =Z r+1
r
f (x)¯g(x)ω2(x)dx ≥ 1 − ǫ
and we can obtain (2.4) choosing ǫ = 1/k. Passing to the proof of (2.3), we get
φ(iλ + a) ≤ φ(iλ + a) < fk, gk > +
φ(iλ + a)
1
k
≤(cid:12)(cid:12)(cid:12)ZR
< φ(t)e−i(a+2πm)tfk,(cid:16)e
< φ(t)St(cid:16)e−i(a+2πm)tfk(cid:17), gk > dt(cid:12)(cid:12)(cid:12) +
1
k
¯(λ−2πmi)t − (St)∗(cid:17)gk > dt(cid:12)(cid:12)(cid:12)
φ(iλ + a) = J ′
1
k
k → 0 as k → ∞. For I ′
k + I ′
k +
From the argument above we deduce that J ′
+(cid:12)(cid:12)(cid:12)ZR
same argument as in the proof of Lemma 2 and we deduce
φ(iλ + a).
k we apply the
φ(iλ + a) ≤ kMφk. (cid:3)
WIENER-HOPF OPERATORS
9
For the proof of Theorem 1 we apply the argument in [7] and Lemmas 2-3. There
exists eλ0 ∈ σ(S) such that Re λ0 = α0. Then for every z ∈ C with Im z = α0 we have
ϕ(z) ≤ kMϕk.
Also there exists e−λ1 ∈ σ((S−1)∗) with Re λ1 = −α1 and for every z ∈ C with Im z = −α1
we have
ϕ(z) ≤ kMϕk.
Applying Phragmen-Lindeloff theorem for the Fourier transform of ϕ ∈ C ∞
domain {z ∈ C : −α1 ≤ Im z ≤ α0}, we deduce
c (R) in the
ϕ(z) ≤ kMϕk
for z ∈ U. Next we exploit the fact that M can be approximated by Mϕ with respect
to the strong operator topology (see [6] for a very general setup covering our case). We
complete the proof repeating the arguments from [6], [7] and since this leads to minor
modifications, we omit the details. To obtain Theorem 2 we follow the same argument
as in [8] and the proof is omitted.
To see that in (1.2) the inclusion may be strict, consider a measure η on R such that
the operator
Mη : f −→ZR
Sx(f )dη(x)
is bounded on L1(R). For this it is enough to have RR dη(x) < ∞. Then Mη is a
multiplier on L1(R) with symbol
η(t) =ZR
e−ixtdη(x).
On the other hand, there exists a bounded measure η on R such that
η(R) 6= σ(Mη)
(see for details [2]). In L1(R) we have α0 = α1 = 0 and U = R. So we have not the
property (1.2) in Theorem 2 for every multiplier even in the case L1(R).
We need the following lemmas.
3. Wiener-Hopf operators
Lemma 4. Let φ ∈ Cc(R+). The operator Tφ commutes with St, ∀t > 0, if and only if
the support of φ is in R+.
Proof. Consider φ ∈ Cc(R+) and suppose that Tφ commutes with St, t ≥ 0. We write
φ = φχR− + φχR+.
10
VIOLETA PETKOVA
If Tφ commutes with St, t ≥ 0, then the operator TφχR− commutes too. Let ψ = φχR−
and fix a > 0 such that ψ has a support in [−a, 0]. Setting f = χ[0,a], we get Saf = χ[a,2a].
For x ≥ 0 we have
P +(ψ ∗ Saf )(x) =Z 0
ψ(t)χ{a≤x−t≤2a}dt =Z min(x−a,0)
−a
ψ(t)dt.
max(−a,−2a+x)
Since P +(ψ ∗ Saf ) = SaP +(ψ ∗ f ), for x ∈ [0, a], we deduce P +(ψ ∗ Saf )(x) = 0 and
Z x−a
−a
ψ(t)dt = 0, ∀x ∈ [0, a].
This implies that ψ(t) = 0, for t ∈ [−a, 0] hence supp(φ) ⊂ R+.
(cid:3)
Next we establish the following
Lemma 5. Let Tφ, φ ∈ Cc(R). Then Tφ commutes with P +(S−t), ∀t > 0 if and only if
supp(φ) ⊂ R−.
Proof. For φ ∈ Cc(R), suppose that Tφ commutes with P +(S−t), ∀t > 0. Set ψ =
φχR+. There exists a > 0 such that supp(ψ) ⊂ [0, a]. We have P +(ψ ∗ P +S−aχ[0,a]) = 0
and then P +S−a(P +ψ ∗ χ[0,a]) = 0. This implies that
On the other hand, for x > a we have
(ψ ∗ χ[0,a])(x) = 0, ∀x > a.
(ψ ∗ χ[0,a])(x) =ZR
ψ(t)χ[0,a](x − t)dt =Z min(a,x)
max(0,x−a)
ψ(t)dt =Z a
x−a
ψ(t)dt.
ǫ ψ(t)dt = 0, ∀a > ǫ > 0 and we get ψ = 0. Thus we conclude that supp(φ) ⊂ R−.
HenceR a
(cid:3)
It is clear that (St)t≥0 and (P +(S−t))t≥0 form continuous semigroups and these semi-
groups preserve positive functions. Moreover, by using the equality
h(P +St)h, gi = hh, (P +S−t)∗gi,
we conclude that the semigroup (P +S−t)∗ preserve positive functions. The issue is that for
St and (P +S−t)∗ the spectral mapping theorem holds and we may repeat the arguments
used in section 2. Thus we obtain the following
Lemma 6. 1) For all φ ∈ C ∞
and Re λ = a0, we have
c (R) such that supp(φ) ⊂ R+, for λ such that eλ ∈ σ(S)
φ(iλ + a) ≤ kTφk, ∀a ∈ R.
2) For all φ ∈ C ∞
Re λ = −a1, we have
c (R) such that supp(φ) ⊂ R− and for λ such that e−¯λ ∈ σ((S−1)∗) and
φ(iλ + a) ≤ kTφk, ∀a ∈ R.
WIENER-HOPF OPERATORS
11
Proof. Let A be the generator of the semi-group (St)t≥0. First we obtain using the
same arguments as in the proof of Lemma 1 that for λ such that eλ ∈ σ(S) and Re λ = a0,
there exists a sequence (fn) of functions of E and an integer k ∈ Z so that
lim
n→∞(cid:13)(cid:13)(cid:16)etA − e(λ+2kπi)t(cid:17)fn(cid:13)(cid:13) = 0, ∀t ∈ R+, kfnk = 1, ∀n ∈ N.
Then we notice that
Thus
So we have
k(P +S−t − e−(λ+2kπi)t)fnk = k(P +S−t − e−(λ+2kπi)tP +S−tSt)fnk
≤ kP +S−tke−(λ+2kπi)tk(e(λ+2kπi)t − St)fnk, ∀t ∈ R+.
lim
n→+∞
k(P +S−t − e−(λ+2kπi)t)fnk = 0, ∀t ∈ R+.
lim
n→+∞
k(cid:16)P +St − e(λ+2kπi)t(cid:17)fnk = 0, ∀t ∈ R.
Using the same arguments as in the proof of Lemma 2, we obtain
φ(iλ + a) ≤ kTφk, ∀a ∈ R, ∀φ ∈ C ∞
c (R)
and λ such that eλ ∈ σ(S) and Re λ = a0.
semi-group ((P +S−t)∗)t≥0. (cid:3)
In the same way we prove 2) using the
To establish Theorem 3, we use Lemma 6 and we follow with trivial modifications the
arguments in [5], [7], [8]. We omit the details. For the proof of Theorem 4 we repeat the
arguments in [9].
Now we pass to the proof of Theorem 5.
Proof of Theorem 5.
Let A be the commutative algebra generated by Tφ for all φ in Cc(R+) with support
in R+ and Sx, for all x ∈ R+.
Denote by bA the set of the characters on A. Let β ∈ σ(Tφ) \ {0}. Then there exists
γ ∈ bA such that β = γ(Tφ). We will prove the following equality
This result il not trivial because we cannot commute γ with the Bochner integralRR+ φ(x)Sxdx.
γ(Tφ) =ZR+
φ(x)γ(Sx)dx.
Set
θγ(x) = γ(Sx) =
γ(Sx ◦ Tφ)
γ(Tφ)
, ∀x ∈ R+.
Let ψ ∈ Cc(R+) and let supp(ψ) ⊂ K, where K is a compact subset of R+.
Suppose that (ψn)n≥0 ⊂ CK(R+) is a sequence converging to ψ uniformly on K.
For every g ∈ E, we get
kTψng − Tψgk ≤ kψn − ψk∞ sup
y∈K
kSykkgk
12
VIOLETA PETKOVA
and this implies that limn→+∞ kTψn − Tψk = 0. This shows that the linear map ψ −→ Tψ
is sequentially continuous and hence it is continuous from Cc(R+) into A. Since the map
is continuous from R+ into Cc(R+), we conclude that the map
x −→ Sx ◦ Tφ = TSx(φ)
x −→ Sx(ψ)
is continuous from R+ into A. Consequently, the function θγ is continuous on R+. Intro-
duce
η : Cc(R+) ∋ ψ −→ γ(Tψ).
The map η is a continuous linear form on Cc(R+) and applying Riesz representation
theorem, there exists some Borel measure µ (see for instance, [11]) such that
η(ψ) =ZR+
ψ(x)dµ(x), ∀ψ ∈ Cc(R+).
This implies that for all f , ψ ∈ Cc(R+), we have
(ψ ∗ f )(t)dµ(t)
ψ(x)f (t − x)dx(cid:17)dµ(t).
f (t − x)dµ(t)(cid:17)dx =ZR+
γ(Tψ ◦ Tf ) =ZR+
=ZR+(cid:16)ZR+
ψ(x)(cid:16)ZR+
γ(Tφ) =ZR+
Using the Fubini theorem, we obtain
γ(Tψ ◦ Tf ) =ZR+
and replacing f and ψ by φ, we get
ψ(x)γ(Sx ◦ Tf )dx
φ(x)θγ(x)dx, ∀φ ∈ Cc(R+).
(3.1)
Notice that θγ(x + y) = θγ(x)θγ(y), ∀x, y ∈ R+ . We will prove that θγ(x) 6= 0, ∀x ∈
n ) =
= 0 and θγ( x0
R+. Suppose θγ(x0) = 0, for x0 > 0. Then γ(Sx0) = (cid:16)γ(S x0
γ(cid:16)S x0
n(cid:17) = 0 for every n ∈ N. Since θγ is continuous on R+,
n(cid:17) = θγ(0) = 1
θγ(cid:16)x0
lim
n→+∞
)(cid:17)n
n
and we obtain a contradiction. Consequently, we have θγ(x) = γ(Sx) 6= 0, for all x ∈ R+.
Now define θγ(−x) = 1
θγ(x) , ∀x ∈ R+. It is easy to check that θγ is a morphism on R.
It is clear that θγ(x + y) = θγ(x)θγ(y), for (x, y) ∈ R+ × R+ and for (x, y) ∈ R− × R−.
Suppose that x > y > 0,
θγ(x − y) = γ(SxS−y) =
γ(SxS−ySy)
γ(Sy)
=
θγ(x)
θγ(y)
= θγ(x)θγ(−y).
WIENER-HOPF OPERATORS
13
Moreover,
θγ(y − x) =
1
θγ(x − y)
=
1
θγ(x)θγ(−y)
= θγ(y)θγ(−x).
Since θγ satisfies θγ(x + y) = θγ(x)θγ(y), for all (x, y) ∈ R2, it is well known that this
implies that there exists λ ∈ C such that θγ(x) = eλx, for all x ∈ R.
On the other hand, we have γ(Sx) ∈ σ(Sx) and γ(S1) = eλ ∈ σ(S). Thus (3.1) implies
with λ ∈ O. We conclude that
β = γ(Tφ) = φ(−iλ)
σ(Tφ) \ {0} ⊂ φ(O).
Now, suppose that supp(φ) ⊂ R−. Let B be the commutative Banach algebra generated
by Tψ for all ψ ∈ Cc(R−) and by P +S−x, for all x ∈ R+. Let κ ∈ σ(Tφ). Using the same
arguments as above, and the set of characters bB of B, we get
φ(x)eδxdx,
κ =ZR−
with −iδ ∈ V. This completes the proof of Theorem 5.
(cid:3)
4. Comments and open problems
Following the general schema of the proof of the existence of symbols for multipliers
developed in [6] for locally compact abelian groups, it is natural to conjecture that an
analog of Theorem 1 holds for general Banach spaces of functions under some hypothesis
as we have proved this for general Hilbert space of functions in [7], [9]. Using the notations
of Section 2, the crucial point is the inequality
ϕ(z) ≤ kMϕk, ∀ϕ ∈ C ∞
c (R), Im z = α0.
(4.1)
and a similar inequality for Im z = −α1. To establish (4.1), we introduced the factor
hfk, gki (see proof of Lemma 1) close to 1 and we want to estimate ϕ(z)hfk, gki. Here
the sequence fk, kfkk = 1, must be chosen so that for some integers nk ∈ Z and eλ ∈
σ(S), Re λ = α0, we have
lim
k→∞
k(St − e(λ+2πnki)t)fkk = 0, ∀t ∈ R.
(4.2)
If the spectral mapping theorem is true for the group St = eAt, we have s(A) = α0 and
(4.2) can be obtained as in Section 2. On the other hand, if s(A) < α0, we may construct
(fk) assuming that
k(A − α0 − 2πmi)−1k = +∞.
sup
m∈R
(4.3)
For Hilbert spaces (4.3) holds (see [3], [4], [1]) and author has exploited this property in
[8], [7] to complete the proof of (4.2). For semigroups in Banach spaces s(A) < α0 does
not implies in general (4.3)
14
VIOLETA PETKOVA
(see a counter-example in Chapter V in [1] and the relation between the resolvent of
A and the spectrum of St in [4]). Consequently, it is not possible to use (4.3) and to
construct a sequence fk for which (4.2) holds. Of course another proof of (4.1) could be
possible, and in Banach spaces of functions for which s(A) < α0 this is an open problem.
References
[1] K. J. Engel and R. Nagel, A short course on operator semigroups, Springer, Berlin, 2006.
[2] E. Fasangov´a and P.J. Miana, Spectral mapping inclusions for Phillips functional calculus in Banach
spaces and algebras, Studia Math. 167 (2005), 219-226.
[3] L. Gearhart, Spectral theory for contraction semigroups on Hilbert space, Trans. AMS, 236 (1978),
385-394.
[4] Y. Latushkin and S. Montgomery-Smith, Evolutionary Semigroups and Lyapunov Theorems in Ba-
nach Spaces, J. Funct. Anal. 127 (1995), no. 1, 173-197.
[5] V. Petkova, Wiener-Hopf operators on L2
[6] V. Petkova, Multipliers on Banach spaces of functions on a locally compact abelian group, J. London
ω(R+), Arch. Math.(Basel), 84 (2005), 311-324.
Math. Soc. 75 (2007), 369-390.
[7] V. Petkova, Multipliers on a Hilbert space of functions on R, Serdica Math. J. 35 (2009), 207-216.
[8] V. Petkova, Spectral theorem for multipliers on L2
[9] V. Petkova, Spectrum of the translations and Wiener-Hopf operators on L2
ω(R), Arch. Math. (Basel), 93 (2009), 357-368.
ω(R+), Preprint 2011,
(arXiv.math: 1106.4769).
[10] W. C. Ridge, Approximative point spectrum of a weighted shift, Trans. AMS, 147 (1970), 349-356.
[11] W. Rudin, Fourier analysis on groups, Interscience, New York, 1962.
[12] L. Weis, The stability of positive semigroups on Lp-spaces, Proc. Amer. Math. Soc. 123 (1995),
3089-3094.
[13] L. Weis, A short proof for the stability theorem for positive semigroups on Lp(µ), Proc. Amer. Math.
Soc. 126 (1998), 325-3256.
LMAM, Universit´e de Metz, UMR 7122,Ile du Saulcy 57045, Metz Cedex 1, France.
E-mail address: [email protected]
|
1909.13536 | 1 | 1909 | 2019-09-30T09:01:26 | Lebesgue-type inequalities in greedy approximation | [
"math.FA"
] | We present new results regarding Lebesgue-type inequalities for the Weak Chebyshev Greedy Algorithm (WCGA) in uniformly smooth Banach spaces. We improve earlier bounds in Temlyakov (Forum Math Sigma 2014), for dictionaries satisfying a new property introduced here. We apply these results to derive optimal bounds in two natural examples of sequence spaces. In particular, optimality is obtained in the case of the multivariate Haar system in Lp with 1<p<2, under the Littlewood-Paley norm. | math.FA | math |
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
Abstract. We present new results regarding Lebesgue-type inequalities for the
Weak Chebyshev Greedy Algorithm (WCGA) in uniformly smooth Banach spaces.
We improve earlier bounds in [19] for dictionaries satisfying a new property in-
troduced here. We apply these results to derive optimal bounds in two natural
examples of sequence spaces. In particular, optimality is obtained in the case of the
multivariate Haar system in Lp with 1 < p ≤ 2, under the Littlewood-Paley norm.
1. Introduction
This paper is devoted to theoretical aspects of sparse approximation. The main
motivation for the study of sparse approximation is that many real world signals
can be well approximated by sparse ones.
In a general setting we are working in
a Banach space X with a redundant system of elements (dictionary) D. There is a
solid justification of the importance of a Banach space setting in numerical analysis
in general, and in sparse approximation in particular; see, for instance, [18, Preface].
An element (function, signal) f ∈ X is said to be N-sparse with respect to D if it has
j=1 cjgj, where gj ∈ D and cj is a scalar, j = 1, . . . , N. The
set of all N-sparse elements is denoted by ΣN (D). For a given element f we introduce
the error of best N-term approximation
a representation f = PN
σN (f, D) := inf
a∈ΣN (D)
kf − ak.
In a general setting one studies algorithms (approximation methods) A = {AN (·, D)}∞
with respect to a given dictionary D. These mappings must satisfy that AN (f, D) ∈
ΣN (D), for all f ∈ X; in other words, AN provides an N-term approximant with
respect to D. It is clear that for any f ∈ X and any N we have kf − AN (f, D)k ≥
σN (f, D). We are interested in such pairs (D, A ) for which the algorithm A provides
approximation close to the best N-term approximation. We introduce the correspond-
ing definition (see [21], p.423). Let ϑ(u) be a function such that ϑ(u) ≥ 1.
N =1
DEFINITION 1.1. We say that D is a ϑ-greedy dictionary with respect to A if there
exists a constant C such that for any f ∈ X and all N ∈ N we have
(1.2)
kf − Aϑ(N )N (f, D)k ≤ CσN (f, D).
2010 Mathematics Subject Classification. 41A65, 41A25, 41A46, 46B15, 46B20.
Key words and phrases. Non-linear approximation, weak Chebyshev greedy algorithm, threshold-
ing greedy algorithm, uniformly smooth Banach space, greedy basis.
1
2
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
In the case ϑ(u) = C0 ≥ 1 is a constant we call D an almost greedy dictionary with
respect to A , and in the case C0 = 1 we call D a greedy dictionary. If D = Ψ is a
basis then in the above definitions we replace dictionary by basis. Inequalities of the
form (1.2) are called Lebesgue-type inequalities.
In the case that A = {GN (·, Ψ)}∞
N =1 is the Thresholding Greedy Algorithm (TGA),
the theory of greedy and almost greedy bases is well developed (see [18] and [20]). We
remind that if Ψ = {ψk}∞
k=1 ckψk ∈ X, then
the TGA at the Nth iteration gives an approximant
k=1 is a normalized basis in X, and f =P∞
GN (f, Ψ) :=
N
Xj=1
ckj ψkj , where
ck1 ≥ ck2 ≥ . . . .
In particular, it is known (see [18], p.17) that the univariate Haar system is a greedy
basis with respect to TGA in Lp, for all 1 < p < ∞. Also, it is known that the TGA
does not work well with respect to the trigonometric system (see, for instance, [21,
Ch. 8]). It was demonstrated in the paper [19] (see also [21], Ch.8) that the Weak
Chebyshev Greedy Algorithm (WCGA), which we define momentarily, works very well
for a special class of dictionaries, which includes the trigonometric system.
In this paper we further develop recent results from [19] concerning Lebesgue-type
inequalities for the WCGA in the context of uniformly smooth Banach spaces. We
mostly concentrate on the case when WCGA is applied with respect to a basis. We
also emphasize that, although the theory of these algorithms is typically stated for real
Banach spaces, our results are actually valid for both, real or complex Banach spaces;
in particular they can be applied to the standard (complex) trigonometric basis.
We now recall how this algorithm is defined; see [17] or [18, Chapter 6.2]. Let
(X, k · k) be a Banach space over K = R or C. Given f ∈ X \ {0}, let Ff be an
associated norming functional in X∗, that is,
(1.3)
kFf kX∗ = 1,
and Ff (f ) = kf k.
Such functionals always exist by the Hahn-Banach theorem, and are unique provided
the norm k · k is smooth. Here we shall assume the stronger property that k · k is
uniformly smooth of power type, that is, there exists q > 1 and a constant γ > 0 such
that
(1.4)
ρ(t) ≤ γ tq,
t > 0,
where ρ(t) is the associated modulus of smoothness, given by
(1.5)
2ρ(t) = sup
kf k=kgk=1(cid:16)kf + tgk + kf − tgk − 2kf k(cid:17),
t ∈ R.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
3
Let D = {ϕi}i∈I ⊂ X be a fixed dictionary in X, that is, a subset of unit vectors
with dense span. We also fix a weakness parameter τ ∈ (0, 1]. The WCGA associated
with (X, k · k, D, τ ) is defined as follows.
Weak Chebyshev Greedy Algorithm (WCGA). Given f ∈ X \ {0}, we let
f0 := f and define inductively vectors ϕi1, . . . , ϕin in D and f1, . . . , fn ∈ X by the
following procedure: at step n + 1 we pick any ϕin+1 ∈ D such that
Ffn(ϕin+1) ≥ τ sup
ϕ∈D
Ffn(ϕ),
and let Gn+1(f ) be any element in [ϕi1, . . . , ϕin+1] such that
kf − Gn+1(f )k = dist(cid:0)f, [ϕi1, . . . , ϕin+1](cid:1).
Then we set fn+1 = f − Gn+1(f ), and iterate the process (indefinitely, or until the
remainder fn+1 = 0).
It is known from [17] that kf − Gn(f )k → 0 as n → ∞, for all f ∈ X, provided
that (X, k · k) is uniformly smooth. More recently, it has been shown in [19] that
Lebesgue-type inequalities hold for this algorithm. In this paper it is convenient to
formulate them as follows: we search for functions φ : N → N and constants C > 0
such that, for all N ≥ 1 it holds
(1.6)
(cid:13)(cid:13)f − Gφ(N )(f )k ≤ C σN (f ),
∀ f ∈ X.
Note, in particular, that (1.6) implies exact recovery for each f ∈ ΣN after φ(N) steps.
Since the dictionary D is fixed we omit it from the notation, that is, ΣN = ΣN (D)
and σN (f ) = σN (f, D).
We now state a result from [19] which gives bounds for such functions φ in terms
of suitable parameters depending on (X, k · k, D); see also [21, Section 8.7]. These
parameters are quantified by the following properties.
Property A2. We say that (X, D) satisfies property A2(U), with parameter U ≥ 1,
if
(cid:13)(cid:13)Xi∈A
aiϕi(cid:13)(cid:13) ≤ U(cid:13)(cid:13)Xi∈B
aiϕi(cid:13)(cid:13),
for all finite sets A ⊂ B, and all scalars ai ∈ K. When (1.7) holds only for sets with
A ≤ K, we say that ΣK has the property A2(U).
Property A3. We say that (X, D) satisfies property A3(r, V ), with parameters r ∈
(0, 1] and V ≥ 1, if
(1.7)
(1.8)
Xi∈A
ai ≤ V Ar(cid:13)(cid:13)Xi∈B
aiϕi(cid:13)(cid:13),
for all finite sets A ⊂ B, and all scalars ai ∈ K. When (1.8) holds only for sets with
A ≤ K, we say that ΣK has the property A3(r, V ).
4
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
A brief discussion on the meaning of these properties and some examples, is given
in [19]; see also [21, section 8.7] or §2.2 below. The next result is a special case of
Theorem 2.8 in [19]; see also [21, Theorem 8.7.18].
THEOREM 1.9. Let N ≥ 1 and assume that (X, k · k, D) has the following properties
(i) ρ(t) ≤ γtq, for some γ > 0 and q ∈ (1, 2];
(ii) ΣN satisfies property A3(r, V ), for some r ∈ (0, 1] and V > 0;
(iii) ΣN satisfies property A2(U), for some U ≥ 1.
Then, there is a universal constant C ≥ 1 such that, for every τ ∈ (0, 1], the Lebesgue-
type inequality in (1.6) holds with
(1.10)
where C1(τ, γ, q) = C2(q) γq′/q τ −q′.
φ(N) = (cid:4)C1(τ, γ, q)(cid:0) log(U + 1)(cid:1) V q′
N rq′(cid:5),
REMARK 1.11. A similar result holds with U replaced by V N in (1.10), since A3(V, r)
for ΣN implies A2(U) with U = V N; see [19, Theorem 2.7].
In this paper we elaborate further on these results in the following directions. First
we replace the condition (i) on (X, k · k), by a less demanding condition involving as
well the system D.
DEFINITION 1.12. We say that (X, k · k, D) has the property D(s, c1), for some
parameters c1 > 0 and s > 1, if
(1.13)
for all f ∈ X \ {0} and all ϕ ∈ D.
dist(f, [ϕ]) ≤ kf k(cid:0)1 − c1Ff (ϕ)s(cid:1),
We shall show in Proposition 2.6 below that ρ(t) ≤ γtq implies (1.13) with s = q′
and some c1 = c1(γ, q), actually for all kϕk = 1. However, for certain (X, D) it may
happen D(s, c1) holds with a better (smaller) value of s. For example, if X = ℓp,
1 < p < ∞, and D is the canonical basis, then D(s, c1) holds with s = p′, while the
modulus of smoothness has power type q = min{p, 2}. This, and the more general
example X = ℓp(ℓq), will be discussed in §4 below.
Using this new concept we shall show the following improvement over Theorem 1.9.
THEOREM 1.14. Let N ≥ 1 and assume that (X, k·k, D) has the following properties
(i) (X, D) satisfies property D(s, c1), for some s > 1 and c1 > 0;
(ii) ΣN satisfies property A3(r, V ), for some r ∈ (0, 1] and V > 0;
(iii) ΣN satisfies property A2(U), for some U ≥ 1.
Then, there is a universal constant C ≥ 1 such that, for every τ ∈ (0, 1], the Lebesgue-
type inequality in (1.6) holds with
(1.15)
φ(N) = (cid:4)C ′
1(τ, c1, s)(cid:0) log(U + 1)(cid:1) V s N rs(cid:5),
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
5
where C ′
1(τ, c1, s) = C ′
2(s) c−1
1 τ −s.
As an application, we shall find optimal Lebesgue type inequalities for some special
pairs (X, D). We remark that no lower bounds for functions φ satisfying (1.6) seemed
to appear earlier in the literature (besides the trivial φ(N) ≥ N). Our first example
concerns the space X = ℓp(ℓq) with D the canonical basis.
THEOREM 1.16. Let X = ℓp(ℓq), with 1 < p, q < ∞, and D be the canonical basis.
Let
Then, there exists c = c(p, q, τ ) > 0 such that the WCGA satisfies (1.6) with
q′
p′o.
q′ ,
β = β(p, q) = maxn p′
φ(N) = (cid:4)c N β(cid:5).
Moreover, suppose that
(1.17)
kx − Gψ(N )xk ≤ C σN (x),
∀ N ≥ 1, x ∈ X.
Then ψ(N) ≥ c′ N β, for some c′ > 0.
Concerning this example, we shall also compare the WCGA with the usual TGA,
showing that the former performs better when p > q′; see Figure 4.1 below.
Our second application of Theorem 1.14 regards the d-variate Haar system in X =
Lp([0, 1]d). Here, however, we must renorm the space to obtain new results. We
consider the Littlewood-Paley renorming, defined as follows. Let Hd
p = Hp×. . .×Hp =
{Hn}n≥0 denote the Lp-normalized d-variate Haar system (ie, the tensor product of
(1.18)
the 1-dimensional Haar basis Hp in [0, 1]). Given f =Pn≥0 cn(f )Hn ∈ Lp, we let
1
"
"f "
"p :=(cid:13)(cid:13)S(f )(cid:13)(cid:13)Lp, where S(f ) =(cid:16)Xn≥0
cn(f )Hn2(cid:17)
2 .
This defines an equivalent norm in Lp([0, 1]d), provided 1 < p < ∞. Our next result
gives the optimal growth for the Lebesgue type functions when 1 < p ≤ 2.
THEOREM 1.19. Let X = Lp([0, 1]d), 1 < p ≤ 2, endowed with the norm "
let D = Hd
p be the Haar basis. Let
" · "
"p, and
Then, the WCGA satisfies (1.6) with
1
1
2
.
−
h(p, d) = (d − 1)(cid:12)(cid:12)(cid:12)
p(cid:12)(cid:12)(cid:12)
φ(N) = (cid:4)c (1 + log N)p′h(p,d) N(cid:5),
for some c = c(p, d, τ ) > 0. Moreover, suppose that
(1.20)
"
"f − Gψ(N )f "
"p ≤ C σN (f ),
∀ N ≥ 1, f ∈ X.
Then ψ(N) ≥ c′ (1 + log N)p′h(p,d) N, for some c′ > 0.
6
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
REMARK 1.21. In the case 2 < p < ∞, it is known that an analog of (1.20), with
(cid:129) · (cid:129)p replaced by the standard k · kLp, holds with
ψ(N) ≤ c N 2/p′
;
see [19, Example 4]. However, we do not know whether this power is optimal.
The proof of Theorem 1.19, presented in §6, is obtained as the special case q = 2
of a more general statement for a class of sequence spaces fp,q. These are defined by
norms as in (1.18), but with the square function Sf replaced by an ℓq-function Sqf ;
see §5 below for precise statements. The spaces ℓp(ℓq) and fp,q are discrete analogs
of the Besov and Triebel-Lizorkin spaces, so our results can be transferred as well to
these settings, see Remarks 6.1 and 6.2 below.
Finally, in the last part of the paper we discuss the significance of property A3 for
suitable classes of bases D = Ψ = {ψn}n≥1 in X = Lp. As a consequence of our results
we shall obtain the following.
THEOREM 1.22. Let X = Lp[0, 1], 1 < p < ∞, and let D = Ψ be a greedy basis with
respect to the TGA. Let
(1.23)
α(p) = 1
if 1 < p ≤ 2,
and α(p) = 2/p′
if 2 < p < ∞.
Then, there exists a universal constant C ≥ 1 such that
for some c = c(p, τ ) > 0.
(cid:13)(cid:13)f − G
cN α(p)f(cid:13)(cid:13)p ≤ C σN (f ),
∀ N ≥ 1, f ∈ Lp,
In particular, when 1 < p ≤ 2, greedy bases with respect to the TGA are almost
greedy with respect to the WCGA (in the sense of Definition 1.1). For p > 2, however,
we do not know whether the power α(p) = 2/p′ may be improved. These resuls are
discussed in §7, together with various additional statements under weaker assump-
tions in Ψ (such as almost-greedy, quasi-greedy, etc...), where the best exponents r in
property A3 are found for each class of bases. We remark that, in the case of X = Lp
with p > 2, we do not know whether, for some basis (or dictionary) D, the Lebesgue
inequality for the WCGA may hold with φ(N) = O(N).
2. Preliminaries
2.1. Norming functionals and distance to subspaces. Let (X, k · k) be a Banach
space over K = R or C. Let ρ(t) be the associated modulus of smoothness, given in
(1.5). Recall that k · k is called uniformly smooth if ρ(t) = o(t) as t → 0. Recall also
from (1.3) that Ff ∈ X∗ denotes the norming functional of a vector f ∈ X \ {0}. In
all our proofs below, the functional ρ(t) will only appear via the following inequality.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
7
PROPOSITION 2.1. For all f, g ∈ X with kf k = kgk = 1 it holds
(2.2)
0 ≤ kf + tgk − kf k − tℜ[Ff (g)] ≤ 2ρ(t),
t ∈ R.
In particular, if ρ(t) = o(t), then ℜ[Ff (·)] is the Fr´echet derivative of k · k at f .
Proof. The left inequality in (2.2) follows easily from
(2.3)
kf + tgk ≥ Ff (f + tg) ≥ ℜ[Ff (f + tg)] = kf k + t ℜ[Ff (g)],
t ∈ R.
Define
σ(t, f, g) := kf + tgk + kf − tgk − 2kf k ≥ 0,
∀ t ∈ R.
Then, using (2.3) with t replaced by −t we have
kf + tgk = 2kf k − kf − tgk + σ(t, f, g) ≤ 2kf k − kf k + tℜ[Ff (g)] + σ(t, f, g),
and therefore
kf + tgk − kf k − tℜ[Ff (g)] ≤ σ(t, f, g) ≤ 2ρ(t),
t ∈ R.
This establishes the upper bound in (2.2).
(cid:3)
By homogeneity one deduces
COROLLARY 2.4. For every (non-null) f, g ∈ X with kgk = 1 it holds
(2.5)
0 ≤ kf + tgk − kf k − tℜ[Ff (g)] ≤ 2kf kρ(t/kf k),
t ∈ R.
An interesting application of (2.5) gives the following result.
PROPOSITION 2.6. Suppose that 2ρ(t) ≤ γ tq, for some γ > 0 and q > 1. Then,
there exists c1 = c1(γ, q) > 0 such that
for all f, ϕ ∈ X \ {0} with kϕk = 1.
dist(f, [ϕ]) ≤ kf k(cid:0)1 − c1Ff (ϕ)q′(cid:1),
Proof. Let ν = sign Ff (ϕ) and λ > 0. Then Corollary 2.4 implies that
kf − λνϕk ≤ kf k − λℜ[Ff (νϕ)] + γkf k(λ/kf k)q
(2.7)
(2.8)
= kf k(cid:16)1 − λ
kf kFf (ϕ) + γ( λ
kf k )q(cid:17).
Now, it is easily checked that the 1-dimensional function h(λ) = γλq − λF reaches a
minimum at λ1 = [F/(γq)]q′−1, and that
h(λ) = −c1F q′
, with c1 =
min
λ>0
1
q′(γq)q′−1 .
So, minimizing over λ > 0 in (2.8) one obtains (2.7).
(cid:3)
Thus, recalling Definition 1.12, we recover the following assertion from §1.
8
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
COROLLARY 2.9. Suppose that ρ(t) ≤ γ tq, for some γ > 0 and q > 1. Then, for
every dictionary D, the pair (X, D) has the property D(s, c1) with s = q′ and some
c1 = c1(γ, q) > 0.
REMARK 2.10. Observe that ρ(t) ≤ γ tq can only hold for q ≤ 2, and hence q′ ≥ 2.
However, property D(s, c1) may hold in some cases with s close to 1. For instance, if
X = ℓp with p > 2 and D is the canonical basis, then it is not hard to verify that
(X, D) has property D(s, c1) with s = p′ and c1 = 1/p; see Proposition 4.12 below for
a more general class of such examples.
REMARK 2.11. Observe that the property D(s, c1) can also be written as
One may ask whether this property could hold for some non-uniformly smooth Banach
space (X, k · k) and some dictionary D.
,
∀ kf k = 1, ϕ ∈ D.
≤ c−1
1 (cid:16)1 − dist(f, [ϕ])(cid:17)1/s
(cid:12)(cid:12)(cid:12)
Ff (ϕ)(cid:12)(cid:12)(cid:12)
Below we shall also use the following known lemma.
LEMMA 2.12. Assume that the norm k · k is Gateaux differentiable in X. Consider
a vector f ∈ X, a finite dimensional subspace Y ⊂ X, and an element g ∈ Y such that
kf − gk = dist(f, Y) > 0. Then
(2.13)
Ff −g(h) = 0,
∀ h ∈ Y.
Proof. Let h ∈ Y and let ν = sign Ff −g(h). Then, for all λ > 0 we have
dist(f, Y) ≤ kf − g − λνhk ≤ kf − gk − λℜ[Ff −g(νh)] + o(λ)
= dist(f, Y) − λFf −g(h) + o(λ).
Thus, Ff −g(h) ≤ o(λ)/λ, which letting λ ց 0 implies that Ff −g(h) = 0.
(cid:3)
REMARK 2.14. As a consequence of Lemma 2.12, the vectors ϕi1, ϕi2, ... chosen by
the WCGA are always linearly independent.
2.2. Properties A2 and A3 for bases. Suppose that D = Ψ = {ψn}∞
n=1 is a (normal-
ized) Schauder basis in X. We denote its dual system by Ψ∗ = {ψ∗
n}n≥1. We briefly
discuss the meaning of properties A2 and A3 in this case, and relate them with more
familiar concepts from the theory of thresholding greedy algorithms.
In the first result we use the standard notation
SA(f ) :=Xn∈A
ψ∗
n(f )ψn,
f ∈ X,
for each finite set A ⊂ N. The next lemma is immediate from the definitions.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
9
LEMMA 2.15. Let Ψ be a (normalized) Schauder basis in X. Then, for each N ≥ 1,
ΣN satisfies the property A2 with parameter
(2.16)
U = kN := sup
A≤N
kSAk.
In particular, (X, Ψ) has the property A2(U) if and only if Ψ is an unconditional basis,
with suppression unconditionality constant U.
The second result relates property A3 with the right-democracy function of the dual
basis Ψ∗. We use the convenient notation
when A ⊂ N is a finite set, and ε = (εn)∞
n=1 ⊂ K is such that εn = 1 for all n.
1
∗
εA :=Xn∈A
εnψ∗
n,
LEMMA 2.17. Let Ψ be a (normalized) Schauder basis in X. Then (X, Ψ) satisfies
property A3(r, V ) if and only if
(2.18)
for all finite A ⊂ N, and all ε = (εn)∞
(cid:13)(cid:13)1
∗
εA(cid:13)(cid:13)X∗ ≤ V Ar
n=1 ⊂ K with εn = 1.
Proof. Assume first (2.18), and take sets A ⊂ B and scalars an ∈ K. If εn = sign an
then
Xn∈A
∗
an = 1
εA(cid:0)Pn∈B anψn(cid:1) ≤ k1
∗
εAkX∗(cid:13)(cid:13)Pn∈B anψn(cid:13)(cid:13) ≤ V Ar(cid:13)(cid:13)Pn∈B anψn(cid:13)(cid:13).
If f ∈ X has a finite
Conversely, assume property A3(r, V ), and pick A and ε.
expansion with respect to Ψ we have
∗
1
εA(f ) ≤Xn∈A
ψ∗
n(f ) ≤ V Ar(cid:13)(cid:13)
εAkX∗ ≤ V Ar, and (2.18) holds.
∗
Therefore, k1
∞
Xn=1
ψ∗
n(f )ψn(cid:13)(cid:13) = V Ar kf k.
(cid:3)
3. Proof of Theorem 1.14
We use the definition of WCGA given in §1. We shall follow closely the arguments
given in [19, §3]; see also [21, §8.7]. As in that paper, the proof is split into two main
steps.
3.1. The iteration inequality. In the first step we prove the following variant of
[19, Theorem 2.3]; see also [21, Theorem 8.7.12]. The proof makes use of the property
D(s, c1), and has been rewritten to simplify some steps from [19].
THEOREM 3.1. Let K ≥ 1 and assume that
(i) (X, D) has property D(s, c1), for some s > 1
(ii) ΣK has property A3(r, V ), for some r ≤ 1.
10
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
Then, for all f ∈ X, Φ ∈ ΣK and all m, M ≥ 0 it holds
(3.2)
kfm+M k ≤ e− c2 M
Krs kfmk + 2 kΦ − f k,
where c2 = c(s) c1 τ s V −s, for some c(s) > 0.
Proof. The case fm+M = 0 is trivial, so we assume kfm+M k > 0. If n ∈ [m, m + M)
then also kfnk > 0. From the definition of the algorithm and property D(s, c1),
(3.3)
kfn+1k ≤ dist(fn, [ϕin+1]) ≤ kfnk(cid:16)1 − c1Ffn(ϕin+1)s(cid:17).
When Φ =PT aiϕi ∈ ΣK, we use the notation
aiϕi,
for each A ⊂ T.
ΦA :=XA
We also denote
(3.4)
Γn := supp Gn(f ) = {i1, . . . , in},
and Tn = T \ Γn.
By Lemma 2.12 and the definition of the algorithm we have
(3.5)
By property A3,
τ Ffn(ΦA) = τ Ffn(ΦA∩Tn) ≤ XA∩Tn
ai Ffn(ϕin+1).
XA∩Tn
ai ≤ V Ar kΦ − Gnf k ≤ V Ar(cid:0)kΦ − f k + kfnk(cid:1).
Therefore, inserting these estimates into (3.3) we obtain
(3.6)
kfn+1k ≤ kfnkh1 − c1(cid:16)
τ Ffn(ΦA)
V Ar (kΦ − f k + kfnk)(cid:17)si,
for all (non-empty) sets A ⊂ T . One wants to pick the best possible set A in (3.6).
For later estimates we shall allow sets A ⊂ Tk, for a fixed k ∈ [0, m], and define
B = Tk \ A.
In this theorem it will suffice to consider k = 0 and A = T (so B = ∅), but we keep
on with the general case. Since Γk ⊂ Γn, using Lemma 2.12 we have
Ffn(ΦA) = Ffn(ΦA + ΦΓk∩A − Gnf ) = Ffn(Φ − ΦB − f + fn)
≥ kfnk − kΦ − f k − kΦBk.
Inserting the above inequalities into (3.6), and denoting c3 = c1(τ /(2V ))s, we obtain
kfn+1k ≤ kfnk(cid:16)1 − c3
≤ kfnk(cid:16)1 − c3
[kfnk − kΦ − f k − kΦBk]s
+
[2−1Ar (kΦ − f k + kfnk)]s(cid:17)
(cid:17),
[kfnk − kΦ − f k − kΦBk]s
+
[Ar kfnk]s
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
11
where the last inequality is trivial when kfnk ≤ kΦ−f k+kΦBk, and in the complement
case follows from kfnk ≥ kΦ − f k. We now use the elementary inequality
(1 − u)s
+ ≥ δs(1 − 2u),
u > 0,
which holds for some δs > 0. Applying this with u = (kΦ − f k + kΦBk)/kfnk, and
letting c2 = c3δs, we obtain
(3.7)
kfn+1k ≤ kfnk(cid:16)1 −
c2
Ars(cid:17) +
2 c2
Ars (kΦ − f k + kΦBk).
This expression can be rewritten as
Denoting β = c2/Ars, one also has
kfn+1k − 2(kΦ − f k + kΦBk) ≤ (cid:16)kfnk − 2 (kΦ − f k + kΦBk)(cid:17)(cid:16)1 −
(cid:16)kfn+1k − 2(kΦ − f k + kΦBk)(cid:17)+
Ars(cid:17).
≤ (1 − β)(cid:16)kfnk − 2 (kΦ − f k + kΦBk)(cid:17)+
after checking the trivial negative cases. One can now iterate this inequality for
m ≤ n < m + M to obtain
c2
,
kfm+M k − 2(kΦ − f k + kΦBk) ≤ (1 − β)M(cid:16)kfmk − 2 (kΦ − f k + kΦBk)(cid:17)+
≤ (1 − β)M kfmk,
and therefore
(3.8)
kfm+M k ≤ (1 − β)M kfmk + 2(kΦ − f k + kΦBk).
Since 1 − β ≤ e−β and β ≥ c2/K rs, letting B = ∅ in (3.8) gives (3.2).
(cid:3)
REMARK 3.9. The above proof actually shows the validity of the following more
general inequality
(3.10)
kfm+M k ≤ e−c2M /Ars kfmk + 2(kΦ − f k + kΦBk),
for any sets A ⊂ Tk and B = Tk \ A, and any 0 ≤ k ≤ m. This is the analog of the
corresponding expression in [19, (3.7)]; see also [21, (8.7.22)]. Observe that (3.10) also
makes sense for A = ∅, in which case the inequality follows trivially from
kfkk ≤ kf − ΦΓkk = kf − Φ + ΦBk ≤ kΦ − f k + kΦBk.
3.2. Proof of Theorem 1.14. The second part of the proof of Theorem 1.14 is
then carried exactly as in the paper [19]; see also [21, pp. 433-440]. It is based on
appropriate choices of sets Aj and Bj in inequality (3.10), at various inductive steps.
We only remark that the quantities Ajrq′ that appear in [19] can all be replaced
by Ajrs, according to inequality (3.10), as it is only via this inequality that the
uniform smoothness of the norm (or the property D(s, c1)) is used in this proof. All
the arguments in [19] can then be applied verbatim, so we do not write down the
details.
12
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
(cid:3)
4. The WCGA in the spaces ℓp(ℓq)
In this section we illustrate the performance of the WCGA in the space X = ℓp(ℓq),
when 1 < p, q < ∞. This is the set of all sequences x = (xj,k)∞
j,k=1 ⊂ K such that
(4.1)
kxk :=h
∞
Xj=1(cid:16)
∞
Xk=1
xj,kq(cid:17)
p
qi
1
p < ∞.
Throughout this section we let D = {ej,k}j,k≥1, the canonical basis in the space X.
4.1. Norming functionals in ℓp(ℓq). It will be convenient to use the following no-
tation
∆j(x) :=(cid:16)
∞
Xk=1
1
q ,
xj,kq(cid:17)
so that kxk = (Pj≥1 ∆j(x)p)1/p.
LEMMA 4.2. When x ∈ ℓp(ℓq) \ {0}, the norming functional Fx is given by
(4.3)
Fx(y) =
1
kxkp−1
∆j(x)p−q
∞
Xj=1
∞
Xk=1
xj,kq−2¯xj,kyj,k.
Proof. It is straightforward to check that Fx(x) = kxk. So we only need to verify that
Fx(y) ≤ kyk. This follows from a double use of Holder's inequality
Fx(y) ≤ kxk−(p−1)Xj
= kxk−(p−1)Xj
≤ kxk−(p−1)(cid:16)Xj
∆j(x)p−q(cid:16)Xk
∆j(x)p−1 ∆j(y)
xj,k(q−1)q′(cid:17)1/q′
∆j(y)
∆j(x)(p−1)p′(cid:17)1/p′
kyk = kyk.
(cid:3)
We shall apply the functionals Fx to the elements ej,k of the canonical basis.
COROLLARY 4.4. If x ∈ ℓp(ℓq) \ {0} and j, k ∈ N, then
(4.5)
As special cases notice that
∆j(x)p−q xj,kq−1
kxkp−1
.
(cid:12)(cid:12)Fx(ej,k)(cid:12)(cid:12) =
• if supp x ⊂ {j0} × N, then Fx(ej,k) = δj,j0 xj0,kq−1/kxkq−1.
• if supp x ∩ ({j} × N) ≤ 1, for all j, then Fx(ej,k) = xj,kp−1/kxkp−1.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
13
In particular, for vectors x in these two cases, the WCGA (with τ = 1) coincides
with the usual TGA; that is, after N steps of the algorithm it picks the N largest
coefficients, and the remainder xN is the projection of x on the subspace spanned by
the remaining basis vectors.
4.2. Performance of WCGA in ℓp(ℓq). Here we compute the relevant parameters
in (i)-(iii) of Theorems 1.9 and 1.14. Since the canonical basis is unconditional we see
that property A2 holds with U = 1. Property A3 is given by the next lemma, which
follows immediately from the inclusion X ֒→ ℓmax{p,q}.
LEMMA 4.6. For every finite set A ⊂ N × N and every x ∈ X we have
xj,k ≤ Ar kxk, with r = max{ 1
p′ , 1
q′ }.
X(j,k)∈A
In particular, property A3 holds with parameters r and V = 1.
The power of uniform smoothness for the standard norm (4.1) in ℓp(ℓq) can be
estimated as follows; see [5, Proposition 17] or [13, Theorem 3.5].
LEMMA 4.7. If 1 < p, q < ∞, then ρ(t) ≤ γ tσ with σ = min{p, q, 2}.
If at this point one applies Theorem 1.9, a simple computation easily leads to the
following.
COROLLARY 4.8. Let β = max{ p′
(4.9)
p′ }. Assume also that
q′ , q′
min{p, q} ≤ 2.
Then there are constants C1 and C2 such that
(4.10)
kx − GC1 N β (x)k ≤ C2 σN (x),
∀ N ≥ 1, x ∈ X.
REMARK 4.11. If one attempts to use Theorem 1.9 in the case min{p, q} > 2, then
σ = 2 in Lemma 4.7, and one only obtains kx − G
C1 N β1 (x)k ≤ C2 σN (x) with
β1 = max{ 2
p′ , 2
q′ } > max{ p′
q′ , q′
p′ }.
As we shall see below, this result can be improved using instead Theorem 1.14.
4.3. An improvement of the previous bound. We begin by establishing the va-
lidity of property D(s, c1).
PROPOSITION 4.12. Let 1 < p, q < ∞, and s = max{p′, q′}. Then there exists
c1 = c1(p, q) > 0 such that,
(4.13)
dist(x, [ej,k]) ≤ kxk(cid:0)1 − c1Fx(ej,k)s(cid:1),
for all (j, k) ∈ N × N and all x ∈ X = ℓp(ℓq) \ {0}. In particular, the canonical basis
in ℓp(ℓq) has the property D(s, c1) with s = max{p′, q′}.
14
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
Proof. We may assume that kxk = 1 and that j = k = 1. Using the notation
∆j = ∆j(x) = (Pk xj,kq)1/q, we can write
j +(cid:16)∆q
dist(x, [e1,1])p =
(4.14)
∆p
Xj=2
∞
1 − x1,1q(cid:17)
p
q
= 1 − ∆p
1 + ∆p
1(cid:16)1 − x1,1
∆1
p
q
.
q(cid:17)
We shall often make use of the following elementary inequality.
LEMMA 4.15. Let α > 0. Then
(cid:0)1 − u(cid:1)α ≤ 1 − min{1, α}u,
u ∈ [0, 1].
In the last expression of (4.14) we apply Lemma 4.15, with u = x1,1/∆1q and
α = p/q, so if c = min{1, p/q} we obtain
(4.16)
dist(x, [e1,1])p ≤ 1 − c ∆p
1 x1,1
∆1
q = 1 − c Fx(e1,1) x1,1,
the last equality due to (4.5). Letting s = max{p′, q′}, note that min{p − 1, q − 1} =
s′ − 1. So, using that both ∆1 and x1,1/∆1 are ≤ 1, it follows that
p−1
Fx(e1,1)s−1 = ∆
s′−1
1
Therefore, we have shown
q−1
s′ −1 ≤ x1,1.
x1,1
∆1(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
dist(x, [e1,1])p ≤ 1 − cFx(e1,1)s.
Finally, using again Lemma 4.15 we obtain (4.13) with c1 = c/p = min{ 1
p, 1
q }.
(cid:3)
Thus, Theorem 1.14 will produce the following improvement over Corollary 4.8.
COROLLARY 4.17. Let 1 < p, q < ∞, and let β = max{ p′
constants C1 and C2 such that
q′ , q′
p′ }. Then there are
(4.18)
kx − GC1N β (x)k ≤ C2σN (x),
∀ N ≥ 1, x ∈ X.
This proves the first assertion in Theorem 1.16.
4.4. Optimality of the bound in (4.18). We now prove the last assertion in Theo-
rem 1.16. The lower bounds will be obtained by testing with suitable examples. For
simplicity we assume τ = 1 and C = 1.
PROPOSITION 4.19. Let 1 < p, q < ∞. Suppose that
(4.20)
kx − Gψ(N )xk ≤ σN (x),
∀ N ≥ 1, x ∈ X.
Then, there exists cp,q > 0 such that
(4.21)
ψ(N) ≥ cp,q N β, with β = max{ p′
q′ , q′
p′ }.
Proof. Pick x = 1A + nα
1B, where
• A = m and A ∩ ({j} × N) = 1, for 1 ≤ j ≤ m, and 0 otherwise.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
15
• B = n and B ⊂ {m + 1} × N.
• α = p′( 1
q′ − 1
p′ ).
Observe that with this choice we have
(4.22)
Fx(ej,k) = kxk−(p−1),
(j, k) ∈ A ∪ B.
Indeed, if (j, k) ∈ A then xj,k = ∆j(x) = 1 and (4.22) follows from (4.5). If (j, k) ∈ B
then j = m + 1 and
∆m+1(x)p−qxm+1,kq−1 = nα(p−q) n
p−q
q nα(q−1) = 1.
In the case p′ ≥ q′, for each 1 ≤ ℓ ≤ m we shall pick Gℓx = 1Aℓ, where Aℓ−1 ⊂ Aℓ ⊂ A
with Aℓ = ℓ. Observe that this is possible because the equality in (4.22) continues
to hold when x is replaced by each remainder xℓ = x − Gℓx. So if we let m = ψ(n),
then (4.20) implies
nαk1Bk = kx − Gmxk ≤ σn(x) ≤ kx − nα
1Bk = k1Ak = m1/p.
Since k1Bk = n1/q, this implies
(4.23)
ψ(n) = m ≥ np(α+ 1
q ) = n
p′
q′ ,
where in the last equality we use the expression of α.
In the case q′ ≥ p′ we use the same example, except that we take B = 2n. For
1Bℓ where Bℓ−1 ⊂ Bℓ ⊂ B with Bℓ = ℓ.
each 1 ≤ ℓ ≤ n, this time we pick Gℓx = nα
To justify this choice at each step, observe that, if (j, k) ∈ B \ Bℓ then,
kxℓkp−1Fxℓ(ej,k) = ∆m+1(xℓ)p−q nα(q−1) = nα(p−1) (2n − ℓ)
p−q
q
= (cid:16) 2n − ℓ
n (cid:17)
p
q −1
≥ 1,
while for (j, k) ∈ A we have kxℓkp−1Fxℓ(ej,k) = 1. So, if we let n = ψ(m), from
(4.20) we have
m1/p = k1Ak ≤ kx − Gnxk ≤ σm(x) ≤ kx − 1Ak = nαk1Bk = 2
1
q nα+ 1
q ,
which implies
(4.24)
ψ(m) = n ≥ cp,q m
1
p(α+1/q) = cp,q m
q′
p′ .
Combining (4.23) and (4.24) one obtains (4.21).
PROOF of Theorem 1.16:
Combine Corollary 4.17 and Proposition 4.19.
(cid:3)
(cid:3)
16
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
4.5. WCGA vs TGA in ℓp(ℓq). Let Gn(x) denote the usual thresholding greedy
algorithm (TGA) applied to x, with regard to the canonical basis in X = ℓp(ℓq). The
following result is a consequence of [12, Theorem 5].
PROPOSITION 4.25. Let 1 < p, q < ∞. Then there are constants C1, C2 such that
(4.26)
kx − GC1N b(x)k ≤ C2 σN (x),
∀ N ≥ 1, x ∈ X,
with b = max{ p
q , q
p}.
REMARK 4.27. The exponent b = max{ p
[12, Theorem 3.2].
q , q
p } in (4.26) is best possible, in view of
So one may pose the question of when WCGA performs better than TGA, in the
sense that the power β in Corollary 4.17 is smaller than the power b in (4.26). As-
suming p > q, this is equivalent to
β =
q′
p′ ≤
p
q
= b ⇐⇒ qq′ ≤ pp′ ⇐⇒ q ≥ p′.
If p < q, then β ≤ b iff p ≥ q′. Thus, overall the WCGA performs better than the
TGA, in the sense that β ≤ b, if and only if q ≥ p′ (or q = p).
1
q
TGA N
p
q
WCGA
q′
p′
N
TGA
q
p
N
WCGA N
p′
q′
1
p
Figure 4.1. Number of iterations (modulo constants) of TGA or
WCGA to reach σN (x) in ℓp(ℓq).
5. The WCGA in the spaces fp,q
In this section we consider a class of sequence spaces fp,q related with the family of
Triebel-Lizorkin spaces. In the next section we shall specialize to the case q = 2, and
deduce the results for the Haar system in (Lp, (cid:129) · (cid:129)p) asserted in Theorem 1.19.
Throughout this section we make use of the following notation. We fix d ≥ 1 and
let Rd be the set of all dyadic rectangles I ⊂ [0, 1]d. That is, I = I1 × . . . × Id, with
each Ii = 2−ji(ki + [0, 1)), for some ji, ki ∈ N0 and 0 ≤ ki < 2ji, when i = 1, . . . , d.
When d = 1 we will write R1 = D, so that Rd = D × . . . × D.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
17
For fixed d ≥ 1 and 1 < p, q < ∞, we consider the space X = fp,q = fp,q(Rd), defined
as the set of all sequences x = (xI)I∈Rd such that
(5.1)
where 1I,p = I−1/p
use the notation
kxk = kxkfp,q :=(cid:13)(cid:13)(cid:13)(cid:0) XI∈Rd(cid:12)(cid:12)xI 1I,p(cid:12)(cid:12)
q(cid:1)1/q(cid:13)(cid:13)(cid:13)Lp([0,1]d)
,
1I is the Lp-normalized characteristic function of I. We shall often
(Sqx)(u) :=(cid:0) XI∈Rd(cid:12)(cid:12)xI 1I,p(u)(cid:12)(cid:12)
q(cid:1)1/q,
u ∈ [0, 1]d.
Throughout this section, D = {eI}I∈Rd will be the canonical basis in fp,q.
5.1. Norming funcionals in fp,q.
LEMMA 5.2. When x ∈ fp,q \ {0}, the norming functional Fx is given by
(5.3)
Fx(y) =
1
kxkp−1 Z [(Sqx)(u)]p−q(cid:16) XI∈Rd
xIq−2 xI yI 1I,p(u)q(cid:17)du .
Proof. It is easy to see that Fx(x) = kxk, so we only need to verify that Fx(y) ≤ kyk.
This follows easily from a double use of Holder's inequality
kxkp−1Fx(y) ≤ Z (Sqx)p−qhXI
q′ hXI
= Z (Sqx)p−1 (Sqy) du ≤ kxkp−1 kyk .
xI(q−1)q′
q
I,pi
1
1
yIq
1
1
q du
q
I,pi
(cid:3)
If we specialize Fx to the elements eI of the canonical basis we obtain.
COROLLARY 5.4. If x ∈ fp,q \ {0} and I ∈ Rd, then
(5.5)
xIq−1
kxkp−1
1
Iq/pZI
Sqxp−q.
(cid:12)(cid:12)Fx(eI)(cid:12)(cid:12) =
REMARK 5.6. As a special case, if supp x consists of pairwise disjoint I's, one has
So, in this case WCGA and TGA coincide.
(cid:12)(cid:12)Fx(eI)(cid:12)(cid:12) = xIp−1/kxkp−1.
5.2. Distance to subspaces. We study the property D(s, c1) in the spaces fp,q(Rd).
This will make unnecessary to compute the modulus of smoothness for k · kfp,q (which
seems not to appear in the literature).
LEMMA 5.7. Let 1 < p, q < ∞, and let s = s(p, q) = max{p′, q′}. Then, for all
x ∈ fp,q \ {0} it holds
(5.8)
s(cid:17),
dist(x, [eI]) ≤ kxk(cid:16)1 − cp,q(cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
∀ I ∈ Rd,
18
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
with cp,q = min{1/p, 1/q}. In particular, the canonical basis in fp,q has the property
D(s, c1) with s = max{p′, q′}.
Proof. We may assume that kxk = 1. Then,
dist(x, [eI])p = kx − xIeIkp =Z (Sqx)(u)ph1 −
xI 1I,p(u)q
(Sqx)(u)q ip/q
du,
where it is understood that the integral is taken over the set {u :
Setting c := min{1, p/q} and applying Lemma 4.15 we obtain
(Sqx)(u) 6= 0}.
(5.9)
dist(x, [eI])p ≤Z (Sqx)ph1 − c(cid:12)(cid:12)(cid:12)
xI 1I,p
Sqx (cid:12)(cid:12)(cid:12)
qi du = 1 − c xI(cid:12)(cid:12)Fx(eI)(cid:12)(cid:12),
using in the last step Corollary 5.4 (and kxk = 1). We now separate two cases:
a) Case 1 < p ≤ q. Since xI 1I,p(u) ≤ (Sqx)(u) and p − q ≤ 0, we have
(Sqx)(u)p−q ≤ xI 1I,p(u)p−q .
Integration over I gives
Z (Sqx)p−q
Then, using (5.5) we see that
1
q
I,p ≤ xIp−q Z 1
p
I,p = xIp−q.
Fx(eI) = xIq−1 Z (Sqx)p−q
q
I,p ≤ xIp−1.
1
Thus, we have obtained a lower bound for xI, which inserted into (5.9) gives
Finally, since c = p/q, a last use of Lemma 4.15 (with α = 1/p < 1) gives
dist(x, [eI])p ≤ 1 − c (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
p−1 = 1 − c (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
p′
.
1+ 1
b) Case q < p < ∞. Holder's inequality with exponents (p/(p − q), p/q) gives
(5.10)
Then, from (5.5),
p′
.
dist(x, [eI]) ≤ 1 − q−1 (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
p hZ 1
I,pi
I,p ≤ hZ (Sqx)pi
p−q
p
q
1
Z (Sqx)p−q
q
p = 1.
Fx(eI)
≥ Fx(eI)
xIq−1 =
R (Sqx)p−q
dist(x, [eI])p ≤ 1 − (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
q
I,p
1
1+ 1
q−1 = 1 − (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
q′
.
Inserted into (5.9) (and using c = 1) it gives
Hence, a last use of Lemma 4.15 (with α = 1/p < 1) implies
(5.11)
dist(x, [eI]) ≤ 1 − p−1 (cid:12)(cid:12)Fx(eI)(cid:12)(cid:12)
q′
.
(cid:3)
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
19
REMARK 5.12. Testing with explicit examples, it is possible to show that the power
s = max{p′, q′} in (5.8) cannot be replaced by any smaller number.
5.3. Property A3. Here, ΣN is the set of N-term combinations from the canonical
basis of X = fp,q(Rd). We define the exponent
h = h(p, q; d) = (d − 1)(cid:16)1
p
−
.
1
q(cid:17)+
LEMMA 5.13. Let 1 < p, q < ∞ and N ≥ 1. Then ΣN satisfies property A3(r, V )
with V = c(log N)h and r = 1/p′, for some constant c = c(p, q, d) > 0. That is
(5.14)
XI∈A
xI ≤ c (1 + log N)(d−1)( 1
p − 1
q )+ A1/p′
xI eIk.
kXI∈B
for all sets A ⊂ B ⊂ Rd with A ≤ N and B finite.
Proof. In view of Lemma 2.17 it suffices to estimate
Since (fp,q)∗ = fp′,q′, then (5.14) reduces to show that
(5.15)
≤ c (1 + log N)(d−1)( 1
q − 1
p )+ N
1
p .
eI(cid:13)(cid:13)(fp,q)∗ .
sup
A≤N(cid:13)(cid:13)Xn∈A
eI(cid:13)(cid:13)fp,q
sup
A≤N(cid:13)(cid:13)Xn∈A
When q = 2 this was proved in [23, Proposition 10], see also [15, Theorem A]. A small
modification of those proofs gives also the case q 6= 2. For completeness, we sketch
the arguments in Appendix 1 below.
(cid:3)
5.4. The WCGA in fp,q. A direct application of Theorem 1.14, with the parameters
obtained in Lemmas 5.7 and 5.13, gives the following.
THEOREM 5.16. Let X = fp,q(Rd), 1 < p, q < ∞, and let D be the canonical basis.
Define
h(p, q; d) = (d − 1)(cid:16)1
p
−
,
1
q(cid:17)+
Then, the WCGA satisfies (1.6) with
and α(p, q) =( 1
q′/p′
if p ≤ q
q ≤ p.
if
(5.17)
for some c = c(p, q, d, τ ) > 0.
φ(N) = (cid:4)c (1 + log N)p′h(p,q;d) N α(p,q)(cid:5),
REMARK 5.18. As in Remark 4.11 above, if in the case min{p, q} > 2 we had used
Theorem 1.9 rather than Theorem 1.14, then one would have obtained a power 2/p′,
which is worse than α(p, q) (assuming that the estimate for the modulus of smoothness
in Lemma 4.7 also applies when X = fp,q).
20
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
5.5. A lower bound for the WCGA in fp,q when p ≤ q. We show in this sub-
section that, when p ≤ q, the function φ(N) = ⌊c (1 + log N)p′(d−1)( 1
q ) N⌋ in (5.17)
cannot be replaced by a slower growing one.
p − 1
THEOREM 5.19. In the conditions of Theorem 5.16, suppose that p ≤ q and that
(5.20)
∀ N ≥ 1, x ∈ fp,q(Rd).
Then ψ(N) ≥ c′ (1 + log N)p′(d−1)( 1
q ) N, for some c′ > 0.
(cid:13)(cid:13)x − Gψ(N )x(cid:13)(cid:13) ≤ C σN (x),
p − 1
Proof. Fix n ∈ N and ~m = (m1, . . . , md) ∈ Nd, and let m := m1 + . . . + md. Consider
the following sets
An =nI ∈ Rd : I ⊂ [0, 1/2]d
and I = 2−no
and
B ~m =nI1 × . . . × Id ∈ Rd : Ii ⊂ [1/2, 1] and Ii = 2−mi, i = 1, . . . , do.
Observe that
(5.21)
An ≈ nd−1 2n
≈ n
d−1
q 2n/p,
u ∈ (0, 1/2)d,
and (cid:16) XI∈An
1
q
I,p(u)(cid:17)1/q
I,p(u)(cid:17)1/q
q
1
B ~m = 2m−1
and (cid:16) XI∈B ~m
while
(5.22)
≈ 2m/p,
u ∈ (1/2, 1)d,
We build a vector
x = xAn + xB ~m := a XI∈An
eI + b XI∈B ~m
eI,
for suitable a, b > 0 to be chosen later. Using the expression for the fp,q-norming
functionals in (5.5), and the estimates in (5.21), (5.22), one sees that
(5.23) kxkp−1(cid:12)(cid:12)Fx(eI)(cid:12)(cid:12) = xIq−1 ZI
We now pick b such that
(Sqx)p−q
1
q
I,p ≈ ( ap−1n
bp−1,
(d−1)(p−q)
q
,
if
if
I ∈ An
I ∈ B ~m.
(5.24)
bp−1 ≈ ap−1 n(d−1)(p−q)/q.
The constants can be adjusted so that the WCGA always chooses
Gk(x) = b
k
Xj=1
eIj , when k ≤ B ~m,
for some enumeration I1, I2, . . . of B ~m. At this point one should notice that (5.23)
continues to hold with x replaced by each remainder x − Gk(x), as long as k < B ~m
(so one can still pick elements I from B ~m \ {I1, . . . , Ik}).
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
21
Finally, given N ≫ 1, we select the largest n and the smallest m such that
(5.25)
An ≤ N and B ~m ≥ ψ(N).
Then, N ≈ nd−12n and ψ(N) ≈ 2m. The assumption in (5.20) gives
kxAnk ≤ kx − Gψ(N )(x)k ≤ C σN (x) ≤ C kxB ~mk.
By (5.21) and (5.22) this implies
a n
d−1
q 2n/p ≈ kxAnk ≤ C kxB ~mk ≈ b 2m/p ≈ b ψ(N)1/p.
Recalling (5.24) this amounts to
ψ(N) & (a/b)p n
(d−1)p
q
2n ≈ n
(d−1)p′
q′
2n ≈ [log N](d−1)( p′
q′ −1) N,
using in the last step the relation N ≈ nd−12n. This proves the theorem since the
exponent in the power of the log can also be written as p′/q′ − 1 = p′( 1
p − 1
q ).
(cid:3)
6. The WCGA for the d-variate Haar basis in Lp
We recall the definition of the Haar system. Consider the 1-dimensional functions
h0 = 1[0,1) and h = 1[0,1/2) − 1[1/2,1). For every dyadic interval I = 2−j(k + [0, 1)) in
[0, 1], define
hI,p(x) = 2j/ph(2jx − k),
and h0,p(x) = h0(x).
Then, if 1 < p < ∞, the system Hp = {h0, hI,p}I∈D is a (normalized) unconditional
basis of Lp[0, 1]. With a slight abuse of notation we write Hp = {hI,p}I∈ ¯D with
¯D = D ∪ {0}. The d-variate Haar system is then
Hd
p = Hp × . . . × Hp = {HI,p}I∈ ¯Rd,
with ¯Rd = ¯D × . . . × ¯D, that is,
HI,p(x1, . . . , xd) =
d
Yi=1
hIi,p(xi),
if
I = I1 × . . . × Id ∈ ¯Rd.
The system D = Hd
p is an unconditional basis of Lp([0, 1]d), and moreover, for every
function f =PI cI(f )HI,p ∈ Lp, the expression
"
"f "
"p :=(cid:13)(cid:13)S(f )(cid:13)(cid:13)Lp, where S(f ) =(cid:16) XI∈ ¯Rd
1
2 ,
cI(f )HI,p2(cid:17)
defines an equivalent norm in Lp([0, 1]d), provided 1 < p < ∞. Here, the coefficients
are given by
cI(f ) := h f , HI,p′ i,
I ∈ ¯Rd.
22
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
Clearly, (Lp, (cid:129) · (cid:129)p) is related with the spaces fp,2 from the last section. More precisely,
the coefficient map defines an isometric isomorphism
(cid:0)Lp([0, 1]d), (cid:129)·(cid:129)p(cid:1) −→ fp,2( ¯Rd)
f 7−→ (cid:0)cI(f )(cid:1)I∈ ¯Rd
with the latter space defined as in §5, with the only minor change that Rd is replaced
by ¯Rd (so a few additional terms appear in the norm (5.1)). This change does not
affect the proofs, and all the results in §5 continue to hold for the spaces fp,q( ¯Rd). In
particular, Theorem 1.19 from the introduction becomes a corollary of Theorems 5.16
and 5.19, in the special case q = 2.
REMARK 6.1. One can use the previous isometric isomorphism to transfer the results
for X = fp,q in §5 into results for (univariate) wavelet bases Ψ = {ψI} in the class of
Triebel-Lizorkin spaces F r
p,q. Indeed, one can define equivalent norms in the latter of
the form
kf kF r
p,q :=(cid:13)(cid:13)(cid:13)(cid:0)XI
cI(f )1I,pq(cid:1)
,
1
q(cid:13)(cid:13)(cid:13)Lp
where cI(f ) = h f , ψ∗
I i = I−r h f , ψI,p′ i, with ψI,p′ = ψI /kψIkp′. Just remark that
in §5 one should modify the index set D acccording to the underlying space (say, over
R rather than [0, 1]). In the d-variate case, the use of tensor wavelet bases, gives rise
to an isometry between fp,q(Rd) and Triebel-Lizorkin classes with dominating mixed
smoothness (sometimes denoted Sr
p,qF ; see e.g. [22, Th. 1.12]).
REMARK 6.2. In a similar fashion, one may transfer the results for X = ℓp(ℓq) in
§4 into results for wavelet bases in the class of Besov spaces Br
q,pB in the
d-variate case), using the equivalent norms
q,p (or Sr
kf kBr
q,p :=hXj (cid:16)Xk
cI(j,k)(f )q(cid:17)
,
p
qi1/p
with I(j, k) = 2−j(k + [0, 1)), j, k ∈ Z.
7. Property A3 for classes of bases in Lp
In this section we find optimal exponents for the property A3 in certain classes
of bases of Lp, 1 < p < ∞. These classes (greedy, almost greedy, quasi-greedy,...)
behave well with respect to the TGA, and we study the performance with respect to
the WCGA. Results of this section complement the corresponding results from [19]
(see also [21], Section 8.7.4).
Below we shall use the following known lemma, which is valid for Lp over an arbi-
K the subspace of K-valued functions
trary (σ-finite) measure space. We denote by Lp
in Lp.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
23
LEMMA 7.1. Let 1 < p < ∞, and let X = Lp = Lp
such that
C. Then, there exists γ = γ(p) > 0
ρLp(t) :=
sup
kf kp=kgkp=1(cid:16)kf + tgkp + kf − tgkp − 2kf kp(cid:17)/2 ≤ γ tq,
with q = min{p, 2}.
∀ t > 0,
Proof. This result is well-known when X = Lp
that in the complex-valued case one has
R; see e.g. [14, Vol II, p. 63]. We remark
ρLp
C
(t) = ρLp
R
(t) ≤ γ tq,
∀ t > 0 ,
since Lp
C can be isometrically embedded into some Lp
R; see [3, Example 4.6].
(cid:3)
7.1. Greedy bases. The univariate Haar system Hp is a greedy basis of Lp[0, 1],
1 < p < ∞, and it was shown in [19, Example 4] that a Lebesgue inequality holds
with φ(N) = O(N) if 1 < p ≤ 2, and φ(N) = O(N 2/p′) if 2 ≤ p < ∞. Here we
generalize this result to any greedy basis Ψ = {ψn}n≥1 of Lp, as stated in Theorem
1.22 from §1.
We begin with a computation of the parameters for the A3 property. We assume that
Ψ is normalized in Lp, and recall that a greedy basis is unconditional and democratic.
LEMMA 7.2. If Ψ is a greedy basis in Lp, then A3(r, V ) holds with r = 1/p′ and some
constant V > 0.
Proof. Let Ψ∗ = {ψ∗
suffices to show that
n}∞
n=1 be the dual system to Ψ in Lp′. In view of Lemma 2.17 it
(7.3)
(cid:13)(cid:13)Xn∈Λ
ψ∗
n(cid:13)(cid:13)Lp′ ≤ C Λ1/p′
,
for all finite sets Λ ⊂ N. We shall use a result from functional analysis [11, Theorem
4b] which says that for any (seminormalized) unconditional basis B = {bk}k≥1 of Lq,
1 < q < ∞, there is a subsequence kj, j = 1, 2, . . . , such that
In particular, if B is unconditional and democratic, then necessarily
(7.4)
bkkq ≍ Λ1/q,
∞
Xj=1
1
q .
αkj q(cid:17)
∞
Xj=1
(cid:13)(cid:13)
αkj bkj(cid:13)(cid:13)Lq ≍(cid:16)
kXk∈Λ
with the constants of equivalency depending at most on B and q. Now, if Ψ is a
greedy basis in Lp, 1 < p < ∞, then it was shown in [2, Theorem 5.1] that Ψ∗ is also
a greedy basis in the dual space (Lp)∗ = Lp′. Therefore (7.3) is just a consequence of
(7.4) with q = p′.
(cid:3)
24
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
PROOF of Theorem 1.22:
The proof is now a direct application of Theorem 1.9.
Indeed, A2(U) holds because Ψ is unconditional, while A3(r = 1/p′, V ) was proved
in the previous lemma. By Lemma 7.1 we also have ρ(t) ≤ γtq with q = min{p, 2}.
Therefore, Theorem 1.9 implies
(7.5)
kf − G
C(t,p,Ψ)mα(p)(f )kp ≤ Cσm(f, Ψ)p,
with α(p) = rq′, which agrees with (1.23).
(cid:3)
7.2. Tensor products of greedy bases. We now extend the first part of Theorem
1.19 from the multivariate Haar system to an arbitrary tensor product of univariate
greedy bases. Let Ψ be a normalized basis in Lp([0, 1)). In the space Lp([0, 1)d) we
define the system
Ψd := Ψ × · · · × Ψ =nψn(x) = ψn1(x1) · · · ψnd(xd) : n = (n1, . . . , nd) ∈ Ndo.
Clearly, if Ψ is unconditional, so is Ψd. Democracy, however, does not transfer, but
one has the following result from [15] (see also [21, Ch. 8]).
PROPOSITION 7.6. Let 1 < q < ∞ and let Ψ be a greedy basis for Lq[0, 1]. Then
for any Λ with Λ = m ≥ 2, we have: for 2 ≤ q < ∞
C 1
q,d m1/q min
n∈Λ
and for 1 < q ≤ 2
cn ≤ kXn∈Λ
cnψnkq ≤ C 2
q,d (log m)h(q,d) m1/q max
n∈Λ
cn,
C 3
q,dm1/q(log m)−h(q,d) min
n∈Λ
where h(q, d) := (d − 1)1/2 − 1/q.
cn ≤ kXn∈Λ
cnψnkq ≤ C 4
q,dm1/q max
n∈Λ
cn
Note that h(p, d) = h(p′, d). We now derive the property A3 for Ψd.
LEMMA 7.7. Consider the system Ψd in Lp([0, 1]d) defined above, where Ψ is a uni-
variate greedy basis in Lp[0, 1]. Then, for each N ≥ 1, the set ΣN (Ψd) satisfies property
A3(r, V ) with r = 1/p′ and
V = C(p, d) (1 + log N)(d−1)( 1
p − 1
2 )+ .
Proof. By Lemma 2.17, it suffices to compute
sup
Λ≤N(cid:13)(cid:13)Xn∈Λ
ψ∗
n(cid:13)(cid:13)(cid:13)Lp′ ,
n = ψ∗
n1 ⊗ . . . ⊗ ψ∗
where ψ∗
nd, n ∈ Nd, are the elements of the dual system (Ψd)∗. Since
Ψ∗ is a univariate greedy basis (by [2, Thm 5.1]), we can apply Proposition 7.6 with
q = p′, and obtain
sup
Λ≤N(cid:13)(cid:13)Xn∈Λ
ψ∗
n(cid:13)(cid:13)(cid:13)Lp′ ≤ Cp,d (1 + log N)(d−1)( 1
p − 1
2 )+ N 1/p′
.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
25
(cid:3)
THEOREM 7.8. Let 1 < p < ∞. Consider the system Ψd in Lp([0, 1]d) defined above,
where Ψ is a univariate greedy basis in Lp[0, 1]. Then, the WCGA in Lp([0, 1]d) applied
to Ψd satisfies
(7.9)
with φ(m) = C(t, p, d, Ψ) (1 + log m)(d−1)p′( 1
p − 1
kf − Gφ(m)(f )kp ≤ Cσm(f )p,
2 )+ mα(p), and with α(p) as in (1.23).
Proof. This is again a direct application of Theorem 1.9 and the previous lemma. (cid:3)
7.3. Nikol'skii ℓ1X property. When Ψ = {ψk}∞
k=1 is a basis in Lp := Lp([0, 1)d),
1 < p < ∞, then the A3 property essentially amounts to compute the following
operator norms
kSΛkp,A := kSΛkLp→A := sup
f 6=0
kSΛ(f )kA/kf kp ,
where we denote, for each finite set Λ ⊂ N,
SΛ(f ) := SΛ(f, Ψ) :=Xk∈Λ
ck(f )ψk,
kSΛ(f )kA :=Xk∈Λ
ck(f ),
and as usual ck = h f , ψ∗
k i. Then A3(r, V ) holds if and only if
kSΛkp,A ≤ V Λr.
The notation refers to A = {f ∈ L1 : Pk∈Λ ck(f ) < ∞}, which for the trigono-
metric system is the usual Wiener algebra.
7.4. General lower bounds. We say that a Banach space X has type t if there exists
a universal constant C such that for fk ∈ X and all n ≥ 1,
(7.10)
Aveεk=±1k
εkfkkt!1/t
n
Xk=1
≤ C n
Xk=1
kfkkt!1/t
.
So, given Λ ⊂ N, if we set fk = ψk, k ∈ Λ, then there exists a function f =Pk∈Λ εkψk
with kf kp ≤ CΛ1/t. Thus,
kSΛf kA = Λ = Λ1/t′
Λ1/t ≥ C −1 Λ1/t′
kf kp.
For X = Lp it is well-known that t = min{p, 2} if 1 < p < ∞. Therefore, the following
lower bounds always hold
(7.11)
kSΛkp,A & Λ1/p′
if 1 < p ≤ 2,
and kSΛkp,A & Λ1/2 if p > 2.
As we shall see below, there are bases Ψ for which the symbols "&" in (7.11) can be
replaced by "≈" for all Λ. Notice, from Theorem 1.9, that the WCGA will be most
effective in those situations, since the exponents will satisfy rq′ = 1, and therefore we
have φ(N) = c(t, p) (log U) N.
26
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
7.5. Schauder bases. Let Ψ be a (normalized) Schauder basis in Lp, 1 < p < ∞.
Then, the following general upper bound holds.
PROPOSITION 7.12. There exists ε = ε(p, Ψ) > 0 and c > 0 such that
kSΛkp,A ≤ c Λ1−ε,
∀ Λ ⊂ N.
Proof. Indeed, by a well-known theorem of Gurari-Gurari [10], there exists s < ∞ and
c > 0 such that
k
∞
Xn=1
anψnkp ≥
1
c (cid:16)
∞
Xn=1
ans(cid:17)1/s
,
for all finite sequences of scalars an. So if f =Pn anψn we have
∞
kSΛf kA =Xn∈Λ
an ≤ Λ1/s′(cid:16)
ans(cid:17)1/s
Xn=1
≤ cΛ1− 1
s kf kp.
Thus, the result holds with ε = 1/s.
(cid:3)
REMARK 7.13. When no further assumption is made on the basis, the bound in the
proposition cannot be improved, even if p = 2. Indeed, for every α < 1, consider the
basis Ψ in L2 constructed in [8, Proposition 3.10]. This basis satisfies the following
property: if f =P2N
n=1 ψn and Λ = {1, 3, . . . , 2N − 1}, then
kSΛf k2 ≥ cN α kf k2.
Since kSΛf k2 ≤ kSΛf kA (by the triangular inequality), we conclude that
kSΛk2,A ≥ c Λα,
∀ Λ ⊂ 2N − 1.
7.6. Uniformly bounded orthogonal bases. Suppose that Ψ = {ψk}k≥1 is an
orthonormal system in L2([0, 1]d) such that cΨ := supk≥1 kψkkL∞ < ∞. These systems
satisfy c−1
Ψ ≤ kψkkp ≤ cΨ, for all k ≥ 1 and all 1 ≤ p ≤ ∞.
7.6.1. Case 2 ≤ p < ∞.
PROPOSITION 7.14. If Ψ is a uniformly bounded orthogonal basis of Lp[0, 1], 2 ≤
p < ∞, then
kSΛ(·, Ψ)kLp→A ≍ Λ1/2,
∀ Λ.
Proof. The lower bound was shown in (7.11). The upper bound follows from an
elementary argument, as in [19, Example 1].
(cid:3)
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
27
7.6.2. Case 1 < p ≤ 2. In this case the following holds
(7.15)
c1 Λ1/2 ≤ kSΛkp,A ≤ c2 Λ1/p,
∀ Λ,
and some constants ci(p, Ψ) > 0. The lower bound is due to kSΛkp,A ≥ kSΛk2,A =
Λ1/2. The upper bound follows from the argument in [19, Example 1q]; see also [21,
Ch. 8]. The example of the univariate trigonometric system shows that the bounds
in (7.15) cannot be improved for this class of bases.
REMARK 7.16. It is known that uniformly bounded orthogonal bases cannot be
unconditional in Lp, for any p 6= 2; see [6]. Thus, when p > 2 we cannot hope to
obtain φ(N) = O(N) from Theorem 1.9 using these examples. However, it is possible
to construct such bases with φ(N) = O(N log log N); see Proposition 7.25 below.
7.7. Unconditional bases. Recall that a Banach space X has cotype q if there exists
c > 0 such that
(7.17)
Aveεk=±1 hk
N
Xn=1
εnfnki ≥ ch
N
Xn=1
1
q .
kfnkqi
If Ψ is an unconditional basis, and we let fn = anψn, then we have
(7.18)
k
N
Xn=1
anψnk ≥ c2h
N
Xn=1
1
q ,
anqi
with a constant c2 = c2(Ψ, c) > 0. Thus, if f =Pn anψn, for a finite sequence an, we
have
(7.19)
kSΛf kA =Xn∈Λ
an ≤ Λ1/q′hXn
anqi
1
q ≤ c−1
2
Λ1/q′
kf k.
Now, specializing to the case when X = Lp, we know that q = max{2, p}. Therefore,
from (7.19) and (7.11) we obtain the following
Case 1 < p ≤ 2.
(7.20)
c1 Λ1/p′
≤ kSΛkp,A ≤ c2 Λ1/2,
∀ Λ,
Case 2 ≤ p < ∞.
(7.21)
c1 Λ1/2 ≤ kSΛkp,A ≤ c2 Λ1/p′
,
∀ Λ,
These inequalities cannot be improved for the whole class of unconditional bases.
Indeed, when Ψ is the univariate Haar basis we have, for all 1 < p < ∞,
(7.22)
kSΛkp,A ≍ Λ1/p′
,
∀ Λ.
On the other hand, it is known that Lp ≈ Lp ⊕ ℓ2, when 1 < p < ∞. Consider
n=1 in Lp ⊕ ℓ2, where {ϕ2n−1} is the Haar system in Lp, and {ϕ2n} is
the basis {ϕn}∞
28
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
the canonical basis in ℓ2. Then, the above isomorphim produces a (seminormalized)
unconditional basis Ψ in Lp, with the property
(7.23)
kSΛkp,A ≈ Λ1/2,
∀ Λ ⊂ 2N.
7.8. Quasi-greedy bases. If Ψ is a (normalized) quasi-greedy basis in a Banach
space X of cotype q, then letting fn = ψn in (7.17) we obtain
ψnk ≥ c1 Λ1/q,
kXn∈Λ
for some c1(Ψ) > 0. This is weaker than (7.18), but using [9, Lemma 4.1] we have
k(an)kℓq,∞ ≤ C kXn
If f =Pn anψn, for a finite sequence (an), and (a∗
ment, then
anψnk.
j ) denotes its decreasing rearrange-
kSΛf kA =Xn∈Λ
(7.24)
an ≤
a∗
j ≤ k(an)kℓq,∞
Λ
Xj=1
≤ C ′ Λ1/q′
kf k.
j− 1
q
Λ
Xj=1
When X = Lp, using the corresponding value for the cotype q, one obtains again the
bounds in (7.20) and (7.21).
From the examples in (7.22) and (7.23) it is clear that these bounds cannot be
improved in the class of all quasi-greedy bases. Notice further that within this class
one can replace the example in (7.23) by the following construction.
PROPOSITION 7.25. There exists a uniformly bounded orthonormal system Ψ =
{ψk}∞
k=1, consisting of trigonometric polynomials, which is a quasi-greedy basis in
Lp[0, 1], for all 1 < p < ∞. Moreover, Ψ is democratic with kPk∈Λ ψkkp ≍ Λ1/2, for
all finite Λ ⊂ N.
For details on this construction, we refer to [16] and [4] (see also [20, Ch 3] and [21,
Ch 3]). The example Ψ in Proposition 7.25 produces the bounds
(7.26)
kSΛkp,A ≈ Λ1/2,
∀ Λ ⊂ N.
Also, quasi-greedness implies that ΣN has property A2 with U . log N; see [1, Lemma
8.2]. In particular, when p > 2, the Lebesgue inequality for the WCGA which one
obtains from Theorem 1.9 holds with φ(N) = O(N log log N); see [19] or [21, p. 445].
So far, we do not know any example of a basis (or even a dictionary) in Lp, p > 2,
with φ(N) = O(N).
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
29
7.9. Almost greedy bases. This class is a subset of the previous case, and therefore
kSΛkp,A satisfies the same bounds (7.20) and (7.21). The examples in (7.22) and
(7.26) show that the assumption that Ψ is additionally democratic does not imply
any improvement in those bounds.
7.10. Greedy bases. This class is also a subset of the previous three cases, but this
time we obtain the following improvement. The proof follows easily from the same
arguments we already gave in (7.4).
PROPOSITION 7.27. Let Ψ be a greedy basis of Lp, 1 < p < ∞. Then
(7.28)
kSΛkp,A ≍ Λ1/p′
,
∀ Λ ⊂ N.
8. Appendix 1
In this section we prove the following result, which was asserted in (5.15).
PROPOSITION 8.1. Let X = fp,q(Rd), 1 < p, q < ∞, with norm defined as in (5.1).
Then
sup
A≤N(cid:13)(cid:13)Xn∈A
eI(cid:13)(cid:13)fp,q(Rd) ≤ c N
1
p (1 + log N)(d−1)( 1
q − 1
p )+ .
For q = 2 this was proved in [23, 15]. Here we adapt the arguments to the case
of a general q. We split the proof into several lemmas, which have an independent
interest. Recall that, for a sequence x = (xλ)λ∈Λ, over a set of indices Λ, we define its
support as supp x = {λ ∈ Λ : xλ 6= 0}.
LEMMA 8.3. Let p ≥ q. If x0, x1, x2, . . . have disjoint supports, then
(8.2)
(8.4)
Proof. The support condition implies that
Then, the definition of norm in (5.1) and Minkowski's inequality (since p/q ≥ 1) give
.
q
fp,q(cid:17)1/q
(cid:13)(cid:13)(cid:13)Pn≥0 xn(cid:13)(cid:13)(cid:13)fp,q
≤ (cid:16)Pn≥0(cid:13)(cid:13)xn(cid:13)(cid:13)
Sq(Pn≥0 xn)q =Pn≥0 Sq(xn)q.
= (cid:13)(cid:13)(cid:13)
Sq(Pn≥0 xn)(cid:13)(cid:13)(cid:13)Lp
≤ (cid:16)Pn≥0(cid:13)(cid:13)Sq(xn)q(cid:13)(cid:13)Lp/q(cid:17)1/q
1/q
Lp/q
= (cid:13)(cid:13)(cid:13)Pn≥0 Sq(xn)q(cid:13)(cid:13)(cid:13)
= (cid:16)Pn≥0(cid:13)(cid:13)xn(cid:13)(cid:13)
q
fp,q(cid:17)1/q
.
(cid:3)
(cid:13)(cid:13)(cid:13)Pn≥0 xn(cid:13)(cid:13)(cid:13)fp,q
In the next two lemmas we shall assume that d = 1, so that R1 = D. We denote
by ℓp,q = ℓp,q(D) the discrete Lorentz space indexed by D. We shall use the follow-
j = xIj )j≥1 is its decreasing
ing (equivalent) quasi-norm:
if x = (xI)I∈D and if (x∗
30
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
rearrangement, then
(cid:13)(cid:13)x(cid:13)(cid:13)ℓp,q := (cid:16)Xj≥0
2j/px∗
2j q(cid:17)1/q
.
LEMMA 8.5. Let d = 1 and p ≥ q. Then, ℓp,q(D) ֒→ fp,q(D) ֒→ ℓp(D), that is
(8.6)
kxkℓp ≤ kxkfp,q ≤ c kxkℓp,q.
Proof. The left inequality is trivial, since q ≤ p implies fp,q ֒→ fp,p = ℓp(D). We prove
the right inequality. Let x = (xI )I∈D and let xI1 ≥ xI2 ≥ . . . be its decreasing
rearrangement. Define
xn = X2n≤j<2n+1
xIj eIj ,
so that x = Pn≥0 xn and the vectors xn have disjoint supports. Then, (8.4) holds.
Moreover, for each n ≥ 0
kxnkfp,q ≤ x∗
2n(cid:13)(cid:13)P2n≤j<2n+1 eIj(cid:13)(cid:13)fp,q
≤ cp,q x∗
2n 2n/p,
where the last inequality is due to the p-democracy of the canonical basis in fp,q(D)
when d = 1; see [7, Prop. 3.2]. Thus,
kxkfp,q ≤ (cid:16)Xn≥0
kxnkq
fp,q(cid:17)1/q
. (cid:16)Xn≥0
2n/px∗
2nq(cid:17)1/q
= kxkℓp,q.
(cid:3)
LEMMA 8.7. Let d = 1 and 1 < q < p < ∞. Then, for all A ⊂ D with A ≤ N, it
holds
(cid:13)(cid:13)(cid:13)XI∈A
xIeI(cid:13)(cid:13)(cid:13)fp,q(D)
≤ c′(cid:2)1 + log N(cid:3)
1
q − 1
p (cid:16)XI∈A
xIp(cid:17)1/p
.
Proof. By (8.6) and Holder's inequality we have
(8.8)
kxkfp,q . kxkℓp,q ≤ kxk1−θ
ℓp,1 kxkθ
ℓp,p,
with θ = (1 − 1
q )/(1 − 1
p ). Another use of Holder's inequality gives
kxkℓp,1 =
⌊log2 A⌋
Xj=0
2j/px∗
2j ≤ (1 + log2 A)1/p′
kxkℓp,p.
Inserting this into (8.8) and using that kxkℓp,p ≈ kxkℓp the result follows.
(cid:3)
PROOF of Proposition 8.1:
and
The proof for p ≤ q is trivial, since ℓp = fp,p ֒→ fp,q,
(cid:13)(cid:13)Xn∈A
eI(cid:13)(cid:13)fp,q
≤(cid:13)(cid:13)Xn∈A
eI(cid:13)(cid:13)ℓp = A1/p.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
31
So from now on we consider p > q. The result is known for d = 1 by the p-democracy of
fp,q(D); see [7, Prop. 3.2]. So we proceed by induction, and will prove (8.2) assuming
its validity with d replaced by d − 1. Let A ⊂ Rd with A ≤ N, and define
A1 = {I1 ∈ D : ∃ I ′ ∈ Rd−1 s.t. I1 × I ′ ∈ A}
and for each I1 ∈ A1,
A′(I1) = {I ′ ∈ Rd−1 : I1 × I ′ ∈ A}.
Then
(cid:13)(cid:13)XI∈A
eIkp
fp,q(Rd) = Zu′Zu1(cid:12)(cid:12)(cid:12) XI1∈A1(cid:0) XI ′∈A′(I1)
(1 + log N)( 1
q − 1
. Zu′
p/q
1
q
q
I ′,p(cid:1)1
I1,p(cid:12)(cid:12)(cid:12)
p )p XI1∈A1(cid:12)(cid:12)(cid:12) XI ′∈A′(I1)
du1 du′
p/q du′,
1
q
I ′,p(cid:12)(cid:12)
using Lemma 8.7 in the inner integral, since A1 ≤ N. The last displayed expression
equals
(1 + log N)( 1
q − 1
p/q du′
1
q
. (1 + log N)( 1
p )p XI1∈A1Zu′(cid:12)(cid:12)(cid:12) XI ′∈A′(I1)
I ′,p(cid:12)(cid:12)
p )p XI1∈A1
= h(1 + log N)(d−1)( 1
q − 1
q − 1
p )A1/pip
,
(1 + log N)(d−2)( 1
q − 1
p )pA′(I1)
using in the middle step the induction hypothesis (since A′(I1) ≤ N).
(cid:3)
Acknowledgments
The authors would like to thank the Isaac Newton Institute for Mathematical Sci-
ences, Cambridge, for support and hospitality during the program Approximation,
Sampling and Compression in Data Science where some work on this paper was un-
dertaken; this work was supported by EPSRC grant no EP/K032208/1.
G.G. was supported in part by grants MTM2016-76566-P, MTM2017-83262-C2-
2-P and Programa Salvador de Madariaga PRX18/451 from Micinn (Spain), and
grant 20906/PI/18 from Fundaci´on S´eneca (Regi´on de Murcia, Spain). E.H. was
supported by grant MTM2016-76566-P (Spain), and by the European Union's Horizon
2020 research and innovation programme under the Marie Sk lodowska-Curie grant
agreement No 777822. D.K. was supported by Simons Foundation Collaborative Grant
No 636954. V.T. was supported by the Russian Federation Government Grant No.
14.W03.31.0031.
32
S. DILWORTH, G. GARRIG ´OS, E. HERN ´ANDEZ, D. KUTZAROVA, AND V. TEMLYAKOV
References
[1] 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.
[2] S.J. Dilworth, N.J. Kalton, Denka Kutzarova, and V.N. Temlyakov, The Thresholding
Greedy Algorithm, Greedy Bases, and Duality, Constr. Approx. 19 (4) (2003), 575 -- 597.
[3] S.J. Dilworth, D. Kutzarova, K. Shuman, V. Temlyakov, and P. Wojtaszczyk, Weak
convergence of greedy algorithms in Banach spaces. J. Fourier Anal. Appl. 14 (2008), no. 5-6,
609 -- 628.
[4] S.J. Dilworth, M. Soto-Bajo, and V.N. Temlyakov, Quasi-greedy bases and Lebesgue-
type inequalities, Studia Math. 211 (1) (2012), 41 -- 69.
[5] T. Figiel, On the moduli of convexity and smoothness. Studia Math. 56 (2) (1976), 121 -- 155.
[6] V.F. Gaposkin, On unconditional bases in Lp (p > 1) spaces. Uspehi Mat. Nauk 13 (1958) no.
4 (82), 179184.
[7] G. Garrig´os, E. Hern´andez, Sharp Jackson and Bernstein Inequalities for n-term Approxi-
mation in Sequence spaces with Applications, Indiana Univ. Math. J. 53 (6) (2004), 1739 -- 1762.
[8] G. Garrig´os, P. Wojtaszczyk, Conditional quasi-greedy bases in Hilbert and Banach spaces.
Indiana Univ. Math. J. 63 (4) (2014), 1017 -- 1036.
[9] R. Gribonval, M. Nielsen, Some remarks on non-linear approximation with Schauder bases.
East J. Approx. 7 (3) (2001), 267 -- 285.
[10] V.I. Gurarii and N.I. Gurarii, On bases in uniformly convex and uniformly smooth spaces.
Izv. Akad. Nauk SSSR Ser. Mat. 35 (1971), 210 -- 215.
[11] M.I. Kadec and A. Pelczynski, Bases, lacunary sequences, and complemented subspaces in
the spaces Lp. Studia Math. 21 (1962), 161-176.
[12] A. Kamont, V.N. Temlyakov, Greedy approximation and the multivariate Haar system.
Studia Math. 161 (3) (2004), 199 -- 223.
[13] K. Kazimierski, On the smoothness and convexity of Besov spaces. J. Inverse Ill-Posed Probl.
21 (3) (2013), 411 -- 429.
[14] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I,II. Springer-Verlag, Berlin,
1979.
[15] G. Kerkyacharian, D. Picard, V.N. Temlyakov, Some inequalities for the tensor product
of greedy bases and weight-greedy bases. East J. Approx. 12 (1) (2006), 103 -- 118.
[16] M. Nielsen, An example of an almost greedy uniformly bounded orthonormal basis for Lp(0, 1).
J. Approx. Theory, 149 (2) (2007), 188 -- 192.
[17] V.N. Temlyakov, Greedy algorithms in Banach spaces. Adv. Comput. Math. 14 (3) (2001),
277 -- 292.
[18] V.N. Temlyakov. Greedy Approximation. Cambridge University Press, Cambridge, 2011.
[19] V.N. Temlyakov, Sparse approximation and recovery by greedy algorithms in Banach spaces.
Forum Math, Sigma 2 (12) (2014), 26 p.
[20] V.N. Temlyakov, Sparse Approximation with Bases, Advanced Courses in Mathematics CRM
Barcelona, Birkhauser, Springer Basel 2015.
[21] V.N. Temlyakov. Multivariate Approximation. Cambridge University Press, Cambridge, 2018.
[22] H. Triebel. Function spaces with dominating mixed smoothness. European Mathematical So-
ciety, Zurich, 2019.
[23] P. Wojtaszczyk, Greedy Algorithm for General Biorthogonal Systems, J. Approx. Theory
107 (2) (2000), 293-314.
LEBESGUE-TYPE INEQUALITIES IN GREEDY APPROXIMATION
33
S. J. Dilworth, Department of Mathematics, University of South Carolina, Columbia,
SC, USA
E-mail address: [email protected]
Gustavo Garrig´os, Departamento de Matem´aticas, Universidad de Murcia, 30100
Murcia, Spain
E-mail address: [email protected]
Eugenio Hern´andez, Departamento de Matem´aticas, Universidad Aut´onoma de Madrid,
28049 Madrid, Spain
E-mail address: [email protected]
D. Kutzarova, Department of Mathematics, University of Illinois Urbana-Champaign,
Urbana, IL 61807, USA; Institute of Mathematics and Informatics, Bulgarian Acad-
emy of Sciences, Sofia, Bulgaria
E-mail address: [email protected]
V. N. Temlyakov, Department of Mathematics, University of South Carolina,
Columbia, SC, USA; Steklov Institute of Mathematics, Moscow, Russia; Lomonosov
Moscow State University, Moscow, Russia
E-mail address: [email protected]
|
1901.05245 | 1 | 1901 | 2019-01-16T12:03:12 | c-numerical range of operator products on B(H) | [
"math.FA"
] | Let H be a complex Hilbert space of dimension no less than 2 and B(H) be the algebra of all bounded linear operators on H. We give the form of surjective maps on B(H) preserving c-numerical range of operator products when the maps satisfy preserving weak zero products. As a result, we obtain the characterization of surjective maps on Mn(C) preserving c-numerical range of operator products. The proof of the results depends on some propositions of operators in B(H), which are of different interest. | math.FA | math |
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
YANFANG ZHANG,1 and XIAOCHUN FANG2∗
Abstract. Let H be a complex Hilbert space of dimension ≥ 2 and B(H)
be the algebra of all bounded linear operators on H. We give the form of
surjective maps on B(H) preserving c-numerical range of operator products
when the maps satisfy preserving weak zero products. As a result, we obtain
the characterization of surjective maps on Mn(C) preserving c-numerical range
of operator products. The proof of the results depends on some propositions
of operators in B(H), which are of different interest.
1. Introduction and preliminaries
Let H be a complex Hilbert space with inner product h·, ·i, B(H) and Bs(H)
be the algebra of all bounded linear operators and the real linear space of all self-
adjoint operators on H, respectively. For c = (c1, ..., ck) ∈ Rk and k ≤ dim H, the
c-numerical range and c-numerical radius of A ∈ B(H) are respectively defined
as
Wc(A) =( kXj=1
cjhAej, eji : {e1, ..., ek} is an orthonormal subset in H) ,
rc(A) = sup{λ : λ ∈ Wc(A)}.
Obviously, if k = 1, we get the classical numerical range W (A) and the nu-
merical radius w(A) of A. If (c1, ..., ck) = (1, ..., 1), Wc(A) and rc(A) reduce to
k-numerical range and k-numerical radius of A, respectively ( [4, 8, 10]). Numer-
ical range (radius) and c-numerical range (radius) are important concepts and
have many applications in pure and applied mathematics, especially in quantum
control and quantum information. Readers can refer to [8, 21] for more informa-
tion.
It is always of interest to characterize mappings with some special properties
such as leaving certain functions, subsets or relations invariant, which are called
preserving problems. There is a considerable interest about preservers of numer-
ical range and generalized numerical range (radius)[3, 6, 15, 16]. Li gave the
form of linear map on Mn(C) preserving c-numerical range (resp. radius) [15]. It
is quite different from linear maps preserving numerical range (resp. radius)on
B(H)([1]). Hou and Di [13] gave the characterization of surjective maps pre-
serving numerical range of operator products on B(H) and Bs(H) and showed
that such maps have the form A → ±U AU∗ for some unitary U ∈ B(H). It is
2010 Mathematics Subject Classification. Primary 47B49; Secondary 47A12.
Key words and phrases. Preserver, c-numerical range, and elliptical ranges.
1
2
YANFANG ZHANG, and XIAOCHUN FANG
natural to ask what is the form of surjective maps preserving c-numerical range
of operator products.
Let A, B be two algebras and Φ : A → B is a map. For A, B ∈ A, AB = 0 ⇒
Φ(A)Φ(B) = 0 (AB = 0 ⇔ Φ(A)Φ(B) = 0), Φ is called a map preserving zero
product (preserving zero products on both sides). The characterization of surjec-
tive maps preserving zero products on both sides on standard operator algebras
was given in [5, Lemma 2.2].
If c-numerical radius is a norm, maps preserv-
ing c-numerical radius of operator products on B(H) are maps preserving zero
products on both sides. Then by [5, Theorem 2.3], the form of maps preserving
c-numerical radius of operator products for dim(H) ≥ 3 can get when rc(·) is a
norm. The same result can also be obtained from [2, Theorem 3.2 and Theorem
3.3 ]. From these, it is possible to get the form of surjective maps preserving
c-numerical range of operator products on B(H) only when rc(·) is a norm and
the dimension of H is no less than 3.
In this article, we give the characterization of surjective maps on Mn(C) pre-
serving c-numerical range of matrix products. Moreover, we give the definition
of maps preserving weak zero products (on both sides). Then we obtain the form
of surjective maps on B(H) preserving c-numerical range of operator products
when the maps satisfy preserving weak zero products. From this result, the char-
acterization of surjective maps on B(H) preserving c-numerical range of operator
products can get easily when rc(·) is a norm. Here dim H just needs to be greater
than 1. Then the characterization of surjective maps preserving numerical range
of operator products in [13] is just a special case of this result. It is worth noting
that we calculate the c-numerical range of some special operators in B(H) and
show a type of operators with symmetric c-numerical range whenever H is of
finite or infinite dimension. They are of different interest.
Next we introduce some notations and state the assumptions in the paper.
(1) Assume c = (c1, ..., ck) ∈ Rk and c1 ≥ ... ≥ ck with not all ci's equal.
(2)To avoid trivial considerations, we always assume k ≥ 2.
(3) If dim H = n < ∞, we regard B(H) as Mn(C), which is the set of all
complex n × n matrices. Also we assume n ≥ k.
(4) F1(H), S and T are respectively the set of rank-1 operators, the set of
operators with symmetrical c-numerical range and the set satisfying {A ∈ S :
AB ∈ S and BA ∈ S f or B ∈ B(H)} in B(H).
The following are our main results.
Theorem 1.1. Let Mn(C) be the set of all n × n complex matrices and Φ :
Mn(C) → Mn(C) be a surjective map. Then Wc(Φ(A)Φ(B)) = Wc(AB) for all
A, B ∈ Mn(C) if and only if Φ has one of the following forms:
(i) when c1 + ck 6= 0, then there is a unitary matrix U ∈ Mn(C) such that
Φ(A) = ±U AU∗ for all A ∈ Mn(C);
(ii) when there exists an integer p (1 < p < k) satisfying cp + ck+1−p 6= 0 and
cj + ck+1−j = 0 for j = 1, ..., p − 1, then there are a unitary matrix U ∈ Mn(C),
functions φ : T → {−1, 1} and φ′ : Mn(C) → {−1, 1} satisfying φ′(A)φ′(B) =
such that
sgn(AB) =(cid:26) {1}
Φ(A) =
φ(A)U AU∗
φ′(A)U AU∗
±U AU∗
if AB 6∈ S
{−1, 1} if AB ∈ S
if A ∈ T
if A ∈ S \ T
if A ∈ Mn(C) \ S
for all A ∈ Mn(C).
(iii) when cj + ck+1−j = 0 for j = 1, ..., k, then there are a unitary matrix
U ∈ Mn(C) and a function g : Mn(C) → {1, −1} such that
or
Φ(A) = g(A)U AU∗
for all A ∈ Mn(C)
Φ(A) = ig(A)U AU∗
for all A ∈ Mn(C).
Next result is about surjective maps on B(H) preserving c-numerical range of
operator products when the maps satisfy preserving weak zero products.
Theorem 1.2. Let B(H) be the algebra of all bounded linear operators on complex
Hilbert space H. Suppose Φ : B(H) → B(H) is a surjective map preserving weak
zero products on both sides. Then Wc(Φ(A)Φ(B)) = Wc(AB) for all A, B ∈ B(H)
if and only if Φ has one of the following forms:
(i) when c1 + ck 6= 0, there is a unitary operator U ∈ B(H) such that Φ(A) =
±U AU∗ for all A ∈ B(H);
(ii) When there exists an integer p (1 < p < k) satisfying cp + ck+1−p 6= 0 and
cj + ck+1−j = 0 for j = 1, ..., p − 1, then there are a unitary operator U ∈ B(H),
functions φ : T → {−1, 1} and φ′ : B(H) → {−1, 1} satisfying φ′(A)φ′(B) =
sgn(AB) for any A, B ∈ B(H) where
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
3
sgn(AB) for any A, B ∈ Mn(C) where
sgn(AB) =(cid:26) {1}
Φ(A) =
φ(A)U AU∗
φ′(A)U AU∗
±U AU∗
if AB 6∈ S
{−1, 1} if AB ∈ S
if A ∈ T
if A ∈ S \ T
if A ∈ B(H) \ S
such that
for all A ∈ B(H).
(iii) When cj + ck+1−j = 0 for j = 1, ..., k, then there are a unitary operator
U ∈ B(H) and a function g : B(H) → {1, −1} such that
Φ(A) = g(A)U AU∗
for all A ∈ B(H)
Φ(A) = ig(A)U AU∗
for all A ∈ B(H).
ci 6= 0, then rc(·) is a norm, the following result can be obtained imme-
or
If
nPk=1
diately.
4
YANFANG ZHANG, and XIAOCHUN FANG
Corollary 1.3. Suppose Φ : B(H) → B(H) is a surjective map satisfying
Wc(Φ(A)Φ(B)) = Wc(AB) for all A, B ∈ B(H). If
is either (i) or (ii) in Theorem 1.2.
nPk=1
ci 6= 0, the form of Φ
The paper is organized as follows. In Section 2, we show some properties of
c-numerical range, which are true whether H is finite or not. We calculate c-
numerical range of rank-1, some rank-2 operators in B(H) and show a type of
operators with symmetrical c-numerical range. In Section 3, the proof of main
results will be given.
2.
c-numerical range of operators on B(H)
In [15], Li showed the properties of c-numerical range and c-numerical radius
on Mn(C). First we list some of them which are also true in B(H) when H is of
infinite dimension.
Proposition 2.1. Let c = (c1, ..., ck) ∈ Rk. For A ∈ B(H), then
(1) [20]The set Wc(A) is convex for all A ∈ B(H).
(2) Wc(U AU∗) = Wc(A) for any unitary U ∈ B(H).
(3) Wc(λA) = λWc(A) for any λ ∈ C.
(4) Wc(λI + A) = λ
ci + Wc(A) for any λ ∈ C.
kPi=1
(5) Wβc(A) = βWc(A) for any β ∈ R.
(6) Suppose ci's are not equal. Then Wc(A) is a singleton if and only if A is
a scalar multiple of the identity.
(7) The c-numerical radius rc(A) is a norm on B(H) if and only if
and not all ci's are equal.
ci 6= 0
kPi=1
(8) Suppose ci's are not equal. Then Wc(A) ⊂ R if and only if A is hermitian.
In [17], Li described the c-numerical range for self-adjoint operators especially
when H is infinite dimensional. Let S be a self-adjoint operator in B(H).
If
dimH = k, denote by λ1(S) ≥ ... ≥ λk(S) the eigenvalues of S. If H is infinite
dimensional, define
λm(S) = sup{λm(X∗SX) : X∗X = Im}.
Also λm of an infinite dimensional operator S can be defined as follows. Let
σe(A) = ∩{σ(A + F ) : F ∈ B(H) has finite rank}
be the essential spectrum of A ∈ B(H), and let
which also equals the supremum of the set
λ∞(S) = sup σe(S),
σ(S)\{S − µI has a non − trivial finite dimensional null space}.
Then S = σ(S)∩(λ∞(S), ∞) has only isolated points, and we arrange the elements
in descending order, say, λ1 ≥ λ2 ≥ ... counting multiplicities. If S is infinite,
then λj(S) = λj for each positive integer j. If S has m elements, then λj(S) = λj
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
5
for j = 1, ..., m, and λj = λ∞ for j > m. Then the following property is about
c-numerical range of self-adjoint operators in B(H).
Proposition 2.2. (See [15, 17]) Let S ∈ Bs(H).
(1) If dim H = k, Wc(S) =" kPj=1
(2) If H is of infinite dimension, cl(Wc(S)) = [mc(S), Mc(S)], where
cjλk+1−j(S),
cjλj(S)#.
kPj=1
ck+1−jλj(S) : 0 ≤ l ≤ k)
k−lXj=1
ck+1−jλj(−S) : 0 ≤ l ≤ k) .
k−lXj=1
cjλj(−S) +
cjλj(S) −
cj
if dim H = k
max{cj, 0} if dim H > k
cj
if dim H = k
min{cj, 0} if dim H > k
and
mc(S) = inf(−
lXj=1
Mc(S) = sup( lXj=1
¯cj =(cid:26)
ecj =(cid:26)
Next we denote
and
for j = 1, .., k. This will be used to describe c-numerical range of some operators
in B(H).
Proposition 2.3. Suppose that T is a rank-1 operator in B(H), then Wc(T ) is an
elliptical disk with foci ¯c1tr(T ),ecktr(T ) and minor axis (¯c1−eck)pk T k2 − tr(T ) 2
or a line segment with end points ¯c1tr(T ) andecktr(T ).
Proof. If dim H = k, the conclusion can be found in [6]. If dim H = n > k, adding
zero to c, we get d = (d1, ..., dn) with d1 ≥ ... ≥ dn and Wc(A) = Wd(A) for all
A ∈ B(H). Obviously, d1 = max{c1, 0} and dn = min{ck, 0}. The conclusion
also can be obtained. Next we prove it when H is infinite dimensional.
Assume T = x ⊗ f with k x k= 1, then there is unit x′ in [x]⊥ and complex a, d
such that f = ax + dx′. then with respect to a suitable orthogonal basis in H =
[x] ⊕ [x′] ⊕ [x, x′]⊥, T =(cid:18) a d
0 0 (cid:19) ⊕ 0 with k T k=p a 2 + d 2 and tr(T ) = a.
2(eiθT + e−iθT ∗).
For 0 ≤ θ < 2π, The real part of eiθT is denoted by Re(eiθT ) = 1
We calculate the non-zero eigenvalues of Re(eiθT ), that is
and
λ1 =
1
2
Re(eiθa) +
λ2 =
1
2
Re(eiθa) −
1
2p(Re(eiθa))2 + d2 > 0
2p(Re(eiθa))2 + d2 < 0.
1
6
YANFANG ZHANG, and XIAOCHUN FANG
From Proposition 2.2(2), we know Wc(Re(eiθT )) = [mc(Re(eiθT ), Mc(Re(eiθT )],
ckλ2
c1λ1
where
if c1 ≤ 0
c1λ1 + ckλ2
if c1 > 0 and ck < 0
Writing a as a = aeiφ, then
Mc(Re(eiθT )) =
With the notation of ¯cj andecj, it can be written Mc(Re(eiθT )) = ¯c1λ1 +eckλ2.
Mc(Re(eiθT )) = ¯c1λ1 +eckλ2
2 pa2 cos2(θ + φ) + d2
2(¯c1 +eck)Re(eiθa) + (¯c1−eck)
2(¯c1 +eck)a cos(θ + φ) + (¯c1−eck)
It is shown in ([6, 7]) that Ω is an elliptical disc centered at peiφ with foci
2 p(Re(eiθa))2 + d2
(p − β)eiφ and (p + β)eiφ, and semi-minor axis α if and only if
= 1
= 1
if ck ≥ 0.
(2.1)
Mc(Re(eiθΩ)) = p cos(θ + φ) +pβ2 cos2(θ + φ) + α2.
1
(cid:3)
Proposition 2.4. Let A ∈ B(H) be rank-2 and there is a suitable orthonormal
0 b (cid:19) ⊕ 0 with a, b, d ∈ C.
Then we can get Wc(T ) is an ellipse with ¯c1tr(T ) andecktr(T ) as foci, and
2(¯c1 −eck)pk (T ) k2 −tr(T )2 as semi-minor axis. Specially, if tr(T ) =k T k,
Wc(T ) degenerates to a line segment with end points ¯c1tr(T ) andecktr(T ).
basis such that A =(cid:18) a d
(1) If a ≥ b > 0 and d = 0, then Wc(A) is the line segment [ecka +eck−1b, ¯c1a +
(2) If a > 0 > b and d = 0, then Wc(A) is the line segment [ecka+¯c1b, ¯c1a+eckb].
(3) If 0 > a ≥ b and d = 0, then Wc(A) is the line segment [¯c1b + ¯c2a,eckb +
eck−1a].
¯c1a +eckb and ¯c1b +ecka and with minor axis (¯c1 −eck)d.
(5) If dim H ≥ 3 and c satisfying cj + ck+1−j = 0 for j = 1, 2, and a, b ∈ R+
(4) If a, b ∈ R and d2 > 4ab, then Wc(A) is the elliptical disk with foci
with 0 < d2 < 4ab, then Wc(A) is not an elliptical disk.
¯c2b].
Proof. (1) For the case dim H < ∞, the conclusion can get from Proposition
2.2 easily. If dim H = ∞, A is self-adjoint and Wc(A) = [m, M], where M =
max{c1a + c2b, c1a, 0} and m = min{cka + ck−1b, cka, 0}. That is
c1a + c2b
if c2 ≥ 0
c1a
0
if c1 > 0 and c2 < 0
if c1 ≤ 0
cka + ck−1b
cka
0
if ck−1 ≤ 0
if ck−1 > 0 and ck < 0
if ck ≥ 0.
and
M =
m =
There are 9 combinations of m, M, five of which can be possible with c1 ≥
c2 ≥ ... ≥ ck as follows:
(a) if ck ≥ 0, then m = 0, M = c1a + c2b;
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
7
(b) if c2 > 0 and ck−1 < 0, then m = cka + ck−1b and M = c1a + c2b;
(c) if c1 ≤ 0, then m = cka + ck−1b and M = 0;
(d) if c1 > 0 and c2 < 0, then m = cka + ck−1b and M = c1a;
(e) if ck−1 > 0 and ck < 0, then m = cka and M = c1a + c2b.
(2),(3) can get as the proof of (1) similarly.
(4) Assume a, b ∈ R and d2 > 4ab. For 0 ≤ θ < 2π, the real part of eiθA is
2 (eiθA + e−iθA∗). We calculate the non-zero eigenvalues of Re(eiθA)
Re(eiθA) = 1
denoting λ1, λ2. Then
From the cases, we get m =ecka +eck−1b and M = ¯c1a + ¯c2b.
(a + b) cos θ +p(a − b) cos2 θ + d2
(a + b) cos θ −p(a − b) cos2 θ + d2
(¯c1 +eck)(a + b) cos θ
Mc(Re(eiθA)) =
λ1 =
λ2 =
2
2
+
2
(¯c1 −eck)p(a − b)2 cos2 θ + d2
2
Together with Equation (2.1), it follows that Wc(A) is an elliptical disk with
foci ¯c1a +eckb and ¯c1b +ecka and minor axis (¯c1 −eck)d.
(5) Without loss of generality, suppose a ≥ b. For 0 ≤ θ < 2π, the non-zero
eigenvalues of Re(eiθA) is same as that (2.2) and (2.3). Besides, for d2 < 4ab,
there must be φ ∈ (0, π
2 ) such that cos2 φ = d2
4ab . Then
Obviously, λ1 > 0 > λ2. So we have Mc(Re(eiθA)) = ¯c1λ1 + ckλ2. Then
(2.2)
(2.3)
.
where S1 = [2π − φ, 2π) ∪ [0, φ], S2 = (φ, π − φ) ∪ (π + φ, 2π − φ) and S3 =
[π − φ, π + φ].
We know W (Re(eiθA)) is interval [m(θ), M(θ)]. From the results in (1), (2)
and (3) above, we get
λ1 ≥ λ2 > 0 if θ ∈ S1
λ1 > 0 > λ2
if θ ∈ S2
if θ ∈ S3.
0 > λ1 ≥ λ2
¯c1λ1 + ¯c2λ2
¯c1λ1 +eckλ2
eck−1λ1 +eckλ2
c1λ1
c1(λ1 − λ2)
−c1λ2
c1λ1 + c2λ2
c1(λ1 − λ2)
−c2λ1 − c1λ2
if θ ∈ S1
if θ ∈ S2
if θ ∈ S3.
if θ ∈ S1
if θ ∈ S2
if θ ∈ S3.
if θ ∈ S1
if θ ∈ S2
if θ ∈ S3.
M(θ) =
M(θ) =
M(θ) =
With c1 + ck = 0 and c2 + ck−1 = 0, we get
when k = 2 and dim H ≥ 3;
or
when k ≥ 3.
8
YANFANG ZHANG, and XIAOCHUN FANG
Denote
M(θ) can be written as
M(θ) =
if k ≥ 3
if k = 2.
2
c1
2
2
c1
2
if k ≥ 3
if k = 2
and c =(cid:26) c1−c2
c =(cid:26) c1+c2
c(a + b) cos θ + cp(a − b)2 cos2 θ + d2
− c(a + b) cos θ + cp(a − b)2 cos2 θ + d2
M(θ) = p cos(θ + ω) +pq2 cos2(θ + ω) + t2.
( c + c)p(a − b)2 cos2 θ + d2
if θ ∈ S1
if θ ∈ S2
if θ ∈ S3.
If Wc(A) is an ellipse, then M(θ) has the form as (2.1), i.e. there is p, q, t ∈ R
and ω ∈ [0, 2π) such that
From (2.4), we get M(θ) = M(θ + π). Then it can be deduced p = 0 in (2.5).
So (2.5) can be written as
(2.3′)
Next we show ω = 0. For θ ∈ (φ, π − φ) ∪ (π + φ, 2π − φ), (2.4) and (2.3′) tell
M(θ) =pq2 cos2(θ + ω) + t2.
that
q2 cos2(θ + ω) + t2 = ( c + c)2(a − b)2 cos2 θ + ( c + c)2d2.
(2.6)
Taking the derivative about θ on both sides of (2.6), we get
q2 sin(2θ + 2ω) = ( c + c)2(a − b)2 sin 2θ.
(2.7)
Let θ = π
2 , (2.7) shows sin 2ω = 0 i.e. ω = 0 or π
2 . Again take the
derivative on both sides of (2.6) and let θ = π
2 , we get q2 cos 2ω = ( c + c)2(a − b)2.
So cos 2ω has to be non-negative and then ω = 0. Then (2.3′) can be written as
2 or 3π
M(θ) = ( c + c)p(a − b)2 cos2 θ + d2.
Let θ = 0 in (2.4) and (2.3′′), we have
c(a + b) + cp(a − b)2 + d2 = ( c + c)p(a − b)2 + d2.
which get d2 = 4ab, a contradiction. So Wc(A) is not an ellipse.
Proposition 2.5. Suppose T, S ∈ B(H) are rank-1 with Wc(T ) = Wc(S), then
(1) If c1 + ck 6= 0, tr(T ) = tr(S);
(2) If c1 + ck = 0, tr(T ) = tr(S) or tr(T ) = −tr(S).
Proof. By Proposition 2.3, Wc(T ) = Wc(S) implies one of the following equations
true:
(2.4)
(2.5)
(2.3′′)
(cid:3)
(2.8)
(2.9)
(cid:26) ¯c1tr(T ) = ¯c1tr(S)
ecktr(T ) =ecktr(S)
(cid:26) ¯c1tr(T ) =ecktr(S)
ecktr(T ) = ¯c1tr(S)
With the fact that c1 and ck are not 0 at the same time, we can get tr(T ) =
tr(S) from (2.8). For (2.9), it shows tr(T ) = tr(S) = 0 when c1 + ck 6= 0 and
tr(T ) = −tr(S) when c1 + ck = 0. Hence, when c1 + ck 6= 0, tr(T ) = tr(S). When
c1 + ck = 0, either tr(T ) = tr(S) or tr(T ) = −tr(S) holds.
(cid:3)
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
9
Next we consider the symmetry of c-numerical range for operators in B(H).
Proposition 2.6. Let A ∈ B(H) be a self-adjoint operator with finite ranks. For
c = (c1, ..., ck), there is an integer p (1 < p < k) such that cp + ck+1−p 6= 0 and
cj + ck+1−j = 0 for j = 1, ..., p − 1.
(1) If rank A < p, then Wc(A) = −Wc(A);
(2) If rank A = p and A or −A is positive, then Wc(A) 6= −Wc(A).
Proof. Assume A ∈ B(H) is self-adjoint with rank A = a (a ≤ p). Let λ1 ≥ ... ≥
λl be the positive eigenvalues of A and λk ≤ ... ≤ λk+1−(a−l) be the negative ones.
From Proposition 2.3, there is ξ, η ∈ R such that Wc(A) = [ξ, η].
(1) If 1 ≤ a < p, we can calculate ξ =
ck+1−jλj +
cjλa+1−j and η =
lPj=1
a−lPj=1
cjλj +
ck+1−jλa+1−j. Obviously, ξ + η =
(cj + ck+1−j)λj +
ck+1−j)λa+1−j = 0. So Wc(A) = −Wc(A).
(2) If a = p and A is positive, the eigenvalues of A, saying λ1, ..., λp, are all
(cj +
a−lPj=1
positive. When dim H = k, then ξ =
ck+1−jλj and η =
cjλj. Easily,
a−lPj=1
lPj=1
lPj=1
pPj=1
ξ + η =
ck+1−jλj +
cjλj 6= 0.
When H is infinite dimensional, we know
pPj=1
ξ =
and
pPj=1
pPj=1
k−tPj=1
pPj=1
pPj=1
pPj=1
tPj=1
η =
ck+1−jλj
ck+1−jλj
ck+1−jλj
if p ≤ t and p < k − t
if p ≤ t and p > k − t
if p > t and p ≤ k − t
cjλj
if p ≤ t and p < k − t
cjλj
if p ≤ t and p > k − t
cjλj
if p > t and p ≤ k − t
pPj=1
p−1Pj=1
k−tPj=1
where t is the integer satisfying 1 ≤ t ≤ k and ct ≥ 0 > ct+1. Next we show
ξ + η 6= 0 according to different cases of t:
(a) If p ≤ t and p < k − t, then ξ + η =
(cj + ck+1−j)λj + (cp + ck+1−p)λp 6= 0.
(b) If p ≤ t and p > k − t, then ξ + η =
(cj + ck+1−j)λj + (ck−t+1λk−t+1 + ... +
cpλp) = ck−t+1 + ... + cp. If we assume ξ + η = 0, then ck−t+1λk−t+1 + ... + cpλp = 0,
and with ck−t+1 ≥ ... ≥ cp ≥ ct ≥ 0, we get ck−t+1 = ... = cp = ... = ct = 0. Also
cp + ck+1−p 6= 0, so ck+1−p 6= 0. By p > k − t, we can get k + 1 − p ≤ t. That is
10
YANFANG ZHANG, and XIAOCHUN FANG
ck+1−p ≥ ct = 0. But from the assumption p ≤ k +1−p, it comes ck+1−p ≤ cp = 0.
We get a contradiction and ξ + η 6= 0.
(c) If p > t and p ≤ k − t, then ξ + η =
(cj + ck+1−j)λj + (ck+1−(t+1)λt+1 +
... + ck+1−pλp) = ck+1−(t+1)λt+1 + ... + ck+1−pλp < 0, so ξ + η 6= 0. This means
Wc(A) 6= −Wc(A) if a = p. As to −A is positive, similar results can get.
(cid:3)
tPj=1
For the self-adjoint operators with rank more than p, the symmetry of c-
numerical range is associated with cj and ck+1−j for j > p, hence we failed to give
a fixed conclusion.
3. Maps preserve c-numerical range of operator products on B(H)
In this section, first we give the definition of maps preserving weak zero prod-
ucts.
Definition 3.1. Let A, B be two subalgebras of B(H) containing F1(H) and
Φ : A → B be a map. Φ is said to preserve weak zero products if for any A and
rank-1 T ∈ A, AT = 0 implies Φ(A)Φ(T ) = 0. Φ is said to preserve weak zero
products on both sides if it preserves weak zero products and for any B and rank-1
S ∈ B, BS = 0 implies B S = 0 where Φ( B) = B and Φ( S) = S.
Obviously, a map preserving zero products must be one preserving weak zero
products. Next we give an example to show the inverse is not true.
Example 3.2. Let H = H1 ⊕ H2 with dim H1 = 2 and U =(cid:18) 0 1
1 0 (cid:19) ⊕ I be an
unitary operator. Define Φ : B(H) → B(H) by
Φ(A) =(cid:26) U AU∗
AU∗
if rankA ≤ 1
if rank A ≥ 2
It is easy to check that Φ is preserving left-weak zero product. Take non-zero
B0 ∈ B(H) and let A =(cid:18) 0 1
However, Φ(A)Φ(B) = U(AU∗B)U∗ = U((cid:18) 1 0
0 0 (cid:19) ⊕ 0 and B =(cid:18) 1 0
0 0 (cid:19) ⊕ 0)U∗ 6= 0.
0 0 (cid:19) ⊕ B1, then AB = 0.
From Theorem 1.2, it can be see that for maps preserving c-numerical range
of operator products on B(H), they are equivalent between preserving weak zero
products and preserving zero products.
Next we show some lemmas which are useful to the proof of main theorems .
Lemma 3.3. Let A ∈ B(H). The following conditions are equivalent.
(1) A is a rank-1 operator.
(2) For any B ∈ B(H) with AB 6= 0, there is tB ∈ C such that Wc(AB) is
an elliptical disk (a line segment) with ¯c1tB andecktB as foci (end points).
Proof. Since(1)⇒(2) is clear, next we show (2)⇒(1). Suppose rank A ≥ 2,
then there are linear independent x1,x2 ∈ H such that Ax1⊥Ax2 and kAx1k =
kAx2k = 1. We finish the proof by three cases:
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
11
(a): If c1 + ck 6= 0, we let B1 = αx1 ⊗ Ax1 + βx2 ⊗ Ax2, where α > 0 > β.
AB1 is a self-adjoint operator in B(H). By proposition 3(2), Wc(AB1) = [¯c1β +
(cid:26) ¯c1β +eckα =ecktB1
¯c1α +eckβ = ¯c1tB1
(b): If c1 + ck = 0 and c2 + ck−1 6= 0, we let B2 = ξx1 ⊗ Ax1 + ηx2 ⊗ Ax2,
where ξ > η > 0 . AB is a self-adjoint operator in B(H). By proposition 3(1),
eckα, ¯c1α +eckβ]. On the other hand, there is a complex tB1 such that Wc(AB1) is
a line segment with end points ¯c1tB1 andecktB1. So we get
(cid:26) ¯c1β +eckα = ¯c1tB1
¯c1α +eckβ =ecktB1
It deduces ¯c1 +eck = 0, impossible.
Wc(AB2) = [eckξ +eck−1η, ¯c1ξ +¯c2η]. Also there is a complex tB2 such that Wc(AB2)
is a line segment with end points ¯c1tB2 andecktB2. So
(cid:26) eckξ +eck−1η = ¯c1tB2
¯c1ξ + ¯c2η =ecktB2
It deduces ¯c2 +eck−1 = 0, impossible.
(c): If c1 + ck = 0 and c2 + ck−1 = 0, we let B3 = x1 ⊗ Ax1 + x2 ⊗ Ax2 when
dim H = 2. Then Wc(AB3) is a singleton, a contradiction. When dim H ≥ 3, let
B4 = ax1 ⊗ Ax1 + εx2 ⊗ Ax1 + bx2 ⊗ Ax2 where a > b > 0 and 0 < ε2 < 4ab, then
we know Wc(AB4) is not an ellipse from proposition (5), a contradiction. So A
must be rank-1.
(cid:3)
or
or
(cid:26) eckξ +eck−1η =ecktB2
¯c1ξ + ¯c2η = ¯c1tB2
Lemma 3.4. Let T ∈ B(H), there is a function h : H × H → {1, −1} satisfying
hT x, f i = h(x, f )hx, f i for each x, f ∈ H, then T = ±I.
Proof. Firstly, we will show hT x, xi = hx, xi for any unit x ∈ H or hT x, xi =
−hx, xi for any unit x ∈ H. If not, there are linearly independent unit x1, x2 ∈ H
such that hT x1, x1i = hx1, x1i and hT x2, x2i = −hx2, x2i, then
hT (x1 + x2), x1 + x2i = hT x1, x1i + hT x2, x2i + hT x1, x2i + hT x2, x2i
= hx1, x1i + hx1, x2i − hx2, x2i − hx2, x1i.
We know either hT (x1 + x2), x1 + x2i = hx1 + x2, x1 + x2i or hT (x1 + x2), x1 + x2i =
−hx1 + x2, x1 + x2i. Hence, we get hx1, x2i = kx1k2 or hx2, x1i = kx2k2, a
contradiction with the linear independence of x1, x2.
Assume hT x, xi = hx, xi for x ∈ H. For each f ∈ H, there is x ∈ [x] and
f ∈ [x]⊥ such that f = x + f , where [x] is the subspace spanned by x and
[x]⊥ is orthogonal subspace of [x]. Then hT x, f i = hT x, x + f i = hx, xi. So
hT x, f i = hx, f i for any x, f ∈ H. Then T = I. If hT x, xi = −hx, xi holds for
x ∈ H, T = −I can be get similarly.
(cid:3)
Lemma 3.5. [18, Wigner Theorem] Let H be a complex Hilbert space and T :
H → H be a bijective satisfying hT x, T yi = hx, yi for any x, y ∈ H, then there
are a unitary or anti-unitary operator U : H → H and a function θ : H → C
with θ(x) ≡ 1 such that T x = θ(x)U x for any x ∈ H.
Proof of Theorem 3.4. The sufficiency is obvious, here we only show the
necessity, i.e. Φ is a surjective map preserving weak zero products on both sides
and Wc(Φ(A)Φ(B)) = Wc(AB) for all A, B ∈ B(H).
12
YANFANG ZHANG, and XIAOCHUN FANG
Claim 1. Φ preserving rank-1 on both sides.
Assume T ∈ B(H) is rank-1 and T satisfying Φ( T ) = T . For any A ∈ B(H)
with AΦ(T ) 6= 0, there is A such that Φ( A) = A. Then Wc(Φ( A)Φ(T )) =
Wc(AΦ(T )) = Wc( AT ). Because Φ preserves weak zero products on both sides,
then AT 6= 0. Thus Wc(AΦ(T )) is an ellipse (line segment) with foci(end points)
rank-1. Next we will finish the proof by two cases.
¯c1tr( AT ) and ecktr( AT ). By Lemma 3.3, Φ(T ) is rank-1. Similarly, T is also
Case I when c1 + ck 6= 0.
Proof of this case will be finished by 3 steps.
Step 1.1. Φ is injective.
Assume A1, A2 ∈ B(H) satisfying Φ(A1) = Φ(A2). From Wc(Φ(Ai)Φ(x⊗f )) =
Wc(Aix⊗f ) for i = 1, 2, we can get Wc(A1x⊗f ) = Wc(A2x⊗f ) for any x, f ∈ H.
With c1 + ck 6= 0 and Proposition ??, hA1x, f i = hA2x, f i for any x, f ∈ H. So
A1 = A2 and Φ is injective.
Step 1.2. Φ is linear.
Let A1 and A2 in B(H) and T ∈ B(H) be rank-one, we know tr(Φ(A1 +
A2)Φ(T )) = tr((A1 + A2)T ). So tr(Φ(A1 + A2)Φ(T )) = tr(A1T ) + tr(A2T ) =
tr((Φ(A1) + Φ(A2))Φ(T )). By the surjection of Φ, Φ(T ) can run over all rank-
one operators. So Φ(A1 + A2) = Φ(A1) + Φ(A2). Similarly, we can check Φ is
homogeneous.
Step 1.3. There is a unitary operator U ∈ B(H) such that Φ(A) = U AU∗ for
any A ∈ B(H) or Φ(A) = −U AU∗ for any A ∈ B(H) .
Now we get that Φ is a linear bijection preserving rank-one operators in both
directions. By [11, Lemma 1.2], Φ has one of the following forms:
(i) There exist bijective linear operators U and V such that Φ(x⊗f ) = U x⊗V f
for any x, f ∈ H.
(ii) There exist bijective conjugate operators U and V such that Φ(x ⊗ f ) =
U f ⊗ V x for any x, f ∈ H.
If (i) holds, that is, there exist bijective linear operators U and V such that
Φ(x ⊗ f ) = U x ⊗ V f for any x, f ∈ H.
By considering Wc((x ⊗ f )2) and Proposition ??2.5(1), we get tr((x ⊗ f )2) =
tr(Φ(x ⊗ f )2). That is hx, f i2 = hU x, V f i2 for any x, f ∈ H. It follows that U, V
are bounded. With Lemma 3.4, V ∗U = ±I.
For any A ∈ B(H), Wc(Ax⊗f ) = Wc(Φ(A)Φ(x⊗f )) implies hV ∗Φ(A)U x, f i =
hAx, f i. Hence Φ(A) = ±U AU−1.
For any unit x ∈ H, Wc((x ⊗ x)2) = W ((U x ⊗ (U−1)∗x)2) implies U x ⊗ (U−1)∗x
is self-adjoint. So there is a positive number r such that U−1x = rU x and
kxk2 = rkU xk2 for unit x. Thus we get U U∗ = rI. Let U1 = 1√r U and U1 is
unitary such that Φ(A) = ±U1AU∗1 for any A ∈ B(H).
If (ii) holds, we can get there is conjugate unitary U such that Φ(A) = ±U A∗U∗
for all A ∈ B(H) similarly as (i).
Pick an orthogonal basis {ejj ∈ J} and define a conjugate unitary operator
J : H → H by Jx =Pj∈J
¯ξjej if x =Pj∈J ξjej. Clearly, J 2 = I, J∗ = J and
A∗ = JAtJ, where At is the transpose of A respect to the basis {ej : j ∈ J}.
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
13
Let U1 = U J, Then U1 is a unitary operator and Φ(A) = ±U1AtU∗1 for all
A ∈ B(H). Thus for any A, B ∈ B(H), we have Wc(AB) = Wc(Φ(A)Φ(B)) =
Wc(U1AtBtU∗1 ) = Wc((BA)t) = Wc(BA), which is impossible.
Case II when c1 + ck = 0.
Step 2.1. For any x, y ∈ H, there are ux, uy and λx,y ∈ T = {λ : λ = 1} such
that Φ(x ⊗ y) = λx,yux ⊗ uy and hx, yi = hux, uyi. Specially, λx,x ∈ {−1, 1}
for any x ∈ H or λx,x ∈ {−i, i} for any x ∈ H.
For any x ∈ H, assume Φ(x ⊗ x) = ux ⊗ vx with kuxk = kxk. By considering
x ⊗ x, we get kxk2Wc(x ⊗ x) = hux, vxiWc(ux ⊗ vx). Hence kxk4 = hux, vxi2or −
hux, vxi2 and hux, vxiux ⊗ vx is self-adjoint. So there exists t ∈ C such that
ux = tvx. Then kxk4 = hux, vxi2 = ¯t−2kuxk4 i.e. t ∈ {−1, 1, −i, i}. Then there is
λx,x ∈ {−1, 1, −i, i} such Φ(x ⊗ x) = λx,xux ⊗ ux for all x ∈ H.
For any y ∈ H, next we will show there is λx,y ∈ T such that Φ(x ⊗ y) =
λx,yux ⊗ uy and hux, uyi = hx, yi. Obviously, Φ preserves orthogonality of
rank-1 operators in both directions. So the ranges of Φ(x ⊗ y) and Φ(x ⊗ y)∗
contain in span[ux, uy], which is the linear subspace spanned by ux, uy. Assume
Φ(x ⊗ y) = vx,y ⊗ wx,y. From Wc(Φ(x ⊗ y)Φ(x ⊗ x)) = Wc((x ⊗ y)(x ⊗ x)), we
If x⊥y, obviously ux⊥wx,y,
get λx,xhux, wx,yiWc(vx,y ⊗ ux) = hx, yiWc(x ⊗ x).
then wx,y is linearly dependant with uy.
vx,y ⊗ ux is
self-adjoint, then vx,y is also linearly dependant with ux. Similarly by considering
(x ⊗ y)(x ⊗ x), we get wx,y is linearly dependant with uy. So there is a complex
λx,y such that Φ(x ⊗ y) = λx,yux ⊗ uy.
If hx, yi 6= 0, λx,xhux,wx,yi
hx,yi
If x ⊥ y, hux, uyi = hx, yi = 0. For Wc(ux⊗uy) is an circle disc with diameter
λx,y = 1 and hux, uyi = hx, yi.
x)). If x 6⊥ y, considering Φ(x ⊗ y)Φ(x ⊗ x), we know λx,yλx,xhux, uyi = hx, yior −
( ¯c1 −eck)kuxkkuyk, we get λx,y = 1 from Wc(Φ(x ⊗ y)Φ(x ⊗ x)) = Wc((x ⊗ y)(x ⊗
hx, yi and pkxk2kyk2 − hx, yi2 = pλx,y2kuxk2kuyk2 − λx,yhux, uyi2. Hence
In fact, between λx,x ∈ {−1, 1} and λx,x ∈ {i, −i}, only one holds for any
x ∈ H. If not, assume there are x0 and y0 satisfying Φ(x0 ⊗ x0) = iux0 ⊗ ux0
and Φ(y0 ⊗ y0) = uy0 ⊗ uy0. Take z0 ∈ span[x0, y0]. Considering Wc(Φ(x0 ⊗
x0)Φ(z0 ⊗z0)), we get λz0 ∈ {i, −i}. However, we get λz0 ∈ {1, −1} by considering
Wc(Φ(y0 ⊗ y0)Φ(z0 ⊗ z0)). That is a contradiction.
Step 2.2. There are a unitary operator U : H → H and a function g : B(H) →
{−1, 1} such that Φ(A) = g(A)U AU∗ for any A ∈ B(H) or Φ(A) = ig(A)U AU∗
for any A ∈ B(H).
By Step 2.1, there are ux with kuxk = kxk such that Φ(x ⊗ x) = λx,xux ⊗ ux for
any x ∈ H. Then define a map V : H → H by V x = ux and V (eiθx) = eiθV x for
any θ ∈ [0, 2π). Next we show V is injective. Assume there is x, y such that V x =
V y. By considering (x⊗x)(y⊗y), we get hy, xiWc(x⊗y) = hV y, V xiWc(V x⊗V y).
then x = ty and t2 = 1 or t2 = −1. Since V (tx) = tV x, it deduces V x 6= V y for
t 6= 1. So x = y. Because Φ is surjective and preserves rank-1 on both sides, we
get V is surjective. Then there is a bijective map V with hV x, V yi = hx, yisuch
that Φ(x ⊗ y) = λx,yV x ⊗ V y for any x, y ∈ H where λx,y ∈ T may be not same
as above and we still use the same symbol. By Lemma 3.5, there are a unitary
14
YANFANG ZHANG, and XIAOCHUN FANG
or anti-unitary operator U : H → H and θ : H → C with θ(x) = 1 such that
V x = θ(x)U x. So Φ(x ⊗ y) = λx,yθ(x)θ(y)U x ⊗ U x for any x, y ∈ H.
Let h(x, y) = θ(x)θ(y)λx,y, then Φ(x⊗y) = h(x, y)U x⊗U x for any x, y ∈ H. It
is obvious that h(x, x) = λx,x. Next we show h(x, y) ∈ {−1, 1} for any x, y ∈ H
or h(x, y) ∈ {−i, i} for any x, y ∈ H. Take z ∈ span[x, y] with hz, xi 6= 0
and hz, yi 6= 0. From Wc(Φ(z ⊗ z)Φ(x ⊗ y)) = Wc((z ⊗ z)(x ⊗ y)), we get
h(z, z)h(x, y)Wc(U z ⊗ U x) = Wc(z ⊗ x). For Wc(z ⊗ x) is an ellipse but not a
circle, h(z, z)h(x, y) = 1 or − 1. With the fact h(x, x) ∈ {−1, 1} for all x ∈ H
or h(x, x) ∈ {−1, 1} for all x ∈ H, we get h(x, y) ∈ {−1, 1} for any x, y ∈ H or
h(x, y) ∈ {−i, i} for any x, y ∈ H. Define a function f : F1(H) → {−1, 1} with
f (x ⊗ y) = h(x, y), then either
Φ(x ⊗ y) = f (x ⊗ y)U x ⊗ U y
for any x, y ∈ H
(3.1)
or
for any x, y ∈ H.
Φ(x ⊗ y) = if (x ⊗ y)U x ⊗ U y
(3.2)
If (3.1) holds, for any A with rankA ≥ 2 and x 6∈ ker A, Wc((Φ(A)U x⊗U Ax) =
Wc(Ax ⊗ Ax). It implies f (x ⊗ Ax)Φ(A)U x ⊗ U Ax is self-adjoint. Then there is
tx,A ∈ {−1, 1} such that U∗Φ(A)U x = tx,AAx. In fact we can denote tx,A = tA
because it is independent with x. For any x, y 6∈ ker A, U∗Φ(A)U x = tx,AAx and
U∗Φ(A)U y = ty,AAy. On the other hand, U∗Φ(A)U(x + y) = tx+y,AA(x + y). So
(ty,A − tx+y,A)Ay = (tx+y,A − tx,A)Ax. tx,A 6= ty,A will deduce Ax = 0 or Ay = 0,
impossible. Define a function g : B(H) → {−1, 1} with
g(A) =(cid:26) f (A)
tA
rankA = 1
rankA ≥ 2,
then Φ(A) = g(A)U AU∗ for any A ∈ B(H).
If (3.2) holds, −iΦ is accord with (3.1). Then Φ(A) = ig(A)U AU∗ for any
A ∈ B(H).
Similarly as Step 1.3 of the proof in Case I, it is impossible that U is anti-
unitary.
Step 2.3. Φ has the form in Theorem 1.2.
When there exists an integer p (1 < p < k) satisfying cp + ck+1−p 6= 0 and
cj + ck+1−j = 0 for j = 1, ..., p − 1, the form Φ(A) = ig(A)U AU∗ for A ∈ B(H)
is impossible. If not, let D ∈ B(H) is a project with rank D = p. Wc(Φ(D)2) =
Wc(D2) deduces −Wc(D) = Wc(D), which is contradict with Proposition 2.6.
Remembering S = {A ∈ B(H) : Wc(A) = −Wc(A)} and T = {A ∈ S : AB ∈
S and BA ∈ S for any B ∈ B(H)}, we know Wc(Φ(I)Φ(B)) = Wc(B) for any
operator B 6∈ S. Thus g(B) = g(I) = 1 or − 1 for any operator B ∈ B(H) \ S.
For any T ∈ T , define the function φ : T → {−1, 1} with φ(T ) = g(T ). For any
S ∈ S \ T , define the function φ′ : B(H) → {−1, 1} with φ′(S) = g(S). Because
Wc(Φ(S)Φ(A)) = Wc(SA) for any A ∈ B(H), φ′ must satisfy φ′(S)φ′(A) =
sgn(SA) for any A ∈ B(H) where
sgn(SA) =(cid:26) {1}
SA 6∈ S
{−1, 1} SA ∈ S.
The proof is finished.
(cid:3)
c-NUMERICAL RANGE OF OPERATOR PRODUCTS ON B(H)
15
Remark 3.6. For the form (iii) in Theorem 1.2, it is not easy to show all the
elements in S and T . Here we show some subsets of them.
Lemma 3.7. For c = (c1, ..., ck), if there exists an integer p (1 < p < k) sat-
isfying cp + ck+1−p 6= 0 and cj + ck+1−j = 0 for j = 1, ..., p − 1. Let V1 = {T :
T is a positive or non-positive operator with rank T = p } and V2 = {T : T is a
self-adjoint operator with rank T < p}, then V1 ⊂ B(H) \ S, V2 ⊂ S, F1(H) ⊂ T
and T 6= S.
Proof. From Proposition 2.6(1),(2), we know V1 ⊂ B(H) \ S, V2 ⊂ S. For each
rank-1 operator T and any operator A ∈ B(H), it comes Wc(AT ) = −Wc(AT ).
Then F1(H) ⊂ T . Let D = Ip−1 ⊕ (−1) ⊕ 0 and G = Ip−1 ⊕ (−2) ⊕ 0. Wc(D) =
cj −ck]. Obviously, Wc(D) = −Wc(D). Then D ∈ S. However
[
ck+1−j −c1,
DG is an positive operator with rank (DG) = p, then Wc(DG) 6= −Wc(DG). So
D 6∈ T and T 6= S.
(cid:3)
p−1Pj=1
At the end, we finish the proof of Theorem 1.1.
Proof of Theorem 1.1.
Claim 1. Φ(A) = 0 ⇔ A = 0.
If
cj 6= 0, then Wc(A) = {0} if and only if A = 0 and the claim is true. If
p−1Pj=1
kPj=1
kPj=1
kPj=1
cj = 0, Wc(T ) = {0} implies T is multiple of the identity. Assume Φ(A) = 0,
then AB is a multiple of the identity for any B ∈ Mn(C). Let B run over all
rank-1 matrices, we get AD = 0 for any rank-1 matrices. Because any matrix in
Mn(C) is a combination of rank-1 matrices. Thus AB = 0 for all B ∈ Mn(C).
Then A = 0.
Claim 2. Φ preserves zero products on both sides.
If
cj 6= 0, obviously Φ preserves zero products. Next we show if
cj = 0, Φ
kPj=1
also preserves zero products. When
cj = 0, On contrary assuming that there
are non-zero A, B ∈ Mn(C) with AB = 0 such that Φ(A)Φ(B) = tI 6= 0. Then
Φ(A) and Φ(B) are both invertible. Taking x, f ∈ Cn such that x ∈ ker A and
f 6∈ ker A∗, we see Ax ⊗ f = 0 and x ⊗ f A 6= 0. This deduces Φ(A)Φ(x ⊗ f ) = sI
where s ∈ C.
If s 6= 0, then
Φ(x ⊗ f )Φ(A) = tI and x ⊗ f A = 0, impossible. So AB = 0 implies Φ(A)Φ(B) =
0.
If s = 0, we get Φ(x ⊗ f ) = 0, impossible.
So from Theorem 1.2, Φ has the form in Theorem 1.1.
(cid:3)
References
1. Z.F. Bai, J.C. Hou, Numerical radius distance preserving maps on B(H), Proceedings of
the American Mathematical Society, 132 (2004), no. 5, 1453 -- 1461.
2. M. Bendaoud, A. Benyouness, M. Sarih, et al, Preservers of radial unitary similarity func-
tions on products of operators, Linear Algebra and Its Applications, 542 (2017), no. 5,
484-500.
kPj=1
16
YANFANG ZHANG, and XIAOCHUN FANG
3. K. Chan, c -Numerical radius isometries on matrix algebras and triangular matrix algebras,
Linear Algebra and Its Applications, 466 (2015), no. 2, 160-181.
4. C.Q. Chen, F.F. Lu, Maps that preserve higher-dimensional numerical ranges with operator
Jordan products, Linear and Multilinear Algebra, 132 (2017), no. 12, 2530C2537.
5. J.L. Cui, J.C. Hou, Maps leaving functional values of operator products invariant, Linear
Algebra and Its Applications, 428 (2008), no. 7, 1649-1663.
6. M.T. Chien, The c -numerical range of a rank-one matrix, Applied Mathematics Letters,
14 (2001), no.2, 167-170.
7. M.T. Chien, et al, On geometric properties of the numerical range, Linear Algebra and Its
Applications, 274 (1998), no. 1, 389-410.
8. K.E. Gustafson, D.K.M. Rao, Numerical Range, Springer New York, 1997.
9. H.L. Gau, C.K. Li, C*-isomorphisms, Jordan isomorphisms, and numerical range preserv-
ing maps, Proceedings of the American Mathematical Society, 135 (2007), no. 9, 2907-2914.
10. P.R. Halmos, A Hilbert Space Problem Book, Grad,Texts in Math., Spring-Verlag, New
York, 1982.
11. J.C.Hou, Rank-preserving linear maps on B(H), Sci. in China (Ser.A), 32 (1989), no. 15,
929-940.
12. J.C. Hou, C.K. Li, X.F. Qi, Numerical Range of Lie Product of Operators, Integral Equa-
tions and Operator Theory, 83 (2015), no. 4, 497-516.
13. J.C. Hou, Q.H. Di, Maps preserving numerical ranges of operator products, Proceedings of
the American Mathematical Society, 134 (2006), no. 5, 1435-1446.
14. L.L. Liu, Linear transformations preserving log-convexity, Ars Combinatoria -Waterloo then
Winnipeg-, 100 (2011), no. 11, 473-483.
15. C.K. Li, N.K.Tsing, Linear operators that preserve the c-numerical range or radius of ma-
trices, Linear and Multilinear Algebra, 23 (1988), no. 1 , 27-46.
16. C.K. Li, The c-spectral, c-radial and c-convex matrices, Linear and Multilinear Algebra, 15
(1986), no. 1, 5-15.
17. C.K. Li, N.K. Tsing, N.S.Sze , Elliptical range theorems for generalized numerical ranges
of quadratic operators, Rocky Mountain Journal of Mathematics, 41 (2011), no.3, 813-832.
18. L. Molnr, A generalization of Wigner's unitary-antiunitary theorem to Hilbert modules,
Journal of Mathematical Physics, 40 (1999), no. 11, 5544-5555.
19. Y.T.Poon, Another proof of a result of Westwick, Linear and Multilinear Algebra, 9 (1980),
no. 1, 181-186.
20. Westwick R, A theorem on numerical range, Linear and Multilinear Algebra, textbf2 (1975),
no. 4, 311-315.
21. T. Schulte-herbrggen, Gunther Dirr, Uwe Helmke, et al, The Significance of the C-
Numerical Range and the Local C-Numerical Range in Quantum Control and Quantum
Information, Linear and Multilinear Algebra, 56 (2008), no. 1, 3-26.
1School of Mathematical Sciences, Tongji University, Shanghai 200092, P.R.
China.
E-mail address: [email protected]
2School of Mathematical Sciences, Tongji University, Shanghai 200092, P.R.
China.
E-mail address: [email protected]
|
1807.04012 | 1 | 1807 | 2018-07-11T09:05:52 | On the existence of vectors dual to a set of linear functionals | [
"math.FA"
] | We give a simple proof of a crucial lemma that is established in [1, Lemma 2.1] by induction, and plays important roles in that paper and [2]. | math.FA | math |
ON THE EXISTENCE OF VECTORS DUAL TO A SET OF LINEAR FUNCTIONALS
SHIBO LIU
ABSTRACT. We give a simple proof of a crucial lemma that is established in [1, Lemma
2.1] by induction, and plays important roles in that paper and [2].
In a recent paper [1], Brezhneva and Tret'yakov give an elementary proof of the
Karush -- Kuhn -- Tucker Theorem (that is, [1, Theorem 1.1]) in normed linear spaces.
This theorem is about the minimization problems with a finite number of inequality
constraints. In their proof of [1, Theorem 1.1], an important step is to establish the
following lemma.
Lemma 0.1 ([1, Lemma 2.1]). Let X be a linear space. Let ξi : X → R, i = 1, . . . , n,
be linear functionals which are linearly independent. Then there exists a set {ηi} of n
linearly independent vectors in X such that the matrix
A =
· · ·
〈ξ1, η1〉
...
〈ξ1, ηn〉
...
〈ξn, η1〉
· · ·
〈ξn, ηn〉
is invertible.
This lemma also plays a crucial role in [2], where an elementary proof of the La-
grange multiplier theorem in normed linear spaces is given.
We emphasize that 0.1 is equivalent to the following seemingly stronger result.
Theorem 0.2. Let X be a vector space. Let ξi : X → R, i = 1, . . . , n, be linear functionals
which are linearly independent. Then there exist ǫ1, . . . , ǫn in X such that 〈ξi, ǫ j〉 = δi
j,
where δi
j is the Kronecker delta.
In fact, let
Then AA−1 = In, the n × n identity matrix, means
A−1 =
〈ξi, ǫ j〉 =ξi,
b1
1
...
bn
1
· · ·
· · ·
b1
n
bn
n
,
n
Xk=1
bk
j ηk =
n
Xk=1
ǫ j =
n
Xk=1
bk
j ηk
〈ξi, ηk〉bk
j =AA−1i
j
= δi
j.
Therefore, Lemma 0.1 and Theorem 0.2 are equivalent.
In [1], Lemma 0.1 is proved by induction on n. In this note, we present a much
simple proof of Theorem 0.2.
Proof of Theorem 0.2. We define a linear map ϕ : X → Rn by
ϕ(x ) =ξ1(x ), . . . , ξn(x ) .
1991 Mathematics Subject Classification. 15A03, 15A04.
Key words and phrases. linear functionals, linearly independent, linear map, perpendicular.
This work was supported by NSFC (11671331). It was done while the author was visiting the Abdus
Salam International Centre for Theoretical Physics. The author is grateful to ICTP for its hospitality.
1
2
SHIBO LIU
We claim that ϕ is surjective. Otherwise, ϕ(X ) is a proper subspace of Rn, there exists
λ ∈ Rn\ {0} which is perpendicular to ϕ(X ). Let λ = (λ1, . . . , λn), then for all x ∈ X ,
0 = λ · ϕ(x ) =
λiξi(x ).
n
Xi=1
Hence the set of linear functionalsξi is linearly dependent, a contradiction.
Let e j = (0, . . . , 1, . . . , 0) be the j-th standard base vector of Rn and take ǫ j ∈ ϕ−1(e j),
then
It follows that ξi(ǫ j) = δi
j.
ξ1(ǫ j), . . . , ξn(ǫ j) = ϕ(ǫ j) = e j.
REFERENCES
(cid:3)
[1] O. Brezhneva, A. A. Tret'yakov, An elementary proof of the Karush-Kuhn-Tucker theorem in normed linear spaces for problems with a finite number of inequality constraints,
Optimization 60 (2011) 613 -- 618.
[2] O. Brezhneva, A. A. Tret'yakov, An elementary proof of the Lagrange multiplier theorem in normed linear spaces,
Optimization 61 (2012) 1511 -- 1517.
SCHOOL OF MATHEMATICAL SCIENCES, XIAMEN UNIVERSITY, XIAMEN 361005, CHINA
E-mail address: [email protected]
|
1703.02792 | 2 | 1703 | 2017-05-29T08:27:58 | On sum of two subnormal kernels | [
"math.FA"
] | We show, by means of a class of examples, that if $K_1$ and $K_2$ are two positive definite kernels on the unit disc such that the multiplication by the coordinate function on the corresponding reproducing kernel Hilbert space is subnormal, then the multiplication operator on the Hilbert space determined by their sum $K_1+K_2$ need not be subnormal. This settles a recent conjecture of Gregory T. Adams, Nathan S. Feldman and Paul J. McGuire in the negative. We also discuss some cases for which the answer is affirmative. | math.FA | math |
ON SUM OF TWO SUBNORMAL KERNELS
SOUMITRA GHARA AND SURJIT KUMAR
Abstract. We show, by means of a class of examples, that if K1 and K2 are two positive
definite kernels on the unit disc such that the multiplication by the coordinate function on the
corresponding reproducing kernel Hilbert space is subnormal, then the multiplication operator
on the Hilbert space determined by their sum K1 + K2 need not be subnormal. This settles
a recent conjecture of Gregory T. Adams, Nathan S. Feldman and Paul J. McGuire in the
negative. We also discuss some cases for which the answer is affirmative.
1. Introduction
Let H be a complex separable Hilbert space. Let B(H) denote the Banach algebra of bounded
linear operators on H. Recall that an operator T in B(H) is said to be subnormal if there exists
a Hilbert space K ⊃ H and a normal operator N in B(K) such that N (H) ⊂ H and NH = T.
For the basic theory of subnormal operators, we refer to [12].
Completely hyperexpansive operators were introduced in [6]. An operator T ∈ B(H) is said
to be completely hyperexpansive if
n
(−1)j(cid:18)n
Xj=0
j(cid:19)T ∗jT j ≤ 0
(n ≥ 1).
The theory of subnormal and completely hyperexpansive operators are closely related with
the theory completely monotone and completely alternating sequences (cf. [5], [6] ).
Let Z+ denote the set of non-negative integers. A sequence {ak}k∈Z+ of positive real numbers
is said to be a completely monotone if
n
(−1)j(cid:18)n
Xj=0
j(cid:19)am+j ≥ 0
(m, n ≥ 0).
It is well-known that a sequence {ak}k∈Z+ is completely monotone if and only if the sequence
{ak}k∈Z+ is a Hausdorff moment sequence, that is, there exists a positive measure ν supported
in [0, 1] such that ak =R[0,1] xkdν(x) for all k ∈ Z+ (cf. [9]).
Similarly, a sequence {ak}k∈Z+ of positive real numbers is said to be completely alternating
if
n
(−1)j(cid:18)n
Xj=0
j(cid:19)am+j ≤ 0
(m ≥ 0, n ≥ 1).
2010 Mathematics Subject Classification: Primary 46E20, 46E22; Secondary 47B20, 47B37
Key words and phrases: Completely alternating, Completely hyperexpansive, Completely monotone, Positive
definite kernel, Spherically balanced spaces, Subnormal operators
Work of the first author was supported by CSIR SPM Fellowship and work of the second author was supported
by Inspire Faculty Fellowship.
1
ON SUM OF TWO SUBNORMAL KERNELS
2
Note that {ak}k∈Z+ is completely alternating if and only if the sequence {∆ak}k∈Z+ is completely
monotone, where ∆ak := ak+1 − ak.
open unit disc D := {z ∈ C : z < 1} with reproducing kernel K(z, w) = Pk∈Z+
Let H(K) be a reproducing kernel Hilbert space consisting of holomorphic functions on the
ak(z ¯w)k.
Thus K is holomorphic in the first variable and anti-holomorphic in the second. We make
the assumption that a kernel function is always holomorphic in the first variable and anti-
holomorphic in the second throughout this note. Consider the operator Mz of multiplication by
the coordinate function z on H(K). As is well-known, such a multiplication operator is unitarily
equivalent to a weighted shift operator W on the sequence space
xk2
ak
ℓ2(a) := {x := (x0, x1, . . .) : kxk2 =
< ∞}
∞
Xk=0
an+1
with W en = q an
en+1, where en is the standard unit vector. Assume that Mz is bounded.
Then Mz is a contractive subnormal if and only if the sequence { 1
ak }k∈Z+ is a Hausdorff moment
sequence. On the other hand, Mz on H(K) is completely hyperexpansive if and only if the
sequence { 1
ak }k∈Z+ is completely alternating (cf. [6, Proposition 3]).
For any two positive definite kernels K1 and K2, their sum K1 +K2 is again a positive definite
kernel and therefore determines a Hilbert space H(K1 + K2) of functions. It was shown in [4]
that
H(K1 + K2) = {f = f1 + f2 : f1 ∈ H(K1), f2 ∈ H(K2)},
H(K1+K2) := infnkf1k2
is a Hilbert space with the norm given by
kfk2
H(K1) + kf2k2
Sum of two kernel functions is also discussed by Salinas in [19]. He proved that if K1 and K2
are generalized Bergman kernels (for definition, refer to [16] ), then so is K1 + K2. Although
not explicitly stated in [4], it is not hard to verify that the multiplication operator Mz on
H(K1 + K2) is unitarily equivalent to the operator PM⊥(Mz,1 ⊕ Mz,2)M⊥ , where Mz,i is the
operator of multiplication by the coordinate function z on H(Ki) and
H(K2) : f = f1 + f2, f1 ∈ H(K1), f2 ∈ H(K2)o .
M = {(g,−g) ∈ H(K1) ⊕ H(K2) : g ∈ H(K1) ∩ H(K2)} ⊆ H(K1) ⊕ H(K2).
Evidently, if Mz,1 and Mz,2 are subnormal, then so is Mz,1 ⊕ Mz,2. Here, we discuss the subnor-
mality of the compression PM⊥(Mz,1 ⊕ Mz,2)M⊥ for a class of kernels. In particular, we show
that the subnormality of Mz,1 and Mz,2 need not imply PM⊥(Mz,1 ⊕ Mz,2)M⊥ is subnormal.
A similar question on subnormality involving the point-wise product of two positive definite
kernels was raised in [19]. Recall that the product K1K2 of two positive definite kernels defined
on, say the unit disc D, is also a positive definite kernel on D. Indeed, if H(K1) ⊗ H(K2) ⊆
Hol(D × D) be the usual tensor product of the two Hilbert spaces H(K1) and H(K2), and N is
the subspace {h ∈ H(K1)⊗H(K2) : h(z, z) = 0, z ∈ D}, then the operator Mz,1⊗I acting on the
Hilbert space H(K1) ⊗ H(K2) compressed to N⊥ is unitarily equivalent to the multiplication
operator Mz on the Hilbert space H(K1K2). If Mz,1 is subnormal on H(K1), then so is the
operator Mz,1 ⊗ I.
ON SUM OF TWO SUBNORMAL KERNELS
3
The answer to the question of subnormality, both in the case of the sum as well as the
product, is affirmative in several examples. For over thirty years, the question of whether the
compression to N⊥ is subnormal had remained open. Recently, a counter-example has been
found, see [3, Theorem 1.5]. The conjecture below is similar except that it involves the sum of
two kernels.
Conjecture 1.1. ([1, pp. 22]). Let K1(z, w) =Pk∈Z+
be any two reproducing kernels satisfying:
ak(z ¯w)k and K2(z, w) =Pk∈Z+
bk(z ¯w)k
= lim bk
bk+1
(a) lim ak
ak+1
(b) lim ak = lim bk = ∞
(c)
1
ak
= 1
=R[0,1] tkdν1(t) and 1
bk
=R[0,1] tkdν2(t) for all k ∈ Z+.
Then the multiplication operator Mz on H(K1 + K2) is a subnormal operator.
An equivalent formulation, in terms of the moment sequence criterion, of the conjecture is
bk }k∈Z+ are Hausdorff moment sequences, does it necessarily
ak }k∈Z+ and { 1
ak+bk }k∈Z+ is also a Hausdorff moment sequence?
the following. If { 1
follow that {
1
Like the case of the product of two kernels, here we give a class of counter examples to the
conjecture stated above. Paul McGuire, in a private communication, has informed the authors
of an example that they had found. In fact, he says that their example is one of the examples
discussed in this note.
These two cases suggest that it may be fruitful to ask when the compression of a subnormal
operator to an invariant subspace is again subnormal.
The paper is organized as follows. In section 2, we provide a class of counter-examples which
settles the Conjecture 1.1. We also discuss some cases for which answer to this conjecture is
affirmative.
In the last section, we try to answer analogously in a certain class of weighted
multi-shifts.
2. Sum of two subnormal reproducing kernels need not be subnormal
For the construction of counter-examples to the conjecture, we make use of the following
result, borrowed from [2, Proposition 4.3].
Proposition 2.1. For distinct positive real numbers a0, ..., an and non-zero real numbers b0, ..., bn,
consider the polynomial p(x) = Πn
p(l)}l∈Z+
is never a Hausdorff moment sequence.
k=0(x + ak + ibk)(x + ak − ibk). Then the sequence { 1
For r > 0, let Kr be a positive definite kernel given by
Kr(z, w) := Xk∈Z+
k + r
r
(z ¯w)k
(z, w ∈ D).
The case r = 1 corresponds to the Bergman kernel. It is easy to see that the multiplication
operator Mz on H(Kr) is a contractive and subnormal and the representing measure is rxr−1dx.
ON SUM OF TWO SUBNORMAL KERNELS
4
For s, t > 0, consider the multiplication operator Mz on H(Ks,t), where
Ks,t(z, w) := Xk∈Z+
(k + s)(k + t)
st
(z ¯w)k
(z, w ∈ D).
The case s = 1 and t = 2, corresponds to the kernel (1 − z ¯w)−3. Note that Mz is a contractive
subnormal with the representing measure ν is given by
dν(x) = (cid:26) −s2xs−1 log x dx if s = t
if s 6= t.
st xt−1−xs−1
dx
s−t
One easily verifies that Kr and Ks,t both satisfy all the conditions (a), (b) and (c) of the
Conjecture 1.1. But the multiplication operator on their sum need not be subnormal for all
possible choices of s, t > 0. This follows from the following theorem.
Theorem 2.2. The multiplication operator Mz on H(Kr + Ks,t) is subnormal if and only if
(2.1)
Proof. Notice that
(rs + st + tr)2 ≥ 8r2st.
(Kr + Ks,t)(z, w) = Xk∈Z+
k2 + (s + t + st
r )k + 2st
st
(z ¯w)k
(z, w ∈ D).
The roots of the polynomial x2 + (s + t + st
r )x + 2st are
:= −(s + t + st
:= −(s + t + st
r ) +q(s + t + st
r ) −q(s + t + st
2
2
r )2 − 8st
r )2 − 8st
and
.
x1
x2
Suppose that (rs + st + tr)2 ≥ 8r2st. Then the kernel Kr + Ks,t will be of the form 2Ks′,t′ where
s′ = −x1 and t′ = −x2. Hence, Mz on H(Kr + Ks,t) is a subnormal operator.
Conversely, assume that (rs + st + tr)2 < 8r2st. By Proposition 2.1, it follows that Mz on
(cid:3)
H(Kr + Ks,t) can not be subnormal.
Remark 2.3 : If we choose s = 1, t = 2 and r > 2, then the inequality (2.1) is not valid.
We also point out that if K1 and K2 are any two reproducing kernels such that the multi-
plication operators on H(K1) and H(K2) are hyponormal, then the multiplication operator on
H(K1 + K2) need not be hyponormal. An example illustrating this is given below.
Example 2.4. For any s, t > 0, consider the reproducing kernel K s,t given by
K s,t(z, w) := 1 + sz ¯w + s2(z ¯w)2 + t
(z ¯w)3
1 − z ¯w
.
Note that K s,t defines a reproducing kernel on the unit disc D and the multiplication operator
t , 1, 1,··· ).
Mz on H(K s,t) can be seen as a weighted shift operator with weights (q 1
s ,q 1
Thus, it follows that Mz on H(K s,t) is hyponormal if and only if s2 ≤ t ≤ s3.
s ,q s2
ON SUM OF TWO SUBNORMAL KERNELS
5
Observe that Mz on H(K s,t + K s′,t′
s+s′ ,q s+s′
(q 2
necessary that
s2+s′2 ,q s2+s′2
s+s′ ≤ s+s′
2
) can be realized as a weighted shift operator with weights
, 1, 1,··· ). For the hyponormality of this weighted shift operator, it is
t+t′
s2+s′2 , which is true only when s = s′.
We remark that this is different from the case of the product of two kernels, where, the
hyponormality of the multiplication operator on the Hilbert space H(K1K2) follows as soon as
we assume they are hyponormal on the two Hilbert spaces H(K1) and H(K2), see [10].
If T ∈ B(H) is left invertible then the operator T ′ given by T ′ = T (T ∗T )−1 is said to be the
operator Cauchy dual to T. The following result has been already recorded in [6, Proposition
6], which may be paraphrased as follows:
Theorem 2.5. Let K(z, w) =Pk∈Z+
ak(z ¯w)k be a positive definite kernel on D and Mz be the
multiplication operator on H(K). Assume that Mz is left invertible. Then the followings are
equivalent:
(i) {ak}k∈Z+ is a completely alternating sequence.
(ii) The Cauchy dual M′z of Mz is completely hyperexpansive.
(iii) For all t > 0, {
(iv) For all t > 0, the multiplication operator Mz on H(tK + (1− t)S) is contractive subnor-
t(ak−1)+1}k∈Z+ is a completely monotone sequence.
1
mal, where S is the Szego kernel on D.
Remark 2.6 : If {ak}k∈Z+ is a completely alternating sequence then by putting t = 1 in part
(iii) of Theorem 2.5, it follows that { 1
bk(z ¯w)k be any two
Corollary 2.7. Let K1(z, w) = Pk∈Z+
reproducing kernels such that {ak}k∈Z+ and {bk}k∈Z+ are completely alternating sequences, then
the multiplication operator Mz on H(K1 + K2) is subnormal.
ak(z ¯w)k and K2(z, w) = Pk∈Z+
ak }k∈Z+ is a completely monotone sequence.
Proof. It is easy to verify that the sum of two completely alternating sequences is completely
alternating. The desired conclusion follows immediately from Remark 2.6.
(cid:3)
Remark 2.8 : Note that { k+r
{ (k+s)(k+t)
r }k∈Z+ is a completely alternating sequence but the sequence
}k∈Z+ is not completely alternating. So, the reproducing kernels Kr and Ks,t dis-
st
cussed in Theorem 2.2 does not satisfy the hypothesis of Corollary 2.7.
Proposition 2.9. Let K(z, w) = Pk∈Z+
ak(z ¯w)k be any positive definite kernel such that the
multiplication operator Mz on H(S + K) is subnormal. Then the multiplication operator on
H(K) is subnormal.
Proof. From subnormality of Mz on H(S + K), it follows that { 1
monotone sequence. Thus, {1 − 1
1+ak}k∈Z+ is completely alternating. Note that
1+ak }k∈Z+ is a completely
(ak)−1 = (1 + ak)−1(1 − 1
1+ak
)−1
=
∞
Xj=1(cid:16) 1
1+ak(cid:17)j
.
ON SUM OF TWO SUBNORMAL KERNELS
6
Observe that {
(1+ak)j }k∈Z+ is a completely monotone sequence for all j ≥ 1. Now, being the
limit of completely monotone sequences, {a−1
k }k∈Z+ is completely monotone.
1
(cid:3)
Remark 2.10 : We have following remarks:
(i) The converse of the Proposition 2.9 is not true (see the example discussed in part (ii) of
the Remark 2.13).
(ii) If we replace Szego kernel S by Bergman kernel then the conclusion of the Proposition
2.9 need not be true. For example, by using Proposition 2.1, one may choose α > 0
k2+αk+1}k∈Z+ is not completely monotone but the sequence
such that the sequence {
{
k2+(α+1)k+2}k∈Z+ is completely monotone.
1
1
We use the convenient Pochhammer symbol given by (x)n := Γ(x+n)
Γ(x)
, where Γ denotes the
gamma function.
For λ, µ > 0, consider the positive definite kernel
Kλ,µ(z, w) = Xk∈Z+
(λ)k
(µ)k
(z ¯w)k
(z, w ∈ D).
It is easy to see that the case µ = 1 corresponds to the kernel (1 − z ¯w)−λ. Note that the
multiplication operator Mz on H(Kλ,µ) may be realized as a weighted shift operator with
weight sequence {q k+µ
First part of the following theorem is proved in [15] and the representing measure is given in
k+λ}k∈Z+.
[13, Lemma 2.2]. Here, we provide a proof for the second part only.
Theorem 2.11. The multiplication operator Mz on H(Kλ,µ) is
(i) subnormal if and only if λ ≥ µ. In this case, the representing measure ν of Mz is given
by
dν(x) = (
Γ(λ)
Γ(µ)Γ(λ−µ) xµ−1(1 − x)λ−µ−1dx if λ > µ
if λ = µ,
δ1(x)dx
where δ1 is the Dirac delta function.
(ii) completely hyperexpansive if and only if λ ≤ µ ≤ λ + 1.
Proof. The multiplication operator Mz on H(Kλ,µ) is completely hyperexpansive if and only if
the sequence { (µ)k
(λ)k }k∈Z+ is completely alternating. Here
∆(cid:18) (µ)k
(λ)k(cid:19) =
(µ)k+1
(λ)k+1 −
(µ)k
(λ)k
=
=
(µ)k+1 − (µ)k(λ + k − 1)
(λ)k+1
(µ)k
.
(λ + 1)k
µ − λ
λ
By first part of this theorem, { µ−λ
if λ ≤ µ ≤ λ + 1. This completes the proof.
λ
(µ)k
(λ+1)k }k∈Z+ is a completely monotone sequence if and only
(cid:3)
ON SUM OF TWO SUBNORMAL KERNELS
7
The following proposition gives a sufficient condition for the subnormality of multiplication
operator on Hilbert space determined by sum of two kernels belonging to the class Kλ,µ.
Proposition 2.12. Let 0 < µ ≤ λ′ ≤ λ ≤ λ′ + 1. Then the multiplication operator Mz on
H(Kλ,µ + Kλ′,µ) is contractive subnormal.
Proof. Observe that
(Kλ,µ + Kλ′,µ)(z, w) = Xk∈Z+
(λ)k + (λ′)k
(µ)k
(z ¯w)k
(z, w ∈ D)
and
(µ)k
(λ)k + (λ′)k
=
(µ)k
(λ′)k
1
1 + (λ)k
(λ′)k
.
Since µ ≤ λ′, it follows from part (i) of Theorem 2.11 that { (µ)k
If λ′ ≤ λ ≤ λ′ + 1, then by part (ii) of Theorem 2.11, the sequence { (λ)k
(λ′)k}k∈Z+ is completely monotone.
(λ′)k}k∈Z+ is a completely
alternating sequence. So is the sequence {1 + (λ)k
(λ′)k }k∈Z+. Hence, by Remark 2.6,( 1
1+
(λ′)k)k∈Z+
(λ)k
is a completely monotone sequence. Thus, being a product of two completely monotone se-
quences, {
Remark 2.13 : Here are some remarks:
(λ)k+(λ′)k}k∈Z+ is a completely monotone sequence. This completes the proof.
(µ)k
(cid:3)
(i) The case when µ < λ′ and λ = λ′ + 1. The representing measure for the sequence
( 1
1+
(λ′)k)k∈Z+
(λ)k
is λ′x2λ′−1dx. The representing measure for the sequence { (µ)k
(λ′)k }k∈Z+ is
given in part (i) of Theorem 2.11. Thus, using Remark 2.4 of [2], one may obtain the
representing measure for Mz on H(Kλ,µ + Kλ′,µ) to be given by
dν(x) =
λ′Γ(λ′)
Γ(µ)Γ(λ′ − µ)
x2λ′−1(cid:18)Z 1−x
0
tλ′−µ−1(1 − t)µ−2λ′−1dt(cid:19) dx.
But in general, when λ < λ′ + 1, we do not know the representing measure for the
sequence ( 1
1+
(λ′)k)k∈Z+
(λ)k
as well as for the sequence n (µ)k
(ii) Consider the kernel (K1,1+K3,1)(z, w) =Pk∈Z+
from Proposition 2.1 that the sequence n
k2+3k+4ok∈Z+
2
2
k2+3k+4
(λ)k+(λ′)kok∈Z+.
(z ¯w)k for all z, w ∈ D. It follows
is not completely monotone.
Consequently, the multiplication operator Mz on H(K1,1 + K3,1) is not subnormal.
(1)k
For λ > 1, consider the kernel Kλ,1 + K3,1. We claim that there exists a λ0 > 1 such
is not completely monotone. If not, assume that it is a completely
(λ0)k+(3)kok∈Z+
thatn
monotone sequence for all λ > 1. As λ goes to 1, one may get that n
completely monotone, which is a contradiction. Therefore, we conclude that there exists
a λ0 > 1 such that the multiplication operator Mz on H(Kλ0,1 + K3,1) is not subnormal.
By using properties of gamma function, one may verify that Kλ0,1 and K3,1 both satisfy
k2+3k+4ok∈Z+
is
2
ON SUM OF TWO SUBNORMAL KERNELS
8
(a), (b) and (c) of the Conjecture 1.1. This also provides a class of counterexamples for
the Conjecture 1.1.
ap
k(z ¯w)k and K2(z, w) =
aq
k(z ¯w)k are any two reproducing kernels such that {ak}k∈Z+ is a completely alternating
Proposition 2.14. Let 0 < p ≤ q ≤ p + 1. Suppose K1(z, w) =Pk∈Z+
Pk∈Z+
sequence. Then the multiplication operator Mz on H(K1 + K2) is subnormal.
Proof. Note that
1
ap
k + aq
k
=
1
ap
k(1 + aq−p
k
)
.
k
}k∈Z+ is also completely alternating. Thus so is {1 + aq−p
Since 0 ≤ q − p ≤ 1 and {ak}k∈Z+ is completely alternating, it follows from [7, Corollary 1]
that {aq−p
}k∈Z+ . Hence, by Remark
2.6, {(1 + aq−p
k }k∈Z+ is completely
monotone. Now the proof follows as the product of two completely monotone sequences is also
completely monotone.
)−1}k∈Z+ is completely monotone. By [8, Corollary 4.1], {a−p
(cid:3)
k
k
Example 2.15. For any p > 0, let Kp(z, w) be the positive definite kernel given by
Kp(z, w) := Xk∈Z+
(k + 1)p(z ¯w)k
(z, w ∈ D).
Then it is known that the multiplication operator Mz on H(Kp) is subnormal with the repre-
senting measure dν(x) = (− log x)p−1
dx (cf. [14, Theorem 4.3]). By Proposition 2.14, it follows
that Mz on H(Kp + Kq) is subnormal if p ≤ q ≤ p + 1.
Γ(p)
The next result also provides a class of counter-examples to the Conjecture 1.1.
Theorem 2.16. Consider the positive definite kernel Kp given in Example 2.15. Then the
multiplication operator Mz on H(Kp + Kp+2) is subnormal if and only if p ≥ 1.
Proof. For x ∈ (0, 1], let g(x) := 1
1
(− log x − y)p−1 sin y dy and dν(x) = g(x)dx. Then
0
xkdν(x) =
Z 1
0
Γ(p)R − log x
Γ(p)Z ∞
Γ(p)Z ∞
Γ(p)Z ∞
y=0
1
1
1
x=0
y=0Z e−y
e−(k+1)yZ ∞
e−(k+1)y Γ(p)
u=0
y=0
(k + 1)p
(k + 1)2 + 1
1
.
=
=
=
1
xk(− log x − y)p−1dx sin y dy
e−(k+1)uup−1du sin y dy
(k + 1)p sin y dy
1
(k+1)p
Thus the sequence {
(k+1)2+1}k∈Z+ is completely monotone if and only if the function g(x)
is non-negative a.e. Note that the function g(x) is non-negative on (0, 1] a.e. if and only if the
function h(x) := g(e−x) is non-negative a.e. on (0,∞). Now
Γ(p)Z 1
Γ(p)Z x
h(x) =
xp
1
0
0
By [17, Chapter 3, pp 439], we have h(x) =
function of first kind. Thus, the sequence {
(k+1)p
1
2
1
2 ,
1
(x) is the Lommel's
p−
(k+1)2+1}k∈Z+ being completely monotone is
p−
1
2 ,
1
2
(x − y)p−1 sin y dy =
√x
Γ(p) s
1
(1 − y)p−1sin(xy) dy.
(x), where s
ON SUM OF TWO SUBNORMAL KERNELS
9
equivalent to the non-negativity of the function s
(x). If p ≥ 1 then by [20, Theorem A],
(x) ≥ 0 for all x > 0. The converse follows from [20, Theorem 2], which
p−
1
2 ,
1
2
we get that s
1
2 ,
1
2
p−
completes the proof.
3. Multi-variable case
(cid:3)
Let Zd
+ denote the cartesian product Z+ × ··· × Z+ (d times). Let α = (α1,··· , αd) ∈ Zd
+,
we write α := α1 + ··· + αd and α! = α1!··· αd!.
If T = (T1,··· , Td) is a d-tuple of commuting bounded linear operators Tj (1 ≤ j ≤ d) on H
then we set T ∗ to be (T ∗1 ,··· , T ∗d ) and T α to be T α1
1
··· T αd
d .
Given a commuting d-tuple T of bounded linear operators T1,··· , Td on H, set
QT (X) :=
d
Xi=1
T ∗i XTi
(X ∈ B(H)).
For X ∈ B(H) and k ≥ 1 , one may define Qk
Recall that T is said to be
T (X) := QT (Qk−1
T
(X)), where Q0
T (X) = X.
(i) spherical contraction if QT (I) ≤ I.
(ii) jointly left invertible if there exists a positive number c such that QT (I) ≥ cI.
For a jointly left invertible T, the spherical Cauchy dual T s of T is the d-tuple (T s
1 , T s
d ),
2 ··· , T s
i := Ti(QT (I))−1 (i = 1, 2,··· , d). We say that T is a joint complete hyperexpansion if
where T s
Bn(T ) :=
n
(−1)k(cid:18)n
Xk=0
k(cid:19)Qk
T (I) ≤ 0
(n ≥ 1).
Throughout this section B denotes the open unit ball {z ∈ Cd : z12 + ··· + zd2 < 1} and
be a multi-sequence of positive numbers. Consider the Hilbert space H 2(β) of
∂B denotes the unit sphere {z ∈ Cd : z12 + ··· + zd2 = 1} in Cd.
Let {βα}α∈Zd
formal power series f (z) =Pα∈Zd
kfk2
f (α)zα such that
f (α)2β2
α < ∞.
+
+
H 2(β) = Xα∈Zd
+
The Hilbert space H 2(β) is said to be spherically balanced if the norm on H 2(β) admits the
slice representation [ν, H 2(γ)], that is, there exist a Reinhardt measure ν and a Hilbert space
H 2(γ) of formal power series in one variable such that
kfk2
H 2(β) =Z∂B kfzk2
H 2(γ)dν(z)
(f ∈ H 2(β)),
where γ = {γk}k∈Z+ is given by the relation βα = γαkzαkL2(∂B,ν) for all α ∈ Zd
+. Here, by
the Reinhardt measure, we mean a Td-invariant finite positive Borel measure supported in ∂B,
where Td denotes the the unit d-torus {z ∈ Cd : z1 = 1,··· ,zd = 1}. For more details on
spherically balanced Hilbert spaces, we refer to [11].
ON SUM OF TWO SUBNORMAL KERNELS
10
The following lemma has been already recorded in [11, Lemma 4.3]. We include a statement
for ready reference.
Lemma 3.1. Let H 2(β) be a spherically balanced Hilbert space and let [ν, H 2(γ)] be the slice rep-
resentation for the norm on H 2(β). Consider the d-tuple Mz = (Mz1,··· , Mzd) of multiplication
by the co-ordinate functions z1,··· , zd on H 2(β). Then for every n ∈ Z+ and α ∈ Zd
+,
L2(∂B,ν).
n
n
hBn(Mz)zα, zαi =
Mz (I)zα, zαi =
(−1)k(cid:18)n
Xk=0
k(cid:19)γ2
k+αkzαk2
(−1)k(cid:18)n
k(cid:19)hQk
Xk=0
If the interior of the point spectrum σp(M∗z ) of M∗z is non-empty then H 2(β) may be realized
as a reproducing kernel Hilbert space H(K) [18, Propositions 19 and 20], where the reproducing
kernel K is given by
K(z, w) = Xα∈Zd
+
zα ¯wα
β2
α
(z, w ∈ σp(M∗z )).
This has lead to the following definition.
Definition 3.2 : Let H(K) be a reproducing kernel Hilbert space defined on the open unit ball
aαzα ¯wα for all z, w ∈ B. We say that K is balanced
kernel if H(K) is a spherically balanced Hilbert space. Further, the multiplication d-tuple Mz
on H(K) may be called as balanced multiplication tuple.
B with reproducing kernel K(z, w) =Pα∈Zd
+
Remark 3.3 : The spherical Cauchy dual M s
tuple Mz can be seen as a multiplication d-tuple M s
co-ordinate functions z1,··· , zd on H 2(β s), where
kzαkL2(∂B,ν)
β s
α =
1
γα
(α ∈ Zd
+).
z of a jointly left invertible balanced multiplication
zd) of multiplication by the
z = (M s
z1,··· , M s
In other words, the norm on H 2(β s) admits the slice representation [ν, H 2(γ′)], where γ′k = 1/γk
for all k ∈ Z+.
bαzα ¯wα are any two
Proposition 3.4. If K1(z, w) = Pα∈Zd
balanced kernels with the slice representations [ν, H 2(γ1)] and [ν, H 2(γ2)] respectively. Then
K1 + K2 is a balanced kernel with the slice representation [ν/2, H 2(γ)], where γ = {γk} is given
by the relation
aαzα ¯wα and K2(z, w) = Pα∈Zd
+
+
γk =
√2γk,1γk,2
(γ2
k,1 + γ2
k,2)1/2
(k ∈ Z+).
Proof. For every α ∈ Zd
+, we have
aα + bα =
Therefore
1
γ2
α,1kzαk2
L2(∂B,ν)
+
1
γ2
α,2kzαkL2(∂B,ν)
=
α,1 + γ2
(γ2
α,2kzαk2
γ2
α,1γ2
α,2)
L2(∂B,ν)
.
kzαk2
H(K1+K2) =
α,1γ2
2γ2
α,2
α,1 + γ2
(γ2
α,2)kzαk2
L2(∂B,ν/2) = γ2
αkzαk2
L2(∂B,ν/2)
ON SUM OF TWO SUBNORMAL KERNELS
11
for all α ∈ Zd
immediately.
+. Since {zα}α∈Zd
+
forms an orthogonal subset of L2(∂B, ν/2), the conclusion follows
(cid:3)
Remark 3.5 : The conclusion of the Proposition 3.4 still holds even if we chose two different
Reinhardt measures ν1 and ν2 in the slice representations of K1 and K2, such that for some
sequence of positive real numbers {hk}k∈Z+, kzαkL2(∂B,ν1) = hαkzαkL2(∂B,ν2) for all α ∈ Zd
+.
For every j = 1, 2, it is easy to verify that
d
Xi=1
kzα+εik2
kzαk2
L2(∂B,νj)
L2(∂B,νj )
= 1.
This implies that {hk}k∈Z+ is a constant sequence, say c. Now, a routine argument, using the
Stone-Weierstrass theorem, we conclude that µ1 = c2µ2.
Let Kν denote the class of all balanced kernels with the following properties:
(i) For all K ∈ Kν, the norm on H(K) admits the slice representations with fixed Reinhardt
measure ν.
(ii) For every member K of Kν, the multiplication operator Mz defined on H(K) is jointly
left invertible.
(iii) The Cauchy dual tuple M s
z of is a joint complete hyperexpansion.
Lemma 3.6. For every member K of Kν, the multiplication operator tuple Mz defined on H(K)
is a subnormal spherical contraction.
Proof. Let K ∈ Kν and [ν, H 2(γ)] be the slice representation for the norm on H(K). Note
z of Mz is a balanced multiplication tuple with slice representation
that the Cauchy dual M s
[ν, H 2(1/γ)] (see Remark 3.3). Since M s
z is a joint complete hyperexpansion. It follows from
Lemma 3.1 that {1/γ2
k}k∈Z+ is a completely alternating sequence. Therefore, by Remark 2.6,
{γ2
k}k∈Z+ is completely monotone sequence. Now again by applying Lemma 3.1, we conclude
that the multiplication operator Mz is a subnormal spherical contraction.
(cid:3)
Theorem 3.7. If K1 and K2 are any two members of Kν then the multiplication operator Mz
on H(K1 + K2) is a subnormal spherical contraction.
k,1γ2
γ2
k,2
k,1+γ2
γ2
Proof. Note that the norm on H(K1 + K2) admits the slice representation [ν/2, H 2(γ)], where
γ2
for all k ∈ Z+ (see Proposition 3.4). It follows from the proof of Lemma 3.6 that
k = 2
{1/γ2
k}k∈Z+ is
a completely alternating sequence. Now the conclusion follows by imitating the argument given
in Lemma 3.6.
k,2}k∈Z+ are completely alternating. So their sum, that is, {1/γ2
k,1}k∈Z+ and {1/γ2
(cid:3)
k,2
For λ > 0, consider the positive definite kernel Kλ given by
Kλ(z, w) =
1
(1 − hz, wi)λ
The norm on H(Kλ) admits the slice representation [σ, H 2(γ)], where σ denotes the normalized
surface area measure on ∂B and γ2
for all k ∈ Z+. It is well known that the multiplication
k = (d)k
(λ)k
(z, w ∈ B).
ON SUM OF TWO SUBNORMAL KERNELS
12
operator Mz,λ on H(Kλ) is a subnormal contraction if and only if λ ≥ d. The same can also
be verified by using Lemma 3.1 and part (i) of Theorem 2.11. Similarly, by using Lemma 3.1
and part (ii) of Theorem 2.11, one may conclude that the Cauchy dual tuple M s
z,λ is a joint
complete hyperexpansion if and only if d ≤ λ ≤ d + 1. Thus, if we choose λ and λ′ are such
that d ≤ λ, λ′ ≤ d + 1. Then Kλ and Kλ′ ∈ Kσ. It now follows from Theorem 3.7 that the
multiplication operator Mz on H(Kλ + Kλ′) is subnormal. This is also included in the following
example.
Example 3.8. Let 0 < d ≤ λ′ ≤ λ ≤ λ′ + 1. Note that the norm on H(Kλ + Kλ′) admits
the slice representation [σ/2, H 2(γ)], where γ2
for all k ∈ Z+. From the proof of
Proposition 2.12, it is clear that {γ2
k}k∈Z+ is completely monotone. Hence, the multiplication
operator Mz on H(Kλ + Kλ′) is subnormal.
k = 2(d)k
(λ)k+(λ′)k
A d-tuple S = (S1,··· , Sd) of commuting bounded linear operators S1,··· , Sd in B(H) is a
spherical isometry if S∗1 S1 + ··· + S∗dSd = I. In other words, QS(I) = I. The most interesting
example of a spherical isometry is the Szego d-shift; that is, the d-tuple Mz of multiplication
operators Mz1,··· , Mzd on the Hardy space H 2(∂B) of the unit ball.
Let ν be a Reinhardt measure. Consider the multiplication d-tuple Mz on a reproducing
kernel Hilbert space H(K ν ) determined by the reproducing kernel
(3.2)
K ν(z, w) = Xα∈Zd
+
zα ¯wα
kzαk2
L2(∂B,ν)
(z, w ∈ B).
Note that Mz is a spherical isometry. In this case, the norm on H(K ν ) admits the slice repre-
sentation [ν, H 2(D)], where H 2(D) is the Hardy space of the unit disc.
Theorem 3.9. Let K ν be the reproducing kernel given as in equation (3.2) and K be any
balanced kernel with the slice representation [ν, H 2( γ)]. Assume that the multiplication operator
Mz on H(K ν + K) is subnormal. Then the multiplication operator on H( K) is subnormal.
Proof. Observe that the norm on H(K ν + K) admits the slice representation [ν/2, H 2(γ)],
k )−1 for all k ∈ Z+. Since Mz on H(K ν + K) is subnormal, it follows from
k = 2(1 + 1/γ2
where γ2
k)−1}k∈Z+ is
k}k∈Z+ is a completely monotone sequence. Hence, {(1 + 1/γ2
Lemma 3.1 that {γ2
a completely monotone sequence. If we replace ak by 1/γ2
k in the proof of the Proposition 2.9,
we get that {γ2
k}k∈Z+ is completely monotone. Now, by applying Lemma 3.1, we conclude that
the multiplication operator on H( K) is subnormal.
(cid:3)
We conclude the paper with the following questions:
Question 3.10. In view of Proposition 2.12 and Theorem 2.16, it is natural to ask that
(i) what is the necessary and sufficient condition for the multiplication operator Mz on
H(Kλ,µ + Kλ′,µ) to be subnormal?
(ii) what is the necessary and sufficient condition for the multiplication operator Mz on
H(Kp + Kq) to be subnormal?
ON SUM OF TWO SUBNORMAL KERNELS
13
Question 3.11. Let K ν be the reproducing kernel given as in equation (3.2) and K be any
positive definite kernel given by
K(z, w) := Xα∈Zd
+
aαzα ¯wα
(z, w ∈ B).
Assume that the d-tuple Mz = (Mz1,··· , Mzd) of multiplication by the co-ordinate functions
z1,··· , zd on H(K ν + K) is subnormal.
Is it necessary that the multiplication operator on
H( K) subnormal?
Acknowledgments. We express our sincere thanks to Prof. G. Misra for many fruitful conver-
sations and suggestions in the preparation of this paper. We would also like to thank Prof. S.
Chavan for his many useful comments and careful reading of the manuscript.
References
[1] G. T. Adams, N. S. Feldman and P. J. McGuire, Tridiagonal reproducing kernels and subnormality, J.
Operator Theory 70, no. 2 (2013), 477-494.
[2] A. Anand and S. Chavan, A moment problem and joint q-isometry tuples, Complex Anal. Oper. Theory 11
(2017), 785-810.
[3] A. Anand and S. Chavan, Module tensor product of subnormal modules need not be subnormal, J. Funct.
Anal. 272 (2017), 4752-4761.
[4] N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc. 68 (1950), 337-404.
[5] A. Athavale, Holomorphic kernels and commuting operators, Trans. Amer. Math. Soc. 304, no. 1 (1987),
101-110.
[6] A. Athavale, On completely hyperexpansive operators, Proc. Amer. Math. Soc. 124 (1996), 3745-3752.
[7] A. Athavale and A. Ranjekar, Bernstein functions, complete hyperexpansivity and subnormality-I, Integr.
Equ. Oper. Theory 43, no. 3 (2002), 253-263.
[8] C.
Benhida,
R. Curto
and G.
Exner, Moment
infinitely
divisible
weighted
shift,
http://homepage.divms.uiowa.edu/∼rcurto/MID.pdf
[9] C. Berg, J. P. R. Christensen and P. Ressel, Harmonic Analysis on Semigroups, Springer-Verlag, Berlin
1984.
[10] M. Badri and P. Szeptycki,Cauchy products of positive sequences Rocky Mountain J. Math. 20 (1990),
351-357.
[11] S. Chavan and S. Kumar, Spherically balanced Hilbert spaces of formal power series in several variables-I,
J. Operator Theory 72 (2014), 405-428.
[12] J. Conway, The Theory of Subnormal Operators, Math. Surveys Monographs, vol 36, Amer. Math. Soc.
Providence, RI 1991.
[13] J. Cui and Y. Duan, Berger measure for S(a,b,c,d), J. Math. Anal. Appl. 413 (2014), 202-211.
[14] R. Curto and G. Exner, Berger measure for some transformations of subnormal weighted shifts, Integr.
Equ. Oper. Theory 84, no. 3 (2016), 429-450.
[15] R. Curto, Y. Poon, and J. Yoon, Subnormality of Bergman-like weighted shifts, J. Math. Anal. Appl. 308
(2005), 334-342.
[16] R. Curto and N. Salinas, Generalized Bergman Kernels and the Cowen-Douglas Theory, Amer. J. Math.
106, no. 2 (1985), 447-488.
[17] I.S. Gradshteyn and I.M. Ryzhik, Table of integrals, series, and products, 7th edition, Elsevier Academic
Press, New York, 2007.
[18] N. Jewell and A. Lubin, Commuting weighted shifts and analytic function theory in several variables, J.
Operator Theory 1 (1979), 207-223.
[19] N. Salinas, Products of kernel functions and module tensor products, Topics in operator theory, 219-241,
Oper. Theory Adv. Appl. 32, Birkhuser, Basel, 1988.
[20] J. Steinig, The sign of Lommel's function, Trans. Amer. Math. Soc. 163 (1972), 123-129.
ON SUM OF TWO SUBNORMAL KERNELS
14
(S. Ghara) Department of Mathematics, Indian Institute of Science, Bangalore 560012, India
E-mail address: [email protected]
(S. Kumar) Department of Mathematics, Indian Institute of Science, Bangalore 560012, India
E-mail address: [email protected]
|
1001.0755 | 1 | 1001 | 2010-01-05T20:09:24 | Sharp estimates for the commutators of the Hilbert, Riesz transforms and the Beurling-Ahlfors operator on weighted Lebesgue spaces | [
"math.FA",
"math.CA"
] | We prove that the operator norm on weighted Lebesgue space L2(w) of the commutators of the Hilbert, Riesz and Beurling transforms with a BMO function b depends quadratically on the A2-characteristic of the weight, as opposed to the linear dependence known to hold for the operators themselves. It is known that the operator norms of these commutators can be controlled by the norm of the commutator with appropriate Haar shift operators, and we prove the estimate for these commutators. For the shift operator corresponding to the Hilbert transform we use Bellman function methods, however there is now a general theorem for a class of Haar shift operators that can be used instead to deduce similar results. We invoke this general theorem to obtain the corresponding result for the Riesz transforms and the Beurling-Ahlfors operator. We can then extrapolate to Lp(w), and the results are sharp for 1 < p < 1. | math.FA | math |
Sharp estimates for the commutators of the Hilbert, Riesz
transforms and the Beurling-Ahlfors operator on weighted
Lebesgue spaces
Daewon Chung
June 8, 2018
Abstract
We prove that the operator norm on weighted Lebesgue space L2(w) of the commu-
tators of the Hilbert, Riesz and Beurling transforms with a BM O function b depends
quadratically on the A2-characteristic of the weight, as opposed to the linear dependence
known to hold for the operators themselves. It is known that the operator norms of these
commutators can be controlled by the norm of the commutator with appropriate Haar
shift operators, and we prove the estimate for these commutators. For the shift operator
corresponding to the Hilbert transform we use Bellman function methods, however there
is now a general theorem for a class of Haar shift operators that can be used instead to
deduce similar results. We invoke this general theorem to obtain the corresponding result
for the Riesz transforms and the Beurling-Ahlfors operator. We can then extrapolate to
Lp(w), and the results are sharp for 1 < p < ∞. 1 2
1 Introduction
We say w is a weight if it is positive almost everywhere and locally integrable. The norm of
f ∈ Lp(w) is
kfkLp(w) :=(cid:18)ZR f (x)pw(x)dx(cid:19)1/p
.
Helson and Szegö first found necessary and sufficient conditions for the boundedness of the Hilbert
transform
Hf (x) = p.v.
1
πZ f (y)
x − y
dy
in weighted Lebesgue spaces in [11]. Hunt, Muckenhoupt, and Wheeden showed in [13] that a new
necessary and sufficient condition for boundedness of the Hilbert transform in Lp(w) is that the
weight satisfies the Ap condition, namely:
[w]Ap := sup
I hwiIhw−1/(p−1)ip−1
I < ∞ ,
(1.1)
where we denote the average over the interval I by h·iI , and we take the supremum over all intervals
in R . After one year, Coifman and Fefferman extended in [5] the result to a larger class of convolution
singular integrals with standard kernels. Recently, many authors have been interested in finding the
1Key words and phrases: Operator-weighted inequalities, Commutator, Hilbert transform, Paraproduct
22000 Mathematics Subject Classification. Primary 42A50; Secondary 47B47.
1
sharp bounds for the operator norms in terms of the Ap-characteristic [w]Ap of the weight. That is,
one looks for a function φ(x), sharp in terms of its growth, such that
kT fkLp(w) ≤ Cφ(cid:0)[w]Ap(cid:1)kfkLp(w) .
For T = M, the Hardy-Littlewood maximal function, S. Buckley [3] showed that φ(x) = x1/(p−1)
is the sharp rate of growth for all 1 < p < ∞ . He also showed in [3] that φ(x) = x2 works for
the Hilbert transform in L2(w) . S. Petermichl and S. Pott improved the result to φ(x) = x3/2 ,
for the Hilbert transform in L2(w) in [28]. More recently, S. Petermichl proved in [26] the linear
dependence, φ(x) = x, for the Hilbert transform in L2(w)
kHfkL2(w) ≤ C[w]A2kfkL2(w) ,
by estimating the operator norm of the dyadic shift. Linear bounds in L2(w) were also obtained
by O. Beznosova for the dyadic paraproduct [2]. Most recently there are new proofs for the linear
estimates in L2(w) for some operators including the Hilbert transform and the dyadic paraproduct,
[17] and [7]. The conjecture is that all the Calderón-Zygmund singular integral operator obey linear
bounds in weighted L2 . So far this is known only for the Beurling-Ahlfors operator [9], [30], the
Hilbert transform [26], Riesz transforms [27], the martingale transform [32], the square function [12],
[24] and well localized dyadic operators [15], [17], [7]. It is now also known for Calderón-Zygmund
convolutions operators that are smooth averages of well localized operators [31].
In this paper, we are interested in obtaining sharp weight inequalities for the commutators of
the Hilbert, Riesz, and Beurling transforms with multiplication by a locally integrable function
b ∈ BM O .
1.1 Commutators: main results
Commutator operators are widely encountered and studied in many problems in PDEs, and Har-
monic Analysis. One classical result of Coifman, Rochberg, and Weiss states in [6] that, for the
Calderón-Zygmund singular integral operator with a smooth kernel, [b, T ]f := bT (f ) − T (bf ) is a
bounded operator on Lp
Rn, 1 < p < ∞ , when b is a BMO function. Weighted estimates for the
commutator have been studied in [1], [20], [21], and [23]. Note that the commutator [b, T ] is more
singular than the associated singular integral operator T , in particular, it does not satisfy the cor-
responding weak (1, 1) estimate. However one can find a weaker estimate in [21]. In 1997, C. Pérez
[21] obtained the following result concerning commutators of singular integrals, for 1 < p < ∞ ,
k[b, T ]fkLp(w) ≤ CkbkBM O[w]2
A∞kM 2fkLp(w) ,
where M 2 = M ◦ M denotes the Hardy-Littlewood maximal function iterated twice. With this
result and Buckley's sharp estimate for the maximal function [3] one can immediately conclude that
2
k[b, T ]kLp(w)→Lp(w) ≤ C[w]2
A∞ [w]
p−1
Ap kbkBM O .
In this paper we show that for T the Hilbert, Riesz, Beurling transform, for 1 < p ≤ 2 one can drop
[w]A∞ term, in the above estimate, and this is sharp. However for p > 2 , the Lp(w)-norm of [b, T ]
is bounded above by kbkBM Od[w]2
. For T = H the Hilbert transform we prove, using Bellman
function techniques similar to those used in [2], [26], the following Theorem.
A2
p
2
Theorem 1.1. There exists a constant C > 0, such that
k[b, H]fkLp(w)→Lp(w) ≤ CkbkBM O[w]
2 max{1, 1
Ap
p−1 }
kfkLp(w) ,
and this is sharp for 1 < p < ∞.
Most of the work goes into showing the quadratic estimate for p = 2, sharp extrapolation [8]
then provides the right rate of growth for p 6= 2. An example of C. Pérez shows the rates are sharp.
Our method involves the use of the dyadic paraproduct πb and its adjoint π∗
b , both of which obey
linear estimates in L2(w), see [2], like the Hilbert transform. It also uses Petermichl description of
the Hilbert transform as an average of dyadic shift operators S, [25] and reduces the estimate to
obtaining corresponding estimates for the commutator [b, S]. After we decompose this commutator
in three parts:
[b, S]f = [πb, S]f + [π∗
b , S]f + [λb, S]f ,
we estimate each commutator separately. This decomposition has been used before to analyze the
commutator, [25], [15], [16]. For precise definitions and detail derivations, see Section 1.2. The first
two commutators immediately give the desired quadratic estimates in L2(w) from the known linear
bounds of the operators commuted. For the third commutator we can prove a better than quadratic
bound, in fact a linear bound. The following Theorem will be the crucial part of the proof.
Theorem 1.2. There exists a constant C > 0, such that
k[λb, S]kL2(w)→L2(w) ≤ C[w]Ad
for all b ∈ BM O d and w ∈ Ad
2.
2kbkBM O d ,
(1.2)
This theorem is an immediate consequence of results in [14], [17] and [7], since the operator
[λb, S] belongs to the class of Haar shift operators for which they can prove linear bounds. We
present a different proof of this result and others, using Bellman function techniques and bilinear
Carleson embedding theorems, very much in the spirit of [25] and [2]. These arguments were found
independently by the author, and we think they can be of interest.
We then observe that for any Haar shift operator T as defined in [17] the commutator [λb, T ]
is again a Haar shift operator, and therefore it obeys linear bounds in A2-characteristic of the
weight as in Theorem 1.2. As a consequence, we obtain quadratic bounds for all commutators of
Haar shift operators and BMO function b . In particular, this holds true for Haar shift operators
in Rn whose averages recover the Riesz transforms [27] and for martingale transforms in R2 whose
averages recover the Beurling-Ahlfors operator [30], [9]. Extrapolation will provide Lp(w) bounds
which turn out to be sharp for the Riesz transforms and Beurling-Ahlfors operators as well. The
following Theorem holds
Theorem 1.3. Let Tτ be the first class of Haar shift operators of index τ . Its convex hull include
the Hilbert transform, Riesz transforms, the Beurling-Ahlfors operator and so on. Then, there exists
a constant C(τ, n, p) which only depend on τ , n and p such that
k[b, Tτ ]kLp(w)→Lp(w) ≤ C(τ, n, p)[w]
2 max{1, 1
Ap
p−1 }
kbkBM O
See Section 10 for definitions and precise statements of these results.
3
1.2 The Hilbert transform case
The bilinear operator
f H(g) + H(f )g
maps L2 × L2 into H 1, here H is the Hilbert transform and H 1 is the real Hardy space defined by
H 1(R) := {f ∈ L1(R) : Hf ∈ L1(R)}
with norm
Because the dual of H 1 is BM O, we will pair with a BM O function b . Using that H ∗ = −H, we
obtain that
kfkH 1 = kfkL1 + kHfkL1 .
h f H(g) + H(f )g, bi = h f, H(g)b − H(gb)i .
Hence the operator g 7→ H(g)b − H(gb) should be L2 bounded. This expression H(g)b − H(gb) is
called the commutator of H with the BM O function b . More generally, we define as follows.
Definition 1.4. The commutator of the Hilbert transform H with a function b is defined as
[b, H](f ) = bH(f ) − H(bf ) .
Our main concern in this paper is to prove that the commutator [b, H], as an operator from
R(w) is bounded by the square of the A2-characteristic, [w]A2 , of the weight times the
R(w) into L2
L2
BM O norm, of b , kbkBM O , where
kbkBM O := sup
I
1
IZI b(x) − hbiIdx .
The supremum is taken over all intervals in R . Note that when we restrict the supremum to dyadic
intervals this will define BM O d and we denote this dyadic BM O norm by k·kBM O d . We now state
our main results.
Theorem 1.5. There exists C such that for all w ∈ A2 ,
k[b, H]kL2(w)→L2(w) ≤ C[w]2
A2kbkBM O ,
for all b ∈ BM O .
Once we have boundedness and sharpness for the crucial case p = 2, we can carry out the power
of the Ap-characteristic, for any 1 < p < ∞ , using the sharp extrapolation theorem [8] to obtain
Theorem 1.1. Furthermore, an example of C. Pérez [22] shows this quadratic power is sharp. In
[25], S. Petermichl showed that the norm of the commutator of the Hilbert transform is bounded by
the supremum of the norms of the commutator of certain shift operators. This result follows after
writing the kernel of the Hilbert transform as a well chosen averages of certain dyadic shift operators
discovered by Petermichl. More precisely, S. Petermichl showed there is a non zero constant C such
that
k[b, H]k ≤ C sup
α,r k[b, Sα,r]k ,
(1.3)
where the dyadic shift operator Sα,r is defined by
Sα,rf = XI∈Dα,rhf, hIi(hI− − hI+) .
4
π∗
b (f )(x) :=XI∈D
πb(f )(x) :=XI∈D
λb(f )(x) :=XI∈D
hb, hIihf, hIih2
I (x) ,
hb, hIihfiI hI (x) ,
hbiIhf, hIihI (x) .
S(f ) =XI∈D
hf, hIi(hI− − hI+)
Thus, we have
where
[b, S] = [π∗
b , S] + [πb, S] + [λb, S] ,
(1.5)
Where hI denotes the Haar function associated to the interval I, I± denote the left and right
halves of I, and Dα,r is a shifted and dilated system of dyadic intervals. Denote by S the shift op-
erator corresponding to the standard dyadic intervals D . See the next section for a precise definition.
Let us consider a compactly supported b ∈ BM O d and f ∈ L2 . Expanding b and f in the Haar
system associated to the dyadic intervals D ,
b(x) =XI∈D
hb, hIihI (x), f (x) = XJ∈D
hf, hJihJ (x) ;
formally, we get the multiplication of b and f to be broken into three terms,
bf = π∗
b (f ) + πb(f ) + λb(f ) ,
(1.4)
where πb is the dyadic paraproduct, π∗
b is its adjoint and λb(·) = π(·)b , defined as follows
and we can estimate each term separately. Notice that both πb and π∗
b are bounded operators in
Lp for b ∈ BM O , despite the fact that multipication by b is not a bounded operator in L2 unless
b is bounded (L∞) . Therefore, λb can not be a bounded operator in Lp . However [λb, S] will be
bounded on Lp(w) and will be better behaved than [b, S] . Decomposition (1.5) was used to analyze
the commutator with the shift operator first by Petermichl in [25], but also Lacey in [15] and authors
in [16] to analyze the iterated commutators. Since all estimates are independent on the dyadic grid,
through out this paper we only deal with the dyadic shift operator S associated to the standard
dyadic grid D . For a single shift operator the hypothesis required on b and w are that they belong
to dyadic BM O d and Ad
2 with respect to the underlying dyadic grid defining the operator. However
since ultimately we want to average over all grids, we will need b and w belonging to BM O d and
Ad
2 for all shifted and scaled dyadic grids, that we will have if b ∈ BM O and w ∈ A2 , non-dyadic
BM O and A2 . Beznosova has proved linear bounds for πb and π∗
b , [2], together with Petermichl's
[26] linear bounds for S , this immediately provides the quadratic bounds for [πb, S] and [π∗
b , S].
Theorem 1.5 will be proved once we show the quadratic estimate holds for [λb, S]. We can actually
obtain a better linear estimate as in Theorem 1.2. Some terms in (1.5) do also obey linear bounds.
Theorem 1.6. There exists C such that
kπ∗
b SkL2(w) + kSπbkL2(w) ≤ C[w]Ad
2kbkBM Od .
for all b ∈ BM Od .
5
Note the three operators [λb, S] , π∗
b S and Sπb are generalized Haar shift operators for which
there are now two different proofs of linear bounds on L2(w) with respect to [w]Ad
, [17] and [7],
and in this paper we present a third proof. We are now ready to explain the organization of this
paper. In Section 2 we will introduce notation and discuss some useful results about weighted Haar
systems. In Section 3 we will start our discussion on how to find the linear bound for the term
[λb, S] , most of which will be very similar to calculations performed in [26]. In Section 4 we will
introduce a number of Lemmas and Theorems that will be used. In Section 5 we will finish the
linear estimate for the term [λb, S] , and prove Theorem 1.5. In Section 6 we prove the linear bound
for π∗
b S . In Section 7 we reduce the proof of the linear bound for Sπb to verifying three embedding
conditions, two are proved in this section, the third is proved in Section 8 using a Bellman function
argument. In Section 9 we present the Lp(w) estimate of the commutator with a Haar shift operator,
Theorem 1.3. Finally, in Section 10 we provide the sharpness for the commutators of Hilbert, Riesz
transforms and Beurling-Ahlfors operators.
2
2 Preliminaries and Notation
Let us now introduce the notation which will be used frequently through this paper. Even though the
Ap conditions have already been introduced in (1.1), we will state the special case of this condition
when p = 2 , namely Ad
2 since we will refer repeatedly to this. We say w belongs to Ad
2 class, if
[w]Ad
2
:= sup
I∈DhwiIhw−1iI < ∞ .
(2.1)
Here we take the supremum over all dyadic interval in R . Note that if w ∈ A2 then w ∈ Ad
2 and
2 ≤ [w]A2 . Intervals of the form [k2−j , (k + 1)2−j ) for integers j, k are called dyadic intervals.
[w]Ad
Again, let us denote D the collection of all dyadic intervals, and let us denote D(J) the collection
of all dyadic subintervals of J . For any interval I ∈ D, there is a Haar function defined by
hI (x) =
1
I1/2 (χI+(x) − χI−(x)) ,
where χI denotes the characteristic function of the interval I , χI (x) = 1 if x ∈ I , χI (x) = 0
otherwise. It is a well known fact that the Haar system {hI}I∈D is an orthonormal system in L2
R .
We also consider the different grids of dyadic intervals parametrized by α, r, defined by
D α,r = {α + rI : I ∈ D} ,
for α ∈ R and positive r . For each grid D α,r of dyadic intervals, there are corresponding Haar func-
tions hα,r
R . Let us introduce a proper orthonormal
system for L2
, I ∈ D α,r that are an orthonormal system in L2
I
R(w) defined by
hw
I :=
1
w(I)1/2(cid:20) w(I−)1/2
w(I+)1/2
χI+ −
w(I+)1/2
w(I−)1/2
χI−(cid:21) ,
where w(I) =RI w . We define the weighted inner product by hf, giw =RR f gw . Then, every function
f ∈ L2(w) can be written as
hf, hw
I iwhw
I ,
where the sum converges a.e. in L2(w) . Moreover,
f =XI∈D
L2(w) =XI∈D
kfk2
6
hf, hw
I iw2 .
(2.2)
Again D can be replaced by D α,r and the corresponding weighted Haar functions are an orthonormal
system in L2(w) . For convenience we will observe basic properties of the disbalanced Haar system.
First observe that hhK , hw
I iw could be non-zero only if I ⊇ K, moreover, for any I ⊇ K,
hhK , hw
Here is the the calculation that provides (2.3),
I iw ≤ hwi1/2
K .
(2.3)
1
hhK , hw
I iw =(cid:12)(cid:12)(cid:12)(cid:12)Z
K1/2w(I)1/2 (χK+(x) − χK−(x))(cid:20) w(I−)1/2
K1/2w(I)1/2ZK(cid:12)(cid:12)(cid:12)(cid:12)
{z
If K ⊂ I+, then A ≤ w(I−)1/2w(I+)−1/2w(K). Thus
w(I−)1/2
w(I+)1/2
χI+(x) +
w(I+)1/2 χI+(x) −
w(I+)1/2
w(I−)1/2
w(I+)1/2
w(I−)1/2 χI−(x)(cid:21)w(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
χI−(x)(cid:12)(cid:12)(cid:12)(cid:12)w(x)dx
}
≤
1
A
.
hhK , hw
I iw ≤ hwi1/2
K s w(K)w(I−)
w(I+)w(I) ≤ hwi1/2
K .
Similarly, if K ⊂ I− . If K = I, then A ≤ 2 w(K−)1/2w(K+)1/2 . Thus
2pw(K−)w(K+)
I iw = hhK , hw
Kiw ≤ hwi1/2
hhK , hw
w(K)
K
≤ hwi1/2
K .
Estimate (2.3) implies that hh J , hw−1
J
iw−1hhJ , hw
J iw ≤ √2[w]1/2
Ad
2
, where J is the parent of J ,
hh J , hw−1
J
iw−1hhJ , hw
w(cid:19)1/2
J (cid:18) 1
JZJ
J = hw−1i1/2
J iw ≤ hw−1i1/2
J hwi1/2
J (cid:18) 2
w(cid:19)1/2
JZJ
√2hw−1i1/2
= hw−1i1/2
J hwi1/2
√2 [w]1/2
≤
≤
Ad
2
J
.
Also, one can deduce similarly the following estimate
hhJ , hw−1
J
iw−1hhJ−, hw
J iw ≤
√2[w]1/2
Ad
2
.
(2.4)
(2.5)
For I ) J, hw
value of hw
J
J
I is constant on J. We will denote this constant by hw
on J and hw
(J) ≤ w(J)−1/2 , as can be seen by the following estimate,
(J) =(cid:26) w( J−)1/2/w( J )1/2w( J+)1/2 ≤ w(J)−1/2 if J = J+
w( J+)1/2/w( J )1/2w( J−)1/2 ≤ w(J)−1/2 if J = J− .
I (J) . Then hw
J
hw
J
Let us define the weighted averages, hgiJ,w := w(J)−1RJ g(x)w(x)dx . As with the standard
Haar system, we can write the weighted averages
(J) is the constant
(2.6)
hgiJ,w = XI∈D:I)J
7
hg, hw
I iwhw
I (J) .
(2.7)
In fact, here is the derivation of (2.7).
hgiJ,w =
=
1
w(J)ZJXI∈D
w(J)ZJ
hg, hw
I iwhw
w(x)dx XI∈D:I)J
1
1
w(J)ZJXI∈D
I (J) = XI∈D:I)J
hg, hw
I iwhw
hg, hw
I iwhw
I (J) .
I (x)w(x)dx =
hg, hw
I iwhw
I (J)w(x)dx
Also, we will be using system of functions {H w
I }I∈D defined by
H w
I = hIpI − Aw
I χI where Aw
I = hwiI+ − hwiI−
.
(2.8)
2hwiI
I } is orthogonal in L2 with norms satisfying the inequality kw1/2H w
Then, {w1/2H w
refer to [2]. Moreover, by Bessel's inequality we have, for all g ∈ L2 ,
L2 .
1
IhwiI hg, w1/2H w
I i2 ≤ kgk2
XI∈D
I kL2 ≤pIhwiI ,
(2.9)
3 Linear bound for [λb, S] part I
In general, when we analyse commutator operators, a subtle cancelation delivers the result one wants
to find. In the analysis of the commutator [b, S], the part [λb, S] will allow for certain cancelation.
First, let us rewrite [λb, S] .
[λb, S](f ) = λb(Sf ) − S(λbf )
hλbf, hJi(hJ− − hJ+)
hbiIhSf, hIihI −XJ∈D
hbiIhf, hJihhJ− − hJ+, hIihI −XJ∈DXI∈D
hbiJ+hf, hJihJ+ −XJ∈D
hbiJ−hf, hJihJ− −XJ∈D
hbiJ− − hbiJ+
2
hf, hJihJ− −XJ∈D
hbiJ+ − hbiJ−
2
∆J bhf, hJi(hJ+ + hJ−) ,
=XI∈D
=XI∈DXJ∈D
=XJ∈D
=XJ∈D
= −XJ∈D
hbiIhf, hIihhI , hJi(hJ− − hJ+)
hbiJ+ + hbiJ−
2
hf, hJi(hJ− − hJ+)
hf, hJihJ+
recall the notation ∆J b = (hbiJ+ −hbiJ−)/2 . To find the L2(w) operator norm of [λb, S], it is enough
to deal with the operator
Sb(f ) =XI∈D
∆I bhf, hIihI− .
We shall state the weighted operator norm of Sb as a Theorem and give a detailed proof. Theorem
1.2 is a direct consequence of the following Theorem. We will prove the following Theorem by the
technique used in [26].
Theorem 3.1. There exists a constant C > 0, such that
for all b ∈ BM O d and w ∈ Ad
kSbkL2(w)→L2(w) ≤ C[w]Ad
2 for all f ∈ L2(w).
2kbkBM O d
8
(3.1)
One can easily check that choosing b = I1/2hI , yields kSkL2(w)→L2(w) ≤ C[w]Ad
(3.1) is equivalent to the following inequality for any positive function f, g
2
. Inequality
hSb,w−1f, giw ≤ C[w]Ad
2kbkBM O dkfkL2(w−1)kgkL2(w) ,
(3.2)
where Sb,w−1(f ) = Sb(w−1f ) . Expanding f and g in the disbalanced Haar systems respectively for
L2(w−1) and L2(w) yields for (3.2),
hSb,w−1f, giw =(cid:12)(cid:12)(cid:12)(cid:12)Z Sb,w−1(cid:0)XI∈D
=(cid:12)(cid:12)(cid:12)(cid:12)XI∈DXJ∈D
=(cid:12)(cid:12)(cid:12)(cid:12)XI∈DXJ∈D
I
hf, hw−1
I
hf, hw−1
I
iw−1hg, hw
hf, hw−1
iw−1hg, hw
hg, hw
I
(cid:1)(cid:0)XJ∈D
iw−1hw−1
J iwZ hw
J iwhSb,w−1hw−1
I
J Sb,w−1(hw−1
I
J iwhw
J(cid:1)wdx(cid:12)(cid:12)(cid:12)(cid:12)
)wdx(cid:12)(cid:12)(cid:12)(cid:12)
J iw(cid:12)(cid:12)(cid:12)(cid:12) .
, hw
(3.3)
In (3.3),
hSb,w−1hw−1
I
, hw
∆LbhhL, w−1hw−1
J iw =DXL∈D
JEw
i hL− , hw
iw−1 6= 0, only when L ⊆ I and hhL−, hw
I
= XL∈D
∆LbhhL, hw−1
I
iw−1hhL−, hw
J iw
I
Since hhL, hw−1
J iw 6= 0 only when L− ⊆ J , then we have
non-zero terms if I ⊆ J or J ⊆ I in the sum of (3.3). Thus we can split the sum into four parts,
PI=J , PI= J , P J (I , andPI(J . Let us now introduce the truncated shift operator
SI
b(f ) := XL∈D(I)
∆Lbhf, hLihL− ,
and its composition with multiplication by w−1,
SI
b,w−1(f ) := XL∈D(I)
∆Lbhw−1f, hLihL− .
We will see that the weighted norm kSI
b,w−1χIkL2(w), plays a main role in our estimate for hS b,w−1hw−1
I
, hw
J iw .
4 Theorems and Lemmas
To prove inequality (3.2) we need several theorems and lemmas. One can find the proof in the
indicated references. First we recall some embedding theorems. Another main result in [26] is a
two-weighted bilinear embedding theorem, which was proved by a Bellman function argument.
Theorem 4.1 (Petermichl's Bilinear Embedding Theorem). Let w and v be weights so that hwiIhviI ≤
Q for all intervals I and let {αI} be a non-negative sequence so that the three estimates below hold
for all J
XI∈D(J)
XI∈D(J)
αI
hwiI ≤ Q v(J)
hviI ≤ Q w(J)
αI
9
(4.1)
(4.2)
XI∈D(J)
αI ≤ QJ .
(4.3)
Then there is c such that for all f ∈ L2(w) and g ∈ L2(v)
XI∈D
αIhfiI,whgiI,v ≤ cQkfkL2(w)kgkL2(v) .
Corollary 4.2 (Bilinear Embedding Theorem). Let w and v be weights. Let {αI} be a sequence of
nonnegative numbers such that for all dyadic intervals J ∈ D the following three inequalities hold
with some constant Q > 0 ,
XI∈D(J)
XI∈D(J)
XI∈D(J)
αIhviI I ≤ Q v(J)
αIhwiI I ≤ Q w(J)
αIhwiIhviI ≤ QJ .
(4.4)
(4.5)
(4.6)
Then for any two nonnegative function f, g ∈ L2
XI∈D
αIhf v1/2iIhgw1/2iI I ≤ CQkfkL2kgkL2
holds with some constant C > 0 .
Both Bilinear Embedding Theorems are key tools in our estimate. One version of such a theorem
appeared in [19]. The original version of the next lemma also appeared in [26]. Using the fact that,
for all I ∈ D , ∆I b ≤ 2kbkBM Od , one can prove the following lemma similarly to its original proof.
w−1(I)1/2 for all
Lemma 4.3. There is a constant c such that kSI
intervals I and weights w ∈ Ad
2 .
b,w−1χIkL2(w) ≤ ckbkBM O d[w]Ad
2
We will also need the Weighted Carleson Embedding Theorem from [19], and some other in-
equalities for weights.
Theorem 4.4 (Weighted Carleson Embedding Theorem). Let {αJ} be a non-negative sequence
such that for all dyadic intervals I
XJ∈D(I)
αJ ≤ Qw−1(I) .
Then for all f ∈ L2(w−1)
XJ∈D
αJhfi2
J,w−1 ≤ 4Qkfk2
L2(w−1) .
Theorem 4.5 (Wittwer's sharp version of Buckley's inequality). There exist a positive constant C
such that for any weight w ∈ Ad
2 and dyadic interval I ∈ D ,
1
J XI∈D(J)(cid:0)hwiI+ − hwiI−(cid:1)2
hwiI
I ≤ C[w]Ad
2hwiJ .
10
We refer to [32] for the proof. You can find extended versions of Theorem 4.4 and 4.5 to the
doubling positive measure σ in [24]. One can find the Bellman function proof of the following three
Lemmas in [2].
Lemma 4.6. For all dyadic interval J and all weights w .
1
J XI∈D(J)
I∆I w2 1
hwi3
I ≤ hw−1iJ .
Lemma 4.7. Let w be a weight and {αI} be a Carleson sequence of nonnegative numbers. If there
exist a constant Q > 0 such that
then
∀ J ∈ D ,
1
J XI∈D(J)
αI ≤ Q ,
∀ J ∈ D ,
1
J XI∈D(J)
αI
hw−1iI ≤ 4QhwiJ
and therefore if w ∈ Ad
2 the for any J ∈ D we have
Lemma 4.8. If w ∈ Ad
∀J ∈ D ,
1
J XI∈D(J)
hwiI αI ≤ 4Q[w]Ad
2hwiJ .
2 then there exists a constant C > 0 such that
1
J XI∈D(J)(cid:18)hwiI+ − hwiI−
hwiI
(cid:19)2
IhwiIhw−1iI ≤ C[w]Ad
2
.
5 Linear bound for [λb, S] part II
We will continue to estimate the sum (3.3) in four parts.
5.1 PI= J
For this case, it is sufficient to show that
hSb,w−1hw−1
J
, hw
J iw ≤ ckbkBM O d[w]Ad
2
Since hhk, hw
hSb,w−1hw−1
J
I iw could be non-zero only if K ⊆ I,
, hw
∆Lbhw−1hw−1
L , hLihL−, hw
has non-zero term only when L ⊆ J. Thus
∆Lbhhw−1
hSb,w−1hw−1
, hw
J
L
, hLiw−1hhL−, hw
J iw =(cid:12)(cid:12)(cid:12)DXL∈D
J iw =(cid:12)(cid:12)(cid:12) XL∈D( J )
=(cid:12)(cid:12)(cid:12) XL∈D(J)
∆Lbhhw−1
+ ∆ J bhhw−1
J
b,w−1hw−1
≤ hSJ
J
J
, hLiw−1hhL−, hw
, h Jiw−1hh J−, hw
J iw
J iw + ∆ J bhhw−1
, hw
J
11
.
∆Lbhhw−1
J
, hLiw−1hhL−, hw
J iw(cid:12)(cid:12)(cid:12)
∆Lbhhw−1
J
, hLiw−1hhL−, hw
J iw(cid:12)(cid:12)(cid:12)
JEw(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)XL∈D
J iw(cid:12)(cid:12)(cid:12)
J iw(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) XL∈D(J s)
, h Jiw−1hh J−, hw
J iw
in the second equality, J s denotes the sibling of J, so for all L ⊆ J s, hhL−, hw
(2.4),
J iw = 0 . Then, by
∆ J bhhw−1
J
, h Jiw−1hh J−
, hw
J iw ≤
So for the remaining part:
√2kbkBM O d[w]1/2
Ad
2
.
(5.1)
(J)hSJ
here the last inequality uses (2.6) and Lemma 4.3.
J iw = hw−1
, hw
b,w−1hw−1
hSJ
J
J
b,w−1χJ , hw
J iw ≤ ckbkBM O d[w]Ad
2
,
(5.2)
5.2 PI=J
In this case, the argument is similar to the argument in Section 5.1. We have
hSb,w−1hw−1
J
, hw
J iw = XL∈D
∆Lbhhw−1
J
, hLiw−1hhL−, hw
J iw
here we have zero summands, unless L ⊆ J. Thus,
hSb,w−1hw−1
J
J
J iw = hSJ
b,w−1hw−1
b,w−1hw−1
, hw
, hw
J iw
b,w−1hw−1
≤ hSJ+
J iw + hSJ−
, hw
≤ ckbkBM O d[w]Ad
J
J
.
2
J iw + ∆J bhhw−1
, hw
J
, hJiw−1hhJ−, hw
J iw
In the last inequality, we use same arguments as in (5.2) for the first two term, and (2.5) for the
last term.
To obtain our desired results, we need to understand the supports of Sb(w−1hw−1
Since
I
) and S∗
b (whw
J ) .
5.3 P J (I and PI(J
and hhw−1
I
I
Sb(w−1hw−1
) = XL∈D
∆Lbhw−1hw−1
, hLihL− = XL∈D
, hLi can be non-zero only when L ⊆ I , Sb(w−1hw−1
b (whw
), hw
, S∗
I
I
hSb(w−1hw−1
I
J iw = hhw−1
I
J )iw−1 ,
∆Lbhhw−1
I
, hLiw−1hL− ,
) is supported by I . Also,
(5.3)
12
J , hL−ihL is supported by J . Let us now consider the sum
yield that S∗
J ( I . Then
b (whw
J ) = PL∈D ∆Lbhwhw
(cid:12)(cid:12)(cid:12)(cid:12) XI,J: J(I
iw−1hg, hw
hf, hw−1
I
J iwhSb,w−1hw−1
I
, hw
J iw(cid:12)(cid:12)(cid:12)(cid:12)
hf, hw−1
I
iw−1hg, hw
J iwhhw−1
I
, S∗
b (whw
= (cid:12)(cid:12)(cid:12)(cid:12) XI,J: J(I
= (cid:12)(cid:12)(cid:12)(cid:12)XJ∈D XI:I) J
= (cid:12)(cid:12)(cid:12)(cid:12)XJ∈D
≤ kgkL2(w)(cid:18)XJ∈D
hf, hw−1
I
iw−1hg, hw
J iwhw−1
I
( J)hSb,w−1χ J , hw
hfi J,w−1hg, hw
J iwhSb,w−1χ J , hw
J iw(cid:12)(cid:12)(cid:12)(cid:12)
w(cid:19)1/2
J i2
J,w−1hSb,w−1χ J , hw
,
hfi2
J )iw−1(cid:12)(cid:12)(cid:12)(cid:12)
J iw(cid:12)(cid:12)(cid:12)(cid:12)
(5.4)
(5.5)
(5.6)
here (5.3) and the fact that S∗
(2.7) and (5.3). If we show that
b (whw
J ) is supported by J are used for equality (5.4), and (5.5) uses
XJ∈D
then we have
hfi2
J,w−1hSb,w−1χ J , hw
J i2
w ≤ ckbk2
BM O d[w]2
Ad
2kfk2
L2(w−1) ,
(5.7)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) XI,J: J(I
hf, hw−1
I
iw−1hg, hw
J iwhSb,w−1hw−1
I
, hw
2kfk2
L(w−1)kgkL2(w) .
J iw(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ CkbkBM O d[w]Ad
To prove the inequality (5.7), we apply the Theorem 4.4. The embedding condition becomes
XJ∈D:J (I
hSb,w−1χ J , hw
J i2
w ≤ ckbk2
BM O d[w]2
Ad
2
w−1(I)
after shifting the indices. Since hhL−, hw
J ( I, we can write
J iw = 0 unless L ⊆ J , and we will sum over J such that
∆Lbhw−1χ J , hLihhL−, hw
J iw
b,w−1χI, hw
J iw .
hSb,w−1χ J , hw
J iw = XL∈D( J )
J iw = hSI
∆Lbhw−1χ J , hLihhL− , hw
J iw = XL∈D
= XL∈D(I )
w = XJ∈D:J (I
J i2
hSb,w−1χ J , hw
XJ∈D:J (I
∆Lbhw−1χI , hLihhL−, hw
Thus,
b,w−1χIk2
the sum J ( I. The partPI(J is similar toP J (I . One uses that Sb(w−1hw−1
last inequality due to (2.2). By Lemma 4.3, the embedding condition holds. Hence we are done for
) is supported by I
b,w−1χI , hw
J i2
and Theorem 4.4.
w ≤ kSI
L2(w) ,
hSI
I
13
5.4 Proof of Theorem 1.5
We refer to [2] for following theorem.
Theorem 5.1. The norm of dyadic paraproduct on the weighted Lebesgue space L2(w) is bounded
from above by a constant multiple of the product of the Ad
2 characteristic of the weight w and the
BM O d norm of b.
To break [b, S] into three parts, as in (1.5), we assumed that b ∈ BM O d is compactly supported.
However, we need to replace such a b with a general BM O d function.
In order to pass from a
compactly supported b to general b ∈ BM O d , we need the following lemma which is suggested in
[10] .
Lemma 5.2. Suppose φ ∈ BM O . Let eI be the interval concentric with I having length eI = 3I .
Then there is ψ ∈ BM O such that ψ = φ on I, ψ = 0 on R \eI and kψkBM O ≤ ckφkBM O .
Proof. Without loss of generality, we assume hφiI = 0 . Write I =S ∞
Jn, as in following figure.
n=0 Jn where dist(Jn, ∂I) =
J3
J1
J0
J2
J4
K1
{z
I
K2
}
Then J0 is the middle third of I . For n > 0, let Kn be the reflection of Jn across the nearest
endpoint of I and set
This construction of ψ satisfies Lemma 5.2.
ψ(x) =
φ(x) if x ∈ I
hφiJn if x ∈ Kn
0 otherwise .
By Theorem 3.1, Corollary 1.2, Theorem 5.1, Lemma 5.2 and using the fact kπbk = kπ∗
bk, we
can prove Theorem 1.5.
Proof of Theorem 1.5. For any compactly supported b ∈ BM O ,
k[b, H]kL2(w)→L2(w) ≤ C sup
≤ C sup
α,r k[b, S α,r]kL2(w)→L2(w)
α,r (cid:0)k[πb, S α,r]kL2(w)→L2(w) + k[π∗
b , S α,r]kL2(w)→L2(w) + k[λb, S α,r]kL2(w)→L2(w)(cid:1)
≤ C(cid:0)4kπbkL2(w)→L2(w) sup
≤C[w]2
A2kbkBM O .
α,r kS α,rkL2(w)→L2(w) + C[w]A2kbkBM O(cid:1)
For fixed b, we consider the sequence of intervals Ik = [−k, k] and the sequence of BM O functions
bk which are constructed as in Lemma 5.2. Then, there is a constant c, which does not depend on
k , such that kbkkBM O ≤ ckbkBM O . Furthermore, there is a uniform constant C such that
k[bk, H]kL2(w)→L2(w) ≤ C[w]2
(5.8)
Therefore, for some subsequence of integers kj and f ∈ L2(w) , [bkj , H](f ) converges to [b, H](f )
almost everywhere. Letting j → ∞ and using Fatou's lemma, we deduce that (5.8) holds for all
b ∈ BM O .
A2kbkBM O .
14
6 Linear bound for π∗b S
It might be useful to know what is the adjoint operator of S. Let us define sgn(I) = ±1 if I = I∓ .
Then, for any function f, g ∈ L2 ,
hf, hIihg, hJihhI− − hI+, hJi
hf, hIihg, hI+i
hf, hIihg, hI−i −XI∈D
hf, hIi(cid:0)sgn(I−)hg, hI−i + sgn(I+)hg, hI+i(cid:1)
sgn(I)hg, hIih IE = hf, S∗gi .
hf, h Iisgn(I)hg, hIi
hSf, gi =XI∈DXJ∈D
=XI∈D
=XI∈D
=XI∈D
=Df,XI∈D
S∗f (x) =XI∈D
sgn(I)hf, hIih I (x) .
Now, we see the adjoint operator of dyadic shift operator S is
The following lemma provides the bound we are looking for the term π∗
Lemma 6.1. Let w ∈ Ad
2 and b ∈ BM O d. Then, there exists C so that
b S .
kπ∗
b SkL2(w)→L2(w) ≤ C[w]Ad
2kbkBM O d .
Proof. In order to prove Lemma 6.1 it is enough to show that for any positive square integrable
function f, g
Using the system of functions {H w
b S(f w−1/2), gw1/2i ≤ C[w]Ad
hπ∗
(6.1)
I }I∈D defined in (2.8), we can rewrite the left hand side of (6.1)
2kbkBM O dkfkL2kgkL2 .
b S(f w−1/2), gw1/2i = hS(f w−1/2), πb(gw1/2)i =XI∈D
hπ∗
hgw1/2iIhb, hIihS(f w−1/2), hIi
=XI∈D
=XI∈D
hgw1/2iIhb, hIi sgn(I)hf w−1/2, h Ii
sgn(I)hgw1/2iIhb, hIihf w−1/2, H w−1
+XI∈D
q I
sgn(I)hgw1/2iIhb, hIihf w−1/2, Aw−1
i
1
I
I
.
(6.2)
χ Ii
1
q I
Our claim is that both sums in (6.2) are bounded by [w]Ad
2kbkBM O dkfkL2kgkL2 , i.e.
sgn(I)hgw1/2iIhb, hIihf w−1/2, H w−1
I
≤ C[w]Ad
2kbkBM O dkfkL2kgkL2
(6.3)
XI∈D
and
XI∈D
sgn(I)hgw1/2iIhb, hIihf w−1/2, Aw−1
I
χ Ii
≤ C[w]Ad
2kbkBM O dkfkL2kgkL2 .
(6.4)
1
i
q I
q I
1
15
First let us verify the bound for (6.3). Using Cauchy-Schwarz inequality,
XI∈D
sgn(I)hgw1/2iIhb, hIihf w−1/2, H w−1
I
1
i
q I
Ihb, hIi2hw−1i I(cid:19)1/2(cid:18)XI∈D
Ihb, hIi2hw−1i I(cid:19)1/2
hgw1/2i2
hgw1/2i2
≤(cid:18)XI∈D
≤ kfkL2(cid:18)XI∈D
hf, w−1/2H w−1
I
i2(cid:19)1/2
1
Ihw−1i I
.
Thus, for (6.3), it is enough to show that
XI∈D
hgw1/2i2
Ihb, hIi2hw−1i I ≤ C[w]2
Ad
2kbk2
BM O dkgk2
L2 .
It is clear that 2hwi I ≥ hwiI , thus
XI∈D
hgw1/2i2
Ihb, hIi2hw−1i I =XI∈D
Ihb, hIi2hw−1i IhwiIhwi−1
hgw1/2i2
Ihb, hIi2hwi−1
.
I
I
hgw1/2i2
2XI∈D
≤ 2[w]Ad
If we show for all J ∈ D ,
1
J XI∈D(J)
hb, hIi2hwi−1
I hwi2
I =
1
J XI∈D(J)
hb, hIi2hwiI ≤ [w]Ad
2kbk2
BM O dhwiJ ,
(6.5)
(6.6)
(6.7)
then by Weighted Carleson Embedding Theorem 4.4 with w instead of w−1 , we will have (6.6) .
Since b ∈ BM O d , {hb, hIi2}I∈D is a Carleson sequence with constant kbk2
BM O d that is
1
J XI∈D(J)
hb, hIi2 ≤ kbk2
BM O d .
Applying Lemma 4.7 with αI = hb, hIi we have inequality (6.7). We now concentrate on the estimate
(6.4), we can estimate the left hand side of (6.4) as follows.
XI∈D
1
sgn(I)hgw1/2iIhb, hIihf w−1/2, Aw−1
I
χ Ii
hgw1/2iI hb, hIihf w−1/2i I Aw−1
q I
I q I
sgn(I)hgw1/2iIhb, hIihf w−1/2i I Aw−1
q I
q I
hgw1/2i I hb, hIihf w−1/2i I Aw−1
hgw1/2iI(cid:0)hb, hI−i + hb, hI+i(cid:1) hf w−1/2iI Aw−1
=XI∈D
≤XI∈D
≤ 2XI∈D
= 2XI∈D
I
I
I
pI .
16
By Bilinear Embedding Theorem, inequality (6.4) holds provided the following three inequalities
hold,
∀ J ∈ D,
∀ J ∈ D,
∀ J ∈ D,
I
1
1
J XI∈D(J)(cid:0)hb, hI−i + hb, hI+i(cid:1)Aw−1
J XI∈D(J)(cid:0)hb, hI−i + hb, hI+i(cid:1)Aw−1
J XI∈D(J)(cid:0)hb, hI−i + hb, hI+i(cid:1)Aw−1
1
I
I
2
,
pIhw−1iIhwiI ≤ CkbkBM O d [w−1]Ad
pIhw−1iI ≤ CkbkBM O d [w−1]Ad
2hw−1iJ ,
pIhwiI ≤ CkbkBM O d [w−1]Ad
2hwiJ .
(6.8)
(6.9)
(6.10)
For (6.8), by Cauchy-Schwarz inequality
I
1
J XI∈D(J)(cid:0)hb, hI−i + hb, hI+i(cid:1)Aw−1
pIhw−1iIhwiI
≤(cid:18) 1
J XI∈D(J)(cid:0)hb, hI−i + hb, hI+i(cid:1)2hw−1iIhwiI(cid:19)1/2(cid:18) 1
(hb, hI−i + hb, hI+i)2 ≤ 3XI∈D
XI∈D
Since
1
J XI∈D(J)
and by Lemma 4.8,
1
J XI∈D(J)
(Aw−1
I
(hb, hI−i + hb, hI+i)2hw−1iIhwiI ≤ C[w−1]Ad
2
hb, hIi2 ≤ C[w−1]Ad
2kbk2
BM O d ,
)2Ihw−1iIhwiI(cid:19)1/2
.
(6.11)
(Aw−1
I
J XI∈D(J)
hb, hIi2 ,
J XI∈D(J)
)2Ihw−1iIhwiI ≤ C [w−1]Ad
1
2
.
Thus embedding condition (6.8) holds. For (6.9), by Cauchy-Schwarz inequality and (6.11) we have
1
J XI∈D(J)
I
(hb, hI−i + hb, hI+i)Aw−1
≤ C(cid:18) 1
J XI∈D(J)
pIhw−1iI
hb, hIi2hw−1iI(cid:19)1/2(cid:18) 1
J XI∈D(J)
J XI∈D(J)(cid:18)hw−1iI+ − hw−1iI−
(cid:19)2
2hw−1iI
)2Ihw−1iI =
1
By Theorem 4.5,
1
J XI∈D(J)
(Aw−1
I
(Aw−1
I
)2Ihw−1iI(cid:19)1/2
.
Ihw−1iI ≤ C[w−1]Ad
2hw−1iJ .
Similarly with (6.7), we have
1
J XI∈D(J)
hb, hIi2hw−1iI ≤ [w−1]A2kbk2
BM O dhw−1iJ .
To finish, we must estimate (6.10). In a similar way with (6.9), we need to estimate
(cid:18) 1
J XI∈D(J)
hb, hIi2hwiI(cid:19)1/2(cid:18) 1
J XI∈D(J)
(Aw−1
I
)2IhwiI(cid:19)1/2
.
17
By Lemma 4.6, we have
1
J XI∈D(J)
(Aw−1
I
)2IhwiI ≤ [w−1]Ad
2
= [w−1]Ad
2
I
1
(Aw−1
)2Ihw−1i−1
J XI∈D(J)
J XI∈D(J)(cid:18)hw−1iI+ − hw−1iI−
hw−1i3
1
(cid:19)2
I
I
I
≤ C[w−1]Ad
2hw−1iJ .
This completes the proof of Lemma 6.1.
Due to the almost self adjoint property of the Hilbert transform, a certain bound for π∗
b H
immediately returns the same bound for Hπb. However we have to prove the boundedness of Sπb
independently because S is not self adjoint.
7 Linear bound for Sπb
Lemma 7.1. Let w ∈ Ad
2 and b ∈ BM O d. Then, there exists C so that
kSπbkL2(w)→L2(w) ≤ C[w]Ad
2kbkBM O d .
Proof. We are going to prove Lemma 7.1 by showing
hSπb(w−1f ), giw ≤ C[w]Ad
2kbkBM O dkfkL2(w−1)kgkL2(w) ,
(7.1)
for any positive function f, g ∈ L2 . Since hSπb(f ), hIi = sgn(I)hπb(f ), h Ii = sgn(I)hfi Ihb, h Ii , We
have
By expanding g in the disbalanced Haar system for L2(w) ,
Sπb(f ) =XI∈D
sgn(I)hfi Ihb, h IihI .
hw−1fi Ihb, h Iisgn(I)hhI , giw
hSπb(w−1f ), giw =XI∈D
=XI∈DXJ∈D
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
J iwhhI , hw
J iw .
Since hhI , hw
J iw could be non zero only if J ⊇ I , we can split above sum into three parts,
and
XI∈D
XI∈D
XI∈D XJ:J ) I
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
I iwhhI , hw
I iw,
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
I iwhhI , hw
I iw,
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
J iwhhI , hw
J iw .
18
(7.2)
(7.3)
(7.4)
We claim that all sums, (7.2), (7.3), and (7.4), can be bounded with a bound that depends on
[w]Ad
2kbkBM O d at most linearly. Since hhI , hw
I iw ≤ hwi1/2
, we can estimate (7.2)
I
(cid:12)(cid:12)(cid:12)(cid:12)XI∈D
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
≤(cid:18)XI∈D
hw−1i2
Ihfi2
≤ CkgkL2(w)[w]1/2
Ad
I iwhhI , hw
I iw(cid:12)(cid:12)(cid:12)(cid:12)
I,w−1hb, h Ii2hwiI(cid:19)1/2(cid:18)XI∈D
2(cid:18)XI∈D
hg, hw
I,w−1hb, hIi2hw−1iI(cid:19)1/2
hfi2
.
I i2(cid:19)1/2
By Weighted Carleson Embedding Theorem 4.4,
XI∈D
is provided by
hfi2
I,w−1hb, hIi2hw−1iI ≤ C[w]Ad
2kbk2
BM O dkfk2
L2(w−1)
1
J XI∈D(J)
hb, hIi2hw−1iI ≤ [w]Ad
2kbk2
BM O dhw−1iJ
which we already have in (6.7) . Thus, we have
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
XI∈D
Similarly to (7.2), we can estimate (7.3) using hhI , hw
I iwhhI , hw
I iw ≤ hwi1/2
I ≤ √2hwi1/2
I
I iw ≤ C[w]Ad
2kbkBM O dkfkL2(w−1)kgkL2(w) .
(7.5)
(cid:12)(cid:12)(cid:12)(cid:12)XI∈D
sgn(I)hw−1i Ihfi I,w−1hb, h Iihg, hw
I iw(cid:12)(cid:12)(cid:12)(cid:12)
I iwhhI , hw
hw−1i Ihfi I,w−1 hb, h Iihg, hw
hw−1iIhfiI,w−1 hb, hIihg, hw
I
I iwhwi1/2
I iwhwi1/2
I
≤
√2XI∈D
= 2√2XI∈D
≤ 2√2(cid:18)XI∈D
hw−1i2
Ihfi2
2(cid:18)XI∈D
≤ CkgkL2(w)[w]1/2
≤ C[w]Ad
Ad
I,w−1hb, hIi2hwiI(cid:19)1/2(cid:18)XI∈D
hg, hw
I,w−1hb, hIi2hw−1iI(cid:19)1/2
hfi2
I i2(cid:19)1/2
2kbkBM O dkfkL2(w−1)kgkL2(w) .
Since hw
J is constant on I, for J ) I and we denote this constant by hw
J ( I) . Then we know by (2.7),
XJ:J ) I
J iw = XJ:J ) I
hg, hw
J iwhhI , hw
hg, hw
J iwhw
J ( I)hhI , wi = hgi I,whhI , wi .
19
Thus, we can rewrite (7.4)
(cid:12)(cid:12)(cid:12)(cid:12)XI∈D
sgn(I)hw−1i Ihfi I,w−1hb, h Iihgi I,whhI , wi(cid:12)(cid:12)(cid:12)(cid:12)
≤XI∈D
=XI∈D
hw−1i Ihfi I,w−1 hb, h Iihgi I,whhI , wi
hw−1iI hb, hIi (hhI− , wi + hhI+, wi)hfiI,w−1hgiI,w .
(7.6)
(7.7)
2kbkBM O dkfkL2(w−1)kgkL2(w) . We are going to prove
We claim the sum (7.7) is bounded by [w]Ad
it using Petermichl's Bilinear Embedding Theorem 4.1. Thus, we need to show that the following
three embedding conditions hold,
∀ J ∈ D,
∀ J ∈ D,
∀ J ∈ D,
1
1
J XI∈D(J)
J XI∈D(J)
J XI∈D(J)
1
hb, hIihw−1iI (hhI−, wi + hhI+, wi)
hb, hIihw−1iI (hhI−, wi + hhI+, wi)
1
hwiI ≤ C[w]Ad
hw−1iI ≤ C[w]Ad
1
2kbkBM O d hw−1iJ , (7.8)
2kbkBM O d hwiJ , (7.9)
hb, hIihw−1iI (hhI−, wi + hhI+, wi) ≤ C[w]Ad
2kbkBM O d .
(7.10)
After we split the sum in (7.8):
1
J XI∈D(J)
hb, hIihw−1iIhhI−, wi
1
hwiI
+
1
J XI∈D(J)
hb, hIihw−1iIhhI+, wi
1
hwiI
,
we start with Cauchy-Schwarz inequality to estimate the first sum of embedding condition (7.8),
1
J XI∈D(J)
hb, hIihw−1iIhhI−, wi
≤(cid:18) 1
J XI∈D(J)
hb, hIi2hw−1i2
A2(cid:18) 1
J XI∈D(J)
≤ C[w]1/2
2kbkBM O dhw−1iJ .
≤ C[w]Ad
1
=
1
hwiI
J XI∈D(J)
IhwiI(cid:19)1/2(cid:18) 1
hb, hIi2hw−1i(cid:19)1/2(cid:18) 1
hb, hIihw−1iIpI−∆I−w
hwiI
I(cid:19)1/2
J XI∈D(J)
I−∆I−w2 1
hwi3
I−(cid:19)1/2
J XI∈D(J)
1
hwi3
I−∆I−w2
(7.11)
Inequality (7.11) due to Lemma 4.6 and (6.7) . Also, the other sum can be estimated by exactly the
same method. Thus we have the embedding condition (7.8). To see the embedding condition (7.9),
it is enough to show
1
J XI∈D(J)
hb, hIihhI−, wi ≤ C[w]Ad
2kbkBM O dhwiJ ,
20
as we did above. We use Cauchy-Schwarz inequality for embedding condition (7.9), then
1
J XI∈D(J)
hb, hIihhI−, wi =
1
J XI∈D(J)
≤(cid:18) 1
J XI∈D(J)
≤ CkbkBM O dhwi1/2
hb, hIipI−∆I−w
hw−1iI(cid:19)1/2(cid:18) 1
hb, hIi2
J (cid:18)[w]A2
J XI∈D(J)
1
J XI∈D(J)
I−∆I−w2
I−∆I−w2hw−1iI(cid:19)1/2
hwiI−(cid:19)1/2
1
(7.12)
≤ C[w]Ad
2kbkBM O dhwiJ .
(7.13)
Here inequality (7.12) uses Lemma 4.7, and inequality (7.13) uses the fact that hwi−1
and Theorem 4.5 after shifting the indices.
I ≤ 2hwi−1
I−
If we show the embedding condition (7.10), then we can immediately finish the estimate for (7.7)
2kbkBM O dkfkL2(w−1)kgkL2(w) . Combining this and (7.5) will give us our desire
with bound C[w]Ad
result.
8 Proof for embedding condition (7.10)
The following lemma lies at the heart of the matter for the proof of the embedding condition (7.10) .
Lemma 8.1. There is a positive constant C so that for all dyadic interval J ∈ D
J hw−1i1/4
I (cid:18)∆I+w + ∆I−w
≤ Chwi1/4
Ihwi1/4
hw−1i1/4
(cid:19)2
J
I
hwiI
1
J XI∈D(J)
,
(8.1)
whenever w is a weight. Moreover, if w ∈ Ad
2 then for all J ∈ D
1
J XI∈D(J)
IhwiIhw−1iI(cid:18)∆I+w + ∆I−w
hwiI
(cid:19)2
≤ C [w]Ad
2
.
Proof of condition (7.10). By using Cauchy-Schwarz inequality and Lemma 8.1, we have:
1
J XI∈D(J)
1
hb, hIihw−1iI (hhI−, wi + hhI+, wi)
J XI∈D(J)
√2(cid:18) 1
hb, hIihw−1iIrI
J XI∈D(J)
hb, hIi2hw−1iIhwiI(cid:19)1/2(cid:18) 1
1
2 (cid:0)hwiI−+ − hwiI−− + hwiI++ − hwiI+−(cid:1)
Ihw−1iIhwi−1
J XI∈D(J)
=
≤
I (∆I+w + ∆I−w)2(cid:19)1/2
≤ C[w]Ad
2kbkBM O d .
21
We turn to the proof of Lemma 8.1. In the first place, we need to revisit some properties of
function B(u, v) := 4√uv on the domain D0 which is given by
It is known, we refer to [2], that B(u, v) satisfies the following differential inequality in D0
{(u, v) ∈ R2
+ : uv ≥ 1/2} .
− (du, dv)d2B(u, v)(du, dv)t ≥
1
8
v1/4
u7/4du2 .
Furthermore, this implies the following convexity condition. For all (u, v), (u±, v±) ∈ D0 ,
B(u, v) −
B(u+, v+) + B(u−, v−)
2
v1/4
u7/4
≥ C1
(u+ − u−)2 ,
where u = (u+ + u−)/2 and v = (v+ + v−)/2 . Let us define
A(u, v, ∆u) := aB(u, v) + B(u + ∆u, v) + B(u − ∆u, v) ,
(8.2)
(8.3)
on the domain D1 with some positive constant a > 0 . Here (u, v, ∆u) ∈ D1 means all pairs
(u, v), (u + ∆u, v), (u − ∆u, v) ∈ D0 . Then A has the size property and the convexity property:
if (u, v, ∆u) ∈ D1,
then 0 ≤ A(u, v, ∆u) ≤ (a + 2) 4√uv ,
(8.4)
(8.5)
and
A(u, v, ∆u) −
1
2(cid:2)A(u+, v+, ∆u1) + A(u−, v−, ∆u2)(cid:3) ≥ C2
v1/4
u7/4
(∆u2
1 + ∆u2
2) ,
where u = (u+ + u−)/2 , v = (v+ + v−)/2, and ∆u = (u+ − u+)/2 . The property (8.5) is directly
from the definition of function B(u, v) . At the end, ∆u will play the role of ∆Iw, ∆u1 is ∆I+w,
and ∆u2 is ∆I−w . We can rewrite left hand side of the inequality (8.5) as follows
1
A(u, v, ∆u) −
= aB(u, v) + B(u + ∆u, v) + B(u − ∆u, v) −
2(cid:2)A(u+, v+, ∆u1) + A(u−, v−, ∆u2)(cid:3)
1
2(cid:2)aB(u+, v+) + B(u+ + ∆u1, v+) + B(u+ − ∆u1, v+)
= aB(u, v) −
+ aB(u−, v−) + B(u− + ∆u1, ∆u−) + B(u− − ∆u1, v−)(cid:3)
2hB(u + ∆u + ∆u1, v + ∆v)
+ B(u + ∆u − ∆u1, v + ∆v) + B(u − ∆u + ∆u2, v − ∆v) + B(u − ∆u − ∆u2, v − ∆v)i .
(B(u+, v+) + B(u−, v−)) + B(u+, v) + B(u−, v) −
a
2
1
(8.6)
Using Taylor's theorem:
B(u + u0, v + v0) = B(u, v) +∇B(u, v)(u0, v0)t +Z 1
0
(1− s)(u0, v0)d2B(u + su0, v + sv0)(u0, v0)tds ,
22
and the convexity condition of B(u, v), we are going to estimate the lower bounds of (8.6).
B(u + ∆u + ∆u1, v + ∆v)
1
−
2
= −
0
1
1
1
2(cid:16)B(u, v) + ∇B(u, v)(∆u + ∆u1, ∆v)t(cid:17)
−Z 1
(1 − s)(∆u + ∆u1, ∆v)d2B(u + s(∆u + ∆u1), v + s∆v)(∆u + ∆u1, ∆v)tds
2(cid:16)B(u, v) + ∇B(u, v)(∆u + ∆u1, ∆v)t(cid:17) +
2(cid:16)B(u, v) + ∇B(u, v)(∆u + ∆u1, ∆v)t(cid:17) +
2(cid:16)B(u, v) + ∇B(u, v)(∆u + ∆u1, ∆v)t(cid:17) +
2(cid:16)B(u, v) + ∇B(u, v)(∆u + ∆u1, ∆v)t(cid:17) +
(1 − s)(v + s∆v)1/4ds
Z 1
∆v
v
(1 − s)
(∆u + ∆u1)2
0
(∆u + ∆u1)2v1/4
8(4u)7/4
v1/4
1
u7/4
(u + s(∆u + ∆u1))7/4
(1 − s)(1 + s
8Z 1
(∆u + ∆u1)2 .
(v + s∆v)1/4
72 · 43/4
Z 1
8(4u)7/4
1
1
1
0
0
≥ −
≥ −
≥ −
≥ −
Inequality (8.7) is due to the following inequalities
(∆u + ∆u1)2ds
)1/4ds
(8.7)
(8.8)
∆u = u+ − u−
2
≤ u+ + u−
2
= u and ∆u1 = u++ − u+−
2
≤ u+ ≤ 2u .
Since (1 − s)1/4 ≤ (1 − βs)1/4 ≤ (1 + βs)1/4 for any β < 1, it is clear that
Z 1
0
(1 − s)(1 + βs)1/4ds ≥Z 1
0
(1 − s)5/4ds =
4
9
,
and this allows the inequality (8.8). With the same arguments, we also estimate the following lower
bounds:
−
1
2hB(u + ∆u − ∆u1, v + ∆v) + B(u − ∆u + ∆u2, v − ∆v) + B(u − ∆u − ∆u2, v − ∆v)i
≥ −
−
−
1
1
2(cid:16)B(u, v) + ∇B(u, v)(∆u − ∆u1, ∆v)t(cid:17) +
2(cid:16)B(u, v) + ∇B(u, v)(−∆u + ∆u2,−∆v)t(cid:17) +
2(cid:16)B(u, v) + ∇B(u, v)(−∆u − ∆u2,−∆v)t(cid:17) +
1
1
72 · 43/4
1
1
72 · 43/4
72 · 43/4
v1/4
u7/4 (∆u − ∆u1)2
v1/4
u7/4
v1/4
u7/4
(−∆u + ∆u2)2
(∆u + ∆u2)2 .
(8.9)
We can have the following inequality by combining (8.8), (8.9) and (8.6),
A(u, v, ∆u) −
1
2(cid:2)A(u+, v+, ∆u1) + A(u−, v−, ∆u2)(cid:3)
a
2
1
≥ (a − 2)B(u, v) −
v1/4
u7/4 (2∆u2 + ∆u2
(∆u2
2) .
36 · 43/4
1
1 + ∆u2
+
v1/4
u7/4
≥
36 · 43/4
1 + ∆u2
2)
(B(u+, v+) + B(u−, v−)) + B(u+, v) + B(u−, v)
(8.10)
23
To see the inequality (8.10), using convexity condition of B(u, v) = 4√uv and inequality: (1− s)u ≤
u − s∆u ≤ u + s∆u ,
a
(a − 2)B(u, v) −
1
2
0
2(cid:16)B(u+, v+) + B(u−, v−)(cid:17) + B(u+, v) + B(u−, v)
= a(cid:18)B(u, v) −
−(cid:18) 3
∆u2Z 1
∆u2Z 1
∆u2 −
∆u2 −
2(cid:19) v1/4
3
(B(u+, v+) + B(u−, v−)(cid:19)
(1 − s)v1/4(u + s∆u)−7/4ds +
6
16
∆u2 v1/4
3
u7/4
2
=(cid:18)aC1 −
u7/4Z 1
16
v1/4
u7/4
v1/4
u7/4
(1 − s)−3/4ds
0
∆u2 v1/4
≥ aC1
= aC1
∆u2 .
u7/4
3
16
0
(1 − s)v1/4(u − s∆u)−7/4ds(cid:19)
(8.11)
Choosing a constant a sufficiently large so that aC1 > 3/2, quantity in (8.11) remains positive.
This observation and discarding nonnegative terms yield inequality (8.10). Choosing the constant
C2 = 1/(36 · 43/4) in (8.5) completes the proof of the concavity property of A(u, v, ∆u) . We now
turn to the proof of Lemma 8.1.
Proof of Lemma 8.1. Let uI := hwiI , vI := hw−1iI , u± = uI±, v± = vI±, ∆uI = ∆I w, ∆u1 =
∆uI+, and ∆u2 = ∆uI− . Then by Hölder's inequality (u, v, ∆u), (u+, v+, ∆u1), and (u−, v−, ∆u2)
belong to D1 . Fix J ∈ D , by properties (8.4) and (8.5)
(a + 2)J 4phwiJhw−1iJ ≥ JA(uJ , vJ , ∆uJ )
≥
1
2(cid:18)J+A(u+, v+, ∆u1) + J−A(u−, v−, ∆u2)(cid:19) + JC hw−1iJ
hwi7/4
J
Since A(u, v, ∆u) ≥ 0, iterating the above process will yield
(∆J+u2 + ∆J−u2) .
J 4phwiJhw−1iJ ≥ C XI∈D(J)
Ihw−1i1/4
I
hwi−7/4
I
(∆I+w2 + ∆I−w2) .
(8.12)
Also, one can easily have
and
J 4phwiJhw−1iJ ≥ C XI∈D(J)
J 4phwiJhw−1iJ ≥ C XI∈D(J)
Ihw−1i1/4
I
hwi−7/4
I ∆I+w2 ,
Ihw−1i1/4
I
hwi−7/4
I ∆I−w2 .
(8.13)
(8.14)
24
Then,
1
J XI∈D(J)
Ihw−1i1/4
I
hwi−7/4
I
(∆I+w + ∆I−w)2
=
≤
≤
I
1
1
J(cid:18) XI∈D(J)
Ihw−1i1/4
+ ∆I−w2) + 2 XI∈D(J)
J XI∈D(J)
Ihw−1i1/4
+ 2(cid:18) XI∈D(J)
4phwiIhw−1iI .
Ihw−1i1/4
3
C
I
I
hwi−7/4
I
(∆I+w2
Ihw−1i1/4
I
hwi−7/4
I
(∆I+w∆I−w)(cid:19)
hwi−7/4
I
(∆I+w2 + ∆I−w2)
hwi−7/4
I ∆I+w2(cid:19)1/2(cid:18) XI∈D(J)
Ihw−1i1/4
I
I ∆I−w2(cid:19)1/2!
hwi−7/4
9 Recently developed tools and their applications
The dyadic shift operator was first introduced in [20] to replace the weighted norm estimate for
the Hilbert transform. It was also encountered in [29], so Riesz transforms can be obtained as the
result of averaging some dyadic shift operator. Recently, in [17] and [7], a more general class of
dyadic shift operators, so called the Haar shift operators were introduced. The Hilbert transform,
Riesz transforms, and Beurling-Ahlfors operator are in the convex hull of this class, as they can
be written as appropriate averages of Haar shift operators. Let Dn denote the collection of dyadic
cubes in Rn , Dn(Q) denotes dyadic subcubes of Q , and Q denotes the volume of the dyadic cube
Q . We start with some definitions.
Definition 9.1. A Haar function on a cube Q ⊂ Rn is a function HQ such that
(a) HQ is supported on Q, and is constant on Dn(Q) .
(b) kHQk∞ ≤ Q−1/2 .
(c) HQ has a mean zero.
Definition 9.2. Given an integer τ > 0 , we say an operator of the following form is in the first
class of Haar shift operators of index τ
Tτ f (x) = XQ∈Dn XQ′,Q′′∈D(Q)
2−τ nQ≤Q′,Q′′
aQ′,Q′′hf, HQ′iHQ′′(x) ,
where the constant aQ′,Q′′ satisfy the following size condition:
aQ′,Q′′ ≤ C(cid:18)Q′
Q · Q′′
Q(cid:19)1/2
25
.
(9.1)
Note that once a choice of Haar functions has been made {HQ}Q∈Dn , then this is an orthogonal
family, such that kHQkL2 ≤ 1 , so one could normalize in L2 . Note that one can easily see that the
dyadic shift operator S belongs to the first class of a Haar shift operator of index τ = 1 with
aI ′,I ′′ =(cid:26) ±1 for I ′ = I, I ′′ = I∓
otherwise .
0
One of the main result in [17] and [7] is the following
Theorem 9.3 ([17], [7]). Let T be in the first class of Haar shift operators of index τ . Then for all
w ∈ Ad
2 , there exists C(τ, n) which only depends on τ and n such that
kTkL2(w)→L2(w) ≤ C(τ, n)[w]Ad
2
.
As a consequence of this Theorem, linear bounds for the Hilbert transform, Riesz transforms,
and the Beurling-Ahlfors operator are recovered. There are now two different proofs of Theorem
(9.3) in [17] and [7]. The commutator [λb, S] is also in the first class of Haar shift operators of index
τ = 1 . Recall the observation in the section 4,
Then we can see
[λb, S](f ) = −XI∈D
aI ′,I ′′ =(cid:26) −∆I b
0
∆Ibhf, hIi(hI+ + hI−) .
for I ′ = I, I ′′ = I±
otherwise ,
moreover aI ′,I ′′ = ∆I b ≤ 2kbkBM O d this means the constant aI ′,I ′′ satisfy the size condition (9.1)
with C = 2√2kbkBM O d . These observations, Theorem 5.1, and Theorem 9.3 immediately recover
the quadratic bound for the commutator of the Hilbert transform which was proved in this paper.
We now define the second class of Haar shift operators of index τ .
Definition 9.4. Given an integer τ > 0 , we say an operator T of the form in Definition 9.2 is in the
second class of Haar shift operators of index τ , if T is bounded on L2 and the function HQ satisfy
the condition (a) and (b) in Definition 9.2.
The second class of Haar shift operators is more general than the first class. One can easily
observe that the operators πb, Sπb and π∗
b S do not satisfy the condition (c) on Definition 9.1, however
these operators satisfy the conditions of Definition 9.4. Note that the n-variable paraproduct is a
sum of 2n − 1 of the second class of Haar shift operator of index 1, the restricted n-variable dyadic
paraproduct
πbf = XQ∈DnhfiQhb, HQiHQ .
In [14], the linear estimate for the maximal truncations of these operators is presented. This also
recovers our linear bound estimates for Sπb and π∗
b S. On the other hand, authors in [7] also
reproduce the linear estimate for the dyadic paraproduct with different technique.
Lemma 9.5. Let Tτ a Haar shift operator of the first class, then [λb, Tτ ] is an operator of the same
class.
Proof. We are going to use the restricted multi-variable λb operator which is
λbf = XQ∈DnhbiQhf, HQiHQ .
26
One can get the n-variable λb operator by summing over 2n − 1 of restricted λb operator. Observe
that,
[λb, Tτ ]f = λb(Tτ f ) − Tτ (λbf )
= XQ∈Dn XQ′,Q′′∈D(Q)
= XQ∈Dn XQ′,Q′′∈D(Q)
2−τ nQ≤Q′,Q′′
2−τ nQ≤Q′,Q′′
aQ′,Q′′hbiQ′′hf, HQ′iHQ′′ − XQ∈Dn XQ′,Q′′∈D(Q)
2−τ nQ≤Q′,Q′′
aQ′,Q′′hbiQ′hf, HQ′iHQ′′
aQ′,Q′′(hbiQ′′ − hbiQ′)hf, HQ′iHQ′′ .
Since
[λb, Tτ ] remains in the same class of Tτ .
(cid:12)(cid:12) aQ′,Q′′(hbiQ′′ − hbiQ′)(cid:12)(cid:12) ≤ kbkBM O aQ′,Q′′ ,
Theorem 9.3 and Lemma 9.5 allow to extend our result to more general class of commutators
including the Riesz transforms and the Beurling-Ahlfors operator as in Theorem 1.3.
10 Sharp bounds
In this section, we start proving that the quadratic estimate in Theorem 1.5 is sharp, by showing
an example which returns quadratic bound. This example was discovered by C. Peréz [22] who
is kindly allowing us to reproduce it in this paper. The same calculations show that the bounds
in Theorem 1.1 are also sharp for p 6= 2 and 1 < p < ∞ . Variations over this example will then
show that the bounds in Theorem 1.3 are sharp for the Riesz transforms and the Beurling-Ahlfors
operator as well.
10.1 The Hilbert transform
Consider the weight, for 0 < δ < 1 :
It is well known that w is an A2 weight and
w(x) = x1−δ .
[w]A2 ∼
1
δ
.
We now consider the function f (x) = x−1+δχ(0,1)(x) and BMO function b(x) = log x . We claim
that
For 0 < x < 1 , we have
k[b, H]f (x) ≥
1
δ2 f (x) .
[b, H]f (x) =Z 1
0
log x − log y
x − y
= x−1+δZ 1/x
0
y−1+δdy =Z 1
t(cid:17)
log(cid:16) 1
t−1+δdt .
0
1 − t
log(x/y)
x − y
y−1+δdy
27
Now,
0
and since log(1/t)
1−t
Z 1/x
log(1/t)
1 − t
But since
Z 1
0
log(1/t)
1 − t
our claim follows and
is positive for (0, 1) ∪ (1,∞) we have for 0 < x < 1
t−1+δdt .
log(1/t)
1 − t
0
log(1/t)
1 − t
t−1+δdt =Z 1
[b, H]f (x) > x−1+δZ 1
t−1+δdt >Z 1
0
0
k[b, H]fkL2(w) ≥
t−1+δdt +Z 1/x
1
log(1/t)
1 − t
t−1+δdt ,
log(1/t)t−1+δdt =Z ∞
1
δ2kfkL2(w) ∼ [w]2
0
A2kfkL2(w) .
se−sδds =
1
δ2 ,
(10.1)
(10.2)
A first approximation of what the bounds in Lp(w) is given by an application of the sharp extrap-
olation theorem for the upper bound, paired with the knowledge of the sharp bound on L2(w) to
obtain a lower bound.
Proposition 10.1. For 1 < p < ∞ there exist constants c and C only depending on p such that
2 min{1, 1
Ap
p−1 }
kbkBM O ≤ k[b, H]kLp(w)→Lp(w) ≤ C[w]
c[w]
for all b ∈ BM O .
Proof. Because the upper bound in (10.3) is the direct consequence of the quadratic bound in the
Theorem 1.5 and sharp extrapolation theorem, we will only prove the lower bound. Let us assume
that, for 1 < r < 2 and α < 1 ,
kbkBM O ,
(10.3)
2 max{1, 1
Ap
p−1 }
This and the sharp extrapolation theorem return
k[b, H]kLr (w)→Lr(w) ≤ C[w]2α
ArkbkBM O .
This contradicts to the sharpness (p = 2) . Similarly, one can conclude for p > 2 .
k[b, H]kL2(w)→L2(w) ≤ C[w]2α
A2kbkBM O .
We now consider the weight w(x) = x(1−δ)(p−1) then w is an Ap weight with [w]Ap ∼ δ1−p . By
(10.1) and (10.2) we have
k[b, H]fkLp(w) ≥
1
δ2 kfkLp(w) = (δ1−p)
2
p−1kfkLp(w) ∼ [w]
p−1
Ap kfkLp(w) .
2
This shows the upper bound in (10.3) is sharp for 1 < p ≤ 2 . We use the duality argument to
see the sharpness of the quadratic estimate for p > 2 . Note that the commutator is a self-adjoint
operator:
hbH(f ) − H(bf ), gi = hf, H ∗(bg)i − hf, bH ∗(g)i = hf, bH(g) − H(bg)i .
Consider 1 < p ≤ 2 and set u = w1−p′
, then
k[b, H]kLp′
(u)→Lp′
(u) = k[b, H]kLp′
(w1−p′
)→Lp′
(w1−p′
) = k[b, H]∗kLp′
2
(w1−p′
)→Lp′
(w1−p′
)
= k[b, H]kLp(w)→Lp(w) ≤ CkbkBM Od[w]
= CkbkBM O[w1−p′
Ap′ = CkbkBM O[u]2
]2
1−p
Ap
Ap′ .
(10.4)
Since the inequality in (10.4) is sharp, we can conclude that the result of Theorem 1.1 is also sharp
for p > 2 .
28
10.2 Beurling-Ahlfors operator
Recall the Beurling-Ahlfors operator B is given by convolution with the distributional kernel p.v.1/z2:
Bf (x, y) = p.v.
f (x − u, y − v)
(u + iv)2
dudv .
Then the commutator of the Beurling-Ahlfors operator can be written:
[b,B]f (x, y) = p.v.
b(x, y) − b(s, t)
((x − s) + i(y − t))2 f (s, t) dsdt .
It was observed, in [9], that the linear bound for the Beurling-Ahlfors operator is sharp in L2(w) ,
with weights w(z) = zα and functions f (z) = z−α where α < 2 . Similarly, we consider weights
w(z) = z2−δ where 0 < δ < 1. Note that w(z) = z2−δ : C → [0,∞) is a A2-weight with [w]A2 ∼
δ−1 . We also consider a BMO function b(x) = log z . Let E = {(r, θ) 0 < r < 1, 0 < θ < π/2}
and Ω = {(r, θ) 1 < r < ∞, π < θ < 3π/2} We are going to estimate [b,B]f (z) for z ∈ Ω with a
function f (z) = zδ−2χE(z) . Let z = x + iy and ζ = s + ti . Then, for z ∈ Ω ,
1
πZR2
πZR2
1
[b,B]f (z) =
=
=
1
1
π(cid:12)(cid:12)(cid:12)(cid:12)ZE
π(cid:12)(cid:12)(cid:12)(cid:12)ZE
π(cid:12)(cid:12)(cid:12)(cid:12)ZE
1
dζ(cid:12)(cid:12)(cid:12)(cid:12)
(b(z) − b(ζ))f (ζ)
(z − ζ)2
b(x, y) − b(s, t)
((x − s) + i(y − t))2 f (s, t)dsdt(cid:12)(cid:12)(cid:12)(cid:12)
(log z − log ζ) ζδ−2((x − s)2 − (y − t)2)
+ iZE
((x − s)2 + (y − t)2)2
(log z − log ζ) ζδ−2(2(x − s)(y − t))
((x − s)2 + (y − t)2)2
dsdt
dsdt(cid:12)(cid:12)(cid:12)(cid:12) .
4
For z ∈ Ω and ζ ∈ E , we have (x− s)(y− t) ≥ xy and by triangle inequality ((x− s)2 + (y− t)2)2 =
z − ζ4 ≤ ( z + ζ)4 . After neglecting the positive term (real part), we get
drdθ(cid:19)2
zdt(cid:19)2
π2(cid:18)xyZ π/2
Z 1
z4Z 1/z
log(1/t)( zt)δ−1
( z + ζ)4
log( z/r)rδ−1
log( z/ ζ) ζδ−2
log( z/r)rδ−2r
[b,B]f (z)2 ≥
π2(cid:18)xyZE
= x2y2(cid:18) 1
z4Z 1
= x2y2(cid:18) 1
z4−δ Z 1/z
dsdt(cid:19)2
(1 + r/ z)4 dr(cid:19)2
= x2y2(cid:18) 1
dt(cid:19)2
log(1/t)tδ−1
( z + r)4
(1 + t)4
(1 + t)4
=
4
0
0
0
0
0
.
Since z/( z + 1) ≤ 1/(1 + t) , for t < 1/ z , we have
[b,B]f (z)2 ≥ x2y2(cid:18)
1
z−δ( z + 1)4Z 1/z
0
log(1/t)tδ−1 dt(cid:19)2
x2y2
z−2δ( z + 1)8(cid:18) z−δ(1 + δ log z)
δ2
(cid:19)2
=
=
x2y2
( z + 1)8
(1 + δ log z)2
.
δ4
29
Then, we can estimate the L2(w)-norm as follows.
k[b,B]fk2
L2(w) ≥
=
=
=
=
r4 cos2 θ sin2 θ(1 + δ log r)2
z2−δ dxdy
(r + 1)8
π
1
1
π
( z + 1)8
δ4ZΩ
x2y2(1 + δ log z)2
δ4Z ∞
1 Z 3π/2
δ416Z ∞
δ4212Z ∞
δ4212(cid:18) 1
δ3 (cid:19) =
2δ
δ2 +
(r + 1)8
2δ2
+
π
π
δ
1
1
r7−δ(1 + δ log r)2
dr ≥
(1 + δ log r)2r−1−δ dr
5π
212 ·
1
δ5 .
r3−δdrdθ
π
δ416Z ∞
1
r7−δ(1 + δ log r)2
(2r)8
dr
L2(w) = π/2δ , we have that k[b,B]fkL2(w)/kfkL2(w) ∼ δ−2 , which allows to
Combining with kfk2
conclude that the quadratic bound for the commutator with the Beurling-Ahlfors operators is sharp
in L2(w) . Same calculations with weights w(z) = z(2−δ)(p−1) and functions f (z) = z(δ−2)(p−1)
will provide the sharpness for 1 < p ≤ 2 , and it is sufficient to conclude for all 1 < p < ∞ because
the Beurling-Ahlfors operator is essentially self adjoint operator (B∗ = eiφB) , so the commutator
of the Beurling-Ahlfors operator is also self adjoint.
10.3 Riesz transforms
Consider weights w(x) = xn−δ and functions f (x) = xδ−nχE(x) where E = {x x ∈ (0, 1)n ∩
B(0, 1)} , and a BMO function b(x) = log x . It was observed that xn−δ is an A2-weight in Rn
with [w]A2 ∼ δ−1 . We are going to estimate [b, Rj]f over the set Ω = {y ∈ B(0, 1)c yi < 0 for all i =
1, 2, ..., n} , where Rj stands for the j-th direction Riesz transform on Rn and is defined as follows:
Rjf (x) = cnp.v.ZRn
yj
yn+1 f (x − y) dy ,
1 ≤ j ≤ n ,
where cn = Γ((n + 1)/2)/π(n+1)/2 . One can observe that, for all x ∈ E and fixed y ∈ Ω ,
yj − xj ≥ yj and y − x ≤ y + x .
Cn denotes a constant depending only on the dimension that may change from line to line. Then,
[b, Rj]f (y) =(cid:12)(cid:12)(cid:12)(cid:12)ZE
(yj − xj)(log y − log x) xδ−n
y − xn+1
log( y/r)rδ−nrn−1
≥ yjZE∩Sn−1Z 1
yn+1−δZ 1/y
y−δ( y + 1)n+1(cid:18) y−δ(1 + δ log y)
( y + r)n+1
log(1/t)tδ−1
(1 + t)n+1 dt ≥
Cn yj
Cn yj
δ2
=
=
0
0
dx
log( y/ x) xδ−n
( y + x)n+1
dx(cid:12)(cid:12)(cid:12)(cid:12) ≥ yjZE
drdσ = Cn yjZ 1/y
yn+1−δ(cid:18) y
y + 1(cid:19)n+1Z 1/y
Cn yj
0
0
log(1/t)(t y)δ−1 y
( y + yt)n+1
dt
log(1/t)tδ−1 dt
(cid:19) =
Cn yj
( y + 1)n+1
1 + δ log y
.
δ2
30
We now can bound from below the L2(w)-norm as follows.
y2
j (1 + δ log y)2
( y + 1)2n+2
yn−δdy
j r2(1 + δ log r)2rn−δrn−1
γ2
(r + 1)2n+2
drdσ(γ)
k[b, Rj ]f (x)k2
L2(w) >
≥
≥
=
Cn
Cn
Cn
Cn
δ4 ZΩ
δ4 ZΩ∩Sn−1Z ∞
δ4 Z ∞
δ4 Z ∞
1
1
1
(1 + δ log r)2r2n−δ+1
dr
r2n+2
(1 + δ log r)2r−δ−1 dr =
Cn
δ5 ,
which establishes sharpness for the commutator of the Riesz transform when p = 2 . Since R∗
j =
−Rj, one can easily check that the commutator of Riesz transforms are also self-adjoint operators.
Furthermore, choosing weight w(x) = x(n−δ)(p−1), we will obtain the sharpness for 1 < p < ∞ by
the same argument we used in the case of the Hilbert transform.
11 Acknowledgments
This work is part of the author's Ph.D dissertation [4]. The author like to thank Professor Carlos
Pérez for his comments and useful suggestions. Finally, the author is also very grateful to his
graduate adviser María Cristina Pereyra for her suggestions and helpful interaction.
References
[1] J. Alvarez, R. J. Bagby, D. S. Kurtz and C. Pérez, Weighted estimates for commutators of
linear operators Studia Math. 104(2) (1993) 195-209.
[2] O. Beznosova, Linear bound for dyadic paraproduct on weighted Lebesgue space L2(w) J. of Fun.
Anal. 255 4 (2008), 994-1007.
[3] S. Buckley, Summation condition on Weights, Michigan Math. J., 40(1) (1993) 153-170.
[4] D. Chung, Ph.D Dissertation, University of New Mexico, (2010).
[5] R. Coifman and C. Fefferman, Weighted norm inequalities for maximal functions and singular
integrals Studia Math. 51 (1974), 241-250.
[6] R. Coifman, R. Rochberg, and G. Weiss, Factorization theorems for Hardy spaces in several
variables Ann. of Math. (2) 103 (1976), 611-635.
[7] D. Cruz-Uribe, SFO, J. Martell and C. Pérez, Sharp weighted estimates for classical operators
Preprint.
[8] O. Dragicević, L. Grafakos, M. C. Pereyra, and S. Petermichl, Extrapolation and sharp norm
estimates for classcical operators on weighted Lebesgue spaces Publ. Mat. 49 (2005), 73-91.
[9] O. Dragicević and A. Volberg, Sharp estimate of the Ahlfor-Beurling operator via averaging
Martingale transforms, Michigan Math. J. 51 (2003)
[10] J. B. Garnett, Bounded analytic functions Acad. Press, NY, 1981.
31
[11] H. Helson and G. Szegö, A problem in prediction theory, Ann. Math. Pura. Appl. 51 (1960),
107-138.
[12] S. Hukovic, S. Treil, A. Volberg, The Bellman functions and the sharp weighted inequalities for
square functions Operator Theory : Advances and Applications, the volume in memory of S.
A. Vinogradov, v. 113, Birkhauser Verlag, (2000)
[13] R. Hunt, B. Muckenhoupt, and R. Wheeden, Weighted norm inequalities for conjugate function
and the Hilbert transform Trans. Amer. Math. Soc. 176 (1973), 227-251.
[14] T. Hytönen, M. Lacey, M. Reguera, and A. Vagharshakyan, Weak and Strong-type esti-
mates for Haar Shift Operators: Sharp power on the Ap characteristic (2009), available at
http://arxiv.org/abs/0911.0713
[15] M. Lacey, Haar shifts, Commutators, and Hankel operators CIMPA-UNESCO conference on
Real Analysis and it Applications, La Falda Argentina, (2008).
[16] M. Lacey, S. Petermichl, J. Piper and B. Wick, Iterated Riesz Commutators: A simple proof of
boundedness Proc. of El Escorial (2008).
[17] M. Lacey, S. Petermichl and M. Reguera, Sharp A2 inequality for Haar shift operators Math.
Ann. to appear.
[18] B. Muckenhoupt, Weighted norm inequalities for the Hardy maximal function Trans. Amer.
Math. Soc. 165 (1972), 207-226.
[19] F. Nazarov, S. Treil, and A. Volberg, The Bellman functions and two-weight inequalities for
Haar multipliers J. Amer. Math. Soc. 12 (1999), 909-928.
[20] C. Pérez, Endpoint estimates for commutators of singular integral operators J. Funct. Anal.
128 (1995), 163-185.
[21] C. Pérez, Sharp estimats for commutators of singular integrals via iterations of the Hardy-
Littlewood maximal functions J. Fourier Anal. Appl. 3 (1997), 743-756.
[22] C. Pérez, Personal communication (2009).
[23] C. Pérez and G. Pradolini, Sharp weighted Endpoint Estimates for commutators of singular
integrals Michigan Math. H. 49 (2001) 23-37.
[24] M. C. Pereyra, Haar Multipliers meet Bellman Functions Revista Mat. Iber. 25 3 (2009), 799-
840.
[25] S. Petermichl, Dyadic shifts and a logarithmic estimate for Hankel operators with matrix symbol
C. R. Acad. Sci. Paris 330 (2000), 455-460.
[26] S. Petermichl, The sharp bound for the Hilbert transform on weighted Lebesgue spaces in terms
of the classical Ap Characteristic, Amer. J. of Math. 129 (2007) 1355-1375.
[27] S. Petermichl, The sharp weighted bound for the Riesz transforms Proc. Amer. Math. Soc. 136
(2008), 1237-1249.
[28] S. Petermichl and S. Pott, An estimate for weighted Hilbert transform via square functions
Trans. Amer. Math. Soc. 354 (2002), 1699-1703.
32
[29] S. Petermichl, S. Treil and A. Volberg, Why the Riesz transforms are averages of the dyadic
shifts? Proc. of the 6th Inter. Conf. on Harmonic Analysis and Partial Differential Equations
(El Escorial, 2000) (2002) 209-228.
[30] S. Petermichl and A. Volverg, Heating of the Ahlfors-Beurling operator: Weakly quasiregular
maps on the plane are quasiregular Duke Math J. 112 (2002), 281-305.
[31] A. Vagharshakyan Recovering Singular Integrals from Haar Shifts (2009), available at
http://arxiv.org/abs/0911.4968
[32] J. Wittwer, A shap estimate on the norm of Martingale transform Math. Res. Lett. 7 (2000)
1-12.
Department of Mathematics and Statistics, 1 University of New Mexico, Albu-
querque, NM 87131-0001, E-mail: [email protected] and [email protected]
33
|
1511.04568 | 1 | 1511 | 2015-11-14T15:11:21 | Reducibility of invertible tuples to the principal component in commutative Banach algebras | [
"math.FA"
] | Let $A$ be a complex, commutative unital Banach algebra. We introduce two notions of exponential reducibility of Banach algebra tuples and present an analogue to the Corach-Su\'arez result on the connection between reducibility in $A$ and in $C(M(A))$. Our methods are of an analytical nature. Necessary and sufficient geometric/topological conditions are given for reducibility (respectively reducibility to the principal component of $U_n(A)$) whenever the spectrum of $A$ is homeomorphic to a subset of $\mathbb C^n$. | math.FA | math |
REDUCIBILITY OF INVERTIBLE TUPLES TO THE
PRINCIPAL COMPONENT IN COMMUTATIVE BANACH
ALGEBRAS
RAYMOND MORTINI AND RUDOLF RUPP
Abstract. Let A be a complex, commutative unital Banach algebra.
We introduce two notions of exponential reducibility of Banach algebra
tuples and present an analogue to the Corach-Su´arez result on the con-
nection between reducibility in A and in C(M (A)). Our methods are
of an analytical nature. Necessary and sufficient geometric/topological
conditions are given for reducibility (respectively reducibility to the prin-
cipal component of Un(A)) whenever the spectrum of A is homeomorphic
to a subset of Cn.
1. Introduction
The concepts of stable ranges and reducibility of invertible tuples in rings
originate from Hyman Bass's work [2] treating problems in algebraic K-
theory. Later on, due to work of L. Vasershtein [29], these notions also
turned out to be very important in the theory of function algebras and
topology because of their intimate relations to extension problems. This
direction has further been developed by Corach and Su´arez, [4], [5]. Function
theorists have also been interested in this subject and mainly computed the
stable ranks for various algebras of holomorphic functions. For example,
P.W. Jones, D. Marshall and T. Wolff [9] determined the stable rank of
the disk algebra A(D), and Corach and Su´arez [7] the one for the polydisk
and ball algebras. The whole culminated in S. Treil's work on the stable
rank for the algebra H ∞ of bounded analytic functions on the unit disk
[28]. Recent work includes investigations of stable ranks for real-symmetric
function algebras (see for instance [17] and [15]). The subject of the present
paper is linked to the theory developped by Corach and Su´arez and provides
a detailed analysis of the fine structure of the set Un(A) of invertible tuples
within the realm of commutative Banach algebras. The main intention is
the introduction of a new concept, the exponential reducibility of n-tuples,
and to present a new view on the structure of the connected components of
Un(A).
In contrast to the work of Corach and Su´arez (and Lin), we use an ana-
lytic framework (and methods) instead of the powerful algebraic-topological
setting. We think that this makes the theory accessible to a larger reader-
ship.
1
2
RAYMOND MORTINI AND RUDOLF RUPP
1.1. Notational background and scheme of the paper. Let A be a
commutative unital Banach algebra over K = R or K = C, the identity
element (or multiplicatively neutral element) being denoted by 1. Then the
spectrum (=set of nonzero, multiplicative K-linear functionals on A) of A is
denoted by M (A), and the set of all n×n-matrices over A by Mn(A). If f ∈
C(X, K), the space of all K-valued continuous functions on the topological
space X, then Z(f ) := {x ∈ X : f (x) = 0}. If f = (f1, . . . , fn) ∈ C(X, Kn),
j=1 Z(fj) is the joint zero-set. Moreover, if f ∈ An, then
j=1 fjgj and, when viewed as an
j=1 fj2, hf , gi := f · g := Pn
then Z(f ) := Tn
f = qPn
element in An, e1 := (1, 0, . . . , 0). Finally, for f ∈ C(X, K),
f ∞ = f X = sup{f (x) : x ∈ X}.
Let us begin with the pertinent definitions.
Definition 1.1. • An n-tuple (f1, . . . , fn) ∈ An is said to be invertible (or
unimodular), if there exists (x1, . . . , xn) ∈ An such that the B´ezout equation
j=1 xjfj = 1 is satisfied. The set of all invertible n-tuples is denoted by
Un(A). Note that U1(A) = A−1.
Pn
• An (n + 1)-tuple (f1, . . . , fn, g) ∈ Un+1(A) is called reducible (in A) if
there exists (a1, . . . , an) ∈ An such that (f1 + a1g, . . . , fn + ang) ∈ Un(A).
• The Bass stable rank of A, denoted by bsr A, is the smallest integer n
such that every element in Un+1(A) is reducible. If no such n exists, then
bsr A = ∞.
The following two results due to Corach and Su´arez are the key to the
theory of stable ranks.
Lemma 1.2. ([4, p. 636] and [6, p. 608]). Let A be a commutative, unital
Banach algebra over K.Then, for g ∈ A, the set
Rn(g) := {f ∈ An : (f , g) is reducible}
is open-closed inside the open set
In(g) := {f ∈ An : (f , g) ∈ Un+1(A)}.
In particular, if φ : [0, 1] → In(g) is a continuous map and (φ(0), g) is
reducible, then (φ(1), g) is reducible. Moreover, Rn(g) = gAn + Un(A).
The next assertion, which gives us a relation between reducibility in a
Banach algebra A and the associated uniform algebra C(M (A)) of all con-
tinuous complex-valued functions on the spectrum M (A) of A, actually is
one of the most important theorems in the theory of the Bass stable rank:
Theorem 1.3 (Corach-Su´arez). ([5, p. 4]). Let A be a commutative uni-
tal complex Banach algebra and suppose that (f1, . . . , fn, g) is an invertible
(n + 1)-tuple in A. Then (f1, . . . , fn, g) is reducible in A if and only if
(bf1, . . . , bfn,bg) is reducible in C(M (A)).
REDUCIBILITY TO THE PRINCIPAL COMPONENT
3
Here is now the scheme of the paper. In Section two we have a look at
the principal components of Mn(A) and Un(A) and in Section three we give
a connection between reducibility and the extension of invertible rows to
invertible matrices in the principal component of Mn(A).
In the forth section of our paper we are concerned with the analogues of
the results quoted above for our new notion of "reducibility of (n + 1)-tuples
in A to the principal component of Un(A)" (see below for the definition).
In the fifth section we apply these results and give geometric/topological
conditions under which (n + 1)-tuples in C(X, K) for X ⊆ Kn are reducible,
respectively reducible to the principal component of Un(C(X, K)). Let us
point out that due to Vasershtein's work, the Bass stable rank of C(X, K) is
less than or equal to n + 1; hence every invertible (n + 2)-tuple in C(X, K)
is reducible, but in general, not every tuple having length less than n + 1 is
reducible. In the sixth section we apply our results to the class of Euclidean
Banach algebras. In Section 8 we give a simple proof of a result by V. Ya.
Lin telling us that a left-invertible matrix L over A can be complemented
to an invertible matrix over A if and only if the matrix L of its Gelfand
transforms can be complemented in the algebra C(M (A)).
2. The principal components of Mn(A) and Un(A)
In this section we expose for the reader's convenience several results nec-
essary to develop our theory, and whose proofs we could not locate in the
literature (in particular for the case of real algebras). First, let us recall
that if A = (A, · ) is a commutative unital Banach algebra over K, then
the principal component Exp Mn(A) (=the connected component of the
identity matrix In) of the group of invertible n × n-matrices over A is given
by
Exp Mn(A) = {eM1 · · · eMk : Mj ∈ Mn(A)}.
(see [19, p. 201])
Our description of the connected components of the set Un(A), viewed as
a topological subspace of An, is based on the following classical result giving
a relation between two invertible tuples that are close to each other. An
elementary proof of that result (excepted the addendum) is given in [26].
Theorem 2.1. Let A = (A, · ) be a commutative unital Banach algebra
over K. Suppose that f = (f1, . . . , fn) is an invertible n-tuple in A. Then
there exists ε > 0 such that the following is true:
j=1 gj − fj < ε there
is a matrix H ∈ Mn(A) such that
For each g = (g1, . . . , gn) ∈ An satisfying Pn
.
= (exp H)
g1
...
gn
f1
...
fn
4
RAYMOND MORTINI AND RUDOLF RUPP
In particular, g itself is an invertible n-tuple. Moreover, if u · f t = 1, then
ue−H · gt = 1.
Addendum: if fn = gn, then H can be chosen so that its last row is the zero
vector and
eH =(cid:18) eM ∗
0n−1 1(cid:19)
for some matrix M ∈ Mn−1(A).
Theorem 2.2. Let A be a commutative unital Banach algebra over K. If
f = (f1, . . . , fn) ∈ Un(A), then the connected component, C(f ), of
f in
Un(A) equals the set
In particular, C(f ), is path-connected.
f · Exp Mn(A).
Proof. Let C = f · Exp Mn(A); that is
C =
f ·(cid:16) kYj=1
exp Mj(cid:17) : Mj ∈ Mn(A), k ∈ N
.
We first note that C is path-connected. To see this, let
g = f · exp(M1) · · · exp(Mk),
for some Mj ∈ Mn(A). Then the map φ : [0, 1] → Un(A), given by
φ(t) := f · exp(tM1) · · · exp(tMk),
is a continuous path joining f to g within Un(A).
We claim that C is open and closed in Un(A). In fact, let g ∈ C. According
to Theorem 2.1, there is ε > 0 so that for every h ∈ An with
nXj=1
gj − hj < ε,
there is matrix M ∈ Mn(A) such that ht = (exp M ) gt. That is h =
g · exp M t. Therefore h ∈ C. Hence g is an interior point of C. Thus C is
open in An. Since Un(A) itself is open in An, we conclude that C is open in
Un(A).
Theorem 2.1 again, there is ε > 0 so that for every h ∈ An withPn
To show that C is (relatively) closed in Un(A), we take a sequence (gk) in
C that converges (in the product topology of An) to g ∈ Un(A). Applying
j=1 gj −
hj < ε, there is matrix M ∈ Mn(A), depending on h, such that h =
g · exp M t. This holds in particular for h = gk, whenever k is large. Thus
g = gk · exp Mk for some Mk ∈ Mn(A), from which we conclude that g ∈ C.
Hence C is closed in Un(A).
Being open-closed and connected now implies that C is the maximal con-
nected set containing itself. Hence, with f ∈ C, we deduce that C(f ) =
C.
(cid:3)
REDUCIBILITY TO THE PRINCIPAL COMPONENT
5
We are now able to define the main object of this paper:
Definition 2.3. Let A be a commutative unital Banach algebra over K.
Then the principal component of Un(A) is the connected component of e1
in Un(A) and is given by the set
P(Un(A)) := e1 · Exp Mn(A).
Remark. • If n = 1, then U1(A) = A−1 and
P(Un(A)) = exp A := {ea : a ∈ A}.
• Let us note that in the representation e1 · Exp Mn(A) of the principal
component of Un(A) any other "canonical" element ej, with
ej := (0, . . . , 0, 1{z}j
is admissible, too. In fact, for i 6= j,
, 0, . . . , 0),
is a path in Un(A) joining ei with ej, because
γi,j(t) := (1 − t)ei + tej
D(1 − t)ei + t ej , ei + ejE = 1.
Hence ei and ej belong to the same connected component of Un(A).
3. Extension of invertible rows to the principal component
An interesting connection between reducibility and extension of rows to
invertible matrices (resp. to matrices in the principal component) is given in
the following theorem (see for example [21, p. 311] and [20, p. 1129]). The
additional property of being extendable to finite products of exponential
matrices (hence to the principal component of Mn(A)), seems not to have
been considered in the literature before (as far as we know).
Theorem 3.1. Let A be a commutative unital Banach algebra over K with
unit element 1. Suppose that u := (f1, . . . , fn, g) ∈ Un+1(A) is reducible.
Then there is an invertible matrix W ∈ Mn+1(A) with determinant 1 and
which is a finite product of exponential matrices such that u W = e1. In
other words, u ∈ P(Un(A)). Moreover, if M = W −1, then u is the first row
of M .
Proof. Let f = (f1, . . . , fn). Since u = (f , g) is reducible, there exists x =
(x1, . . . , xn) ∈ An with f + g x ∈ Un(A). Hence there is y = (y1, . . . , yn) ∈
An such that 1
y · (f + gx)t = 1 − g.
1 Note that we do want the element 1 − g here on the right-hand side.
6
RAYMOND MORTINI AND RUDOLF RUPP
Consider the matrices
1
1
0
...
· · ·
0
x1 x2
1
. . .
· · ·
1
. . . xn
1
y1
y2
...
yn
y · xt + 1
. . .
−(f1 + gx1)
· · · −(fn + gxn)
1
W1 =
W2 =
Since
W1 =
1
0
· · ·
1
0
...
· · ·
0
x1 x2
. . .
· · ·
. . .
· · ·
0
...
...
0
1
xn 1
·
1
0
...
0
. . .
· · ·
. . .
0 · · ·
· · ·
y1
y2
...
1 yn
0
1
=: M1M2,
it is easy to see that W1 and W2 are invertible matrices in Mn+1(A) with
determinant 1 satisfying 2
(f1, . . . , fn, g) W1W2 = (0, . . . , 0, 1) ∈ An+1.
Let W3 = h et
(−1)net
1
(when identifying R · 1 with R),
n+1
n i
et
2
. . . et
∈ Mn+1(A); that is
W3 =
0 (−1)n
0
...
...
0
1
0
...
...
0
0
0
1
0
0
. . .
. . .
. . .
. . .
. . .
. . .
. . .
0
0
1
0
.
Note that en+1W3 = e1 and det W3 = 1. If we put W = W1W2W3, then
W is invertible in Mn+1(A), det W = 1, and uW = e1, where e1 ∈ An+1.
Write W −1 =(cid:20) w
V (cid:21), where w is the first row. It is easy to see that u = w.
Thus
(cid:20) u
V (cid:21) W = In+1.
Hence the row u has been extended by V to an invertible matrix M := W −1.
2 The matrix multiplication here is preferably done from the left to the right: first
multiply the one-row matrix with W1, then go on.
REDUCIBILITY TO THE PRINCIPAL COMPONENT
7
Note that M1 and M2 in the decomposition W1 = M1M2, as well as W2,
have the form In+1 + N , where N is a nilpotent matrix. Hence, In+1 + N =
eB for some B ∈ Mn+1(A) (just use an appropriate finite section of
the power series expansion of the real logarithm log(1 + x)). Moreover,
W3 ∈ Mn+1(R) and det W3 > 0. Thus, W3 is a product of exponential
matrices over R 3. Consequently, W = W1W2W3 is a finite product of
exponential matrices over A.
(cid:3)
4. Reducibility to the principal component
Definition 4.1. Let A be a commutative unital Banach algebra over K = R
or K = C. An invertible pair (f, g) ∈ U2(A) is said to be reducible to the
principal component exp A of A−1 if there exists u, v ∈ A such that
f + ug = ev.
It is clear that if A−1 = U1(A) is connected, then the notions of "re-
ducibility of pairs" and "reducibility of pairs to the principal component"
coincide. Our favourite example is the disk algebra A(D). If U1(A) is dis-
connected, as it is the case for the algebra C(T, C) for example, then for
every f ∈ A−1 \ exp A, the pair (f, 0) is reducible, but not reducible to the
principal component of A−1. This notion seems to have appeared for the
first time in Laroco's work [10] in connection with the stable rank of H ∞.
Criteria for various function algebras have been established by the second
author of this note in [22, 23, 24, 25].
Now we generalize this notion to tuples, a fact that never before has been
considered. We propose two different settings. Here is the first one (the
second one will be dealt with in Section 7).
Definition 4.2. Let A be a commutative unital Banach algebra over K.
An invertible (n + 1)-tuple (f , g) ∈ Un+1(A) is said to be reducible to the
principal component of Un(A) if there exists h ∈ An such that
f + g h ∈ P(Un(A)).
The following Proposition is pretty clear in the case of complex Banach
algebras, since every permutation matrix P ∈ Mn(C) has a complex loga-
rithm in Mn(C) (see [16]). So what does matter here, is that we consider
real algebras, too.
Proposition 4.3. Let A be a commutative unital Banach algebra over K
and let (f , g) ∈ Un+1(A) be an invertible (n + 1)-tuple in A which is re-
ducible to the principal component P(Un(A)) of Un(A). Suppose that f is
a permutation of f . Then also the tuple ( f , g) is reducible to the principal
component of Un(A).
3Actually, two exponentials will suffice; see [16].
8
RAYMOND MORTINI AND RUDOLF RUPP
Proof. Without loss of generality, n ≥ 2.
Case 1 Let f = (f1, . . . , fn), f = (fn, f2, . . . , fn−1, f1) and
0 · · ·
1
. . .
S =
1
.
1
· · ·
0
1
0
0
Note that S = S−1 and det S = −1. The action of S in A 7→ AS is to
interchange the first and last column. Let W ∈ Mn(R) be given by
(−1)n−1
W =
· · ·
0
1
1
0
0
. . .
. . .
1
.
Then det W = 1 and en = e1W . In particular W ∈ Exp Mn(R). Now,
by assumption, f + g x = e1M for some M ∈ Exp Mn(A). Hence with
x := x S,
f + g x = (f + g x)S = e1 M S
= (enS)(M S) = en(SM S)
= (e1W ) (SM S) = e1(W SM S)
∈ e1 · Mn(A) = P(Un(A)),
where we have used that M ∈ Exp Mn(A) if and only if S−1M S ∈
Exp Mn(A) for every invertible matrix S; just observe that
S−1 (cid:0) kYj=1
eMj(cid:1)S =
kYj=1
(S−1eMj S) =
kYj=1
eS−1Mj S.
Case 2 Let f be an arbitrary permutation of f . Hence f = f P for some
permutation matrix P . If det P > 0 then, P = eP1eP2 for some matrices
Mj ∈ Mn(R) (see for example [16]).
If det P < 0, then we aditionally
interchange via S the first coordinate with the last one in f . Let us call this
new n-tuple F . Then F = f Q for some permutation matrix Q with det Q >
0, and again Q = eQ1eQ2 for some Qj ∈ Mn(R). Now, by assumption, there
exists x ∈ An such that
f + g x = e1 eM1 . . . eMk
for some matrices Mj ∈ Mn(A). Hence, by multiplying at the right with Q,
f Q + g xQ = e1 eM1 . . . eMk eQ1eQ2.
Thus (F , g) is reducible to the principal component of Un(A). The first case
now implies that the same holds for ( f , g).
(cid:3)
REDUCIBILITY TO THE PRINCIPAL COMPONENT
9
A sufficient condition for reducibility to the principal component is given
in the following technical result:
Lemma 4.4. Let A be a commutative unital Banach algebra over K. Sup-
pose that (f , g) ∈ Un+1(A) and n ≥ 2. Then (f , g) is reducible to the prin-
cipal component of Un(A) if there exist two vectors x ∈ An and v ∈ Un(A)
such that v is reducible itself with respect to some of its coordinates 4 and
(f + x g) · vt = 1.
Proof. Suppose that i0 6= n. Then we interchange the i0-th coordinate with
the n-th coordinate in the three vectors f , x and v appearing here. The
new vectors f , x and v still satisfy the B´ezout equation
(f + x g) · vt = 1.
Since (f , g) is reducible to the principal component of Un(A) if and only if
(f , g) does (Proposition 4.3), we may assume, right at the beginning, that
v is reducible with respect to its last coordinate.
By Theorem 3.1, the reducibility of the row vector v implies the existence
of a finite product P1 of exponential matrices over A such that vt is the first
column of a matrix P1 ∈ Exp Mn(A). Hence (as matricial products)
(f + x g) P1 = (f + x g) (vt ∗ ∗ ∗) = (1, x2, . . . , xn)
for some xj ∈ A. If we let
P2 =
1 −x2
. . . −xn
1
. . .
1
,
then P2 ∈ Exp Mn(A) (because it has the form In + N , where N is nilpo-
tent), and
(f + x g) P1P2 = (1, x2, . . . , xn)P2 = e1.
Hence f + x g ∈ e1 · Exp Mn(A) = P(Un(A)).
(cid:3)
In order to study the reducibility to the principal component, we introduce
a certain equivalence relation on the set of n-tuples, reminiscent of that in [4].
Corach and Su´arez considered diagonal matrices M all of whose diagonal
g ⇐⇒ f − g M ∈ a An for
entries were invertible elements in A: f
such a matrix M . The equivalence classes of that relation, though, do not
seem to be compatible with the connected components of In(a); openness
for example fails.
CS∼
a
4 This means that there exists i0 and aj ∈ A, (j = 1, . . . , n − 1), such that for v =
(v1, . . . , vi0 , . . . , vn), the vector (v1 + a1vi0 , . . . , vi0−1 + ai0−1vi0 , vi0 +1 + ai0+1vi0 , . . . , vn +
anvi0 ) belongs to Un−1(A).
10
RAYMOND MORTINI AND RUDOLF RUPP
Theorem 4.5. Let A be a commutative unital Banach algebra over K, a ∈
A, and consider the open set
In(a) := {f ∈ An : (f , a) ∈ Un+1(A)}.
Given f , g ∈ An, define the relation
f
exp
∼
a
Then
(1)
(2)
g ⇐⇒ ∃ x ∈ An, ∃B1, . . . , Bk ∈ Mn(A) : f + a x = g eB1 . . . eBk .
is an equivalence relation on An.
exp
∼
a
If f ∈ In(a), then [f ] ⊆ In(a), where
exp
∼
a
[f ] := {h ∈ An : h
f }
(3)
is the equivalence class associated with f .
If f ∈ In(a), then
(3i)
(3ii)
(3iii)
[f ] is open in An,
[f ] is a closed-open subset of In(a),
[f ] is a (path)-connected set within In(a).
(4) The connected components of In(a) are the equivalence classes [f ],
where f ∈ In(A).
Proof. (1) •
exp
∼
a
is reflexive: just take x = 0 and Bj = O.
is symmetric: if f + a x = g eB1 . . . eBk , then
•
exp
∼
a
g − a(cid:0)x e−Bk . . . e−B1(cid:1) = f e−Bk . . . e−B1.
•
exp
∼
a
is transitive (here we use that in the definition of the relation
exp
∼
a
products of exponential matrices appear; a single exponential matrix would
f 3, then there exist xj ∈ An and
not be sufficient): let f 1
Ej ∈ Exp Mn(A) such that
f 2 and f 2
exp
∼
a
exp
∼
a
f 1 + a x1 = f 2 E1 = (f 3 E2 − a x2)E1.
f 1 + a(x1 + x2E1) = f 3 E2E1.
f 3.
Then
Hence f 1
exp
∼
a
(2) Let f ∈ In(a). Then there is x ∈ An such that f + a x ∈ Un(A). Now
if f ∈ [f ] then,
for some E ∈ Exp Mn(A). Hence
f + a x = f E
f + a (x + xE) = (f + a x)E ∈ Un(A),
from which we conclude that ( f , a) ∈ Un+1(A). In other words, f ∈ In(a).
Thus [f ] ⊆ In(a).
REDUCIBILITY TO THE PRINCIPAL COMPONENT
11
(3i) To show the openness of [f ] whenever f ∈ In(a), let h ∈ [f ]. By (2),
(h, a) ∈ Un+1(A). We claim that there is ε > 0 such that every h′ ∈ An with
h′ − h < ε is equivalent to f . To see this, choose according to Theorem
2.1, ε > 0 so small that (h′, a) = (h, a) eH for some H ∈ Mn+1(A).
In
particular h′ ∈ In(a). By that same Theorem, H may be chosen so that the
last column of H is zero and that
for some K ∈ Mn(A) and x ∈ An. Since
x
eH =(cid:18)eK 0t
1(cid:19)
(h′, a) = (h, a) (cid:18)eK 0t
1(cid:19) ,
x
n
n
we conclude that
h′ = h eK + a x.
In other words, h′ ∈ [h] = [f ]. Hence [f ] is open in An.
(3ii) Let (hj) be a sequence in [f ] ⊆ In(a) converging to some h′ ∈ In(a).
As in the previous paragraph, if n is sufficently large, we may conclude
that h′ ∈ [hj] = [f ] for j ≥ j0. Hence [f ] is (relatively) closed in In(a).
Furthermore, since [f ] ⊆ In(a), we deduce from (3i) that [f ] is also open in
In(a).
(3iii) Let f ∈ [f ]; say f + a x = f eM1 · · · eMk for some x ∈ An and
Mj ∈ Mn(A). Then the map H : [0, 1] → An given by
H(t) = f etM1 · · · etMk − ta x
is a continuous path joining f with f . By definition of
equivalent to f ; that is H(t) ∈ [f ]. Thus [f ] is path connected.
exp
∼
a
, each H(t) is
(4) This follows immediately from (3i)-(3iii).
(cid:3)
Here is the counterpart to Lemma 1.2.
Theorem 4.6. Let A be a commutative unital Banach algebra over K.
Then, for g ∈ A, the set
Rexp
:= {f ∈ An : (f , g) is reducible to the principal component of Un(A)}
= g An + P(Un(A))
n (g)
is open-closed inside In(g). In particular, if F : [0, 1] → In(g) is a contin-
uous map for which (F (0), g) is reducible to the principal component, then
(F (1), g) is reducible to the principal component, too.
Proof. We first note that, by definition, Rexp
n (g) ⊆ In(g) and that the re-
ducibility of (f , g) ∈ Un+1(A) to the principal component of Un(A) is equiv-
alent to the assertion that f
e1. Thus
exp
∼
g
(4.1)
f ∈ Rexp
n (g) ⇐⇒ f ∈ [e1] ⇐⇒ [f ] = [e1].
12
RAYMOND MORTINI AND RUDOLF RUPP
In other words, Rexp
4.5.
n (g) = [e1]. The assertion then follows from Theorem
Now if F : [0, 1] → In(g) is a curve in In(g), then C := F ([0, 1]) is
(cid:3)
n (g), we deduce that C ⊆ Rexp
connected. Since F (0) ∈ Rexp
n (g).
The following corollaries are immediate (the second one is originally due
to Corach and Su´arez [4]).
Corollary 4.7. Let A be a commutative unital Banach algebra over K and
g ∈ A. Then the following assertions are equivalent:
(1) Every invertible (n + 1)-tuple (f , g) ∈ Un+1(A) is reducible to the
principal component of Un(A); that is In(g) = Rexp
n (g).
(2) In(g) is connected.
Proof. Just note that by equation (4.1), Rexp
n (g) = [e1] and that [e1] is a
connected set which is contained in In(g) for every g ∈ A. The result now
follows from Theorem 4.5.
(cid:3)
Corollary 4.8. Let A be a commutative unital Banach algebra over K and
g ∈ A. Then the following assertions are equivalent:
(1) In(g) = Rn(g);
(2) Each component of In(g) meets Un(A).
Proof. Since the connected components of In(g) are the equivalence classes
with f ∈ In(g) (Theorem 4.5), we have the following equivalent
[f ] for
exp
∼
g
assertions for a given f ∈ In(g):
(i)
(ii)
(iii)
[f ] ∩ Un(A) 6= ∅,
there exists u ∈ Un(A) such that u
there exists u ∈ Un(A), x ∈ An, and B1, . . . , Bk ∈ Mn(A) such
that
exp
∼
g
f ,
(iv)
(f , g) is reducible.
f + g x = u eB1 · · · eBk ,
Note that we actually proved a stronger result than stated, because the
assertions (i)-(iv) are valid for each individual f .
The following two Lemmas are very useful to check examples upon re-
ducibility. They roughly say that the reducibility of (f , g) depends only on
the behaviour of the Gelfand transforms of the coordinates fj of f on the
(cid:3)
Lemma 4.9. Let A be a commutative unital Banach algebra over C. Sup-
zero set of bg. We use the following notation: bf := (bf1, . . . , bfn).
pose that (f , g) ∈ Un+1(A). Let E := Z(bg). If for some u ∈ Un(A) and
matrices Mj ∈ Mn(A)
sup
x∈E(cid:12)(cid:12)bf (x) −bu(x) expcM1(x) · · · expcMm(x)(cid:12)(cid:12) < ε,
REDUCIBILITY TO THE PRINCIPAL COMPONENT
13
where ε is sufficiently small, then (f , g) is reducible in A.
Proof. Let δ := min{bf (x) : x ∈ E}. Since (f , g) ∈ Un(A), we have δ > 0.
Fix ε ∈ ]0, δ/2], and let b := u exp M1 · · · exp Mm ∈ Un(A) be chosen so
that
Consider the path ψ : [0, 1] → An given by
sup
x∈E
bf (x) −bb(x) < ε.
ψ(t) = (1 − t) f + t b.
On E we then have the following estimates:
(cid:12)(cid:12)(1 − t)bf + tbb(cid:12)(cid:12) = (cid:12)(cid:12)t(bb −bf ) +bf(cid:12)(cid:12)
≥ bf − t bb −bf
≥ δ − δ/2 = δ/2.
Hence the tuples (ψ(t), g) are invertible in A for every t. Since for t = 1,
ψ(1) = b ∈ Un(A), the tuple (ψ(1), g) is reducible in A. By Lemma 1.2,
(ψ(0), g) then is reducible which in turn implies the reducibility of (f , g). (cid:3)
Lemma 4.10. Let A be a commutative unital Banach algebra over C.
If for some matrices
Suppose that (f , g) ∈ Un+1(A). Let E := Z(bg).
Mj ∈ Mn(A)
sup
x∈E(cid:12)(cid:12)bf (x) − e1 · expcM1(x) · · · expcMm(x)(cid:12)(cid:12) < ε,
where ε is sufficiently small, then (f , g) is reducible to the principal compo-
nent P(Un(A)) of Un(A).
Proof. Consider the path ψ : [0, 1] → An given by
ψ(t) = (1 − t) f + t b,
where b := e1 · exp M1 · · · exp Mm. If
then (ψ(t), g) ∈ Un+1(A) for every t ∈ [0, 1]. Now (ψ(1), g) = (b, g) is
reducible to the principal component of Un(A) since
0 < ε < (1/2) min{bf (x) : x ∈ Z(bg)},
b + 0 · g = e1 · exp M1 · · · exp Mm ∈ P(Un(A)).
Hence, by Theorem 4.6, (ψ(0), g) = (f , g) is reducible to the principal com-
ponent of Un(A), too.
(cid:3)
We close this section with our main theorem, which is the analogue to
the Corach-Su´arez result Theorem 1.3 ([5]).
It is based on the Arens-
Novodvorski-Taylor theorem ([1],[18], [27]), a version of which we recall here.
Theorem 4.11 (Arens-Novodvorski-Taylor). Let A be a commutative unital
complex Banach algebra, X := M (A) its spectrum and Mn(A) the Banach
algebra of n × n matrices over A.
14
RAYMOND MORTINI AND RUDOLF RUPP
(1i) Suppose that for some M ∈ Mn(A) there is M ∈ Mn(C(X)) such
5 Then M = exp L1 · · · exp Lm for some Lj ∈
that cM = exp M .
Mn(A).
(1ii) Let M ∈ Mn(A)−1. If cM belongs to the principal component of
Mn(C(X))−1, then M already belongs to the principal component of
Mn(A)−1.
(2) Let f ∈ Un(C(X)). Then there exist g ∈ Un(A) and G1, . . . , Gm ∈
(3) Let u and v be in Un(A). Suppose that there are matrices Gj ∈
Mn(C(X)) such that f = bg exp G1 · · · exp Gm.
Mn(C(X)) such that bu = bv exp G1 · · · exp Gm. Then u and v be-
(4) The Gelfand transform induces a group isomorphism between the
long to the same connected component of Un(A).
quotient groups
Mn(A)−1/Exp Mn(A) and Mn(C(X))−1/Exp Mn(C(X)).
Item (2), in particular, says that every connected component of Un(C(X))
path in Un(C(X)), then u and v can be joined by a path in Un(A). Item (4)
also says that every element in Mn(C(X))−1 is homotopic in Mn(C(X))−1
contains an element of the form bf := (bf1, . . . , bfn), where f ∈ Un(A). More-
over, (3) is equivalent to the assertion that if bu and bv can be joined by a
to cM for some M ∈ Mn(A)−1.
Theorem 4.12. Let A be a commutative unital Banach algebra over C.
Given an invertible tuple (f , g) ∈ Un+1(A), the following assertions are
equivalent:
(1)
(f , g) is reducible to the principal component of Un(A);
(3)
(2) bf Z(bg) belongs to the principal component of Un(C(Z(bg))).
(bf ,bg) is reducible to the principal component of Un(C(M (A))).
Proof. (1) =⇒ (2) Let E := Z(bg). By assumption there is h ∈ Un(A) such
that
u := f + g h ∈ P(Un(A)).
That is, there are matrices Mj ∈ Mn(A) such that
u = e1 · exp M1 · · · exp Mm.
If we apply the Gelfand transform and restrict to Z(bg), then
Hence bf E ∈ P(Un(C(E))).
bf E = e1 ·(cid:0) expcM1 · · · expcMm(cid:1)E.
(2) =⇒ (1) By assumption, there exist matrices Cj ∈ Mn(C(E)) such
that
5 Here cM is the matrix whose entries are the Gelfand transforms of the entries of the
bf E = e1 · exp C1 · · · exp Ck.
matrix M .
REDUCIBILITY TO THE PRINCIPAL COMPONENT
15
·∞
belongs to the principal component P(Un(C(E))) of Un(C(E)), which is
"generated" by e1, we conclude from the Arens-Novodvorski-Taylor The-
be the uniform closure of the restriction algebra bAE in
Let AE = bAE
C(E). Since Z(g) is A-convex, M (AE) = E ([8]). Because bf E ∈ (AE)n
orem 4.11(3) that bf E belongs to the same component of Un(AE) as e1;
namely the principal component P(Un(AE)) of Un(AE). Hence there are
matrices Bj ∈ Mn(AE) such that
bf E = e1 · exp B1 · · · exp Bm.
Now, we uniformly approximate on E the matrices Bj by matrices cMj with
Mj ∈ Mn(A); say
sup
x∈E(cid:12)(cid:12)(cid:12)bf (x) − e1 · expcM1(x) · · · expcMm(x)(cid:12)(cid:12)(cid:12) < ε
By Lemma 4.10, (f , g) is reducible to the principal component P(Un(A)) of
Un(A).
(2) =⇒ (3) follows from Lemma 4.10 and (3) =⇒ (2) is clear.
(cid:3)
If n = 2, then the previous result reads as follows:
Corollary 4.13. Let A be a commutative unital Banach algebra over C.
Given an invertible pair (f, g) ∈ U2(A), the following assertions are equiva-
lent:
(1) There exist a, h ∈ A such that f + ag = eh;
(2) bf Z(bg) = ev for some v ∈ C(Z(bg)).
5. Reducibility in C(X, K) with X ⊆ Kn
In this section we study the reducibility in C(X, K) with X ⊆ Kn in
detail.
Definition 5.1.
a) Let K ⊆ Rn be compact. A bounded connected
component of Rn \ K is called a hole of K.
b) Let K, L be two compact sets in Rn with K ⊆ L. The pair (K, L)
is said to satisfy the hole condition if every hole of K contains a hole
of L.
The following concepts were introduced in C by the second author in [22].
Definition 5.2. Let K ⊆ Rn be compact and g ∈ C(K, K). Then g is said
to satisfy the boundary principle if for every nonvoid open set G in Rn with
G ⊆ K the following condition holds:
(B1) If g ≡ 0 on ∂G, then g ≡ 0 on G.
Proposition 5.3. Condition (B1) is equivalent to the following assertion:
(B2) If G is an open set in Rn such that G ⊆ K \ Z(g), then there exists
x0 ∈ ∂G such that g(x0) 6= 0.
16
RAYMOND MORTINI AND RUDOLF RUPP
Proof. Assume that g satisfies (B1) and let G ⊆ K \ Z(g) be open. Then
g cannot vanish identically on ∂G since otherwise (B1) would imply that
g ≡ 0 on G. A contradiction to the assumption that G ∩ Z(g) = ∅. Hence g
satisfies (B2).
Conversely, let g satisfy (B2). Suppose, to the contrary, that g does not
satisfy condition (B1). Then there is an open set G in Rn with G ⊆ K,
g ≡ 0 on ∂G, but such that g does not vanish identically on G. Hence
U := {x ∈ G : g(x) 6= 0}
is an open, nonvoid set in Rn which is contained in K \Z(g). But ∂U ⊆ Z(g),
because
∂U = U \ U ⊆ G \ U ⊆ (G \ U ) ∪ ∂G ⊆ Z(g).
This contradicts condition (B2) for U . Hence such a set G cannot exist and
we deduce that g has property (B1).
(cid:3)
The following result gives an interesting connection between the hole con-
dition (Definition 5.1) and the boundary principle (Definition 5.2).
Theorem 5.4. Let K ⊆ Rn be compact and g ∈ C(K, K). The following
assertions are equivalent:
(1) g satisfies the boundary principle (B1).
(2)
(Z(g), K) satisfies the hole condition; that is, every hole of Z(g)
contains a hole of K 6.
Proof. (1) =⇒ (2) Suppose that there is a component Ω of Rn \ Z(g) with
Ω ⊆ K. Then Ω is open in Rn and ∂Ω ⊆ Z(g). Condition (B1) now implies
that g ≡ 0 on Ω (note that by assumption Ω ⊆ K). Hence Ω ⊆ Z(g). A
contradiction.
(2) =⇒ (1) We show that the equivalent condition (B2) is satisfied. So let
G be open in Rn and assume that G ⊆ K \ Z(g). Let Ω be a component of
G. Then Ω is bounded and open in Rn. In view of achieving a contradiction,
suppose that g ≡ 0 on ∂Ω. Then ∂Ω ⊆ Z(g). Since Ω ∩ Z(g) = ∅, we deduce
that Ω belongs to a connected component C of Rn \ Z(g). If Ω is a proper
subset of C, then a path connecting z0 ∈ Ω and w ∈ C \ Ω within C would
pass through a boundary point z1 of Ω. But then g(z1) = 0, contradicting
z1 ∈ C. Hence Ω = C. Since Ω is bounded, we conclude that Ω is a hole
of Z(g). But Ω ⊆ G ⊆ K; thus (Z(g), K) cannot satisfy the hole condition
(2). This is a contradiction. Hence there is x0 ∈ ∂Ω such that g(x0) 6= 0.
Since ∂Ω ⊆ ∂G (note that Ω is supposed to be a component of G), we are
done.
(cid:3)
Let us emphasize that for Z(g) ⊆ K ⊆ Rn the pair (Z(g), K) automat-
ically satisfies the hole condition (and equivalently the boundary principle
(B1)) if K ◦ = ∅. An important class of zero-sets satisfying the equivalent
conditions (1) and (2) is given in the following example:
6 or which is the same, no hole of Z(g) is entirely contained in K.
REDUCIBILITY TO THE PRINCIPAL COMPONENT
17
Example 5.5. Let K ⊆ Rn be compact and g ∈ C(K, K). Then Z(g) has
property (B1) if:
(i) Rn \ Z(g) is connected whenever n ≥ 2;
(ii) R \ Z(g) has exactly two components whenever n = 1.
The next two theorems give nice geometric/topological conditions for re-
ducibility of n-tuples, respectively for reducibility to the principal compo-
nent of Un(C(X, K)). They generalize the corresponding facts for pairs
developed by the second author in [22] for certain planar compacta.
Theorem 5.6. Let K ⊆ Kn be compact and g ∈ C(K, K). The following
assertions are equivalent:
(1) C(K, K) is n-stable at g, that is (f , g) is reducible for every f =
(f1, . . . , fn) ∈ C(K, Kn) such that (f , g) ∈ Un+1(C(K, K)).
(2) f Z(g) admits a zero-free extension to K for all f ∈ C(K, Kn) with
Z(f ) ∩ Z(g) = ∅.
(3) g satisfies the boundary principle (B1).
(4)
(Z(g), K) satisfies the hole condition.
Proof. i) The equivalence of (1) with (2) is well-known (see [5] or [26]).
The equivalence of (2) with (4) is [14, Theorem 5.6], provided we identify
(u1 + iv1, . . . un + ivn) in the complex-valued case with the real-valued (2n)-
tuple (u1, v1, . . . , un, vn). The equivalence of (3) with (4) is Theorem 5.4. (cid:3)
Theorem 5.7. Let K ⊆ Kn be compact and g ∈ C(K, K). The following
three assertions are equivalent:
(1)
(2)
(f , g) is reducible to the principal component of Un(C(K, K)) for
every f ∈ C(K, Kn) such that (f , g) ∈ Un+1(C(K, K));
there exist matrices Bj ∈ Mn(C(Z(g), K)) such that
f Z(g) = e1 · eB1 · · · eBk
for every f ∈ C(K, Kn) with Z(f ) ∩ Z(g) 6= ∅;
(3) Z(g) has no holes in Kn.
Proof. First we note that (1) and (2) are equivalent in view of Theorem 4.12.
Since we have to deal here only with the special case C(K, K), the following
simple proof is available:
(1) =⇒ (2) If f + g h ∈ P(Un(C(K, K))) for some h ∈ C(K, Kn) then,
by using the representation of the principal component,
f + g h = e1 · eB1 · · · eBk
for some matrices Bj ∈ Mn(C(K, K)). Restricting this identity to Z(g)
yields the assertion (2).
(2) =⇒ (1) This follows from Lemma 4.10.
(2) =⇒ (3) Suppose, to the contrary, that Z(g) admits a bounded com-
plementary component G ⊆ Kn. Then ∂G ⊆ Z(g). Let a ∈ G and f (z) =
18
RAYMOND MORTINI AND RUDOLF RUPP
z − a, z ∈ Z(g). Since (f , g) ∈ Un+1(C(Z(g), K)), there exist by hypothesis
(2) matrices Bj ∈ Mn(C(Z(g), K)) such that
f Z(g) = e1 · eB1 · · · eBk .
Extending via Tietze's result the matrices Bj continuously to Kn, would
yield a zero-free extension of the n-tuple (z − a)Z(g) to G. This contradicts
a corollary to Brouwer's fixed point theorem (see [3, Chapter 4]).
(3) =⇒ (2) By a standard result in vector analysis (see for example [14,
Corollary 5.8]), the connectedness of Kn \ Z(g) implies that the invertible
tuple f Z(g) ∈ Un(C(Z(g), K)) admits a zero-free extension F to Kn. Let
B ⊆ Kn be a closed ball whose interior contains X. Note that B is a
contractible Hausdorff space. Hence, the set Un(C(B, K)) is connected.
eB1 · · · eBk for some matrices Bj ∈ Mn(C(B, K)).
Thus, F B = e1
Restricting to Z(g) yields the assertion (2).
(cid:3)
·
6. Reducibility in Euclidean Banach algebras
Let us call a complex commutative unital Banach algebra A a Euclidean
Banach algebra if the spectrum M (A) of A is homeomorphic to a compact
set in Cn. This class of algebras includes every finitely generated Banach
algebra over C, for example the algebras
and certain algebras 7 of type
P (K) = C[z1, . . . , zn](cid:12)(cid:12)K
·K
A(K) = {f ∈ C(K, C) : f holomorphic in K ◦},
K ⊆ Cn compact (for example K = D
or K = Bn, the closed unit ball in
Cn). Using the Arens-Novodvorski-Taylor theorem we can now generalize
Theorems 5.6 and 5.7 to Euclidean Banach algebras.
n
Theorem 6.1. Let A be a Euclidean Banach algebra with spectrum X ⊆
Cn and let g ∈ A. Then the assertions (1)-(3), respectively (4)-(5), are
equivalent:
(f , g) is reducible for every n-tuple f ∈ An with (f , g) ∈ Un+1(A).
(1)
(2)
(4)
(Z(bg), X) satisfies the hole condition.
(f , g) is reducible to the principal component of Un(A) for every
n-tuple f ∈ An with (f , g) ∈ Un+1(A).
(3) Z(bg) satisfies the boundary principle in Cn.
(5) Z(bg) has no holes in Cn.
need to show that for every f ∈ C(X, Cn) with Z(f )∩ Z(bg) = ∅, the (n + 1)-
tuple (f ,bg) is reducible in C(X, C) (resp. reducible to P(Un(C(X, C))) ).
(1) =⇒ (2) (resp. (4) =⇒ (5) ) By Theorem 5.6 (resp. Theorem 5.7) we
Proof. First we note that by Theorem 5.4, (2) and (3) are equivalent.
7 We do not know whether every algebra A(K) is finitely generated.
REDUCIBILITY TO THE PRINCIPAL COMPONENT
19
·E
Let E := Z(bg). Consider the algebra B := bAE
closure of the restriction algebra bAE to E. Since E is the zero set of the
Gelfand transform of a function in A, E is A-convex and so the spectrum,
M (B), of B coincides with E (see [8]). Note that f E ∈ Un(C(E, C)). By
the Arens-Novodvorski-Taylor Theorem 4.11 (2), there exist h ∈ Un(B) and
G1, . . . , Gm ∈ Mn(C(E, C)) such that
, that is the uniform
f E = h exp G1 · · · exp Gm.
In particular h ≥ δ > 0 on E. By Tietze's extension theorem, we may
assume that Gj ∈ Mn(C(X, C)). Using the definition of B, choose a ∈ An
so that 8
h(x) −ba(x) <
ε
M HS
,
(6.1)
sup
x∈E
where M := exp G1 · · · exp Gm and where ε is so small that Z(a) ∩ Z(bg) =
Z(a) ∩ E = ∅. Thus (a, g) ∈ Un+1(A). The hypothesis (1) (resp. (4) implies
that (a, g) is reducible in A (resp. reducible to the principal component of
Un(A)). That is, there is x ∈ An such that u := a + x g ∈ Un(A) (resp.
u ∈ P(Un(A))). In particular, bu = ba on E. Since vM ≤ v · M HS for
every vector v, we have the following estimates on E:
(cid:12)(cid:12)f E −bu exp G1 · · · exp Gm(cid:12)(cid:12) = (cid:12)(cid:12)f E −ba exp G1 · · · exp Gm(cid:12)(cid:12)
= (cid:12)(cid:12)(h −ba) exp G1 · · · exp Gm(cid:12)(cid:12)
≤ h −ba M HS < ε.
from Lemma 4.9 (resp. Lemma 4.10), applied to the algebra C(X, C), that
(f , g) is reducible in C(X, C) (resp. reducible to the principal component
of C(X, C) (whenever ε > 0 is small).
Since bu ∈ Un(C(X, C)) (resp. bu ∈ P(Un(C(X, C)))), we deduce
(2) =⇒ (1) (resp. (5) =⇒ (4) ) Let (f , g) ∈ Un+1(A). Hence bf andbg have
(2) (resp. (5)) and the assumption M (A) = X imply that (bf ,bg) is reducible
in C(X, C) (resp. reducible to the principal component of Un(C(X, C))).
By the Corach-Su´arez Theorem 1.3 (resp. Theorem 4.12) (f , g) is reducible
in A (resp. reducible to the principal component of Un(A)).
(cid:3)
no common zeros on X. By Theorem 5.6, (resp. Theorem 5.7) hypothesis
7. Exponential reducibility II
Here we introduce our second notion of exponential reducibility.
8Here M HS is the Hilbert-Schmidt norm.
20
RAYMOND MORTINI AND RUDOLF RUPP
Definition 7.1. Let A be a commutative unital Banach algebra over K with
identity element 1. Given
(a, g) := (a1, . . . , an, g) ∈ Un+1(A),
we call (a, g) exponentially reducible if there exists xj, bj ∈ A such that
nXj=1
exj (aj + bjg) = 1.
Observation 7.2. Let A be a commutative unital Banach algebra over K
such that
(1) bsr A = 1,
(2) U1(A) is connected.
Then every invertible pair (a, g) ∈ A is exponentially reducible.
Proof. By (1), a + bg ∈ U1(A) for some b ∈ A. Since U1(A) = exp A, we
arrive at a + bg = ex for some x ∈ A. This is of course equivalent to say
that e−x(a + bg) = 1.
(cid:3)
The following result gives a relation between exponential reducibility and
reducibility to the principal component.
Proposition 7.3. Let A be a commutative unital Banach algebra over K
and let (a, g) ∈ Un+1(A). Suppose that (a, g) is exponentially reducible.
Then (a, g) is reducible to the principal component of Un(A).
Proof. By assumption,Pn
Un(A) and b := (b1, . . . , bn) ∈ An; that is
j=1 exj (aj+bjg) = 1 for some v := (ex1 , . . . , exn) ∈
(a + b g) · vt = 1.
Since v ∈ Un(A) is reducible in A (if n ≥ 2), we deduce from Lemma 4.4
that
a + b g ∈ e1 · Exp Mn(A) = P(Un(A)).
In other words, (a, g) is reducible to the principal component of Un(A). The
case n = 1 is obvious.
(cid:3)
Remark Whereas for invertible pairs both notions coincide, exponential
reducibility of tuples of length at least three is, in general, a much stronger
requirement than reducibility to the principal component. As an example,
we take the disk algebra A(D). Let (z, f ) ∈ U2(D) be an invertible pair that
is not totally reducible (this means that there do not exist two invertible
functions u and v in A(D) such that uz + vf = 1, see [13] for the existence).
Then (z, f, 0) ∈ U3(A(D)). Since bsr A(D) = 1, U2(A(D)) is connected
([5]), and so U2(A(D)) coincides with its principal component. In particular
every invertible triple in A(D) is reducible to the principal component of
U2(A(D)). On the other hand, (z, f, 0) cannot be exponentially reducible,
since otherwise
ea1(z)(z + b1(z) · 0) + ea2(z)(f (z) + b2(z) · 0) = 1
REDUCIBILITY TO THE PRINCIPAL COMPONENT
21
for some functions aj, bj ∈ A(D). But an equation of the form ea1(z)z +
ea2(z)f (z) = 1 is not possible because by our choice, (z, f ) is not totally
reducible.
Examples of exponentially reducible tuples appeared in [10] (for pairs)
and [12] (for tuples):
Example 7.4. Let fj ∈ H ∞ and let b be an interpolating Blaschke product.
Suppose that (f1, . . . , fn, b) ∈ Un+1(H ∞). Then (f1, . . . , fn, b) is exponen-
tially reducible.
We do not know yet a characterization of the exponentially reducible
tuples (in none of the standard algebras).
8. Complementing left-invertible matrices
Based on the Arens-Novodvorski-Taylor theorem, we conclude our paper
by giving a simple analytic proof of a result by V. Ya. Lin [11, p. 127]
concerning extension of left invertible matrices. Although this proof seems
to be known among the specialists in the field, it never appeared explicitely
in print (see also the footnote in [11, p. 127]). One may view this result as
a companion result to Theorem 4.12 and Theorem 3.1.
Theorem 8.1 (Lin). Let A be a commutative unital Banach algebra over
C. Then a left-invertible matrix L over A can be complemented/extended to
an invertible matrix over A if and only if L 9 can be complemented in the
algebra C(M (A)).
Proof. It is sufficient to consider invertible rows (see [30, p. 345/346]).
So let a ∈ Un(A). Suppose that there exists an invertible matrix M ∈
By the Arens-Novodvorski-Taylor Theorem 4.11 (4), there is Q ∈ Mn(A)−1
Mn(C(M (A))) such that ba = e1M ; that is, ba is the first row of M .
such that bQ is homotopic in Mn(C(M (A)))−1 to M . Hence, for some
Gj ∈ Mn(C(M (A))),
and so
M = bQ eG1 . . . eGk ,
ba = (e1bQ) eG1 . . . eGk .
Let b be the first row of Q; that is b = e1Q. Then bb = e1bQ.
Since b ∈ Un(A), we see that ba and bb belong to the same component
of Un(C(M (A))). Hence, by another application of the Arens-Novodvorski
Theorem 4.11 (3), a and b belong to the same component of Un(A). Thus,
there exist Hj ∈ Mn(A) such that
a = b eH1 · · · eHk .
9 Here L = (dai,j) is the matrix formed with the Gelfand-transforms of the entries of L.
22
RAYMOND MORTINI AND RUDOLF RUPP
Consequently,
a = (e1Q) eH1 · · · eHk
= e1 R
for some R ∈ Mn(A)−1.
(cid:3)
Acknowledgements We thank Alexander Brudnyi for valuable com-
ments on the Arens-Novodvorski-Taylor Theorem 4.11 and the referee for
his suggestions.
References
[1] Arens, R., To what extent does the space of maximal ideals determine the algebras?
in: Function Algebras, Proc. Internat. Sympos. on Function Algebras, Tulane Univ.,
(1965-1966), 164 -- 168. 13
[2] Bass, H., K-theory and stable algebra, Publications Math´ematiques de L'I.H. ´E.S., 22
(1964), 5 -- 60. 1
[3] Burckel, R.B.,
An Introduction to Classical Complex Analysis, Birkhauser,
Basel+Stuttgart, 1979. 18
[4] Corach, G. and Su´arez, F. D., Stable rank in holomorphic function algebras, Illinois
J. Math. 29 (1985), 627 -- 639. 1, 2, 9, 12
[5] Corach, G. and Su´arez, F. D., Extension problems and stable rank in commutative
Banach algebras, Topology and its Applications 21 (1985), 1 -- 8. 1, 2, 13, 17, 20
[6] Corach, G. and Su´arez, F. D., On the stable range of uniform algebras and H ∞,
Proc. Amer. Math. Soc. 98 (1986), 607 -- 610. 2
[7] Corach, G. and Su´arez, F. D., Dense morphisms in commutative Banach algebras,
Trans. Amer. Math. Soc. 304 (1987), 537 -- 547. 1
[8] Gamelin, T.W., Uniform algebras, Chelsea, New York, 1984. 15, 19
[9] Jones, P.W., Marshall, D., and Wolff, T.H., Stable rank of the disc algebra, Proc.
Amer. Math. Soc. 96 (1986), 603-604. 1
[10] Laroco, L. A., Stable rank and approximation theorems in H ∞, Trans. Amer. Math.
Soc. 327 (1991), 815 -- 832. 7, 21
[11] Lin, V. Ya., Holomorphic fiberings and multivalued functions of elements of a Banach
algebra, Funct. Anal. Appl. 7 (1973), 122 -- 128; translation from Funkts. Anal. Prilozh. 7
(1973), 43 -- 51. 21
[12] Mortini, R., An example of a subalgebra of H ∞ on the unit disk whose stable rank
is not finite, Studia Math. 103 (1992), 275 -- 281. 21
[13] Mortini, R. and Rupp, R., Totally reducible elements in rings of analytic functions,
Comm. Algebra 20 (1992), 1705 -- 1713. 20
[14] Mortini, R. and Rupp, R., Mappings into the Euclidean sphere, Amer. Math.
Monthly 119 (2012), 485 -- 494. 17, 18
[15] Mortini, R. and Rupp, R., The Bass stable rank for the real Banach algebra A(K)sym,
J. Funct. Anal. 261 (2011), 2214 -- 2237. 1
[16] Mortini, R. and Rupp, R., Logarithms and exponentials in Banach algebras, Banach
J. Math. Anal. 9 (2015), 164 -- 172. 7, 8
[17] Mortini, R. and Wick, B., The Bass and topological stable ranks of H ∞
R (D) and
AR(D), Crelle Journal (Journal fur Reine und Angewandte Mathematik) 636 (2009),
175 -- 191. 1
[18] Novodvorski, M.E., Certain homotopical invariants of spaces of maximal ideals,
Mat. Zametki 1 (1967), 487 -- 494. 13
[19] Palmer, T.W., Banach algebras and the general theory of ∗-algebras, Vol 1, Cam-
bridge Univ. Press, London, 1994. 3
REDUCIBILITY TO THE PRINCIPAL COMPONENT
23
[20] Quadrat, A. and Robertz, D., Computation of bases free modules over the Weyl
algebras, J. Symbolic Comp. 42 (2007), 1113 -- 1141. 5
[21] Rieffel, M., Dimension and stable rank in the K-theory of C ∗-algebras, Proc. Lon-
don Math. Soc. 46 (1983), 301 -- 333. 5
[22] Rupp, R., Stable rank and boundary principle, Topology Appl. 40 (1991), 307 -- 316.
7, 15, 17
[23] Rupp, R., Stable rank and the ∂-equation, Canad. Math. Bull. 34 (1991), 113 -- 118. 7
[24] Rupp, R., Stable rank in Banach algebras, In: Function spaces (Edwardsville, IL,
1990), Lecture Notes in Pure and Appl. Math., 136, Dekker, New York, 1992, 357 -- 365. 7
[25] Rupp, R., Analytic functions on circular domains, Travaux mathmatiques, IV, 1 -- 81,
S´em. Math. Luxembourg, Centre Univ. Luxembourg, Luxembourg, 1992. 7
[26] Rupp, R., Zerofree extension of continuous functions on a compact Hausdorff space,
Topology Appl. 93 (1999), 65 -- 71. 3, 17
[27] Taylor, J.L., Topological invariants of the maximal ideal space of a Banach algebra,
Advances in Math. 19 (1976), 149 -- 206. 13
[28] Treil, S., The stable rank of the algebra H ∞ equals 1, J. Funct. Anal. 109 (1992),
130 -- 154. 1
[29] Vasershtein, L., Stable rank of rings and dimensionality of topological spaces, Funct.
Anal. Appl. 5 (1971), 102 -- 110; translation from Funkts. Anal. Prilozh. 5 (1971), No.2,
17 -- 27. 1
[30] Vidyasagar, M., Control System Synthesis: A Factorization Approach, MIT Press
Series in Signal Processing, Optimization, and Control, 7, MIT Press, Cambridge, MA,
1985. 21
Raymond Mortini, D´epartement de Math´ematiques et Institut ´Elie Cartan
de Lorraine, UMR 7502, Universit´e de Lorraine, Metz, France, raymond.mortini@univ-
lorraine.fr and Rudolf Rupp, Fakultat fur Angewandte Mathematik, Physik
und Allgemeinwissenschaften, TH-Nurnberg, Germany, [email protected]
|
1709.07060 | 1 | 1709 | 2017-09-20T20:05:24 | Continuity and growth of free multiplicative convolution semigroups | [
"math.FA"
] | Let $\mu$ be a compactly supported probability measure on the positive half-line and let $\mu^{\boxtimes t}$ be the free multiplicative convolution semigroup. We show that the support of $\mu^{\boxtimes t}$ varies continuously as $t$ changes. We also obtain the asymptotic length of the support of these measures. | math.FA | math | CONTINUITY AND GROWTH OF FREE MULTIPLICATIVE
CONVOLUTION SEMIGROUPS
XIAOXUE DENG AND PING ZHONG
Abstract. Let µ be a compactly supported probability measure on the positive half-line
and let µ⊠t be the free multiplicative convolution semigroup. We show that the support of
µ⊠t varies continuously as t changes. We also obtain the asymptotic length of the support
of these measures.
7
1
0
2
p
e
S
0
2
]
.
A
F
h
t
a
m
[
1
v
0
6
0
7
0
.
9
0
7
1
:
v
i
X
r
a
1. Introduction
Let µ and ν be probability measures on [0, ∞). The free convolution µ ⊠ ν represents the
distribution of product of two positive operators in a tracial W ∗-probability space whose
distributions are µ and ν respectively. We refer to [14] for an introduction to free probability
theory.
Given a probability measure µ on [0, ∞) not being a Dirac measure, it is known [4] that,
for any t > 1, the fractional free convolution power µ⊠t is defined appropriately, such that
it interpolates the discrete convolution semigroup {µ⊠n}n∈Z+ , where µ⊠n = µ ⊠ · · · ⊠ µ is
the n-fold free convolution. This is the multiplicative analogue of Nica-Speicher semigroup
[12] defined firstly for the free additive convolution. The free convolution semigroups obey
many regularity properties and has been studied extensively. See [15] for a survey on free
convolutions and other topics in free probability theory.
Denote by supp(µ) the support of the measure µ. We will prove the following result
about continuity of the support of µ⊠t, which is the analogue of the work [16] for free
multiplicative convolution on the positive half line.
Theorem 1.1. Let µ be a compactly supported probability measure on [0, ∞). Then the
supports of measures {supp(µ⊠t} change continuously in the Hausdorff metric with respect
to the parameter t.
We then study the asymptotic size of the support.
Theorem 1.2. Let µ be a compactly supported probability measure on [0, ∞) with mean
m1(µ) = R ∞
0 s dµ(s) = 1. We denote by µ⊠t = max{m : m ∈ supp(µ⊠t)}. Then
µ⊠t/t = eV.
lim
t→∞
where V is the variance of µ.
Theorem 1.2 generalizes Kargin's work [11] (see also [2]) to continuous semigroup. Our
proof is different from Kargin's proof, but uses the density formula for free convolution
semigroups [10, 18]. The free multiplicative convolution on the unit circle is usually studied
together with the positive half line case.
It was shown in [1] that many results can be
1
2
deduced from results on free additive convolutions. Since analogue results were known in
additive case, a separate work on free multiplicative convolution on the unit circle become
unnecessary. Hence we focus on measures on the positive half line in our article.
An estimation for the size of support of free additive convolution semigroups was obtained
in [3, 8].
In light of the proof of Theorem 1.2, it is very likely higher order asymptotic
expansion can be obtained for free additive convolution semigroups. We plan to investigate
it in a forthcoming work as it may have some applications in quantum information theory.
The paper is organized as follows. In Section 2, we collect some known regularity results
about free multiplicative convolution semigroups. In Section 3, we give the proof for The-
orem 1.1. We study the asymptotic behavior of free multiplicative convolution semigroups
and give the proof for Theorem 1.2 in Section 4.
2. Free convolution on the positive half line
Let µ be a probability measure on [0, ∞). The ψ-transform of µ is the moment generating
function of µ defined as
ψµ(z) = Z ∞
0
zs
1 − zs
dµ(s),
which is analytic on Ω = C\[0, ∞). The η-transform of µ is defined as ηµ = ψµ/(1 + ψµ) on
the same domain as the ψ-transform.
Any probability measure µ on [0, ∞) can be recovered from its η-transform by Stieltjes
inversion formula. Indeed, we have the identity
(2.1)
Gµ(cid:18) 1
z(cid:19) =
z
1 − ηµ(z)
,
z ∈ Ω,
where Gµ is the Cauchy transform of µ.
µ(z) > 0
for z < 0, and therefore ηµ(−∞, 0) is invertible. Let η−1
µ be the inverse of ηµ and set
Σµ(z) = η−1
µ (z)/z, where z < 0 is sufficiently small. The free convolution of two such
probability measures µ and ν is determined by Σµ⊠ν(z) = Σµ(z)Σν (z). In particular, the
n-th order free multiplicative convolution power µ⊠n of µ satisfies the identity
If µ is not a Dirac measure, then η′
(2.2)
Σµ⊠n (z) = Σn
µ(z),
where z < 0 is sufficiently small.
We now briefly recall the construction of µ⊠t which interpolates the relation (2.2) as
follows. Let κµ(z) = z/ηµ(z) for z ∈ Ω and one can write κµ(z) = exp(u(z)), where u is an
analytic function on Ω and can be expressed as
(2.3)
u(z) = a +Z ∞
0
1 + zs
z − s
dρ(s),
where a = − log ηµ(i) and ρ is a finite positive Borel measure on [0, ∞) following [10,
Proposition 4.1].
We define Φt(z) := z exp[(t − 1)u(z)].
It turned out that the function Φt is the right
inverse of Voiculescu subordination function ωt [4]. More precisely, we have
Φt(ωt(z)) = z,
and ηµ⊠t (z) = ηµ(ωt(z))
3
for all z ∈ Ω and t > 1. It turns out that the function ωt can be regarded as the η-transform
of a ⊠-infinitely divisible measure on [0, ∞) and the function ηµ⊠t can be retrieved from ωt.
We refer to [4, 10] for more details.
The following result was proved in [4].
Theorem 2.1. Let µ be a probability measure on [0, ∞) and t > 1.
(1) A point x ∈ (0, ∞) satisfies ηµ⊠t (x) = 1 if and only if x−1/t is an atom of µ with
mass µ({x−1/t}) ≥ (t − 1)/t. If µ({x−1/t}) > (t − 1)/t, then 1/x is an atom of µ⊠t,
and
µ⊠t ({1/x}) = tµ({x−1/t}) − (t − 1).
(2) The nonatomic part of µ⊠t is absolutely continuous and its density is continuous
except at the finitely many points x such that η⊠t(x) = 1.
(3) The density of µ⊠t is analytic at all points where it is different from zero.
The study of regularity property of free convolutions relies on Voiculescu's subordination
result [4, 5, 7, 13]. By a careful study of boundary behavior of subordination functions, we
were able to give a formula for the density function of absolutely continuous part of µ⊠t in
[10]. To describe our result, we need some auxiliary functions studied in [10].
Let g be a function defined on (0, ∞) × (0, π) by
(2.4)
g(r, θ) = −
ℑu(reiθ)
θ
=
r sin θ
θ
Z ∞
0
s2 + 1
r2 − 2rs cos θ + s2 dρ(s).
The function g(r, θ) is decreasing on (0, π) for any r ∈ (0, ∞) fixed and limθ→π− g(r, θ) = 0.
We then set
g(r) = lim
θ→0
g(r, θ) = Z ∞
0
At(r) = inf (cid:26)θ ∈ (0, π) : g(r, θ) <
and
r(s2 + 1)
(r − s)2 dρ(s)
t − 1(cid:27) .
1
It is clear that if At(r) > 0, then g(r, At(r)) = 1/(t − 1). We further let ht(r) := Φt(reiAt(r))
and
t = (cid:26)r ∈ (0, ∞) : g(r) >
V +
1
t − 1(cid:27) .
The function ht is a homeomorphism of [0, ∞) and limr→∞ ht(r) = ∞.
The functions defined above can be used to describe the image set Ωt = ωt(C+). The set
t (C+ ∪ (−∞, 0)) having the negative half plane
Ωt is in fact the connected component of Φ−1
as part of its boundary. Moreover, we proved that
and
Ωt = {reiθ : At(r) < θ < π, r ∈ (0, ∞)},
∂Ωt = (−∞, 0] ∪ {reiθ : θ = At(r), r ∈ (0, ∞)}.
The following result is one of main results in [10].
Theorem 2.2. Suppose that µ is a probability measure on [0, ∞) not being a Dirac measure
and t > 1. Let St = (cid:8)1/ht(r) : r ∈ V +
(1) The measure (µ⊠t)ac is concentrated on the closure of St.
t (cid:9). Then the following statements hold.
4
(2) The density of (µ⊠t)ac is analytic on the set St and is given by
d(µ⊠t)ac
dx
(cid:18) 1
ht(r)(cid:19) =
1
π
ht(r)lt(r) sin θt(r)
1 − 2lt(r) cos θt(r) + l2
t (r)
,
r ∈ V +
t
,
where lt(r) = r exp ℜu(reiAt(r)) and θt(r) = tAt(r)/(t − 1) for r ∈ V +
t .
(3) The number of components in supp(µ⊠t)ac is a decreasing function of t.
3. Continuity of free convolution semigroups
In this section, we assume that the probability measure µ on [0, ∞) is compactly sup-
ported.
Lemma 3.1. Let µ be a probability measure on [0, ∞). Set κµ(z) = z/ηµ(z) and write
κµ(z) = exp[u(z)], where u is give by (2.3). Then limx→0− κ(x) = 1/m1(µ) and
1
s
Moreover, when m1(µ) = 1, we have
Z ∞
0
dρ(s) = log m1(µ) + a.
Z ∞
0
1 + s2
s2
dρ(s) = V,
where V is the variance of µ.
Proof. Observe that limr→0− ηµ(r)/r = m1(µ) by the definition of ηµ. We write u as
u(z) = a − zZ ∞
and hence limr→0− u(r) = a −R ∞
We calculate
0
1 dρ(s) − (z2 + 1)Z ∞
0
1
s − z
dρ(s),
0 1/s dρ(s). We then deduce the first equation.
u′(r) = −Z ∞
s2 + 1
(r − s)2 dρ(s),
0
for all r < 0 and, by Monotone Convergence Theorem,
s2 + 1
lim
r→0−
u′(r) = −Z ∞
0
dρ(s).
s2
On the other hand, when m1(µ) = 1, we have u(z) = − ln (ηµ(z)/z) = − ln(1 + V z + o(z)),
which yields that u′(0−) = −V . Therefore, we deduce that
V = Z ∞
0
s2 + 1
s2
dρ(s).
This finishes the proof.
The following result is from [10].
(cid:3)
Lemma 3.2. Let I be an open interval contained in (V +
convex on I.
t )c, then ρ(I) = 0 and g is strictly
Proposition 3.3. Let µ be a compactly supported probability measure on [0, ∞) and 1 <
α < β. Write κµ(z) = exp[u(z)], where u is an analytic function given by (2.3).
(1) Then there exists a > 0 such that [0, a)∩supp(ρ) = ∅. Moreover, the set V +
t
bounded away from zero and increasing for all t ∈ (α, β).
5
is uniformly
(2) Given any b > 0, the sets V +
t ∩ [0, b] are Hausdorff continuous with respect to t.
Proof. By the definition of η-transform, we have ηµ(0) = 0 and η′
identity
µ(0) = m1(µ) < ∞. The
z
1 − ηµ(z)
= Gµ(cid:18) 1
z(cid:19) = Z ∞
0
1
1/z − x
dµ(x),
implies that ηµ is real on some interval [0, a) under the assumption that µ is compactly
supported.
It follows that u is also real on [0, a) and hence ρ([0, a)) = 0 by Stieltjes
inversion formula. Hence, the function g is finite in a neighborhood of zero and g(r) =
is bounded away from 0. The definition of
0
V +
t
r(s2+1)
(r−s)2 dρ(s) → 0 as r → 0. We see that V +
immediately implies that V +
R ∞
t1 ⊂ V +
Assume now tn → t ∈ (α, β) and (V +
t
t ∩ [0, b]) is
tn ∩ [0, b])\Bǫ(V +
a ǫ-neighborhood of the set V +
t ).
We may assume that rn → r by passing to a subsequence if necessary. Lemma 3.3 implies
that supp(ρ) ⊂ Bǫ(V +
t ∩ [0, b]. We then have a series rn ∈ (V +
t ) and hence we can take the limit
t2 if t1 < t2. This proves the first assertion.
t ∩ [0, b]), where Bǫ(V +
tn ∩ [0, b]) 6⊂ Bǫ(V +
g(r) = lim
n→∞
g(rn) = lim
n→∞
1
tn − 1
.
On the other hand, g ≤ 1
(V +
t )c. This contradiction proves the second claim.
t−1 and g is strictly convex on any open interval contained in
(cid:3)
Proposition 3.4. Given b > 0, the graphs {reiAt(r) : r ∈ V +
Hausdorff metric for t ∈ (1, ∞).
t ∩ (0, b)} are continuous in the
Proof. For 0 < c < π, we define V +
prove that there exists δ > 0 such that
t,c = {r ∈ V +
t
: At(r) ≥ c}. Given ǫ > 0, we will start to
(3.1)
At(r) < As(r) ≤ (1 + ǫ)At(r)
for all r ∈ V +
t,c ∩ (0, b) if 0 < s − t < δ. The first inequality follows from the fact that the
function g(r, θ) is a decreasing function of θ. To prove the second inequality by contradiction,
we assume that there exists a series tn > t, tn → t and rn, r ∈ V +
t,c ∩ (0, b) such that rn → r
and Atn (rn) > (1 + ǫ)At(rn). We then have
g(rn, (1 + ǫ)At(rn)) ≥ g(rn, Atn(rn)) =
1
tn − 1
.
As At(rn) ≥ c, we can take the limit and obtain
g(r, (1 + ǫ)At(r)) ≥
1
t − 1
,
due to the fact that the integrand in (2.4) is bounded away from zero and At is continuous.
On the other hand, we have g(r, At(r)) = 1
t−1 and g(r, θ) is a strictly decreasing function of
θ. This contradiction yields the second inequality in (3.1).
6
Given ǫ > 0, using the similar argument as above, we can prove that there exists δ > 0
such that
(3.2)
(1 − ǫ)At(r) ≤ As(r) < At(r)
for all r ∈ V +
t,c ∩ (0, b) if −δ < s − t < 0.
We claim that
sup{As(r) : r ∈ (0, b)\V +
t,c} ≤ 2c
if s − t is small enough. Assume that is not the case, then there exists a series tn → t and
rn → r, where rn /∈ V +
t,c ∩ (0, b), such that Atn(rn) > 2c. We have
g(rn, 2c) ≥ g(rn, Atn(rn)) =
1
tn − 1
.
Taking the limit, we have g(r, 2c) ≥ 1/(t − 1), which implies that At(r) ≥ 2c. As the cluster
set of rn /∈ V +
t,c, it must satisfy At(r) ≤ c. This
contradiction proves our claim.
t,c or an end point of V +
t,c, r is either not in V +
The desired assertion follows by above results and applying Proposition 3.4.
(cid:3)
Proposition 3.5. Let µ be a compactly supported probability measure on [0, ∞). If µ({0}) >
1, then µ⊠t({0}) = µ({0}) for all t > 1.
If 0 ∈ supp((µ⊠s)ac) for some s > 1, then
0 ∈ supp((µ⊠t)ac) for all t > 1.
Proof. It follows from [6] that rGµ(r) → µ({0}) as r → 0− on the negative half line. Hence,
µ({0}) = 1 + limr→−∞ ψµ(r). For any t > 0, as limr→−∞ ωt(r) = −∞, we have
µ⊠t({0}) = 1 + lim
r→−∞
ψµ(ωt(r)) = 1 + lim
r→−∞
ψµ(r) = µ({0}).
Assuming that 0 ∈ supp((µ⊠s)ac) for some s > 1, we claim that ρ is not compactly
supported. Indeed, if ρ is compactly supported, then limr→∞ g(r) = 0 and hence the set
As is compact and the set supp((µ⊠s)ac) is bounded away from zero by Theorem 2.2, which
contradicts to the assumption. The fact that the measure ρ is not compactly supported in
turn implies that Vt is not compact and 0 ∈ supp((µ⊠t)ac) for all t > 1 by according to
Theorem 2.2.
(cid:3)
We are now in a position to prove the main result of this section.
Proof of Theorem 1.1. We first focus on the absolutely continuous part. Given t0 > 1,
δ ∈ (0, t0 − 1) and choose a such that V +
t0+δ ⊂ [a, ∞). For s ∈ (t0 − δ, t0 + δ), we will prove
that
(cid:12)(cid:12)(cid:12)(cid:12)
ht0 (x)
hs(y) (cid:12)(cid:12)(cid:12)(cid:12)
= (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Φt0(xeiAt0 (x))
Φs(yeiAs(y)) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= O(cid:18) x
y(cid:19) ,
(3.3)
for x ∈ V +
t0 and y ∈ V +
s ∩ V +
t0 . We fix x0 ∈ V +
t0 and write
Φt0(x0eiAt(x0))
Φs(yeiAs(y))
=
Φt0(yeiAs(y))
Φs(yeiAs(y))
Φt0(x0eiAt(x0))
Φt0(yeiAs(y))
.
We observe that the function (s, x) → As(x) is continuous in a neighborhood of (t0, x0)
such that x0 ∈ V +
t0 by the definition (2.4). When δ is sufficiently small, we thus have
7
= exp[(t0 − s)u(yeiAs(y))] ≤ exp[(t0 − s)C],
where C is a constant depending on δ and y. To estimate the second factor, denote z1 =
x0eiAt(x0) and z2 = yeiAs(y). We have
= (cid:18) x0
y (cid:19) · exp[(t0 − 1)(u(z1) − u(z2))]
Φt0(z1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Φt0(yeiAs(y))
Φs(yeiAs(y)) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
Φt0(z2)(cid:12)(cid:12)(cid:12)(cid:12)
and
u′(z) = −Z ∞
0
1 + s2
(z − s)2 dσ(s).
For any z = reiθ ∈ Ωs ∩ Ωt0, we have that
r sin(θ)
θ
Z ∞
0
s2 + 1
z − s2 dσ(s) <
1
t0 − δ − 1
,
which yields that
u′(z) ≤ Z ∞
0
1 + s2
z − s2 dz
1
θ
·
≤
r sin(θ)
t0 − δ − 1
.
We now choose a curve γ ⊂ Ωs ∩ Ωt0 connecting z1 and z2, and obtain
u(z1) − u(z2) = (cid:12)(cid:12)(cid:12)(cid:12)
Zγ
u′(z)dz(cid:12)(cid:12)(cid:12)(cid:12)
≤
1
t0 − δ − 1 Zγ
θ
r sin(θ)
dz .
The curve γ can be chosen such that θ/(r sin(θ)) is bounded along γ when δ is chosen to be
small thanks to Proposition 3.4. The length of γ can be arbitrarily small if x0/y is. We then
deduce (3.3). Moreover, this estimation is uniform over compact subsets of V +
t0 ∩ [a, ∞].
Hence, for any point ht0(x) = Φt0(xeiAt0 (x)) with x ∈ V +
there exists y ∈ V +
and choose b such that ǫ > max{1/hs(x) : x ∈ V +
t0 > 1, we thus have
t0 ∩ [a, b] and δ small enough,
s such that ht0 (x)/hs(y) is sufficiently small. Let ǫ > 0 be a small number
t0+δ ∩ [a, b], s − t0 < δ}. Given ǫ > 0 and
and
supp((µ⊠t0)ac) ∩ [ǫ, ∞) ⊂{1/x : x = ht0 (r), r ∈ V +
t0 ∩ [a, b]}
⊂ Bǫ(supp(µ⊠s)ac),
supp((µ⊠s)ac) ∩ [ǫ, ∞) ⊂{1/x : x = hs(r), r ∈ V +
s ∩ [a, b]}
⊂ Bǫ(supp(µ⊠t0)ac),
if δ is sufficiently small.
To discuss the behavior of supp((µ⊠t)ac) close to zero, we have two cases.
supp((µ⊠t)ac), then 0 ∈ supp((µ⊠s)ac) by Proposition 3.5.
If 0 ∈
If 0 /∈ supp((µ⊠t)ac), then ρ
8
is bounded away from ∞ and thus supp((µ⊠t)ac) is bounded away zero.
above inclusions imply the sets {supp(µ⊠t)ac} is Hausdorff continuous.
In both cases,
Finally, atoms of µ⊠t change continuously as time evloves by Theorem 2.1 and Proposition
(cid:3)
3.5. This completes the proof.
4. Estimation of norm of free multiplicative convolution semigroups
We give an estimation of the size of the support of µ⊠t for a compactly supported prob-
ability measure on [0, ∞).
Proof of Theorem 1.2. Let a > 0 be such that supp(ρ) ⊂ (a, ∞). As limr→0 g(r) = 0, the
set V +
t } > 0. By Theorem 1.2, we
t
have
is bounded away from zero. Let αt = min{r : r ∈ V +
By the choice of αt, we have At(αt) = 0,
µ⊠t =
1
.
ht(αt)
and
g(αt) = Z ∞
0
αt(s2 + 1)
(αt − s)2 dρ(s) =
1
t − 1
ht(αt) = Φt(αt) = αt exp[(t − 1)u(αt)].
By (2.3) and the assumption m1(µ) = 1, we have
u(αt) = a +Z ∞
0 (cid:18) 1
s
0
1 + αts
αt − s
dρ(s)
+
1 + αts
αt − s (cid:19) dρ(s)
αt(1 + s2)
(αt − s)2
αt − s
s
dρ(s)
= Z ∞
= Z ∞
= −Z ∞
0
0
1
= −
t − 1
αt(1 + s2)
(αt − s)2 dρ(s) + αtZ ∞
+ αtZ ∞
αt(1 + s2)
(αt − s)2
1
s
0
0
dρ(s).
αt(1 + s2)
(αt − s)2
1
s
dρ(s)
It is clear that limt→∞ αt = 0 and hence
1 = lim
t→∞
g(αt)(t − 1) = lim
t→∞
(tαt) ·(cid:18)Z ∞
0
s2 + 1
s2
dρ(s)(cid:19) ,
which yields that limt→∞(tαt) = 1/V thanks to Lemma 3.1. We then obtain
lim
t→∞
u(αt)(t − 1) = −1 + lim
t→∞
[(t − 1)αt] · αtZ ∞
0
1 + s2
(αt − s)2
1
s
dρ(s)
as supp(ρ) is bounded below from zero.
= −1
Finally, we have
µ⊠t
t
lim
t→∞
= lim
t→∞
1
tht(αt)
= lim
t→∞(cid:18) 1
= e ·Z ∞
0
t · αt
s2 + 1
s2
· exp[−(t − 1)u(αt)](cid:19)
dρ(s)
This proves the desired result.
= eV.
9
(cid:3)
Proposition 4.1. Let µ be a probability measure supported on [c, d] not being a Dirac
measure with c, d > 0. There exists T such that the sets V +
t and supp(µ⊠t) have only one
connected component for all t > T .
Proof. The measure ρ determined by (2.3) is compactly supported since supp(µ) ⊂ [c, d].
Hence limr→0 g(r) = limr→∞ g(r) = 0 and V +
t
is bounded away from zero and ∞.
t0 , ρ(I) = 0 by Lemma 3.2. Hence the function g(r) = R ∞
Let [a, b] ⊂ (0, ∞) be a finite interval such that supp(ρ) ⊂ [a, b]. Fix t0 > 1. For any open
r(s2+1)
(r−s)2 dρ(s)
t0 . Therefore, we can choose T large enough, so
T . The set V +
T has only one connected component as supp(ρ) ⊂ [a, b]. It then
t and hence supp(µ⊠t) have only one connected component for all t > T by
(cid:3)
interval I ⊂ [a, b]\V +
has a positive local minimum at [a, b]\V +
that [a, b] ⊂ V +
follows that V +
Theorem 2.2.
0
Proposition 4.2. Let µ be a probability measure on [c, d] not being a Dirac measure for
some c, d > 0. Define
and
Then
at = sup{a : a < x for
all x ∈ supp(µ⊠t)},
bt = inf{b : b > x for
all x ∈ supp(µ⊠t)}.
lim
t→∞
(at)1/t = (cid:18)Z ∞
0
x−1dµ(x)(cid:19)−1
,
and
lim
t→∞
(bt)1/t = Z ∞
0
x dµ(x).
Proof. The set V +
t
βt = max{r : r ∈ V +
By adapting the argument in the proof of Theorem 1.2, we have
is bounded away from zero and ∞. Let αt = min{r : r ∈ V +
t } > 0 and
t } < ∞, we have limt→∞ αt = 0 and limt→∞ βt = 0 by Theorem 2.2.
lim
t→∞
tαt = m1(µ)/V
and
lim
t→∞
(t/βt)Z ∞
0
(s2 + 1) dρ(s) = 1.
On the other hand, we have
and
lim
t→∞
exp[u(αt)] = lim
z→0
exp[u(z)] = lim
z→0
z/ηµ(z) = 1/m1(µ)
lim
t→∞
exp[u(βt)] = lim
z→∞
exp[u(z)] = lim
z→∞
z/ηµ(z) = Z ∞
0
x−1dµ(x).
10
Therefore, we have
lim
t→∞
(at)1/t = lim
t→∞
(ht(βt))−1/t
{βt exp[(t − 1)u(βt)]}−1/t
= lim
t→∞
= lim
t→∞
exp[−u(βt)]
and
= (cid:18)Z ∞
0
x−1dµ(x)(cid:19)−1
;
lim
t→∞
(bt)1/t = lim
t→∞
(ht(αt))−1/t
= lim
t→∞
= lim
t→∞
{αt exp[(t − 1)u(αt)]}−1/t
exp[−u(αt)] = m1(µ).
This finishes the proof.
(cid:3)
Remark 1. Proposition 4.2 should be compared with the main result in [9], where weak limit
of rescaled discrete semigroups were studied. Our method is not suitable to study weak limit,
but works to estimate the asymptotic bound of free multiplicative convolution semigroups.
The interested reader can easily generalize results in [9] to continuous free multiplicative
convolution semigroups using their method.
Remark 2. Free multiplicative convolution semigroup can also be defined for measures on
the unit circle [4]. Let µ be a probability measure on the unit circle not being a Dirac
measure, it is known in [17, Proposition 3.26] that µ⊠t converges to the uniform measure
on the unit circle as t → ∞.
5. Acknowledgements
The second author wants to thank Hari Bercovici, Alexandru Nica, Jiun-Chau Wang
and John Williams for useful discussions. This work was partially supported by NSFC no.
11501423, 11431011.
References
[1] Michael Anshelevich and Octavio Arizmendi, The exponential map in non-commutative probability, Int.
Math. Res. Notices 17 (2017), 5302 -- 5342.
[2] Octavio Arizmendi and Carlos Vargas, Products of free random variables and k-divisible non-crossing
partitions, Electron. Commun. Probab. 17 (2012), no. 11, 13. MR 2892410
[3] Serban Belinschi, Benoıt Collins, and Ion Nechita, Eigenvectors and eigenvalues in a random subspace
of a tensor product, Invent. Math. 190 (2012), no. 3, 647 -- 697. MR 2995183
[4] Serban T. Belinschi and Hari Bercovici, Partially defined semigroups relative to multiplicative free con-
volution, Int. Math. Res. Not. (2005), no. 2, 65 -- 101. MR 2128863
[5]
, A new approach to subordination results in free probability, J. Anal. Math. 101 (2007), 357 -- 365.
MR 2346550
[6] Hari Bercovici and Dan Voiculescu, Regularity questions for free convolution, Nonselfadjoint operator
algebras, operator theory, and related topics, Oper. Theory Adv. Appl., vol. 104, Birkhauser, Basel,
1998, pp. 37 -- 47. MR 1639647
11
[7] Philippe Biane, Processes with free increments, Math. Z. 227 (1998), no. 1, 143 -- 174. MR 1605393
[8] Benoıt Collins, Motohisa Fukuda, and Ping Zhong, Estimates for compression norms and additivity
violation in quantum information, Internat. J. Math. 26 (2015), no. 1, 1550002, 20. MR 3313648
[9] Uffe Haagerup and Soren Moller, The law of large numbers for the free multiplicative convolution,
Operator algebra and dynamics, Springer Proc. Math. Stat., vol. 58, Springer, Heidelberg, 2013, pp. 157 --
186. MR 3142036
[10] Hao-Wei Huang and Ping Zhong, On the supports of measures in free multiplicative convolution semi-
groups, Math. Z. 278 (2014), no. 1-2, 321 -- 345. MR 3267581
[11] Vladislav Kargin, On asymptotic growth of the support of free multiplicative convolutions, Electron.
Commun. Probab. 13 (2008), 415 -- 421. MR 2424965
[12] Alexandru Nica and Roland Speicher, On the multiplication of free N -tuples of noncommutative random
variables, Amer. J. Math. 118 (1996), no. 4, 799 -- 837. MR 1400060
[13] Dan Voiculescu, The analogues of entropy and of Fisher's information measure in free probability theory.
I, Comm. Math. Phys. 155 (1993), no. 1, 71 -- 92. MR 1228526
[14] Dan Voiculescu, Ken Dykema, and Alexandru Nica, Free random variables, CRM Monograph Series,
vol. 1, American Mathematical Society, Providence, RI, 1992, A noncommutative probability approach
to free products with applications to random matrices, operator algebras and harmonic analysis on free
groups. MR 1217253 (94c:46133)
[15] Dan Voiculescu, Nicolai Stammeier, and Moritz Weber, Free probability and operator algebras, Munster
Lectures in Mathematics, European Mathematical Society, 2016.
[16] John Williams, On the hausdorff continuity of free L´evy processes and free convolution semigroups, J.
Math. Anal. Appl. (online 2016).
[17] Ping Zhong, Free Brownian motion and free convolution semigroups: multiplicative case, Pacific J.
Math. 269 (2014), no. 1, 219 -- 256. MR 3233917
[18]
, On the free convolution with a free multiplicative analogue of the normal distribution, J. Theoret.
Probab. 28 (2015), no. 4, 1354 -- 1379. MR 3422934
Xiaoxue Deng: School of Mathematics and Statistics, Wuhan University, No. 299 BaYi
Road, Wuhan, Hubei 430074, China
E-mail address: [email protected]
Ping Zhong: School of Mathematics and Statistics, Wuhan University, No. 299 BaYi Road,
Wuhan, Hubei 430074, China; and Department of Pure Mathematics, University of Waterloo,
Waterloo, ON, Canada
E-mail address: [email protected]
|
1806.06691 | 1 | 1806 | 2018-06-18T13:54:25 | Improved Ingham-type result on $\mathbb R^d$ and on connected, simply connected nilpotent Lie Groups | [
"math.FA"
] | In \cite{BRS} we have characterized the existance of a non zero function vanishing on an open set in terms of the decay of it's Fourier transform on the $d$-dimensional Euclidean space, the $d$-dimensional torus and on connected, simply connected two step nilpotent Lie groups. In this paper we improved these results on $\mathbb R^d$ and prove analogus results on connected, simply connected nilpotent Lie groups. | math.FA | math |
IMPROVED INGHAM-TYPE RESULT ON Rd AND ON CONNECTED, SIMPLY
CONNECTED NILPOTENT LIE GROUPS
MITHUN BHOWMIK
Abstract. In [3] we have characterized the existance of a non zero function vanishing on an open set
in terms of the decay of it's Fourier transform on the d-dimensional Euclidean space, the d-dimensional
torus and on connected, simply connected two step nilpotent Lie groups. In this paper we improved
these results on Rd and prove analogus results on connected, simply connected nilpotent Lie groups.
1. Introduction
It is a well known fact in harmonic analysis that if the Fourier transform of an integrable function on
the real line is very rapidly decreasing then the function can not vanish on a nonempty open set unless it
vanishes identically. For instance if, let f ∈ L1(R) and a > 0 be such that its Fourier transform satisfies
the estimate
bf (ξ) ≤ Ce−aξ,
for all ξ ∈ R.
If f vanishes on a nonempty open set then f is identically zero. This initial observation motivates one
to endeavour for a more optimal decay of the Fourier transform bf for such a conclusion to hold. For
instance we may ask: if bf decays faster than 1/(1 + · )n for all n ∈ N but slower than the function e−a·,
can f vanish on a nonempty open set without being identically zero? A more precise question could be:
is there a nonzero integrable function f on R vanishing on a nonempty open set such that its Fourier
transform satisfies the following estimate
(1.1)
log ξ ,
for large ξ?
bf (ξ) ≤ Ce− ξ
The answer to the above question is in the negative and follows from a classical result due to Ingham
[7]. Analogous results were also obtained by Paley-Wiener, Levinson and Hirschman [11, 10, 9, 6]. In
[3], we have obtained analogues results on d-dimensioan Euclidean spaces, d-dimensional torus and on
connected, simply connected two step nilpotent Lie groups. In the context of Euclidean space we have
the following theorem [3]. For f ∈ L1(Rd), we shall define its Fourier transform bf by
f (x) e−2πix·ξ dx,
for ξ ∈ Rd.
For 1 < p ≤ 2, the definition of Fourier transform of f ∈ Lp(Rd) is extended in the usual way.
bf (ξ) =ZRd
2010 Mathematics Subject Classification. Primary 22E25; Secondary 22E30, 43A80.
Key words and phrases. nilpotent Lie group.
This work was supported by Indian Statistical Institute, India (Research fellowship).
1
2
MITHUN BHOWMIK
Theorem 1.1. Let f ∈ Lp(Rd), for some p with p ∈ [1, 2] and ψ : [0, ∞) → [0, ∞) be a locally integrable
function. Suppose that the Fourier transform bf of f satisfies the following estimate
(1.2)
for almost every ξ ∈ Rd, where P (ξ) is a polynomial and we define
bf (ξ) ≤ CP (ξ)e−ψ(kξk),
I =Z ∞
ψ(t)
t2 dt.
1
(a) Let ψ(t) = θ(t)t, for t ∈ [0, ∞), where θ : [0, ∞) → [0, ∞) is a decreasing function which decreases
to zero as t goes to infinity. If f vanishes on any nonempty open set in Rd and I is infinite then
f is zero on Rd. Conversely, if I is finite, then given any positive real number l there exists a
nonzero even function f ∈ Cc(Rd) supported in ball B(0, l) of radius l, centered at zero satisfying
(1.2).
(b) Let ψ be an increasing function on [0, ∞). If f vanishes on any nonempty open set in Rd and I
is infinite then f is zero on Rd. Conversely, if I is finite, given any positive real number l there
exists a nonzero f ∈ Cc(Rd) supported in B(0, l) satisfying (1.2).
(c) If
supp f ⊆ {x ∈ Rd x · η ≤ s},
and I is infinite then f is zero on Rd. Conversely, if I is finite then there exists a nonzero
f ∈ L2(Rd) vanishing on the set {x ∈ Rd : x · η ≥ x0}, for some x0 ∈ R and η ∈ Sd−1 satisfying
(1.2).
In [3] we have shown that the original statements of the results of Ingham, Levinson and Paley-Wiener
on R are follows form the theorem above. We shall show in this paper that it is possible to prove analogues
of Theorem 1.1 under the following variant of (1.2)
(1.3)
ZRd
bf (ξ)qeqψ(kξk)
(1 + kξk)N
dξ < ∞,
where q ∈ [1, ∞] (with obvious modification for q = ∞) and N is a positive real number. Our next main
result in this paper is an analogue of Theorem 1.1, c) on the connected, simply connected nilpotent Lie
group G.
We shall use the following standard notation in this paper: supp f denotes the support of the function
f and C denotes a constant whose value may vary. For a finite set A we shall use the symbol #A to
denote the number of elements in A. B(0, r) denotes the open ball of radius r centered at 0 in Rd and
¯B(0, r) denotes its closure. For x, y ∈ Rd, we shall use kxk to denote the norm of the vector x and x · y
to denote the Euclidean inner product of the vectors x and y. A function ψ : Rd → [0, ∞) is said to
be radially increasing (or radially decreasing) if ψ is radial and satisfies the condition ψ(x) ≥ ψ(y) (or
ψ(x) ≤ ψ(y)) whenever x ≥ y. We shall often consider a radial function on Rd as an even function on
R or equivalently, as a function on [0, ∞).
2. Results of Ingham, Levinson and Paley-Wiener on Rd
In this section our aim is to improve Theorem 1.1 in the following way. Instead of the pointwide decay
(1.2) on the Fourier transform side we will consider the weighted Lq estimate for 1 ≤ q ≤ ∞ of the form
(1.3). Using convolution techniques we will be able to show that it is enough to prove the result for q = 1
and N = 0 and rest of the proof goes like the proof of Theorem 1.1. First we state and prove an analogue
of Theorem 1.1 (a) on Rd.
IMPROVED INGHAM-TYPE RESULT ON RD AND ON NILPOTENT LIE GROUPS
3
Theorem 2.1. Let θ : Rd → [0, ∞) be a radially decreasing measurable function with limkξk→∞ θ(ξ) = 0
and
I =Zkξk≥1
θ(ξ)
kξkd dξ.
(a) Let f ∈ Lp(Rd), p ∈ [1, 2], be such that
ZRd
bf (ξ)q eqθ(ξ)kξk
(1 + kξk)N
for some q ∈ [1, ∞) and some N ≥ 0 or
dξ < ∞,
(2.1)
(2.2)
(2.3)
bf (ξ) ≤ C P (ξ) e−θ(ξ)kξk,
for almost every ξ ∈ Rd,
where P is a polynomial. If f vanishes on a nonempty open subset in Rd and I is infinite, then
f is zero almost everywhere on Rd.
(b) If I is finite then there exists a nontrivial f ∈ C∞
c (Rd) satisfying the estimate (2.2) (or (2.3) if
q = ∞).
Proof. We shall first prove the assertion (b). For q = ∞ it follows from Theorem 1.1, (a). For q ∈ [1, ∞)
we choose φ1 ∈ C∞
c (Rd) with supp φ1 ⊆ B(0, l/2) and consider the function f = f0 ∗ φ1. Clearly, support
of the function f is contained in B(0, l) and
This, in particular, proves (b). It now remains to prove (a).
ZRd
bf (ξ)q eqθ(ξ)kξk
(1 + kξk)N
dξ ≤ CZRd
cφ1(ξ)q dξ < ∞.
For part (a), we first show that it suffices to prove the case q = 1, N = 0. Suppose f ∈ Lp(Rd)
vanishes on an open set U ⊆ Rd and satisfies (2.2), for some q > 1 and N ∈ N. Since the condition (2.2)
is invariant under translation of f we can assume that f vanishes on B(0, l), for some positive l. If we
choose φ ∈ C∞
c (Rd) supported in B(0, l/2) then f ∗ φ vanishes on the ball B(0, l/2). In fact, if kxk < l/2
then
ZRd
f (x − y)φ(y) dy =ZB(0, l
2 )
f (x − y)φ(y) dy = 0,
as
kx − yk ≤ kxk + kyk < l.
By using Holder's inequality we get
ZRd
≤ ZRd
< ∞,
\(f ∗ φ)(ξ) eθ(ξ)kξk dξ
(1 + kξk)N ! 1
bf (ξ)q eqθ(ξ)kξk
q
k(1 + kξk)Nbφ(ξ)kLq′ (Rd)
as N/q is smaller than N . Here q′ satisfies the relation
1
q
+
1
q′ = 1.
4
MITHUN BHOWMIK
Hence, by the case q = 1, N = 0 it follows that f ∗ φ vanishes identically. As φ ∈ C∞
c (Rd) we have that bφ
is nonzero almost everywhere. This implies that bf vanishes almost everywhere and so does f . The same
technique can be applied to reduce the case q = 1 and N ∈ N to the case q = 1 and N = 0 by using
Holder's inequality. For the case q = ∞, we get from (2.3) that
ZRd
[f ∗ φ(ξ) eθ(ξ)kξk dξ ≤ZRd
P (ξ) bφ(ξ) dξ < ∞.
So, without loss of generality, we assume that f ∈ Lp(Rd) such that bf satisfies the condition
(2.4)
ZRd
bf (ξ) eθ(ξ)kξk dξ = C′ < ∞.
Rest of the proof goes similarly as the proof of Theorem 2.2 in [3].
(cid:3)
As in Theorem 2.1 we will have the following version of Theorem 1.1, (b) on Rd.
Theorem 2.2. Let ψ : Rd → [0, ∞) be a radially increasing function and
I =Z ∞
1
ψ(t)
t2 dt.
(a) Let f ∈ Lp(Rd), p ∈ [1, 2], be such that
supp f ⊆ {x ∈ Rd x · η ≤ s},
for some η ∈ Sd−1 and s ∈ R and bf satisfies the estimate
dξ < ∞.
ZRd
(1 + kξk)N
bf (ξ)q eqψ(ξ·η)
bf (ξ) ≤ C (1 + kξk)N e−ψ(ξ·η),
If the integral I is infinite then f is the zero function.
for some q ∈ [1, ∞) and some N ≥ 0, or
for almost every ξ ∈ Rd.
(2.7)
(2.8)
ψ(ξ)
kξkd+1 dξ.
I =Zkξk≥1
ZRd
bf (ξ)q eqψ(ξ)
(1 + kξk)N dξ < ∞,
(a) Let f ∈ Lp(Rd), p ∈ [1, 2], be a function satisfying the estimate
(2.5)
(2.6)
for some q ∈ [1, ∞) and some N ≥ 0 or
bf (ξ) ≤ C P (ξ) e−ψ(ξ),
for almost every ξ ∈ Rd,
where P is a polynomial. If f vanishes on a nonempty open subset in Rd and I is infinite, then
f is zero almost everywhere on Rd.
(b) If I is finite then there exists a nontrivial f ∈ C∞
c (Rd) satisfying the estimate (2.5) (or (2.6) if
q = ∞).
In the following, we shall improved Theorem 1.1, c) which will be used in the next section to prove an
analogous result for connected simply connected nilpotent Lie groups (see Theorem 3.1).
Theorem 2.3. Let ψ : [0, ∞) → [0, ∞) be a locally integrable function and
IMPROVED INGHAM-TYPE RESULT ON RD AND ON NILPOTENT LIE GROUPS
5
(b) If ψ is non-decreasing and I is finite then there exists a nontrivial f ∈ Cc(Rd) satisfying the
estimate (2.7), for some q ∈ [1, ∞) and all η ∈ Sd−1 or (2.8), for q = ∞ and all η ∈ Sd−1.
Proof. We shall first prove a) for the case p = 2. Then for all p ∈ [1, 2) the result will follow by reducing
it to the case p = 2 as was done in the proof of Theorem 1.1, (c). As in the proof of Theorem 2.1 it
suffices to prove the case q = 1 and N = 0. Now, by using translation and rotation of the function
f , we can assume without loss of generality that η = e1 = (1, 0, · · · , 0) and s = 0. Then, by writing
ξ = (ξ1, ξ2, · · · , ξd) the hypothesis (2.7) becomes
(2.9)
For y ∈ Rd−1, we now define gy by
ZRd
bf (ξ) eψ(ξ1) dξ < ∞.
gy(x) = Fd−1f (x, y),
for almost every x ∈ R.
It then follows that for almost every y ∈ Rd−1, gy ∈ L2(R) with
supp gy ⊆ {x ∈ R x ≤ 0},
and by (2.9)
(2.10)
As y varies over a set of full (d − 1) dimensional Lebesgue measure, we just need to prove that gy is the
zero function. By the Paly-Wiener theorem (Theorem 2.7, [3]) it suffices to show that
ZR
1 + t2
1 + t2
dt = ∞.
bgy(t) eψ(t) dt < ∞.
ZR
log(bgy(t))
ZR
log(bgy(t)eψ(t))
dt = ZR
≥ ZR
ZR
log(bgy(t))
1 + t2
dt
dt < ∞
dt
1 + t2
log(bgy(t)eψ(t)) − ψ(t)
1 + t2 dt −ZR
log(bgy(t)eψ(t))
1 + t2
ψ(t)
dt.
If
(2.11)
then
ZR
log(bgy(t))
1 + t2
As I is infinite, it follows from (2.11) that
is divergent. Hence, by the Paly-Wiener theorem (Theorem 2.7, [3]) it follows that gy is the zero function.
Now, suppose
(2.12)
ZR
log(bgy(t)eψ(t))
1 + t2
dt = ∞.
6
MITHUN BHOWMIK
For a measurable function F on Rd, we define
and hence
log+ F (x) = max{log F (x), 0}
log− F (x) = − min{log F (x), 0},
log F (x) = log+ F (x) + log− F (x).
As log+ F (x) is always smaller than F (x) we get that
1 + t2
1 + t2
dt = ∞.
ZR
bgy(t)eψ(t)
1 + t2
log+(cid:0)bgy(t)eψ(t)(cid:1)
dt ≤ZR
ZR
log−(bgy(t)eψ(t))
log−(cid:0)bgy(t)eψ(t)(cid:1)
dt = Z(cid:8)t
cgy(t)eψ(t)≤1(cid:9)
≤ ZR
log− bgy(t)
bgy(t) ≤ bgy(t)eψ(t) ≤ 1,
log− (bgy(t)) ≥ log−(cid:16)bgy(t)eψ(t)(cid:17) .
ZR
ZR
log− (bgy(t))
log (bgy(t))
dt
dt
dt,
1 + t2
1 + t2
1 + t2
as on the set {t bgy(t)eψ(t) ≤ 1} we have
and hence
Therefore the integral
is divergent. Hence, the integral
by (2.10). From (2.12) we now conclude that
But
ZR
1 + t2
dt < ∞,
log−(cid:0)bgy(t)eψ(t)(cid:1)
1 + t2
dt
is divergent. By the Paly-Wiener theorem (Theorem 2.7, [3]) it now follows that gy is the zero function.
This completes the proof of part a). To prove Part b) we observe that since ψ is non decreasing
ψ(ξ · η) ≤ ψ(kξk),
for ξ ∈ Rd.
Therefore the proof follows from Theorem 2.2, (b).
(cid:3)
Remark 2.4.
(1) It is clear from the statement of Theorem 2.3 that if f is compactly supported con-
tinuous function then η ∈ Sd−1 can be taken to be arbitrary.
(2) If ψ is assumed to be a radial increasing function then Theorem 1.1, (c) follows from the case
p = ∞ of Theorem 2.3 and also from Theorem 2.2.
In this section, our aim is to prove an analogue of Theorem 2.3 in the context of connected, simply
connected nilpotent Lie groups.
3. Nilpotent Lie Groups
IMPROVED INGHAM-TYPE RESULT ON RD AND ON NILPOTENT LIE GROUPS
7
3.1. Preliminaries on Nilpotent Lie Groups. In this section, we shall discuss the preliminaries and
notation related to connected, simply connected nilpotent Lie groups. These are standard and can be
found, for instance, in [1, 2, 5].
Let G be a connected, simply connected nilpotent Lie group with Lie algebra g and let g
∗ be the vector
space of real-valued linear functionals on g. In this case, the exponential map exp : g → G becomes an
analytic diffeomorphism, which enables us to identify G with g ∼= Rd, where d = dim g. Let
(3.1)
{0} = g0 ⊂ g1 ⊂ · · · ⊂ gd = g
be a Jordan-Holder series of ideals of the nilpotent Lie algebra g such that dim gj = j, for j = 0, · · · , d
and ad(X)gj ⊆ gj−1, for j = 1, · · · , d and for all X ∈ g ([5], Theorem 1.1.9). We choose from this
sequence a Jordan-Holder basis
(3.2)
for j = 1, · · · , d. In particular, RX1 is contained in the center of g. Let {X ∗
dual to {X1, · · · , Xd}. We define coordinates on G by
{X1, · · · , Xd}, where Xj ∈ gj\gj−1,
1 , · · · , X ∗
d } be the basis of g
∗
(x1, · · · , xd) ∼= exp(x1X1 + · · · + xdXd).
By the Baker-Campbell-Hausdorff formula exp(X) exp(Y ) = exp(Z), where
(3.3)
Z = X + Y +
1
2
[X, Y ] +
1
12
[X, [X, Y ]] −
1
12
[Y, [X, Y ]] + · · ·
([5], P.12), which is a finite sum as G is a nilpotent Lie group. We introduce a 'norm function' on G
defined by
The composite map
kxk =(cid:0)x2
d(cid:1) 1
1 + · · · + x2
2 , x = exp(x1X1 + · · · + xdXd) ∈ G, xj ∈ R.
Rd −→ g −→ G,
(x1, · · · , xd) 7→
dXj=1
xjXj 7→ exp(
dXj=1
xjXj),
is a diffeomorphism and maps the Lebesgue measure on Rd to a Haar measure on G. We shall also
identify g
∗ with Rd via the map
ν = (ν1, · · · , νd) →
νjX ∗
j ,
νj ∈ R,
dXj=1
where {X ∗
1 , · · · , X ∗
d } is the dual basis of g
∗. On g
∗ we define the Euclidean norm relative to this basis by
k
dXj=1
νjX ∗
j k =(cid:0)ν2
d(cid:1) 1
1 + · · · + ν2
2 = kνk.
For ν ∈ g
∗ we have a natural bilinear map
This map is antisymmetric and the radical of this map is
(X, Y ) 7→ Bν(X, Y ) = ν([X, Y ]).
{X ∈ g Bν(X, Y ) = 0,
for all Y ∈ g} = rν .
8
MITHUN BHOWMIK
Clearly, Bν descends to nondegenerate, antisymmetric bilinear map Bν : g/rν × g/rν → R. Hence, rν is
of even codimension in g ([5], Lemma 1.3.2). Since the coadjoint orbit Oν of ν is diffeomorphic to G/Rν,
it follows that Oν is even dimensional. An index j ∈ {1, 2, · · · , d} is called a jump index for ν ∈ g
∗ if
rν + gj 6= rν + gj−1,
where the Jordan-Holder sequence gj are as in (3.2) (see [1]). Let e(ν) be the set of jump indices for
∗. This set contains dim(Oν ) indices, which is an even number. Moreover, there are disjoint sets
ν ∈ g
∗
of indices P and Q such that P ∪ Q = {1, · · · , d} and a G-invariant nonempty Zariski open set U ⊆ g
such that e(ν) = P , for all ν ∈ U ([5], Theorem 3.1.2; [1]). The elements of U are called generic linear
functionals. For ν ∈ U, we define
Then the Pfaffian P f (ν) is given by
Bν,P = ν([Xi, Xj])i,j∈P .
(3.4)
We define
P f (ν)2 = det Bν,P ,
for all ν ∈ U.
VP = span {X ∗
i : i ∈ P }, and VQ = span{X ∗
i : i ∈ Q}.
Let dν be the Lebesgue measure on VQ normalized so that the unit cube spanned by {X ∗
i : i ∈ Q} has
∗ = VQ ⊕ VP and W = U ∩ VQ is a cross-section for the coadjoint orbits through points
volume 1. Then g
in U.
For φ ∈ L1(G) we define its operator-valued Fourier transform by
π(φ) =ZG
φ(g)π(g−1) dg,
for π ∈ bG,
where dg denotes the Haar measure of G and bG is the unitary dual of G. If φ ∈ L1(G) ∩ L2(G) then π(φ)
is a Hilbert-Schmidt operator and kπ(φ)kHS will denote the Hilbert-Schmidt norm of π(φ). If dν denotes
the element of Lebesgue measure on W, then µ is the Plancherel measure for G, where dµ = P f (ν)dν.
The Plancherel formula is thus given by
kφk2
2 =ZG
φ(g)2 dg =ZW
kπν(φ)k2
HS dµ(ν),
for all φ ∈ L1(G) ∩ L2(G).
We shall now state and prove the following analogue of Theorem 1.1, (c) on a connected, simply
connected nilpotent Lie group G.
Theorem 3.1. Let ψ : [0, ∞) → [0, ∞) be a locally integrable function and
I =Z ∞
1
ψ(t)
t2 dt.
Suppose f ∈ Cc(G) is such that
(3.5)
ZW
kπν(f )k2
HS e2ψ(ν1) P f (ν) dν < ∞,
where ν1 = ν(X1) and X1 is given by (3.2). If I is infinite then f vanishes identically on G. Conversely,
if ψ is nondecreasing and I is finite, then there exists a nontrivial f ∈ Cc(G) such that (3.5) holds.
IMPROVED INGHAM-TYPE RESULT ON RD AND ON NILPOTENT LIE GROUPS
9
To prove the theorem we need some more notation. If r is the dimension of coadjoint orbits Oν , ν ∈ U,
then Q has d − r elements and so VQ can be identified with Rd−r. In an abuse of notation we write
VQ = RX ∗
1 ⊕ Rd−r−1 and let
p∗ : VQ → RX ∗
1 ,
ν 7→ ν1X ∗
1
denote the canonical projection. As W is a Zariski open set in VQ, p∗(W) = O is also a nonempty Zariski
open subset of R. In particular, this is a subset in R with full Lebesgue measure. So, it will be convenient
to write elements ν ∈ W as (ν1, ν′), where ν1 ∈ O and ν′ is in the set Wν1 defined by
Wν1 = {ν′ ∈ Rn−d−1 : (ν1, ν′) ∈ W}.
It turns out that Wν1 is also a Zariski open subset of Rn−d−1, for each fixed ν1 ∈ O. Given f ∈ Cc(G)
and the Jordan-Holder basis {X1, · · · , Xd} given in (3.2) we define for each y ∈ Rd−1
and f ∗
y (t) = fy(−t). We now consider the function
fy(t) = fexptX1 +
g(t) =ZRd−1
dXj=2
yjXj ,
for t ∈ R,
(fy ∗ f ∗
y )(t)dy,
for t ∈ R.
Lemma 3.2. ([2], Lemma 3.2) For all ν1 ∈ O,
bg(ν1) =ZWν1
kπν(f )k2
HS P f (ν) dν′,
wherebg is the one dimensional Fourier transform of g.
We now present the proof of the theorem.
Proof of Theorem 3.1. It follows from the definition of g that
Since f ∈ Cc(G) it follows that g ∈ Cc(R). In particular, g ∈ L1(R). We now have the following lemma.
ξ ∈ R,
bfy(ξ)2dy,
fy(ξ)2dy dξ = kf k2
2.
and consequentlybg ≥ 0 with
bg(ξ) =ZRd−1
ZRbg(ξ) =ZRd
ZRbg(ν1) eψ(ν1) dν1 = ZOZWν1
= ZW
kπν(f )k2
HS eψ(ν1) P f (ν) dν′ dν1
kπν(f )k2
HS eψ(ν1) P f (ν) dν.
Therefore the function f is identically zero on G if and only if g is zero on R. We now complete the proof
by showing that g vanishes identically. In order to do this we shall apply Theorem 2.3 to the function g.
By Lemma 3.2, it follows that
By the hypothesis (3.5) the right-hand side integral is finite. Therefore by Theorem 2.3 we conclude that
g is identically zero on R and hence f vanishes identically.
10
MITHUN BHOWMIK
Conversely, if ψ is non decreasing and the integral I is finite then by Theorem 1.1, (b), for a given δ
positive there exists g ∈ Cc(R) with supp g ⊂ [−δ, δ] satisfying the estimate
(3.6)
y ∈ R.
Let Z = exp(RX1) denote a central subgroup of G and we identify Z with R. Let h ∈ Cc(G) with
supp h ⊂ V where V is a symmetric neighbourhood of the identity element in G. We now define a
function f on G by
bg(y) ≤ Ce−ψ(y),
f (x) =ZZ
g(t) h(t−1x) dt, x ∈ G.
Clearly, f is continuous. Moreover, f ∈ Cc(G), which follows from the fact that
as t ∈ Z. It can be shown ([8], P. 490) that
kt−1xk ≥ kxk − ktk
(3.7)
kπν (f )k2
HS,
for all ν ∈ W.
Therefore, from (3.6) and (3.7) it follows that
kπν(f )k2
HS e2ψ(ν1) P f (ν) dν
HS = bg(ν1)2 kπν(h)k2
ZW
= ZW
≤ CZW
bg(ν1)2 kπν(h)k2
kπν(h)k2
= C khkL2(G) < ∞.
HS P f (ν) dν
HS e2ψ(ν1) P f (ν) dν
This completes the proof.
(cid:3)
If the function ψ is assumed to be nondecreasing then we have the following corollary.
Corollary 3.3. Let ψ : [0, ∞) → [0, ∞) be a nondecreasing function and
I =Z ∞
1
ψ(t)
t2 dt.
Suppose f ∈ Cc(G) is such that
ZW
kπν(f )k2
HS e2ψ(kνk) P f (ν) dν < ∞.
If the integral I is infinite then f vanishes identically on G.
Proof. Since ψ is nondecreasing ψ(kνk) ≥ ψ(ν1). Therefore, the result follows from Theorem 3.1.
(cid:3)
Acknowledgement. We would like to thank Swagato K. Ray and Suparna Sen for the many useful
discussions during the course of this work.
IMPROVED INGHAM-TYPE RESULT ON RD AND ON NILPOTENT LIE GROUPS
11
References
[1] Baklouti, A; Ben Salah, Nour . On theorems of Beurling and Cowling-Price for certain nilpotent Lie groups, Bull. Sci.
Math. 132 (2008), no. 6, 529-550. MR2445579 (2009f:22007)
[2] Baklouti, A; Thangavelu, S. Variants of Miyachi's theorem for nilpotent Lie groups, J. Aust. Math. Soc. 88 (2010),
no. 1, 1-17. MR2770922 (2012a:22012)
[3] Bhowmik, M.; Ray, S. 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.
[4] Bochner, S. Quasi-analytic functions, Laplace operator, positive kernels Ann. of Math. (2) 51 (1950), 68-91. MR0032708
(11,334g)
[5] Corwin,L. J.; Greenleaf, F. P. Representations of nilpotent Lie groups and their applications Cambridge University
Press, Cambridge, 1990. MR1070979 (92b:22007)
[6] 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)
[7] Ingham, A. E. A Note on Fourier Transforms J. London Math. Soc. 9 (1934) no. 1 , 29-32. MR1574706
[8] Kaniuth, E.; Kumar, A. Hardy's theorem for simply connected nilpotent Lie groups Math. Proc. Cambridge Philos.
Soc. 131 (2001), no. 3, 487-494. MR1866390 (2002j:22007)
[9] Levinson, N. Gap and Density Theorems American Mathematical Society Colloquium Publications, v. 26. American
Mathematical Society, New York, 1940. MR0003208 (2,180d)
[10] Paley, R. E. A. C.; Wiener, N. Fourier transforms in the complex domain (Reprint of the 1934 original) American
Mathematical Society Colloquium Publications, 19. American Mathematical Society, Providence, RI, 1987. MR1451142
(98a:01023)
[11] 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, 348355. MR1501688
Stat-Math Unit, Indian Statistical Institute, 203 B. T. Road, Kolkata - 700108, India.
E-mail address: [email protected]
|
1607.02161 | 1 | 1607 | 2016-07-07T20:21:28 | Unconditionally $p$-converging operators and Dunford-Pettis Property of order $p$ | [
"math.FA"
] | In the present paper we study unconditionally $p$-converging operators and Dunford-Pettis property of order $p$. New characterizations of unconditionally $p$-converging operators and Dunford-Pettis property of order $p$ are established. Six quantities are defined to measure how far an operator is from being unconditionally $p$-converging. We prove quantitative versions of relationships of completely continuous operators,unconditionally $p$-converging operators and unconditionally converging operators. We further investigate possible quantifications of the Dunford-Pettis property of order $p$. | math.FA | math |
UNCONDITIONALLY p-CONVERGING OPERATORS AND
DUNFORD-PETTIS PROPERTY OF ORDER p
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
Abstract. In the present paper we study unconditionally p-converging operators
and Dunford-Pettis property of order p. New characterizations of unconditionally
p-converging operators and Dunford-Pettis property of order p are established. Six
quantities are defined to measure how far an operator is from being unconditionally
p-converging. We prove quantitative versions of relationships of completely continu-
ous operators,unconditionally p-converging operators and unconditionally converg-
ing operators. We further investigate possible quantifications of the Dunford-Pettis
property of order p.
1. Introduction and notations
Throughout the paper, p∗ denotes the conjugate number of p for 1 ≤ p < ∞; if
p = 1, lp∗ plays the role of c0. X, Y will denote real (or complex) Banach spaces and
L(X, Y ) the space of all the operators (=continuous linear maps) between X and Y .
K(X, Y ) denotes the space of all the compact operators between X and Y . Let X
be a Banach space, 1 ≤ p < ∞ and we denote lp(X) by the set of all p-summable
p (X) be the
p (X) is a Banach space with
sequences in X with the natural norm k(xn)nkp = (P∞n=1 kxnkp)
set of all weakly p-summable sequences in X. Then lw
the norm
p . Let lw
1
k(xn)nkw
p = sup{(
∞X
n=1
< x∗, xn > p)
1
p : x∗ ∈ BX ∗},
∀ (xn)n ∈ lw
p (X).
It is a well-known result of A. Grothendieck ([15],[12,Proposition 2.2])that the canon-
ical correspondence T 7→ (T en)n provides an isometric isomorphism of L(lp∗, X) onto
lw
p (X). A sequence (xn)n ∈ lw
p (X) is unconditionally p-summable if
sup{(
∞X
n=m
< x∗, xn > p)
1
p : x∗ ∈ BX ∗} → 0 as m → ∞.
Date: August 27, 2018.
Dongyang Chen's project was supported by the Natural Science Foundation of Fujian Province
of China(No. 2015J01026). Lei Li is the corresponding author and his research is partly supported
by the NSF of China (11301285).
1
2
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
We denote the set of all unconditionally p-summable sequences on X by lu
p (X). It
is obvious that (xn)n is unconditionally 1-summable if and only if (xn)n is uncondi-
tionally summable. J. H. Fourie and J. Swart proved that the same correspondence
T 7→ (T en)n provides an isometric isomorphism of K(lp∗, X) onto lu
p (X) (see [14]). Let
us recall that an operator T : X → Y is unconditionally converging if T takes weakly
1-summable sequences to unconditionally 1-summable sequences. For p = ∞, the
space lu
∞(X) is identical to c0(X), the space of all norm null sequences in X. Hence-
forth, for p = ∞, we refer to consider the space cw
0 (X) of weakly null sequences in X,
instead of lw
∞(X) = l∞(X). Recall that an operator T : X → Y is completely contin-
uous if T takes weakly null sequences to norm null sequences. It is well-known that
p-summing operators are precisely those operators which take weakly p-summable se-
quences(unconditionally p-summable sequences) to p-summable sequences. A natural
question arises: what are operators which take weakly p-summable sequences to un-
conditionally p-summable sequences? This is the starting point of our investigation.
The paper is organized as follows:
In Section 2, we introduce the concept of unconditionally p-converging operators(1 ≤
p ≤ ∞), which is the extension of unconditionally converging operators and com-
pletely continuous operators. It is proved that unconditionally p-converging opera-
tors coincide with the p-converging operators introduced by J. M. F. Castillo and
F. S´anchez in [7] although their original definitions are different. New concepts of
weakly p-Cauchy sequences and weakly p-limited sets are introduced to characterize
unconditionally p-converging operators. We establish characterizations of weakly p-
limited sets and investigate connections between weakly p-limited sets and relatively
norm compact sets. A counterexample is constructed to show that an operator is
unconditionally p-converging not precisely when its second adjoint is.
Section 3 is concerned with Dunford-Pettis property of order p (DP Pp for short)
introduced in [7], which is a generalization of the classical Dunford-Pettis property.
It turns out that many classical spaces failing Dunford-Pettis property enjoy DP Pp,
such as Hardy space H 1 and Lorentz function spaces Λ(W, 1). In this section, we use
weakly p-Cauchy sequences and weakly p-limited sets to characterize DP Pp. New
characterizations of DP Pp in dual spaces are obtained. We also introduce the notion
of hereditary Dunford-Pettis property of order p and establish its characterizations.
In particular, we prove that a Banach space X has the hereditary DP Pp if and only if
every weakly p-summable sequence in X admits a weakly 1-summable subsequence.
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p3
Finally, the surjective Dunford-Pettis property of order p, a formally weaker property
than DP Pp, is introduced and its characterizations are obtained.
In the last two sections of the present paper we investigate possibilities of quantify-
ing unconditionally p-converging operators and the Dunford-Pettis property of order
p. This is inspired by a large number of recent results on quantitative versions of
various theorems and properties of Banach spaces (see [1,3,13,17,18,19]). Section 4
contains quantitative versions of the implications among three classes of operators-
completely continuous,unconditionally p-converging and unconditionally converging
ones. M. Kacena, O. F. K. Kalenda and J. Spurn´y have already defined a quan-
tity measuring how far an operator is from being completely continuous in [17]. In
this section, we define another equivalent quantity measuring complete continuity
of an operator. We further define six quantities measuring how far an operator is
from being unconditionally p-converging. Moreover, we show that one of the six new
quantities is equal to the quantity defined in [20] to measure how far an operator is
unconditionally converging in case of p = 1.
In Section 5 we introduce a new locally convex topology and give two topolog-
ical characterizations of Dunford-Pettis property of order p. Using the introduced
quantity measuring unconditional p-convergence of an operator and the new locally
convex topology, we show that the Dunford-Pettis property of order p is automati-
cally quantitative in a sense. We also define two quantities measuring how far a set
is weakly p-limited. One of the two new quantities is used to quantify the Dunford-
Pettis property of order p. The other is used to define a stronger quantitative version
of Dunford-Pettis property of order p. Several characterizations of this quantitative
version of Dunford-Pettis property of order p are established.
The reader is referred to [12] and [22] for any unexplained notation or terminology.
2. Unconditionally p-converging operators
Definition 2.1. Let 1 ≤ p ≤ ∞. We say that an operator T : X → Y is uncondi-
tionally p-converging if T takes a weakly p-summable sequence (xn)n ∈ lw
p (X)((xn)n ∈
cw
0 (X) for p = ∞) to an unconditionally p-summable sequence (T xn)n ∈ lu
p (Y )((xn)n ∈
c0(Y ) for p = ∞).
We begin with a simple, but extremely useful, characterization of unconditionally
p-converging operators.
4
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
Theorem 2.1. Let 1 ≤ p < ∞. The following are equivalent for an operator T :
X → Y :
(1) T is unconditionally p-converging;
(2) T S is compact for any operator S ∈ L(lp∗, X)(L(c0, X) for p = 1).
Proof. (1) ⇒ (2). Let S ∈ L(lp∗, X)(1 < p < ∞)(L(c0, X) for p = 1). By the
ideal property of unconditionally p-converging operators, T S is unconditionally p-
converging. Since (en)n is weakly p-summable in lp∗(1 < p < ∞)(c0 for p = 1),
(T Sen)n is unconditionally p-summable. Then there exists a compact operator R :
lp∗ → X such that Ren = T Sen(n = 1, 2, ...). Thus T S is compact.
(2) ⇒ (1). Let (xn)n ∈ lw
p (X). Then there exists an operator S : lp∗ → X(1 < p <
∞)(S : c0 → X for p = 1) such that Sen = xn(n = 1, 2, ...). By (2), we get (T Sen)n
is unconditionally p-summable. Thus T S is unconditionally p-converging.
(cid:3)
Before another frequently useful characterization of unconditionally p-converging
operators is given, we recall the notion of weakly p-convergent sequences introduced
in [8]. A sequence (xn)n in a Banach space X is said to be weakly p-convergent to
x ∈ X(1 ≤ p ≤ ∞) if the sequence (xn − x)n is weakly p-summable in X. Weakly
∞-convergent sequences are simply the weakly convergent sequences. It is natural to
generalize weakly Cauchy sequences to the general case 1 ≤ p ≤ ∞.
Definition 2.2. Let 1 ≤ p ≤ ∞. We say that a sequence (xn)n in a Banach space X
is weakly p-Cauchy if for each pair of strictly increasing sequences (kn)n and (jn)n of
positive integers, the sequence (xkn − xjn)n is weakly p-summable in X.
Obviously, every weakly p-convergent sequence is weakly p-Cauchy, and the weakly
∞-Cauchy sequences are precisely the weakly Cauchy sequences.
Theorem 2.2. Let 1 ≤ p ≤ ∞. The following statements about an operator T : X →
Y are equivalent:
(1) T is unconditionally p-converging;
(2) T sends weakly p-convergent sequences onto norm convergent sequences;
(3) T sends weakly p-Cauchy sequences onto norm convergent sequences.
Proof. (1) ⇒ (2). Suppose that (xn)n is weakly p-convergent in X. We may assume
that (xn)n is weakly p-summable. Then there exists an operator S : lp∗ → X, 1 < p <
∞(S : c0 → X for p = 1) such that Sen = xn(n = 1, 2, ...). By Theorem 2.1, T S is
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p5
compact and hence (T Sen)n is relatively compact. Consequently, limn→∞ kT Senk =
0.
(2) ⇒ (3). Let (xn)n be a weakly p-Cauchy sequence in X. By (2), for each
pair of strictly increasing sequences (kn)n and (jn)n of positive integers, the sequence
(T xkn − T xjn)n converges to 0 in norm and hence (T xn)n converges in norm.
(3) ⇒ (1). Suppose that T is not unconditionally p-converging. By Theorem
2.1, the operator T S is non-compact for some operator S ∈ L(lp∗, X)(1 < p <
∞)(L(c0, X) for p = 1). Then there exists a weakly null sequence (zn)n in lp∗(1 < p <
∞)(c0 for p = 1) such that kT Sznk > ǫ0 > 0(n = 1, 2, ...). By passing to subsequences,
we may assume that the sequence (zn)n is equivalent to the unit vector basis (en)n in
lp∗. Let R : lp∗ → lp∗ be an isomorphic embedding with Ren = zn(n = 1, 2, ...). Let
xn = SRen. Then (xn)n is weakly p-summable in X and hence weakly p-Cauchy. By
the assumption, (T xn)n converges to 0 in norm, but kT xnk > ǫ0 > 0(n = 1, 2, ...),
which is a contradiction.
(cid:3)
It should be noted that Theorem 2.2(2) is the definition of the so called p-converging
operators defined by J. M. F. Castillo and F. S´anchez in [7]. In this note, we use the
terminology unconditionally p-converging operators instead of p-converging operators.
Recall that a subset K of a Banach space X is relatively weakly p-compact (1 ≤
p < ∞) if K is contained in S(Blp∗ ) for 1 < p < ∞(S(Bc0) for p = 1) for some
operator S from lp∗(c0 for p = 1) into X (see [25]). A subset K of a Banach space X
is said to be relatively weakly p-precompact if every sequence in K admits a weakly
p < ∞. But the converse needs not to be true. Let X = (P∞n=1 ln
p-convergent subsequence (see [6]). Bessaga-Pe lczy´nski Selection Principle yields that
every relatively weakly p-compact set is relatively weakly p-precompact for any 1 <
1 )p∗(1 < p < ∞).
It follows from Bessaga-Pe lczy´nski Selection Principle that BX is relatively weakly p-
precompact. But BX is not relatively weakly p-compact because X is not isomorphic
to a quotient of lp∗. Another counterexample is Lp(1 < p < ∞, p 6= 2). For each
1 < p < ∞, p 6= 2, BLp is relatively weakly r-precompact, where r = max(p∗, 2), but
is not relatively weakly r-compact because such Lp is not isomorphic to a quotient of
lr∗.
By using the weakly p-Cauchy sequences, we can correspondingly define the con-
ditionally weakly p-compact sets as follows:
6
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
Definition 2.3. Let 1 ≤ p ≤ ∞. We say that a subset K of a Banach space X
is conditionally weakly p-compact if every sequence in K admits a weakly p-Cauchy
subsequence.
The following result,which follows from Theorem 2.2, says that unconditionally
p-converging operators are precisely those operators that send conditionally weakly
p-compact subsets onto relatively norm compact subsets.
Theorem 2.3. Let T ∈ L(X, Y ) and 1 ≤ p < ∞. The following statements are
equivalent:
(1) T is unconditionally p-converging;
(2) T maps relatively weakly p-precompact subsets onto relatively norm compact sub-
sets;
(3) T maps conditionally weakly p-compact subsets onto relatively norm compact sub-
sets;
(4) T maps relatively weakly p-compact subsets onto relatively norm compact subsets.
Definition 2.4. Let X be a Banach space and 1 ≤ p < ∞. We say that a bounded
subset K of X∗ is weakly p-limited if limn→∞ supx∗∈K < x∗, xn > = 0 for every
(xn)n ∈ lw
p (X).
The following result, an immediate consequence of Theorem 2.2, is a characteriza-
tion of unconditionally p-converging operators in terms of weakly p-limited subsets.
Theorem 2.4. Let 1 ≤ p < ∞. The following are equivalent for an operator T :
X → Y :
(1) T is unconditionally p-converging;
(2) T ∗ maps bounded subsets of Y ∗ onto weakly p-limited subsets of X∗.
J.M.F.Castillo and F.S´anchez said that a Banach space X ∈ Wp(1 ≤ p < ∞) if
any bounded sequence in X admits a weakly p-convergent subsequence (see [8]). We
use this notion to characterize weakly p-limited sets.
Theorem 2.5. Let 1 < p < ∞ and X be a Banach space. The following statements
are equivalent about a bounded subset K of X∗:
(1) K is weakly p-limited;
(2) For all spaces Y ∈ Wp and for every operator T from Y into X, the subset T ∗(K)
is relatively norm compact;
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p7
(3) For every operator T from lp∗ into X, the subset T ∗(K) is relatively norm compact.
Proof. (1) ⇒ (2). Let T be an operator from Y ∈ Wp into X such that T ∗(K) is
not relatively norm compact. Then there exists a sequence (x∗n)n in K such that
(T ∗x∗n)n admits no norm convergent subsequences. Since Y ∗ is reflexive, by passing
to a subsequence if necessary we may assume that (T ∗x∗n)n converges weakly to some
y∗ ∈ Y ∗ and kT ∗x∗n − y∗k > ǫ0 for some ǫ0 > 0 and for all n ∈ N. For each n,
choose yn with kynk ≤ 1 such that < T ∗x∗n − y∗, yn > > ǫ0. Since Y ∈ Wp, by
passing to a subsequence again if necessary one can assume that the sequence (yn)n is
weakly p-convergent to some y ∈ Y . Thus, by hypothesis, we get limn→∞ supx∗∈K <
x∗, T yn − T y > = 0. Note that, for each n ∈ N,
< T ∗x∗n−y∗, yn > ≤ < x∗n, T yn−T y > + < x∗n, T y > − < y∗, y > + < y∗, y−yn > .
This implies that limn→∞ < T ∗x∗n − y∗, yn >= 0, which is a contradiction.
(2) ⇒ (3) is immediate because lp∗ ∈ Wp;
(3) ⇒ (1). Let (xn)n ∈ lw
p (X). Then there exists an operator T from lp∗ into X
such that T en = xn for all n ∈ N. It follows from (3) that T ∗(K) is relatively norm
compact. By the well-known characterization of relatively norm compact subsets of
lp, one can derive that limn→∞ supx∗∈K < x∗, xn > = 0.
(cid:3)
By Theorem 2.5, we see that relatively norm compact sets are weakly p-limited.
But Theorem 2.4 demonstrates that there are many weakly p-limited sets which are
not relatively norm compact. Indeed, for each 1 < p < ∞ and for each 1 < r < p∗,
the identity map Ir on lr is unconditionally p-converging and hence the unit ball Blr∗
of lr∗ is weakly p-limited. In the following result, we use biorthogonal sequences to
characterize weakly p-limited sets which are not relatively norm compact.
Theorem 2.6. Suppose that X is reflexive and K is a weakly p-limited subset of X∗.
If K is not relatively norm compact, then there exits a seminormalized biorthogonal
sequence (xn, x∗n)n in X × (K − K) such that (x∗n)n is a basic sequence and (xn)n has
no weakly p-Cauchy subsequence.
Proof. Suppose that K is not relatively norm compact, and let (fn)n be a sequence
in K with no norm convergent subsequence. Since X is reflexive, we may assume
that the sequence (fn)n converges weakly. Then there exist two strictly increasing
sequences (kn)n and (jn)n of positive integers and ǫ0 > 0 such that kfkn − fjnk > ǫ0
8
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
for all n ∈ N. Let x∗n = fkn − fjn ∈ (K − K). Then (x∗n)n is weakly null. By Bessaga-
Pe lczy´nski Selection Principle, we can assume that (x∗n)n is a basic sequence. Let
(x∗∗n )n be the associated sequence of coefficient functionals, and for each n ∈ N, let
xn ∈ X be a Hahn-Banach extension of x∗∗n to all of X∗. Then the sequence (xn, x∗n)n
is seminormalized and biorthogonal.
It remains to show that (xn)n has no weakly p-Cauchy subsequence.
If (yn)n is
a weakly p-Cauchy subsequence of (xn)n, then (yn+1 − yn)n is weakly p-summable.
Since K is weakly p-limited, the subset K − K is also weakly p-limited, which implies
that limn→∞ supk < x∗k, yn+1 − yn > = 0. This is impossible because (xn, x∗n)n is
biorthogonal.
(cid:3)
A consequence of Theorem 2.6 is that for any 1 < p < ∞, there exists a relatively
weakly compact sequence that admits no weakly p-Cauchy subsequence. Moreover, it
should be noted that the converse of Theorem 2.6 is true. Actually, it is easy to verify
that if K is a subset of X∗ and the sequence (xn, x∗n)n in X × (K − K) is biorthogonal
with supn kxnk < ∞, then K is not relatively norm compact.
The following result shows that an operator is unconditionally p-converging not
precisely when its second adjoint is.
Theorem 2.7.
(1) Let T ∈ L(X, Y ) and 1 ≤ p ≤ ∞. If T ∗∗ is unconditionally p-converging, then T
is unconditionally p-converging;
(2) For each 1 ≤ p ≤ ∞, there exists an unconditionally p-converging operator T , but
T ∗∗ is not unconditionally p-converging.
Proof. (1). By the ideal property of unconditionally p-converging operators, JY T is
unconditionally p-converging, where JY : Y → Y ∗∗ is the canonical mapping. Let
S ∈ L(lp∗, X)(1 < p < ∞)(L(c0, X) for p = 1). By Theorem 2.1, JY T S is compact
and hence T S is compact. Again by Theorem 2.1, T is unconditionally p-converging.
(2). J. Bourgain and F. Delbaen (see [5]) constructed a Banach space XBD such
that XBD has the Schur property and X∗∗BD is isomorphically universal for separable
Banach spaces. Since XBD has the Schur property, every operator from lp(1 < p < ∞)
and from c0 into XBD is compact. By Theorem 2.1, every operator with domain XBD
is unconditionally p-converging for each 1 ≤ p < ∞.
In particular, the identity
map IXBD on XBD is unconditionally p-converging. But since X∗∗BD is isomorphically
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p9
universal for separable Banach spaces, there exists a closed subspace Xp∗(X0 for p = 1)
of X∗∗BD such that Xp∗ is isomorphic to lp∗ for 1 < p < ∞ (X0 is isomorphic to c0 for
p = 1). This implies that I∗∗XBD = IX ∗∗
is not lp∗-strictly singular for 1 < p < ∞ (c0-
strictly singular for p = 1). Thus I∗∗XBD = IX ∗∗
is not unconditionally p-converging.
For p = ∞, the identity map IXBD is obviously completely continuous, but I∗∗XBD =
IX ∗∗
(cid:3)
is not completely continuous because X∗∗BD has not the Schur property.
BD
BD
BD
3. Dunford-Pettis Property of order p
Let us recall that a Banach space X has the Dunford-Pettis property (in short,
DPP) if for every Banach space Y , every weakly compact operator T : X → Y
is completely continuous (see [16]). An operator T : X → Y is said to be weakly
compact if T BX is relatively weakly compact in Y . J. M. F. Castillo and F. S´anchez
extended the classical Dunford-Pettis property to the general case for 1 ≤ p ≤ ∞ in
[7]. Let 1 ≤ p ≤ ∞. A Banach space X is said to have the Dunford-Pettis property
of p (in short, DP Pp) if for every Banach space Y , every weakly compact operator
T : X → Y is unconditionally p-converging. Many classical spaces failing the DPP
enjoy the DP Pp. A simple observation is that if a Banach space X has cotype q < ∞,
then X has the DP Pp for any 1 < p < q∗. Thus, the classical Hardy space H 1, which
fails the DPP (see [10]), has the DP Pp for any 1 < p < 2.
It is known that all
the Lorentz function spaces Λ(W, 1)'s fail the DPP (see [10]). But there are certain
positive results for DP Pp. For example, if we take W (t) = 1
2√t , t ∈ (0, 1], then the
space Λ(W, 1) has the DP Pp for some 1 < p ≤ 2. Another non-reflexive space failing
the DPP is the interesting space L built in [21]. Indeed, it was shown in [4]that even
duals of L fail the DPP and odd duals of L fail the surjective DPP, which is genuinely
weaker than the DPP. Moreover, F. Bombal, P. Cembranos and J. Mendoza proved
that for any 1 ≤ p < ∞,every operator from L into lp is compact (see [4]). This
means that L∗ has the DP Pp for any 1 < p < ∞. More examples can be found in [7].
Let us start with a characterization of the DP Pp by means of weakly p-limited sets.
Theorem 3.1. Let 1 < p < ∞. A Banach space X has the DP Pp if and only if each
relatively weakly compact subset of X∗ is weakly p-limited.
Proof. The sufficient part follows immediately from Theorem 2.4. On the other hand,
let K be a relatively weakly compact subset of X∗. By the Davis-Figiel-Johnson-
Pe lczy´nski factorization lemma (see [9]), there exists a reflexive space Z, which is a
10
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
linear subspace of X∗, such that the inclusion map J : Z → X∗ is bounded and the
unit ball BZ of Z contains K. Since Z is reflexive, there is an operator T : X → Z∗
such that T ∗ = J. By the assumption, T is unconditionally p-converging. By Theorem
2.4, the set T ∗(BZ) = J(BZ) = BZ is weakly p-limited in X∗. Thus K is also weakly
p-limited.
(cid:3)
Let us remark that for each 1 < p < ∞, there exists a weakly p-limited set which
is not relatively weakly compact. Indeed, we take X = L∗, where the space L is built
in [21]. As mentioned above, the identity IX on X is unconditionally p-converging
for each 1 < p < ∞. It follows from Theorem 2.4 that the unit ball BX ∗ is weakly
p-limited, but it is not weakly compact because the space L is non-reflexive.
The following result is an internal characterization of the DP Pp. It is a refinement
of [7,Proposition 3.2].
Theorem 3.2. Let 1 < p < ∞ and X be a Banach space. The following are equiva-
lent:
(1) X has the DP Pp;
(2) Every weakly compact operator T from X into c0 is unconditionally p-converging;
(3) limn→∞ < x∗n, xn >= 0, for every weakly p-Cauchy sequence (xn)n in X and every
weakly null sequence (x∗n)n in X∗;
(4) limn→∞ < x∗n, xn >= 0, for every (xn)n ∈ lw
(x∗n)n in X∗;
(5) limn→∞ < x∗n, xn >= 0, for every (xn)n ∈ lw
(x∗n)n in X∗.
p (X) and every weakly Cauchy sequence
p (X) and every weakly null sequence
Proof. (1) ⇒ (2) is trivial.
(2) ⇒ (3). Given a weakly p-Cauchy sequence (xn)n
in X and a weakly null sequence (x∗n)n in X∗. Define an operator T : X → c0 by
T x = (< x∗n, x >)n. Since (x∗n)n converges to 0 weakly, T ∗ is weakly compact and so
is T . By (2), T is unconditionally p-converging. By Theorem 2.2, (T xn)n converges
to some ξ = (ξk)k ∈ c0 in norm. Let ǫ > 0. There exists a positive integer N1 such
that kT xn − ξk < ǫ
2 for all n > N1. Choose another positive integer N2 such that
ξk < ǫ
2 for all k > N2. By the definition of T , we have < x∗n, xn > < ǫ for all
n > max(N1, N2). Thus limn→∞ < x∗n, xn >= 0.
(3) ⇒ (4) is trivial.
(4) ⇒ (5).
If (xn)n is weakly p-summable in X and (x∗n)n is weakly Cauchy in
X∗, yet (< x∗n, xn >)n does not converge to 0. By passing to subsequences, we may
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p11
assume that < x∗n, xn > > ǫ0 for some ǫ0 > 0 and all n ∈ N. Since (xn)n is weakly
p-summable and in particular weakly null, there exists a subsequence (xkn)n of (xn)n
such that < x∗n, xkn > < ǫ0
2 for all n ∈ N. Since (x∗n)n is weakly Cauchy, we see
that (x∗kn − x∗n)n is weakly null. By (3), limn→∞ < x∗kn − x∗n, xkn >= 0. This implies
that < x∗kn − x∗n, xkn > < ǫ0
3 for n large enough. But for such n's, we have
ǫ0 < < x∗kn, xkn > ≤ < x∗kn − x∗n, xkn > + < x∗n, xkn > <
5ǫ0
6
.
(5) ⇒ (1). Let T : X → Y be a weakly compact operator. Let us suppose that
T is not unconditionally p-converging. Appealing again to Theorem 2.2, we obtain
a weakly p-summable sequence (xn)n in X and ǫ0 > 0 such that kT xnk > ǫ0(n =
1, 2, ...). Pick y∗n ∈ Y ∗ such that < y∗n, T xn >= kT xnk and ky∗nk = 1 for all n ∈ N.
Since T is weakly compact, so is T ∗. Hence there is a subsequence (y∗kn)n of (y∗n)n
such that the sequence (T ∗y∗kn)n converges weakly and hence is weakly Cauchy. The
assumption ensures that the sequence (< T ∗y∗kn, xkn >)n = (kT xknk)n converges to 0,
which is a contradiction.
(cid:3)
Corollary 3.3. Let 1 < p < ∞. If X∗∗ has the DP Pp, then so is X.
The converse of Corollary 3.3 is not true. In fact, the Banach space X = (Pn ln
2 )c0
enjoys the DPP, but X∗∗ = (Pn ln
2 )l∞ contains a complemented copy of l2. Since l2
fails the DP Pp for any 2 ≤ p < ∞, X∗∗ also fails the DP Pp for any 2 ≤ p < ∞. In
the case of the classical DPP, there is a result better than Corollary 3.3: If X∗ has
the DPP, then X has the DPP too (see [10]). The analogous result is not true for
the DP Pp: for each 1 < p < ∞, every operator from lp into Tsirelson's space T is
compact, hence T has the DP Pp for any 1 < p < ∞. But, for each 1 < p < ∞, there
is a non-compact operator from lp into T ∗. Thus, for each 1 < p < ∞, T ∗ fails the
DP Pp.
Corollary 3.4. Suppose that a Banach space X contains no copy of l1 and let 1 <
p < ∞. The following statements are equivalent:
(1) X∗ has the DP Pp;
(2) For all Banach spaces Y , every weakly compact operator T : Y → X has the
unconditionally p-converging adjoint;
(3) limn→∞ < x∗n, xn >= 0, for every (x∗n)n ∈ lw
quence (xn)n in X;
p (X∗) and every weakly Cauchy se-
12
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
(4) limn→∞ < x∗n, xn >= 0, for every weakly p-Cauchy sequence (x∗n)n in X∗ and every
weakly null sequence (xn)n in X;
(5) limn→∞ < x∗n, xn >= 0, for every (x∗n)n ∈ lw
(xn)n in X.
p (X∗) and every weakly null sequence
Proof. We only prove (2) ⇒ (3) and (5) ⇒ (1).
(2) ⇒ (3). Assuming the contrary, we can find (x∗n)n ∈ lw
p (X∗) and a weakly
Cauchy sequence (xn)n in X such that < x∗n, xn > > ǫ0 for some ǫ0 > 0 and all
n ∈ N. Since (x∗n)n is weakly null, there exists a subsequence (x∗kn)n of (x∗n)n such
that < x∗kn, xn > < ǫ0
2 for all n ∈ N.
Define an operator S : X∗ → c0 by
2 for all n ∈ N. Thus < x∗kn, xn − xkn > > ǫ0
Sx∗ = (< x∗, xn − xkn >)n,
x∗ ∈ X∗.
It is easy to check that S∗en = xn − xkn(n = 1, 2, ...), where (en)n is the unit vector
basis of l1. Thus the operator S∗ maps l1 into X and is weakly compact. By (2), the
operator S∗∗ is unconditionally p-converging. Moreover, an easy verification shows
that S∗∗ = S. By Theorem 2.2, we get limn→∞ kSx∗knk = 0.
It follows from the
definition of the operator S that limn→∞ < x∗kn, xn − xkn > = 0, which is a
contradiction.
(5) ⇒ (1). By Theorem 3.2, it is enough to verify that for every (x∗n)n ∈ lw
p (X∗)
and every weakly null sequence (x∗∗n )n in X∗∗, the sequence (< x∗∗n , x∗n >)n converges
to 0. Now we suppose that it is false. Then, by passing to subsequences, we may
assume that < x∗∗n , x∗n > > ǫ0 for some ǫ0 > 0 and all n ∈ N. Of course, we
may also assume that kx∗∗n k ≤ 1 for all n ∈ N. It follows from Goldstine's Theorem
that for each n ∈ N, there exists an xn ∈ BX such that < xn − x∗∗n , x∗n > < ǫ0
2 .
Then < x∗n, xn > > ǫ0
2 for all n ∈ N. By Rosenthal's Theorem, (xn)n has a
weakly Cauchy subsequence, which is still denoted by (xn)n. Then there exists a
subsequence (x∗kn)n of (x∗n)n such that < x∗kn, xn > < ǫ0
3 for all n ∈ N. By (5),
we get limn→∞ < x∗kn, xkn − xn >= 0, which implies that < x∗kn, xkn − xn > < ǫ0
for n large enough. It is easy to verify that for such n's, < x∗kn, xkn > < ǫ0
2 . This
contradiction completes the proof.
6
Definition 3.1. Let 1 < p < ∞. We say that a Banach space X has the heredi-
tary Dunford-Pettis property of order p (in short, hereditary DP Pp)if every (closed)
subspace of X has the DP Pp.
(cid:3)
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p13
We present a useful characterization of hereditary DP Pp. We need a J. Elton's
result that can be found in [11].
Lemma 3.5. [11] If (xn)n is a normalized weakly null sequence of a space X such that
no subsequence of it is equivalent to the unit vector basis (en)n of c0, then (xn)n has
a subsequence (yn)n for which given any subsequence (zn)n of (yn)n and any sequence
(αn)n∈c0 we have supn kPn
k=1 αkzkk = +∞.
Theorem 3.6. Let X be Banach space and 1 < p < ∞. The following are equivalent:
(1) X has the hereditary DP Pp;
(2) Every normalized weakly p-summable sequence in X admits a subsequence that is
equivalent to the unit vector basis of c0;
(3) Every weakly p-summable sequence in X admits a weakly 1-summable subsequence;
(4) Every weakly p-summable sequence in X admits a subsequence (yn)n such that
supN kPN
n=1 ynk < ∞.
Proof. (1) ⇒ (2). Let (xn)n be a normalized weakly p-summable sequence in X
such that it admits no subsequence that is equivalent to the unit vector basis (en)n
of c0.
It follows from Lemma 3.5 that (xn)n has a subsequence (yn)n as stated in
Lemma 3.5. By Bessaga-Pe lczy´nski Selection Principle, we may assume that (yn)n is
a basic sequence. Let X0 = span{yn : n = 1, 2, ...}. Let (y∗n)n ⊂ X∗0 be the coefficient
functionals of the basic sequence (yn)n. For each N, define a projection PN : X0 → X0
by
PN (y) =
NX
n=1
< y∗n, y > yn,
y ∈ X0.
Then the projection PN 's are uniformly bounded in operator norm. An easy verifi-
cation shows that P ∗∗N y∗∗ = PN
n=1 < y∗∗, y∗n > yn for all y∗∗ ∈ X∗∗0 . Lemma 3.5 and
the uniform boundedness of the projection PN 's imply that (< y∗∗, y∗n >)n ∈ c0 for all
y∗∗ ∈ X∗∗0 , that is, (y∗n)n is weakly null. Since < y∗n, yn >= 1 for all n ∈ N, it follows
from Theorem 3.2 again that X0 fails the DP Pp.
(2) ⇒ (3) and (3) ⇒ (4) are obvious.
(4) ⇒ (1). Take a subspace X0 of X that fails the DP Pp. Appealing to Theorem
3.2, we obtain a weakly compact operator T : X0 → c0 which is not unconditionally
p-converging. Applying Theorem 2.2, we get a normalized weakly p-summable se-
quence (xn)n in X such that kT xnk ≥ ǫ0 for all n ∈ N. Bessaga-Pe lczy´nski Selection
14
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
Principle allows us to assume that the sequence (T xn)n is equivalent to the unit vec-
tor basis (en)n of c0. By the weak compactness of T , the sequence (xn)n admits no
subsequence equivalent to the unit vector basis (en)n. By Lemma 3.5, the sequence
(xn)n admits a subsequence (yn)n for which given any subsequence (zn)n of (yn)n, one
has supN kPN
n=1 znk = ∞.
A direct consequence of Theorem 3.6 is the following corollary:
(cid:3)
Corollary 3.7. If a Banach space X has the hereditary DP Pp, then each weakly p-
summable sequence in X admits a subsequence (xn)n such that limn→∞ kPn
k=1 xkk/n
1
p∗ =
0.
We close this section with the surjective DP Pp, a formally weaker property than
the DP Pp. By the Davis-Figiel-Johnson-Pe lczy´nski's factorization theorem (see [9]),
a Banach space X has the DP Pp if and only if for all reflexive spaces Y , every
operator from X into Y is unconditionally p-converging. We introduce the surjective
DP Pp by imposing that every surjective operator from X onto the reflexive space Y
is unconditionally p-converging. The motivation for introducing the surjective DP Pp
was to extend the surjective DPP introduced in [21].
Definition 3.2. Let 1 < p < ∞. We say that a Banach space X has the surjec-
tive DP Pp if for all reflexive spaces Y , every surjective operator from X onto Y is
unconditionally p-converging.
The following are the internal characterizations of the surjective DP Pp.
Theorem 3.8. The following are equivalent for a Banach space X and 1 < p < ∞:
(1) X has the surjective DP Pp;
(2) limn→∞ < x∗n, xn >= 0, for every weakly p-Cauchy sequence (xn)n in X and every
weakly null sequence (x∗n)n in X∗ such that span{x∗n : n = 1, 2, ...} is reflexive;
(3) limn→∞ < x∗n, xn >= 0, for every (xn)n ∈ lw
(x∗n)n in X∗ such that span{x∗n : n = 1, 2, ...} is reflexive;
(4) limn→∞ < x∗n, xn >= 0, for every (xn)n ∈ lw
(x∗n)n in X∗ such that span{x∗n : n = 1, 2, ...} is reflexive.
p (X) and every weakly Cauchy sequence
p (X) and every weakly null sequence
Proof. (1) ⇒ (2). Let (xn)n ⊂ X and (x∗n)n ⊂ X∗ be as in (2). Let Z = span{x∗n :
n = 1, 2, ...}. Then (Z⊥)⊥ = Z, where Z⊥ := {x ∈ X :< x∗, x >= 0 for all x∗ ∈ Z}
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p15
and (Z⊥)⊥ := {x∗ ∈ X∗ :< x∗, x >= 0 for all x ∈ Z⊥}. Let Q : X → X/Z⊥ be the
natural quotient. Then Q∗ : (X/Z⊥)∗ → Z is a surjective isometrical isomorphism.
Let Q∗fn = x∗n, fn ∈ (X/Z⊥)∗ for all n ∈ N. By (1), the quotient Q is unconditionally
p-converging. By Theorem 2.2, the sequence (Qxn)n converges in norm to Qx for
some x ∈ X. Thus
< x∗n, xn − x > = < fn, Qxn − Qx > ≤ (sup
n
kfnk)kQxn − Qxk → 0 (n → ∞).
Since (x∗n)n is weakly null,
x∗n, xn >= 0.
limn→∞ < x∗n, x >= 0.Therefore we have limn→∞ <
(2) ⇒ (3) is obvious.
(3) ⇒ (4). Suppose that (4) is false. Then there exist a sequences (xn)n ∈ lw
p (X)
and a weakly Cauchy sequence (x∗n)n in X∗ such that span{x∗n : n = 1, 2, ...} is
reflexive so that < x∗n, xn > > ǫ0 > 0 for all n ∈ N. Since the sequence (xn)n
converges to 0 weakly, there is a subsequence (xkn)n of (xn)n such that < x∗n, xkn >
< ǫ0
for all n ∈ N. Since the space span{x∗n : n = 1, 2, ...} is reflexive, the
space span{x∗n − x∗kn : n = 1, 2, ...} is reflexive too. By the hypothesis, limn→∞ <
x∗n − x∗kn, xkn >= 0. Thus, < x∗n − x∗kn, xkn > < ǫ0
2 for n large enough, which implies
that for such n's, < x∗kn, xkn > < ǫ0, a contradiction.
2
(4) ⇒ (1). Suppose that X fails the surjective DP Pp. Then there exists a surjective
operator T from X onto a reflexive space Y such that T is not unconditionally p-
converging. By Theorem 2.2, there exists a normalized weakly p-summable sequence
(xn)n in X such that kT xnk > ǫ0 for all n ∈ N. For each n, choose y∗n ∈ Y ∗ with
ky∗nk = 1 such that < y∗n, T xn >= kT xnk. By the reflexivity of Y , we may assume that
the sequence (y∗n)n converges to 0 weakly by passing to subsequences if necessary. Let
x∗n = T ∗y∗n. Then the sequence (x∗n)n converges to 0 weakly too. Since T is surjective,
the operator T ∗ : Y ∗ → X∗ is an isomorphic embedding. This implies that the
space span{x∗n : n = 1, 2, ...} is contained in T ∗(span{y∗n : n = 1, 2, ...}) and hence is
reflexive. By (4), limn→∞ < x∗n, xn >= 0, a contradiction because < x∗n, xn >> ǫ0 for
all n ∈ N. This concludes the proof.
(cid:3)
An immediate consequence of Theorem 3.8 is the following:
Corollary 3.9. Let 1 < p < ∞. If X∗∗ has the surjective DP Pp, then so is X.
16
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
We also use the space X = (Pn ln
2 )c0 to show that the converse of Corollary 3.9
is not true. The same argument shows that the space X = (Pn ln
2 )c0 enjoys the
surjective DP Pp for any 1 < p < ∞, but X∗∗ also fails the surjective DP Pp for any
2 ≤ p < ∞.
The following result analogous to Theorem 3 in [4] shows that the surjective DP Pp
and the DP Pp coincide for certain classes of Banach spaces.
Theorem 3.10. If a Banach space X contains a complemented copy of l1, then X
has the DP Pp if and only if X has the surjective DP Pp.
4. Quantifying unconditionally p-converging operators
As discussed above, we see that unconditionally p-converging operators are in-
termediate between completely continuous operators and unconditionally converging
operators. Precisely, we have the following implications:
T completely continuous ⇒ T unconditionally p-converging ⇒ T unconditionally
converging.
In this section, we quantify these implications. We need some necessary quantities.
Let (xn)n be a bounded sequence in a Banach space X. Set
ca((xn)n) = inf
n
sup{kxk − xlk : k, l ≥ n}.
This quantity is a measure of non-Cauchyness of the sequence (xn)n. More precisely,
ca((xn)n) = 0 if and only if (xn)n is norm Cauchy. In [17], an important quantity
measuring how far an operator T : X → Y is from being completely continuous,
denoted as cc(T ), is defined by
cc(T ) = sup{ca((T xn)n) : (xn)n ⊂ BX weakly Cauchy }.
Obviously, T is completely continuous if and only if cc(T ) = 0. In this note, we
define another equivalent quantity measuring the complete continuity of an operator
T : X → Y as follows:
ccn(T ) = sup{lim supn kT xnk : (xn)n ⊂ BX weakly null }.
Obviously, T is completely continuous if and only if ccn(T ) = 0. The following
theorem demonstrates these two quantities are equivalent.
Theorem 4.1. Let T ∈ L(X, Y ). Then ccn(T ) ≤ cc(T ) ≤ 2ccn(T ).
To prove Theorem 4.1, we need the following lemma.
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p17
Lemma 4.2. Let X be a Banach space and (xn)n be a weakly null sequence in BX .
Let ǫ > 0 be such that kxnk > ǫ for all n ∈ N. Then, for every δ > 0, there is a
subsequence (xkn)n of (xn)n such that ca((xkn)n) ≥ ǫ − δ.
Proof. We set xk1 = x1. Choose x∗1 ∈ SX ∗ such that < x∗1, xk1 >= kxk1k. Since (xn)n
is weakly null, there exists k2 > k1 such that < x∗1, xk2 > < δ. Then
kxk1 − xk2k ≥ < x∗1, xk1 − xk2 > ≥ < x∗1, xk1 > − < x∗1, xk2 > ≥ ǫ − δ.
Suppose that we have obtained {xk1, xk2, ..., xkn} such that kxki − xknk ≥ ǫ − δ for
i = 1, 2, ..., n − 1. Let Yn = span{xk1, xk2, ..., xkn}. Pick a c-net {z1, z2, ..., zm} ⊂ SYn
for SYn, where 0 < c < δ
2 . Choose z∗1, z∗2, ..., z∗m in SX ∗ such that < z∗i , zi >= 1
for i = 1, 2, ..., m. Since (xn)n is weakly null, there exists kn+1 > kn such that
< z∗i , xkn+1 > < c for all i = 1, 2, ..., m. Then, for each 1 ≤ j ≤ n, there exists
1 ≤ i ≤ m such that k
− zik < c. Thus
xkj
kxkj k
kxkj − xkn+1k ≥ < z∗i , xkj − xkn+1 >
≥ 1 − < z∗i , xkn+1 > − < z∗i , xkj − zi >
≥ 1 − c − kxkj − zik
≥ 1 − c − (1 + c − ǫ) = ǫ − 2c
≥ ǫ − δ
By induction, we get a subsequence (xkn)n such that kxkn − xkmk ≥ ǫ − δ(n 6=
m, n, m = 1, 2, ...). This yields that ca((xkn)n) ≥ ǫ − δ.
(cid:3)
Proof of Theorem 4.1. Step 1. cc(T ) ≤ 2ccn(T ).
We may suppose that cc(T ) > 0 and fix any c > 0 satisfying cc(T ) > c. Then there
is a weakly Cauchy sequence (xn)n in BX such that ca((T xn)n) > c. It follows that
there exist two strictly increasing sequences (kn)n, (ln)n of positive integers such that
kT xkn − T xlnk > c for all n ∈ N. Set zn = (xkn − xln)/2. Then (zn)n is a weakly null
sequence in BX and kT znk > c/2 for each n ∈ N. Hence lim supn kT znk ≥ c/2 and
then ccn(T ) ≥ c/2. Since c < cc(T ) is arbitrary, we get cc(T ) ≤ 2ccn(T ).
Step 2. ccn(T ) ≤ cc(T ).
We may suppose that kT k = 1 and ccn(T ) > 0. Suppose that ccn(T ) > ǫ > 0.
Then there is a weakly null sequence (xn)n in BX such that lim supn kT xnk > ǫ. This
18
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
yields a subsequence of (xn)n, still denoted by (xn)n, so that kT xnk > ǫ for each
n ∈ N. By Lemma 4.2, for every δ > 0, there is a subsequence (xkn)n of (xn)n such
that ca((T xkn)n) ≥ ǫ − δ. This means that cc(T ) ≥ ǫ − δ. Since δ > 0 is arbitrary,
we get cc(T ) ≥ ǫ. By the arbitrariness of ǫ < ccn(T ), we obtain ccn(T ) ≤ cc(T ). This
completes the proof of Theorem 4.1.
(cid:3)
To quantify unconditionally p-converging operators, we will need two measures of
non-compactness. Let us fix some notations. If A and B are nonempty subsets of a
Banach space X, we set
d(A, B) = inf{ka − bk : a ∈ A, b ∈ B},
bd(A, B) = sup{d(a, B) : a ∈ A}.
Thus, d(A, B) is the ordinary distance between A and B, and bd(A, B) is the non-
symmetrized Hausdorff distance from A to B.
Let A be a bounded subset of a Banach space X. The Hausdorff measure of non-
compactness of A is defined by
χ(A) = inf{bd(A, F ) : F ⊂ X finite},
χ0(A) = inf{bd(A, F ) : F ⊂ A finite}.
Then χ(A) = χ0(A) = 0 if and only if A is relatively norm compact. It is easy to
verify that
(4.1)
χ(A) ≤ χ0(A) ≤ 2χ(A).
Now we define five quantities which measure how far an operator is from being
unconditionally p-converging. Let T ∈ L(X, Y ) and 1 ≤ p < ∞. We set
uc1
p(T ) = sup{lim supn kT xnk : (xn)n ∈ lw
p(T ) = sup{ca((T xn)n) : (xn)n ⊂ BX weakly p-Cauchy },
p (X), (xn)n ⊂ BX},
uc2
uc3
p(T ) = sup{ca((T xn)n) : (xn)n ⊂ BX weakly p-convergent },
uc4
p(T ) = sup{χ0(T L) : L ⊂ BX relatively weakly p-compact },
uc5
p(T ) = sup{χ0(T L) : L ⊂ BX relatively weakly p-precompact }.
Clearly, uc1
p(T ) = 0 if and only if T is un-
conditionally p-converging. It turns out that the above five quantities are equivalent.
p(T ) = uc3
p(T ) = uc4
p(T ) = uc2
p(T ) = uc5
Theorem 4.3. Let T ∈ L(X, Y ) and 1 < p < ∞. Then
uc5
p(T ) ≤ uc3
p(T ) ≤ uc2
p(T ) ≤ 2uc1
p(T ) ≤ 2uc4
p(T ) ≤ 2uc5
p(T ).
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p19
Proof. Step 1. uc5
p(T ) ≤ uc3
p(T ).
We may assume that uc5
p(T ) > 0. Let us fix any 0 < c < uc5
p(T ). Then there exists a
relatively weakly p-precompact subset L ⊂ BX such that χ0(T L) > c. By induction,
we can construct a sequence (xn)n in L such that kT xn − T xmk > c, n 6= m, n, m =
1, 2, ... Since L is relatively weakly p-precompact, the sequence (xn)n admits a weakly
p-convergent subsequence that is still denoted by (xn)n. Thus we get ca((T xn)n) ≥ c,
which yields uc3
p(T ) ≥ c. By the arbitrariness of c, we get uc5
p(T ) ≤ uc3
p(T ).
Step 2. uc2
We assume that uc2
p(T ) ≤ 2uc1
p(T ).
p(T ) > 0 and fix any 0 < c < uc2
p(T ). Then there is a weakly p-
Cauchy sequence (xn)n in BX such that ca((T xn)n) > c. By induction, there exist two
strictly increasing sequences (kn)n, (ln)n of positive integers such that kT xkn −T xlnk >
c for all n ∈ N. Set zn = (xkn − xln)/2. Then (zn)n is a weakly p-summable sequence
in BX and kT znk > c/2 for each n ∈ N. Hence uc1
p(T ) ≥ c/2. Since c is arbitrary, we
get Step 2.
Step 3. uc1
Suppose uc1
p(T ) ≤ uc4
p(T ) > c > 0. Then there exists a weakly p-summable sequence (xn)n in
BX such that kT xnk > c for all n ∈ N. We claim that χ0((T xn)n) ≥ c. If this is false,
p(T ).
we can find a finite subset F of (T xn)n such that bd((T xn)n, F ) < c. Since F is finite,
there exist y ∈ F and a subsequence (T xkn)n of (T xn)n such that kT xkn − yk ≤ c
for each n ∈ N. Since the sequence (T xkn)n is weakly null, we get kyk ≤ c. This
contradiction completes the proof Step 3.
p(T ) ≤ uc2
The remaining inequalities uc3
p(T ) are immediate.
p(T ) ≤ uc5
p(T ), uc4
It should be mentioned that a quantity is defined in [20] to measure how far an
operator is unconditionally converging as follows:
(cid:3)
uc(T ) = sup{ca((
nX
i=1
T xi)n) : (xn)n ∈ lw
1 (X), k(xn)nkw
1 ≤ 1}.
Obviously, uc(T ) = 0 if and only if T is unconditionally converging. Inspired by this
quantity, we define the sixth quantity measuring how far an operator is uncondition-
ally p-converging as follows:
uc6
p(T ) = sup{lim sup
n
kT xnk : (xn)n ∈ lw
p (X), k(xn)nkw
p ≤ 1}.
20
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
It is obvious that uc6
p(T ) = 0 if and only if T is unconditionally p-converging. This
new quantity will be used in next section to prove a quantitative version of the
Dunford-Pettis property of order p.
Theorem 4.4. Let T ∈ L(X, Y ). Then uc6
1(T ) = uc(T ).
Proof. Step 1. uc6
1(T ) ≤ uc(T ).
i=1 T xi)n).
1 ≤ 1. It aims to show lim supn kT xnk ≤ ca((Pn
1 (X) with k(xn)nkw
i=1 T xi)n). Then there exists n ∈ N such that kPk
Let (xn)n ∈ lw
Let c > ca((Pn
i=1 T xi−Pl
i=1 T xik <
c for all k, l ≥ n. In particular, we have kT xkk = kPk
i=1 T xik < c for all
k ≥ n + 1. Thus one can derive that lim supn kT xnk ≤ c. Since c > ca((Pn
i=1 T xi)n)
is arbitrary, we get lim supn kT xnk ≤ ca((Pn
i=1 T xi −Pk−1
i=1 T xi)n).
1(T ).
1 (X) with k(xn)nkw
1 ≤ 1 such that ca((Pn
Step 2. uc(T ) ≤ uc6
We can suppose that uc(T ) > 0 and fix an arbitrary 0 < c < uc(T ). Then there
exists (xn)n ∈ lw
i=1 T xi)n) > c. By induction,
we can find two strictly increasing sequences (kn)n, (ln)n, ln < kn of positive integers
i=ln+1 xi(n = 1, 2, ...). It is
1 ≤ 1 such that kT znk > c for
1(T ) ≥ c and the proof of Step
such that kPkn
easy to see that (zn)n belongs to lw
1 (X) with k(zn)nkw
all n ∈ N, which yields lim supn kT znk ≥ c. Hence uc6
2 is completed.
i=ln+1 T xik > c for all n ∈ N. Let zn = Pkn
(cid:3)
Combining Theorem 4.1 with Theorem 4.4, we get the promised quantitative ver-
sions of the above implications.
Theorem 4.5. Let T ∈ L(X, Y ) and 1 ≤ p < ∞. Then uc(T ) ≤ uc6
p(T ) ≤ cc(T ).
5. Quantifying Dunford-Pettis property of order p
Let X be a Banach space and let F be the family of all weakly compact subsets of
BX ∗. For F ∈ F , define a semi-norm qF on X∗∗ by
qF (x∗∗) = sup
x∗∈F
< x∗∗, x∗ > ,
x∗∗ ∈ X∗∗.
The locally convex topology generated by the family of semi-norms {qF : F ∈ F }
is called the Mackey topology, denoted by τ (X∗∗, X∗). The restriction to X of the
Mackey topology τ (X∗∗, X∗) is called the Right topology in [23]. This topology is
denoted by ρX or simply ρ when X is obvious.
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p21
In this section, we introduce a new locally convex topology. Let X be a Banach
space and let 1 ≤ p < ∞. Let Fp be the family of all relatively weakly p-compact
subsets of X. For F ∈ Fp, we define a semi-norm qF on X∗ by
qF (x∗) = sup
x∈F
< x∗, x > ,
x∗ ∈ X∗.
The locally convex topology generated by the family of semi-norms {qF : F ∈ Fp}
is denoted by ρ∗p when X is obvious. Applying Grothendieck's Completeness Theo-
rem([24, p.148]), we obtain that the space (X∗, ρ∗p) is complete. Hence, a bounded
subset A of X∗ is relatively ρ∗p-compact if and only if A is totally bounded, equiva-
lently, the set AF = {x∗F : x∗ ∈ A} is totally bounded in l∞(F ) for each relatively
weakly p-compact subset F ⊂ BX . So, if we set
χp
m(A) = sup{χ0(AF ) : F ∈ Fp, F ⊂ BX},
then A is relatively ρ∗p-compact if and only if χp
m(A) = 0. The following result,
which is immediate from [19, Lemma 4.4], implies that an operator T : X → Y is
unconditionally p-converging if and only if T ∗BY ∗ is relatively ρ∗p-compact.
Theorem 5.1. Let T ∈ L(X, Y ) and 1 ≤ p < ∞. Then 1
2uc4
p(T ).
2 uc4
p(T ) ≤ χp
m(T ∗BY ∗) ≤
Let (x∗n)n be a bounded sequence in X∗. We set
caFp((x∗n)n) =
sup
F∈Fp,F⊂BX
inf
n
sup{qF (x∗k − x∗l ) : k, l ≥ n},
and
ecaFp((x∗n)n) = inf{caFp((x∗kn)n) : (x∗kn)n is a subsequence of (x∗n)n}.
The quantity caFp measures how far the sequence (x∗n)n is from being ρ∗p-Cauchy. In
particular, caFp((x∗n)n) = 0 if and only if the sequence (x∗n)n is ρ∗p-Cauchy.
The following result contains two topological characterizations of DP Pp.
Theorem 5.2. The following are equivalent about a Banach space X and 1 < p < ∞:
(1) X has the DP Pp;
(2) Every weakly p-summable sequence in X is ρ-null;
(3) Every weakly convergent sequence in X∗ is ρ∗p-convergent.
Proof. The equivalence of (1) and (2) is essentially Theorem 3.1. The implication
(3) ⇒ (1) follows from Theorem 3.2. It remains to prove (1) ⇒ (3).
22
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
Let (x∗n)n be weakly null in X∗. Define an operator T : X → c0 by
T x = (< x∗n, x >)n,
x ∈ X.
Since (x∗n)n is weakly null, T is weakly compact. By (1), we get T is unconditionally
p-converging. Let F ∈ Fp. It follows from Theorem 2.3 that T F is relatively norm
compact in c0. By the well-known characterization of relatively norm compact subsets
of c0, we get
lim
n→∞
qF (x∗n) = lim
n→∞
which implies that (x∗n)n is ρ∗p-null.
< x∗n, x > = 0,
sup
x∈F
(cid:3)
To quantify the DP Pp, we will need several measures of weak non-compactness.
Let A be a bounded subset of a Banach space X. The de Blasi measure of weak
non-compactness of A is defined by
ω(A) = inf{bd(A, K) : ∅ 6= K ⊂ X is weakly compact }.
Then ω(A) = 0 if and only if A is relatively weakly compact. It is easy to verify
that
ω(A) = inf{ǫ > 0 : there exists a weakly compact subset K of X such that
A ⊂ K + ǫBX}.
Other commonly used quantities measuring weak non-compactness are:
wkX(A) = bd(A
w∗
, X), where A
w∗
denotes the weak∗ closure of A in X∗∗.
wckX(A) = sup{d(clustX ∗∗((xn)n), X) : (xn)n is a sequence in A}, where
clustX ∗∗((xn)n) is the set of all weak∗ cluster points in X∗∗ of (xn)n.
γ(A) = sup{ limn limm < x∗m, xn > − limm limn < x∗m, xn > : (xn)n is a sequence in
A, (x∗m)m is a sequence in BX ∗ and all the involved limits exist}.
It follows from [1,Theorem 2.3] that for any bounded subset A of a Banach space X
we have
wckX(A) ≤ wkX(A) ≤ γ(A) ≤ 2wckX(A),
wkX(A) ≤ ω(A).
For an operator T , ω(T ), wkY (T ), wckY (T ), γ(T ) denote ω(T BX), wkY (T BX), wckY (T BX)
and γ(T BX),respectively. C. Angosto and B. Cascales([1])proved the following in-
equality:
γ(T ) ≤ γ(T ∗) ≤ 2γ(T ), for any operator T .
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p23
So,putting these inequalities together, we get,for any operator T ,
(5.1)
1
2
wkY (T ) ≤ wkX ∗(T ∗) ≤ 4wkY (T ).
Let X be a Banach space and A be a bounded subset of X∗. For 1 ≤ p < ∞, we
set
ιp(A) = sup{lim sup
n
ηp(A) = sup{lim sup
n
sup
x∗∈A
sup
x∗∈A
< x∗, xn > : (xn)n ∈ lw
p (X), (xn)n ⊂ BX},
< x∗, xn > : (xn)n ∈ lw
p (X), k(xn)nkw
p ≤ 1}.
These two quantities measure how far A is weakly p-limited. Obviously, ηp(A) =
in
ιp(A) = 0 if and only if A is weakly p-limited. The following theorem says,
particular, that weakly p-limited sets coincide with relatively ρ∗p-compact sets.
Its
proof is similar to [19, Lemma 5.6].
Theorem 5.3. Let X be a Banach space, 1 ≤ p < ∞ and A be a bounded subset of
X∗. Then
1
8
χp
m(A) ≤ ιp(A) ≤ χp
m(A).
In the following theorem, we quantify the DP Pp by using the quantities ω(·), ιp(·),
ecaFp(·) and χp
m(·).
Theorem 5.4. Let X be a Banach space and 1 < p < ∞. The following are equiva-
lent:
(1) X has the DP Pp;
(2) uc1
(3) ιp(A) ≤ ω(A) for every bounded subset A of X∗;
(4) ecaFp((x∗n)n) ≤ 2ω((x∗n)n) whenever (x∗n)n is a bounded sequence in X∗;
p(T ) ≤ ω(T ∗) for every operator T from X into any Banach space Y ;
m(A) ≤ 2ω(A) for every bounded subset A of X∗.
(5) χp
Proof. (2) ⇒ (1) is obvious. (3) ⇒ (1) and (5) ⇒ (1) follow from Theorem 3.1.
(1) ⇒ (2). Let Y be a Banach space and let T ∈ L(X, Y ). Let ǫ > 0 be such that
T ∗BY ∗ ⊂ K + ǫBX ∗, K ⊂ X∗ is weakly compact. Let (xn)n ∈ lw
p (X) and (xn)n ⊂ BX .
Since X has the DP Pp, it follows from Theorem 3.1 that limn→∞ supx∗∈K < x∗, xn >
= 0. Let c > 0. Then there exists a positive integer N such that supx∗∈K < x∗, xn >
< c for each n ≥ N. For each n ∈ N, pick y∗n ∈ BY ∗ with kT xnk =< y∗n, T xn >.
Since T ∗BY ∗ ⊂ K + ǫBX ∗, then, for each n ∈ N, there exists x∗n ∈ K such that
24
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
kT ∗y∗n − x∗nk ≤ ǫ. Then, for n ≥ N, we get
kT xnk =< T ∗y∗n, xn >
≤ ǫ + < x∗n, xn >
≤ ǫ + sup
x∗∈K
< x∗, xn >
≤ ǫ + c.
This yields lim supn kT xnk ≤ ǫ+c. Since c > 0 is arbitrary, we obtain lim supn kT xnk ≤
ǫ and hence uc1
p(T ) ≤ ǫ. This proves uc1
p(T ) ≤ ω(T ∗).
(1) ⇒ (3). Let (xn)n be a weakly p-summable sequence in BX . Let ǫ > 0 be such
that A ⊂ K +ǫBX ∗, K ⊂ X∗ is weakly compact. For each x∗ ∈ A, there exists z∗ ∈ K
such that kx∗ − z∗k ≤ ǫ. This yields
< x∗, xn > ≤ ǫ + sup
x∗∈K
< x∗, xn >
(n = 1, 2, ...).
Since X has the DP Pp, it follows from Theorem 3.1 that limn→∞ supx∗∈K < x∗, xn >
= 0. Thus we get lim supn supx∗∈A < x∗, xn > ≤ ǫ, which completes the proof
(1) ⇒ (3).
(1) ⇒ (4). Let (x∗n)n be a bounded sequence in X∗. Let ǫ > 0 be such that
(x∗n)n ⊂ K + ǫBX ∗, K ⊂ X∗ is weakly compact. For each x∗n, there exists z∗n ∈ K such
that kx∗n − z∗nk ≤ ǫ. Since K is weakly compact, there exists a weakly convergent
subsequence (z∗kn)n of (z∗n)n. By Theorem 5.2, we see that the sequence (z∗kn)n is
ρ∗p-convergent and hence caFp((z∗kn)n) = 0. Note that for any F ∈ Fp, F ⊂ BX, we
have
qF (x∗ki − x∗kj ) ≤ qF (x∗ki − z∗ki) + qF (z∗ki − z∗kj ) + qF (z∗kj − x∗kj )
≤ 2ǫ + qF (z∗ki − z∗kj ), i, j = 1, 2, ...
This yields
caFp((x∗kn)n) ≤ 2ǫ + caFp((z∗kn)n) = 2ǫ.
(4) ⇒ (1). Let (xn)n ∈ lw
Hence, we get ecaFp((x∗n)n) ≤ 2ǫ and then ecaFp((x∗n)n) ≤ 2ω((x∗n)n).
p (X) and let (x∗n)n be weakly null in X∗. By (4), we
get ecaFp((x∗n)n) = 0. A classical diagonal argument yields a subsequence (x∗kn)n of
(x∗n)n which is ρ∗p-Cauchy. By the completeness of the topology ρ∗p, we see that the
subsequence (x∗kn)n is ρ∗p-convergent. Since (x∗n)n is weakly null, (x∗kn)n is ρ∗p-null.
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p25
Since (xn)n is weakly p-summable, one has
< x∗kn, xkn > ≤ sup
i
< x∗kn, xi > → 0 (n → ∞).
Then Theorem 3.2 gives (1).
(1) ⇒ (5). Let c > ω(A). Then there exists a weakly compact subset K of X∗
such that bd(A, K) < c. Since X has the DP Pp, it follows from Theorem 3.1 that
χp
m(K) = 0. Let ǫ > 0 and L ∈ Fp, L ⊂ BX. Then there exists a finite subset
F ⊂ K such that bd(KL, F L) < ǫ, so χ(AL) ≤ c + ǫ. Since ǫ > 0 is arbitrary, we
get χ(AL) ≤ c. By (4.1), we get χ0(AL) ≤ 2c. This implies that χp
m(A) ≤ 2c, which
completes the proof.
(cid:3)
The following quantitative version obviously strengthens the Dunford-Pettis prop-
erty of order p.
p(T ) ≤ C · wkX ∗(T ∗) for every operator T from X into
Theorem 5.5. Let X be a Banach space and 1 < p < ∞. The following are equiva-
lent:
(1) There is C > 0 such that uc6
any Banach space Y ;
(2) There is C > 0 such that uc6
l∞;
(3) There is C > 0 such that ηp(A) ≤ C · wkX ∗(A) for each bounded subset A of X∗;
(4) There is C > 0 such that uc6
p(T ) ≤ C · wkY (T ) for every operator T from X into
any Banach space Y ;
(5) There is C > 0 such that uc6
l∞.
p(T ) ≤ C · wkX ∗(T ∗) for every operator T from X into
p(T ) ≤ C · wkl∞(T ) for every operator T from X into
Proof. The implication (1) ⇒ (2) is trivial with the same constant.
(2) ⇒ (3). Assume that there is C > 0 such that uc6
p(T ) ≤ C · wkX ∗(T ∗) for
every operator T from X into l∞. We'll show that (3) holds with the constant 32C.
Let A be a bounded subset of X∗. We may assume that ηp(A) > 0. Let us fix any
0 < ǫ < ηp(A). By the definition of ηp(A), there exist a sequence (x∗n)n in A and a
p ≤ 1 such that < x∗n, xn > > ǫ for each
sequence (xn)n in lw
n ∈ N. Let us define an operator S : l1 → X∗ by
p (X) with k(xn)nkw
S((αn)n) = X
αnx∗n,
n
(αn)n ∈ l1.
26
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
As in the proof of Theorem 5.4 in [17], the set S(Bl1) is contained in the closed
absolutely convex hull of (x∗n)n and so wkX ∗(S) ≤ 2wkX ∗((x∗n)n). Let T = S∗JX :
X → l∞. By (2) and (5.1), we get uc6
p(T ) ≤ C · wkX ∗(T ∗). Thus
ǫ ≤ lim sup
< x∗n, xn > ≤ lim sup
n
kT xnk
n
≤ uc6
p(T ) ≤ C · wkX ∗(T ∗)
≤ 4C · wkl∞(T ) ≤ 4C · wkl∞(S∗)
≤ 16C · wkX ∗(S) ≤ 32C · wkX ∗((x∗n)n)
≤ 32C · wkX ∗(A)
Since ǫ < ηp(A) is arbitrary, we get the assertion (3).
(3) ⇒ (1). Let us suppose that (3) holds with a constant C > 0. Let T ∈ L(X, Y ).
p ≤ 1. For each n ∈ N, pick y∗n ∈ BY ∗ so that
Let (xn)n ∈ lw
kT xnk =< y∗n, T xn >. Applying (3) to A = (T ∗y∗n)n, we get
p (X) with k(xn)nkw
lim sup
kT xnk = lim sup
n
n
≤ lim sup
n
sup
x∗∈A
≤ C · wkX ∗(A) ≤ C · wkX ∗(T ∗),
< T ∗y∗n, xn >
< x∗, xn > ≤ ηp(A)
which yields uc6
p(T ) ≤ C · wkX ∗(T ∗).
Finally, the equivalences of (1) ⇔ (4) and (2) ⇔ (5) follow from estimate (5.1).
(cid:3)
It should be mentioned that the assertion (3) of Theorem 5.5 is a quantitative
version of Theorem 3.1.
Definition 5.1. We say that a Banach space X has the quantitative Dunford-Pettis
property of order p if X satisfies the equivalent conditions of Theorem 5.5.
The following Theorem 5.7 is a quantitative version of Corollary 3.3. To prove it,
we need a simple lemma.
Lemma 5.6. Let X be a closed subspace of a Banach space Y and let A be a bounded
subset of X. Then
(5.2)
wkY (A) ≤ wkX(A) ≤ 2wkY (A).
UNCONDITIONALLY p-CONVERGING OPERATORS AND DUNFORD-PETTIS PROPERTY OF ORDER p27
Proof. We can identify X∗∗ with X⊥⊥ ⊂ Y ∗∗. Under this identification, the weak∗
closure of A in X∗∗ is equal to the weak∗ closure of A in Y ∗∗. This yields the
left inequality immediately. To prove the right inequality of (5.2), let us fix any
. Then there exists y ∈ Y such that ky∗∗ − yk ≤ c.
c > wkY (A). Take any y∗∗ ∈ A
Choose y∗ ∈ X⊥ with ky∗k = 1 so that d(y, X) = < y∗, y > . Then we get
w∗
d(y∗∗, X) ≤ ky∗∗ − yk + d(y, X) ≤ c + < y∗, y > = c + < y∗, y∗∗ − y > ≤ 2c.
Thus wkX(A) ≤ 2c. By the arbitrariness of c > wkY (A), we obtain wkX(A) ≤
2wkY (A).
(cid:3)
Theorem 5.7. If X∗∗ has the quantitative Dunford-Pettis property of order p, then
so is X. More precisely,
(a) If X∗∗ satisfies one of the conditions (1),(2),(4) and (5) of Theorem 5.5 with a
given constant C, then X satisfies the respective condition of Theorem 5.5 with 16C;
(b) If X∗∗ satisfies the condition (3) of Theorem 5.5 with a given constant C, then X
satisfies the respective condition (3) of Theorem 5.5 with C.
Proof. The assertion (a) follows immediately from the inequality (5.1) and the easy
p(T ∗∗) for each operator T . The assertion (b) is a direct conse-
fact that uc6
quence of (5.2).
(cid:3)
p(T ) ≤ uc6
Acknowledgements. This work is done during the first author's visit to Depart-
ment of Mathematics, Texas A&M University. We would like to thank Professor W.
B. Johnson for helpful discussions and comments.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
C. Angosto and B. Cascales, Measures of weak non-compactness in Bananch spaces, Topol-
ogy Appl.156(2009), 1412-1421.
F. Albiac and N. J. Kalton, Topics in Banach space theory, Springer, 2005.
H. Bendov´a, O. F. K. Kalenda and J. Spurn´y, Quantification of the Banach-Saks property,
J.Funct. Anal.268(2015), 1733-1754.
F. Bombal, P. Cembranos and J. Mendoza, On the surjective Dunford-Pettis property,
Math.Z.204
(1990), 373-380.
J. Bourgain and F. Delbaen, A class of special L∞ spaces, Acta Math.145(1980),no.3 -- 4,
155 -- 176.
J. M. F. Castillo and F. S´anchez, Remarks on some basic properties of Tsirelson's space,
Note di Matematica.XIII(1993), 117-122.
J. M. F. Castillo and F. S´anchez, Dunford-Pettis-like properties of continuous vector func-
tion spaces, Revista Matematica.6(1993), 43-59.
28
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
DONGYANG CHEN, J. ALEJANDRO CH ´AVEZ-DOM´INGUEZ, AND LEI LI
results
related to the Dunford-Pettis property, Con-
J. M. F. Castillo and F. S´anchez, Weakly p-compact, p-Banach-Saks and super-reflexive
Banach spaces, J.Math.Anal.Appl.185(1994), 256-261.
W.J. Davis, T. Figiel, W. B. Johnson, and A. Pe lczy´nski, Factoring weakly compact opera-
tors, J. Funct. Anal. 7(1974), 311-327.
J. Diestel, A survey of
temp.Math.2(Amer.Math.
Soc.,Providence,RI,1980), 15-60.
J. Diestel, Sequences and Series in Banach Spaces, Springer, New York, 1984.
J. Diestel, H. Jarchow and A. Tonge, Absolutely summing operators, Cambridge Studies in
Adv. Math., Vol.43, Cambridge Univ. Press, Cambridge, 1995.
M. Fabian, P. H´ajek, V. M. Santaluc´ıa and V. Zizler, A quantitative version of Krein's
theorem, Rev.Mat.Iberoam.21(2005), 237-248.
J. H. Fourie and J. Swart, Banach ideals of p-compact operators, Manuscripta
Math.26(1979), 349-362.
A. Grothendieck, Sur certaines classes des suites dans les espaces de Banach, et le th´eor´eme
de Dvoretzky-Rogers, Bol.Soc.Mat.Sao Paulo.8(1956), 81-110.
A. Grothendieck, Sur les applications lin´eaires faiblement compactes d'espaces du type
C(K), Canad J.Math.5(1953), 129-173.(French)
M. Kacena, O. F. K. Kalenda and J. Spurn´y, Quantitative Dunford-Pettis property,
Adv.Math. 234 (2013), 488-527.
O. F. K. Kalenda, H. Pfitzner and J. Spurn´y, On quantifications of weak sequential com-
pleteness, J.Funct. Anal. 260(2011), 2986-2996.
O. F. K. Kalenda and J. Spurn´y, Quantifications of the reciprocal Dunford-Pettis property,
Studia Math.210(2012), 261-278.
H. Krulisov´a, Quantification of Pe lczy´nski's property (V), to appear.
D. Leung, Uniform convergence of operators and Grothendieck spaces with the Dunford-
Pettis property, Math.Z.197(1988), 21-32.
J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I, Sequence Spaces, Springer,
Berlin, 1977.
A. M. Peralta, I. Villanueva, J. D. M. Wright and K. Ylinen, Topological characterisation
of weakly compact operators, J.Math.Anal.Appl.325(2007), 968-974.
H. H. Schafer, Topological vector spaces, New York: Springer 1971.
D. P. Sinha and A. K. Karn, Compact operators whose adjoints factor through subspaces of
lp, Studia Math.150(2002), 17-53.
School of Mathematical Sciences, Xiamen University, Xiamen,361005,China
E-mail address: [email protected]
Department of Mathematics, University of Oklahoma, Norman, OK 73019-3103,USA
E-mail address: [email protected]
School of Mathematical Sciences and LPMC, Nankai University, Tianjin, 300071,
China
E-mail address: [email protected]
|
1511.08929 | 1 | 1511 | 2015-11-28T21:02:14 | On ergodic operator means in Banach spaces | [
"math.FA"
] | We consider a large class of operator means and prove that a number of ergodic theorems, as well as growth estimates known for particular cases, continue to hold in the general context under fairly mild regularity conditions. The methods developed in the paper not only yield a new approach based on a general point of view, but also lead to results that are new, even in the context of the classical Cesaro means. | math.FA | math |
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
ALEXANDRU ALEMAN AND LAURIAN SUCIU
Abstract. We consider a large class of operator means and prove that a number of ergodic
theorems, as well as growth estimates known for particular cases, continue to hold in the
general context under fairly mild regularity conditions. The methods developed in the paper
not only yield a new approach based on a general point of view, but also lead to results that
are new, even in the context of the classical Ces`aro means.
1. Introduction
Let X and Y be two Banach spaces, and B(X , Y) be the Banach space of all bounded linear
operators from X into Y. Consider B(X ) = B(X , X ) as a Banach algebra and I ∈ B(X ) the
identity operator on X . The range and the kernel of T ∈ B(X ) will be denoted by R(T ) and
N (T ), respectively. Also, σ(T ) and σp(T ) stand for the spectrum and the point spectrum of
T , and we denote by X ∗ and T ∗ the dual space of X and the adjoint operator of T .
The present paper is concerned with sequences {Tn} in the closed convex hull of the powers
of a fixed operator T ∈ B(X ) of the form
(1.1)
tnjT j,
Tn =Xj≥0
tnj ≥ 0 , Xj≥0
tnj = 1,
and if tnj > 0 for infinitely many values of j, the above sum converges in norm.
We shall denote from now on by κ(T ) the set of such, possibly infinite, convex combinations
of powers of T and we refer to sequences in κ(T ). Actually, we formally apply an infinite
matrix to the operator sequence {T n}.
We let
(1.2)
and
(1.3)
X0 = {x ∈ X : kTnxk → 0, n → ∞},
Xc = {x ∈ X : {Tnx} converges in norm}.
This general type of operator mean has already been considered in the literature, for
example in [C], [Eb], [Kr], and more recently, in [R] and [Sc], but only under the additional
assumption that T is power bounded, that is {T n} is bounded in norm.
2010 Mathematics Subject Classification. Primary 47A10, 47A35; Secondary 47B20.
Key words and phrases. Cesaro mean, supercyclic operator, Kreiss bounded operator.
1
2
A. ALEMAN AND L. SUCIU
Let us notice that in [C] such sequences {Tn} are defined with respect to a regular matrix
{tnj} which implies, in particular, that limn→∞ tnj = 0 for j = 0, 1, ....
The basic idea behind the results presented in this paper is that a great deal of estimates,
or ergodic theorems for specific operator means, continue to hold in a fairly general context
than the power boundedness, and they can be deduced from either:
1) A regularity condition for the sequence {Tn} (which does not require the regularity of
the matrix {tnj}),
or,
2) Conditions invariant to rotations.
The regularity condition considered here is that for some fixed, nonnegative integer n0, we
have T Tn − Tn+n0 → 0 either strongly, or in the operator norm. It is used in order to exhibit
useful isometries related to {Tn} that are acting on quotient spaces of X. This is a method
that appears also in the works of [K´e] (concerning the generalized Banach limits), and is a
powerful tool in the study of asymptotic behavior (see for instance [Al], [V]).
By a rotational invariant condition we mean a growth restriction obeyed not only by the
sequence {Tn}, but also by all ”rotated” sequences Tnλ obtained by applying the same convex
combinations to the operator λT . More precisely, for Tn as in (1.1) we set
(1.4)
tnjλjT j,
λ = 1.
Tnλ =Xj≥0
This type of condition brings complex analysis in the picture, a powerful tool in the study of
bounded linear operators.
The paper is organized as follows. Section 2 serves to an introductory purpose. We
present some general technical tools, including properties of X0, as well as a list of examples
of sequences in κ(T ), satisfying the regularity conditions. Moreover, we introduce and discuss
the backward iterate sequence {T (−1)
} of a sequence {Tn} in κ(T ), which is inspired by the
well known relations between Ces´aro means of order p ≥ 1 and the means of order p + 1. The
n
backward iterate is determined by the identity
T (−1)
n
(T − I) =
Xj≥1
jtnj
−1
(Tn − I) ,
and plays an important role in our considerations.
In Section 3 we investigate sequences in κ(T ) for supercyclic operators T and we refer to
the Ansari-Bourdon theorem [AB], which asserts that for a power bounded and supercyclic
operator T , the powers must converge strongly to zero. The result may fail for general
sequences in κ(T ), because of the presence of eigenvalues on the unit circle of the adjoint
T ∗. However, we are able to give a fairly complete picture for the subspaces X0 and Xc
defined above (Theorem 3.3), under the regularity condition. Our general method based
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
3
on the intertwining of T with an isometry gives even in the power boundedness context,
an improvement of Ansari-Bourdon’s result, namely the fact that if T is supercyclic and
power bounded, then every sequence in κ(T ) which satisfies the regularity condition strongly
converges to zero. Concrete examples of supercyclic operators which actually are Ces´aro
bounded but not power bounded, are deferred to Section 5.
In particular, this answers a
question raised by J. Bonet in a recent public communication concerning the existence of a
Ces`aro bounded hypercyclic operator.
Section 4 is devoted to conditions invariant to rotations. Our first main theorem (Theorem
4.1) is a far-reaching generalization of a result of O. Nevanlinna [N]. More precisely, for
operators T ∈ B(X ) with peripheral spectrum σ(T ) ∩ T of arclength measure zero we derive
an estimate for the powers T n, under the assumption that for some sequence {Tn} in κ(T ),
both sequences {Tn} and {T (−1)
} satisfy uniform rotational invariant growth restrictions. To
avoid technicalities we shall only mention some applications of this theorem. For example,
n
when applied to (discretized) Abel means the result yields the following Nevanlinna-type
theorem: If u : (1, ∞) → (0, ∞) is decreasing, then such a behavior of the resolvent
implies
k(T − λI)−1k ≤
u(λ)
λ − 1
,
λ > 1 ,
kT nk = o(cid:18)nu(cid:18) n
n − 1(cid:19)(cid:19) , n → ∞ .
The result applies to other means as well, in particular, to other functions of T . We pay
special attention to the Ces`aro means. We prove that certain rotational invariant growth
conditions for Ces`aro means of any order p ≥ 2 are actually equivalent to a Kreiss-type
condition (Theorem 4.7). When p = 2, a special case of this result was proved by J. Strikwerda
and B. Wade in [SW]. Other similar considerations on the Kreiss resolvent condition were
obtained by E. Berkson and T. A. Gillespie [BG] for trigonometrically well-bounded operators,
and by O. El-Fallah and T. Ransford [FR] for operators satisfying a generalized Kreiss-type
condition with respect to a compact subset of the unit circle.
In the case p = 1, our conditions are equivalent to uniform estimates for the partial sums of
the Taylor expansion at infinity of the resolvent function of T (Theorem 4.7). This is related
to the uniform Kreiss boundedness condition, a concept introduced recently in [MSZ]. For
such conditions a weaker conclusion (in the sense of strong convergence) holds under an
additional technical assumption (Theorem 4.8), and this can be applied to the supercyclic
operators, for instance.
Section 5 concerns examples which require more work. In addition to the construction of
a Ces`aro bounded hypercyclic operators already mentioned above, we have even a Ces`aro
bounded 3-isometry (in the sense of [AS]). This operator has the remarkable property that
for any x 6= 0, the Ces`aro means of T converge weakly, but not strongly, to zero. This is a
4
A. ALEMAN AND L. SUCIU
somewhat surprising fact which is also related to a well-known mean ergodic theorem (see
[Kr]). It shows that even if the regularity condition holds, weak convergence of the means
fails to imply their strong convergence in a fairly dramatic way.
Finally, by modifying an example of A. L. Shields [Sh] we show that the conclusion of the
generalized O. Nevanlinna theorem (Theorem 4.1) may fail if the peripheral spectrum of T
has positive arclength measure on the unit circle.
Let us notice that at this moment we know only an example, which appear also in [Sh], of
a uniformly Kreiss bounded operator on a Hilbert space, which is not power bounded. The
operator in Subsection 5.1 is another such example which is also hypercyclic.
2. Preliminaries and terminology
2.1. General considerations. Given T ∈ B(X ), we begin with two observations about
sequences {Tn} in κ(T ).
Proposition 2.1. Assume that the sequence {Tn} in κ(T ) is bounded on R(T − I)m, for
some nonnegative integer m. Then X0 is closed and
R(T − I)k ,
X0 ⊂ \k≥0
Proof. Clearly, the closure Y of X0 is invariant for T, Tn and if S denotes the restriction of T
to Y, it is easy to see that S −I has dense range. Indeed, if f ∈ R(S −I)⊥, then f (T x) = f (x)
for all x ∈ Y, so that from the T -invariance and the fact that Tn ∈ κ(T ), we deduce that
f (Tnx) = f (x) ,
for all x ∈ Y and all n ∈ N. But then f vanishes on the dense subspace X0, i.e. f = 0. Since
S is bounded on Y it follows that (S − I)k has dense range for all k ∈ N, and we obtain
X0 ⊂ (S − I)kY ⊂ R(T − I)k .
In particular, this gives X0 ⊂ R(T − I)m . and by the uniform boundedness principle we
conclude that X0 = Y.
(cid:3)
A direct consequence of this result is that if X0 = R(T − I)k for some nonnegative integer
k then
(2.1)
R(T − I)p = R(T − I)k ,
p ≥ k .
A simple additional condition that determines X0 is given below.
Corollary 2.2. Assume that there exist nonnegative integers k, m with k ≥ m, such that the
sequence {Tn} in κ(T ) is bounded on R(T − I)m, and that {Tn(T − I)k} converges strongly
to zero. Then
X0 = R(T − I)k .
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
5
Our second observation provides a tool to estimate X0, and it applies to more general
sequences {Tn}. We shall use the following notations.
Given a continuous seminorm γ on X we denote its kernel by N (γ), and consider the
quotient space X/N (γ) as a normed space with
k[x]k = γ(x),
[x] = x + N (γ) .
We denote by Xγ the completion of this space, and by Qγ the quotient map from X into Xγ.
Obviously, Qγ ∈ B(X , Xγ) with
N (Qγ) = N (γ) , R(Qγ) = Xγ .
Proposition 2.3. Assume that {Tn} ⊂ B(X ) is bounded on R(T −I)m, for some nonnegative
integer m, and that each Tn commutes with T ∈ B(X ). Let
(2.2)
γ(x) = lim sup
n→∞
kTn(T − I)mxk,
x ∈ X .
Then there exists a unique operator Vγ ∈ B(Xγ), such that QγT = VγQγ. Moreover, σ(Vγ) ⊂
σ(T ) and if F is analytic in a neighborhood of σ(T ) and satisfies F (Vγ) = 0, then X0 ⊂
R(F (T )(T − I)m).
Proof. The statement is almost self-explanatory. The equality QγT = VγQγ defines Vγ on a
dense subspace, and since N (Qγ) = N (γ) is invariant for T , it follows that Vγ is bounded on
this subspace, hence it has a unique bounded extension to Xγ. In a similar way we obtain
bounded linear operators Vλ, λ ∈ C \ σ(T ) with Qγ(T − λI)−1 = VλQγ, which implies that
Vλ = (Vγ − λI)−1. Finally, for functions F as in the statement we have QγF (T ) = F (Vγ)Qγ,
hence QγF (T ) = 0 whenever F (Vγ) = 0. This also gives the required inclusion, which finishes
the proof.
(cid:3)
Due to the last assertion above, we are of course interested in situations where Vγ is as
simple as possible, in particular, when it has a large functional calculus algebra.
An immediate application for sequences in κ(T ) is given below.
Corollary 2.4. Assume that the sequence {Tn} in κ(T ) is bounded on R(T − I)m, for some
nonnegative integer m, and let γ be the seminorm given by (2.2). Then X0 = R(T − I)k for
some k ≥ m if and only if (Vγ − I)k−m = 0.
Proposition 2.3 allows us to extend to the ergodic context the method developed in [Al],
[K´e], [V] and [SZ] in the study of asymptotic properties of scalar weighted powers of an
operator in B(X ).
6
A. ALEMAN AND L. SUCIU
2.2. A regularity assumption. Given T ∈ B(X ) we shall consider sequences {Tn} in B(X )
which are bounded on R(T − I)m, and satisfy the condition
(2.3)
lim
n→∞
k(T Tn − Tn+n0)xk = 0 ,
x ∈ R(T − I)m .
for some non-negative integers m, n0.
A sequence {Tn} satisfying the condition (2.3) will be shortly called (n0, m)-regular. In
particular, we say that {Tn} is regular when n0 = 1 and m = 0, and {Tn} is ergodic (according
to [Kr]) if it verifies (2.3) with n0 = m = 0. For different values of m, n0, these conditions
of regularity are in general not related. In particular, ergodicity cannot be obtained from
the regularity, as we shall see in the following examples (including the one constructed in
Subsection 5.2).
Some direct consequences of the regularity condition are listed below.
Theorem 2.5. Assume that {Tn} ⊂ B(X ) is (n0, m)-regular. Then
(i) If {Tn} ⊂ κ(T ) is bounded on R(T − I)m, then X0 = R(T − I)k for some k > m if and
only if
lim
n→∞
k−m
Xl=0
(−1)l(cid:18)k − m
l (cid:19)Tn+ln0x = 0 ,
x ∈ R(T − I)m .
Moreover, Xc ∩ R(T − I)m is the direct topological sum
Xc ∩ R(T − I)m = X0 ⊕ (N (T − I) ∩ R(T − I)m) .
(ii) If each Tn commutes with T , {Tn} is bounded on R(T − I)m and γ is the seminorm
given by (2.2), then the operator Vγ from Proposition 2.3 is an isometry.
Proof. (i) A repeated application of (2.3) gives for x ∈ R(T − I)m
k−m
(T − I)k−mTnx −
Xl=0
(−1)l(cid:18)k − m
l (cid:19)Tn+ln0x → 0 , n → ∞ ,
which proves the first part. Also from (2.3) we see that if x ∈ Xc then
lim
n→∞
Tnx ∈ N (T − I) ,
which together with Proposition 2.1 proves that Xc ∩ R(T − I)m is the algebraic direct sum
of X0 and N (T − I) ∩ R(T − I)m. If x ∈ X0, y ∈ N (T − I) ∩ R(T − I)m then we can choose
n such that
Since Tny = y we obtain
kTnxk < kx − yk .
kyk ≤ kTnxk + kTn(y − x)k < ky − xk(1 + sup
n
kTnR(T − I)mk) ,
and (i) follows.
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
7
To see (ii) just apply again (2.3) to conclude that γ(T x) = γ(x), and this assures that Vγ
(cid:3)
is an isometry.
One motivation for the condition (2.3) is a result in [Ku] which asserts that for m = 0, n0 =
0, 1, this condition is necessary for the strong convergence of {Tn} to the ergodic projection
PT of T , that is PT ∈ B(X ) with N (T − I) = R(PT ) and R(T − I) = N (PT ). More precisely,
the result in [Ku] states that Tn → PT strongly if and only if {Tn} satisfies (2.3) with m = 0,
n0 = 0, 1 and
(2.4)
lim
k→∞
sup
n∈N
k
1
k + 1
k
Xj=0
Tn+jx − PT xk = 0,
x ∈ X .
The last condition is called almost convergence and was introduced by G. G. Lorentz in [L].
One can obtain similar characterizations of the strong convergence of {Tn} for m = 0 and
n0 > 1, by using (2.3) and an appropriate almost convergence of step n0 > 1.
2.3. Some examples. A large variety of examples of sequences satisfying (2.3) arise from
matrix summability methods (see [B]), more precisely, by applying such matrices to the
sequence {T n}. The simplest choice of such matrices is the identity which gives the sequence
{T n}.
It satisfies (2.3) with n0 = 1. The so-called Zweier matrix (see [B]) leads to the
sequence {Tn} with
(2.5)
Tn =
1
2
(T n−1 + T n), n ≥ 1
which also satisfies (2.3) with n0 = 1, that is {Tn} is regular. Remark that {Tn} is ergodic if
and only if {T n} converges strongly on R(T 2 − I).
The most common examples considered in this area are the Ces`aro means of order p ∈ N of
0 (T ) = I,
n (T )defined for n ∈ N by : M (p)
an operator T ∈ B(X ), which are the operators M (p)
M (0)
n (T ) = T n, while for n, p ≥ 1 as
(2.6)
M (p)
n (T ) =
=
=
p
(n + 1)...(n + p)
p
(n + 1)...(n + p)
(j + p − 1)!
j!
M (p−1)
j
(T )
(n − j + p − 1)!
(n − j)!
T j
n
n
Xj=0
Xj=0
j
n + k
)T j
p
n + p
n
p−1
Xj=0
(1 −
Yk=1
In the case p = 1 we write
Mn(T ) := M (1)
n (T ) =
1
n + 1
T j
n
Xj=0
and it is usually called the Ces`aro mean of T .
8
A. ALEMAN AND L. SUCIU
If {M (p)
n (T )} is bounded or it strongly converges in B(X ), we say that T is p-Ces`aro
bounded (ergodic), respectively T is Ces`aro bounded (ergodic) when p = 1.
These operator means have a considerably richer algebraic structure. They satisfy the
following recurrent identities valid for n, p ≥ 1:
(2.7)
(2.8)
(2.9)
M (p)
n (T )(T − I) =
p
n + 1
(M (p−1)
n+1 (T ) − I) ,
T M (p)
n (T ) =
n + p + 1
n + 1
M (p)
n+1(T ) −
p
n + 1
I ,
n + p + 1
n + 1
M (p)
n+1(T ) − M (p)
n (T ) =
p
n + 1
M (p−1)
n+1 (T ) .
It is easy to see by the relations (2.7) and (2.8) before that the regularity of {M (p)
means the ergodicity of {M (p+1)
(by using (2.9)), hence each bounded mean {M (p)
n (T )} just
n (T )} is ergodic it is also regular
n (T )} satisfies (2.3) with m = 0 and n0 = 1.
It is interesting to remark that these means satisfy the regularity assumption in the operator
(T )}. In addition, if {M (p)
n
norm when they are bounded on R(T − I)m.
Another example of a sequence {Tn} in κ(T ) is given by the binomial means
(2.10)
Tn = 2−n
n
Xk=0(cid:18)n
k(cid:19)T k = Sn.
where S = 1
2 (T + I). Clearly, R(T − I)m = R(S − I)m for m ≥ 0, and {Tn} satisfies (2.3) for
some integer n0 ≥ 1 if and only if it is ergodic. For example, as it is proved in [DyS], every
contraction on a Hilbert space will satisfy (2.3).
Clearly, one may consider also continuous summability methods, for example, the Abel
means defined as
(2.11)
A(T, r) = (1 − r)Xj≥0
rjT j = (1 − r)(I − rT )−1 ,
0 < r < 1 ,
where T ∈ B(X ) has spectral radius at most one. Using the equality
(I − rT )(I − ρT )−1 = I + (ρ − r)(I − ρT )−1 ,
it is easy to verify that the above sums are bounded if and only if the sequence {Tn} with
Tn = A(T, 1 − 1
n ) is bounded. Moreover, since A(T,r)
1−r = (1 − rT )−1 we have
1
n − 1
Tn(T − I) =
(Tn − I)
and it is again easy to see that
T Tn − Tn+1 =
1
(n + 1)2 (cid:18) n(n + 3)
n − 1
Tn + I(cid:19) .
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
9
Hence the sequence {Tn} is ergodic if and only if it is regular, and this equivalently means
that 1
n Tn → 0 strongly on X . In this case {Tn} also satisfies (2.3) for any positive integer n0
(with m = 0) even in operator norm if it is bounded.
Finally, one can deal with more complicated continuous summability methods as well. If
F is analytic in the unit disc D and has positive Taylor coefficients, then for any operator
T ∈ B(X ) with spectral radius at most 1, and any strictly increasing sequence {rn} tending
to 1, the identity
(2.12)
Tn =
1
F (rn)
F (rnT )
defines a sequence in κ(T ). Of course, the regularity of such sequences is more involved and
depends on the Taylor coefficients of F , as well as on the choice of {rn}.
2.4. The backward iterate. Given a sequence {Tn} in κ(T ) with Tn =Pj≥0 tnjT j, where
(2.13)
jtnj < ∞
tn0 6= 1 , Xj≥1
for all n ∈ N, its backward iterate {T (−1)
n
} is defined by
(2.14)
Then T (−1)
n
(2.15)
T (−1)
n
=Xk≥0
snkT k , where
snk = Pj≥k+1 tnj
Pj≥1 jtnj
.
belongs to κ(T ) as well, and satisfies the identity
T (−1)
n
−1
(T − I) =
Xj≥1
jtnj
(Tn − I) .
It is a simple exercise to verify that if Tn = M (p)
n (T ) then by (2.7), for n > 1, we have
T (−1)
n
= M (p+1)
n−1 (T ).
Another direct computation reveals that for the discretized Abel means introduced before
T (−1)
n
coincides with Tn.
Also, if Tn is a sequence of the form (2.12) then T (−1)
F (rnz) − F (rn)
rnF ′(rn)(z − 1)
Gn(z) =
n
.
(2.16)
= Gn(T ), where
Similarly, for the sequence {Tn} given by (2.5) we have for n ≥ 2,
(2.17)
T (−1)
n
=
=
2
2n − 1 n−2
Xk=0
2
2n − 1
T k +
1
2
T n−1!
(nMn−1(T ) −
T n−1).
1
2
Also, it is easy to verify that T (−1)
n
to zero on R(T 2 − I), and it is regular if and only if 1
case, one can see that if {T (−1)
is ergodic if and only if {Mn(T )} converges strongly
n → 0 strongly on X . In this
In
} is ergodic then it is also regular, but not conversely.
n T (−1)
n
10
A. ALEMAN AND L. SUCIU
n (T )}. In fact, {Mn(T )} is ergodic if and only if {T (−1)
n
n
n
particular, the ergodicity, or regularity of {T (−1)
respectively of {M (2)
n T n(T − I)x → 0 for x ∈ X . On the other hand, {M (2)
1
is regular and 1
of ergodicity or regularity of {T (−1)
bounded, respectively.
} are assured by the ergodicity of {Mn(T )},
} is ergodic and
n (T )} is ergodic if and only if {T (−1)
}
n2 T nx → 0 for x ∈ X . Remark that in these last remarks, the conditions
n (T ) are
} are superfluous, if the means Mn(T ) or M (2)
n
Finally, we note that in certain cases the regularity condition can imply ergodicity. Such
an example is the sequence {Tn} with Tn = Mn(T )2. We omit the details.
3. Operators from linear dynamics
In this section we refer to two classes of operators related to linear dynamics, which have
been considered in [AB], [BM], [BC], [GLM], [He] and [K´e].
Recall that an operator T ∈ B(X ) is called supercyclic if there exists x ∈ X such that
the set {λT nx : λ ∈ C, n ∈ N} is dense in X . In this case, any such vector x is called a
supercyclic vector for T .
According to the terminology of [BC] we say that T ∈ B(X ) is (weakly) hypercyclic with
support N (N ∈ N \ {0}) if there exists x0 ∈ X such that the set
(3.1)
{T k1x0 + T k2x0 + ... + T kN x0 : k1, ..., kN ≥ 0}
is (weakly) dense in X . Clearly, these concepts only make sense when X is separable. The
motivation for considering such conditions is the Ansari-Burdon theorem ([AB], Theorem
2.2) which asserts that any supercyclic power bounded operator T must satisfy the condition
kT nxk = 0 ,
x ∈ X .
lim
n→∞
Let us start with the weakly hypercyclic case and record the following simple observation.
Lemma 3.1. Let Y be a separable Banach space and let V ∈ B(Y) be a contraction. If λV
is weakly hypercyclic with support N , for some λ ∈ C with λ ≤ 1, then Y = {0}.
Proof. If x0 ∈ Y, the set in (3.1) is contained in the ball centered at the origin and of radius
N kx0k (as kV k ≤ 1), hence it cannot be weakly dense in Y, unless Y = {0}.
(cid:3)
Throughout in what follows, for a bounded sequence {Tn} in κ(T ) we let γ the seminorm
from (2.2) with m = 0, and consider the operator Vγ obtained by an application of Proposition
2.3 to this seminorm. As a consequence of Lemma 3.1 we have
Corollary 3.2. Assume that λT is weakly hypercyclic with support N , for some λ ∈ C with
λ ≤ 1. If the sequence {Tn} in κ(T ) is bounded and the operator Vγ is a contraction then
{Tn} converges strongly to zero on X . In particular, this holds when {Tn} satisfies (2.3) with
m = 0 and n0 ≥ 0.
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
11
Proof. The assumption on λT assures for the operators Vγ and Qγ given by (the proof of)
Proposition 2.3 that if x0 is the vector in (3.1), the set
{Qγ(λT )k1x0 + Qγ(λT )k2x0 + ... + Qγ(λT )kN x0 : k1, ..., kN ≥ 0}
= {λVγ)k1Qγx0 + (λVγ)k2Qγx0 + ... + (λVγ)kN Qγx0 : k1, ..., kN ≥ 0}
is weakly dense in Xγ, because Q∗
γ into X ∗. Thus λVγ is hypercyclic with support
N , and by the previous lemma we must have Xγ = {0} which means Tnx → 0 for x ∈ X . (cid:3)
γ maps X ∗
Let us now turn to supercyclic operators T , where the situation is more subtle. This is
due to the fact that for a supercyclic operator T , its adjoint T ∗ may have eigenvalues. A
simple consequence of the definition (see also [BM]) shows that such an eigenvalue is unique,
and the corresponding eigenspace has dimension one. Due to this fact the Ansari-Bourdon
theorem does not extend directly to more general means, in particular Corollary 3.2 does
not necessarily hold for supercyclic operators. In fact, in Subsection 5.1 we shall construct
a Ces`aro (even uniformly Kreiss) bounded and supercyclic operator T such that both T and
T ∗ have the eigenvalue 1, hence {Mn(T )} does not converge strongly to zero. It could also
happens that T ∗ has an eigenvalue, but T does not.
The following theorem describes the possibilities that can occur in the general context,
where the condition (2.3) is essential.
Theorem 3.3. Let T ∈ B(X ) be supercyclic and assume that the sequence {Tn} ⊂ κ(T ) is
bounded and satisfies condition (2.3) for some n0 ≥ 0. If T ∗ has no eigenvalues on the unit
circle then {Tn} converges strongly to zero on X , that is X0 = X .
If T ∗ has the (unique) eigenvalue µ ∈ T then :
(i) {Tny} converges to zero for some y ∈ X \ R(T − µI) if and only if X0 = X . In this case
µ 6= 1.
(ii) {Tny} converges to a nonzero limit for some y ∈ X \ R(T − µI) if and only if X0 6= Xc =
X . This is further equivalent to the fact that both T and T ∗ have the eigenvalue 1.
(iii) {Tny} diverges for some (and hence for all) y ∈ X \ R(T − µI) if and only if X0 = Xc 6=
X .
Proof. Let γ be the seminorm given by (2.2) with m = 0. From the fact that T is supercyclic
it follows immediately that the isometry Vγ on Xγ given by condition (2.3) and Theorem
2.5 is supercyclic, as well. Thus, by the Ansari-Bourdon theorem mentioned before, we have
dim Xγ ≤ 1. Clearly, Xγ = {0} means X0 = X and this certainly occurs when T ∗ has not
eigenvalues on T, as we see below.
Suppose that dim Xγ = 1 . Then there exists µ ∈ T such that
Vγ[x] = µ[x] ,
[x] = x + N (γ),
x ∈ X ,
12
A. ALEMAN AND L. SUCIU
and since N (γ) = X0, we obtain that {Tn(T − µI)} converges strongly to zero, i.e. {Tn} con-
verges strongly to zero on R(T − µI). If T ∗ has no eigenvalues on T we have R(T − µI) = X
and this is a contradiction which shows that X0 = X .
Suppose that T ∗ has the eigenvalue µ ∈ T. As pointed out above, the corresponding
eigenspace has dimension one. Thus for every y ∈ X \ R(T − µI) there exists f ∈ X∗, T ∗f =
µf 6= 0, such that
x − f (x)y ∈ R(T − µI) ,
x ∈ X .
With respect to the direct sum decomposition X = R(T − µI) ⊕ Cy the operator T has the
matrix representation
T = T R(T − µI)
0
(T − µI)Cy
µI2
! ,
where I2 denotes the identity on Cy. A simple computation shows that
T n = T nR(T − µI)
0
(T n − µnI)Cy
µnI2
! ,
so that, if
and we denote
we obtain
tnjT j ,
tnjµj ,
Tn =Xj≥0
τn(µ) =Xj≥0
Tn = TnR(T − µI)
0
(Tn − τn(µ)I)Cy
τn(µ)I2
! .
(i) If we can choose y as above such that Tny → 0, then from
f (Tny) = τn(µ)f (y),
we see that τn(µ) → 0 and we obtain that {Tn} converges strongly to zero on X . The converse
is obvious. Note also that since τn(µ) → 0 and τn(1) = 1, we must have µ 6= 1.
(ii) We show the equivalence of the first and the third assertion. If we can choose y such
that Tny → y0 6= 0, then by (2.3) we have
T y0 = lim
n→∞
TnT y = lim
n→∞
Tn+n0y = y0 .
from the fact that Tn(T − µI)y → 0 we deduce that T y0 = µy0, hence µ = 1 and from above,
1 is an eigenvalue for T . Conversely, if µ = 1 and (T − I)y0 = 0, y0 6= 0, then
i.e. y0 ∈ X \ R(T − µI) and {Tny0} converges to y0. Moreover, this assumption implies
Tny0 = τn(1)y0 = y0 ,
Tn = TnR(T − I)
0
0
I2! → 0
0 I2! ,
0
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
13
in the strong operator topology, that is X0 6= Xc = X . If the last relation holds it is obvious
that we can choose y as in the statement.
(iii) follows immediately from the fact that {Tn} converges strongly to zero on R(T − µI)
and that this space has codimension one.
(cid:3)
If one of the alternatives (i) or (ii) occur then {Tn(T − I)} converges strongly to zero,
that is {Tn} satisfies the regularity assumption (2.3) with m = n0 = 0. Some interesting
examples appear in the study of rotational invariant conditions, like Kreiss, or uniform Kreiss
boundedness, and will be discussed in the last section.
Within this class of operators the situation becomes more transparent, as the following
result shows.
Corollary 3.4. Let T ∈ B(X ) be supercyclic and we assume that the sequence {Tn} ⊂ κ(T )
is bounded and ergodic. If T ∗ has no eigenvalues on T then {Tn} converges strongly to zero
on X that is X0 = X .
If T ∗ has the (unique) eigenvalue µ ∈ T then :
(i) X0 = X if and only if µ 6= 1.
(ii) X0 6= Xc = X if and only if both T and T ∗ have the eigenvalue 1.
(iii) X0 = Xc 6= X if and only if σp(T ∗) \ σp(T ) = {1}.
Proof. In view of Theorem 3.3 we only need to prove that if T ∗ has the unique eigenvalue
µ ∈ T \{1} then X0 = X . This is however obvious, since in this case R(T − I) = X = X0. (cid:3)
One reason why the Ansari-Bourdon theorem holds is a result in [BM] which asserts that if
T ∗ has the eigenvalue µ then T R(T − µI) is hypercyclic, hence T cannot be power bounded.
However, even in the context of power boundedness our method based on the intertwining of
T with Vγ (as in the previous proof) leads to an improvement of the Ansari-Bourdon result.
For a supercyclic power bounded operator T it shows that every sequence in κ(T ) which
satisfies (2.3) must converge strongly to zero.
Despite the problems that occur in the general context considered here, one can prove the
following interesting extension of the Ansari-Bourdon result.
Theorem 3.5. Let T ∈ B(X ) be supercyclic and assume that the sequence {Tn} ⊂ κ(T ) is
bounded and satisfies the regularity assumption (2.3) for some integer n0 ≥ 0. If there exists
a supercyclic vector x0 ∈ X satisfying one of the following conditions
(i)
or
(ii)
kT nx0k = 0 ,
inf
n∈N
kT nx0k < ∞ ,
sup
n∈N
then {Tn} converges strongly to zero on X .
14
A. ALEMAN AND L. SUCIU
Proof. If (i) holds, there is a subsequence {mj} ⊂ N with T mj x0 → 0 as j → ∞.
If Qγ
and Vγ are the operators associated to the seminorm γ given by (2.2) with m = 0, from the
relation V mj
γ Qγx0 = QγT mj x0 → 0 it follows that Qγx0 = 0 (Vγ being an isometry). This
gives Qγ = 0, since x0 is a supercyclic vector for T . Hence X0 = N (Qγ) = X which means
that Tn → 0 strongly on X .
If (ii) holds, we can use (i) to assume, without loss of generality, that α := inf n∈N kT nx0k >
0. Let 0 6= x ∈ X and {λk} ⊂ C \ {0}, {nk} ⊂ N be such that x = limk→∞ λkT nkx0. Then
kxk = lim
k→∞
λkkT nk x0k ≥ α lim sup
k→∞
λk,
therefore the sequence {λk} is bounded, and so it contains a convergent subsequence, let’s
say λkj → λ, j → ∞. Obviously, one has λ 6= 0 because {T nkj x0} is bounded and
limj→∞ λkj T nkj x0 = x 6= 0. Then for m ∈ N we have that
kT mxk = lim
j→∞
λkj kT mT nkj x0k ≤ cλ.
Hence T is power bounded and so T m → 0 strongly. In turn, this gives
kTnxk = lim
j→∞
kλkj TnT nkj x0k ≤ λ sup
m∈N
kTmk lim
j→∞
kT nkj x0k = 0,
and consequently Tnx → 0 for any x ∈ X .
(cid:3)
4. Conditions invariant to rotations
Given a sequence {Tn} in κ(T ) and λ ∈ C, we denote by Tnλ the sequence obtained by
applying the same convex combinations to the operator λT , that is
(4.1)
Tn =Xj≥0
tnjT j =⇒ Tnλ =Xj≥0
tnjλjT j ,
where tnj ≥ 0, Pj≥0 tnj = 1. We are interested in the effect of growth restrictions for Tnλ,
that are uniform in λ ∈ T.
Note that if Tn = T n, then Tnλ = λnTn, λ ∈ T. If Tn = 1
2 (T n−1 + T n), it easy to verify
that {Tnλ} are uniformly bounded on the unit circle, if and only if T is power bounded. The
situation become much more interesting for other sequences. For example, if {Tn} is the
sequence of discretized Abel sums
(4.2)
Tn =
1
n
∞
Xj=0
(1 −
1
n
)jT j ,
then the uniform boundedness of the sequences {Tnλ}, λ ∈ T, is equivalent to the Kreiss
boundedness condition
k(T − λI)−1k ≤
C
λ − 1
,
λ > 1 ,
for some constant C > 0.
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
15
4.1. A generalized Nevanlinna theorem. O. Nevanlinna [N] proved the interesting result
that for operators T ∈ B(X ) satisfying the Kreiss boundedness condition, and with σ(T ) ∩ T
of arclength measure zero, one has
kT nk = o(n), n → ∞ .
The aim of the present subsection is to show that this type of estimate holds in a very
general context and can be derived from rotational invariant conditions. This requires a
different approach which makes use of the backward iterate T (−1)
defined in (2.14).
n
Theorem 4.1. Let {Tn} be a sequence in κ(T ) with Tn = Pj≥0 tnjT j, that satisfies (2.13)
and
jtnj = ∞ .
lim
n→∞Xj≥1
If {wn} is an increasing sequence of positive numbers such that
kTnλk + kT (−1)
nλ k ≤ wn ,
for all λ ∈ T, and σ(T ) ∩ T has arclength measure zero, then for every strictly increasing
sequence of positive integers {mn}, we have
Proof. We obviously have for λ ∈ T,
kT nkPj≥n+1 tmnj
Pj≥1 jtmnj
nλ (λT − I) =
Xj≥1
T (−1)
= o(wmn) , n → ∞ .
−1
jtnj
(Tnλ − I) ,
hence, by assumption we obtain
kT (−1)
nλ R(λT − I)k = o(wn), n → ∞ .
Since R(λT − I) = X a.e. on T with respect to arclength measure, it follows that
(4.3)
kT (−1)
nλ k = o(wn) ,
a.e. on T. With the notation in (2.14) we also have
w−1
mnsmnnT n = w−1
mneite−int dt
T (−1)
,
2π
0
mnZ 2π
mnZ 2π
0
kT (−1)
mneitk
dt
2π
.
in the sense of Bochner integral, so that
(4.4)
w−1
mnsmnnkT nk ≤ w−1
Again by assumption we have that
sup
t∈[0,2π]
mnkT (−1)
w−1
mneitk = O(1), n → ∞ ,
16
A. ALEMAN AND L. SUCIU
and by (4.3) we can apply the dominated convergence theorem in (4.4) to obtain the conclu-
sion.
(cid:3)
Remark 4.2. The reason for considering arbitrary increasing sequences of integers {mn} in
Theorem 4.1 is that for a general sequence {Tn} in κ(T ), the n-th power T n may not appear
in the n-term Tn. Moreover, different choices of {mn} may lead to different estimates of
kT nk. Both aspects occur in the case of Ces`aro means.
A very natural class of sequences {Tn} for which Theorem 4.1 applies directly, are those
generated by power series with positive coefficients, as defined in (2.12), since their backward
iterates are easily determined by (2.16). However, in order to avoid technicalities we shall
only consider the particular case of the discretized Abel sums for which we have T (−1)
= Tn.
As a direct consequence we find the following extended version of Nevanlinna’s theorem.
n
Corollary 4.3. Let u : (1, ∞) → (0, ∞) be a decreasing function. If T ∈ B(X ) satisfies the
condition
k(T − λI)−1k ≤
u(λ)
λ − 1
,
λ > 1 ,
and σ(T ) ∩ T has arclength measure zero, then
kT nk = o(cid:18)nu(cid:18) n
n − 1(cid:19)(cid:19) , n → ∞.
Proof. We have that the Abel means (4.2) are in κ(T ) and satisfy
kTnk ≤ u(cid:18) n
n − 1(cid:19) .
By the observation preceding the corollary, T (−1)
follows by an application of Theorem 4.1.
n
satisfy the same inequality and the result
(cid:3)
With a specific choice of the function u above we obtain :
Corollary 4.4. If T ∈ B(X ) satisfies for some r ≥ 0 the condition
(λ − 1)r+1
λr
sup
λ>1
k(T − λI)−1k < ∞ ,
and σ(T ) ∩ T has arclength measure zero, then
kT nk = o(nr+1) , n → ∞ .
The next corollary provides only spectral conditions (on the resolvent function of T ) in
order for {M (r)
n (T )} to be uniformly convergent to the ergodic projection, for r ≥ 1.
Corollary 4.5. Let T ∈ B(X ). If 1 is a simple pole of the resolvent of T , σ(T ) ∩ T has
arclength measure zero and for an integer r ≥ 1 we have
(λ − 1)r
λr−1 k(T − λI)−1k < ∞,
sup
λ>1
(λ − 1)k(T − λI)−1k < ∞ ,
sup
λ>1
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
17
then
kM (r)
n (T ) − PT k → 0, as n → ∞.
Proof. This is obtained by using Corollary 4.4, Theorem 6 in [Hi], Theorem 1 [Ed] and the
uniform Abel ergodic theorem in [LSS].
(cid:3)
We should point out here that the hypothesis that σ(T ) ∩ T has arclength measure zero
cannot be relaxed, in general (see [N] and Subsection 5.3 below).
4.2. Ces`aro means and resolvent estimates. As mentioned in Section 2, if Tn = M (p)
then T (−1)
n−1(T ). Then Theorem 4.1 applies to Ces`aro means in the following way.
= M p+1
n
n (T ),
Corollary 4.6. Let {wn} be an increasing sequence of positive numbers. If T ∈ B(X ) satisfies
the condition
w−1
n kM (p)
n (λT )k < ∞ ,
sup
n∈N
λ∈T
for some integer p ≥ 1, and σ(T ) ∩ T has arclength measure zero, then
kT nk = o(nwpn) , n → ∞ .
Proof. By (2.6) we have that
sup
n∈N
λ∈T
w−1
n kM (p+1)
n
(λT )k < ∞ ,
hence, by the remarks above, we can apply Theorem 4.1 to Tn = M (p)
From the second equality in (2.6) we have
n (T ) with mn = pn.
smnn = Pj≥n+1 tmnj
Pj≥1 jtmnj
=
p + 1
p (cid:19)p
pn + 1(cid:18) p − 1
αn ,
where αn → 1+. The conclusion follows from Theorem 4.1.
(cid:3)
For the special choice wn = nr, n ≥ 1, with r ≥ 0 fixed, Corollary 4.6 yields the implication
(4.5)
sup
n≥1
λ∈T
n−rkM (p)
n (λT )k < ∞ =⇒ kT nk = o(nr+1), n → ∞ ,
whenever p ≥ 1 and σ(T ) ∩ T has arclength measure zero. It turns out that such ”regular”
growth restrictions for the Ces`aro means can be reformulated in terms of resolvent estimates.
This has been first proved by Strikwerda and Wade [SW] in the case when p = 2 and r = 0.
The Ces`aro means of order one are related to the partial sums of the Taylor expansion at
infinity of the resolvent function of T . The uniform boundedness of these partial sums in the
exterior of the unit disc is called the uniform Kreiss boundedness condition (see [MSZ] and
[GH] for the boundedness only on the real line) and is equivalent to the uniform boundedness
of {Mn(λT )} on the unit circle. These last two conditions are more restrictive. The result
below provides an extension of the theorems we just mentioned.
18
A. ALEMAN AND L. SUCIU
Theorem 4.7. Given T ∈ B(X ) we have:
(i) If p ≥ 2 and r ≥ 0 then
n−rkM (p)
n (λT )k < ∞ ⇐⇒ sup
λ>1
sup
n≥1
λ∈T
(λ − 1)r+1
λr
k(T − λI)−1k < ∞ ,
(ii) If p = 1 and r ≥ 0 then
sup
n≥1
λ∈T
n−rkMn(λT )k < ∞ ⇐⇒ sup
n∈N
λ>1
(λ − 1)r+1
λr
< ∞ .
λ−k−1T k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=0
p (cid:19)M (p)
Xn=0(cid:18)n + p
Proof. (=⇒) The argument uses the ideas in [SW]. We start with the identity proved in [SW],
formula (6.3), for 0 < ρ < 1 and λ ∈ T, namely
∞
(I − ρλT )−1 = (1 − ρ)p
n (λT )ρn .
For p ≥ 2 the resolvent estimate follows directly from
∞
Xn=0(cid:18)n + p
p (cid:19)nrρn ≤ C(1 − ρ)−p−r−1 .
By a straightforward argument we see that the above equality holds for all ρ ∈ D and by
comparing the coefficients of ρk we obtain for p = 1,
n
n−1
ρkλkT k = (1 − ρ)
(k + 1)Mk(λT )ρk + (n + 1)Mn(λT )ρn .
Xk=0
In this case the desired estimate follows from the inequalities
(k + 1)r+1ρk ≤ C(1 − ρ)−r−2 ,
(1 − ρ)r+1ρn ≤ C ′(n + 1)−r−1 ,
Xk=0
n−1
Xn=0
for 0 < ρ < 1.
λ ∈ T,
so that
(n + 1)(n + 2)
2
(⇐=) We begin with (i) and assume first that p = 2. Use (2.6) to write for 0 < ρ < 1,
(n + 1)(n + 2)
2
(cid:13)(cid:13)(cid:13)(cid:13)
M (2)
0
n (λT ) =Z 2π
n (λT )(cid:13)(cid:13)(cid:13)(cid:13)
M (2)
≤
=
(I − ρλeitT )−1
Xj=0
(1 − ρ)r+1 ρ−nZ 2π
C
0
C
(1 − ρ)r+1(1 − ρ2)
(1 − ρ)r+2 ρ−n .
n we obtain the estimate for M (2)
≤
C
n
(n − j + 1)e−ijtρ−j dt
2π
,
1
1 − ρe−it2
dt
2π
ρ−n
If we now choose ρ = 1 − 1
this follows immediately from the first equality in (2.6).
n (T ) in (i) for all r ≥ 0. For p > 2,
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
19
Let us turn to the case when p = 1. For n ≥ 1, µ ∈ T and λ ∈ C with λ > 1 we have by
the Abel summation formula
Mn(µT ) =
=
1
n + 1
n + 1
λn+1
λn+1µ
1
n
Xk=0
Xk=0
n
=
From the assumption
we obtain
kMn(µT )k ≤ C
and if λ = 1 + 1
n ,
n
1
n + 1
n
1
n−1
n−1
n + 1
Xk=0
(µT )k =
λ−k−1(µT )k + (1 − λ)
(λµ)−k−1T k + (1 − λ)
Xk=0
Xk=0
Xk=0
λ−k−1T k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(λ − 1)r+1(n + 1)(cid:18)λn +
(λ − 1)r+1
Xk=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
λr+1
λr
n
sup
n∈N
λ>1
λk+1λ−k−1(µT )k
λk+1
λk+1µ
k
Xj=0
Xj=0
k
λ−j−1(µT )j
(λµ)−j−1T j
.
< ∞ ,
1 − λ(λn − 1)
λ − 1
(cid:19) ,
kMn(µT )k ≤ 2r(2e − 1)Cnr
for n ≥ 1 and µ ∈ T, which concludes the proof.
(cid:3)
In Subsection 5.3 below we construct examples which show that the conditions in part (ii)
are stronger than those in part (i) of the theorem. In fact, by a similar argument as in the
previous proof one can show that if one (any) of the conditions in (i) is satisfied then
(4.6)
kMn(λT )k = O(nr log(n)), n → ∞.
sup
λ∈T
On the other hand, the conditions in (ii) do not imply a similar growth condition for the
powers T n. Indeed, let us consider T = I − V , where V is the classical Volterra operator
(V f )(t) :=Z t
0
f (s)ds,
f ∈ Lp[0, 1],
1 ≤ p ≤ ∞,
p 6= 2.
It is known from [MSZ] that T is uniformly Kreiss bounded and kT nk ∼ n 1
stands for ”comparable”. In particular this gives that { T n
log(n) } is unbounded.
4 − 1
2p , where ∼
4.3. The case p = 1. We want to investigate the growth restrictions
(4.7)
n−rkMn(λT )k < ∞ ,
sup
n≥1
λ∈T
in more detail. The somewhat surprising result we are going to prove below asserts that
under these conditions, the conclusion of Theorem 4.1 in the sense of strong convergence only
holds under an additional assumption. In particular, our theorem provides a partial answer
to Question 3 in [MSZ].
20
A. ALEMAN AND L. SUCIU
Theorem 4.8. Let T ∈ B(X ) satisfying for some integer r ≥ 0 the condition (4.7), and
let Qγ be the operator induced by the seminorm γ given by (2.2) for Tn = M (r+1)
m = r + 1. If R(Qγ) is closed then
(T ) and
n
kT nxk = o(nr+1) , n → ∞
for x ∈ X .
Proof. Note first that a direct application of (2.7) shows that under our hypothesis we have
that {n−r−1T n} is bounded, so one can define the seminorm on X
γ(x) = lim sup
n→∞
n−r−1kT nxk = lim sup
n→∞
kTn(T − I)r+1xk,
x ∈ X ,
n
(T ). By (4.7) we have 1
where Tn = M (r+1)
n Tnx → 0 for x ∈ R(T − I)r, while by (2.8)
this means that {Tn} satisfies (2.3) with n0 = 1 and m = r. By applying Theorem 2.5 (ii)
and Proposition 2.3 to Tn and γ, one obtains an isometry Vγ on the corresponding quotient
space Xγ satisfying VγQγ = QγT , where Qγ is the corresponding quotient map of X into
Xγ = QγX . The following observation reveals a remarkable property of these operators.
Namely, we have
(4.8)
Indeed, if
α = sup
n≥1
n
λ∈T(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=0
then for j ≥ 1 we have
sn(λT ) =
λkT k ,
< ∞ .
λkV k
γ Qγ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=0
n
T jsn(λT ) = λ
j
[(n + j + 1)Mj+n(λT ) − jMj−1(λT )] .
Divide both sides by jr+1 to obtain for x ∈ X and n ≥ 1 fixed
kQγsn(λT )xk = γ(sn(λT )x) ≤ 2 lim sup
m→∞
m−rkMm(λT )kkxk ,
which gives (4.8).
We claim that (4.8) implies Qγ = 0, that is Xγ = {0}, under our assumption. To see this,
assume that R(Qγ) is closed and Qγ 6= 0. Since Vγ is a non zero isometry it must have an
approximate eigenvalue ζ ∈ T. Then there exist sequences {xj} in Xγ, {lj} in X ∗
γ , with
and
kxjkγ = 1 ,
lim
j→∞
(Vγ − ζI)xj = 0 ,
kljk = 1 ,
lj(xj) >
1
2
.
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
21
Now by (4.8) we have for j, n ≥ 1
Z 2π
0
and Parseval’s formula gives
Since for fixed n we have
lim
j→∞
by letting j → ∞ above we obtain
2
dt
2π
≤ α2 ,
lj(eiktV k
γ xj)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
n
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xk=0
n
lj(V k
γ xj)2 ≤ α2 .
Xk=0
kV k
γ xj − ζ kxjk2 = 0 ,
n
Xk=0
lim sup
j→∞
n
Xk=0
lj(xj)2 ≤ α2 ,
and we arrive at the inequality n ≤ 4α2 for all positive integers n, which gives a contradiction.
Thus Qγ = 0 and consequently, Xγ = {0}, that is
kT nxk = o(nr+1) , n → ∞ ,
for all x ∈ X . This ends the proof.
(cid:3)
The examples in the next section show that the exponent of n cannot be improved.
As an application of this theorem, let us remark that if T is supercyclic and it satisfies (4.7)
then Vγ is also supercyclic, hence dim(Xγ) ≤ 1 because Vγ is an isometry. So R(Qγ) = Xγ,
and by Theorem 4.8 we have Xγ = {0}. In particular, if r = 0 in (4.7) then one obtains
kT nxk = o(n) as n → ∞ for x ∈ X . Thus we infer by Corollary 3.4 the following extension
of the Ansari-Bourdon theorem
Corollary 4.9. If T ∈ B(X ) is supercyclic and uniformly Kreiss bounded such that 1 /∈
σp(T ∗) \ σp(T ), then for every λ ∈ T the Ces`aro means Mn(λT ) converge strongly on X .
Notice that in reflexive spaces, the spectral condition in this corollary is superfluous (by
Corollary 3.4), therefore the sequence {Mn(λT )} always converges strongly in this case, the
limit being non zero if and only if 1 ∈ σp(T ) ∩ σp(T ∗). This last spectral condition can be
satisfied, as we shall see in Subsection 5.1. Thus a version of Lorch’s theorem [Kr, Theorem
2.1.2, p. 73] on one hand, and of Ansari-Bourdon on the other hand, is obtained for supercyclic
and uniformly Kreiss bounded operators, in reflexive spaces.
A question arises which is suggested by Theorem 4.8.
Question. For T ∈ B(X ) uniformly Kreiss bounded does k 1
n T nk converge to zero as
n → ∞?
22
A. ALEMAN AND L. SUCIU
5. Examples
5.1. Uniformly Kreiss bounded hypercyclic and supercyclic operators. Consider the
weighted Dirichlet spaces Dα, α > −1 consisting of analytic functions f (z) = Pn≥0 fnznin
the unit disc D with the property that
(5.1)
(n + 1)1−αfn2 < ∞ ,
kf k2 =Xn≥0
Clearly, the monomials en(z) = (n + 1)(1−α)/2zn form an orthonormal basis in this space.
Therefore, the operator Mz of multiplication by the independent variable is a forward weighted
shift of the form
Its adjoint M ∗
z is a hypercyclic backward shift, and for every µ ∈ T the matrix operator
.
Mzen =(cid:18) n + 2
Tµ = µM ∗
n + 1(cid:19)(1−α)/2
µIC!
0
0
z
is supercyclic on Dα ⊕ C, with σp(Tµ) ∩ σp(T ∗
µ ) = {µ} (see [BM, Theorem 1.40]).
The following result is well known. For example, it can be deduced from the results in
z Dα, α > 0, and hence all corresponding operators
[AP]. A direct consequence of it is that M ∗
Tµ are uniformly Kreiss bounded.
Proposition 5.1. If α > 0 then the operator Mz is uniformly Kreiss bounded on Dα.
Proof. A straightforward computation based on Parseval’s formula yields
kf k2 ∼ f (0)2 +ZD
f ′(z)2(1 − z2)αdA(z) ,
f ∈ Dα ,
where A denotes the area measure. Then for λ = 1 we have
Since the functions
Mn(λMz)f (ζ) =
1 − (λζ)n+1
(n + 1)(1 − λζ)
f (ζ) .
Fn(λ, ζ) =
1 − (λζ)n+1
(n + 1)(1 − λζ)
are uniformly bounded in D × D and satisfy
for some C > 0 and all λ, ζ ∈ D and all positive integers n, we obtain
kMn(λMz)f k2 ∼
1 − λζ
C
≤
(cid:12)(cid:12)(cid:12)(cid:12)
∂
∂ζ
Fn(λ, ζ)(cid:12)(cid:12)(cid:12)(cid:12)
(n + 1)2 (cid:18)f (0)2 +ZD
(n + 1)2 f (0)2 +ZD
f (z)2
+ZD
1 − λz2 (1 − z2)αdA .
.
1
1
(Mn(λMz)f )′2(1 − z2)αdA(cid:19)
f (z)′2(1 − z2)αdA
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
Thus, the statement will follow if the operator
is bounded on Dα. As pointed out before, a proof of this fact can be found in [AP].
f 7→Z z
0
f (ζ)dζ
1 − λζ
,
23
(cid:3)
This example shows that one can has in view some non-trivial ergodic properties (as in
Section 3) for supercyclic operators T which are not power bounded, but with some bounded
sequences in κ(T ), as well the means M (p)
n (T ), p ≥ 1.
5.2. A Ces`aro bounded 3-isometry. Given a Hilbert space H and T ∈ B(H), we say that
T is an N -isometry, N ∈ N, (see [AS]), if it satisfies
N
(−1)k(cid:18)N
Xk=0
k(cid:19)kT kxk2 = 0 ,
x ∈ H .
We shall be concerned with the cases when N = 2, 3. Our aim is to construct a Ces`aro
bounded 3-isometry T on the Hilbert space H for which the sequence { 1
n T nx} is bounded
below for all x ∈ H \ {0}. In particular, {Mn(T )x} diverges for each f ∈ H \ {0}. Such an
example is unknown in literature.
In addition, we shall see that for any x the sequence {Mn(T )x} converges weakly to zero.
We start with the following simple method of constructing a 3-isometry using a given
2-isometry. Throughout in what follows we shall denote for a polynomial p(z) =Pn≥0 pnzn
p(z)2dm(z) ,
kpk2
2 =Xn≥0
pn2 =ZT
where m is the normalized arclength measure on the unit circle and
Lp(z) =Xn≥1
pnzn−1 =
p(z) − p(0)
z
.
Proposition 5.2. Let T be a 2-isometry on the Hilbert space H. Fix x0 ∈ H \ {0} and let
H1 be the completion of the space of analytic polynomials with respect to the norm
kpk2
1 = kpk2
2 +Xn≥0
kLnp(T )x0k2 .
Then the operator of multiplication by the independent variable Mzp = zp extends to a
bounded operator on H1 which is a 3-isometry with the property that
kMzpk2
1 − kpk2
1 = kp(T )x0k2 .
Proof. A straightforward computation gives
kMzpk2
1 = kpk2
2 + kT p(T )x0k2 +Xn≥0
kLnp(T )x0k2 ,
hence
kMzpk2
1 ≤ (1 + kT k2)kp(T )x0k2
24
A. ALEMAN AND L. SUCIU
for all polynomials p. Moreover the equality
kMzpk2
1 = kpk2
1 + kT p(T )x0k2
follows directly from above. Since T is a 2-isometry, it satisfies for every polynomial p,
kT 2p(T )x0k2 − 2kT p(T )x0k2 + kp(T )x0k2 = 0 .
This immediately implies that
kM 3
z pk2
1 − 3kM 2
z pk2
1 + 3kMzpk2
1 − kMzpk2
1 = 0 ,
and the result follows.
(cid:3)
We shall work with a specific choice of the 2-isometry T , namely the weighted shift Mz on
the Dirichlet space D0 as defined in the previous subsection. Now let x0(z) = 1 − z, z ∈ D.
We then have the following description of the space H1 considered in Proposition 5.2.
Lemma 5.3. With the above notations we have
kpk2
1 ∼ kpk2
∗ = kpk2
2 +ZT
p′(z)(1 − z)2dm(z) ,
for all analytic polynomials p
Proof. Note that for every polynomial p(z) =Pk pkzk we have
k(1 − Mz)Lnpk2
pk − pk−12(k − n + 1) ,
1 = pn2 + Xk≥n+1
so that
(5.2)
k(1 − Mz)pk2
1 = 2kpk2
pk − pk−12(k − n + 1)
2 +Xn≥0 Xk≥n+1
2 +Xk≥1
∼ kpk2
k2pk − pk−12 .
By Parseval’s formula
kpk2
∗ = kpk2
2 +ZT
p′(z)(1 − z)2dm(z) = p12 +Xk≥2
kpk − (k − 1)pk−12
and the result follows from the elementary inequalities
k2pk − pk−12 − 2kpk − pk−1pk−1 ≤ kpk − (k − 1)pk−12
≤ 2k2pk − pk−12 + 2pk−12.
(cid:3)
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
25
An immediate consequence of the lemma is that H1 is continuously contained in the Hardy
space H 2, that is
kf k2 ≤ Ckf k1 ,
for some constant C > 0 and all f ∈ H1. With the lemma in hand we can prove the following
result.
Theorem 5.4. The operator Mz on H1 is Ces`aro bounded and satisfies for every f ∈ H1
kM n
z f k1 ≥ k(1 − Mz)f k2rn(n − 1)
2
.
Moreover, the sequences {Mn(Mz)f } converges weakly to zero for every f ∈ H1.
Proof. For every polynomial p we have that
Mn(Mz)p(z) =
1 − zn+1
(n + 1)(1 − z)
p(z) ,
and the inequality 1 − zn+1 ≤ (n + 1)1 − z, z ∈ T easily implies that
kMn(Mz)pk2
∗ ≤ kpk2
2 + 2ZT
p′(z)(1 − z)2dm(z) + 2ZT
p(z)2dm(z) .
By Lemma 5.3 and the remark following it, we obtain that MzH1 is Ces`aro bounded. More-
over, it is well known and easy to prove that the norm on D0 satisfies
kM n
z pk2 − kpk2 = nkpk2
2
hence by Proposition 5.2 we have for f ∈ H1
n
kM k
z f k2
1 − kM k−1
z
f k2
1
kM k−1
z
(1 − Mz)f k2
1
(k − 1)k(1 − Mz)f k2
2
kM n
z f k2
1 − kf k2
1 =
=
≥
=
n
Xk=1
Xk=1
Xk=1
n
n(n − 1)
2
k(1 − Mz)f k2
2
and the required inequality is proved.
Now, by using the boundedness of {Mn(Mz)}, the weak convergence to zero of {Mn(Mz)}
in B(H1) one reduces to the convergence to zero of the scalar products hMn(Mz)p, qi for all
analytical polynomials p, q. But this easily follows from the above expression of Mn(Mz)p
and Lemma 5.3. The proof is complete.
(cid:3)
26
A. ALEMAN AND L. SUCIU
5.3. Modified Shields examples. This is a simple modification of the example in [Sh]
showing that the hypothesis that σ(T ) ∩ T has arclength measure zero cannot be removed
form Corollary 4.4.
Let r be a nonnegative integer and let Xr be the Banach space of analytic functions f in
D with the property that f (r+1) belongs to the Hardy space H 1. The norm on Xr is given by
kf kr =
r
Xj=0
f (j)(0) +ZT
f (r+1)dm .
Then it is well known (see for example [Du]) that there exists C > 0 such that
f (j)dm ≤ Ckf kr
and kf k∞ ≤ Ckf kr ,
ZT
for all j ≤ r and all f ∈ Xr.
Proposition 5.5. The operator T = MzXr has the following properties:
(i) For p ≥ 2 we have that the sequences {n−rM p
(ii) {n−rMn(T )} is unbounded,
(iii) {n−r−1T n} does not converge strongly to zero.
n(λT )} are uniformly bounded in λ ∈ T,
Proof. To see (i) we use another standard estimate (see again [Du]), to obtain for λ < 1,
k(I − λT )−1f kr .
.
.
.
r
r+1
f (j)(0) +
Xj=0ZT
Xj=0
(1 − λ)r+1 +ZT
(1 − λ)r+1 + kf k∞ZT
(1 − λ)r+1 .
kf kr
kf kr
f (j)(z)
1 − λzj+1 dm(z)
kf kr
f (z)
1 − λzr+2 dm(z)
1
1 − λzr+2 dm(z)
Then the assertion follows by an application of Theorem 4.7.
(ii) We have Mn(T )f = Fnf , where
Fn(z) =
1 − zn+1
(n + 1)(1 − z)
,
z ∈ D .
We claim that for f (z) = 1, n−rMn(T )f = n−rFn are unbounded in Xr. Using the Leibniz
rule and the triangle inequality we obtain n > r + 1
n−rF (r+1)
n
(z) ≥
ar
1 − z
− br
n−j
1 − zj+1 ,
r+1
Xj=1
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
27
with ar, br > 0 independent of n and z. The same standard estimates for integrals mentioned
above yield for ρn = 1 − 1
n
n−rZ 2π
0
F (r+1)
n
(ρneit)dt ≥ a′
r log n − b′
r ,
r, b′
with a′
(iii) For n > r + 1 and f (z) = 1 we have
r > 0 independent of n, and the claim together with (ii) follow.
nr+1 . kT nf kr = n(n − 1) . . . (n − r) ,
which gives the desired conclusion.
(cid:3)
In particular, without additional assumptions on the peripheral spectrum the conclusion
of Corollary 4.4 may fail.
Acknowledgements. The authors are grateful to the referee for his suggestions and
useful comments which improve the original version.
The second named author would like to express his gratitude to the University of Lund,
Centre for Mathematical Sciences, Sweden for hospitality and excellent working conditions.
He was also supported by a Project financed from Lucian Blaga University of Sibiu research
grants LBUS-IRG-2015-01.
References
[AS] J. Agler and M. Stankus, m-Isometric transformations of Hilbert space, I, Integr.
Equat. Oper. Th. Vol. 21 (1995), 383-429.
[AP] A. Aleman and A-M. Persson, Resolvent estimates and decomposable extensions of
generalized Ces`aro operators, J. Funct. Anal. 258 (2010), no. 1, 67-98.
[Al] G. R. Allan, Power-bounded elements and radical Banach algebras, Banach Center
Publ. 38, Institute of Mathematics, Polish Academy of Sciences, Warszawa, 1997, 9-16.
[AB] S. I. Ansari and P. S. Bourdon, Some properties of cyclic operators, Acta Sci. Math.
(Szeged) 63 (1997), 195-207.
[BC] F. Bayart and G. Costakis, Cyclic operators with finite support, Israel Journal of
Mathematics, (2012), 1-37.
[BM] F. Bayart and E. Matheron, Dynamics of linear operators, Cambridge University
Press, 2009.
[BG] E. Berkson and T. A. Gillespie Spectral decompositions, ergodic averages, and the
Hilbert transform, Studia Math., 144 (1), 2001, 39-61.
[B] J. Boos, Classical and modern methods in summability Assisted by Peter Cass. Oxford
Mathematical Monographs. Oxford Science Publications. Oxford University Press, Oxford,
2000.
[C] L. W. Cohen, On the Mean Ergodic Theorem, Ann. of Math. Second Series, Vol. 41,
No. 3, (1940), 505-509.
28
A. ALEMAN AND L. SUCIU
[Du] P.L. Duren, Theory of H p spaces, Pure and Applied Mathematics, Vol. 38 Academic
Press, New York-London 1970
[DyS] K. Dykema and H. Schultz, Brown measure and iterates of the Aluthge transform
for some operators arising from measurable actions, Trans. Amer. Math. Soc. 361 (2009),
6583-6593.
[Eb] W. F. Eberlein, Abstract Ergodic Theorems and Weak Almost Periodic Functions,
Trans. Amer. Math. Soc. 67 (1949), 217-240.
[Ed] E. Ed-Dari, On the (C, α) uniform ergodic theorem, Studia Math. 156 (1), 2003, 3-13.
[FR] O. El-Fallah and T. Ransford, Extremal growth of powers of operators satisfying
resolvent conditions of Kreiss-Ritt type, J. Funct. Anal. 196 (2002), 135-154.
[GLM] M. Gonzalez, F. Le´on-Saavedra, A. Montes-Rodr´ıguez, Semi-Fredholm theory: hy-
percyclic and supercyclic subspaces, Proc. London Math Soc. (3) 81 (2000), no. 1, 169-189.
[GH] J. J. Grobler and C. B. Huijsmans, Doubly Abel bounded operators with single spec-
trum, Questiones Mathematicae, 18 (1995), 397-406.
[He] D. A. Herrero, Limits of hypercyclic and supercyclic operators, J. Funct. Anal., 99
(1991), 179-190.
[Hi] E. Hille, Remarks on ergodic theorems, Trans. Amer. Math. Soc. 57 (1945), 246-269.
[K´e] L. K´erchy, Operators with regular norm-sequences, Acta Sci. Math.
(Szeged) 63
(1997), 571-605.
[Kr] U. Krengel, Ergodic Theorems, Walter de Gruyter Studies in Mathematics 6, Berlin -
New York, 1985.
[Ku] M. K. Kuo, Tauberian conditions for almost convergence, Positivity 13 (2009), 611-
619.
[L] G. G. Lorentz, A contribution to the theory of divergent sequences, Acta Math., 80
(1948), 167-190.
[LSS], M. Lin, D. Shoikhet and L. Suciu, Remarks on uniform ergodic theorems, Acta Sci
Math. (Szeged), 81 (2015), 251-283.
[MSZ] A. Montes-Rodr´ıguez, J. S´anchez- ´Alvarez, and J. Zem´anek, Uniform Abel-Kreiss
boundedness and the extremal behaviour of the Volterra operator, Proc. London Math. Soc.
91 (2005), 761-788.
[N] O. Nevanlinna, Resolvent conditions and powers of operators, Studia Math. 145 (2)
(2001), 113-134.
[R] A. V. Romanov, Weak convergence of operator means Izvestiya: Mathematics 75:6,
2011, 1165-1183.
[Sc] M. Schreiber, Uniform families of ergodic operator nets, Semigroup Forum, Vol. 86,
Issue 2, (2013), 321-336.
[Sh] A. L. Shields, On Mobius bounded operators, Acta Sci. Math. (Szeged) 40 (1978),
371-374.
ON ERGODIC OPERATOR MEANS IN BANACH SPACES
29
[SW] J. C. Strikwerda and B. A. Wade, A survey of the Kreiss matrix theorem for power
bounded families of matrices and its extensions, Banach Center Publ. 38, Institute of Math-
ematics, Polish Academy of Sciences, Warszawa, 1994, 339-360.
[SZ], L. Suciu and J. Zem´anek, Growth conditions and Ces`aro means of higher order, Acta
Sci Math. (Szeged), 79 (2013), 545-581.
[V] Q. P. Vu, A short proof of the Y. Katznelson’s and L. Tzafriri’s theorem, Proc. Amer.
Math. Soc., 115 (1992), 1023-1024.
Lund University, Mathematics, Faculty of Sciences, P.O. Box 118 S-221 00, Lund, Sweden
E-mail address: [email protected]
Department of Mathematics, ”Lucian Blaga” University of Sibiu, Dr. Ion Rat¸iu 5-7, Sibiu,
550012, Romania
E-mail address: [email protected]
|
1506.03260 | 1 | 1506 | 2015-06-10T11:35:36 | Estimates for entropy numbers of embedding operators of function spaces on sets with tree-like structure: some limiting cases | [
"math.FA",
"math.CA"
] | In this paper we obtain order estimates for entropy numbers of embeddings of weighted Sobolev spaces into weighted Lebesgue spaces and of weighted summation operators on trees. Here we consider some critical conditions on the parameters. | math.FA | math |
Estimates for entropy numbers of embedding
operators of function spaces on sets with tree-like
structure: some limiting cases
A.A. Vasil'eva
1 Introduction
In [37] order estimates for entropy numbers of the embedding operator of a weighted
Sobolev space on a John domain into a weighted Lebesgue space were obtained, as
well as estimates for entropy numbers of a two-weighted summation operator on a
tree. Here we consider some critical cases.
Recall the definition of entropy numbers (see, e.g., [6, 9, 32]).
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 ,
k ∈ N.
Kolmogorov, Tikhomirov, Birman and Solomyak [3, 16, 34] studied properties of
ε-entropy (this magnitude is related to entropy numbers of embedding operators).
Denote by lm
p (1 6 p 6 ∞) the space Rm with norm
k(x1, . . . , xm)klm
p =(cid:26) (x1p + · · · + xmp)1/p for p < ∞,
max{x1, . . . , xm} for p = ∞.
Estimates for entropy numbers of the embedding operator of lm
q were obtained
in the paper of Schutt [33] (see also [9]). Later Edmunds and Netrusov [7], [8]
generalized this result for vector-valued sequence spaces (in particular, for sequence
spaces with mixed norm).
p into lm
Haroske, Triebel, Kuhn, Leopold, Sickel and Skrzypczak [10 -- 15, 17 -- 24] studied
the problem of estimating entropy numbers of embeddings of weighted sequence
spaces or weighted Besov and Triebel -- Lizorkin spaces.
Triebel [35] 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.
Lifshits and Linde [28] obtained estimates for entropy numbers of two-weighted
Hardy-type operators on a semiaxis (under some conditions on weights). The similar
1
problem for one-weighted Riemann-Liouville operators was considered in the paper
of Lomakina and Stepanov [29]. In addition, Lifshits and Linde [25 -- 27] studied the
problem of estimating entropy numbers of two-weighted summation operators on a
tree.
This paper is organized as follows. In §2 we obtain the general result about
upper estimates for entropy numbers of embedding operators of function spaces
on a set with tree-like structure. The properties of such spaces are almost the
same as properties of function spaces defined in [36, 37] (see Assumptions 1, 2,
3), but there are some differences. In particular, here we suppose that (6) holds;
this condition cannot be directly derived from known results for weighted Sobolev
and Lebesgue spaces. Therefore, first we consider some particular cases of function
spaces satisfying Assumptions A -- C or A, B, D (see §3) and we prove that these
spaces satisfy Assumptions 1 -- 3. In §4 we obtain order estimates for entropy numbers
of embedding operators of weighted Sobolev spaces; to this end, we prove that under
given conditions on weights Assumptions A -- C or A, B, D hold. In §5 we obtain order
estimates for entropy numbers of two-weighted summation operators on a tree in
critical cases.
2 Upper estimates for entropy numbers of embedding
operators of function spaces on sets with tree-like
structure
Let us 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) the vertex set of G. 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.
Let G be a disjoint union of trees (Tj, ξj), 1 6 j 6 k. Then the partial order on
each tree Tj induces the partial order on G.
2
Given j ∈ Z+, ξ ∈ V(T ), we denote
Vj(ξ) := VT
j (ξ) := {ξ′ > ξ : ρT (ξ, ξ′) = j}.
For ξ ∈ V(T ) we denote by Tξ = (Tξ, ξ) the subtree in T with vertex set
{ξ′ ∈ V(T ) : ξ′ > ξ}.
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 vertex set
W if V(G) = W and any two vertices ξ′, ξ′′ ∈ W adjacent in T are also adjacent
in G. Given subgraphs Γ1, Γ2 ⊂ T , we denote by Γ1 ∩ Γ2 the maximal subgraph in
T on the vertex set V(Γ1) ∩ V(Γ2).
Let P = {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).
If Γ ⊂ T is the maximal subgraph on the vertex set W, we set PΓ = {Γ ∩ Tj}j∈N.
We consider the function spaces on sets with tree-like structure from [36, 37].
Let (Ω, Σ, mes) be a measure space, let Θ be a countable partition of Ω into
measurable subsets, let (A, ξ0) be a tree such that
∃c1 > 1 :
card VA
1 (ξ) 6 c1,
ξ ∈ V(A),
(1)
and let F : V(A) → Θ be a bijective mapping.
Throughout we consider at most countable partitions into measurable subsets.
Let 1 6 p, 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 Banach space Yq(E) with norm 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;
3
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
5. if E ∈ Σ, f ∈ Yq(Ω), then f · χE ∈ Yq(Ω).
,
,
f ∈ Xp(E),
f ∈ Yq(E);
(2)
(3)
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}.
(4)
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
(5)
with 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).
In [37] upper estimates for entropy numbers of the embedding operator of the
space Xp(Ω) into Yq(Ω) were obtained under some conditions on these spaces (the
space Xp(Ω) ∼= Xp(Ω)/P(Ω) will be defined later). Some limiting relations between
the parameters were not considered. Here we investigate one of those critical cases.
Throughout we assume that 1 < p < q < ∞ and the following conditions hold.
Assumption 1. There exist a partition {At,i}t>t0, i∈ Jt of the tree A and a number
c > 1 with the following properties.
1. If the tree At′,i′ follows the tree At,i, then t′ = t + 1.
2. For each vertex ξ∗ ∈ V(A) there exists a linear continuous projection Pξ∗ :
Yq(Ω) → P(Ω) such that for any function f ∈ Xp(Ω) and for any subtree
D ⊂ A rooted at ξ∗
kf − Pξ∗f kq
Yq(ΩD) 6 c
∞
Xt=t0 Xi∈Jt,D
here
2(1− q
p )tkf kq
Xp(ΩDt,i );
Jt,D = {i ∈ Jt : V(At,i) ∩ V(D) 6= ∅}, Dt,i = At,i ∩ D.
4
(6)
(7)
Assumption 2. There exist numbers δ∗ > 0 and c2 > 1 such that for each vertex
ξ ∈ V(A) and for any n ∈ N, m ∈ Z+ there exists 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∗2( 1
q − 1
p)tkf kXp(E),
where t > t0 is such that ξ ∈ ∪j∈ Jt
V(At,j).
3. For any E ∈ Tm,n(G)
card {E′ ∈ Tm±1,n(G) : mes(E ∩ E′) > 0} 6 c2.
(8)
(9)
Assumption 3. There exist numbers γ∗ > 0, c3 > 1 and an absolutely continuous
function ψ∗ : (0, ∞) → (0, ∞) such that lim
card V(At,i)
y→∞
∗(y)
yψ′
ψ∗(y) = 0 and for νt := Pi∈ Jt
the following estimate holds:
νt 6 c3 · 2γ∗2t
ψ∗(22t
) =: c3νt,
t > t0.
(10)
Assumption 1 together with the inequality p < q implies that for any t > t0
and for each vertex ξ∗ ∈ VA
t−t0(ξ0) there exists a linear continuous projection Pξ∗ :
Yq(Ω) → P(Ω) such that for any function f ∈ Xp(Ω) and for any subtree D ⊂ A
rooted at ξ∗
kf − Pξ∗f kYq(ΩD) 6 c · 2( 1
q − 1
p )tkf kXp(ΩD).
(11)
Hence, Assumptions 1 -- 3 from [36, 37] hold with λ∗ = µ∗ = 1
q and u∗ ≡ 1.
In particular, there exist a linear continuous projection P : Yq(Ω) → P(Ω) and a
number M > 0 such that for any function f ∈ Xp(Ω) the following estimate holds:
p − 1
kf − P f kYq(Ω) 6 Mkf kXp(Ω).
As P we take the operator Pξ0 (recall that ξ0 is the root of A). Similarly as in [37]
we set
Xp(Ω) = {f − P f : f ∈ Xp(Ω)}
and denote by I the embedding operator of Xp(Ω) into Yq(Ω).
We set Z0 = (p, q, c1, c2, c3, c, δ∗, γ∗, ψ∗).
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).
5
Theorem 1. Suppose that Assumptions 1 -- 3 hold. Then
en(I : Xp(Ω) → Yq(Ω)) .
Z0
1
q − 1
p .
n
(12)
Similarly as in [36], [37] we introduce some more notation.
• ξt,i is the minimal vertex of the tree At,i.
• Gt = ∪ξ∈V(Γt) F (ξ) = ∪i∈ JtΩAt,i.
• Γt is the maximal subgraph on the vertex set ∪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 [36]); hence, we may assume that
J t = Jt,
t > t0.
(13)
(14)
The set Jt0 is a singleton. Denote {i0} = Jt0.
In [37] the operators Qt and Pt,m were defined as follows.
Definition of the operator Qt. For each t > t0, i ∈ Jt
(14)
= J t there exists a
linear continuous operator Pt,i : Yq(Ω) → P(Ω) such that for any function f ∈ Xp(Ω)
and for any subtree A′ ⊂ A rooted at ξt,i
kf − Pt,if kYq(ΩA′ )
(11)
.
Z0
2−( 1
p − 1
q )tkf kXp(ΩA′ ).
(15)
As Pt,i we take P ξt,i
it follows that Pt0,i0 Xp(Ω) = 0 (see [37]).
We set
. From the definition of the space Xp(Ω) and of the operator P
Qtf (x) = Pt,if (x) = P ξt,i
f (x) for x ∈ Ut,i, i ∈ J t,
Qtf (x) = 0 for x ∈ Ω\ Ut,
Tt = {Ut+1,i}i∈J t+1.
6
(16)
(17)
Since p < q, we have for any f ∈ B Xp(Ω)
kf − Qtf kYq( Ut) 6 kf − Qtf kYp,q,Tt−1 ( Ut)
(15)
.
Z0
2−( 1
p − 1
q )t.
Notice that if t < t0, then Qtf = Qt+1f = 0 (since Pt,i0 = 0 for t < t0).
Throughout we set log x := log2 x.
Definition of the operators Pt,m. For t > t0 we set
mt = ⌈log νt⌉.
(18)
(19)
In [36] for each m ∈ Z+ the set Gm,t ⊂ Gt, the partition Tt,m of the set Gm,t and
the linear continuous operator
we constructed. Here the following properties hold:
Pt,m : Yq(Ω) → S Tt,m(Ω)
1. Gm,t ⊂ Gm+1,t, Gmt,t = Gt;
2. for any m ∈ Z+
card Tt,m .
Z0
2m;
3. for any function f ∈ Xp(Ω) and for any set E ∈ Tt,m
kf − Pt,mf kYq(E) .
Z0
2−( 1
p − 1
q )tkf kXp(E), m 6 mt,
kf − Pt,mf kYq(E) .
Z0
4. for any set E ∈ Tt,m
2−( 1
p − 1
q )t · 2−δ∗(m−mt)kf kXp(E), m > mt;
card {E′ ∈ Tt,m±1 : mes(E ∩ E′) > 0} .
Z0
1.
Moreover, we may assume that
Tmt,t = { F (ξ)}ξ∈V(Γt),
Pt,mtf F (ξ) = Pξf F (ξ),
ξ ∈ V(Γt)
(it follows from the construction in [36, p. 37 -- 40]).
7
(20)
(21)
(22)
(23)
(24)
(25)
(26)
Let
Then
t∗(n) = min{t ∈ N : ν t > n},
t∗∗(n) = min{t ∈ N : ν t > 2n}.
2t∗(n) ≍
Z0
log n,
2t∗∗(n) ≍
Z0
n
(see [37, formula (49)]).
For any f ∈ Xp(Ω) the following equality holds:
t∗(n)−1
f =
(Qt+1f − Qtf )χ Ut+1+
Pt=t0
t∗(n)−1
Pt=t0
∞
Pm=0
+
(Pt,m+1f − Pt,mf )χGm,t + (f − Qt∗(n)f )χ Ut∗(n)
(it can be proved similarly as formula (82) in [36]).
We set
Qf Gt = Pt,mtf Gt,
t > t0,
Qn,mf Gt = Pt,mt+mf Gt
Qn,mf Gt = 0 for
for
t∗(n) 6 t < t∗∗(n), m ∈ Z+,
t < t∗(n)
or
t > t∗∗(n).
(22),(25)
q )tkf kXp( F (ξ)) for any ξ ∈ V(Γt) and the
Since kf − Pt,mtf kYq( F (ξ))
space Yq(Ω) is Banach, we get from the inequality p < q that for any f ∈ Xp(Ω) the
inclusion Qf ∈ Yq(Ω) holds and
2−( 1
p − 1
.
Z0
kf − Qf kYq( Ut∗∗(n)) .
Z0
Further,
2( 1
q − 1
p )t∗∗(n)kf kXp(Ω)
(29)
.
Z0
n
1
q − 1
p kf kXp(Ω).
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
(f − Qt∗(n)f )χ Ut∗(n)
+
= ( Qf − Qt∗(n)f )χ Ut∗(n)
+
( Qn,m+1f − Qn,mf ) + (f − Qf )χ Ut∗∗(n)
∞
;
Pm=0
indeed, from (23), (31) and (32) it follows that
Qf )χ Ut∗(n)\ Ut∗∗(n)
.
8
( Qn,m+1f − Qn,mf ) = (f −
∞
Pm=0
Lemma 1. There exists a sequence {kt}t∗(n)−1
t=t0 ⊂ N such that
t∗(n)−1
Xt=t0
(kt − 1) .
Z0
n,
t∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
1
q − 1
p .
n
(35)
(36)
Lemma 2. There exists a sequence {kt,m}t06t<t∗(n),m∈Z+ ⊂ N such thatPt,m
n,
(kt,m−1) .
Z0
t∗(n)−1
∞
Xt=t0
Xm=0
ekt,m(Pt,m+1 − Pt,m : Xp(Ω) → Yq(Gm,t)) .
Z0
1
q − 1
p .
n
Lemmas 1 and 2 are proved similarly as Lemmas 6 and 7 in [37].
Lemma 3. We have
en(( Q − Qt∗(n))χ Ut∗(n)
: Xp(Ω) → Yq(Ω)) .
Z0
1
q − 1
p .
n
Notice that for t∗∗(n) 6 t0 this estimate follows from (33). Hence, throughout
we assume that t∗∗(n) > t0.
We need some auxiliary assertions.
For any ν ∈ N we denote by Iν the identity operator on Rν.
Theorem A. [9, 33]. Let 1 6 p 6 q 6 ∞. Then
ek(Iν : lν
p → lν
q ) ≍
p,q
1,
1 6 k 6 log ν,
k )
(cid:18) log(1+ ν
2− k
ν ν
k
q − 1
p ,
1
(cid:19)
1
p − 1
q
,
log ν 6 k 6 ν,
ν 6 k.
The following properties of entropy numbers are well-known (see, e.g., [9, 32]):
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 ).
In particular, from the first property it follows that
ek(ST ) 6 kSkek(T ),
ek(ST ) 6 kT kek(S).
(37)
(38)
9
Theorem B. [25]. 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 .
Lemma 4. [39]. Let (T , ξ∗) be a tree with finite vertex set, let
card V1(ξ) 6 k
for any vertex ξ ∈ V(T ),
(39)
and let the mapping Φ : 2V(T ) → R+ satisfy the following conditions:
Φ(V1 ∪ V2) > Φ(V1) + Φ(V2), V1, V2 ⊂ V(T ), V1 ∩ V2 = ∅,
(40)
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.
Lemma 5. [36]. Let T be a finite partition of a measurable subset G ⊂ Ω, ν =
dim ST (Ω) (see (4)). Then there exists a linear isomorphism A : ST (Ω) → Rν such
that kAkYp,q,T (G)→lν
1, kA−1klν
1.
σp,q
.
σp,q, r0
q →Yq(G) .
q, r0
Lemma 6. [44, formula (60)]. 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 < ∞.
(41)
Given t∗(n) 6 t 6 t∗∗(n), i ∈ Jt, we denote
At,i =(cid:26) At,i
for
At∗∗(n),i
t < t∗∗(n),
for
t = t∗∗(n),
Γt = Γt for t < t∗∗(n), Γt∗∗(n) = Γt∗∗(n). Let D be a subtree in Γt∗(n). We set
Jt,D = {i ∈ Jt : V(At,i) ∩ V(D) 6= ∅}, Dt,i = At,i ∩ D.
(42)
Throughout we take as ε = ε(Z0) > 0 a sufficiently small number (it will be
chosen later by Z0).
10
Proof of Lemma 3. We set t′
∗(n) = max{t∗(n), t0}.
Step 1. Given t < t∗∗(n), we denote by At the subtree in A with vertex set
V(A)\V(Γt+1).
2V( At) → R+ by
Let f ∈ B Xp(Ω). For each t′
∗(n) 6 t < t∗∗(n) we define the mapping Φf,t :
Φf,t(W) = Xξ∈W∩V(Γt)
kf kp
Xp( F (ξ))
.
Then for any disjoint sets W1, W2 we have
Φf,t(W1 ⊔ W2) = Φf,t(W1) + Φf,t(W2).
For each t′
∗(n) 6 t < t∗∗(n) we set
kf kp
Xp( F (ξ))
;
εt = Pξ∈V(Γt)
if
nt = ⌈n · 2−tεt⌉
εt > 0; nt = 1 if
εt = 0.
Then
t∗∗(n)−1
Xt=t′
∗(n)
εt 6 1,
t∗∗(n)−1
Xt=t′
∗(n)
2tnt 6
t∗∗(n)−1
Xt=t′
∗(n)
nεt +
t∗∗(n)−1
Xt=t′
∗(n)
(29),(45)
6
2t
C(Z0)n
with C(Z0) ∈ N. By (44),
Φf,t(V( At)) = εt.
(43)
(44)
(45)
(46)
(47)
Let f 6= 0. It follows from Lemma 4 and (1) that there exists a number C(Z0) ∈ N
and a family of partitions {Tf,t,l}06l6log nt of the tree At, which satisfy the following
conditions:
card Tf,t,l 6 C(Z0)2−lnt,
card {A′′ ∈ Tf,t,l±1 : V(A′) ∩ V(A′′) 6= ∅} .
Z0
1, A′ ∈ Tf,t,l,
and for any subtree A′ ∈ Tf,t,l such that card V(A′) > 2
Φf,t(V(A′))
(47)
6 C(Z0)2ln−1
t εt.
11
(48)
(49)
(50)
Moreover, we may assume that
For f ≡ 0 we set Tf,t,l = { At}.
Given t < t∗∗(n), we denote
Tf,t,⌊log nt⌋ = { At}.
Wt =nξ ∈ V(Γt) : {ξ} ∈ Tf,t,0, kf kp
Xp( F (ξ))
> C(Z0)n−1
t εto ,
Wt = {ξ ∈ Wt : VA
1 (ξ) ∩ { ξt+1,j}j∈ Jt+1
6= ∅},
St = {ξ ∈ V(Γt) : ∃D ∈ Tf,t,0 : ξ ∈ Vmin(D)},
St = St ∪(cid:16)∪ξ∈Wt−1[VA
St∗∗(n) = ∪ξ∈W
1 (ξ) ∩ V(Γt)](cid:17) for
[VA
1 (ξ) ∩ V(Γt∗∗(n))],
t∗∗(n)−1
t < t∗∗(n),
S =(cid:16)∪t∗∗(n)
t=t′
∗(n)
Then
St(cid:17) ∪ {ξt′
∗(n),j}j∈ Jt′
.
∗(n)
S\ ∪t∗∗(n)−1
∗(n){ ξt,j : j ∈ Jt}, Wt ⊂ St.
For each vertex ξ ∈ S we denote by D(ξ) the tree with vertex set
∗(n) St ⊂ ∪t∗∗(n)
t=t′
t=t′
V(D(ξ)) = {ξ′ > ξ :
[ξ, ξ′] ∩ S = {ξ}}.
We set
Tf = {D(ξ) : ξ ∈ S}.
(51)
(52)
(53)
(54)
(55)
(56)
(57)
(58)
(59)
Then Tf is a partition of the graph Γt′
∗(n) into subtrees.
Step 2. We say that D ∈ Tf if D ∈ Tf and for any ξ ∈ Wt we have D 6= {ξ}.
Let (D, ξ∗) ∈ Tf . Then ξ∗ ∈ V(Γt′
∗(n)). From Assumption 1 and the inequality
p < q it follows that
kf − Pξ∗f kq
Yq(ΩD)
(6),(42)
.
Z0
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) Xj∈ Jt,D
kf kq
Xp(ΩDt,j
),
kf − Qf kq
Yq(ΩD)
(31)
=
kf − Pt,mtf kq
Yq(ΩDt,j
(22)
.
Z0
)
t∗∗(n)
Xt=t′
∗(n) Xj∈ Jt,D
12
.
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) Xj∈ Jt,D
kf kq
Xp(ΩDt,j
).
Hence,
X(D, ξ∗)∈ Tf
kf − Pξ∗f kq
Yq(ΩD) .
Z0
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) X(D, ξ∗)∈ Tf Xj∈ Jt,D
kf kq
Xp(ΩDt,j
),
(60)
X(D, ξ∗)∈ Tf
kf − Qf kq
Yq(ΩD) .
Z0
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) X(D, ξ∗)∈ Tf Xj∈ Jt,D
kf kq
Xp(ΩDt,j
).
(61)
Denote by Tf,t the family of trees D ∈ Tf,t,0 such that
Xξ∈V( D)∩V(Γt)
kf kp
Xp( F (ξ))
6 C(Z0)n−1
t εt.
(62)
Assertion 1. Let D ∈ Tf , t′
T ∈ Tf,t such that Dt,j = Tt,j.
∗(n) 6 t < t∗∗(n), j ∈ Jt,D. Then there exists a tree
Proof of Assertion 1. Let D = D(ξ∗), ξ∗ ∈ S. Since j ∈ Jt,D, the vertices ξ∗ and
ξt,j are comparable. There exists a tree T ∈ Tf,t,0 such that
max{ξ∗, ξt,j} ∈ V(T ).
(63)
We claim that V(Dt,j) ⊂ V(T ). Indeed, let ξ ∈ V(Dt,j)\V(T ). Then ξ > max{ ξt,j, ξ∗}.
We set
η∗ = min{η ∈ [max{ξ∗, ξt,j}, ξ] : η /∈ V(T )} > ξ∗.
Then η∗ ∈ V(Γt) and there exists a tree T ′ ∈ Tf,t,0 such that η∗ is the minimal
vertex of T ′. Hence, η
contradiction.
(54)
∈ St
(55),(56)
⊂
S; therefore, ξ
(58)
/∈ V(D), which leads to a
Thus, V(Tt,j) ⊃ V(Dt,j) 6= ∅.
Let us show that V(Tt,j) ⊂ V(Dt,j). Denote by ξ the minimal vertex of the
tree T . Since V(Tt,j) 6= ∅, the vertices ξ and ξt,j are comparable. Let us show that
max{ξ∗, ξt,j} = max{ ξ, ξt,j}. Indeed, by (63) we have max{ξ∗, ξt,j} > max{ ξ, ξt,j}.
If the inequality is strict, then ξ∗ > ξt,j. In addition, ξ∗ ∈ V(At,j) ∩ S, and by (54)
and (57) we get ξ∗ ∈ St; i.e., ξ∗ is the minimal vertex of some tree from the partition
Tf,t,0. By (63), ξ∗= ξ.
Thus,
max{ ξ, ξt,j} = max{ξ∗, ξt,j} ∈ V(Dt,j).
(64)
13
Suppose that there exists a vertex ξ ∈ V(Tt,j)\V(D). We set
η∗∗ = min(cid:16)hmax{ ξ, ξt,j}, ξi \V(D)(cid:17) .
(65)
Then η∗∗ > max{ ξ, ξt,j} and η∗∗ ∈ V(At,j). Hence, η∗∗ /∈ ∪t∗∗(n)
i ∈ Jt′}.
t′=t′
Further, η∗∗ ∈ V(Tt,j); therefore, η∗∗ /∈ ∪t∗∗(n)
∗(n)St′ by (54). From (57) it follows that
t′=t′
η∗∗ /∈ S. This together with (58), (64), (65) yields that [ξ∗, η∗∗] ∩ S = {ξ∗}; applying
(58) once again, we get that η∗∗ ∈ V(D), which leads to a contradiction.
∗(n){ ξt′,i :
It remains to check that T ∈ Tf,t. If T /∈ Tf,t, then V(T )
(50),(62)
=
{ξ},
kf kXp( F (ξ)) > C(Z0)n−1
t εt; i.e., ξ
(52)
∈ Wt. Then either ξ ∈ Wt (in this case,
VA
1 (ξ)
that V(D)
(54),(55)
St ∪ St+1) or VA
⊂
(58)
= {ξ} and D /∈ Tf . This completes the proof of Assertion 1. ⋄
1 (ξ)
1 (ξ)
(54)
⊂ St. Hence, VA
(56)
⊂ S. This implies
By (54), the first inclusion of (57), (58) and (59), for any tree D ∈ Tf we have
Dt∗∗(n),j = At∗∗(n),j, j ∈ Jt∗∗(n). This together with Assertion 1 yields that
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) X(D, ξ∗)∈ Tf Xj∈ Jt,D
kf kq
Xp(ΩDt,j
) 6
t∗∗(n)−1
Xt=t′
∗(n)
2t(1− q
p ) XT ∈ Tf,t Xj∈Jt,T
kf kq
Xp(ΩTt,j )+
+2t∗∗(n)(1− q
+
. n1− q
p +
i.e.,
t∗∗(n)−1
∗(n)
t∗∗(n)−1
Xt=t′
Xt=t′
∗(n)
2t(1− q
2t(1− q
(29)
.
Z0
)
n1− q
p +
Xp(Ω At∗∗(n),j
kf kq
p ) Xj∈ Jt∗∗(n)
p ) XT ∈ Tf,t
Xξ∈V(T )∩V(Γt)
p ) XT ∈ Tf,t
(48)
.
Z0
− q
p
t n
t
ε
q
p
kf kp
Xp( F (ξ))
t∗∗(n)−1
Xt=t′
∗(n)
q
p
(62)
.
Z0
n1− q
p +
2t(1− q
p )ε
q
p
1− q
p
t n
t
;
t∗∗(n)
Xt=t′
∗(n)
2t(1− q
p ) X(D, ξ∗)∈ Tf Xj∈ Jt,D
kf kq
Xp(ΩDt,j
).
Z0
n1− q
p +
2t(1− q
p )ε
t∗∗(n)−1
Xt=t′
∗(n)
If n · 2−tεt > 1, then nt
(44)
≍ n · 2−tεt; hence,
q
p
t n
1− q
p
t =: A. (66)
2t(1− q
p )ε
q
p
t n
1− q
p
t ≍
p,q
2t(1− q
p )ε
q
p
t n1− q
p 2−t(1− q
p )ε
1− q
t = n1− q
p
p εt.
14
If n · 2−tεt < 1, then nt = 1. From (45) it follows that εt 6 1. Therefore,
n1− q
p +
A .
Z0
t∗∗(n)−1
Xt=t′
∗(n)
n1− q
p εt +
t∗∗(n)−1
Xt=t′
∗(n)
p )
2t(1− q
(45)
.
Z0
From (60), (61), (66), (67) we get that
n1− q
p )
p + 2t∗∗(n)(1− q
(29)
.
Z0
n1− q
p . (67)
X(D, ξ∗)∈ Tf
this yields
(kf − Pξ∗f kq
Yq(ΩD) + kf − Qf kq
Yq(ΩD)) .
Z0
n1− q
p ;
X(D, ξ∗)∈ Tf
k Qf − Pξ∗f kq
Yq(ΩD) .
Z0
n1− q
p .
(68)
Step 3. Let T be a partition of Γt′
∗(n) into subtrees. For each (D, ξ∗) ∈ T we set
In addition, we put
Then
PTf ΩD = Pξ∗f − Qt′
∗(n)f.
PTf Ω\ Ut′
∗(n)
= 0.
(69)
(70)
k Qf − Qt′
∗(n)f − PTf kq
Yq( Ut′
∗(n))
k Qf − Pξ∗f kq
Yq(ΩD)
(26),(31)
=
= X(D, ξ∗)∈T
=
X(D, ξ∗)∈T, card V(D)>2
k Qf − Pξ∗f kq
Yq(ΩD).
In particular, from (68) and the definition of Tf (see the beginning of Step 2) it
follows that
k Qf − Qt′
∗(n)f − PTf f kq
Yq( Ut′
∗(n))
n1− q
p .
.
Z0
(71)
Step 4. Let us define the family N of partitions of the graph Γt′
∗(n) into subtrees.
Each of these partitions is constructed as follows. Let C(Z0), C(Z0) be as defined at
Step 1.
1. We choose the sequence {nt}t∗∗(n)−1
∗(n) ⊂ N such that
t=t′
t∗∗(n)−1
Xt=t′
∗(n)
2tnt 6 C(Z0)n.
(72)
15
2. For each t ∈ {t′
∗(n), . . . , t∗∗(n) − 1} we take C(Z0)nt vertices in V(Γt) (some
of them may coincide); we denote this set by Ut.
3. For each t ∈ {t′
∗(n), . . . , t∗∗(n) − 1} we choose an arbitrary subset Vt of
{ξ ∈ Ut : VA
1 (ξ) ∩ V(Γt+1) 6= ∅}.
4. Let
U = ∪j∈ Jt′
∗(n)
{ ξt′
∗(n),j} ∪(cid:16)∪t∗∗(n)−1
∗(n)
t=t′
Ut(cid:17) ∪(cid:16)∪t∗∗(n)−1
t=t′
∗(n) ∪ξ∈ Vt [VA
This vertex set generates the desired partition T of the graph Γt′
subtrees:
1 (ξ) ∩ V(Γt+1)](cid:17) .
∗(n) into
T = {D′
(ξ) : ξ ∈ U}, V(D′
(ξ)) = {ξ′ > ξ :
[ξ, ξ′] ∩ U = {ξ}}.
(73)
Let us estimate the value N .
1. First we estimate the number of choices of {nt}t∗∗(n)−1
∗(n)
(we denote this value
by N1). Let 1 6 l 6 C(Z0)n. The number of choices of nt ∈ N such that
t∗∗(n)−1
t=t′
2tnt = l can be estimated from above by the number of choices of
numbers nt ∈ N such that
nt = l. The last magnitude can be estimated
from above by the number of partitions of {1, . . . , l} into t∗∗(n) − t′
intervals. This value does not exceed (l + t∗∗(n) − t′
∗(n)
∗(n)−1. Hence,
∗(n))t∗∗(n)−t′
t∗∗(n)−1
Pt=t′
∗(n)
Pt=t′
∗(n)
N1 6 X16l6 C(Z0)n
Xk=1
C(Z0)n+t∗∗(n)
6
(l + t∗∗(n) − t′
∗(n))t∗∗(n)−t′
∗(n)−1 6
kt∗∗(n)−1 .
Z0
( C(Z0)n + t∗∗(n))t∗∗(n) =: N ′
1.
2. Given the sequence {nt}t∗∗(n)−1
∗(n) , we estimate the number of choices of a set
t=t′
∪t∗∗(n)−1
∗(n)
t=t′
Ut (we denote this magnitude by N2). We have
(10)
6
N2
t∗∗(n)−1
Yt=t′
∗(n)
(c3ν t)C(Z0)nt =: N ′
2.
3. Let us estimate the number of choices of Vt, t′
∗(n) 6 t 6 t∗∗(n) − 1 (denote
this value by N3). Since card Ut 6 C(Z0)nt, we have
2C(Z0)nt =: N ′
3.
N3 6
t∗∗(n)−1
Yt=t′
∗(n)
16
Thus, N 6 N ′
1N ′
2N ′
3, which yields
log N 6 t∗∗(n) log( C(Z0)n + t∗∗(n)) +
t∗∗(n)−1
Xt=t′
∗(n)
C(Z0)nt log(c3νt)+
+
t∗∗(n)−1
Xt=t′
∗(n)
(10),(29),(41)
C(Z0)nt
.
Z0
(log n)2 +
t∗∗(n)−1
Xt=t′
∗(n)
2tnt
(72)
.
Z0
n;
i.e.,
log N .
Z0
n.
(74)
Step 5. Denote by N ′ the family of partitions Tf , f ∈ B Xp(Ω) (see (59)). Then
N ′ ⊂ N (it follows from (52) -- (56), (58), (73) and from the estimates (46), (48)).
By (74) and Theorem B, there exists l∗ = l∗(Z0) ∈ N such that
el∗n( Q − Qt′
∗(n) : Xp(Ω) → Yq( Ut′
∗(n))) 6
6 sup
f ∈B Xp(Ω)
inf
T∈N ′
k Qf − Qt′
∗(n)f − PTf kYq( Ut′
∗(n)) + sup
T∈N ′
en(PT : Xp(Ω) → Yq( Ut′
∗(n))) 6
6 sup
f ∈B Xp(Ω)
k Qf − Qt′
∗(n)f − PTf f kYq( Ut′
∗(n)) + sup
T∈N ′
en(PT : Xp(Ω) → Yq( Ut′
∗(n))) .
Z0
(71)
.
Z0
n
1
q − 1
p + sup
T∈N ′
en(PT : Xp(Ω) → Yq( Ut′
∗(n))).
Step 6. It remains to prove that
en(PTf : Xp(Ω) → Yq( Ut′
∗(n))) .
Z0
1
q − 1
p ,
n
(75)
where f ∈ B Xp(Ω).
Recall that Tf = {D(ξ) : ξ ∈ S}, where S is defined by (56) and D(ξ) is defined
by formula (58). Also we observe that if ξ = ξt′
∗(n),j for some j ∈ Jt′
∗(n), then
PTf ΩD(ξ)
(69)
= (P ξt′
∗(n),j
− Qt′
∗(n))ΩD
(16)
= 0.
( ξ
t′
∗(n),j
)
(76)
Let t′
∗(n) 6 t < t∗∗(n). Recall that the sets St and Wt are defined by formulas
(54) and (53), respectively. We set
Vf,t = St ∪ {ξ ∈ V(Γt+1)\St+1 : ∃η ∈ Wt : ξ ∈ VA
1 (η)}.
(77)
17
Notice that Vf,t ⊂ S by (55), (56). Denote by Γf,t the maximal subgraph in A on
vertex set
V(Γf,t) = ∪ξ∈Vf,t
V(D(ξ)).
(78)
We set
Denote
Ω(f,t) = ∪ξ∈V(Γf,t)
F (ξ), P(f,t)h = PTf h · χΩ(f,t),
h ∈ Yq(Ω).
(79)
T′
f,t = {D(ξ)}ξ∈Vf,t.
Then T′
f,t is a partition of the graph Γf,t.
We claim that
PTf =
t∗∗(n)−1
Xt=t′
∗(n)
P(f,t).
(80)
(81)
Indeed, let h ∈ Yq(Ω). If x ∈ ∪t∗∗(n)−1
t∗∗(n)−1
t=t′
∗(n) ∪ξ∈Vf,t ΩD(ξ), then by (79) we get PTf h(x) =
t∗∗(n)−1
Pt=t′
∗(n)
Pt=t′
∗(n)
P(f,t)h(x). In other cases we have
P(f,t)h(x) = 0. On the other hand, if
PTf h(x) 6= 0, then x ∈ ΩD(ξ) for some ξ ∈ S by (59) and (70). From (55), (56), (76)
and (77) it follows that ξ ∈ Vf,t for some t ∈ {t′
∗(n), . . . , t∗∗(n) − 1}. This completes
the proof of (81).
Let us prove that there exists a sequence {kt}t′
∗(n)6t<t∗∗(n) such that
t∗∗(n)−1
Xt=t′
∗(n)
(kt − 1) .
Z0
n,
t∗∗(n)−1
Xt=t′
∗(n)
ekt(P(f,t) : Xp(Ω) → Yq(Ω)) .
Z0
1
q − 1
p .
n
(82)
Let ξ ∈ St. If ξ ∈ St\Wt, then we set
D(ξ) = D(ξ),
P(f,t)Ω D(ξ)
= P(f,t)ΩD(ξ)
(69),(79)
=
(Pξ − Qt′
∗(n))ΩD(ξ)
.
(83)
If η ∈ Wt, then we denote by D(η) the tree with vertex set
V( D(η)) = {η} ∪(cid:16)∪ξ∈VA
1 (η)∩V(Γt+1)\St+1
V(D(ξ))(cid:17)
(84)
18
and put
P(f,t)Ω D(η)
= (Pη − Qt′
∗(n))Ω D(η)
.
Observe that by (78) we have
T′′
f,t := { D(ξ)}ξ∈St
is the partition of Γf,t.
Finally, we set
Let
P(f,t)Ω\Ω(f,t) = 0.
T ′
f,t = {ΩD : D ∈ T′
By (69), (79), (83), (85), (87) we have
f,t}, T ′′
f,t = {ΩD : D ∈ T′′
f,t}.
P(f,t) ∈ ST ′
f,t
(Ω),
P(f,t) ∈ ST ′′
f,t
(Ω).
Moreover, the partition T ′
Let h ∈ B Xp(Ω). Then
f,t refines the partition T ′′
f,t.
(85)
(86)
(87)
(88)
kP(f,t)h − P(f,t)hkp
p,q,T ′
f,t
(5),(69),(80),(83),(84),(85)
=
= Xη∈Wt
X
ξ∈VA
1 (η)∩V(Γt+1)\St+1
X
ξ∈VA
1 (η)∩V(Γt+1)\St+1
kPξh − Pηhkp
Yq(ΩD(ξ)
) .
Z0
(kh − Pξhkp
Yq(ΩD(ξ)
) + kh − Pηhkp
Yq(ΩD(ξ)
(11)
.
Z0
))
2−t(1− p
q );
. Xη∈Wt
i.e.,
kP(f,t)h − P(f,t)hkp,q,T ′
f,t
.
Z0
2−t( 1
p − 1
q ).
(89)
We set k′
t = ⌈n · 2−ε(t−t∗(n))⌉, where ε > 0 is a sufficiently small number (it is
chosen by Z0). For any t 6 t∗∗(n) we have
card T′
f,t
(80)
= card Vf,t
(1),(52),(53),(54),(77)
.
Z0
card St
(54)
.
Z0
(44),(45),(48)
card Tf,t,0
.
Z0
⌈2−tn⌉
(29)
≍
Z0
2−tn.
This together with (38), (88), (89), Theorem A and Lemma 5 implies that for some
c∗ = c∗(Z0) > 0
∗(n)
t∗∗(n)−1
Pt=t′
Pt=t∗(n)
.
t∗∗(n)−1
ek′
t(P(f,t) − P(f,t) : Xp(Ω) → Yq(Ω)) .
Z0
2−t( 1
p − 1
q )(2−tn)
1
q − 1
p · 2−c∗·2t−ε(t−t∗(n)) .
Z0
1
q − 1
p .
n
19
(90)
We claim that there exists a sequence {k′′
t }t′
∗(n)6t<t∗∗(n) such that
t∗∗(n)−1
Xt=t′
∗(n)
(k′′
t − 1) .
Z0
n,
t∗∗(n)−1
Xt=t′
∗(n)
ek′′
t ( P(f,t) : Xp(Ω) → Yq(Ω)) .
Z0
1
q − 1
p .
n
(91)
For each t′
∗(n) 6 t < t∗∗(n), 0 6 l 6 log nt we consider the partition Tf,t,l as
defined at Step 1. Let D ∈ Tf,t,l. We set
V∗
t,D = {ξ ∈ V(Γt) ∩ V(D) : VA
1 (ξ) ∩ V(Γt+1) 6= ∅}
and denote by D+ the subtree in A with vertex set
Let
By (48),
V(D+) = V(D) ∪(cid:16)∪ξ∈V∗
t,D
∪ξ′∈VA
1 (ξ)∩V(Γt+1)
V(Aξ′)(cid:17) .
Tf,t,l = {D+ : D ∈ Tf,t,l}.
card Tf,t,l .
Z0
2−lnt.
We claim that for any tree D ∈ Tf,t,l
(92)
(93)
(94)
card {D′
+ : D′ ∈ Tf,t,l±1 : V(D+) ∩ V(D′
+) 6= ∅} .
Z0
1.
(95)
Indeed, by (49), it is sufficient to check that if V(D+) ∩ V(D′
V(D′) 6= ∅. Let ξ ∈ V(D+)∩ V(D′
V(D′)) or ξ ∈ V(Aη), where η ∈ V(Γt) ∩ V(D) ∩ V(D′), VA
+) 6= ∅, then V(D) ∩
+). Then either ξ ∈ V( At) (therefore, ξ ∈ V(D)∩
1 (η) ∩ V(Γt+1) 6= ∅.
Assertion 2. We have Tf,t,0Γf,t
= T′′
f,t. If (T , ξ) ∈ Tf,t,0, ξ /∈ V(Γt), then
V(T+) ∩ V(Γf,t) = ∅.
Proof of Assertion 2.
1. Let (T , ξ) ∈ Tf,t,0. We claim that either there exists a tree D ∈ T′′
f,t such that
V(T+) ∩ V(Γf,t) = V(D) or V(T+) ∩ V(Γf,t) = ∅.
Case ξ ∈ V(Γt). By (54), we have ξ ∈ St. Let us show that
V(T+) ∩ V(Γf,t) = V( D(ξ))
(96)
and apply (86).
20
• We claim that V(T+) ∩ V(Γf,t) ⊃ V( D(ξ)). Since T′′
f,t is a partition of
the graph Γf,t, we have V( D(ξ)) ⊂ V(Γf,t). Let us prove that V( D(ξ)) ⊂
V(T+). Indeed, let ξ ∈ V( D(ξ))\V(T+). We set η∗ = min{η ∈ [ ξ, ξ] :
η /∈ V(T+)} > ξ. Denote by ζ∗ the direct predecessor of η∗. Then ζ∗ /∈
V(Aζ ′) (otherwise, η ∈ V(T+)). Hence, ζ∗ ∈
∪ζ∈V∗
V(Γt). If η∗ /∈ V(Γt), then η∗ ∈ V(Γt+1) (by Assumption 1, condition 1);
once again, we get η∗ ∈ V(T+). Thus, η∗ ∈ V(Γt) and η∗ is the minimal
vertex of some tree T ∈ Tf,t,0. Therefore, η∗
S, which
(54)
∈ St
1 (ζ)∩V(Γt+1)
∪ζ ′∈VA
t,T
(55),(56)
⊂
(58),(83),(84)
implies η∗
which leads to a contradiction.
/∈
V( D(ξ)). On the other hand, η∗ ∈ [ ξ, ξ] ⊂ V( D(ξ)),
• We claim that V(T+)∩V(Γf,t) ⊂ V( D(ξ)). We have V(Γf,t)
(86)
= ∪η∈St
V( D(η)).
If ζ ∈ V( D(η)) ∩ V(T+), then the vertices ξ and η are comparable. Since
ξ ∈ St ⊂ S, the case η < ξ is impossible by (58), (83), (84). Consequently,
η ∈ [ ξ, ζ] ⊂ V(T+); in addition, η ∈ St. By (54), η ∈ V(Γt) is the minimal
vertex of some tree in Tf,t,0. Therefore, η = ξ; i.e., ζ ∈ V( D(ξ)).
This completes the proof of (96).
Case ξ /∈ V(Γt). Then ξ ∈ V(Γt′), t′ < t. Let us check that V(T+) ∩ V(Γf,t) =
(86)
∈ V( D(η)) for some η ∈ St; i.e.,
∅. Indeed, let ξ ∈ V(T+) ∩ V(Γf,t). Then ξ
(54)
∈ V(Γt) is the minimal vertex of some tree from Tf,t,0 (this tree does not
η
coincide with T since ξ /∈ V(Γt) is the minimal vertex of T ). In addition, the
vertices ξ and η are comparable. The case η < ξ is impossible by Assumption
1 (see condition 1). Hence, η ∈ [ ξ, ξ] ⊂ V(T+); i.e., η ∈ V(T ), which leads to
a contradiction.
2. Let D(ξ) ∈ T′′
f,t. We claim that there exists a tree T ∈ Tf,t,0 such that
V( D(ξ)) = V(T+) ∩ V(Γf,t). Indeed, since ξ ∈ St by (86), we have ξ ∈ V(Γt)
and ξ is the minimal vertex of some tree T ∈ Tf,t,0 (see (54)). By (96),
V(T+) ∩ V(Γf,t) = V( D(η)) for some η ∈ St. In addition, ξ ∈ V(T+) ∩ V(Γf,t);
i.e., ξ ∈ V( D(η)) and η = ξ.
This completes the proof of Assertion 2. ⋄
Let D ∈ Tf,t,0, t′
∗(n) 6 t < t∗∗(n). If the minimal vertex of D does not belong to
∗(n)), by Assertion 2 we have V(D)∩V(Γf,t) = ∅. Hence, if V(D)∩V(Γf,t) 6= ∅,
V(Γt′
then Γt′
∗(n) ∩ D is a tree.
Denote by Tn
∗(n) ∩ D, D ∈ Tf,t,l. We have
f,t,l the partition of Γt′
Γt′
∗(n) formed by connected components of graphs
P(f,t)h = P Tn
f,t,0
h · χΩ(f,t),
h ∈ Yq(Ω).
(97)
21
It follows from (69), (79), (83), (85), (87), (88) and Assertion 2.
Given 0 6 l 6 log nt, we set
kt,l = ⌈n · 2−ε(t−t∗(n)+l)⌉,
k′′
(kt,l − 1),
t∗(n) 6 t < t∗∗(n),
t′
∗(n) 6 t < t∗∗(n).
(98)
t = 1 + P06l6log nt
Pt=t′
∗(n)
t∗∗(n)−1
n · 2−ε(t−t∗(n)),
Then k′′
t − 1 .
Z0
(k′′
t − 1) .
Z0
n.
From (51), (92) and (93) it follows that Tf,t,⌊log nt⌋ = {A}, Tn
f,t,⌊log nt⌋ = {At′
∗(n),j}j∈ Jt′
.
∗(n)
From (16) and (69) we get P Tn
f,t,⌊log nt ⌋
= 0. Hence, by (97) and (37),
ek′′
ekt,l(P Tn
t ( P(f,t) : Xp(Ω) → Yq(Ω)) 6
− P Tn
f,t,l+1
f,t,l
: Xp(Ω) → Yq(Ω(f,t))).
(99)
6 P06l<⌊log nt⌋
Denote by Tn
f,t,l+1, and either D′ ∈ Tf,t,l or D′′ ∈ Tf,t,l+1. We set
f,t,l the partition formed by trees D′ ∩ D′′, where D′ ∈ Tn
Tn
f,t,l, D′′ ∈
T n
f,t,l = {ΩD}D∈ Tn
f,t,l
,
T n
f,t,l = {E = ΩD ∩ Ω(f,t) : mes E > 0}D∈ Tn
f,t,l
.
(100)
Let h ∈ B Xp(Ω). We show that
(P Tn
f,t,l
h − P Tn
f,t,l+1
h)χΩ(f,t) ∈ S T n
f,t,l
(Ω).
(101)
Indeed, if D = D′ ∩ D′′, D′ ∈ Tn
f,t,l, D′′ ∈ Tn
f,t,l+1, then P Tn
f,t,l
(69)
∈ P(ΩD′),
hΩD′
P Tn
f,t,l+1
hΩD′′
(69)
∈ P(ΩD′′); therefore,
(P Tn
f,t,l
h − P Tn
f,t,l+1
h)ΩD∩Ω(f,t) ∈ P(ΩD ∩ Ω(f,t)).
(102)
We show that if D′ /∈ Tf,t,l and D′′ /∈ Tf,t,l+1, then (P Tn
h)ΩD∩Ω(f,t) = 0.
Indeed, in this case minimal vertices of the trees D′ and D′′ are equal and coincide
with ξt∗(n),j for some j ∈ Jt∗(n). From (69) it follows that (P Tn
h)ΩD = 0.
From (4) and (102) we get (101).
We claim that for any E′ ∈ T n
h−P Tn
h−P Tn
f,t,l+1
f,t,l, E′′ ∈ T n
f,t,l+1
f,t,l+1
f,t,l
f,t,l
card {E ∈ T n
f,t,l : E ⊂ E′} .
Z0
1,
card {E ∈ T n
f,t,l : E ⊂ E′′} .
Z0
1.
(103)
Let us check the first inequality (the second one is proved similarly). Let E = ΩD′,
f,t,l, E ∈ T n
D′ ∈ Tn
f,t,l is a partition, we have E = ΩD′∩D′′ ∩Ω(f,t),
f,t,l+1, V(D′) ∩ V(D′′) 6= ∅. There exist trees T ′ ∈ Tf,t,l and T ′′ ∈ Tf,t,l+1
D′′ ∈ Tn
f,t,l, E ⊂ E′. Since Tn
22
∗(n), respectively. Observe that the connected component of the graph T ′′ ∩ Γt′
such that D′ and D′′ are connected components of the graphs T ′ ∩ Γt′
∗(n) and T ′′ ∩
Γt′
∗(n)
whose vertex set intersects with V(D′) is unique. This together with (95) implies
(103).
From (94), (103) and the definition of Tn
f,t,l we get that
card T n
f,t,l .
Z0
2−lnt.
(104)
For any h ∈ B Xp(Ω) we have
k(P Tn
f,t,l
h − P Tn
f,t,l+1
h)χΩ(f,t)kp,q, T n
f,t,l
(5),(103)
.
Z0
. k(h − Qt′
∗(n)h − P Tn
f,t,l
h)χ Ut′
∗(n)
kp,q, T n
f,t,l
+
+k(h − Qt′
∗(n)h − P Tn
f,t,l+1
h)χ Ut′
∗(n)
kp,q, T n
f,t,l+1
(5),(69),(100)
=
(11)
.
Z0
(105)
(106)
.
X(D, ξ)∈ Tn
f,t,l
i.e.,
kh − Pξhkp
1/p
Yq(ΩD)
∗(n) .
Z0
+
X(D, ξ)∈ Tn
p )t∗(n) (29)
≍
Z0
. 2( 1
q − 1
p )t′
2( 1
q − 1
(log n)
1
q − 1
p ;
kh − Pξhkp
f,t,l+1
1/p
Yq(ΩD)
k(P Tn
f,t,l
h − P Tn
f,t,l+1
h)χΩ(f,t)kp,q, T n
f,t,l
(log n)
1
q − 1
p .
.
Z0
From (38), (101), (104), (105) and Lemma 5 we get that
ekt,l(P Tn
f,t,l
− P Tn
f,t,l+1
: Xp(Ω) → Yq(Ω(f,t))) .
Z0
1
q − 1
p ekt,l(Ist,l : lst,l
p → lst,l
q ) =: At,l,
. (log n)
where st,l ∈ N,
st,l 6 C∗ · 2−lnt
(44),(45)
6 C∗ · 2−l⌈n · 2−t⌉, C∗ = C∗(Z0) > 1.
For each t 6 t∗∗(n) we have
(98)
6
st,l
kt,l
C∗ · 2−l⌈n · 2−t⌉
n · 2−ε(l+t−t∗(n)) =: σ′
t,l,
23
σ′
t,l
(29)
≍
Z0
2−ln · 2−t
n · 2−ε(l+t−t∗(n)) = 2−l(1−ε)−t(1−ε)−εt∗(n) 6 1.
(107)
together with Theorem A implies that there exists γ0 = γ0(Z0) > 0 such that
decreases not slower than some geometric progression. This
The sequence (cid:8)σ′
t,l(cid:9)l∈Z+
t∗∗(n)−1
Xt=t′
∗(n) X06l<log nt
At,l
(106)
.
Z0
(log n)
1
q − 1
p
. (log n)
1
q − 1
p
1
q − 1
t,l
p
k
t∗∗(n)−1
t∗∗(n)−1
Xt=t′
∗(n) X06l<log nt
Xt=t∗(n)
q − 1
t,0
k
1
p
t∗∗(n)−1
Xt=t′
∗(n) X06l<log nt
s
1
q − 1
t,l 2
p
−
kt,l
st,l .
Z0
· (σ′
t,l)
1
q − 1
p · 2
− 1
σ′
t,l
(98),(107)
.
Z0
. (log n)
1
q − 1
p
· 2(t(1−ε)+εt∗(n))( 1
q − 1
p ) · 2−γ0·2t(1−ε)+εt∗(n)
(98)
.
Z0
. (log n)
i.e.,
1
q − 1
p · n
1
q − 1
p · 2( 1
q − 1
p )t∗(n) · 2−γ0·2t∗(n)
(29)
.
Z0
1
q − 1
p ;
n
t∗∗(n)−1
Xt=t′
∗(n) X06l<log nt
1
q − 1
p .
n
At,l .
Z0
(108)
From (99), (106) and (108) we get (91). This together with (90) yields (82).
Applying (81), we have (75). Taking into account the estimate obtained at Step 5,
we complete the proof of Lemma 3.
It remains to prove that
en ∞
Xm=0
( Qn,m+1 − Qn,m) : Xp(Ω) → Yq(Ω)! .
Z0
1
q − 1
p .
n
(109)
In [23] there were obtained order estimates for entropy numbers of diagonal
operators with weights of logarithmic type. First we give some notations.
Denote by Φ0 the class of non-decreasing functions ϕ : [1, ∞) → (0, ∞) that
satisfy the following condition: there exist c > 0 and α > 0 such that for any
1 6 s 6 t < ∞
ϕ(t)
ϕ(s)
1 + log s(cid:19)α
6 c(cid:18) 1 + log t
.
(110)
We set wϕ(s) = 1 for 0 6 s 6 1 and wϕ(s) = ϕ(s) for s > 1.
24
Let δ > 0, w : R+ → R+ be a continuous function, and let 0 < r, p 6 ∞. For
x = (xm,k)m,k∈Z+ we set
∞
kxlr(2δmlp(w))k :=
Xm=0
2mδr
Xk∈Z+
xm,kw(2−mk)p
1
r
r
p
(appropriately modified if p = ∞ or r = ∞). By lr(2δmlp(w)) we denote the space
of sequences x such that kxlr(2δmlp(w))k < ∞.
Theorem C. [23, p. 11]. Let 0 < p < q 6 ∞, 0 < r, s 6 ∞, δ > 0, ϕ ∈ Φ0, and let
(110) hold with α = 1
p − 1
q . Then
en(id : lr(2δmlp(wϕ)) → ls(lq)) ≍
p,q,r,s,ϕ
1
ϕ(2n)
.
Let us prove (109).
Given m > mt, we denote by Tt,m the partition of the set Gt formed by E′ ∩ E′′,
t,m =
E′ ∈ Tt,m, E′′ ∈ Tt,m+1 (the partitions Tt,m are defined at page 7). Let s′′
dim S Tt,m(Ω). Then there exists M = M(Z0) > 1 such that
(10),(19),(21),(24)
6
s′′
t,m
M · 2m−mt · 2γ∗2t
ψ∗(22t
).
(111)
Given t > 0, m′ > 0, we set
st,m′ = ⌈M · 2m′
· 2γ∗2t
ψ∗(22t
)⌉,
s∗
t,m′ =
sl,m′.
t−1
Xl=0
Denote ϕ(x) = (log(2 + x))
1
p − 1
q , x > 0. Let k = s∗
t,m′ + j, 1 6 j 6 st,m′. Then
2γ∗2t−1
ψ∗(22t−1
) .
Z0
2−m′
k .
Z0
2γ∗2t
ψ∗(22t
),
t > 1,
2−m′k .
Z0
1 if t = 0. Hence,
wϕ(2−m′
k)
(41)
≍
Z0
2( 1
p − 1
q )t,
k = s∗
t,m′ + j,
1 6 j 6 st,m′.
(112)
From Lemma 5 it follows that there is an isomorphism At,m : S Tt,m(Ω) → Rs′′
t,m
such that
kAt,mk
s′′
t,m
(Gt)→l
p
1,
.
Z0
Yp,q, Tt,m
kA
−1
t,mk
s′′
t,m
q →Yq(Gt)
l
1.
.
Z0
(113)
25
Let us define the operator A : Xp(Ω) → l∞(2δ∗m′lp(wϕ)) as follows. Consider a
function f ∈ Xp(Ω). Then ( Qn,m′+1f − Qn,m′f )Gt
(20),(32)
∈
S Tt,mt+m′ (Ω). Let
At,mt+m′(( Qn,m′+1f − Qn,m′f )Gt) = (cm′,t,j)
s′′
t,mt+m′
j=1
.
We set
Then
(Af )m′,s∗
t,m′ +j =(cid:26) cm′,t,j
0 for
for 1 6 j 6 s′′
s′′
t,mt+m′ + 1 6 j 6 st,m′ or t < t0.
t,mt+m′, t > t0,
kAf kl∞(2δ∗m′ lp(wϕ)) = sup
m′>0
wϕ(2−m′
. sup
m′>0
2δ∗m′
Xt>t0
2δ∗m′
Xk∈Z+
Xj=1
2(1− p
q )t
s′′
t,mt+m′
cm′,t,jp
1
p
(112)
.
Z0
k)(Af )m′,kp
1/p
=: N.
From (22), (23) and (24) it follows that for m > mt
kPt,m+1 − Pt,mk Xp(Gt)→Yp,q, Tt,m
(Gt) .
Z0
Hence, by (32), (113) we get
2−δ∗(m−mt) · 2( 1
q − 1
p )tkf kXp(Gt).
s′′
t,mt+m′
Xj=1
cm′,t,jp .
Z0
2−δ∗pm′
2( p
q −1)tkf kp
Xp(Gt);
therefore,
Thus,
N .
Z0 Xt>t0
kf kp
Xp(Gt)!1/p
= kf kXp(Ω).
kAk Xp(Ω)→l∞(2δ∗m′ lp(wϕ)) .
Z0
1.
(114)
Let us define the operator K : l1(lq) → Yq(Ω) by formula
K((cm′,s∗
t,m′ +j)m′∈Z+,t∈Z+,16j6st,m′ ) = Xm′∈Z+Xt>t0
A
−1
t,mt+m′((cm′,s∗
t,m′ +j)
s′′
t,mt+m′
j=1
).
26
Since the sets Gt do not overlap pairwise, we have
kK((cm′,s∗
t,m′ +j)m′∈Z+,t∈Z+,16j6st,m′ )kYq(Ω) 6
∞
6
Xm′=0 ∞
Xt=t0
Xj=1
Xt∈Z+
st,m′
cm′,s∗
t,m′ +jq
∞
Xm′=0
.
Hence,
kA
−1
t,mt+m′((cm′,s∗
t,m′ +j)
s′′
t,mt+m′
j=1
)kq
Yq(Gt)!1/q
(113)
.
Z0
1/q
= k(cm′,s∗
t,m′ +j)m′∈Z+,t∈Z+,16j6st,m′ kl1(lq).
kKkl1(lq)→Yq(Ω) .
Z0
1.
(115)
Let id : l∞(2δ∗m′lp(wϕ)) → l1(lq) be the identity operator. Then
en ∞
Xm=0
( Qn,m+1 − Qn,m) : Xp(Ω) → Yq(Ω)! =
= en(K ◦ id ◦ A : Xp(Ω) → Yq(Ω))
(38),(114),(115)
.
Z0
. en(id : l∞(2δ∗m′
lp(wϕ)) → l1(lq)) =: En.
From Theorem C we get En .
Z0
1
q − 1
p .
n
Applying Lemmas 1, 2, 3 together with (30), (33), (34), (109), we complete the
proof of Theorem 1.
Remark 1. Suppose that instead of Assumption 2 the following condition holds: for
any ξ ∈ V(A) the set F (ξ) is the atom of mes. Then the assertion of Theorem 1
holds as well.
3 Some particular cases
Let (T , ξ0) be a tree, and let u, w : V(T ) → R+. We define the summation operator
Su,w,T by
u(ξ′)f (ξ′),
ξ ∈ V(T ),
f : V(T ) → R.
Let 1 6 p, q 6 ∞. By Sp,q
T ,u,w we denote the minimal constant C in the inequality
Su,w,T f (ξ) = w(ξ)Xξ′6ξ
q
u(ξ′)f (ξ′)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
wq(ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xξ′6ξ
Xξ∈V(T )
1/q
6 C
Xξ∈V(T )
27
1/p
f (ξ)p
, f : V(T ) → R
(appropriately modified for p = ∞ or q = ∞).
j=0
∪N
j (ξ0).
Given a tree (T , ξ0), N ∈ Z+, we denote by [T ]6N a tree with vertex set
VT
Let (A, ξ0) be a tree, let (1) hold, and let the measure space (Ω, Σ, mes), the
partition Θ, the bijection F : V(A) → Θ and the spaces Xp(Ω), Yq(Ω), P(Ω) be as
defined at the page 3. Assumptions 1 -- 3 will be replaced by the following conditions.
Assumption A. There exist functions u, w : V(A) → (0, ∞) and a constant
c2 > 1 with the following property: for any vertex ξ∗ ∈ V(A) there exists a linear
continuous projection Pξ∗ : Yq(Ω) → P(Ω) such that for any vertex ξ > ξ∗ and for
any function f ∈ Xp(Ω)
kf − Pξ∗f kYq( F (ξ)) 6 c2w(ξ) Xξ∗6ξ′6ξ
u(ξ′)kf kXp( F (ξ′)).
(116)
Assumption B. There exists a number δ∗ > 0 such that for any vertex ξ ∈ V(A)
and for any n ∈ N, m ∈ Z+ there exists 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 c2(2mn)−δ∗u(ξ)w(ξ)kf kXp(E).
(117)
3. For any E ∈ Tm,n(G)
card {E′ ∈ Tm±1,n(G) : mes(E ∩ E′) > 0} 6 c2.
(118)
Let V(A) = {ηj,i}j>jmin, i∈ Ij , where jmin > 0, Ijmin = {1}. Suppose that ηjmin,1 is
the minimal vertex of A and VA
j−jmin(ηjmin,1) = {ηj,i}i∈ Ij for any j > jmin.
Assumption C. There exist numbers θ > 0, γ ∈ R, κ > θ
q , αu, αw ∈ R, c3 > 1,
m∗ ∈ N and an absolutely continuous function τ : (0, ∞) → (0, ∞) such that
lim
t→+∞
tτ ′(t)
τ (t) = 0 and the following conditions hold.
1. If κ = θ
q , then αw > 1−γ
q .
2. If κ > θ
q , then αu + αw = 1
p − 1
q ; if κ = θ
q , then αu + αw = 1
p .
3. For any j′ > j > jmin and for any vertex ξ ∈ VA
j−jmin(ηjmin,1)
card VA
j′−j(ξ) 6 c3 · 2θm∗(j′−j) jγτ (m∗j)
j′γτ (m∗j′)
.
(119)
28
4. For any j > jmin, i ∈ Ij
u(ηj,i) = uj = 2κm∗j(m∗j)−αu, w(ηj,i) = wj = 2−κm∗j(m∗j)−αw .
(120)
5. Let κ = θ
q . Then there exists a tree A with the minimal vertex ζ0, which satisfies
the following conditions.
(a) For any j′ > j > jmin and for any vertex ξ ∈ V
A
j−jmin(ζ0)
c−1
3
· 2θm∗(j′−j) jγτ (m∗j)
j′γτ (m∗j′)
6 card V
A
j′−j(ξ) 6 c3 · 2θm∗(j′−j) jγτ (m∗j)
j′γτ (m∗j′)
.
(121)
(b) Let {uj}j>jmin ⊂ (0, ∞), {wj}j>jmin ⊂ (0, ∞) be arbitrary sequences.
Define the functions u, w : V(A) → (0, ∞), u, w : V( A) → (0, ∞) by
uVA
j−jmin
(ηjmin ,1) ≡ uj,
uV
A
j−jmin
(ζ0) ≡ uj,
wVA
j−jmin
(ηjmin ,1) ≡ wj,
wV
A
j−jmin
(ζ0) ≡ wj,
j > jmin.
Then for each N > j > jmin, i ∈ Ij there exists a vertex ξ ∈ V
such that
Sp,q
[Aηj,i ]6N−j ,u, w 6 c3Sp,q
[ A ξ]6N−j ,u, w
.
A
j−jmin(ζ0)
Denote Z0 = (p, q, c1, c2, c3, θ, γ, κ, αu, αw, m∗, δ∗, τ ).
Let ξ∗ ∈ V(A), and let D ⊂ A be a subtree with the minimal vertex ξ∗. Then
from Assumption A it follows that
kf − Pξ∗f kYq(ΩD) 6 c2Sp,q
D,u,wkf kXp(ΩD).
(122)
From Theorem F in [36], Theorem 3.6 in [38] and Assumption C it follows that if
ξ∗ ∈ VA
j−jmin(ξ0), then in the case κ > θ
q
Sp,q
D,u,w .
Z0
sup
s>j
usws ≍
Z0
(m∗j)−αu−αw = (m∗j)
1
q − 1
p ,
(123)
and in the case κ = θ
q
Sp,q
D,u,w .
Z0
sups>j s
Pi=j
(m∗i)−p′αu · 2p′κm∗i!
1
p′ (cid:18)Pi>s
(m∗i)−αwq · sγτ (m∗s)
iγτ (m∗i) · 2−qκm∗s(cid:19)
1
q
.
Z0
. (m∗j)−αu−αw+ 1
q = (m∗j)
1
q − 1
p .
Notice that the proof of (124) depends on condition 5 of Assumption C.
29
(124)
From Assumption B and conditions 2, 4 of Assumption C it follows that if ξ ∈
j−jmin(ξ0), G = F (ξ), E ∈ Tm,n(G), then
VA
kf − PEf kYq(E) 6 c2 · (2mn)−δ∗(m∗j)−αu−αwkf kXp(E) 6
6 c2 · (2mn)−δ∗(m∗j)
1
q − 1
p kf kXp(E).
(125)
We construct the partition {At,i}t>t0, i∈ Jt as follows. Let
t0 = min{t ∈ Z+ : 2t > m∗jmin}.
Given t > t0, we denote by Γt the maximal subgraph in A on the vertex set
V(Γt) = {ηj,s : 2t−1 6 m∗j < 2t, s ∈ Ij},
(126)
and by At,i, i ∈ Jt, the connected components of the graph Γt. By ξt,i we denote the
minimal vertex of the tree At,i. Then
card V(Γt)
(119)
.
Z0
2θ·2t
2−γtτ −1(2t).
(127)
Thus, Assumption 3 holds with νt = 2θ·2t2−γtτ −1(2t). From (125) we obtain that
Assumption 2 holds.
Recall the notations Jt,D = {i ∈ Jt : V(At,i) ∩ V(D) 6= ∅}, Dt,i = D ∩ At,i,
where D ⊂ A is a subtree.
In order to obtain Assumption 1, it is sufficient to prove the following assertion.
Lemma 7. Let D be a subtree in A rooted at ξ∗. Then
kf − Pξ∗f kq
Yq(ΩD) .
Z0
∞
Xt=t0
2(1− q
p )t Xi∈Jt,D
kf kq
Xp(ΩDt,i ).
(128)
Proof. By (116),
kf − Pξ∗f kq
Yq(ΩD) .
Z0 Xξ∈V(D)
wq(ξ) Xξ∗6ξ′6ξ
u(ξ′)kf kXp( F (ξ′))!q
.
Let κ > θ
proof of Lemma 5.1 in [41], we get that
q . Then αu + αw = 1
p − 1
q (see condition 2 of Assumption C). Repeating the
kf − Pξ∗f kq
Yq(ΩD) .
Z0 Xξ∈V(D)
wq(ξ)uq(ξ)kf kq
Xp( F (ξ))
.
If ξ = ηj,i, 2t−1 6 m∗j < 2t, then uq(ξ)wq(ξ)
together with the condition p < q implies (128).
(120)
= (m∗j)− q
p +1 ≍
Z0
2(1− q
p )t. This
30
Now we consider the case κ = θ
q .
Let ξ∗ ∈ V(At,j0), and let ξ ∈ V(Dt+l,jl). Then there exists a sequence { ξt+s,js}l
s=1
such that ξ∗ < ξt+1,j1 < ξt+2,j2 < · · · < ξt+l,jl. Denote
ξt,j0 = ξ∗,
ξt+s,js = ξt+s,js, 1 6 s 6 l.
(129)
We have
u(ξ′)f (ξ′) =
Xξ∗6ξ′6ξ
l−1
Xs=0
Xξt+s,js
6ξ′< ξt+s+1,js+1
u(ξ′)kf kXp( F (ξ′)) + Xξt+l,jl
6ξ′6ξ
u(ξ′)kf kXp( F (ξ′)).
(130)
(131)
Let αw = αw − 1
q . By condition 2 of Assumption C,
We have
αu + αw =
1
p
−
1
q
.
wq(ξ)
Xξ∈V(Dt+l, jl
)
(119),(120),(126)
.
Z0
. X2t+l−16m∗j<2t+l
m∗j !γ
2−θm∗j(m∗j)−αwq · 2θ(m∗j−2t+l−1) 2t+l−1
τ (2t+l−1)
τ (m∗j)
(41)
.
Z0
. 2−κq·2t+l−1 · 2(t+l)(−αwq+1) = 2−κq·2t+l−1 · 2−(t+l)αwq,
which implies
kf − Pξ∗f kq
Yq(ΩDt+l,jl
)
(116)
.
Z0 Xξ∈V(Dt+l,jl
)
wq(ξ) Xξ∗6ξ′6ξ
. Xξ∈V(Dt+l,jl
)
+ Xξ∈V(Dt+l,jl
. 2−qκ·2t+l−1
u(ξ′)f (ξ′)!q
u(ξ′)kf kXp( F (ξ′))
q
q
(130)
.
Z0
+
6ξ′< ξt+s+1,js+1
l−1
wq(ξ)
Xs=0
Xξt+s,js
wq(ξ)
Xξt+l,jl
· 2−q αw(t+l)
Xs=0
6ξ′6ξ
l−1
)
Xξt+s,js
31
.
Z0
(124),(126)
u(ξ′)kf kXp( F (ξ′))
u(ξ′)kf kXp( F (ξ′))
6ξ′< ξt+s+1,js+1
q
+
+2(1− q
p )(t+l)kf kq
Xp(Dt+l,jl
);
i.e.,
kf − Pξ∗f kq
Yq(ΩDt+l,jl
+2−qκ·2t+l−1 · 2−q αw(t+l)
Ps=0
l−1
Denote
2(1− q
p )(t+l)kf kq
Xp(Dt+l,jl
)+
) .
Z0
ξt+s,js
P
6ξ′< ξt+s+1,js+1
q
u(ξ′)kf kXp( F (ξ′))
.
{ζt,i : i ∈ I ′
t} = {ζ ∈ Vmax(D ∩ Γt) : VD
1 (ζ) 6= ∅}.
By (126),
∀i ∈ I ′
t ∃s ∈ I⌈2t/m∗⌉−1 :
ζt,i = η⌈2t/m∗⌉−1,s.
(132)
(133)
(134)
Let us define the tree D with vertex set {ξ∗} ∪(cid:0)∪t>t{ζt,i : i ∈ I ′
order on V( D) is defined as follows. We set
t}(cid:1). The partial
V
D
1 (ζt,i) = {ζt+1,j : ζt+1,j ∈ Dζt,i}.
If ξ∗ /∈ {ζt,i : i ∈ I ′
t}, then we set V
D
1 (ξ∗) = {ζt,i : i ∈ I ′
t}.
Further, we define the functions ϕ, u, w : V( D) → R+. We set
ϕ(ζt,i) =
Xξ′∈V(Γt)∩V(D), ξ′6ζt,i
u(ξ′)
u(ζt,i)
kf kXp( F (ξ′)),
(135)
w(ζt,i) = 2−κ·2t
· 2− αwt,
u(ζt,i) = 2κ·2t
≍
Z0
t}, then we set ϕ(ξ∗) = w(ξ∗) = u(ξ∗) = 0.
· 2−αut (120),(134)
If ξ∗ /∈ {ζt,i : i ∈ I ′
u(ζt,i).
(136)
Let t > t, j ∈ Jt,D. If t > t, then ξt,j ∈ V(D). We take i ∈ I ′
t−1 such that
ξt,j ∈ VD
1 (ζt−1,i). Then
kf − Pξ∗f kq
Yq(Dt,j)
(132),(136)
.
Z0
2(1− q
p )tkf kq
Xp(Dt,j) + wq(ζt−1,i)
Xζ ′∈V( D), ζ ′6ζt−1,i
q
u(ζ ′)ϕ(ζ ′)
(137)
.
Notice that card VD
1 (ζt,i)
(119)
.
Z0
1. Summing (137) over all t > t, j ∈ Jt,D, we obtain
kf − Pξ∗f kq
∞
2(1− q
Pt=t
Yq(ΩD ) .
Z0
p )t Pj∈Jt,D
wq(ζt,i) Pζ ′∈V( D), ζ ′6ζt,i
kf kq
Xp(ΩDt,j )+
u(ζ ′)ϕ(ζ ′)!q
+
∞
Pt=tPi∈I ′
t
(138)
.
32
Let us estimate the second summand (we denote it by L).
Let u(ζ) = u1(ζ)u2(ζ), u1(ζt,i) = 2εt, where ε = ε(Z0) > 0 is sufficiently small
t}, then we set u1(ξ∗) = u2(ξ∗) = 0. By
(it will be chosen later); if ξ∗ /∈ {ζt,i : i ∈ I ′
Holder's inequality,
L = Xζ∈V( D)
6
u1(ζ ′)u2(ζ ′)ϕ(ζ ′)!q
wq(ζ) Xζ ′6ζ
1 (ζ ′)!
wq(ζ) Xζ ′6ζ
2(ζ ′)ϕq(ζ ′) = Xζ ′∈V( D)
Xζ ′6ζ
2(ζ ′)ϕq(ζ ′)Xζ>ζ ′
uq′
uq
uq
q
q′
uq
2(ζ ′)ϕq(ζ ′) .
Z0
6 Xζ∈V( D)
1(ζ)Xζ ′6ζ
wq(ζ)uq
. Xζ∈V( D)
wq(ζ)uq
1(ζ) =: M.
By (134), there exists s ∈ I⌈2t/m∗⌉−1 such that
card V
D
l (ζt,i) = card {i′ ∈ I ′
t+l : ζt+l,i′ > ζt,i} 6
6 card VA
⌈2t+l/m∗⌉−⌈2t/m∗⌉(η⌈2t/m∗⌉−1,s)
(119)
.
Z0
2θ·2t+l2−γ(t+l)τ −1(2t+l)
2θ·2t2−γtτ −1(2t)
.
This together with relations κ = θ
yields that for sufficiently smalll ε > 0
q , αw > 1−γ
q
(see condition 1 of Assumption C)
wq(ζ)uq
1(ζ)
Xζ>ζt,i
(136)
.
Z0 Xl>0
2−qκ·2t+l
· 2−q αw(t+l) · 2qε(t+l) 2θ·2t+l2−γ(t+l)τ −1(2t+l)
2θ·2t2−γtτ −1(2t)
=
= 2−θ·2t
Thus,
· 2γtXl>0
M .
Z0 Xζ ′∈V( D)
i.e.,
∞
Xt=t Xi∈I ′
t
2−q(αw+ γ
q −ε)(t+l) τ (2t)
τ (2t+l)
(41)
.
Z0
2−qκ·2t
· 2−q(αw−ε)t (136)
= wq(ζt,i)uq
1(ζt,i).
wq(ζ ′)uq
1(ζ ′)uq
2(ζ ′)ϕq(ζ ′) = Xζ ′∈V( D)
p )t Xi∈I ′
2(1− q
t: ζt,i∈V(Dt,j )
∞
=
Xt=t Xj∈Jt,D
wq(ζ ′)uq(ζ ′)ϕq(ζ ′)
(131),(136)
=
ϕq(ζt,i);
wq(ζt,i)
Xζ ′∈V( D), ζ ′6ζt,i
q
u(ζ ′)ϕ(ζ ′)
∞
.
Z0
Xt=t Xj∈Jt,D
33
2(1− q
p )t Xi∈I ′
t: ζt,i∈V(Dt,j)
ϕq(ζt,i).
(139)
It remains to prove that for any t > t, j ∈ Jt,D
Xi∈I ′
t: ζt,i∈V(Dt,j )
ϕq(ζt,i) .
Z0
kf kq
Xp(ΩDt,j ).
Then (138), (139), (140) imply (128).
By (129) and (135),
Xi∈I ′
t: ζt,i∈V(Dt,j )
t: ζt,i∈V(Dt,j)
ϕq(ζt,i) = Xi∈I ′
u−q(ζt,i)
= Xi∈I ′
t: ζt,i∈V(Dt,j )
q
kf kXp( F (ξ′))
q
=: S.
u(ξ′)
u(ζt,i)
Xξt,j 6ξ′6ζt,i
u(ξ′)kf kXp( F (ξ′))
Xξt,j 6ξ′6ζt,i
From (133) it follows that ζt,i ∈ Vmax(At,j).
We denote A′
. Then V(Dt,j) ⊂ V(A′
t,j = (At,j) ξt,j
t,j). Given σ > 0, we set
(140)
=
Then
σ for
w(σ)(ξ) =(cid:26) u−1(ξ)
S 6hSp,q
A′
t,j,u,w(σ)iq
for
ξ ∈ V(A′
ξ ∈ Vmax(A′
t,j),
t,j)\Vmax(A′
t,j).
kf kXp(ΩDt,j ).
We claim that for sufficiently small σ > 0 the estimate Sp,q
A′
t,j,u,w(σ)
1 holds. This
.
Z0
implies (140).
Let ξt,j ∈ VA
s∗−jmin(ηjmin,1). From (126) and (129) it follows that
2t−1 6 m∗s∗ < 2t.
(141)
t,j = [A ξt,j
]6⌈2t/m∗⌉−1−s∗ (since ξt+1,i ∈ VA
(ηjmin,1), i ∈ Jt). By
Moreover, A′
A
s∗−jmin(ζ0) such that for
condition 5b of Assumption C, there exists a vertex ζ ∈ V
the tree At,j := [ Aζ]6⌈2t/m∗⌉−1−s∗ and for the functions u, w(σ) : V( At,j) → (0, ∞)
defined by
⌈2t/m∗⌉−jmin
u(ξ) = us,
ξ ∈ V
A
s−s∗(ζ0), w(ξ) =( u−1
⌈2t/m∗⌉−1,
ξ ∈ V
σ,
ξ ∈ V
A
s−s∗(ζ0),
A
⌈2t/m∗⌉−1−s∗(ζ0),
s < ⌈2t/m∗⌉ − 1,
(142)
the inequality Sp,q
A′
t,j ,u,w(σ)
Sp,q
At,j,u,w(σ)
.
Z0
holds.
From (121) it follows that we can apply Theorem 3.6 in [38] and estimate
Sp,q
At,j,u,w(σ)
from above. For sufficiently small σ = σ(Z0, t) > 0 we get
Sp,q
At,j,u,w(σ)
(121),(142)
.
Z0
sup
s∗6s<⌈2t/m∗⌉ s
Xl=s∗
up′
l !1/p′
×
34
×
.
⌈2t/m∗⌉−2
Xl=s
s∗6s6⌈2t/m∗⌉−1 s
Xl=s∗
sup
Since θ = qκ, we have
σq · 2θm∗(l−s) sγτ (m∗s)
lγτ (m∗l)
+ u−q
(120),(141)
1/q
2γtτ (2t)
⌈2t/m∗⌉−1 · 2θ(2t−m∗s) (m∗s)γτ (m∗s)
(cid:16)2−qκ·2t · 2qαut · 2θ(2t−m∗s)(cid:17)1/q
.
Z0
.
2p′κm∗l · (m∗l)−αup′!1/p′
Sp,q
At,j,u,w(σ)
.
Z0
sup
s∗6s6⌈2t/m∗⌉−1
2κm∗s · (m∗s)−αu · 2−κm∗s · 2αut
(141)
1.
.
Z0
This completes the proof.
Thus, Assumptions 1, 2, 3 hold, and by Theorem 1 we get (12).
Now we suppose that instead of Assumption C the following condition holds.
Assumption D. There exist numbers m∗ ∈ N, κ > 0, γ 6 0, ν ∈ R, α∗ ∈ R,
λu ∈ R, λw ∈ R such that λu + λw = 1
q and the following assertions hold.
p − 1
1. For any j′ > j > jmin and for any vertex ξ ∈ VA
card VA
j′−j(ξ) 6 c3
2. For any j > jmin, i ∈ Ij
j−jmin(ξ0)
(m∗j)γ log(m∗j)ν
(m∗j′)γ log(m∗j′)ν .
u(ηj,i) = uj = 2κm∗j(m∗j)α∗ log(m∗j)−λu,
w(ηj,i) = wj = 2−κm∗j(m∗j)−α∗ log(m∗j)−λw.
The partition {At,i}t>t0, i∈ Jt is defined as follows. Let
t0 = min{t ∈ Z+ : 22t
> m∗jmin}.
(143)
(144)
Given t > t0, we denote by Γt the maximal subgraph in A on the vertex set {ηj,s :
22t−1 6 m∗j < 22t, s ∈ Ij}, and by At,i, i ∈ Jt, the connected components of Γt.
Then
card V(Γt) .
Z0
2(1−γ)2t
2−νt.
(145)
Repeating the proof of Lemma 5.1 in [41] and taking into account that p < q,
we get the following assertion.
Lemma 8. Let D be a subtree in A, and let ξ∗ be its minimal vertex. Then
kf − Pξ∗f kq
Yq(D) .
Z0
∞
Xt=t0
2(1− q
p )t Xi∈Jt,D
35
kf kq
Xp(ΩDt,i ).
(146)
From Assumption B, (145) and (146) we obtain Assumptions 1, 2, 3.
Remark 2. Suppose that Assumption B is replaced by the following condition: for
any ξ ∈ V(A) the set F (ξ) is the atom of mes. Then the assertion of Theorem 1
holds as well (see Remark 1).
4 Estimates for entropy numbers of embeddings of
weighted Sobolev spaces
Let us define the weighted Sobolev class W r
Lq,v(Ω).
p,g(Ω) and the weighted Lebesgue space
Let Ω ⊂ Rd be a bounded domain, and let g, v : Ω → (0, ∞) be measurable
functions. For each measurable vector-valued function ψ : Ω → Rl, ψ = (ψk)16k6l,
and for each p ∈ [1, ∞), we put
max
16k6l
kψkLp(Ω) =(cid:13)(cid:13)(cid:13)
=
ZΩ
f defined on Ω we write ∇rf = (cid:16)∂rf /∂xβ(cid:17)β=r
Let β = (β1, . . . , βd) ∈ Zd
ψk(cid:13)(cid:13)(cid:13)p
max
16k6l
1/p
.
ψk(x)p dx
+ := (N ∪ {0})d, β = β1 + . . . + βd. For any distribution
(here partial derivatives are taken
in the sense of distributions), and denote by lr,d the number of components of the
vector-valued distribution ∇rf . We set
W r
p,g(Ω) =(cid:8)f : Ω → R(cid:12)(cid:12) ∃ψ : Ω → Rlr,d : kψkLp(Ω) 6 1, ∇rf = g · ψ(cid:9)
∇rf
(cid:16)we denote the corresponding function ψ by
g (cid:17),
kf kLq,v(Ω)=kf kq,v=kf vkLq(Ω),
Lq,v(Ω) = {f : Ω → R kf kq,v < ∞} .
We call the set W r
∇rf ∈ Lloc
1 (Ω).
p,g(Ω) a weighted Sobolev class. Observe that if g ∈ Lloc
p′ (Ω), then
For x ∈ Rd and a > 0 we shall denote by Ba(x) 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 a
number 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∗,
dγx(t)
dt
(cid:12)(cid:12)(cid:12)
= 1 a.e.,
36
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 [2]). As examples of such domains we can take
1. domains with Lipschitz boundary;
2. the Koch's snowflake;
3. domains Ω = ∪0<t6T int Bct(γ(t)), where γ : [0, T ] → Rd is a curve with natural
parametrization and c > 0.
Domains with zero inner angles do not satisfy the John condition.
We denote by H the set of all nondecreasing positive functions defined on (0, 1].
Definition 4. (see [5]). Let Γ ⊂ Rd be a nonempty compact set and 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
c−1
∗ h(t) 6 µ(Bt(x)) 6 c∗h(t)
(147)
for any x ∈ Γ and t ∈ (0, 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 be the Koch curve. Then Γ is an h-set with h(t) = tlog 4/ log 3
(see [30, p. 66 -- 68]).
Let Ω ∈ FC(a) be a bounded domain, and let Γ ⊂ ∂Ω be an h-set. Below we
consider a function h ∈ H which has the following form near zero:
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.
(148)
(149)
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−αvρv( log t),
(150)
37
where ρg and ρv are absolutely continuous functions such that
tρ′
g(t)
ρg(t)
→
t→+∞
0,
tρ′
v(t)
ρv(t)
→
t→+∞
0.
Moreover, assume that
βv <
d − θ
q
or βv =
d − θ
q
, αv >
1 − γ
q
.
(151)
(152)
Without loss of generality we may consider Ω ⊂(cid:0)− 1
Denote
β = βg + βv, α = αg + αv,
2 , 1
2(cid:1)d.
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.
Denote by Pr−1(Rd) the space of polynomials on Rd of degree not exceeding r−1.
For a measurable set E ⊂ Rd we put
Pr−1(E) = {f E : f ∈ Pr−1(Rd)}.
Observe that W r
p,g(Ω) ⊃ Pr−1(Ω).
In Theorems 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)
(153)
∇rf
g (cid:13)(cid:13)(cid:13)(cid:13)Lp(Ω)
g (cid:13)(cid:13)(cid:13)Lp(Ω)
p,g(Ω), W r
kf − P f kLq,v(Ω) 6 M(cid:13)(cid:13)(cid:13)(cid:13)
p,g(Ω) :=(cid:13)(cid:13)(cid:13)
α0 :=(cid:26) α for βv < d−θ
α − 1
q
∇rf
for βv = d−θ
q .
q ,
(see [40 -- 43]).
p,g(Ω) = span W r
Denote W r
p,g(Ω) be equipped with norm kf k W r
W r
p,g(Ω) = {f − P f :
f ∈ W r
. Denote by I : W r
p,g(Ω)}. Let
p,g(Ω) →
Lq,v(Ω) the embedding operator. From (153) it follows that it is continuous.
First we consider the case 0 < θ < d. We set
In [37] the estimates for entropy numbers en(I : W r
under the following conditions. In the case δ − β > θ(cid:16) 1
d 6= δ−β
Now we obtain estimates for p < q, β − δ = 0, α0 = 1
θ . In the case δ − β = θ(cid:16) 1
p(cid:17)+
q − 1
δ
p − 1
q .
p,g(Ω) → Lq,v(Ω)) were obtained
we assumed that
q − 1
p(cid:17)+
we assumed that α0 6= 1
p − 1
q for p < q.
38
Theorem 2. Suppose that the conditions (148) -- (152) hold, ρg ≡ 1, ρv ≡ 1. Let
0 < θ < d, p < q, β = δ, and let α0 = 1
p − 1
q . Then
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
1
q − 1
p .
n
Now we consider the case θ = 0, βv < d
q . We assume that
ρg(t) = log t−λg,
ρv = log t−λv,
τ (t) = log tν.
(154)
Denote λ = λg + λv.
In [37] estimates of entropy numbers were obtained in the following cases:
q − 1
,
1. δ − β > θ(cid:16) 1
2. δ − β = θ(cid:16) 1
3. δ − β = θ(cid:16) 1
p(cid:17)+
p(cid:17)+
p(cid:17)+
q − 1
q − 1
, α > (1 − γ)(cid:16) 1
, α = (1 − γ)(cid:16) 1
q − 1
q − 1
p(cid:17)+
p(cid:17)+
d 6= α
, δ
1−γ ,
, and if p < q, then λ 6= 1
p − 1
q .
Here we obtain the estimates for p < q, δ − β = 0, α = 0, λ = 1
p − 1
q .
Theorem 3. Let the conditions (148), (150), (154) hold, and let θ = 0, β − δ = 0,
βv < d
q , α = 0, p < q, λ = 1
q . Then
p − 1
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
1
q − 1
p .
n
Proof of Theorems 2 and 3. The lower estimates are proved similarly as in [37].
Let us obtain upper estimates. To this end, we apply Theorem 1. Consider
the tree A with vertex set {ηj,i}j>jmin,i∈ Ij and the partition of the domain Ω into
subdomains Ω[ηj,i], as defined in [40], [41]. We set F (ηj,i) = Ω[ηj,i]. Let the number
s = s(a, d) > 4 be as defined in [40]. Repeating arguments from the proof of Theorem
2 in [41] (without summation over vertices ξ ∈ V(Aξ∗)), we obtain that Assumption
A holds with u(ηj,i) = ϕg(2−sj) · 2−(r− d
q . Assumption
B holds with δ∗ = δ
d (see [39], [36, Lemma 8], [1]). If θ > 0, then Assumption C
holds; condition 5 of this assumption follows from Lemma 2 in [42]. If θ = 0, then
Assumption D holds. In both cases we have κ = d
q − βv, αu = αg, αw = αv. In the
case θ = 0 we have λu = λg, λw = λv. By Theorem 1, we obtain the desired upper
estimates of entropy numbers.
p )sj, w(ηj,i) = ϕv(2−sj) · 2− dsj
39
5 Estimates for entropy numbers of two-weighted
summation operators on a tree
Applying Remark 2, we obtain estimates for entropy numbers of weighted summation
operators on trees.
Let G be a graph. Given a function f : V(G) → R, we set
1/p
kf klp(G) =
Xξ∈V(G)
f (ξ)p
for 1 6 p < ∞,
kf kl∞(G) = sup
ξ∈V(G)
f (ξ). (155)
Denote by lp(G) the space of functions f : V(G) → R with finite norm kf klp(G).
Let A be a tree, V(A) = {ηj,i : j ∈ Z+,
i ∈ Ij}, let η0,1 be the minimal vertex
j (η0,1) = {ηj,i}i∈Ij for any j ∈ Z+. Suppose that for some c∗ > 1,
of A, and let VA
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+,
where the function h ∈ H is defined by (148) near zero. Let u, w : V(A) → (0, ∞),
u(ηj,i) = uj, w(ηj,i) = wj, j ∈ Z+, i ∈ Ij,
uj = 2κm∗j(m∗j + 1)−αu, wj = 2−κm∗j(m∗j + 1)−αw.
(156)
In addition, we suppose that
κ >
θ
q
or κ =
θ
q
, αw >
1 − γ
q
.
(157)
1
q for κ = θ
q .
q , α = αu + αw − 1
We set Z = (p, q, u, w, h, m∗, c∗).
Denote α = αu + αw for κ > θ
In [37] estimates for en(Su,w,A : lp(A) → lq(A)) were obtained in the case α 6=
q . The case α = 1
q , q = ∞ was considered by Lifshits and Linde [25, 27]. To
u,w,A : l1(A) → lp′(A)
p − 1
this end estimates for entropy numbers of the dual operator S∗
were obtained and the result of the paper [4] was applied.
Theorem 4. Let θ > 0, 1 < p < q < ∞, α = 1
q . Then
q − 1
p .
p − 1
p − 1
en(Su,w,A : lp(A) → lq(A)) ≍
Z
1
n
Now we suppose that θ = 0, κ > 0, and instead of (156) the following condition
holds:
uj = 2κm∗j(m∗j + 1)α[log(m∗j + 1)]−λu, wj = 2−κm∗j(m∗j + 1)−α[log(m∗j + 1)]−λw
(158)
with α ∈ R.
Theorem 5. Let θ = 0, γ 6 0, 1 < p < q < ∞, κ > 0, λu + λw = 1
p − 1
q . Then
en(Su,w,A : lp(A) → lq(A)) ≍
Z
1
q − 1
p .
n
40
REFERENCES
[1] O.V. Besov, "Kolmogorov widths of Sobolev classes on an irregular domain",
Proc. Steklov Inst. Math., 280 (2013), 34-45.
[2] 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].
[3] M.Sh. Birman and M.Z. Solomyak, "Piecewise polynomial approximations of
functions of classes W α
p ", Mat. Sb. 73:3 (1967), 331- -- 355.
[4] J. Bourgain, A. Pajor, S. Szarek, N. Tomczak-Jaegermann, "On the duality
problem for entropy numbers of operators", Geometric Aspects of Functional
Analysis, Lecture Notes in Mathematics 1376, 50 -- 63.
[5] M. Bricchi,
"Existence and properties of h-sets", Georgian Mathematical
Journal, 9:1 (2002), 13(cid:22)32.
[6] B. Carl,
I. Stephani, Entropy, Compactness, and the Approximation of
Operators. Cambridge Tracts in Mathematics, V. 98. Cambridge: Cambridge
University Press, 1990.
[7] D.E. Edmunds, Yu.V. Netrusov, "Entropy numbers of operators acting between
vector-valued sequence spaces", Math. Nachr., 286:5 -- 6 (2013), 614 -- 630.
[8] D.E. Edmunds, Yu.V. Netrusov, "Schutt's theorem for vector-valued sequence
spaces", J. Approx. Theory, 178 (2014), 13 -- 21.
[9] D.E. Edmunds, H. Triebel, Function spaces, entropy numbers, differential
in Mathematics, 120 (1996). Cambridge
operators. Cambridge Tracts
University Press.
[10] D.D. Haroske, "Entropy numbers in weighted function spaces and eigenvalue
distributions of some degenerate pseudodifferential operators. I", Math. Nachr.,
167 (1994), 131 -- 156.
[11] D.D. Haroske, "Entropy numbers in weighted function spaces and eigenvalue
distributions of some degenerate pseudodifferential operators. II", Math. Nachr.,
168 (1994), 109 -- 137.
[12] D.D. Haroske, H. Triebel, "Wavelet bases and entropy numbers in weighted
function spaces", Math. Nachr., 278:1 -- 2 (2005), 108 -- 132.
[13] 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.
[14] D.D. Haroske, L. Skrzypczak,
[15] 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.
41
[16] 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.
[17] T. Kuhn, "Entropy numbers of diagonal operators of logarithmic type", Georgian
Math. J. 8:2 (2001), 307-318.
[18] T. Kuhn, "A lower estimate for entropy numbers", J. Appr. Theory, 110 (2001),
120 -- 124.
[19] T. Kuhn, "Entropy numbers of general diagonal operators", Rev. Mat. Complut.,
18:2 (2005), 479 -- 491.
[20] Th. Kuhn, H.-G. Leopold, W. Sickel, and L. Skrzypczak. "Entropy numbers of
embeddings of weighted Besov spaces", Constr. Approx., 23 (2006), 61 -- 77.
[21] 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.
[22] T. Kuhn, "Entropy Numbers in Weighted Function Spaces. The Case of
Intermediate Weights", Proc. Steklov Inst. Math., 255 (2006), 159 -- 168.
[23] 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.
[24] T. Kuhn, "Entropy numbers in sequence spaces with an application to weighted
function spaces", J. Appr. Theory, 153 (2008), 40 -- 52.
[25] M.A. Lifshits, "Bounds for entropy numbers for some critical operators", Trans.
Amer. Math. Soc., 364:4 (2012), 1797 -- 1813.
[26] M.A. Lifshits, W. Linde, "Compactness properties of weighted summation
operators on trees", Studia Math., 202:1 (2011), 17 -- 47.
[27] 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.
[28] 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.
[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] C. Schutt, "Entropy numbers of diagonal operators between symmetric Banach
spaces", J. Appr. Theory, 40 (1984), 121 -- 128.
42
[34] V.M. Tikhomirov, "The ε-entropy of certain classes of periodic functions"
(Russian) Uspehi Mat. Nauk 17:6 (108) (1962), 163 -- 169.
[35] 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.
[36] A.A. Vasil'eva, "Widths of function classes on sets with tree-like structure", J.
Appr. Theory, 192 (2015), 19 -- 59.
[37] A.A. Vasil'eva,
"Entropy numbers of embedding operators of weighted
Sobolev spaces with weights that are functions of distance from some h-set",
arXiv:1503.00144.
[38] 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.
[39] A.A. Vasil'eva, "Widths of weighted Sobolev classes on a John domain", Proc.
Steklov Inst. Math., 280 (2013), 91 -- 119.
[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., 20:3 (2013), 360 -- 373.
[41] 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.
[42] 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.
[43] A.A. Vasil'eva, "Widths of weighted Sobolev classes with weights that are
functions of the distance to some h-set: some limit cases", Russ. J. Math. Phys.,
22:1 (2015), 127 -- 140.
[44] 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.
43
|
1605.01543 | 1 | 1605 | 2016-05-05T10:06:45 | Grothendieck's inequality in the noncommutative Schwartz space | [
"math.FA",
"math.OA"
] | In the spirit of Grothendieck's famous inequality from the theory of Banach spaces, we study a sequence of inequalities for the noncommutative Schwartz space, a Fr\'echet algebra of smooth operators. These hold in non-optimal form by a simple nuclearity argument. We obtain optimal versions and reformulate the inequalities in several different ways. | math.FA | math |
GROTHENDIECK'S INEQUALITY
IN THE NONCOMMUTATIVE SCHWARTZ SPACE
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
Abstract. In the spirit of Grothendieck's famous inequality from
the theory of Banach spaces, we study a sequence of inequalities for
the noncommutative Schwartz space, a Fr´echet algebra of smooth
operators. These hold in non-optimal form by a simple nuclear-
ity argument. We obtain optimal versions and reformulate the
inequalities in several different ways.
1. Introduction
The noncommutative Schwartz space S is a weakly amenable m-
convex Fr´echet algebra whose properties have been investigated in sev-
eral recent papers, see e.g. [2,3,13,14]. It is not difficult to see that as a
Fr´echet space, S is nuclear. From this, we can easily deduce the follow-
ing analogue of Grothendieck's inequality, which we call Grothendieck's
inequality in S: there exists a constant K > 0 so that for any contin-
uous bilinear form u : S × S → C and any n ∈ N, there exists k ∈ N
such that for every m ∈ N and any x1, . . . , xm, y1, . . . , ym ∈ S, we have
(1)
(cid:12)(cid:12)(cid:12)
u(xj, yj)(cid:12)(cid:12)(cid:12)
6 Kkuk∗
n k(xj)kRC
k k(yj)kRC
k
m
X
j=1
The norms on the right hand side arise naturally from the definition
of S, as explained in Section 2 below. Our goal in this note is to show
that in fact k = 2n + 1 always suffices, and that this is best possible.
This appears to be the first result concerning Grothendieck's inequal-
ity in the category of Fr´echet algebras; to the best of our knowledge,
2010 Mathematics Subject Classification. Primary: 47A30, 47A07, 47L10. Sec-
ondary: 47A63.
Key words and phrases. noncommutative Grothendieck inequality, noncommu-
tative Schwartz space, m-convex Fr´echet algebra, bilinear form, state.
The research of the second-named author has been supported by the National
Center of Science, Poland, grant no. UMO-2013/10/A/ST1/00091.
The authors are grateful to the anonymous referee, whose comments greatly
improved the presentation of this paper.
1
2
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
all previous results along these lines concern Banach spaces (includ-
ing C∗-algebras, general Banach algebras and operator spaces). For
Fr´echet algebras, Grothendieck's inequality seems to have a specific
flavour. Every Fr´echet space (and a fortiori, every Fr´echet algebra)
which appears naturally in analysis is nuclear, meaning that all tensor
product topologies are equal. Since Grothendieck's inequality can be
understood as the equivalence of two tensor products, it seems that
we can take inequality (1) for granted. The interesting question that
remains is then optimality.
This paper is divided into four sections.
In the remainder of this
section, we recall a C∗-algebraic version of Grothendieck's inequality
due to Haagerup, and then review the definition and the basic prop-
erties of S which we require. In Section 2 we explain how nuclearity
gives Grothendieck's inequality in S, and we estimate the constants
K and k. Section 3 then settles the optimality question for k via a
matricial construction. We conclude with a short section containing
several reformulations of the inequality.
1.1. Grothendieck's inequality. Pisier's survey article [12] is a com-
prehensive reference for Grothendieck's inequality. This presents many
equivalent formulations and applications of this famous result, and re-
counts its evolution from 'commutative' [5] to 'noncommutative'. Of
these reformulations and extensions, Haagerup's noncommutative ver-
sion most closely resembles (1), and we state it here for the convenience
of the reader.
Theorem 1 ([6], [12, Theorem 7.1]). Let A and B be C∗-algebras.
For any bounded bilinear form u : A × B → C and any finite sequence
(xj, yj) in A × B, we have
X u(xj, yj)(cid:12)(cid:12)(cid:12)
where k(xj)kRC := max(cid:8)(cid:13)(cid:13)P x∗
6 2kuk k(xj)kRC k(yj)kRC
j xj(cid:13)(cid:13)
(cid:12)(cid:12)(cid:12)
2(cid:9).
1.2. The noncommutative Schwartz space. Let
j(cid:13)(cid:13)
2 ,(cid:13)(cid:13)P xjx∗
1
1
s = nξ = (ξj)j∈N ∈ CN : ξn := (cid:16)
+∞
X
j=1
1
2
ξj2j2n(cid:17)
< +∞ for all n ∈ No
denote the so-called space of rapidly decreasing sequences. This space
becomes Fr´echet when endowed with the sequence ( · n)n∈N of norms
defined above. The basis (Un)n∈N of zero neighbourhoods of s is defined
by Un := {ξ ∈ s : ξn 6 1}. The topological dual s′ of s is the so-called
GROTHENDIECK'S INEQUALITY IN S
3
space of slowly increasing sequences, namely
+∞
X
j=1
1
2
n := (cid:16)
ηj2j −2n(cid:17)
< +∞ for some n ∈ No
nη = (ηj)j∈N ∈ CN : η′
where the duality pairing is given by hξ, ηi := Pj∈N ξjηj for ξ ∈ s,
η ∈ s′.
The noncommutative Schwartz space S is the Fr´echet space L(s′, s)
of all continuous linear operators from s′ into s, endowed with the
topology of uniform convergence on bounded sets. The formal identity
map ι : s ֒→ s′ is a continuous embedding and defines a product on
S by xy := x ◦ ι ◦ y for x, y ∈ S. There is also a natural involution
on S given by hx∗ξ, ηi := hξ, xηi for x ∈ S, ξ, η ∈ s′. With these
operations, S becomes an m-convex Fr´echet ∗-algebra. The inclusion
map S ֒→ K(ℓ2) is continuous, and in fact it is a spectrum-preserving
∗-homomorphism [3]. Moreover [13, Proposition 3], an element x ∈ S is
positive (i.e., x = y∗y for some y ∈ S), if and only if the spectrum of x is
contained in [0, +∞), or equivalently hxξ, ξi ≥ 0 for all ξ ∈ s′. On the
other hand, by [3, Cor. 2.4] and [4, Theorems 8.2, 8.3], the topology
of S cannot be given by a sequence of C∗-norms. This causes some
technical inconvenience (e.g. there is no bounded approximate identity
in S) meaning we cannot apply C∗-algebraic techniques directly.
2. The inequality
Let (k · kn)n∈N be a non-decreasing sequence of norms which gives
the topology of S. For u : S × S → C a continuous bilinear form, we
write
kuk∗
n := sup{u(x, y) : x, y ∈ Un}
where Un = {x ∈ S : kxkn 6 1}; similarly, for a functional φ ∈ S ′, we
write
kφk∗
n := sup{φ(x) : x ∈ Un}.
Following Pisier [11, p. 316], for k ∈ N and x1, x2, . . . , xm ∈ S, we write
k(xj)kRC
k = maxn(cid:13)(cid:13)(cid:13)
m
X
j=1
1
2
k
x∗
j xj(cid:13)(cid:13)(cid:13)
,(cid:13)(cid:13)(cid:13)
m
X
j=1
xjx∗
1
2
j(cid:13)(cid:13)(cid:13)
ko.
Relative to our choice of norms k·kn, we have now defined each term in
our hoped-for inequality (1). We will now reformulate it using tensor
products.
For C∗-algebras, such a reformulation is standard.
Indeed, by [6,
Theorem 1.1] (formulated along the lines of [8, Theorem 2.1]), Haagerup's
noncommutative Grothendieck inequality entails the existence of a K >
4
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
0 such that for any C∗-algebras A, B and z in the algebraic tensor prod-
uct A⊗ B, we have kzkπ 6 Kkzkah where k·kπ is the projective tensor
norm and k · kah is the absolute Haagerup tensor norm [8, p. 164]
on A ⊗ B, given by
m
X
j=1
xj2(cid:13)(cid:13)(cid:13)
1
2(cid:13)(cid:13)(cid:13)
m
X
j=1
1
2
.
yj2(cid:13)(cid:13)(cid:13)
kzkah = inf (cid:13)(cid:13)(cid:13)
2(x∗x + xx∗)(cid:1)
1
Here x = (cid:0) 1
2 for x an element of a C∗-algebra, and
the infimum is taken over all representations z = Pm
j=1 xj ⊗ yj where
(xj, yj) ∈ A × B.
We proceed similarly for S. For x ∈ S, let x2 = 1
2(x∗x + xx∗) ∈
S and consider the sequence of absolute Haagerup tensor norms (k ·
kah,n)n∈N on the algebraic tensor product S ⊗ S given by
kzkah,n := inf (cid:13)(cid:13)(cid:13)
m
X
j=1
1
2
n(cid:13)(cid:13)(cid:13)
xj2(cid:13)(cid:13)(cid:13)
m
X
j=1
1
2
n
yj2(cid:13)(cid:13)(cid:13)
where the infimum runs over all ways to represent z = Pm
j=1 xj ⊗ yj in
S ⊗ S. As usual, we write (k · kπ,n)n∈N for the sequence of projective
tensor norms on S ⊗ S.
Just as in the C∗-algebra case, inequality (1) will follow once we show
that the sequences of projective and absolute Haagerup tensor norms
are equivalent on S ⊗S. In fact, the equivalence of these norms follows
immediately from the nuclearity of S (see [9, Theorem 28.15] and [7,
Ch. 21, §2, Theorem 1] for details). On the other hand, the optimal
values of k and K (depending on n and our choice of norms (k·kn)n∈N)
for which (1) hold are not given by such general considerations. These
optimal parameters will be denoted by κ(n) := kbest and Kn := Kbest.
Henceforth, we focus only on the sequence of norms (k·kn)n∈N where
n = {ξ ∈ s′ :
kxkn := sup{xξn : ξ ∈ U ◦
n},
and U ◦
In other words, kxkn is the norm
n := ℓ2((j −n)j)
of x ∈ S, considered as a Hilbert space operator from H ′
to Hn := ℓ2((jn)j). This sequence does indeed induce the topology
of S.
In this context, we will estimate Kn and compute the exact
values of κ(n).
n ∈ N, x ∈ S
ξ′
n 6 1}.
We start with the following result, which can be compared with [13,
Lemma 8]. To fix some useful notation, for n ∈ N we define an infinite
diagonal matrix dn := diag(1n, 2n, 3n, 4n, . . . ) which we consider as an
isometry dn : ℓ2 → H ′
n and simultaneously as an isometry dn : Hn → ℓ2.
GROTHENDIECK'S INEQUALITY IN S
5
Proposition 2. Let n ∈ N. We have
(i) kxkn = sup{hxξ, ξi : ξ ∈ U ◦
(ii) kxk2
(iii) kxk2
n 6 kx2k2n for every self-adjoint x ∈ S; and
n 6 kx∗xk
2n for every x ∈ S.
Moreover, inequalities (ii) and (iii) are sharp.
n} for every positive x ∈ S;
2nkxx∗k
1
2
1
2
Proof. (i) Observe that kxkn = kdnxdnkB(ℓ2). Furthermore, since x is
positive, dnxdn is positive and we have
kxkn = kdnxdnkB(ℓ2) = sup{hxdnξ, dnξi : ξℓ2 6 1}
= sup{hxξ, ξi : ξ′
(ii) For x self-adjoint, we have
n 6 1}.
kx2k2n = kd2nx2d2nkB(ℓ2) = kd2nxk2
B(ℓ2),
and by [1, Proposition II.1.4.2],
kxkn = kdnxdnkB(ℓ2) = ν(dnxdn) = ν(d2nx) 6 kd2nxkB(ℓ2),
where ν(·) denotes the spectral radius. This gives the desired inequal-
ity.
(iii) Since S ֒→ B(ℓ2), any x ∈ S is also a Hilbert space operator,
(x∗x)1/2 i is positive in B(ℓ2⊕ ℓ2)
and the block-matrix operator h (xx∗)1/2
x∗
(see e.g. [10, p. 117]). Equivalently,
x
∀ ξ, η ∈ ℓ2.
hxξ, ηi2 6 h(xx∗)1/2η, ηih(x∗x)1/2ξ, ξi
(2)
For m ∈ N, let us write pm := [ Im 0
matrix. Now fix n ∈ N and choose ξ, η ∈ H ′
all m ∈ N and (2) gives
0 0 ] where Im ∈ Mm is the identity
n. Then pmξ, pmη ∈ ℓ2 for
hpmxpmξ, ηi2 6 hpm(xx∗)1/2pmη, ηihpm(x∗x)1/2pmξ, ξi.
Since (pm)m∈N is an approximate identity in S (see [13, Proposition 2]),
we obtain
hxξ, ηi2 6 h(xx∗)1/2η, ηih(x∗x)1/2ξ, ξi.
Taking the supremum over all ξ, η in the unit ball of H ′
n we get
kxk2
n 6 k(xx∗)1/2knk(x∗x)1/2kn.
Applying (ii) to the positive operators (xx∗)1/2 and (x∗x)1/2 we con-
clude that kxk2
then we have equality in both (ii) and (iii).
For sharpness, observe that if x is a diagonal rank one matrix unit
(cid:3)
n 6 kx∗xk
2nkxx∗k
2n.
1
2
1
2
6
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
Proposition 3. For any n, m ∈ N and x1, . . . , xm, y1, . . . , ym ∈ S, we
have
m
6
1
4
1
4
m
m
m
j=1
j=1
x∗
π2
X
X
X
kxjknkyjkn
j yj(cid:13)(cid:13)(cid:13)
2n+1(cid:13)(cid:13)(cid:13)
j xj(cid:13)(cid:13)(cid:13)
6 (cid:13)(cid:13)(cid:13)
6 and let p ∈ N. We claim that
kxk(cid:13)(cid:13)(cid:13)p+1
X
kxkkp 6 C(cid:13)(cid:13)(cid:13)
kx∗
j(cid:13)(cid:13)(cid:13)
2n+1(cid:13)(cid:13)(cid:13)
X
X
xjx∗
x∗
y∗
k=1
k=1
j=1
j=1
m
m
1
4
Proof. Let C := π2
.
2n+1(cid:13)(cid:13)(cid:13)
m
X
j=1
yjy∗
1
4
j(cid:13)(cid:13)(cid:13)
.
2n+1
By the Cauchy -- Schwarz inequality and Proposition 2(iii) this will then
imply the desired inequality. To establish the claim, let ξ1, . . . , ξm ∈ U ◦
p
and let us write (ej)j∈N for the standard basis vectors in ℓ2. We have
m
X
k=1
m
+∞
hx∗
kxkξk, ξki =
X
k=1
X
i,j=1
hx∗
kxkej, eii(ij)pξk
i i−pξk
j j −p.
Applying the Cauchy -- Schwarz inequality to summation over i, j ∈ N
gives
m
X
k=1
hx∗
kxkξk, ξki 6
m
X
k=1
(cid:16)
+∞
X
i,j=1
hx∗
kxkej, eii2(ij)2p(cid:17)
1
2
.
Since x∗x is positive for any x ∈ S, and for positive operators y ∈ S
we have yij2 6 yiiyjj (where yij := hyej, eii), this implies that
m
X
k=1
hx∗
kxkξk, ξki 6
X
j=1
+∞
+∞
D(cid:16)
m
X
k=1
kxk(cid:17)jpej, jpejE
x∗
m
By Proposition 2(i), for any ε > 0 there are ξ1, . . . , ξm ∈ U ◦
kxk(cid:13)(cid:13)(cid:13)p+1
+ ε.
kxkξk, ξki + ε < C(cid:13)(cid:13)(cid:13)
hx∗
kxkkp <
kx∗
X
X
X
x∗
k=1
k=1
k=1
m
m
p with
Taking the infimum over ε > 0 yields the claim.
(cid:3)
kxk(cid:17)ip+1ei, ip+1eiE
x∗
j −2 sup
6
j=1
X
6 C(cid:13)(cid:13)(cid:13)
m
X
k=1
m
k=1
i∈N D(cid:16)
X
kxk(cid:13)(cid:13)(cid:13)p+1
.
x∗
GROTHENDIECK'S INEQUALITY IN S
7
As a straightforward consequence of Proposition 3, we obtain:
Theorem 4 (Grothendieck's inequality in S). There is a constant
K 6 π2
6 such that kzkπ,n 6 2Kkzkah,2n+1 for any n ∈ N and z ∈
S ⊗ S. Moreover, every continuous bilinear form u : S × S → C
satisfies inequality (1) with k = 2n + 1, for any n, m ∈ N and any
x1, . . . , xm, y1, . . . , ym ∈ S. In particular, taking u(x, y) := φ(x)φ(y)
where φ ∈ S ′, we obtain
(3)
φ(xj)2 6 K(kφk∗
n k(xj)kRC
2n+1)2
X
m
j=1
Remark. This shows that κ(n) 6 2n + 1. On the other hand, it
is easy to show that κ(n) > 2n − 1.
Indeed, if not, then (3) would
hold with 2n + 1 replaced by some ℓ 6 2n − 1. Take m ∈ N, define
ξm := Pm
j=1 jnej and φm ∈ S ′ by φm(x) := hxξm, ξmi. Then for xj :=
ejj, j = 1, . . . , m we get k(xj)kRC
n = m. On the
other hand, Pm
j=1 φm(xj)2 is equivalent (up to a constant) to m4n+1.
Therefore (3) takes the form m4n+1 6 Cm2ℓ+2 for some constant C
(independent of m). Letting m tend to infinity, we obtain ℓ > 2n − 1
2,
a contradiction. Hence κ(n) ∈ {2n, 2n + 1}.
3. Optimality
ℓ = mℓ and kφmk∗
We will now show that κ(n) = 2n+1. For this, we will use the tensor
product formulation, noting that
κ(n) = minnk ∈ N : supn kzkπ,n
kzkah,k
: z ∈ S ⊗ S, z 6= 0o < ∞o.
Recall the diagonal operator dn defined on page 4 above. Since every
x−→
x ∈ S is an operator on ℓ2 via the canonical inclusions ℓ2
dn֒→ ℓ2, it is clear that if x ∈ S, then dnx and xdn are both operators
Hn
on ℓ2. This leads to the following observation.
Proposition 5. If z = Pk
kzkπ,n = (cid:13)(cid:13)(cid:13)
j=1 xj ⊗ yj ∈ S ⊗ S, then
dnxjdn ⊗ dnyjdn(cid:13)(cid:13)(cid:13)π
dn֒→ H ′
X
j=1
n
.
k
Proof. Write
∆nz =
k
X
j=1
dnxjdn ⊗ dnyjdn ∈ B(ℓ2) ⊗ B(ℓ2).
8
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
If ∆nz = Pm
for some ε > 0, then z = Pm
n knkd−1
kzkπ,n 6
l=1 al ⊗ bl ∈ B(ℓ2)⊗B(ℓ2) and Pm
n ⊗ d−1
X
n kn =
n ald−1
n bld−1
kd−1
l=1 d−1
n ald−1
X
m
l=1
l=1 kalkkblk < k∆nzkπ + ε
n bld−1
m
n and
kalkkblk < k∆nzkπ + ε.
l=1
This gives kzkπ,n 6 k∆nzkπ. The reverse inequality is proved similarly.
(cid:3)
We also need the following well-known fact.
Proposition 6. If H is a Hilbert space and x1, . . . , xm ∈ B(H), then
m
X
j=1
(cid:13)(cid:13)(cid:13)
j(cid:13)(cid:13)(cid:13)h
xj ⊗ x∗
= (cid:13)(cid:13)(cid:13)
m
X
j=1
xjx∗
.
j(cid:13)(cid:13)(cid:13)
j=1 xjax∗
Proof. By [15, Theorem 4.3], the Haagerup norm on the left hand side
is equal to the completely bounded norm of the map on B(H) given
by a 7→ Pm
j , which is completely positive, so attains its com-
pletely bounded norm at the identity operator.
(cid:3)
Theorem 7. For every n ∈ N, we have κ(n) = 2n + 1.
Proof. By Theorem 4, it only remains to show that κ(n) > 2n. Choose
kn ∈ N sufficiently large that k + 1 6 2k(k
4n − 1) for all k > kn. This
inequality ensures that for every k > kn, if we define
1
i1 = 2k,
ik+1 = ⌊k
1
4n 2k⌋,
ij = ik+1 + j − (k + 1), 2 6 j 6 k,
then i1 < i2 < · · · < ik+1. Denote by (eij)i,j∈N the standard matrix
units, and for j = 2, . . . , k + 1, consider the self-adjoint operators
Let
xj := ei1,ij + eij ,i1 ∈ Mik+1 ⊂ S.
zk :=
k+1
X
j=2
xj ⊗ xj.
Since dnxjdn = in
by Propositions 5 and 6 we obtain
1 in
j (ei1,ij + eij ,i1) and (dnxjdn)2 = i2n
1 i2n
j (ei1,i1 + eij ,ij ),
kzkkπ,n = (cid:13)(cid:13)(cid:13)
> (cid:13)(cid:13)(cid:13)
k+1
X
j=2
k+1
X
j=2
dnxjdn ⊗ dnxjdn(cid:13)(cid:13)(cid:13)π
dnxjdn ⊗ dnxjdn(cid:13)(cid:13)(cid:13)h
k+1
X
j=2
= (cid:13)(cid:13)(cid:13)
(dnxjdn)2(cid:13)(cid:13)(cid:13)
= i2n
1
k+1
X
j=2
i2n
j .
GROTHENDIECK'S INEQUALITY IN S
9
On the other hand,
xj2 = x2
j = ei1,i1 + eij ,ij
and
k+1
X
j=2
Therefore
d2nx2
j d2n = i4n
1 kei1,i1 +
k+1
X
j=2
i4n
j eij ,ij .
kzkkah,2n 6 (cid:13)(cid:13)(cid:13)
6 i4n
k+1
xj2(cid:13)(cid:13)(cid:13)2n
X
j=2
1 k + i4n
k+1.
k+1
X
j=2
= (cid:13)(cid:13)(cid:13)
d2nx2
j d2n(cid:13)(cid:13)(cid:13)
= max{i4n
1 k, i4n
k+1}
Hence
kzkkπ,n
kzkkah,2n
>
1 Pk+1
j=2 i2n
i2n
j
i4n
1 k + i4n
k+1
>
i−2n
1
i2n
2
1 + k−1i−4n
1
k+1 → ∞ as k → ∞,
i4n
by our choice of i1, . . . , ik+1. So κ(n) > 2n as required.
(cid:3)
4. Reformulations of the inequality
Here we give several different ways of stating our inequality; in each
case, an analogous result may be found in [12]. The methods here are
fairly standard, so full proofs are often omitted. Throughout, we write
K = supn∈N Kn 6 π2/6.
4.1. Grothendieck's inequality with states. Given ξ ∈ U ◦
n, let
φξ ∈ S ′ be given by φξ(x) = hxξ, ξi, x ∈ S. We call an element of
the closed convex hull of {φξ : ξ ∈ U ◦
n} an n-state on S. Note that
by Proposition 2(i), for any positive element x ∈ S we have kxkn =
sup{φ(x) : φ ∈ Vn}, where Vn ⊆ S ′ is the set of all n-states on S. The
next result may be deduced from Theorem 4 by closely following the
Hahn -- Banach Separation argument of [12, §23].
Theorem 8. For any continuous bilinear form u : S × S → C and
n ∈ N, there are (2n + 1)-states φ1, φ2, ψ1, ψ2 on S with
u(x, y) 6 Kkuk∗
n(cid:0)φ1(x∗x) + φ2(xx∗)(cid:1)
1
2(cid:0)ψ1(y∗y) + ψ2(yy∗)(cid:1)
1
2
for all x, y ∈ S.
4.2. 'Little' Grothendieck inequality. As a consequence we ob-
tain the following 'little' Grothendieck inequality in S. Recall that
if T : X → Y is a linear map between Fr´echet spaces, then kTkn,k :=
sup{kT xkk : kxkn 6 1}.
10
RUPERT H. LEVENE AND KRZYSZTOF PISZCZEK
Theorem 9. For any Fr´echet-Hilbert space H, if u1, u2 : S → H are
continuous linear maps, k, m, n ∈ N and x1, . . . , xm, y1, . . . , ym ∈ S,
then
m
X
j=1
(cid:12)(cid:12)(cid:12)
hu1(xj), u2(yj)ik(cid:12)(cid:12)(cid:12)
6 Kku1kn,k ku2kn,k k(xj)kRC
2n+1 k(yj)kRC
2n+1.
Equivalently, for any k, n ∈ N there are (2n + 1)-states φ1, φ2, ψ1, ψ2
such that for all x, y ∈ S we have
hu1(x), u2(y)ik 6 Kku1kn,kku2kn,k
1
× (cid:0)φ1(x∗x) + φ2(xx∗)(cid:1)
2(cid:0)ψ1(y∗y) + ψ2(yy∗)(cid:1)
1
2 .
Proof. Apply Theorems 4 and 8 to uk(x, y) := hu1(x), u2(y)ik for k ∈
N.
(cid:3)
Using the same argument as in the proof of Theorem 8 we can obtain
an equivalent version of the 'little' Grothendieck inequality.
Theorem 10. For any Fr´echet-Hilbert space H, if u : S → H is a
continuous linear map and k, n ∈ N, then there exist (2n + 1)-states
φ1, φ2 on S such that for all x ∈ S we have
kuxkk 6 √Kkukn,k(cid:0)φ1(x∗x) + φ2(xx∗)(cid:1)
1
2 .
References
[1] B. Blackadar. Operator algebras, volume 122 of Encyclopaedia of
Mathematical Sciences. Springer-Verlag, Berlin, 2006.
[2] T. Cia´s. On the algebra of smooth operators. Studia Math.,
218(2):145 -- 166, 2013.
[3] P.
Doma´nski.
Algebra
of
smooth
operators.
http://main3.amu.edu.pl/~domanski/salgebra1.pdf, 2012.
[4] M. Fragoulopoulou. Topological algebras with involution. Elsevier
Science B.V., Amsterdam, 2005.
[5] A. Grothendieck. R´esum´e de la th´eorie m´etrique des produits ten-
soriels topologiques. Resenhas, 2(4):401 -- 480, 1996. Reprint of Bol.
Soc. Mat. Sao Paulo 8 (1953), 1 -- 79.
[6] U. Haagerup. The Grothendieck inequality for bilinear forms on
C ∗-algebras. Adv. in Math., 56(2):93 -- 116, 1985.
[7] H. Jarchow. Locally convex spaces. B. G. Teubner, 1981.
GROTHENDIECK'S INEQUALITY IN S
11
[8] S. Kaijser and A.M. Sinclair. Projective tensor products of C ∗-
algebras. Math. Scand., 55(2):161 -- 187, 1984.
[9] R. Meise and D. Vogt. Introduction to functional analysis, vol-
ume 2 of Oxford Graduate Texts in Mathematics. Oxford Univer-
sity Press, 1997.
[10] V. Paulsen. Completely bounded maps and operator algebras. Cam-
bridge University Press, Cambridge, 2002.
[11] G. Pisier. Introduction to operator space theory, volume 294 of Lon-
don Mathematical Society Lecture Note Series. Cambridge Univer-
sity Press, Cambridge, 2003.
[12] G. Pisier. Grothendieck's theorem, past and present. Bull. Amer.
Math. Soc. (N.S.), 49(2):237 -- 323, 2012.
[13] K. Piszczek. Automatic continuity and amenability in the non-
commutative Schwartz space. J. Math. Anal. Appl., 432(2):954 --
964, 2015.
[14] K. Piszczek. A Jordan-like decomposition in the noncommutative
Schwartz space. Bull. Aust. Math. Soc., 91(2):322 -- 330, 2015.
[15] R. R. Smith. Completely bounded module maps and the Haagerup
tensor product. J. Funct. Anal., 102(1):156 -- 175, 1991.
School of Mathematics and Statistics, University College Dublin,
Belfield, Dublin 4, Ireland
E-mail address: [email protected]
Faculty of Mathematics and Comp. Sci., A. Mickiewicz University
in Pozna´n, Umultowska 87, 61-614 Pozna´n, Poland
E-mail address: [email protected]
|
1501.04513 | 1 | 1501 | 2015-01-19T15:10:25 | Integral inequalities for infimal convolution and Hamilton-Jacobi equations | [
"math.FA"
] | Let $f,g:\Bbb{R}^{N}\rightarrow (-\infty ,\infty ]$ be Borel measurable, bounded below and such that $\inf f+\inf g\geq 0.$ We prove that with $ m_{f,g}:=(\inf f-\inf g)/2,$ the inequality $||(f-m_{f,g})^{-1}||_{\phi }+||(g+m_{f,g})^{-1}||_{\phi }\leq 4||(f\Box g)^{-1}||_{\phi }$ holds in every Orlicz space $L_{\phi },$ where $f\Box g$ denotes the infimal convolution of $f$ and $g$ and where $||\cdot ||_{\phi }$ is the Luxemburg norm (i.e., the $L^{p}$ norm when $L_{\phi }=L^{p}$).
Although no genuine reverse inequality can hold in any generality, we also prove that such reverse inequalities do exist in the form $||(f\Box g)^{-1}||_{\phi }\leq 2^{N-1}(||(\check{f}-m_{f,g})^{-1}||_{\phi }+||(\check{ g}+m_{f,g})^{-1}||_{\phi }),$ where $\check{f}$ and $\check{g}$ are suitable transforms of $f$ and $g$ introduced in the paper and reminiscent of, yet very different from, nondecreasing rearrangement.
Similar inequalities are proved for other extremal operations and applications are given to the long-time behavior of the solutions of the Hamilton-Jacobi and related equations. | math.FA | math |
INTEGRAL INEQUALITIES FOR INFIMAL CONVOLUTION
AND HAMILTON-JACOBI EQUATIONS
PATRICK J. RABIER
This paper is dedicated to the memory of Jean Jacques Moreau
Abstract. Let f, g : RN → (−∞, ∞] be Borel measurable, bounded below
and such that inf f + inf g ≥ 0. We prove that with mf,g := (inf f − inf g)/2,
the inequality (f − mf,g )−1φ + (g + mf,g )−1φ ≤ 4(f (cid:3)g)−1φ holds in
every Orlicz space Lφ, where f (cid:3)g denotes the infimal convolution of f and g
and where · φ is the Luxemburg norm (i.e., the Lp norm when Lφ = Lp).
Although no genuine reverse inequality can hold in any generality, we
also prove that such reverse inequalities do exist in the form (f (cid:3)g)−1φ ≤
2N−1(( f − mf,g )−1φ + (g + mf,g )−1φ), where f and g are suitable trans-
forms of f and g introduced in the paper and reminiscent of, yet very different
from, nondecreasing rearrangement.
Similar inequalities are proved for other extremal operations and applica-
tions are given to the long-time behavior of the solutions of the Hamilton-
Jacobi and related equations.
1. Introduction
If f, g : RN → (−∞, ∞], the infimal convolution f (cid:3)g : RN → [−∞, ∞], first
introduced by Fenchel [9] and Moreau [24], [25], [26], is defined by the formula
(f (cid:3)g)(x) := inf
y∈RN
(f (x − y) + g(y)).
Since then, this operation and its extension to general vector spaces have found an
ever growing variety of applications, including convex functions [13], [29], extension
of Lipschitz functions [12], solutions of the Hamilton-Jacobi equations [2], [20], [31]
and much more (even a proof of the Hahn-Banach theorem [11]). In fact, there are
by now several thousands publications using infimal convolution in areas as diverse
as image processing, economics and finance, information theory, probabilities and
statistics, etc. For a glimpse into some of these problems, see the excellent recent
survey by Lucet [21].
In this paper, we investigate the mathematical properties of infimal convolution
in a new direction, by exploring the existence of integral inequalities involving f, g
and f (cid:3)g. The remark that f (cid:3)g = 0 whenever f ≥ 0 and g ≥ 0 are integrable could
cast serious doubts on the value of this program, but they are quickly dispelled by
the rebuttal that no similar triviality arises from the integrability of f −1 and g−1.
Here and everywhere else, f −1 := 1/f, g−1 := 1/g, etc. This notation will not be
used to denote any set-theoretic inverse.
1991 Mathematics Subject Classification. 26D15, 46E30, 35F25, 49L25.
Key words and phrases. Brunn-Minkowski inequality, enclosing ball, Hamilton-Jacobi equa-
tions, infimal convolution, Orlicz space, rearrangement.
1
2
PATRICK J. RABIER
Omitting technicalities to which we shall return shortly, the first batch of in-
equalities will relate the (Luxemburg) norm (f (cid:3)g)−1φ in any Orlicz space Lφ, to
the norms (f − z)−1φ and (g + z)−1φ for a suitable constant z independent of
φ, to be defined in due time. The only restrictions are that f and g must be Borel
measurable, bounded below and that f (cid:3)g ≥ 0. The proofs depend crucially upon
(a slightly weaker form of) the Brunn-Minkowski inequality.
The setting of Orlicz spaces instead of just the classical Lp spaces introduces only
mild additional technicalities, is more natural in many respects and, as we shall see
in the examples of Section 7, is useful in some applications. It does not even require
any knowledge of Orlicz spaces beyond the definitions of Young functions and of
the Luxemburg norm, which will both be reviewed.
This being said, a simple special case asserts that if f, g ≥ 0 are Borel measurable
and inf f = inf g (see Theorem 3.4 for a full and much more general statement)
f −1p + g−1p ≤ 4(f (cid:3)g)−1p,
(1.1)
for every 1 ≤ p ≤ ∞, where · p is the norm of Lp := Lp(RN ). The constant
4 is best possible among all constants independent of p, as is readily seen when
f = g = 1 and p = ∞.
The Borel measurability requirement has to do with the measurability of f (cid:3)g,
without which (1.1) cannot make sense. Curiously, we were unable to find a dis-
cussion of the measurability properties of the infimal convolution in the classical
literature, but the evidence points to the fact that f and g Lebesgue measurable
does not suffice for the measurability of f (cid:3)g. Indeed, as is well-known, the strict
epigraph of f (cid:3)g is the (vector, also called Minkowski) sum of the strict epigraphs
of f and g and Sierpi´nski [30] showed, almost a century ago, that the sum of two
Lebesgue measurable sets need not be Lebesgue measurable. In contrast, the sum
of two Borel sets is always Lebesgue measurable (but not always a Borel set). See
Section 2 for further details.
A peculiar feature of (1.1) and of more general similar inequalities is that only
the left-hand side is unchanged by modifications of f and g on null sets, as long as
Borel measurability and inf f = inf g > −∞ are preserved.
Most of the paper is actually devoted to perhaps more important -and definitely
more delicate- reverse inequalities which, in a simpler world, would read
(1.2)
(f (cid:3)g)−1p ≤ C(f −1p + g−1p),
with C > 0 independent of f and g in some suitable class of nonnegative functions.
Unfortunately, the main obstacle to (1.2) is that no remotely general converse of
the Brunn-Minkowski inequality holds in any form, even for convex sets. Such a
converse is actually trivially true for Euclidean balls, but a direct application of
this remark only yields (1.2) for a narrow subclass of radially symmetric functions.
To take advantage of the converse of the Brunn-Minkowski inequality for balls
in a much broader setting, we introduce a new function transform, strongly remi-
niscent of, yet very different from, nonincreasing rearrangement. The difference is
that the upper level sets are rounded before being rearranged, the rounding being
performed by using the concept of enclosing ball (see Section 4).
To each function f : RN → [−∞, ∞] (no measurability needed), the aforemen-
tioned transform associates a measurable radially symmetric function f , which in
turn produces another measurable radially symmetric function f := −(−f). In the
special case when f, g ≥ 0 are Borel measurable and inf f = inf g (see Theorem
INEQUALITIES FOR INFIMAL CONVOLUTION
3
6.1 for a full and much more general statement), the reverse inequality in Lp reads
(compare with (1.1) under the same assumptions)
(1.3)
(f (cid:3)g)−1p ≤ 2N −1( f −1p + g−1p).
Such an inequality breaks down completely if f and g are replaced with f and
g, respectively, even if both functions are radially symmetric. For example, if
N = 1, f (x) = x2 + 1 and g(x) = x2 + 1 when x /∈ Q, g(x) = 1 if x ∈ Q, then f and
g are Borel measurable and inf f = inf g = 1. But (f (cid:3)g)−1 = 1/2 is in no Lp space
with p < ∞, whereas f −1 and g−1 = f −1 a.e. are in all of them. (In this example,
it turns out that f = f but g = 1. ) This example also shows that, unlike in (1.1),
neither side of (1.3) is independent of modifications of f or g on null sets that do
not affect Borel measurability or inf f = inf g.
When not trivial (i.e., f = f a.e.), the explicit calculation of f is generally
not possible. Nevertheless, the inequality (1.3) is useful because some simple and
general conditions about f and g ensure the finiteness of the right-hand side (Lemma
6.3). There is certainly more to be discovered in that regard.
The proofs of the inequalities involve two other classical extremal operations
(f ⊼ g)(x) := sup
y
min{f (x − y), g(y)} and (f ⊻ g)(x) := inf
y
max{f (x − y), g(y)}.
Either of these operations fully determines the other, but both notations will be
useful. In general, f ⊻ g = −(−f ) ⊼ (−g) and, for nonnegative functions, f ⊻ g =
(f −1⊼g−1)−1 will be important. We also prove inequalities similar to (1.1) and (1.3)
for the operations ⊼ and ⊻ (both being often referred to as "level sum" operations
in the literature). In fact, a good part of the work will consist in proving integral
inequalities for ⊼, from which those for (cid:3) and ⊻ will be derived.
In the last section, the inequalities are used to obtain Lp (and other) estimates
for the inverses of solutions of the Hamilton-Jacobi equations and variants thereof.
Throughout the paper, µN denotes the N -dimensional Lebesgue measure and,
without a qualifier, measurability always means Lebesgue measurability.
2. Background
The purpose of this short section is to review the basic properties of the opera-
tions mentioned in the Introduction, to set the notation used in future sections and
to settle basic measurability issues.
Recall that if X and Y are subsets of RN , their sum X + Y is defined by
X + Y := (cid:26) {x + y : x ∈ X, y ∈ Y } if X 6= ∅ and Y 6= ∅,
∅ if X = ∅ or Y = ∅.
The following key lemma is well-known. The "proof" below merely makes the
connection with the deep property behind it.
Lemma 2.1. If X and Y are Borel subsets of RN , their sum X + Y is measurable1.
Proof. In Euclidean space (any dimension), the continuous image of a Borel set
is Lebesgue measurable; see Federer [8, p. 69]. Since X × Y is a Borel subset of
RN × RN and the addition is continuous on RN , the result follows.
(cid:3)
1Even a Suslin set, but not necessarily a Borel set.
4
PATRICK J. RABIER
Given two functions f, g : RN → [−∞, ∞] and ξ ∈ R, call F +
ξ and G+
ξ the upper
level sets
(2.1)
F +
ξ := {x ∈ RN : f (x) > ξ},
G+
ξ := {x ∈ RN : g(x) > ξ}
and call W +
ξ
(2.2)
the corresponding upper level set of f ⊼ g :
W +
ξ
:= {x ∈ RN : (f ⊼ g)(x) > ξ}.
It is a standard elementary property that
ξ = F +
(2.3)
W +
ξ + G+
ξ .
By Lemma 2.1 and since f ⊻ g = −(−f ) ⊼ (−g), it follows at once from (2.3) that:
Lemma 2.2. If f, g : RN → [−∞, ∞] are Borel measurable, then f ⊼ g and f ⊻ g
are measurable.
If f : RN → [−∞, ∞] we set
(2.4)
Mf := sup f,
mf := inf f.
The next relations are elementary, but important
(2.5) Mf ⊼g = min{Mf , Mg},
mf ⊼g ≥ min{mf , mg},
Mf ⊻g ≤ max{Mf , Mg},
mf ⊻g = max{mf , mg}.
We now turn to infimal convolution. Given a function f : RN → (−∞, ∞], we
denote by Ef := {(x, ξ) ∈ RN × R : f (x) < ξ} the strict epigraph of f. It is also a
simple well-known property that if g : RN → (−∞, ∞] is another function, then
Ef (cid:3)g = Ef + Eg.
Since a function is Borel measurable (measurable) if and only if its strict epigraph
is a Borel set (measurable), it follows from Lemma 2.1 that
Lemma 2.3. If f, g : RN → (−∞, ∞] are Borel measurable, then f (cid:3)g is measur-
able.
For future use, we also note that if z ∈ R,
(2.6)
f (cid:3)g = (f − z)(cid:3)(g + z).
3. First integral inequalities
The Brunn-Minkowski inequality (see e.g. Gardner's survey [10]) asserts that if
X, Y are nonempty measurable subsets of RN and if X + Y is measurable, then
µN (X + Y )1/N ≥ µN (X)1/N + µN (Y )1/N . Obviously, it fails if X or Y is empty
and the other has positive measure. We shall only need the less sharp form
(3.1)
µN (X + Y ) ≥ µN (X) + µN (Y ),
if X, Y and X + Y are measurable and X 6= ∅, Y 6= ∅.
Lemma 3.1. If f, g : RN → [0, ∞] are measurable and Mf = Mg (possibly ∞; see
(2.4)) and if f ⊼ g is measurable, then
(3.2)
ZRN
f +ZRN
g ≤ ZRN
f ⊼ g.
INEQUALITIES FOR INFIMAL CONVOLUTION
5
Proof. Set M := Mf = Mg ≤ ∞. If M = 0, then f = g = f ⊼ g = 0 and (3.2) is
trivial. In what follows, M > 0.
Since f ≥ 0, it is well-known (see (2.1)) that RRN f = R ∞
ξ = ∅ when ξ ≥ M, this reads RRN f = R M
0 µN (F +
ing F +
R M
0 (µN (F +
ξ ) + µN (G+
0 µN (G+
By (2.2) and (2.3) and since f ⊼ g is measurable by hypothesis, F +
ξ )dξ, so that RRN f +RRN g = R M
ξ ))dξ.
0 µN (F +
ξ )dξ. By us-
ξ )dξ. Likewise, RRN g =
ξ = W +
ξ
ξ 6= ∅ by definition of
ξ ). Therefore,
0 µN (W +
ξ )dξ (by (2.5), the second inequality
ξ )dξ = RRN f ⊼ g since f ⊼ g ≥ 0 and the proof
6= ∅ and G+
ξ + G+
ξ ) = µN (W +
ξ ) ≤ µN (F +
ξ + G+
(cid:3)
is measurable for every ξ. If ξ < M, then F +
ξ
M and so, by (3.1), µN (F +
RRN f +RRN g ≤ R M
0 µN (W +
is even an equality). Now, R ∞
ξ ) + µN (G+
ξ )dξ ≤ R ∞
0 µN (W +
is complete.
Of course, (3.2) does not follow from a pointwise inequality. The condition
Mf = Mg cannot be dropped. For example, if f > 0 and g = 0, then f ⊼ 0 = 0 and
(3.2) fails.
Lemma 3.1 is just the stepping stone for much more general inequalities. Recall
that in the theory of Orlicz spaces, a nonconstant function φ : [0, ∞] → [0, ∞]
is called a Young function if φ(0) = 0 and φ is nondecreasing, convex and left
continuous ([27]; see also [1], [18] for a simplified treatment limited to N -functions).
In particular, φ(∞) = ∞.
Remark 3.1. If φ is a Young function and h : RN → [0, ∞] is measurable, the
monotonicity of φ shows at once that φ(h) is measurable.
If φ is a Young function, the corresponding Orlicz space Lφ consists of all the
measurable functions h on RN such that RRN φ (λh) < ∞ for some λ > 0 (this
makes sense by Remark 3.1). It is a (complete) normed space for the Luxemburg
norm · φ defined by
(3.3)
hφ := inf (cid:26)r > 0 : ZRN
φ(r−1h) ≤ 1(cid:27) .
Since the right-hand side of (3.3) is finite if and only if h ∈ Lφ, it will always be
understood that hφ = ∞ when h is measurable and h /∈ Lφ. Thus, h ∈ Lφ is
equivalent to hφ < ∞. Furthermore, it is readily checked that
(3.4)
hφ ≤ kφ if h ≤ k
and, by the left-continuity of φ and monotone convergence2, that if h ∈ Lφ,
(3.5)
ZRN
φ(h−1
φ h) ≤ 1.
If φ(τ ) := τ p for some 1 ≤ p < ∞, then hφ = hp. On the other hand,
hφ = h∞ when φ is the indicator function of [0, 1] (φ = 0 in [0, 1] and ∞
outside).
Lemma 3.2. If φ is a Young function and if h : RN → [0, ∞], then (see (2.4))
Mφ(h) = φ(Mh).
2This is of course a well-known inequality.
6
PATRICK J. RABIER
Proof. It is plain that h ≤ Mh implies φ(h) ≤ φ(Mh), so that Mφ(h) ≤ φ(Mh).
It only remains to show that φ(Mh) ≤ Mφ(h), which is trivial if Mh = 0. We
henceforth assume Mh > 0.
By the monotonicity of φ and φ(0) = 0, there is τ 1 ∈ [0, ∞] such that φ = ∞
on (τ 1, ∞] and that φ < ∞ on [0, τ 1). Specifically, τ 1 = sup{τ ≥ 0 : φ(τ ) < ∞}.
If τ 1 ∈ (0, ∞), then φ(τ 1) may be finite or infinite. We split the proof into three
cases.
(i) Mh > τ 1. If so, τ 1 < ∞ and then φ(Mh) = ∞. The set {x ∈ RN : h(x) > τ 1}
is not empty and φ(h(x)) = ∞ for every x in that set. Thus, {x ∈ RN : φ(h(x)) =
∞} 6= ∅, so that Mφ(h) = ∞ = φ(Mh).
(ii) Mh = τ 1 = ∞. Then, φ(Mh) = φ(∞) = ∞. If τ > 0 is finite, ∅ 6= {x ∈ RN :
h(x) > τ } ⊂ {x ∈ RN : φ(h(x)) ≥ φ(τ )}(by the monotonicity of φ). As a result,
Mφ(h) ≥ φ(τ ). By letting τ → ∞ = τ 1 and since limτ→∞ φ(τ ) = φ(∞) = ∞ by the
left continuity of φ, it follows that Mφ(h) = ∞ = φ(Mh).
(iii) 0 < Mh ≤ τ 1. If Mh = ∞, then τ 1 = ∞ and (ii) above applies. Assume now
Mh < ∞. For every ε > 0, S := {x ∈ RN : h(x) > Mh − ε} 6= ∅. If ε is small enough,
then Mh − ε > 0 and S ⊂ {x ∈ RN : φ(h(x)) ≥ φ(Mh − ε)} by the monotonicity of
φ. Hence, Mφ(h) ≥ φ(Mh − ε). Since φ is left continuous, Mφ(h) ≥ φ(Mh).
(cid:3)
From Lemma 3.1 and Lemma 3.2, we obtain:
Lemma 3.3. Let φ : [0, ∞] → [0, ∞] be a Young function. If f, g : RN → [0, ∞]
are Borel measurable and if Mf = Mg (possibly ∞), then f ⊼ g is measurable and
(3.6)
max{f φ, gφ} ≤ f ⊼ gφ.
Proof. By Lemma 2.2, f ⊼ g is measurable and so, by Remark 3.1, φ(f ), φ(g) and
φ(f ⊼ g) are measurable. Since φ(min{f (y), g(x − y)}) = min{φ(f (y)), φ(g(x − y))}
by the monotonicity of φ, we infer that supy φ(min{f (y), g(x − y)}) = (φ(f ) ⊼
φ(g))(x). By Lemma 3.2 with h(y) := min{f (y), g(x − y)}, the left-hand side is
φ((f ⊼g)(x)), so that φ(f ⊼g) = φ(f ) ⊼φ(g). In particular, φ(f ) ⊼φ(g) is measurable.
Since Mf = Mg, then Mφ(f ) = Mφ(g), once again by Lemma 3.2. Thus, from the
above and from Lemma 3.1 with f and g replaced with φ(f ) and φ(g), respectively,
(3.7)
ZRN
φ(f ) +ZRN
φ(g) ≤ ZRN
φ(f ⊼ g).
If r > 0, then r−1(f ⊼ g) = r−1f ⊼ r−1g and Mr−1f = Mr−1g. Thus, (3.7) for
r−1f and r−1g yields RRN φ(r−1f ) ≤ RRN φ(r−1(f ⊼ g)), so that f φ ≤ f ⊼ gφ
by (3.3). Likewise, gφ ≤ f ⊼ gφ and (3.6) follows.
(cid:3)
Of course, when Lφ = L1, Lemma 3.1 yields the stronger f 1 + g1 ≤ f ⊼ g1
but (3.6) is optimal when φ is arbitrary (let f = g = 1 and Lφ = L∞).
We are now in a position to prove our first main integral inequality for infimal
convolution. Recall once more the notation (2.4).
Theorem 3.4. Suppose that f, g : RN → (−∞, ∞] are Borel measurable, that
mf , mg ∈ R and that mf + mg ≥ 0. Set
(3.8)
mf,g = (mf − mg)/2.
Then, f − mf,g, g + mf,g and f (cid:3)g are measurable and nonnegative and
(3.9)
(f − mf,g)−1φ + (g + mf,g)−1φ ≤ 4(f (cid:3)g)−1φ,
INEQUALITIES FOR INFIMAL CONVOLUTION
7
for every Young function φ.
Proof. The measurability of f (cid:3)g was established in Lemma 2.3. Next, inf(f −
mf,g) = inf(g +mf,g) = (mf +mg)/2 ≥ 0, whence f (cid:3)g = (f −mf,g)(cid:3)(g +mf,g) ≥ 0
by (2.6). This also implies (f −mf,g)−1 ≥ 0, (g+mf,g)−1 ≥ 0 and sup(f −mf,g)−1 =
2(mf + mg)−1 = sup(g + mf,g)−1. Therefore, the inequality (3.6) is applicable in
the form
(f − mf,g)−1φ + (g + mf,g)−1φ ≤ 2(f − mf,g)−1 ⊼ (g + mf,g)−1φ
and (3.9) follows from 0 ≤ h−1 ⊼ k−1 = (h ⊻ k)−1 ≤ 2(h(cid:3)k)−1 when h and k are
nonnegative, from f (cid:3)g = (f − mf,g)(cid:3)(g + mf,g) ≥ 0 and from (3.4).
(cid:3)
It was noted in the Introduction that the constant 4 in (3.9) is already best
possible when Lφ = L∞.
Remark 3.2. Theorem 3.4 gives a simple necessary condition for the existence of
solutions of infimal convolution equations (see [23], [21] and the references therein):
Suppose that h ≥ 0 is measurable and that g is Borel measurable and bounded below.
If h−1 ∈ Lφ for some Orlicz space Lφ and (g − mg + mh/2)−1φ > 4h−1φ (in
particular, if (g − mg + mh/2)−1 /∈ Lφ), the equation f (cid:3)g = h has no Borel
measurable solution f. Indeed, if f exists, then mf = mh − mg ∈ R and (3.9)
cannot hold.
If z 6= 0 is a constant, there is no simple pointwise relationship between (f − z) ⊻
(g + z) and f ⊻ g. As a result, the method of proof of Theorem 3.4 does not yield
a variant of (3.5) or (3.9) with f (cid:3)g replaced with f ⊻ g. However, if f, g ≥ 0, then
0 ≤ f ⊻ g ≤ f (cid:3)g and such a variant can be obtained as a straightforward corollary
of Theorem 3.4:
Corollary 3.5. Suppose that f, g : RN → [−∞, ∞] are Borel measurable, that g ≥
0 and that f 6 ≡∞ and g 6 ≡∞, so that 0 ≤ mf + + mg < ∞, where f+ := max{f, 0}.
Then, f+ − mf+,g, g + mf+,g (see (3.8 )) and f ⊻ g are measurable and nonnegative
and
(3.10)
(f+ − mf +,g)−1φ + (g + mf+,g)−1φ ≤ 4(f ⊻ g)−1φ,
for every Young function φ.
Proof. Since g ≥ 0, it follows that 0 ≤ f ⊻ g = f+ ⊻ g ≤ f+(cid:3)g. By Lemma 2.2,
f ⊻ g = f+ ⊻ g is measurable. Therefore, the corollary follows from (3.4) and from
Theorem 3.4 for f+ and g.
(cid:3)
The constant 4 is also best possible in (3.10) (among constants independent
of φ): If f = 1 and g = ℓ > 1 is constant, the inequality for the L∞ norm is
4(ℓ + 1)−1 ≤ 4ℓ−1. By letting ℓ → ∞, it follows that, in the right-hand side, 4
cannot be lowered.
4. The radial transforms f and f
The proof of Lemma 3.1 shows that the existence of a converse of the inequality
(3.2), that is,
ZRN
f ⊼ g ≤ C(cid:18)ZRN
f +ZRN
g(cid:19) ,
8
PATRICK J. RABIER
with C > 0 independent of f and g would require µN (F +
ξ ) +
µN (G+
ξ )) for every ξ > 0. However, as pointed out in the Introduction, no converse
of the Brunn-Minkowski inequality or its weaker form (3.1) holds in any generality.
The transforms defined in this section will enable us (in the next section) to
take advantage of the fact that such a converse trivially exists when X and Y are
Euclidean balls. The thought that this case is so special that it cannot have any
broad value would result in a serious oversight.
ξ ) ≤ C(µN (F +
ξ + G+
By a classical theorem of Jung [8, p. 200], [17], every nonempty bounded subset
X of RN is contained in a unique closed ball BX with minimal diameter among all
closed balls containing X, called the enclosing ball of X. If X is unbounded, no
closed ball contains X and we set BX := RN . Lastly, if X = ∅, every singleton {x}
satisfies the "minimal diameter" requirement, whence uniqueness, but not existence,
is lost. For definiteness, we arbitrarily set B∅ := {0}. Evidently, X ⊂ BX in
all cases. Jung's theorem also provides the estimate diam(X) ≤ diam(BX ) ≤
p2N/(N + 1) diam(X), but we shall only make use of the (trivial) first one.
Remark 4.1. An easily overlooked aspect of enclosing balls is that X ⊂ Y implies
only µN (BX ) ≤ µN (BY ) but not BX ⊂ BY , unless N = 1.
The following property of enclosing balls will be important.
Lemma 4.1. If Xn is a nondecreasing sequence of subsets of RN , then µN (B∪Xn ) =
lim µN (BXn ) = sup µN (BXn ).
Proof. Set X := ∪Xn. If X is unbounded, then BX = RN and so µN (BX ) = ∞.
Since X := ∪Xn and Xn ⊂ Xn+1, the diameter of Xn and, hence, that of BXn ,
tends to ∞. Accordingly, lim µN (BXn ) = ∞.
Suppose now that X is bounded, so that BX is a ball. Since Xn ⊂ X im-
plies µN (BXn ) ≤ µN (BX ) and since µN (BXn ) is nondecreasing, it is plain that
lim µN (BXn ) ≤ µN (BX ). To prove the converse, call rn ≥ 0 the radius of BXn .
The sequence rn is nondecreasing and bounded above (by the radius of BX ) and
so it has a limit r ≥ rn for every n. As a result, lim µN (BXn ) is the measure of any
ball with radius r.
Next, call xn the center of BXn . By a simple contradiction argument, the se-
quence xn is bounded (since xn might not be in BX -see Remark 4.1- this is not
totally trivial). After extracting a subsequence, assume that xn → x ∈ RN . Ev-
ery y ∈ X is in Xn for n large enough. Since BXn ⊂ B(xn, r), it follows that
y ∈ B(x, r). Thus, X ⊂ B(x, r), whence µN (B(x, r)) ≥ µN (BX ) by definition
of BX . Since r = lim rn amounts to µN (B(x, r)) = lim µN (BXn ), it follows that
lim µN (BXn ) ≥ µN (BX ).
(cid:3)
Given any function f : RN → [−∞, ∞], we now proceed to constructing a
measurable radially symmetric function f : RN → [−∞, ∞] whose upper level sets
F +
) for every ξ. The construction follows that of
the nonincreasing rearrangement of f.
ξ have measure equal to µN (BF +
ξ
Lemma 4.2. The function µN (BF +
[−∞, ∞].
ξ
) is nonincreasing and right-continuous on
INEQUALITIES FOR INFIMAL CONVOLUTION
9
η ⊂ F +
Proof. If ξ < η, then F +
continuity, let ξn ց ξ, so that F +
lim µN (BF +
).
ξn
ξ , so that µN (BF +
ξ = ∪F +
ξn
η
) ≤ µN (BF +
ξ
). For the right
and then, by Lemma 4.1, µN (BF +
ξ
) =
(cid:3)
Call ρ+
f (ξ) the radius of BF +
follows from Lemma 4.2 that ρ+
ξ
. Since ρ+
f (ξ) is proportional to µN (BF +
)1/N , it
ξ
f is nonincreasing and right-continuous. Therefore,
(4.1)
γ+
f (t) := inf{ξ : ρ+
f (ξ) ≤ t},
is a nonincreasing and right-continuous function on [0, ∞) and
(4.2)
{t ≥ 0 : γ+
f (t) > ξ} = [0, ρ+
f (ξ) is replaced with µN (F +
f (ξ)).
ξ ), γ+
Indeed, when f ≥ 0 and ρ+
f becomes the nonincreas-
ing rearrangement of f and these properties follow uniquely from the monotonicity
and right-continuity of µN (F +
ξ ); see for instance [33, pp. 26-27]. We also point out
that in most modern expositions, the nonincreasing rearrangement of a function f
is defined to be that of f . This has not always been the case (see Day [7] or Lux-
emburg [22]) and the monotonicity and right-continuity properties of nonincreasing
rearrangements are independent of whether f or f is used in their definition.
We now set
(4.3)
f (x) := γ+
f (x).
Some basic properties of f are summarized in the next theorem.
ξ
ξ ) = µN (BF +
ξ denotes the upper ξ-
) for every ξ ∈ [−∞, ∞], where F +
Theorem 4.3. Given f : RN → [−∞, ∞], the function f has the following prop-
erties:
(i) f is measurable and f = f a.e. if and only if f (x) is a.e. equal to a nonincreas-
ing function of x. If also f (x) is a right-continuous function of x, then f = f.
(ii) µN ( F +
level set of f .
(iii) M f ≤ Mf and m f ≥ mf (in particular, f ≥ 0 ⇒ f ≥ 0). Furthermore,
M f = ess sup f and m f = ess inf f .
(iv) (f + z) = f + z for z ∈ R and (f (c·)) = f (c·) for c ∈ R\{0}.
(v) (cf) = c f for every c ≥ 0.
(vi) If h : RN → [−∞, ∞] and h ≤ f, then h ≤ f .
(vii) If f is bounded below on bounded subsets and limx→∞ f (x) = −∞ and if
h : RN → [−∞, ∞] satisfies h(x) ≤ f (x) for x large enough, then h(x) ≤ f (x)
for x large enough. Furthermore, if f (x) is a strictly decreasing function of x
and if h(x) ≤ f (x) when x /∈ B for some open ball B centered at the origin, then
h(x) ≤ f (x) (= f (x) by (i)) for every x /∈ B.
Proof. (i) The measurability of f follows at once from the monotonicity of γ+
f and
the necessity of the given conditions for f = f a.e.
is obvious. Conversely, if
f (x) = γ(x) with γ : [0, ∞) → [−∞, ∞] nonincreasing, the upper level sets of f
are balls centered at the origin (possibly RN ) and ρ+
f is the distribution function
of γ, so that γ+
f = γ∗, the nonincreasing rearrangement of γ. Since γ∗ = γ except
perhaps at the countably many points of discontinuity of γ, it follows that f = f
10
PATRICK J. RABIER
a.e. Clearly, this remains true if f (x) = γ(x) a.e. If γ is right-continuous, then
γ∗ = γ and f = f.
(ii) Just notice that, by (4.2) and (4.3), F +
ξ
is the open ball with center 0 and
radius ρ+
f (ξ).
f = γ+
f (0) = inf{ξ : ρ+
(iii) With no loss of generality, assume Mf < ∞. Since γ+
f (ξ) = 0}. If ξ ≥ Mf , then F +
f is nonincreasing,
= {0}
ξ ) = 0} ≤ Mf . On the other hand,
f = M f . This shows that M f ≤ Mf . Furthermore, M f = ess sup f
f (ξ) = 0. Thus, max γ+
max γ+
and so ρ+
by (4.3), max γ+
by (4.3) and the right-continuity of γ+
f .
f = inf{ξ : µN (F +
ξ = ∅, whence BF +
ξ
That m f ≥ mf is obvious if mf = −∞, or if mf = ∞ (for then f = f = ∞). If
f ≥ mf by (4.1) and so f ≥ mf ,
mf ∈ R and ξ < mf , then ρ+
whence m f ≥ mf . That m f = ess inf f follows from the monotonicity of γ+
f .
f +z(ξ) = ρ+
f (ξ − z),
f +z, i.e., (f +z) = f +z. The proofs that f (c·) = f (c·)
(iv) Since {x ∈ RN : f (x) + z > ξ} = F +
ξ−z, it follows that ρ+
f (ξ) = ∞. Thus, γ+
which in turn yields γ+
if c ∈ R\{0} is equally straightforward.
f +z = γ+
(v) Since this is trivial when c = 0, assume c > 0. Then, ρ+
cf (ξ) = ρ+
f (ξ/c),
whence γ+
cf = cγ+
f , i.e. (cf) = c f .
(vi) If h ≤ f, then (with a self-explanatory notation) H +
h ≤ γ+
f (ξ) and, by (4.1), γ+
). Hence, ρ+
h (ξ) ≤ ρ+
ξ ⊂ F +
f , so that h ≤ f .
µN (BF +
ξ and so µN (BH+
) ≤
ξ
ξ
(vii) Choose an open ball B centered at the origin such that h(x) ≤ f (x) when
x /∈ B. Since f is bounded below on bounded subsets, inf B f is finite and, if
for every ξ ≤ ξ0. By (vi), h is increased when h is
ξ0 < inf B f, then B ⊂ F +
ξ
increased. Thus, if it can be shown that h ≤ f after increasing h, this inequality
also holds before h is increased. In particular, we may increase h on B so that
ξ0 < inf B h and then B ⊂ H +
for every
ξ
ξ ≤ ξ0. On the other hand, {x /∈ B : h(x) > ξ} ⊂ {x /∈ B : f (x) > ξ} since h ≤ f
on RN \B. Altogether, if ξ ≤ ξ0, then H +
f (ξ). As a result,
for every ξ ≤ ξ0. Thus, B ⊂ H +
ξ and so ρ+
h (ξ) ≤ ρ+
ξ ⊂ F +
ξ ∩ F +
ξ
(4.4) γ+
h (t) := inf{ξ : ρ+
h (ξ) ≤ t} ≤ inf{ξ ≤ ξ0 : ρ+
h (ξ) ≤ t} ≤
Since limx→∞ f (x) = −∞, the level set F +
ξ0
f (ξ) ≤ t}.
inf{ξ ≤ ξ0 : ρ+
is bounded, whence ρ+
f (ξ0) < ∞.
f (ξ0) ≤ t by the monotonicity
f (ξ) ≤ t}. By (4.4),
f (ξ0), it follows that h(x) ≤ f (x)
Choose any t ≥ ρ+
of ρ+
f , so that γ+
γ+
h (t) ≤ γ+
when x ≥ ρ+
f (ξ0).
f (ξ0). If ξ > ξ0, then ρ+
f (t) := inf{ξ : ρ+
f (ξ) ≤ ρ+
f (ξ) ≤ t} = inf{ξ ≤ ξ0 : ρ+
f (t). Since this is true for every t ≥ ρ+
To complete the proof, assume in addition that f (x) is a strictly decreasing
function of x. We show that h(x) ≤ f (x) when x /∈ B. If B = ∅, the result follows
from (v). From now on, assume B 6= ∅ (hence B 6= {0} as well since B is open).
By the monotonicity of f in x, inf tB f < inf B f if t > 1. It follows that F +
ξ0
⊂ tB
f (ξ0) ≤ tρ
if t > 1 and ξ0 above is close enough to inf B f. Thus, BF +
ξ0
where ρ is the radius of B. From the above, h(x) ≤ f (x) when x ≥ tρ and, hence,
when x ≥ ρ by first letting t → 1 (which gives only x > ρ) and next using the
⊂ tB, so that ρ+
INEQUALITIES FOR INFIMAL CONVOLUTION
11
right-continuity of f and h with respect to x. Since ρ is the radius of B and B is
centered at the origin, this means that h(x) ≤ f (x) when x /∈ B.
(cid:3)
Even when f ≥ 0, it is not true that f = 0 implies f = 0. The following
characterization is important for the proof of the reverse inequalities (specifically,
of Theorem 5.5 later).
Lemma 4.4. If f : RN → [0, ∞] and f = 0, then either f = 0 or there are x0 ∈ RN
and 0 < z ≤ ∞ such that f (x) = 0 if x 6= x0 and f (x0) = z.
f is nonincreasing, f = 0 means ρ+
has radius 0, which happens only when F +
Proof. By (4.1) and (4.3) and since ρ+
f (0) = 0,
0 = ∅ or
i.e., that the enclosing ball BF +
0 = {x0} is a singleton. If F +
when F +
0 = ∅, then f = 0 since f ≥ 0. Suppose now
that F +
0 = {x0}. This means that x0 is the only point where f is positive, so that
f (x0) = z > 0 (possibly ∞) and that f (x) ≤ 0 when x 6= x0. Since f ≥ 0, it follows
that f (x) = 0 when x 6= x0.
(cid:3)
0
Remark 4.2. By the argument of the above proof, the inequality M f ≤ Mf in
Theorem 4.3 (iii) can be made precise: M f < Mf if and only if some upper level
set of f is a singleton. Put differently, M f = Mf if and only if there is a sequence
of distinct points xn such that f (xn) → Mf .
We shall also need the transform defined by
f := −(−f),
(4.5)
directly given by the formula
(4.6)
where
(4.7)
f (x) = γ−
f (x),
γ−
f (t) := sup{ξ : ρ−
f (ξ) ≤ t}
and ρ−
f (ξ) is the radius of BF −
ξ
with, of course, F −
ξ = {x ∈ RN : f (x) < ξ}.
ξ ) = µN (BF −
ξ
) for every ξ ∈ [−∞, ∞], where F −
Theorem 4.5. Given f : RN → [−∞, ∞], the function f has the following prop-
erties:
(i) f is measurable and f = f a.e. if and only if f (x) is a.e. equal to a nondecreas-
ing function of x. If also f (x) is a right-continuous function of x, then f = f .
(ii) µN ( F −
ξ denotes the lower ξ-
level set of f .
(iii) M f ≤ Mf and m f ≥ mf (in particular, f ≥ 0 ⇒ f ≥ 0). Furthermore,
M f = ess sup f and m f = ess inf f .
(iv) (f + z) = f + z for every z ∈ R and (f (c·)) = f (c·) for c ∈ R\{0}.
(v) (cf) = c f for every c ≥ 0 (and (cf) = c f for every c < 0).
(vi) If h : RN → [−∞, ∞] and h ≤ f, then h ≤ f .
(vii) If h : RN → [−∞, ∞] is bounded above on bounded subsets, limx→∞ h(x) = ∞
and h(x) ≤ f (x) for x large enough, then h(x) ≤ f (x) for x large enough. Fur-
thermore, if h(x) is a strictly increasing function of x and if h(x) ≤ f (x) when
x /∈ B for some open ball B centered at the origin, then h(x) (= h(x) by (i)) ≤ f (x)
for every x /∈ B.
12
PATRICK J. RABIER
(viii) If f ≥ 0, then (f −1) = ( f )−1.
(ix) (f+
) = f+.
Proof. Parts (i) to (vii) follow at once from (4.5) and from the corresponding prop-
erties in Theorem 4.3.
f (ξ) = ∞ and ρ−
f (t) := inf{ξ ≥ 0 : ρ+
(viii) First, note that if f ≥ 0 and ξ < 0, then ρ+
f (ξ) = 0. Hence,
by (4.1) and (4.7), γ+
f (t) := sup{ξ ≥ 0 :
ρ−
f (ξ) ≤ t} ≥ 0. Upon replacing f with f −1 in the former formula, we get (since
passing from f to f −1 changes upper level sets into lower ones) γ+
f −1 (t) := inf{ξ ≥
f (ξ−1) ≤ t}. On the other hand, since inf(S−1) = (sup S)−1 for every subset
0 : ρ−
f (ξ−1) ≤
S ⊂ [0, ∞], we have (γ−
t}. This shows that γ+
f (t))−1 = inf{ξ−1 ≥ 0 : ρ−
f −1 = (γ−
f )−1 and the result follows from (4.3) and (4.6).
f (ξ) ≤ t} = inf{ξ ≥ 0 : ρ−
f (ξ) ≤ t} and γ−
(ix) The lower level sets {x ∈ RN : f+(x) < ξ} are empty if ξ ≤ 0 and coincide
f (ξ) if ξ > 0. Therefore,
(ξ) = 0 if ξ ≤ 0 and ρ−
f+
if ξ > 0. Thus, ρ−
f+
(ξ) = ρ−
with F −
ξ
by (4.7),
(4.8) γ−
f+
(t) = max{sup{ξ ≤ 0 : ρ−
f+
(ξ) ≤ t}, sup{ξ > 0 : ρ−
f+
(ξ) ≤ t}} =
max{0, sup{ξ > 0 : ρ−
f (ξ) ≤ t}}.
Suppose first that γ−
f (t) = sup{ξ : ρ−
f (ξ) ≤ t} ≤ 0. Accordingly, {ξ > 0 : ρ−
f (ξ) ≤ t} = −∞. By (4.8), γ−
f+
t} = ∅ and so sup{ξ > 0 : ρ−
0 = (γ−
sup{ξ > 0 : ρ−
= (γ−
that γ−
f+
f (t) = sup{ξ : ρ−
f )+(t). Suppose next that γ−
(t) = max{0, γ−
f (ξ) ≤ t}. By (4.8), γ−
f+
f )+, whence (f+
) = f+ by (4.6).
f (ξ) ≤
(t) = max{0, −∞} =
f (t) =
f )+(t). This shows
(cid:3)
f (ξ) ≤ t} > 0, so that γ−
f (t)} = (γ−
Remark 4.3. By (4.5) and Remark 4.2, m f = mf if and only if there is a sequence
of distinct points xn such that f (xn) → mf .
5. Reverse inequalities for f ⊼ g
We begin with the (trivial) converse of the Brunn-Minkowski inequality for Eu-
clidean balls.
Lemma 5.1. If B1 and B2 are Euclidean balls in RN , then µN (B1 + B2) ≤
2N −1(µN (B1) + µN (B2)).
Proof. Call ri the radius of Bi, i = 1, 2. It is readily checked that B1 + B2 is
a ball with radius r1 + r2 and the inequality simply follows from (r1 + r2)N ≤
2N −1(rN
(cid:3)
1 + rN
2 ).
In the next lemma, Mf = Mg is not needed (compare with Lemma 3.1).
Lemma 5.2. If f, g : RN → [0, ∞] are Borel measurable, then
ZRN
f ⊼ g ≤ 2N −1(cid:18)ZRN
f +ZRN
g(cid:19) .
Proof. By (2.2) and (2.3) and since f ⊼ g ≥ 0 is measurable (Lemma 2.2), it follows
ξ +
)), where Lemma 5.1 was used.
thatRRN f ⊼g = R ∞
ξ )dξ. Next, F +
and so µN (F +
0 µN (F +
+BG+
ξ +G+
)+µN (BG+
) ≤ 2N −1(µN (BF +
ξ ) ≤ µN (BF +
ξ ⊂ BF +
ξ +G+
G+
+BG+
ξ
ξ
ξ
ξ
ξ
ξ
INEQUALITIES FOR INFIMAL CONVOLUTION
13
This yieldsRRN f ⊼g ≤ 2N −1(cid:16)R ∞
(ii), the right-hand side is 2N −1(cid:16)R ∞
equals 2N −1(cid:16)RRN
0 µN (BF +
ξ
0 µN ( F +
f +RRN g(cid:17) because f , g ≥ 0 by Theorem 4.3 (iii).
)dξ +R ∞
ξ ))dξ +R ∞
0 µN (BG+
0 µN ( G+
ξ
)dξ(cid:17) . By Theorem 4.3
ξ ))dξ(cid:17) , which in turn
(cid:3)
No variant of Lemma 5.2 is true if f (or g) is replaced with f (or g):
Example 5.1. With N = 1, let 0 < f ≤ 1 be integrable with f (n) = 1 for every
n ∈ Z and let g = χ(−1,1) (= g). Then, f ⊼ g = 1, whence RR f ⊼ g = ∞ but
f, g ∈ L1.
If N = 1, f, g ≥ 0 are even and nonincreasing on [0, ∞) and Mf = Mg. Then,
Is there a
RR f ⊼ g = RR f + RR g by Theorem 4.3 (i) and Lemmas 3.1 and 5.2.
different proof ? (If f = g, this follows from (f ⊼ f )(x) = f (x/2).)
To go further, we need a simple property of Young functions.
Lemma 5.3. If φ is a Young function and f : RN → [0, ∞], then (φ(f )) = φ( f ).
Proof. For brevity, we only give the proof in the more important case when 0 <
φ < ∞ on (0, ∞), so that φ is continuous on [0, ∞] and has an inverse ψ. The
general case involves extra technicalities that lengthen the exposition. Recall that
ρ+
f (ξ) denotes the radius of BF +
.
Since φ(f ) ≥ 0, it follows that ρ+
φ(f (x)) > ξ} = {x ∈ RN : f (x) > ψ(ξ)} = F +
Thus, by (4.1) and (4.3), (φ(f ))(x) = inf{ξ ≥ 0 : ρ+
ρ+
f (η) ≤ x} = φ(inf{η ≥ 0 : ρ+
by (4.1) and (4.3) because f ≥ 0 implies ρ+
φ(f )(ξ) = ∞ if ξ < 0. If ξ ≥ 0, then {x ∈ RN :
f (ψ(ξ)).
f (ψ(ξ)) ≤ x} = inf{φ(η) :
f (η) ≤ x} = f (x)
ψ(ξ), so that ρ+
f (η) ≤ x}). Now, inf{η ≥ 0 : ρ+
φ(f )(ξ) = ρ+
f (η) = ∞ if η < 0.
(cid:3)
ξ
Lemma 5.4. If φ is a Young function and f : RN → [0, ∞] is measurable, then
f φ ≤ f φ.
ξ
ξ ) ≤ µN (BF +
) that RRN f = R ∞
Proof. Since f ≥ 0 implies f ≥ 0, it follows from Theorem 4.3 (ii) and from
f . Upon
µN (F +
replacing f by φ(f ) in this inequality, it follows from Lemma 5.3 that RRN φ(f ) ≤
RRN φ( f ). Now, replace f with r−1f where r > 0 and use Theorem 4.3 (v) to
get RRN φ(r−1f ) ≤ RRN φ(r−1 f ) for every r > 0. By (3.3), this implies f φ ≤
fφ.
ξ )dξ ≤ R ∞
ξ )dξ = RRN
0 µN ( F +
0 µN (F +
(cid:3)
Theorem 5.5. If f, g : RN → [0, ∞] are Borel measurable and φ is a Young func-
tion, then φ(f ), φ(g) and φ(f ⊼ g) are measurable and nonnegative and
(5.1)
f ⊼ gφ ≤ 2N −1( f φ + gφ).
Proof. For the measurability of φ(f ), φ(g) and φ(f ⊼g), see Lemma 3.3 and Remark
3.1. That all three are nonnegative is trivial. In the proof of Lemma 3.3, we already
14
PATRICK J. RABIER
established that φ(f ⊼ g) = φ(f ) ⊼ φ(g). Therefore, by Lemma 5.2 for φ(f ) and φ(g)
and by Lemma 5.3, we get
(5.2)
ZRN
φ(f ⊼ g) ≤ 2N −1(cid:18)ZRN
φ( f ) +ZRN
φ(g)(cid:19) .
Suppose first that f φ > 0 and gφ > 0. Since (5.1) is trivial otherwise, we
may and shall assume f φ < ∞ and gφ < ∞. If so, 0 < r := f φ + gφ < ∞
and the inequality (5.2) for r−1f and r−1g is (use Theorem 4.3 (v))
21−N ZRN
φ(r−1(f ⊼ g)) ≤ ZRN
φ(r−1 f ) +ZRN
φ(r−1g).
With λ := r−1 f φ ∈ (0, 1), so that r−1 f = λ f −1
φ
λ)g−1
f and that r−1g = (1 −
φ g, this reads
21−N ZRN
(5.3)
φ(r−1(f ⊼ g)) ≤ ZRN
φ(cid:16)λ f−1
φ
f(cid:17) +ZRN
φ(cid:16)(1 − λ)g−1
φ g(cid:17) .
Since φ is convex and φ(0) = 0, then φ(µτ ) ≤ µφ(τ ) when τ ≥ 0 and 0 ≤ µ ≤ 1.
The choices µ = 21−N in the left-hand side of (5.3) and, next, µ = λ and µ = 1 − λ
in its right-hand side, yield
ZRN
φ(cid:0)21−N r−1(f ⊼ g)(cid:1) ≤ λZRN
φ(cid:16) f −1
φ
f(cid:17) + (1 − λ)ZRN
φ(cid:16)g−1
φ g(cid:17) .
By (3.5), it follows that RRN φ(cid:0)21−N r−1(f ⊼ g)(cid:1) ≤ 1 and so, by (3.3), f ⊼ gφ ≤
2N −1r = 2N −1( f φ + gφ), as claimed in (5.1).
To complete the proof, suppose now that f φ = 0 or gφ = 0. By symmetry,
we may and shall assume that f φ = 0, whence f = 0 a.e. Since f (x) is a
nonincreasing and right-continuous function of x, it follows that f = 0. Thus, by
Lemma 4.4, either f = 0 or there are x0 ∈ RN and 0 < z ≤ ∞ such that f (x0) = z
and f (x) = 0 when x 6= x0.
If f = 0, then f ⊼ g = 0 since g ≥ 0 and (5.1) is trivial.
If f (x0) = z > 0
and f (x) = 0 when x 6= x0, a straightforward calculation shows that (f ⊼ g)(x) =
min{g(x − x0), z}. In particular, 0 ≤ (f ⊼ g) ≤ g(· − x0), whence f ⊼ gφ ≤
g(· − x0)φ = gφ ≤ gφ, the latter by Lemma 5.4. This proves (5.1).
(cid:3)
Since (f ⊼f )(x) ≥ f (x/2), equality holds in (5.1) if Lφ = L1 and 0 ≤ g = f = f ∈
L1. Thus, 2N −1 is best possible among all the constants independent of φ. On the
other hand, even when Lφ = L∞, (5.1) is trivial only under additional assumptions,
namely, Mf = ess sup f, Mg = ess sup g and min{Mf , Mg} = ess sup(f ⊼ g) (which
is not implied by the former two, see Example 5.2 below). If so,
f ⊼ g∞ = min{f ∞, g∞},
(5.4)
which is much better than (5.1) since f ∞ ≤ f∞ and g∞ ≤ g∞ by Lemma
5.4. However, (5.4) is false without the extra assumptions mentioned above and
then (5.1) is no longer trivial in L∞. How badly (5.4) may fail is shown in:
Example 5.2. Let C ⊂ [0, 1] be the Cantor set. It is not hard to see that (x−C)∩C 6=
∅ for every x ∈ [0, 2] (notice that (x − Ck) ∩ Ck 6= ∅ for every k ∈ N, where C1 = [0, 1]
and Ck+1 ⊂ Ck is obtained by removing the open middle thirds of the intervals in
Ck). As a result, if f = ∞ on C and 0 outside, then f is Borel measurable, the
INEQUALITIES FOR INFIMAL CONVOLUTION
15
right-hand side of (5.4) with g = f is 0, but since f ⊼ f = ∞ on [0, 2], its left-hand
side is ∞.
6. Reverse inequalities for f (cid:3)g and f ⊻ g
With the help of Theorem 5.5, it is a simple matter to prove a converse of
Theorem 3.4.
Theorem 6.1. Suppose that f, g : RN → (−∞, ∞] are Borel measurable, that
mf , mg ∈ R and that mf + mg ≥ 0. Then f (cid:3)g ≥ 0, f − mf,g ≥ 0, g + mf,g ≥ 0 and
(6.1)
(f (cid:3)g)−1φ ≤ 2N −1(( f − mf,g)−1φ + (g + mf,g)−1φ),
for every Young function φ.
Proof. Use f − mf,g ≥ 0, g + mf,g ≥ 0 along with f (cid:3)g = (f − mf,g)(cid:3)(g + mf,g) ≥
(f − mf,g) ⊻ (g + mf,g) ≥ 0 to get f − mf,g ≥ 0 and g + mf,g ≥ 0 (by Theorem
4.5 (iii) and (iv)) as well as 0 ≤ (f (cid:3)g)−1 ≤ ((f − mf,g) ⊻ (g + mf,g))−1 = (f −
mf,g)−1 ⊼ (g + mf,g)−1. This yields
(f (cid:3)g)−1φ ≤ (f − mf,g)−1 ⊼ (g + mf,g)−1φ ≤
2N −1(((f − mf,g)−1)φ + ((g + mf,g)−1)φ,
where Theorem 5.5 was used for the second inequality.
Next, ((f − mf,g)−1) = ( f − mf,g)−1 and ((g + mf,g)−1) = (g + mf,g)−1 by
(cid:3)
Theorem 4.5 (viii) and (iv) (in that order), which proves (6.1).
The comments after Theorem 5.5 may be repeated: Without extra assumptions,
(6.1) is not trivial even when Lφ = L∞. The proof of Theorem 6.1 shows that, more
generally, if z ∈ R and −mg ≤ z ≤ mf (whence mf + mg ≥ 0),
(f (cid:3)g)−1φ ≤ 2N −1(( f − z)−1φ + (g + z)−1φ).
(6.2)
However, (6.1) is optimal (to prove (f (cid:3)g)−1 ∈ Lφ) among all the inequalities (6.2).
Indeed, if the right-hand side of (6.2) is finite for some z as above, the right-
hand side of (6.1) is also finite3. To see this, assume first mf + mg = 0. Then,
z = mf = −mg = mf,g is the only possible choice in (6.2) and (6.1) is optimal by
default. Suppose now that mf +mg > 0 and note that, by Theorem 4.5 (iii), f −mf,g
and g + mf,g are both bounded below by (mf + mg)/2 > 0. Thus, ( f − mf,g)−1
∈ L∞ and so, if f − z ≥ 0 and ( f − z)−1 ∈ Lφ for some z, then ( f − mf,g)−1 =
[1 − (z − mf,g)( f − mf,g)−1]( f − z)−1 ∈ Lφ since 1 − (z − mf,g)( f − mf,g)−1 ∈ L∞.
Likewise, if g + z ≥ 0 and (g + z)−1 ∈ Lφ, then (g + mf,g)−1 ∈ Lφ.
It is not clear whether 2N −1 is best possible in (6.1), among all the constants
independent of φ. The remark that (f (cid:3)f )(x) ≤ 2f (x/2) and the choice g = f =
f ≥ 0 with f −1 ∈ L1 = Lφ only shows that the best constant is at least 2N −2.
There is also a converse of Corollary 3.5:
Corollary 6.2. Suppose that f, g : RN → [−∞, ∞] are Borel measurable, that
g ≥ 0 and that f 6 ≡∞, g 6 ≡∞. Then, f ⊻ g ≥ 0 and
(6.3)
(f ⊻ g)−1φ ≤ 2N (cid:0)( f+ − mf+,g)−1φ + (g + mf+,g)−1φ(cid:1) .
3While mostly true, the converse may fail when mf + mg > 0 and z = mf = m f or z =
−mg = −mg .
16
PATRICK J. RABIER
for every Young function φ.
Proof. Since g is nonnegative, f ⊻ g = f+ ⊻ g ≥ (f+(cid:3)g)/2 = (f+/2)(cid:3)(g/2) ≥ 0.
Also, 0 ≤ mf+ , mg < ∞ since f 6 ≡∞ and g 6 ≡∞. Thus, it suffices to use Theorem
6.1 with f+/2 and g/2 along with Theorem 4.5 (v) and (ix).
(cid:3)
In many cases, Theorem 6.1 and Corollary 6.2 can be used to prove that (f (cid:3)g)−1 ∈
Lφ or (f ⊻ g)−1 ∈ Lφ without any calculation of f or g, because there are easily
verifiable sufficient conditions for the finiteness of the right-hand sides of (6.1) and
(6.3). The simplest one is given in the following lemma.
Lemma 6.3. Given f : RN → (−∞, ∞] with mf > −∞ (i.e., f bounded below),
suppose that there are constants c > 0 and α > 0 such that f (x) ≥ h(x) := cxα
for x large enough. Then, for every z < mf ,
( f − z)−1φ ≤ 2h−1φ,RN \B + (mf − z)−1χBφ.
(6.4)
for every open ball B centered at the origin such that4 f ≥ h ≥ 2z outside B and
every Young function φ. If φ is invertible with inverse ψ, this also reads
( f − z)−1φ ≤ 2h−1φ,RN \B + (mf − z)−1[ψ(µN (B)−1)]−1.
(6.5)
In particular, if h−1 ∈ Lφ(RN \B), then ( f − z)−1 ∈ Lφ. (Case in point: Since h is
explicitly known, h−1 ∈ Lφ(RN \B) can often be checked by a calculation.)
Proof. Let B be as in the theorem. By the "furthermore" part of Theorem 4.5
(vii), f ≥ h ≥ 2z outside B. Thus, h ≤ 2( f − z), whence ( f − z)−1 ≤ 2h−1, outside
B. Meanwhile, by Theorem 4.5 (iii), ( f − z)−1 ≤ (m f − z)−1 < (mf − z)−1 < ∞ in
B (and everywhere else) and (6.4) follows. To get (6.5) when φ is invertible with
inverse ψ, just notice that χBφ = [ψ(µN (B)−1)]−1 by (3.3).
(cid:3)
From the proof of Lemma 6.3, xα can be replaced with a function h(x) satisfying
general conditions. For example, if Lφ = Lp with p ≥ 1, Lemma 6.3 yields ( f −
z)−1 ∈ Lp if α > N/p but, if f (x) ≥ cxα(log x)β for large x, the choice of any
continuous strictly increasing function h of x that coincides with cxα(log x)β
for x ≥ 1 (say) shows that ( f − z)−1 ∈ Lp in the limiting case α = N/p if β > p−1.
Remark 6.1. The proof of Lemma 6.3 also shows that, more generally, (6.4) is
true with mf replaced with m f and that this requires only m f > −∞. This may
occasionally be useful, but rarely (see Remark 4.3).
It is more delicate to extend Lemma 6.3 when z = mf = m f . The extra difficulty
loc (because m f = ess inf f ; see Theorem 4.5 (iii)), so
is that ( f − mf )−1 /∈ L∞
that local integrability becomes an issue. This requires further investigation. We
only mention without proof (and will not use later) that if mf is a unique and
nondegenerate minimum of f (plus a mild technical condition) then ( f − mf )−1 ∈
Lp
loc if N ≥ 3 and 1 ≤ p < N/2.
A direct application of Lemma 6.3 yields the following sample result.
Theorem 6.4. Given f, g : RN → (−∞, ∞], suppose that there are constants c > 0
and α > 0 such that f (x) ≥ cxα and g(x) ≥ cxα for x large enough. Then:
(i) If mf , mg ∈ R and mf + mg > 0, then (f (cid:3)g)−1 ∈ Lp if p ≥ 1 and p > N/α.
(ii) If g ≥ 0 and mf+ + mg > 0, then (f ⊻ g)−1 ∈ Lp if p ≥ 1 and p > N/α.
4Since limx→∞ h(x) = ∞, the existence of B is not an issue.
INEQUALITIES FOR INFIMAL CONVOLUTION
17
Proof. (i) Just notice that mf,g < mf (≤ m f ) and −mf,g < mg (≤ mg) since
mf + mg > 0 and use Lemma 6.3 with Lφ = Lp and Theorem 6.1.
(ii) If mf+ = ∞ or mg = ∞, then f ⊻ g = ∞ and the result is trivial. From now
on, mf+ , mg < ∞, whence mf+ , mg ∈ R (because f+, g ≥ 0). By Theorem 4.5 (ix),
), whereas mf+,g < mf+ and −mf+,g < mg since mf+ + mg > 0. Now, use
f+ = (f+
Lemma 6.3 and Corollary 6.2 with Lφ = Lp.
(cid:3)
Lemma 6.3 and its aforementioned variants yield generalizations of Theorem 6.4
to all Orlicz spaces. The proof of Theorem 6.4 does not use the estimate (6.4) but
the next theorem, relevant to the results in the next section, does. As explained
after the proof, there is a good reason to confine attention to Lp spaces.
Theorem 6.5. Suppose that f, g : RN → (−∞, ∞] are Borel measurable, that
mf ≥ 0, mg > −∞ and that mg > 0 if mf = 0. Given 1 ≤ p ≤ ∞, suppose also
that for some α > N/p, there are constants c > 0 and α > 0 such that f (x) ≥ cxα
for x large enough. For t > 0, set
ft(x) := tf (t−1x),
so that mft = tmf . Lastly, assume (g−z)−1 ∈ Lp when z < mg (e.g., if g(x) ≥ cxα
for x large enough by Lemma 6.3, but this is not necessary).
Then, (ft(cid:3)g)−1 ∈ Lp when tmf + mg > 0 (i.e., t > 0 if mf = 0 and t > −m−1
if mf > 0) and:
(i) If mf = 0, then (ft(cid:3)g)−1p = O(tN/p) as t → ∞.
(ii) If mf > 0 and p > N, then limt→∞ (ft(cid:3)g)−1p = 0.
(iii) If mf > 0 and p ≥ N, then (ft(cid:3)g)−1p = O(t−1+N/p) as t → ∞.
f mg
Proof. As in Lemma 6.3, set h(x) := cxα and, in (6.4), let B be a ball centered
at the origin such that f ≥ h and h ≥ mf + mg ≥ 2mf,g = mf − mg outside
B. Evidently, ( f )t ≥ ht (:= th(t−1·)) outside tB and, by Theorem 4.5 (iv) and
) = ( f )t. Thus, ( f )t ≥ ht outside tB. Furthermore, if t ≥ 1, then ht ≥
(v), (ft
tmf + mg ≥ 2mft,g = tmf − mg outside tB.
Since ht(x) = t1−αcxα, it follows from the above that the estimate (6.4) can
be used with f, h and B replaced with ft, ht and tB, respectively, and with z =
mft,g = (tmf − mg)/2, provided that t ≥ 1 and that tmf + mg > 0 (so that
mft,g < mft = tmf ), which holds for large t. Accordingly,
(6.6)
( ft − mft,g)−1p ≤
2t−1+N/ph−1p,RN \B + 2(tmf + mg)−1tN/pµN (B)1/p,
where χtBp = tN/pµN (B)1/p was used. Since h−1p,RN \B < ∞ by the choice
α > N/p, it follows that ( ft − mft,g)−1 ∈ Lp. Also, (g + mft,g)−1 ∈ Lp since
−mft,g = (mg − tmf )/2 < mg when tmf + mg > 0 and since it is assumed that
(g − z)−1 ∈ Lp when z < mg. Thus, by (6.1), (ft(cid:3)g)−1 ∈ Lp.
The estimates (i), (ii) and (iii) follow from (6.1) and (6.6) and from the remarks
that (a) if mf = 0, then (mg > 0 and) (g + mft,g)−1 = (g − mg/2)−1 is independent
of t and (b) if mf > 0, then limt→∞ (g +mft,g)−1p = 0 by dominated convergence
if p < ∞ and by g + mft,g ≥ (tmf + mg)/2 if p = ∞.
(cid:3)
Similar estimates hold when f is replaced with ft in Corollary 6.2 and estimates
can also be worked out in other spaces Lφ, but the technicalities depend on φ. For
18
PATRICK J. RABIER
instance, while Theorem 6.5 remains true if Lp is replaced with L1 + Lp, the proof
is substantially more demanding (recall that L1 + Lp = Lφ with φ(τ ) = τ p in [0, 1]
and φ(τ ) = pτ + 1 − p in (1, ∞) if 1 ≤ p < ∞ and L1 + L∞ = Lφ with φ(τ ) = 0 in
[0, 1] and φ(τ ) = t−1 in (1, ∞); see e.g. [15])). Choices of h other than h(x) = cxα
often lead to challenging calculations.
7. Application to the Hamilton-Jacobi equations
We shall now apply the results of the previous sections to the Hamilton-Jacobi
equations in their simplest form (see Subsection 8.1 for a variant)
(7.1)
(cid:26) ut + H(∇u) = 0 on (0, ∞) × RN ,
u(0, ·) = g on RN ,
where the Hamiltonian H and the initial value g are given functions on RN .
Roughly speaking, when the Hamiltonian H (initial condition g) is convex, the
Hopf-Lax formula (Hopf formula) provides a solution of (7.1 ). In both cases, various
additional conditions are required of H and g and, as always, what constitutes a
solution is somewhat flexible. While the more recent work focuses on viscosity
solutions, other definitions exist as well.
Throughout this section, we assume that g, H : RN → (−∞, ∞], that mg ∈ R
(hence g 6 ≡∞), H 6 ≡∞ and that g is Borel measurable. Further assumptions will
be introduced when needed. It is once and for all understood that t > 0.
We denote by (tH)∗ the Legendre-Fenchel conjugate of tH, that is,
(tH)∗(x) := sup
y∈RN
(x · y − tH(y)) = tH ∗(t−1x)
Since (tH)∗ is always lsc, it is Borel measurable.
7.1. Solutions by the Hopf-Lax formula. In this subsection, H is convex and
H ∗∗(0) ∈ R. The Hopf-Lax formula (Hopf [14], Lax [19])
(7.2)
u(t, ·) = (tH)∗(cid:3)g,
is known to give a solution of (7.1) under various conditions about g. That g is real-
valued and continuous is a common assumption; see Bardi and Faggian [4] and the
references therein. The case when g is lsc and not everywhere finite was considered
by Imbert [16] and Stromberg [32]. Chen and Su [6] show that (7.2) is a solution
when g is real-valued, a.e. continuous and satisfies a condition weaker than upper
semicontinuity. Undoubtedly, other options can be found in the literature.
The inequality (3.9) in Theorem 3.4 can be used with f = (tH)∗ if m(tH)∗ +mg ≥
0. That mg > −∞ was assumed earlier, whereas m(tH)∗ = −tH ∗∗(0) ∈ R. As a
result, the condition m(tH)∗ + mg ≥ 0 is simply
(7.3)
− tH ∗∗(0) + mg ≥ 0,
so that m(tH)∗,g (see (3.8)) is given by
m(tH)∗,g = −(tH ∗∗(0) + mg)/2.
Thus, assuming (7.3), the corresponding inequality (3.9)
(7.4)
(2(tH)∗ + tH ∗∗(0) + mg)−1φ + (2g − tH ∗∗(0) − mg)−1φ ≤
2u(t, ·)−1φ
INEQUALITIES FOR INFIMAL CONVOLUTION
19
and the reverse inequality (6.1) of Theorem 6.1
(7.5)
u(t, ·)−1φ ≤
2N −1(cid:16)(2((tH)∗) + tH ∗∗(0) + mg)−1φ + (2g − tH ∗∗(0) − mg)−1φ(cid:17) ,
hold for every Young function φ.
Given α > 1, call α′ := α/(α − 1) > 1 the Holder conjugate of α. It is easily
checked and certainly folklore that if there is a constant d > 0 such that H(x) ≤
dxα′
for x large enough, then H ∗(x) ≥ cxα for x large enough. Consistent
with Theorem 6.4, it follows that if also g(x) ≥ cxα for x large enough, then
u(t, ·)−1 ∈ Lp for every p ≥ 1, p > N/α.
Furthermore, since (tH)∗ = tH ∗(t−1·), Theorem 6.5 gives estimates for u(t, ·)−1p
as t → ∞ if H ∗∗(0) < 0 or if H ∗∗(0) = 0 and mg > 0. (If H ∗∗(0) > 0, then (7.5)
breaks down when t > mg/H ∗∗(0).) The accuracy (or possible lack thereof) of these
estimates can be evaluated by using the inequality (7.4) with · φ = · p.
We now look at two classical examples in more detail. In both cases, H ∗∗(0) =
H(0) = 0 will make the inequalities simpler, but confines the discussion to mg ≥ 0.
Example 7.1. Suppose that H(x) := x2/2, so that (tH)∗(x) = x2/2t and (7.3)
boils down to mg ≥ 0 since H ∗∗ = H. If 1 ≤ p < ∞, a quick calculation shows that
(2(tH)∗ + mg)−1p = A(p, mg)tN/2p where A(p, mg) > 0 is a constant which is
finite if and only if mg > 0 and p > N/2. Accordingly, from (7.4), u(t, ·)−1 /∈ Lp
for any t > 0 if mg = 0 or if N ≥ 2 and 1 ≤ p ≤ N/2, even if u(0, ·)−1 = g−1 ∈ Lp.
Assume now mg > 0 and p > N/2. By (7.4) with H ∗∗(0) = 0 and since 0 ≤ g ≤
2g − mg ≤ 2g,
(7.6)
u(t, ·)−1p ≥ c(tN/2p + g−1p),
for some constant c > 0 depending only upon N, p and mg. Thus, once again,
u(t, ·)−1 /∈ Lpif g−1 /∈ Lp.
Conversely, since ((tH)∗) = (tH)∗ by Theorem 4.5 (i), it follows from (7.5) and
from 0 ≤ g ≤ 2g − mg ≤ 2g (since mg ≥ mg; see Theorem 4.5 (iii)) that u(t, ·)−1 ∈
Lp for every t > 0 if g−1 ∈ Lp and that u(t, ·)−1p ≤ C(tN/2p + g−1p), where
C > 0 depends only upon N, p and mg. In particular, u(t, ·)−1p = O(tN/2p) as
t → ∞ (which is sharp because of (7.6)). Note that the general estimate of Theorem
6.5 gives only the less precise O(tN/p).
Even though u(t, ·)−1 /∈ Lp if mg = 0, this does not preclude u(t, ·)−1 ∈ Lφ for Orlicz
spaces outside the Lp scale. For instance, if Lφ = L1 + Lp, a simple calculation5
shows that (2(tH)∗ + mg)−1 ∈ L1 + Lp if p > N/2, with the restriction N ≥ 3
if mg = 0. If so, by (7.5), u(t, ·)−1 ∈ L1 + Lp if g−1 ∈ L1 + Lp and, by another
calculation, u(t, ·)−1L1+Lp = O(t) as t → ∞ (optimal by (7.4)). In particular,
this holds if g−1 ∈ Lq with 1 ≤ q ≤ p and p > N/2. This complements the Lp
discussion above, even when mg > 0.
Example 7.2. Suppose that H(x) = x, so that (tH)∗ is the indicator function of
the closed ball B(0, t). Thus, the formula (7.2 ) is simply u(t, x) = inf y≤t g(x − y).
Once again, H ∗∗ = H, so that (7.3) amounts to mg ≥ 0 and, if φ is a Young
function, then (with ωN := µN (B(0, 1))
(7.7)
(2(tH)∗ + mg)−1φ = inf (cid:8)r > 0 : φ(cid:0)r−1m−1
g (cid:1) ωN tN ≤ 1(cid:9) ,
5The formula for φ was given at the end of the previous section.
20
PATRICK J. RABIER
which is ∞ if mg = 0. Thus, by (7.4), u(t, ·)−1 /∈ Lφ for any t > 0 and any φ if
mg = 0, even if u(0, ·)−1 = g−1 ∈ Lφ. (Since u(t, x)−1 = supy≤t g(x − y)−1, this
also follows from the remark that for every ε > 0, u(t, ·)−1 ≥ ε−1 on some ball of
radius t.) Assuming from now on that mg > 0, it follows from (7.7) that if φ has
an inverse ψ on [0, ∞], then (2(tH)∗ + mg)−1φ = m−1
(where, as
usual, ψ−1 = 1/ψ, not φ). In the simple case when φ(τ ) = τ p with p ≥ 1, this
yields
N t−N(cid:1)−1
g ψ(cid:0)ω−1
(7.8)
(2(tH)∗ + mg)−1p = m−1
g ω1/p
N tN/p.
Once again, ((tH)∗) = (tH)∗ by Theorem 4.5 (i). Thus, if g−1 ∈ Lp (equivalent
to (2g − mg)−1 ∈ Lp), it follows from (7.5) and (7.8) that u(t, ·)−1 ∈ Lp and that
u(t, ·)−1p = O(tN/p) as t → ∞ (which is sharp because of (7.4)). This is the
general estimate in Theorem 6.5 (i) which, in this example, is therefore optimal.
In Example 7.2, g−1 ∈ Lp is not enough to get u(t, ·)−1 ∈ Lp : If N = 1 and if
g = f −1/p with p < ∞ and f from Example 5.1, then mg = 1, g−1 = f 1/p ∈ Lp,
but u(t, x) = inf y≤t f −1/p(x − y) = 1 if t ≥ 1, so that u(t, ·)−1 = 1 /∈ Lp.
8. Solutions by the Hopf formula
The Hopf formula (Hopf [14])
u(t, ·) = (tH + g∗)∗
gives a solution of (7.1) when g is convex and various other technical assumptions
are satisfied. See for instance Bardi and Evans [3]. In Penot and Volle [28], g and
H can be extended real-valued. Below, we assume that g is lsc. Since u(t, ·) is
convex, there is no measurability issue.
First, −g∗(0) = inf g = mg (finite, as assumed above) by definition of g∗. Next, if
h, k : RN → (−∞, ∞] and h 6 ≡∞, k 6 ≡∞ (so that h∗ and k∗ are proper), it is well-
known and easily checked that (h + k)∗ ≤ h∗(cid:3)k∗. As a result, u(t, ·) = (tH + g∗)∗ ≤
(tH)∗(cid:3)g∗∗ = (tH)∗(cid:3)g. On the other hand, by using "inf sup ≥ sup inf", we get
inf x(tH +g∗)∗(x) ≥ supy inf x(x·y−(tH +g∗)(y)) = −tH(0)−g∗(0) = −tH(0)+mg.
Indeed, inf x(x·y −(tH +g∗)(y)) = −∞ if y 6= 0 because tH and g∗ are proper. This
shows that if −tH(0) + mg ≥ 0, then 0 ≤ u(t, ·) ≤ (tH)∗(cid:3)g. Furthermore, since
(tH)∗ = (tHC )∗ where HC = H ∗∗ is the closed convex hull of H, it follows that
m(tH)∗ = −tHC(0) ≥ −tH(0). To ensure that m(tH)∗ ∈ R, i.e., that HC (0) > −∞,
it must be assumed that H is bounded below by an affine function. If so, it follows
from Theorem 3.4 with f = (tH)∗ that if −tH(0)+ mg ≥ 0 (hence −tHC(0)+ mg ≥
0), then
(8.1)
(2(tH)∗ + tHC (0) + mg)−1φ + (2g − tHC (0) − mg)−1φ ≤
2u(t, ·)−1φ,
for every Young function φ.
Since the inequality (8.1) depends only upon HC , it remains true when H is
replaced with any proper closed convex function K ≤ H (so that K = KC) under
the same assumption −tH(0) + mg ≥ 0 as above (still needed to ensure u ≥ 0)
because this substitution decreases the left-hand side. This is less accurate, but
often more convenient for practical evaluation.
INEQUALITIES FOR INFIMAL CONVOLUTION
21
As an illustration of this point, it follows from the discussion in Example 7.1 that
u(t, ·)−1 /∈ Lp if H(x) ≥ x2/2 and either 0 < p ≤ N/2 or mg = 0. Alternatively,
from Example 7.2, u(t, ·)−1 /∈ Lφ for any φ if H(x) ≥ x and mg = 0.
In the opposite direction, if L ≥ H is any lsc convex function, then (tH + g∗) ≤
(tL+g∗) and, since both tL and g∗ are proper and lsc, (tL+g∗)∗ = (tL)∗(cid:3)g as soon
as the relative interiors of dom L and dom g∗ have nonempty intersection ([29, p.
145]). If so, (tL)∗(cid:3)g ≤ u(t, ·), so that the reverse inequalities of Theorem 6.1 can be
used with f = (tL)∗ if −tL(0)+mg ≥ 0. Since 0 ≤ −tL(0)+mg = m(tL)∗(cid:3)g ≤ u(t, ·),
it follows that
(8.2)
u(t, ·)−1φ ≤
2N −1(cid:16)(2((tL)∗) + tL(0) + mg)−1φ + (2g − tL(0) − mg)−1φ(cid:17) ,
for every Young function φ. Unlike (8.1), the inequality (8.2) does not require H to
be bounded below by an affine function.
For instance, by Example 7.1, u(t, ·)−1p = O(tN/2p) as t → ∞ if H(x) ≤
L(x) := x2/2, mg > 0 and g−1 ∈ Lp, p > N/2. Also, by Example 7.2, u(t, ·)−1p =
O(tN/p) as t → ∞ if H(x) ≤ L(x) := x, mg > 0 and g−1 ∈ Lp, p ≥ 1.
8.1. Explicit solutions of related problems. Various explicit formulas for the
solution of
(8.3)
(cid:26) ut + H(u, ∇u) = 0 on (0, ∞) × RN ,
u(0, ·) = g on RN ,
when the Hamiltonian depends upon u, have been obtained under suitable (but
restrictive) conditions. The result most directly relevant to this paper can be found
in the work of Barron et al. [5], complemented and generalized in [2]. It is shown in
[2, Theorem 6.11] that if H is continuous on RN +1, with H(s, x) nondecreasing in
s ∈ R, convex and positively homogeneous of degree 1 in x ∈ RN , and if g : RN →
(−∞, ∞] is lsc, then (for t > 0)
(8.4)
u(t, ·) := h[t]
⊻ g,
is the minimal lsc supersolution of (8.3), where6 h[t](x) := h(cid:0)t−1x(cid:1) and
h(x) := inf{s ∈ R : H(s, ·)∗(x) ≤ 0}.
(8.5)
It is shown in [5] that h is quasiconvex and lsc and that mh = −∞.
⊻ g(= h[t]
From now on, we assume g ≥ 0, so that u(t, ·) = h[t]+
⊻ g). Note that
h[t]+ = h+[t]. Since mh = −∞ implies mh[t] = −∞, it follows that mh[t]+ = 0, so
that mh[t]+,g = −mg/2 ≤ 0.
Corollary 3.5 and Corollary 6.2 with f = h[t] can be used to evaluate u(t, ·)−1φ.
Specifically,
(2h[t]+ + mg)−1φ + (2g − mg)−1φ ≤ 2u(t, ·)−1φ
(8.6)
and
(8.7)
u(t, ·)−1φ ≤ 2N +1(cid:16)(2(h[t]+
) + mg)−1φ + (2g − mg)−1φ(cid:17) ,
for every Young function φ.
6This differs from ht previously defined by th(t−1·).
22
PATRICK J. RABIER
Example 8.1. Let H(s, x) = (s+)αx with α > 0. When g ≥ 0, the (nonnegative)
solution (8.4) of (8.3) actually solves ut + uα∇u = 0. By (8.5) and a straightfor-
ward calculation, h(0) = −∞ and h(x) = x1/α if x 6= 0.
If mg = 0, then u(t, ·)−1 /∈ Lp for any p ≥ 1 by (8.6) since h−1
[t]+ = t1/α·−1/α /∈ Lp.
However, it is readily checked that h−1
[t]+ ∈ L1 + Lp if 1 < N α < p. If so, it fol-
lows from (8.7) with mg = 0 that u(t, ·) ∈ L1 + Lp if g−1 ∈ L1 + Lp (hence
g−1 ∈ L1 + Lp by (8.6)) and the calculation of the estimates (8.6) and (8.7) is
) by Theorem 4.5 (i).
trivial since mg = 0 and since h[t]+ = t−1/α · 1/α = (h[t]+
))−1L1+Lp = Ct1/α with C := · −1/α L1+Lp,
Thus, (h[t]+)−1L1+Lp = ((h[t]+
which yields u(t, ·)−1L1+Lp = O(t1/α) as t → ∞ (optimal).
If mg > 0, then (2h[t]+ + mg)−1 ∈ Lp if and only if p > N α. If so and if g−1 ∈ Lp
(equivalent to (2g − mg)−1 ∈ Lp), it follows from (8.7) that u(t, ·)−1 ∈ Lp with
u(t, ·)−1p = O(tN/p) as t → ∞ (optimal by (8.6)).
Example 7.2 is recovered when α = 0 in Example 8.1. If so, h(x) = −∞ on the
closed unit ball and h(x) = ∞ outside, so that h+ is the indicator function of the
closed unit ball. Even though the formulas (7.2) and (8.4) look different, they both
provide the same solution u(t, x) = inf y≤t g(x − y).
Example 8.2. Let H(s, x) = esx, so that, by (8.5), h(0) = −∞ and h(x) = ln x
if x 6= 0. Thus, h+(x) = ln+ x is continuous, radially symmetric and nondecreasing
) = h[t]+ by Theorem 4.5 (i). As always in this
in x, so that, once again, (h[t]+
subsection, g ≥ 0.
Since h[t]+ = 0 on the ball with center 0 and radius t, it follows from (8.6) that
u(t, ·)−1 /∈ Lφ for any t > 0 and any Young function φ if mg = 0. In addition, the
slow growth of ln x reveals that u(t, ·)−1 /∈ Lp, 1 ≤ p < ∞, even if mg > 0.
However, if mg > 0, then (2h[t]+ + mg)−1 ∈ Lφ if φ vanishes fast enough at the
origin. Aside from L∞, one of the simplest examples is given by the Young function
φ(τ ) := eτ −τ −2
. (The growth of φ at infinity could be damped considerably to enlarge
the space Lφ.) Thus, if mg > 0 and g−1 ∈ Lφ (equivalent to (2g − mg)−1 ∈ Lφ),
it follows from (8.7) that u(t, ·)−1 ∈ Lφ for every t > 0. We did not attempt to
estimate u(t, ·)−1φ as t → ∞.
References
[1] Adams, R. A. and Fournier, J. F., Sobolev spaces, 2nd edition, Academic Press, Boston 2003.
[2] Alvarez, O., Barron, E. N. and Ishii, H., Hopf-Lax formulas for semicontinuous data, Indiana
Univ. Math. J. 48 (1999) 993-1035.
[3] Bardi, M. and Evans, L. C., On Hopf's formulas for solutions of Hamilton Jacobi equations,
Nonlinear Anal. TMA 8 (1984) 1373-1381.
[4] Bardi, M. and Faggian, S., Hopf-type estimates and formulas for nonconvex-nonconcave
Hamilton-Jacobi equations, SIAM J. Math. Anal. 29 (1998) 1067-1086.
[5] Barron, E. N., Jensen, R. and Liu, W., Hopf Lax-type formula for ut + H(u, Du) = 0, J.
Differ. Equations 126 (1996) 49-61.
[6] Chen, G.-Q. and Su, B., Discontinuous solutions for Hamilton-Jacobi equations: uniqueness
and regularity, Discrete Contin. Dyn. Syst. 9 (2003) 167-192.
[7] Day, P. W., Rearrangement inequalities, Canadian J. Math. 24 (1972) 930-943.
[8] Federer, H., Geometric measure theory, Springer-Verlag, Berlin (1996).
[9] Fenchel, W., Convex cones, sets and functions, lecture notes, Princeton Univ., Princeton,
N.J. 1953.
[10] Gardner, R. J., The Brunn-Minkowski inequality, Bull. Amer. Math. Soc. 39 (2002) 355-405.
INEQUALITIES FOR INFIMAL CONVOLUTION
23
[11] Glavosits, T. and Sz´az, ´A., The infimal convolution can be used to easily prove the classical
Hahn-Banach theorem, Rostock. Math. Kolloq. 65 (2010) 71-83.
[12] Hiriart-Urruty, J.-B., Extension of Lipschitz functions, J. Math. Anal. Appl. 77 (1980) 539-
554.
[13] Hiriart-Urruty, J.-B. and Lemar´echal, C., Convex analysis and minimization algorithms (2
volumes), Springer-Verlag, Berlin 1996.
[14] Hopf, E., Generalized solutions of non linear equations of first-order, J. Math. Mech. 14
(1965) 951-973.
[15] Hudzik, H., Intersections and algebraic sums of Musielak-Orlicz spaces, Portugaliae Math.
40 (1981) 287-296.
[16] Imbert, C., Convex analysis techniques for Hopf-Lax formulae in Hamilton-Jacobi equations,
J. Nonlinear Convex Anal. 2 (2001) 333-343.
[17] Jung, H. W. E, Uber die kleinste Kugel, die eine raumliche Figur einschliesst, J. Reine Angew.
Math. 123 (1901) 241-257.
[18] Krasnosel'skii, M. A. and Rutickii, Ya. B., Convex functions and Orlicz spaces, P. Noordhoff
Ltd, Groningen 1961.
[19] Lax, P. D., Hyperbolic systems of conservation laws II, Commun. Pure Appl. Math. 10 (1957)
537-566.
[20] Lions, P.-L., Generalized Solutions of Hamilton-Jacobi Equations, Pitman, London, 1982.
[21] Lucet, Y., What shape is your conjugate? A survey of computational convex analysis and its
applications, SIAM Rev. 52 (2010) 505-542.
[22] Luxemburg, W. A. J., Rearrangement invariant Banach function spaces, Queen's Papers on
Pure and Applied Math. 10 (1967) 83-144.
[23] Mazure, M.-L., Equations de convolution et formes quadratiques, Ann. Mat. Pura Appl. 158
(1991) 75-97.
[24] Moreau, J. J., Fonctions `a valeurs dans [−∞, ∞]; notions alg´ebriques, Universit´e de Mont-
pellier, Montpellier 1963.
[25] Moreau, J. J., Inf-convolution des fonctions num´eriques sur un espace vectoriel, C. R. Acad.
Sci. Paris 256 (1963) 5047-5049.
[26] Moreau, J. J., Inf-convolution, sous-additivit´e, convexit´e des fonctions num´eriques, J. Math.
Pures Appl. 49 (1970) 109-154.
[27] O'Neil, R., Fractional integration in Orlicz spaces, Trans. Amer. Math. Soc. 115 (1965)
300-328.
[28] Penot, J.-P. and Volle, M., Explicit solutions to Hamilton-Jacobi equations under mild con-
tinuity and convexity assumptions. J. Nonlinear Convex Anal. 1 (2000) 177-199.
[29] Rockafellar, R. T., Convex analysis, Princeton University Press, Princeton 1970.
[30] Sierpi´nski, W., Sur la question de la mesurabilit´e de la base de M. Hamel, Fund. Math. 1
(1920) 105-111.
[31] Stromberg, T., The operation of infimal convolution, Dissertationes Math. 352 (1996).
[32] Stromberg, T., The Hopf-Lax formula gives the unique viscosity solution, Differential Integral
Equations 15 (2002) 47-52.
[33] Ziemer, W. P., Weakly differentiable functions, Springer-Verlag, New York 1989.
Department of mathematics, University of Pittsburgh, Pittsburgh, PA 15260
E-mail address: [email protected]
|
1306.0358 | 2 | 1306 | 2013-09-17T10:01:40 | Best proximity pair results for relatively nonexpansive mappings in geodesic spaces | [
"math.FA"
] | Given $A$ and $B$ two nonempty subsets in a metric space, a mapping $T : A \cup B \rightarrow A \cup B$ is relatively nonexpansive if $d(Tx,Ty) \leq d(x,y) \text{for every} x\in A, y\in B.$ A best proximity point for such a mapping is a point $x \in A \cup B$ such that $d(x,Tx)=\text{dist}(A,B)$. In this work, we extend the results given in [A.A. Eldred, W.A. Kirk, P. Veeramani, Proximal normal structure and relatively nonexpansive mappings, Studia Math., 171 (2005), 283-293] for relatively nonexpansive mappings in Banach spaces to more general metric spaces. Namely, we give existence results of best proximity points for cyclic and noncyclic relatively nonexpansive mappings in the context of Busemann convex reflexive metric spaces. Moreover, particular results are proved in the setting of CAT(0) and uniformly convex geodesic spaces. Finally, we show that proximal normal structure is a sufficient but not necessary condition for the existence in $A \times B$ of a pair of best proximity points. | math.FA | math |
Best proximity pair results for relatively nonexpansive mappings in
geodesic spaces
Aurora Fern´andez-Le´on1 , Adriana Nicolae2,3
1Dpto. de An´alisis Matem´atico, Universidad de Sevilla, P.O. Box 1160, 41080-Sevilla, Spain
2 Department of Mathematics, Babe¸s-Bolyai University, Kogalniceanu 1, 400084 Cluj-Napoca, Romania
3 Simion Stoilow Institute of Mathematics of the Romanian Academy, Research group of the pro ject PD-3-0152,
P.O. Box 1-764, RO-014700 Bucharest, Romania
E-mail: [email protected], [email protected]
Abstract
Given A and B two nonempty subsets in a metric space, a mapping T : A ∪ B → A ∪ B is
relatively nonexpansive if d(T x, T y ) ≤ d(x, y ) for every x ∈ A, y ∈ B . A best proximity point
for such a mapping is a point x ∈ A ∪ B such that d(x, T x) = dist(A, B ). In this work, we
extend the results given in [A.A. Eldred, W.A. Kirk, P. Veeramani, Proximal normal structure
and relatively nonexpansive mappings, Studia Math. 171, 283–293 (2005)] for relatively nonex-
pansive mappings in Banach spaces to more general metric spaces. Namely, we give existence
results of best proximity points for cyclic and noncyclic relatively nonexpansive mappings in the
context of Busemann convex reflexive metric spaces. Moreover, particular results are proved in
the setting of CAT(0) and uniformly convex geodesic spaces. Finally, we show that proximal
normal structure is a sufficient but not necessary condition for the existence in A × B of a pair
of best proximity points.
MSC: Primary 54E40, 47H10.
Keywords: Relatively nonexpansive mapping, best proximity pair, best proximity point, proxi-
mal normal structure, Busemann convexity.
1
Introduction
Although metric fixed point theory is primary concerned with the existence of fixed points of map-
pings that satisfy certain restrictions, there exist many other related problems that have attracted
a high amount of interest from researchers in the area. One of such problems consists in studying
the existence of approximate solutions of the equation x = T x in the absence of fixed points of the
mapping T . A point x ∈ X is said to be an approximate solution of the equation x = T x if x is
“close” to T x is some sense. Depending on the considered closeness condition between x and T x,
results of different nature have been obtained in the literature. One classical result in this direction
due to Ky Fan [16] states that if A is a compact, convex and nonempty subset of a locally convex
Hausdorff topological vector space X and T is a continuous mapping from A to X , then there exists
a point x ∈ A such that d(x, T x) = d(T x, A), where d is the semi-metric induced by a continuous
semi-norm defined on X . If, instead of considering the condition d(x, T x) = d(T x, A), one requires
that x is an absolute optimal approximate solution, that is, d(x, T x) = dist(A, B ) either for non-self
mappings T : A → B or for mappings T : A ∪ B → A ∪ B such that T (A) ⊆ B , T (B ) ⊆ A or
T (A) ⊆ A, T (B ) ⊆ B , existence, uniqueness and convergence results for such points are known as
1
best proximity point theorems. Note that the notion of best proximity point also refers to such a
type of approximate solution. In the present work we mainly focus on the study of best proximity
points for certain self-mappings T : A ∪ B → A ∪ B satisfying the above inclusion relations. The
first results concerning such mappings were given by Kirk, Srinivasan and Veeramani [21] in 2003.
More precisely, it was proved that if A and B are two nonempty and closed subsets of a complete
metric space, T : A ∪ B → A ∪ B is such that T (A) ⊆ B , T (B ) ⊆ A and there exists k ∈ (0, 1) such
that
d(T x, T y) ≤ kd(x, y) for every x ∈ A, y ∈ B ,
then A ∩ B contains a fixed point of T .
In the last years, many generalizations of this problem have appeared under the assumption
A ∩ B = ∅. In this respect, weaker metric conditions have been considered for the mapping T . This
is, for instance, the case of cyclic contractions [9, 29, 13], cyclic Meir-Keeler contractions [8, 27] or
relatively nonexpansive mappings [10, 11, 28]. Relatively nonexpansive mappings were introduced
by Eldred, Kirk and Veeramani [10] in the following way: a self-mapping T : A ∪ B → A ∪ B is
relatively nonexpansive if
d(T x, T y) ≤ d(x, y) for every x ∈ A, y ∈ B .
If, in addition, T (A) ⊆ B and T (B ) ⊆ A then T is said to be cyclic. Likewise, if T (A) ⊆ A and
T (B ) ⊆ B , then T is called noncyclic.
In [10], several best proximity point results were given in Banach spaces for cyclic and noncyclic
relatively nonexpansive mappings. While many generalizations of best proximity point results from
the linear setting to metric spaces have appeared in the literature (see, among others, the works
[13, 17] on cyclic contractions in metric spaces), no result has been given for relatively nonexpansive
mappings in a nonlinear setting. Here we address this problem extending results proved in [10]
in the context of Banach spaces to Busemann convex reflexive metric spaces. We also give more
particular results in the setting of CAT(0) and uniformly convex geodesic spaces. Furthermore, we
prove an analogue of a result due to Karlovitz [20] showing that proximal normal structure is a
sufficient but not necessary condition for the existence in A × B of a pair of best proximity points.
2 Preliminaries
In this section we compile the main concepts and results we will work with along this paper. We
begin with some basic definitions and notations that are needed. Let (X, d) be a metric space and
consider A and B two subsets of X . Define
d(x, A) = inf {d(x, y) : y ∈ A};
dist(A, B ) = inf {d(x, y) : x ∈ A, y ∈ B };
δ(x, A) = sup{d(x, y) : y ∈ A};
δ(A, B ) = sup{d(x, y) : x ∈ A, y ∈ B }.
From now on, B (a, r) denotes the closed ball in the space X centered at a ∈ X with radius r > 0.
The metric projection PA onto A is the mapping
PA (x) = {z ∈ A : d(x, z ) = dist(x, A)} for every x ∈ X.
When this mapping is well-defined and singlevalued we use the same notation PA (x) to denote the
unique point belonging to this set.
2
In the sequel, we say that a pair of sets (A, B ) has a property if each of the sets A and B has
this property. For instance, we say that the pair (A, B ) is closed and bounded if A and B are
both closed and bounded. A very important property in this paper for a pair of sets is the one of
proximity.
Definition 2.1. A pair (A, B ) of subsets of a metric space is said to be proximal if for each
(a, b) ∈ A × B there exists (a′ , b′ ) ∈ A × B such that d(a, b′ ) = d(a′ , b) = dist(A, B ).
In this context, given a pair of sets (A, B ) in a metric space, we say that the point p ∈ A is
a proximal point of q ∈ B (with respect to A and B ) if d(p, q) = dist(A, B ). Then, (p, q) is also
called pair of proximal points.
In this paper we will mainly work with geodesic spaces. A metric space (X, d) is said to be a
(uniquely) geodesic space if every two points x and y of X are joined by a (unique) geodesic, i.e, a
map c : [0, l] ⊆ R → X such that c(0) = x, c(l) = y , and d(c(t), c(t′ )) = t − t′ for all t, t′ ∈ [0, l].
The image c([0, l]) of such a geodesic forms a geodesic segment which joins x and y and it is not
necessarily unique. If no confusion arises, we use [x, y ] to denote a geodesic segment joining x and
y . A point z in X belongs to a geodesic segment [x, y ] if and only if there exists t ∈ [0, 1] such that
d(x, z ) = td(x, y) and d(y , z ) = (1 − t)d(x, y) and we write z = (1 − t)x + ty for simplicity. Notice
2 , we often use the notation x+y
to denote 1
2 x + 1
that this point may not be unique. When t = 1
2 y .
2
Any Banach space is a geodesic space with usual segments as geodesic segments. A subset A of a
geodesic space X is said to be convex if any geodesic segment that joins each pair of points x and
y of A is contained in A. A geodesic triangle △(x, y , z ) in X consists of three points x, y , z ∈ X
(the vertices of △) and three geodesic segments joining each pair of vertices (the edges of △). For
more about geodesic spaces the reader can check [3, 5, 26].
A metric d : X × X → R is said to be convex if for any x, y , z ∈ X one has
d(x, (1 − t)y + tz ) ≤ (1 − t)d(x, y) + td(x, z ) for all t ∈ [0, 1].
A geodesic space (X, d) is Busemann convex (introduced in [6]) if given any pair of geodesics
c1 : [0, l1 ] → X and c2 : [0, l2 ] → X one has
d(c1 (tl1 ), c2 (tl2 )) ≤ (1 − t)d(c1 (0), c2 (0)) + td(c1 (l1 ), c2 (l2 )) for all t ∈ [0, 1].
It is well-known that Busemann convex spaces are uniquely geodesic and with convex metric.
Given two geodesic segments [x, z ] and [y , w] in a uniquely geodesic space (X, d), we say that
[x, z ] is paral lel to [y , w], and we denote it by [x, z ]k[y , w], if d(x, y) = d(m1 , m2 ) = d(z , w), where
2 and m2 = y+w
m1 and m2 are the midpoints of [x, z ] and [y , w], respectively (that is, m1 = x+z
2 ).
The following property was given by Busemann in [7]. For the convenience of the reader we include
a proof of this fact.
Proposition 2.2. Let x, y , z , w be four points in a Busemann convex geodesic space. Suppose
[x, z ]k[y , w]. Then, [x, y ]k[z , w].
2 , m3 = x+y
2 , m2 = y+w
Proof. Let m1 = x+z
2 and m4 = z+w
2 . Let r = d(x, y) = d(z , w) = d(m1 , m2 ).
By using the Busemann convexity of the space we have that d(m1 , y+z
2 ) ≤ r/2, d(m1 , w+x
2 ) ≤ r/2,
d(m2 , y+z
2 ) ≤ r/2 and d(m2 , w+x
2 ) ≤ r/2, which implies
r = d(m1 , m2 ) ≤ d(cid:18)m1 ,
2 (cid:19) + d(cid:18) y + z
, m2(cid:19) ≤
y + z
2
2 (cid:19) + d(cid:18) w + x
, m2(cid:19) ≤
r = d(m1 , m2 ) ≤ d(cid:18)m1 ,
w + x
2
3
= r.
and
+
= r
r
2
r
2
r
2
+
r
2
Consequently, y+z
2 . Let m5 = m1+m2
2 = m1+m2
= w+x
. Again by the Busemann convexity,
2
2
d(m4 , m5 ) ≤ 1/2d(x, z ) and d(m4 , m5 ) ≤ 1/2d(w, y), which implies d(m4 , m5 ) ≤ 1/2 min{d(x, z ),
d(w, y)}. Since m5 is also the midpoint between m3 and m4 , d(m3 , m4 ) ≤ min{d(x, z ), d(w, y)}.
Suppose now that d(x, z ) < d(y , w). Then d(m3 , m4 ) ≤ d(x, z ) < d(y , w). Since [m3 , m4 ]k[y , w], we
can proceed similarly to get p1 = m3+y
and s1 = m4+w
such that d(p1 , s1 ) ≤ d(x, z ) < d(y , w). By
2
2
repeating the process, we obtain the sequences {pn}, {sn } ⊆ X , where, for n ≥ 2, pn = y+pn−1
and
2
sn = w+sn−1
, with d(sn , w) = d(pn , y) = r/2n+1 and d(pn , sn ) ≤ d(x, z ) < d(y , w).
2
Since d(y , w) ≤ d(y , pn ) + d(pn , sn ) + d(sn , w) for every n ∈ N, we may take superior limit in
the previous inequality to get d(y , w) ≤ d(x, z ) < d(y , w), which is a contradiction and the result
follows.
In the sequel, we will also need the notion of uniformly convex geodesic space (see also [18,
pg. 107]). A geodesic metric space (X, d) is said to be uniformly convex if for any r > 0 and any
ε ∈ (0, 2] there exists δ ∈ (0, 1] such that for all a, x, y ∈ X with d(x, a) ≤ r , d(y , a) ≤ r and
d(x, y) ≥ εr ,
d(m, a) ≤ (1 − δ)r,
where m stands for a midpoint of x and y . A mapping δ : (0, +∞) × (0, 2] → (0, 1] providing such
a δ = δ(r, ε) for a given r > 0 and ε ∈ (0, 2] is called a modulus of uniform convexity of X . If δ
decreases with r (for each fixed ε) we say that δ is a monotone modulus of uniform convexity of X
(introduced in [23]). If δ is lower semicontinuous from the right with respect to r (for each fixed
ε), then we say δ is a lower semicontinuous from the right modulus of uniform convexity of X .
If in the above definition we drop the uniformity conditions then we find the notion of strict
convexity in metric spaces. Consequently, every uniformly convex geodesic space is strictly convex.
Moreover, it is easy to see that every Busemann convex metric space is strictly convex [15] and
that strictly convex metric spaces are uniquely geodesic.
A very important class of geodesic spaces are CAT(0) spaces, that is, metric spaces of non-
positive curvature in the sense of Gromov. These spaces play an essential role in several areas
of mathematics [3] and find applications in other branches of science such as biology and com-
puter science [1, 25]. CAT(0) spaces are defined in terms of comparison with E2 , the Euclidean
plane, as follows: given (X, d) a geodesic metric space, a comparison triangle for a geodesic tri-
angle △(x1 , x2 , x3 ) in (X, d) is a triangle △( ¯x1 , ¯x2 , ¯x3 ) in E2 such that dE2 ( ¯xi , ¯xj ) = d(xi , xj ) for
i, j ∈ {1, 2, 3}. Such a comparison triangle always exists in E2 and is unique up to isometry. A
geodesic triangle △ in X is said to satisfy the CAT(0) inequality if, given ¯△ a comparison triangle
in E2 for △, for all x, y ∈ △
d(x, y) ≤ dE2 ( ¯x, ¯y ),
where ¯x, ¯y ∈ ¯△ are the comparison points of x and y , respectively. A geodesic space X is a CAT(0)
space if all its geodesic triangles satisfy the CAT(0) inequality.
The following four point condition was used by Berg and Nikolaev [2] to characterize CAT(0)
spaces.
Theorem 2.3. Let (X, d) be a geodesic space. X is a CAT(0) space if and only if for every
x, y , z , p ∈ X ,
d(x, z )2 + d(y , p)2 ≤ d(x, y)2 + d(y , z )2 + d(z , p)2 + d(p, x)2 .
In complete CAT(0) spaces, the metric pro jection onto closed and convex subsets behaves as in
Hilbert spaces in a certain sense.
Proposition 2.4 ([3], Proposition 2.4, p. 176). Let X be a complete CAT(0) space, x ∈ X and
C ⊂ X nonempty closed and convex. Then the fol lowing facts hold:
4
1. The metric projection PC (x) of x onto C is a singleton.
2. If y ∈ [x, PC (x)], then PC (x) = PC (y).
3. If x /∈ C and y ∈ C with y 6= PC (x) then ∠
PC (x) (x, y) ≥ π/2.
4. The mapping PC is a nonexpansive retraction from X onto C . Further, the mapping H :
X × [0, 1] → X associating to (x, t) the point at distance td(x, PC (x)) on the geodesic [x, PC (x)]
is a continuous homotopy from the identity map of X to PC .
For a thorough treatment of CAT(0) spaces and related topics the reader can check [3, 19].
In the next section we will also work with reflexive metric spaces which extend the notion of
reflexivity from Banach to metric spaces. A geodesic metric space X is said to be reflexive if for
every decreasing chain {Cα } ⊂ X with α ∈ I such that Cα is closed convex bounded and nonempty
for all α ∈ I we have that \α∈I
Cα 6= ∅. Notice that every complete uniformly convex metric space
with either a monotone or lower semicontinuous from the right modulus of uniform convexity is
reflexive (see [24, 12]). Also note that a reflexive and Busemann convex geodesic space is complete
(see [14, Lemma 4.1]). Moreover, in such a context the metric pro jection onto closed and convex
subsets is well-defined and singlevalued.
Next we give the definition of relatively nonexpansive mapping on the union of two sets.
Definition 2.5. Suppose A and B are two nonempty subsets of a metric space X . A mapping
T : A ∪ B → A ∪ B is relatively nonexpansive if d(T x, T y) ≤ d(x, y) for al l x ∈ A and y ∈ B .
We say that a relatively nonexpansive mapping T is cyclic if T (A) ⊆ B and T (B ) ⊆ A and
noncyclic if T (A) ⊆ A and T (B ) ⊆ B .
In [10], the notion of proximal normal structure was introduced as a counterpart of the well-
known concept of normal structure. This concept plays the same role for relatively nonexpansive
mappings as normal structure plays for nonexpansive mappings. We state below this notion in the
setting of geodesic metric spaces.
Definition 2.6. A convex pair (K1 , K2 ) in a geodesic space is said to have proximal normal struc-
ture if for any closed bounded convex proximal pair (H1 , H2 ) ⊆ (K1 , K2 ) for which dist(H1 , H2 ) =
dist(K1 , K2 ) and δ(H1 , H2 ) > dist(H1 , H2 ), there exists (x1 , x2 ) ∈ H1 × H2 such that
δ(x1 , H2 ) < δ(H1 , H2 )
and
δ(x2 , H1 ) < δ(H1 , H2 ).
As in the linear case, a pair (K, K ) has proximal normal structure if and only if K has normal
structure in the sense of Brodski and Milman [4].
3 Main results
Given a pair of sets (A, B ) in a metric space X , let A0 and B0 be the subsets defined as follows:
A0 = {x ∈ A : d(x, y ′ ) = dist(A, B ) for some y ′ ∈ B },
B0 = {y ∈ B : d(x′ , y) = dist(A, B ) for some x′ ∈ A}.
Proposition 3.1. Let X be a reflexive and Busemann convex metric space and let (A, B ) be a
nonempty closed convex pair of subsets in X . Suppose additional ly B is bounded. Then A0 and B0
are closed, convex, bounded and nonempty.
5
Proof. First we see that B0 is closed, convex, bounded and nonempty. Given any real number
ε > 0, consider the set
A′
ε = {x ∈ X : d(x, A) ≤ dist(A, B ) + ε}.
ε is nonempty and closed. Moreover, A′
It is easy to see that A′
ε is convex. Let x and y be two points
in A′
ε and m the midpoint between them. Since the space is Busemann convex, we have that
(cid:19)
d(m, A) = d(m, PA (m)) ≤ d(cid:18)m,
PA (x) + PA (y)
2
≤ max{d(x, A), d(y , A)} ≤ dist(A, B ) + ε.
Thus m ∈ A′
ε .
Now consider the set Aε = A′
ε ∩ B . It is immediate that Aε is closed, convex and bounded.
Moreover, by definition of A′
ε , Aε is also nonempty. Then, by means of the reflexivity of the space,
we conclude that ∩ε>0Aε 6= ∅. Since B0 = ∩ε>0Aε , we see that B0 is closed, convex, bounded and
nonempty. Notice that A0 is bounded and nonempty since B0 is so. The fact that A0 is also closed
and convex follows by a straightforward verification.
Remark 3.2. Notice that in the previous result we just need one of the sets A and B to be bounded,
no matter which of them.
Theorem 3.3. Let X be a reflexive and Busemann convex metric space and let (A, B ) be a
nonempty closed convex pair of subsets of X such that A is bounded. Let T : A ∪ B → A ∪ B
be a cyclic relatively nonexpansive mapping. Suppose (A, B ) has proximal normal structure. Then
there exists a pair (x, y) ∈ A × B such that d(x, T x) = d(y , T y) = dist(A, B ).
Proof. This result follows by applying similar patterns as in the proof of Theorem 2.1 in [10].
However, in this more general setting, several changes and new techniques must be considered to
get the result. From Proposition 3.1 we have that (A0 , B0 ) is closed, convex, bounded and nonempty.
Moreover, we may notice that this pair is also proximal and satisfies dist(A0 , B0 ) = dist(A, B ). It
is easy to see that T (A0 ) ⊆ B0 and T (B0 ) ⊆ A0 . Since (A, B ) has proximal normal structure, so
does (A0 , B0 ).
Consider the family Γ of sets F ⊆ A0 ∪ B0 such that F ∩ A0 and F ∩ B0 are closed, convex and
nonempty and satisfy T (F ∩A0 ) ⊆ F ∩B0 and T (F ∩B0 ) ⊆ F ∩A0 , dist(F ∩A0 , F ∩B0 ) = dist(A0 , B0 )
and the pair (F ∩ A0 , F ∩ B0 ) is proximal. Since A0 ∪ B0 ∈ Γ, we have that Γ 6= ∅.
Let {Fα }α∈I be a decreasing chain in Γ. We see that F0 = ∩α∈I Fα ∈ Γ. Since F0 ∩ A0 =
∩α∈I (Fα ∩ A0 ) and X is reflexive, we have that F0 ∩ A0 is closed, convex and nonempty. Similarly
F0 ∩ B0 is closed, convex and nonempty. It can be easily proved that T (F0 ∩ A0 ) ⊆ F0 ∩ B0 and
T (F0 ∩ B0 ) ⊆ F0 ∩ A0 . Thus, to see that F0 ∈ Γ, it remains to prove that the pair (F0 ∩ A0 , F0 ∩ B0 )
is proximal and dist(F0 ∩ A0 , F0 ∩ B0 ) = dist(A0 , B0 ). Let p ∈ F0 ∩ A0 . By definition, p ∈ Fα ∩ A0 for
every α ∈ I . Since (A0 , B0 ) is proximal and X is strictly convex, there exists a unique point q ∈ B0
such that d(p, q) = dist(A0 , B0 ). Moreover, by using the proximity of (Fα ∩ A0 , Fα ∩ B0 ), we have
that there exists a point qα ∈ Fα ∩ B0 such that d(p, qα ) = dist(Fα ∩ A0 , Fα ∩ B0 ) = dist(A0 , B0 )
for every α ∈ I . However, since qα ∈ B0 , we have that qα = q for every α ∈ I and therefore
q ∈ F0 ∩ B0 . Consequently, F0 ∈ Γ and applying Zorn’s Lemma we obtain a minimal element K in
Γ.
Let K1 = K ∩ A0 and K2 = K ∩ B0 . If δ(K1 , K2 ) = dist(K1 , K2 ), then d(p, q) = dist(K1 , K2 )
for every pair (p, q) ∈ K1 × K2 . In particular, d(x, T x) = dist(K1 , K2 ) = dist(A0 , B0 ) for every
x ∈ K and the result holds. Suppose now δ(K1 , K2 ) > dist(K1 , K2 ). Since (A0 , B0 ) has proximal
normal structure, there exists (y1 , y2 ) ∈ K1 × K2 and λ ∈ (0, 1) such that δ(y1 , K2 ) ≤ λδ(K1 , K2 )
6
d(z , y1 ) +
and δ(y2 , K1 ) ≤ λδ(K1 , K2 ). Since K ∈ Γ, (K1 , K2 ) is proximal. In fact, we may notice by the
strict convexity of X that for every point p ∈ K there exists only one point q ∈ K such that
2 ) = d(y ′
2 ) ∈ K1 × K2 such that d(y1 , y ′
1 , y ′
d(p, q) = dist(K1 , K2 ). Let (y ′
1 , y2 ) = dist(K1 , K2 ).
1 and y2 ,y ′
Denote by x1 and x2 the midpoints of y1 ,y ′
2 , respectively. By the Busemann convexity it
follows that d(x1 , x2 ) = dist(K1 , K2 ). Since the metric in X is convex, for every z ∈ K2 we have
1
1
1
d(z , y ′
1 ) ≤
d(z , x1 ) ≤
2
2
2
(1 + λ)
≤
2
Similarly, for every z ∈ K1 we have d(z , x2 ) ≤ (1+λ)
2
points (x1 , x2 ) ∈ K1 × K2 and α ∈ (0, 1) satisfying
δ(x1 , K2 ) ≤ αδ(K1 , K2 ) and δ(x2 , K1 ) ≤ αδ(K1 , K2 ).
Now consider the sets L1 ⊆ K1 and L2 ⊆ K2 defined as
L1 = {x ∈ K1 : δ(x, K2 ) ≤ αδ(K1 , K2 ) and for its proximal point y ∈ K2 ,
δ(y , K1 ) ≤ αδ(K1 , K2 )},
δ(K1 , K2 ). Thus, there exists a pair of proximal
(δ(y1 , K2 ) + δ(K1 , K2 ))
δ(K1 , K2 ).
L2 = {y ∈ K2 : δ(y , K1 ) ≤ αδ(K1 , K2 ) and for its proximal point x ∈ K1 ,
δ(x, K2 ) ≤ αδ(K1 , K2 )}.
Since x1 ∈ L1 and x2 ∈ L2 , Li 6= ∅ for i = 1, 2. Next we show that Li is closed and convex
for i = 1, 2. We just give the details for L1 since for L2 the proof follows similar patterns. Let
{vn } ⊆ L1 be a sequence that converges to a point v ∈ K1 . Since d(vn , z ) ≤ αδ(K1 , K2 ) for every
n ∈ N and z ∈ K2 , we get δ(v , K2 ) ≤ αδ(K1 , K2 ). The fact that vn ∈ L1 implies
δ(wn , K1 ) ≤ αδ(K1 , K2 ),
where wn ∈ K2 is the proximal point of vn ∈ K1 . Let w ∈ K2 such that d(v , w) = dist(K1 , K2 ).
By the Busemann convexity, we get
2 (cid:19) = dist(K1 , K2 ).
d(cid:18) vn + v
wn + w
2
Now, by Proposition 2.2, we have d(wn , w) = d(vn , v) for n ∈ N, from where wn → w. Taking limit
in (3.1), we may conclude δ(w, K1 ) ≤ αδ(K1 , K2 ). Consequently v ∈ L1 and then L1 is closed.
Let p1 , q1 ∈ L1 . Next we see that m1 = p1+q1
2 ∈ L1 . Let p2 , q2 ∈ K2 be the proximal points of p1
and q1 , respectively. Consider m2 = p2+q2
. Since (K1 , K2 ) is proximal and the space is Busemann
2
convex, d(m1 , m2 ) = dist(K1 , K2 ). Let z ∈ K2 . The convexity of the metric implies
1
1
d(q1 , z ) ≤ αδ(K1 , K2 ).
d(z , m1 ) ≤
2
2
Thus, δ(m1 , K2 ) ≤ αδ(K1 , K2 ). The fact that δ(m2 , K1 ) ≤ αδ(K1 , K2 ) follows similarly since
δ(p2 , K1 ) and δ(q2 , K1 ) are both ≤ αδ(K1 , K2 ). Then m1 ∈ L1 and so L1 is convex.
From d(x1 , x2 ) = dist(K1 , K2 ) we get dist(L1 , L2 ) = dist(A0 , B0 ). Moreover, by the definition of
the sets Li with i = 1, 2, it is immediate that (L1 , L2 ) is a proximal pair.
d(p1 , z ) +
(3.1)
,
7
In the sequel we see that T (L1 ) ⊆ L2 and T (L2 ) ⊆ L1 . Let x ∈ L1 and y ∈ L2 such that
d(x, y) = dist(L1 , L2 ). We prove that T x ∈ L2 . Let z ∈ K2 . Since d(T x, T z ) ≤ d(x, z ) ≤
δ(x, K2 ) ≤ αδ(K1 , K2 ), we get
T (K2 ) ⊆ B (T x, αδ(K1 , K2 )) ∩ K1 := K ′
1 .
1 is closed, convex and nonempty. Let K ′
Then K ′
2 ⊆ K2 be the set defined as
K ′
2 = {y ′ ∈ K2 : there exists x′ ∈ K ′
1 with d(x′ , y ′ ) = dist(K1 , K2 )}.
Similarly as we proved before that L1 is closed and convex, we get that K ′
2 is closed, convex and
nonempty.
2 ) ⊆ K ′
1 . The fact that T (K ′
2 ) ⊆ K ′
2 and T (K ′
1 ) ⊆ K ′
Now we see that T (K ′
1 is immediate.
1 , K ′
2 ). Then d(T p, T q) = dist(K ′
1 , K ′
2 such that d(p, q) = dist(K ′
1 and q ∈ K ′
Let p ∈ K ′
2 ). Since
1 ) ⊆ K ′
2 . Consequently, T (K ′
1 and therefore T p ∈ K ′
2 , we have T q ∈ K ′
q ∈ K ′
2 . Notice that,
1 , K ′
2 ) is also proximal and satisfies dist(K ′
1 , K ′
by definition, the pair (K ′
2 ) = dist(K1 , K2 ). Thus,
1 = K1 and K ′
2 ∈ Γ and by minimality of K it follows that K ′
1 ∪ K ′
K ′
2 = K2 . Consequently,
K1 ⊆ B (T x, αδ(K1 , K2 )) and therefore δ(T x, K1 ) ≤ αδ(K1 , K2 ). To conclude that T x ∈ L2
it remains to see that the proximal point z ∈ K1 of T x ∈ K2 satisfies δ(z , K2 ) ≤ αδ(K1 , K2 ).
Since T is relatively nonexpansive, we have that z = T y . Thus, we only need to show that
δ(T y , K2 ) ≤ αδ(K1 , K2 ). However, notice that this inequality holds if we repeat the previous
construction of K ′
1 and K ′
2 starting from the point y ∈ L2 and considering any point z ∈ K1 .
Thus, we have T x ∈ L2 and therefore T (L1 ) ⊆ L2 . In a similar way, we may see that T (L2 ) ⊆ L1 .
As a consequence, L1 ∪ L2 ∈ Γ. Since, δ(L1 , L2 ) ≤ αδ(K1 , K2 ), we get a contradiction with the
minimality of K .
Theorem 3.4. Let X be a reflexive and Busemann convex metric space and let (A, B ) be a
nonempty closed convex pair of subsets of X such that A is bounded. Let T : A ∪ B → A ∪ B
be a noncyclic relatively nonexpansive mapping. Suppose (A, B ) has proximal normal structure.
Then there exists a pair (x, y) ∈ A × B such that x = T x, y = T y and d(x, y) = dist(A, B ).
Proof. Proceeding as in the previous theorem, we may see that (A0 , B0 ) is proximal, closed, convex
and nonempty. Moreover, it is also immediate that dist(A0 , B0 ) = dist(A, B ), T (A0 ) ⊆ A0 and
T (B0 ) ⊆ B0 . Now let Γ be the collection of sets F ⊆ A0 ∪ B0 such that F ∩ A0 and F ∩ B0
are closed, convex and nonempty and satisfy T (F ∩ A0 ) ⊆ F ∩ A0 and T (F ∩ B0 ) ⊆ F ∩ B0 ,
dist(F ∩ A0 , F ∩ B0 ) = dist(A0 , B0 ) and the pair (F ∩ A0 , F ∩ B0 ) is proximal. Since A0 ∪ B0 ∈ Γ,
Γ 6= ∅.
Let {Fα }α∈I be a decreasing chain in Γ. Following similar patterns as in the previous proof, we
get that F0 = ∩α∈I Fα ∈ Γ. Then, applying Zorn’s Lemma, we find a minimal element K in Γ.
Let K1 = K ∩ A0 and K2 = K ∩ B0 . If δ(K1 , K2 ) = dist(K1 , K2 ), then d(p, q) = dist(K1 , K2 )
for every pair (p, q) ∈ K1 × K2 . Let (p, q) ∈ K1 × K2 . Since T is relatively nonexpansive,
d(T p, T q) = dist(K1 , K2 ) = dist(A0 , B0 ). Let m ∈ K2 be the midpoint between q and T q . Then,
since d(m, T p) = d(m, p) = dist(A0 , B0 ) and X is strictly convex, we get that T p = p and T q = q
with d(p, q) = dist(A0 , B0 ), so that the result holds. Suppose now δ(K1 , K2 ) > dist(K1 , K2 ).
Repeating the reasoning of the previous theorem, we may find a pair of proximal points (x1 , x2 ) ∈
K1 × K2 such that
δ(x1 , K2 ) ≤ αδ(K1 , K2 ) and δ(x2 , K1 ) ≤ δ(K1 , K2 ).
We consider now the sets L1 ⊆ K1 and L2 ⊆ K2 defined as
L1 = {x ∈ K1 : δ(x, K2 ) ≤ αδ(K1 , K2 ) and for its proximal point y ∈ K2 ,
δ(y , K1 ) ≤ αδ(K1 , K2 )},
8
L2 = {y ∈ K2 : δ(y , K1 ) ≤ αδ(K1 , K2 ) and for its proximal point x ∈ K1 ,
δ(x, K2 ) ≤ αδ(K1 , K2 )}.
Since the definition of these sets is as in Theorem 3.3, we have that (L1 , L2 ) is closed, convex,
nonempty, proximal and satisfies dist(L1 , L2 ) = dist(K1 , K2 ). To see that T (L1 ) ⊆ L1 and T (L2 ) ⊆
L2 we may follow a similar reasoning to the one considered in Theorem 3.3 where the cyclic inclusion
is proved. Although we omit some technical details, we include the proof for completeness. Let
x ∈ L1 and y ∈ L2 such that d(x, y) = dist(L1 , L2 ). We prove that T x ∈ L1 . Let z ∈ K2 . Since
d(T x, T z ) ≤ d(x, z ) ≤ δ(x, K2 ) ≤ αδ(K1 , K2 ), we get
T (K2 ) ⊆ B (T x, αδ(K1 , K2 )) ∩ K2 := K ′
2 .
2 is closed, convex and nonempty. Let K ′
Then K ′
1 ⊆ K1 be the set
K ′
1 = {x ∈ K1 : there exists y ∈ K ′
2 with d(x, y) = dist(K1 , K2 )}.
1 , K ′
2 ) is proximal and satisfies dist(K ′
1 , K ′
2 is closed, convex and nonempty. Moreover, (K ′
Then K ′
2 )
1 ∪ K ′
2 . Therefore, K ′
2 ) ⊆ K ′
1 and T (K ′
1 ) ⊆ K ′
= dist(K1 , K2 ) and T (K ′
2 ∈ Γ and by minimality
of K it follows that K2 ⊆ B (T x, αδ(K1 , K2 )) and therefore δ(T x, K2 ) ≤ αδ(K1 , K2 ). Proceeding
similarly, we may see that δ(T y , K1 ) ≤ αδ(K1 , K2 ). Since T y ∈ K2 is the proximal point of
T x ∈ K1 , we conclude that T x ∈ L1 and therefore T (L1 ) ⊆ L1 . Similarly, T (L2 ) ⊆ L2 . As
a consequence, L1 ∪ L2 ∈ Γ. Since, δ(L1 , L2 ) ≤ αδ(K1 , K2 ), we get a contradiction with the
minimality of K .
As a consequence of any of the two previous results, we get Kirk’s fixed point theorem in the
setting of reflexive and Busemann convex metric spaces when dist(A, B ) = 0. Notice that, in this
particular case, the fact that (A, B ) has proximal normal structure implies that A ∩ B has normal
structure in the sense of Brodski and Milman [4].
Proposition 3.5. Every closed convex pair in a uniformly convex metric space X has proximal
normal structure.
Proof. Let (H1 , H2 ) be a closed convex bounded proximal pair in X with δ(H1 , H2 ) > dist(H1 , H2 ).
Let x, y ∈ H1 such that d(x, y) > 0. Consider the points x′ , y ′ ∈ H2 such that d(x, x′ ) = d(y , y ′ ) =
2 ∈ H1 and m′ = x′+y ′
dist(H1 , H2 ). Let m = x+y
2 ∈ H2 . Since X is strictly convex, d(x′ , y ′ ) > 0.
Let ε = min{d(x, y), d(x′ , y ′ )} and z ∈ H2 . Denote by δX a modulus of uniform convexity of X .
Then,
δ(H1 , H2 ) (cid:19)(cid:19)δ(H1 , H2 ).
d(z , m) ≤ (cid:18)1 − δX (cid:18)δ(H1 , H2 ),
ε
Similarly, if we take z ∈ H1 , we get
δ(H1 , H2 ) (cid:19)(cid:19)δ(H1 , H2 ).
d(z , m′ ) ≤ (cid:18)1 − δX (cid:18)δ(H1 , H2 ),
ε
δ(H1 ,H2 ) (cid:19).
Thus, δ(m, H2 ) ≤ αδ(H1 , H2 ) and δ(m′ , H1 ) ≤ αδ(H1 , H2 ), for α = 1 − δX (cid:18)δ(H1 , H2 ),
ε
Corollary 3.6. Let (A, B ) be a closed convex pair in a complete Busemann convex metric space
X . Suppose that X is uniformly convex with a monotone or lower semicontinuous from the right
modulus of uniform convexity and B is bounded. Let T : A ∪ B → A ∪ B be a cyclic relatively
nonexpansive mapping. Then there exists a pair (x, y) ∈ A × B such that d(x, T x) = d(y , T y) =
dist(A, B ).
9
Corollary 3.7. Let (A, B ) be a closed convex pair in a complete Busemann convex metric space
X . Suppose that X is uniformly convex with a monotone or lower semicontinuous from the right
modulus of uniform convexity and B is bounded. Let T : A ∪ B → A ∪ B be a noncyclic relatively
nonexpansive mapping. Then there exists a pair (x, y) ∈ A × B such that x = T x, y = T y and
d(x, y) = dist(A, B ).
Proposition 3.8. Let (A, B ) be a closed convex pair in a complete Busemann convex metric space
X . Suppose that X is uniformly convex with δX being a monotone or lower semicontinuous from
the right modulus of uniform convexity and B is bounded. Let T : A ∪ B → A ∪ B be a noncyclic
relatively nonexpansive mapping. Let x0 ∈ A0 and define xn+1 = xn+T xn
for every n ≥ 1. Then
2
limn d(xn , T xn ) = 0. Moreover, if T (A) ⊆ C , where C is a compact set in X , then {xn } converges
to a fixed point of T .
Proof. By Corollary 3.7 we can find a point y ∈ B0 such that y = T y . Since the metric of the space
is convex, we get that {d(y , xn )} is nonincreasing and so convergent to some d ≥ 0. Suppose first
that d = 0. In this case, the result is immediate since {T xn} also converges to y . Now we consider
d > 0. Suppose that there exists a subsequence {xnk } of {xn } such that d(xnk , T xnk ) ≥ ε > 0 for
every k ≥ 0.
d+1 ) , 1(cid:27). Then, from the uniform
Suppose first δX is monotone. Let 0 < ρ < min (cid:26) dδX (d+1, ε
d+1 )
1−δX (d+1, ε
convexity of the space, there exists k0 ∈ N such that
d + 1 (cid:19)(cid:19)(d + ρ) for every k ≥ k0 .
d(y , xnk +1 ) ≤ (cid:18)1 − δX (cid:18)d + ρ,
ε
By the definition of ρ, we have d(y , xnk +1 ) < d for every k ≥ k0 , which is a contradiction.
In this case, let ε∗ > 0 such that
Suppose now δX is lower semicontinuous from the right.
ε
ε
ε
d+1 ). For such ε∗ > 0, consider µ(ε∗ ) > 0 such that δX (d,
d+1 ) + ε∗
0 < ε∗ < δX (d,
d+1 ) ≤ δX (r,
for every r ∈ (d, d + µ(ε∗ )). Let 0 < ρ < min (cid:26) dδX (d, ε
d+1 )+ε∗ , 1, µ(ε∗ )(cid:27). By using the uniform
d+1 )−dε∗
1−δX (d, ε
convexity as before, there exists k0 ∈ N such that
d(y , xnk +1 ) ≤ (cid:18)1 − δX (cid:18)d + ρ,
d + 1 (cid:19)(cid:19)(d + ρ) for every k ≥ k0 .
ε
Similarly we get a contradiction. The rest of the proof follows similar patterns to those given in
[10, Proposition 2.3].
Next we provide a bound for the existence of approximate fixed points for the mapping T .
Recall that having a metric space (X, d), a mapping T : X → X and {xn} ⊆ X , a mapping
Φ : (0, ∞) → N is called an approximate fixed point bound for {xn} (see also [22]) if
∀ε > 0, ∃n ≤ Φ(ε) such that d(xn , T xn ) ≤ ε.
We don’t include the proof of the result below since it can be obtained by following a similar
reasoning as in the main result of [23].
Proposition 3.9. Let (A, B ) be a closed convex pair in a complete Busemann convex metric space
X . Suppose that X is uniformly convex with δX being a monotone modulus of uniform convexity
and B is bounded. Let T : A ∪ B → A ∪ B be a noncyclic relatively nonexpansive mapping. Let
10
x0 ∈ A0 and b > 0 such that there exists y ∈ B0 with y = T y for which d(x0 , y) ≤ b. Define
xn+1 = xn+T xn
for every n ≥ 1. Then Φ : (0, ∞) → N,
2
#
Φ(ε) = "
2b
εδX (cid:0)b, ε
b (cid:1)
is an approximate fixed point bound for {xn }.
Proposition 3.10. Let (A, B ) be a compact convex pair in a geodesic space with convex metric.
Then (A, B ) has proximal normal structure.
Proof. Let (H1 , H2 ) ⊆ (A, B ) be a closed convex bounded and proximal pair in X with δ(H1 , H2 ) >
dist(H1 , H2 ). Suppose δ(x, H2 ) = δ(H1 , H2 ) for every x ∈ H1 . Let x0 ∈ H1 . Then there exists y0 ∈
H2 such that δ(x0 , H2 ) = δ(H1 , H2 ) = d(x0 , y0 ). Let x1 ∈ H1 such that d(x1 , y0 ) = dist(H1 , H2 ).
Then d(x1 , x0 ) ≥ d(x0 , y0 ) − d(x1 , y0 ) = δ(H1 , H2 ) − dist(H1 , H2 ). Let y1 ∈ H2 such that
, y1(cid:19) = δ(H1 , H2 ).
d(cid:18) x1 + x0
2
Since the metric is convex, the fact that
1
x1 + x0
) ≤
2
2
implies d(y1 , x0 ) = d(y1 , x1 ) = δ(H1 , H2 ). Let m0,1 = x1+x0
. Take x2 ∈ H1 such that d(y1 , x2 ) =
2
3 m0,1 ) = δ(H1 , H2 ). Let m1,2 = 1
3 x2 + 2
3 x2 + 2
dist(H1 , H2 ) and y2 ∈ H2 such that d(y2 , 1
3 m0,1 . By
using again the convexity of the metric, we obtain δ(H1 , H2 ) = d(y2 , x2 ) = d(y2 , m0,1 ) = d(y1 , m0,1 ).
Moreover, this last equality also implies
(d(y1 , x0 ) + d(y1 , x1 ))
δ(H1 , H2 ) = d(y1 ,
d(y2 , x0 ) = d(y2 , x1 ) = δ(H1 , H2 ).
Suppose we have {x1 , . . . , xn} in H1 , {m0,1 , m1,2 , . . . , mn−1,n} in H1 , where
i
1
mi−2,i−1
xi +
mi−1,i =
i + 1
i + 1
for every i ≥ 2, and {y1 , . . . , yn−1} in H2 such that
d(xi+1 , yi ) = dist(H1 , H2 ) for every i = 1 . . . n − 1,
d(yi , mi−1,i ) = δ(H1 , H2 ) for every i = 1 . . . n − 1
and
d(yi , xj ) = δ(H1 , H2 ) for every i = 0 . . . n − 1 and 0 ≤ j ≤ i.
Now consider the point yn ∈ H2 such that
d(yn , mn−1,n ) = δ(H1 , H2 ).
Take xn+1 ∈ H1 such that
d(xn+1 , yn ) = dist(H1 , H2 ).
By using again the convexity of the metric, we may see that
d(yn , xi ) = δ(H1 , H2 ) for every i = 0 . . . n.
As a consequence,
d(xn+1 , xi ) ≥ d(xi , yn ) − d(xn+1 , yn ) = δ(H1 , H2 ) − dist(H1 , H2 )
for every n ∈ N and for every i = 1 . . . n. Finally, by considering a convergent and, therefore,
Cauchy subsequence of {xn } we get a contradiction.
11
4 CAT(0) spaces
Proposition 4.1. Let (A, B ) be a closed convex pair in a CAT(0) space. Consider the mapping
P : A ∪ B → A ∪ B defined as
P (x) = (cid:26) PB (x) x ∈ A,
PA (x) x ∈ B .
Then P is a cyclic relatively nonexpansive mapping.
Proof. The fact that P is cyclic is immediate. Let x ∈ A and y ∈ B . For simplicity, denote
a = d(x, PA y), b = d(y , PB x), c = d(y , PA y), e = d(x, PB x), h = d(x, y) and r = d(PA y , PB x).
Next we prove that r ≤ h.
Let α = ∠PB x (x, y) and β = ∠PAy (x, y). By Proposition 2.4, cos α, cos β ≤ 0. By the Cosine
Law in CAT(0) spaces, we get
d(x, y)2 = h2 ≥ e2 + b2 and d(x, y)2 = h2 ≥ a2 + c2 .
If we apply Theorem 2.3 to the four points {x, PB x, y , PA y} and consider the two previous inequal-
ities, we obtain
h2 + r2 ≤ a2 + b2 + c2 + e2 ≤ 2h2 ,
and the result follows.
As a consequence of the previous result, we may reason as in [10] to conclude that in the setting
of CAT(0) spaces Theorem 3.4 is a consequence of Theorem 3.3.
Since every pair of closed and convex sets in a CAT(0) space satisfies property U C (see [13] for
more details on this property), we may assert that every noncyclic or cyclic relatively nonexpansive
mapping is also continuous if the pair (A, B ) is in addition proximal. Next we see that, as it
happens in Hilbert spaces [10, Proposition 3.2], a noncyclic relatively nonexpansive mapping is
even nonexpansive if the pair (A, B ) is proximal.
Proposition 4.2. Let (A, B ) be a closed convex bounded proximal pair of sets in a CAT(0) space.
Let T : A ∪ B → A ∪ B be a noncyclic relatively nonexpansive mapping. Then T is nonexpansive.
Proof. Let x1 , x2 ∈ A. Let d1 = d(x1 , PB x2 ), d2 = d(x2 , PB x1 ),
α = ∠PB T x1 (T x1 , T PB x2 ) and β = ∠PB T x2 (T x2 , T PB x1 ).
Note that d(x1 , x2 ) = d(PB x1 , PB x2 ), d(T x1 , T x2 ) = d(PB T x1 , PB T x2 ), d(x1 , PB x1 ) = d(x2 , PB x2 )
= dist(A, B ) and α, β ≥ π/2.
Let △1 = △(T PB x2 , T PB x1 , T x1 ) and △2 = △(T PB x1 , T PB x2 , T x2 ). Since d(T x1 , T PB x1 ) ≤
d(x1 , PB x1 ) it follows by Proposition 2.4, (1) that T PB x1 = PB T x1 . Similarly, T PB x2 = PB T x2 .
If we apply the Cosine Law in CAT(0) spaces to △1 and △2 , we obtain
d(T x1 , T x2 )2 + dist(A, B )2 = d(PB T x1 , PB T x2 )2 + dist(A, B )2
≤ d(T x1 , T PB x2 )2 ≤ d2
1
and
d(T x1 , T x2 )2 + dist(A, B )2 = d(PB T x1 , PB T x2 )2 + dist(A, B )2
≤ d(T x2 , T PB x1 )2 ≤ d2
2 .
12
Thus,
1 , d2
d(T x1 , T x2 )2 + dist(A, B )2 ≤ min{d2
2 }.
If we apply Theorem 2.3 to the four points {x1 , PB x1 , x2 , PB x2 }, we obtain
2 min{d2
1 , d2
2 } ≤ d2
1 + d2
2 ≤ 2(d(x1 , x2 )2 + dist(A, B )2 ).
By using (4.2) and (4.3), we get that T is nonexpansive on A. In a similar way, we get that T
is nonexpansive on B and then the result holds.
(4.2)
(4.3)
Next we see that a similar result also holds for cyclic relatively nonexpansive mappings.
Proposition 4.3. Let (A, B ) be a closed convex bounded proximal pair of sets in a CAT(0) space.
Let T : A ∪ B → A ∪ B be a cyclic relatively nonexpansive mapping. Then T is nonexpansive.
Proof. Let x1 , x2 ∈ A. Let d1 = d(x1 , PB x2 ), d2 = d(x2 , PB x1 )
α = ∠PAT x1 (T x1 , T PB x2 ) and β = ∠PAT x2 (T x2 , T PB x1 ).
Note that d(x1 , x2 ) = d(PB x1 , PB x2 ), d(T x1 , T x2 ) = d(PB T x1 , PB T x2 ), d(x1 , PB x1 ) = d(x2 , PB x2 )
= dist(A, B ) and α, β ≥ π/2.
Let △1 = △(T PB x2 , T PB x1 , T x1 ) and △2 = △(T PB x1 , T PB x2 , T x2 ). By Proposition 2.4, (1),
we get that T PB x1 = PAT x1 and T PB x2 = PAT x2 . By applying now the Cosine Law to △1 and
△2 and Theorem 2.3 to {x1 , PB x1 , x2 , PB x2 } as in the previous theorem we get the result.
5 Proximal normal structure: a sufficient but not necessary con-
dition
In 1979, Karlovitz [20] proved that the normal structure of the domain of a nonexpansive self-
mapping T is a sufficient but not necessary condition to guarantee existence of fixed points of such
a mapping in the context of reflexive Banach spaces. For this aim, a very specific family of reflexive
spaces which originated with R. C. James was considered. In the same setting we see now that
the proximal normal structure behaves similarly with respect to relatively nonexpansive mappings.
First we give an example of a closed convex bounded pair of sets in a reflexive Banach space that
does not have proximal normal structure.
Example 5.1. Let X denote the Banach space given by the set ℓ2 endowed with the norm
kxk = max{kxk∞ , kxk2 /√2}.
Let
and
A = B (θ , 1) ∩ {x = {xn}n≥1 ∈ X : x1 = 1, xi ≥ 0 for every i ≥ 1}
B = B (θ , 2) ∩ {x = {xn}n≥1 ∈ X : x1 = 2, xi ≥ 0 for every i ≥ 1},
where θ denotes the origin of the space ℓ2 . It is easy to see that (A, B ) is closed, convex, bounded
and proximal. From the definition we have kx − yk ≥ 1 for every x ∈ A and y ∈ B . Notice that
e1 ∈ A, 2e1 ∈ B and ke1 − 2e1 k = 1. Then dist(A, B ) = 1. We claim δ(A, B ) = 2. Let x = {xi} ∈ A
and y = {yi} ∈ B . Since x ∈ A, 0 ≤ xi ≤ 1 for every i ≥ 1. Equal ly, y ∈ B implies 0 ≤ yi ≤ 2.
i − 2(x1 y1 ) − 2 P∞
i + P∞
2 = P∞
i=1 x2
i=1 y2
Thus, kx − yk∞ ≤ 2. Note that kx − yk2
i=2 xiyi ≤ 6. Then
kx − yk2 /√2 ≤ √3 < 2 and therefore kx − yk ≤ 2. It is easy to see that {e1 + en }n≥2 ⊆ A and
13
{2e1 + 2en }n≥2 ⊆ B . Since d(e1 + en , 2e1 + 2em ) = 2 for n 6= m, we have δ(A, B ) = 2. On the
other hand, we have δ(x, B ) = 2 for every x = {xi }i≥1 ∈ A . This is a consequence of
d(x, 2e1 + 2en ) ≥ lim
n→∞ xn − 2 = 2.
lim
n→∞
Then (A, B ) does not have proximal normal structure.
Remark 5.2. Let A∗ = conv({e1 + en : n ≥ 2}) and B ∗ = conv({2e1 + 2en : n ≥ 2}). Note that
the pair (A∗ , B ∗ ) ⊆ (A, B ) does not have proximal structure either.
Next we see that proximal normal structure is a sufficient but not necessary condition to obtain
the existence of best proximity points in Theorem 2.1 in [10] and therefore also in Theorem 3.3.
Proposition 5.3. Let (A, B ) be the pair of sets considered in the previous example and suppose
T : A ∪ B → A ∪ B is a cyclic relatively nonexpansive mapping. Then there exists a pair (x, y) ∈
A × B such that d(x, T x) = d(y , T y) = dist(A, B ).
Proof. Consider the point 2e1 ∈ B . Notice that d(2e1 , A) = δ(2e1 , A) = 1. Thus,
1 ≤ d(T (2e1 ), T 2 (2e1 )) ≤ d(2e1 , T (2e1 )) = 1
and the result follows.
Remark 5.4. Notice that the previous result also holds whenever we have a cyclic relatively non-
expansive mapping defined on a pair (A, K ) ⊆ (A, B ) with 2e1 ∈ K .
Acknowledgements
Aurora Fern´andez-Le´on was partially supported by the Plan Andaluz de Investigaci´on de la Junta
de Andaluca FQM-127 and Grant P08-FQM-03543, and by MEC Grant MTM2009-10696-C02-01.
Part of this work was carried out while she was visiting the Babe¸s-Bolyai University in Cluj-Napoca.
She acknowledges the kind hospitality of the Department of Mathematics.
Adriana Nicolae was supported by a grant of the Romanian National Authority for Scientific
Research, CNCS-UEFISCDI, pro ject number PN-II-ID-PCE-2011-3-0383.
References
[1] I. Bartolini, P. Ciaccia, and M. Patella (2002). String matching with metric trees using an
approximate distance. In, String processing and information retrieval (A. H. F. Laender and
A. L. Oliveira, eds.). Lecture Notes in Computer Science vol. 2476, Springer, Berlin, pp. 271–
283.
[2] I. D. Berg and I. G. Nikolaev (2008). Quasilinearization and curvature of Alexandrov spaces.
Geom. Dedicata 133:195–218.
[3] M. R. Bridson and A. Haefliger (1999). Metric Spaces of Non-positive Curvature. Springer-
Verlag, Berlin Heidelberg.
[4] M. S. Brodski and D. P. Milman (1948). On the center of a convex set. Dokl. Akad. Nauk
SSSR (N.S.) 59:837–840 (in Russian).
[5] D. Burago, Y. Burago, and S. Ivanov (2001). A Course in Metric Geometry. Amer. Math.
Soc., Providence, RI.
14
[6] H. Busemann (1955). Geometry of Geodesics. Academic Press, New York.
[7] H. Busemann (1948). Spaces with non-positive curvature. Acta Math. 80:259–310.
[8] C. Di Bari, T. Suzuki, and C. Vetro (2008). Best proximity points for Meir-Keeler contractions.
Nonlinear Anal. 69:3790–3794.
[9] A. A. Eldred and P. Veeramani (2006). Existence and convergence of best proximity points.
J. Math. Anal. Appl. 323:1001–1006.
[10] A. A. Eldred, W. A. Kirk, and P. Veeramani (2005). Proximal normal structure and relatively
nonexpansive mappings. Studia Math. 171:283–293.
[11] R. Esp´ınola (2008). A new approach to relatively nonexpansive mappings. Proc. Amer. Math.
Soc. 136:1987–1995.
[12] R. Esp´ınola, A. Fern´andez-Le´on, and B. Pi¸atek (2010). Fixed points of single- and set-valued
mappings in uniformly convex metric spaces with no metric convexity. Fixed Point Theory
Appl. 2010: Article ID 169837, 16 pages.
[13] R. Esp´ınola and A. Fern´andez-Le´on (2011). On best proximity points in metric and Banach
spaces. Can. J. Math. 63:533–550.
[14] R. Esp´ınola and A. Nicolae (2012). Mutually nearest and farthest points of sets and the Drop
Theorem in geodesic spaces. Monatsh. Math. 165:173–197.
[15] R. Esp´ınola and B. Pi¸atek, Fixed point property and unbounded sets in CAT(0) spaces. J.
Math. Anal. Appl. doi:10.1016/j.jmaa.2013.06.038 (in press).
[16] K. Fan (1969). Extensions of two fixed point theorems of F.E. Browder. Math. Z. 112:234–240.
[17] A. Fern´andez-Le´on (2010). Existence and uniqueness of best proximity points in geodesic
metric spaces. Nonlinear Anal. 73:915–921.
[18] K. Goebel and S. Reich (1984). Uniform Convexity, Hyperbolic Geometry, and Nonexpansive
Mappings. Marcel Dekker, New York and Basel.
[19] M. Gromov (1984). Metric Structure for Riemannian and non-Riemannian Spaces, Birkhauser,
Boston.
[20] L. A. Karlovitz (1976). Existence of fixed points of nonexpansive mappings in a space without
normal structure. Pacific J. Math. 66:153–159.
[21] W. A. Kirk, P. S. Srinivasan, and P. Veeramani (2003). Fixed points for mappings satisfying
cyclical contractive conditions. Fixed Point Theory 4:79–89.
[22] U. Kohlenbach (2005). Some computational aspects of metric fixed point theory. Nonlinear
Anal. 61:823–837.
[23] L. Leu¸stean (2007). A quadratic rate of asymptotic regularity for CAT(0)-spaces. J. Math.
Anal. Appl. 325:386–399.
[24] L. Leu¸stean (2010). Nonexpansive iterations in uniformly convex W -hyperbolic spaces. Con-
temp. Math. 513:193–209.
15
[25] M. Owen and S. Provan (2011). A fast algorithm for computing geodesic distances in tree
space. IEEE/ACM Trans. Computational Biology and Bioinformatics 8:2–13.
[26] A. Papadopoulus (2005). Metric Spaces, Convexity and Nonpositive Curvature. European
Math. Soc., Zurich.
[27] B. Pi¸atek (2011). On cyclic Meir-Keeler contractions in metric spaces. Nonlinear Anal. 74:35–
40.
[28] V. Sankar Ra j and P. Veeramani (2009). Best proximity pair theorems for relatively nonex-
pansive mappings. Appl General Topol. 10:21–28.
[29] T. Suzuki, M. Kikkawa, and C. Vetro (2009). The existence of the best proximity points in
metric spaces with the property UC. Nonlinear Anal. 71:2918–2926.
16
|
1806.03596 | 1 | 1806 | 2018-06-10T07:21:31 | Weighted Riesz bases in g-fusion frames and their perturbation | [
"math.FA"
] | In this paper, we introduce orthonoramal and Riesz bases for g-fusion frames and will show that the weights have basic roles. Next, we prove an effective theorem between frames and g-fusion frames by using an operator. Finally, perturbations of g-fusion frames will be presented. | math.FA | math |
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND
THEIR PERTURBATION
GHOLAMREZA RAHIMLOU, VAHID SADRI, AND REZA AHMADI
Abstract. In this paper, we introduce orthonoramal and Riesz bases
for g-fusion frames and will show that the weights have basic roles.
Next, we prove an effective theorem between frames and g-fusion frames
by using an operator. Finally, perturbations of g-fusion frames will be
presented.
1. Introduction
Bases play a prominent role in discrete frames and their studying can
extract interesting properties from the frames. One of the most important
types of bases are orthonormal bases and also, as their special case, the Riesz
basis. The Riesz basis has been defined in [6] by the image of orthonormal
bases with a bounded bijective operator. Afterwards, this basis has been
equalized in a property with complete sequences and inequalities for the
synthesis operator. Sun in [17] could introduce a Riesz basis for g-frames by
using that property and we will continue his method in Section 3 for g-fusion
frames. In Section 4, we present a useful operator for characterizations of
these frames and finally in Section 5, a perturbation of these frames will be
studied.
2. Preliminaries
Throughout this paper, H and K are separable Hilbert spaces and B(H, K)
is the collection of all bounded linear operators of H into K. If K = H, then
B(H, H) will be denoted by B(H). Also, πV is the orthogonal projection
from H onto a closed subspace V ⊂ H and {Hj}j∈J is a sequence of Hilbert
spaces, where J is a subset of Z. It is easy to check that if u ∈ B(H) is an
invertible operator, then
πuV uπV = uπV .
We define the space H2 := (Pj∈J ⊕Hj)ℓ2 by
H2 = (cid:8){fj}j∈J : fj ∈ Hj, Xj∈J
(1)
kfjk2 < ∞(cid:9),
2010 Mathematics Subject Classification. Primary 42C15; Secondary 46C99, 41A58.
Key words and phrases. g-fusion frame, Dual g-fusion frame, gf-complete, gf-
orthonormal basis, gf-Riesz basis.
1
2
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
with the inner product defined by
h{fj},{gj}i = Xj∈J
hfj, gji.
It is clear that H2 is a Hilbert space with pointwise operations.
Definition 1. Let W = {Wj}j∈J be a collection of closed subspaces of H,
{vj}j∈J be a family of weights, i.e. vj > 0 and Λj ∈ B(H, Hj) for each j ∈ J.
We say Λ := (Wj, Λj , vj) is a generalized fusion frame (or g-fusion frame)
for H if there exists 0 < A ≤ B < ∞ such that for each f ∈ H
(2)
v2
jkΛj πWj fk2 ≤ Bkfk2.
Akfk2 ≤ Xj∈J
We call Λ a Parseval g-fusion frame if A = B = 1. When the right hand
side of (2) holds, Λ is called a g-fusion Bessel sequence for H with bound
B. Throughout this paper, Λ will be a triple (Wj , Λj , vj) with j ∈ J unless
otherwise noted.
The synthesis and analysis operators in the g-fusion frames are defined
by
TΛ : H2 −→ H
TΛ({fj}j∈J) = Xj∈J
vj πWj Λ∗j fj
,
,
T ∗Λ : H −→ H2
T ∗Λ(f ) = {vj ΛjπWj f}j∈J.
Thus, the g-fusion frame operator is given by
SΛf = TΛT ∗Λf = Xj∈J
hSΛf, fi = Xj∈J
j πWj Λ∗j Λj πWj f
v2
v2
jkΛj πWj fk2.
and
(3)
Therefore
AI ≤ SΛ ≤ BI.
This means that SΛ is a bounded, positive and invertible operator (with
adjoint inverse). So, we have the reconstruction formula for any f ∈ H:
(4)
j πWj Λ∗j ΛjπWj S−1
v2
j S−1
v2
Λ πWj Λ∗j Λj πWj f.
f = Xj∈J
Λ f = Xj∈J
For the proof of followings, see [16].
Theorem 2.1. Λ is a g-fusion Bessel sequence for H with bound B if and
only if the operator TΛ is well-defined and bounded operator with kTλk ≤ √B.
Theorem 2.2. Λ is a g-fusion frame for H if and only if
TΛ : H2 −→ H,
TΛ({fj}j∈J) = Xj∈J
vj πWj Λ∗j fj
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 3
is a well-defined, bounded and surjective.
Definition 2. A g-fusion frame Λ := (S−1
frame operator S Λ = T ΛT ∗Λ
Λ.
Λ , vj) with g-fusion
is called the (canonical) dual g-fusion frame of
Λ Wj , Λj πWj S−1
v2
j π Wj
Λj∗Λj πWj f,
Now, we can obtain
j πWj Λ∗j
v2
(5)
Λjπ Wj
f = Xj∈J
Λ Wj , Λj := Λj πWj S−1
Λ .
f = Xj∈J
where Wj := S−1
Definition 3. Let Λj ∈ B(H, Hj) for each j ∈ J. Λ = (Wj , Λj , vj) is called
gf-complete, if
It is easy to check that Λ is gf-complete if and only if
span{πWj Λ∗j Hj} = H.
{f : Λj πWj f = 0, j ∈ J} = {0}.
Proposition 1. If Λ = (Wj , Λj , vj) is a g-fusion frame for H, then Λ is a
complete.
Proof. See [16].
(cid:3)
3. gf-Riesz and Orthonormal Bases
Definition 4. Let W = {Wj}j∈J be a collection of closed subspaces of H
and j ∈ J. We say that (Wj , Λj) is a gf-orthonormal bases for H with respect
to {vj}j∈J, if
(6)
hviπWiΛ∗i gi, vj πWj Λ∗j gji = δi,jhgi, gji ,
i, j ∈ J ,
gi ∈ Hi ,
gj ∈ Hj
(7)
jkΛj πWj fk2 = kfk2 ,
v2
f ∈ H.
Xj∈J
Definition 5. Λ = (Wj , Λj , vj) is called a gf-Riesz basis for H if
(1) Λ is gf-complete,
(2) There exist 0 < A ≤ B < ∞ such that for each finite subset I ⊆ J
and gj ∈ Hj, j ∈ I,
(8)
AXj∈I
kgjk2 ≤ (cid:13)(cid:13)Xj∈I
vjπWj Λ∗j gj(cid:13)(cid:13)
2
≤ BXj∈I
kgjk2.
It is easy to check that if Λ is a gf-Riesz bases for H, then the operator
TΛ which is defined by
is injetive.
TΛ : H2 −→ H
TΛ({gj}j∈J) = Xj∈J
vj πWj Λ∗j gj ,
4
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
Proposition 2. Let Λ = (Wj, Λj , vj) be a g-fusion frame for H and suppose
that (6) holds. Then Λ is a gf-orthonormal basis for H.
Proof. Assume that SΛ is the g-fusion frame operator of Λ and
M := {f ∈ H : SΛf = f}.
It is clear that M is a non-empty closed subspace of H. Let f ∈ H and
k ∈ J. Since
vjvkπWj Λ∗j ΛjπWj πWk Λ∗kΛkπWkf = δj,kπWk Λ∗kΛkπWk f,
then πWk Λ∗kΛkπWkf ∈ M . So, for any h ∈ M⊥ and g ∈ M we have, for all
j ∈ J
hπWj Λ∗j Λj πWj h, gi = hh, πWj Λ∗j Λj πWj gi = 0.
Thus πWj Λ∗j Λj πWj h = 0 for each j ∈ J, and so kΛj πWj hk = 0. By definition
of g-fusion frame, we obtain h = 0. Therefore, M⊥ = {0} and we conclude
H = M . So, SΛ = idH and the proof is completed.
(cid:3)
Theorem 3.1. Λ is a gf-orthonormal bases for H if and only if
(I) vj πWj Λ∗j is isometric for any j ∈ J;
(II) Lj∈J vj πWj Λ∗j (Hj) = H.
Proof. Suppose that Λ is a gf-orthonormal basis for H. If j ∈ J and g ∈ Hj,
we have
hvj πWj Λ∗j g, vj πWj Λ∗j gi = δj,jhg, gi = hg, gi.
Then, vjπWj Λ∗j is isometric for any j ∈ J. So, vjπWj Λ∗j (Hj) is a closed
subspace of H. Therefore, for each j ∈ J and f ∈ H, we have from (7)
hf, fi = Xj∈J
= hXj∈J
v2
jhΛj πWj f, Λj πWj fi
j πWj Λ∗j Λj πWj f, fi.
v2
Thus, for each f ∈ H,
j πWj Λ∗j ΛjπWj f.
v2
f = Xj∈J
By letting gj := vjΛj πWj f , we obtain f = Pj∈J vjπWj Λ∗j gj and
Xj∈J
v2
jkπWj Λ∗j gjk2 = Xj∈J
kgjk2 = Xj∈J
v2
jkΛj πWj fk2 = kfk2.
So, Lj∈J vj πWj Λ∗j (Hj) = H. Conversely, if (I) , (II) are satisfied, then (6) is
clear. Indeed, for each i 6= j, viπWiΛ∗i ⊥ vj πWj Λj and vjπWj Λ∗j is isometric.
Let f ∈ H, we get from (II) for any j ∈ J and some gj ∈ Hj
vj πWj Λ∗j gj
f = Xj∈J
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 5
and
v2
kgjk2.
jkπWj Λ∗j gjk2 = Xj∈J
kfk2 = Xj∈J
Now, let i ∈ J, then for each f, h ∈ Hj,
hvjΛiπWif, hi = hXj∈J
j ΛiπWiπWj Λ∗j gj , hi
v2
= Xj∈J
= hviπWiΛ∗i gi, viπWiΛ∗i hi
= hgi, hi.
jhπWj Λ∗j gj , πWiΛ∗i hi
v2
Hence gi = viΛiπWif for each i ∈ J. So, f = Pj∈J v2
f ∈ H and (7) is proved.
Corollary 1. Every gf-orthonormal basss for H is a gf-Riesz bases for H
with bounds A = B = 1.
j πWj Λ∗j Λj πWj f for all
(cid:3)
Theorem 3.2. Let Θ = (Wj , Θj) be a gf-orthonormal bases with respect to
{vj}j∈J and Λ = (Wj , Λj , vj) be a g-fusion frame for H with same weights.
Then, there exists a operator V ∈ B(H) and surjective such that ΛjπWj =
ΘjπWj V ∗ for all j ∈ J.
Proof. Let
Then, V is well-defined and bounded. Indeed, for each finite subset I ⊆ J
and f ∈ H,
V : H −→ H,
V f = Xj∈J
j πWj Λ∗j Θj πWj f.
v2
kV fk = sup
khk=1(cid:12)(cid:12)(cid:10)Xj∈J
√B(cid:0)Xj∈J
≤
= √Bkfk,
v2
v2
j πWj Λ∗j ΘjπWj f, h(cid:11)(cid:12)(cid:12)
jkΘj πWj fk2(cid:1)
1
2
where B is an upper g-fusion frame bound for Λ. Therefore, the series is
weakly unconditionally Cauchy and so unconditionally convergent in H (see
[7] page 58) and also kV k ≤ √B. Since Θ is a gf-orthonormal bases, then
ΘjπWj πWiΘ∗i g = (vivj)−1δi,jg
and
j πWj Λ∗j ΘjπWj πWiΘ∗i g
v2
V πWiΘ∗i g = Xj∈J
= πWiΛ∗i g,
6
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
for all g ∈ Hj and i ∈ J. Thus ΛjπWj = ΘjπWj V ∗. Now, we show that V
is surjective. Assume that f ∈ H. By Theorem 2.2, there is {gj}j∈J ∈ H2
such that Pj∈J vjπWj Λ∗j gj = f . Let g := TΘ({gj}j∈J), thus
vj πWj Λ∗j gj = f
V g = Xj∈J
V vjπWj Θ∗j gj = Xj∈J
and V is surjective.
(cid:3)
Corollary 2. If Λ is a Parseval g-fusion frame for H, then V ∗ is isometric.
Corollary 3. If Λ is a gf-Riesz bases for H, then V is invertible.
Proof. Let V f = 0 and f ∈ H. Since TΛ is injective and
j πWj Λ∗j Θj πWj f = TΛT ∗Θf,
v2
V f = Xj∈J
therefore, T ∗Θf = 0. So, kfk2 = kTΘfk2 = 0, hence, f = 0.
Corollary 4. If (Wj, Λj) is a gf-orthonormal bases for H with respect to
{vj}j∈J, then V is unitary.
Proof. By Corollaries 1 and 3, the operator V is invertible. Let f ∈ H, we
obtain
(cid:3)
kfk2 = Xj∈J
v2
jkΛj πWj fk2 = Xj∈J
kΘj πWj V ∗fk2 = kV ∗fk2.
Thus, V V ∗ = idH and this means that V is unitary.
(cid:3)
4. Characterizations of g-fusion frames, gf-Riesz and
gf-orthonormal bases
Sun in [17] showed that each g-frame for H induces a sequence in H
dependent on the g-frame and he proved a useful theorem about them. In
this section, we are going to express Sun's method for g-fusion frames.
Let W = {Wj}j∈J be a family of closed subspaces of H, {vj}j∈J be a family
of weights, Λj ∈ B(H, Hj) for each j ∈ J and {ej,k}k∈Kj be an orthonormal
basis for Hj, where Kj ⊆ Z and j ∈ J. Suppose that
ϕ :h −→ C,
ϕ(f ) = hvjΛjπWj f, ej,ki.
We have
therefore, ϕ is a bounded linear functional on H. Now, we can write
kϕfk ≤ vjkΛjkkfk,
So, if
hvjΛjπWj f, ej,ki = hf, vjπWj Λ∗j ej,ki.
(9)
then hf, uj,ki = hvjΛj πWj f, ej,ki for all f ∈ H.
uj,k := vj πWj Λ∗j ej,k,
j ∈ J, k ∈ Kj
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 7
Remark 1. Using (9), we get for each f ∈ H
(10)
vjΛjπWj f = Xk∈Kj
Xk∈Kj
hf, uj,ki2 ≤ v2
hf, uj,kiej,k.
jkΛjk2kfk2.
But
Thus, {uj,k}k∈Kj is a Bessel sequence for H. It follows that for each f ∈ H
and g ∈ Hj
hf, vj πWj Λ∗j gi = hvjΛjπWj f, gi
hvj πWj Λj f, ej,kihej,k, gi
= Xk∈Kj
= Xk∈Kj
= Df, Xk∈Kj
hf, uj,kihej,k, gi
hg, ej,kiuj,kE.
Therefore,
(11)
for all g ∈ Hj.
vj πWj Λ∗j g = Xk∈Kj
hg, ej,kiuj,k
We say {uj,k : j ∈ J, k ∈ Kj} the sequence induced by Λ.
Theorem 4.1. Let Λ = (Wj , Λj , vj) and uj,k be defined by (9). Then we
get the followings.
(I) Λ is a g-fusion frame (resp. g-fusion Bessel sequence, Parseval g-
fusion frame, gf-Riesz basis, gf-orthonormal basis) for H if and only
if {uj,k : j ∈ J, k ∈ Kj} is a frame (resp. Bessel sequence, Parseval
frame, Riesz basis, orthonormal basis) for H.
(II) The g-fusion operator for Λ coincides with the frame operator for
{uj,k : j ∈ J, k ∈ Kj}.
Proof. (I). By (10), Λ is a g-fusion frame (resp. g-fusion Bessel sequence,
: j ∈ J, k ∈ Kj} is a
Parseval g-fusion frame) for H if and only if {uj,k
frame (resp. Bessel sequence, Parseval frame) for H.
Assume that Λ is a gf-Riesz basis for H and gj ∈ Hj. Thus,
gj = Xk∈Kj
where cj,k ∈ ℓ(Kj). Note that, by (10),
cj,kej,k,
{f : ΛjπWj f = 0,
j ∈ J} = {f : hf, uj,ki = 0,
j ∈ J, k ∈ Kji}
8
and
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
AXj∈I
kgjk2 ≤ (cid:13)(cid:13)Xj∈I
cj,k2 ≤ (cid:13)(cid:13)Xj∈I Xk∈Kj
is equivalent to
AXj∈I Xk∈Kj
vj πWj Λ∗j gj(cid:13)(cid:13)
2
≤ BXj∈I
kgjk2
2
cj,kuj,k(cid:13)(cid:13)
≤ BXj∈I Xk∈Kj
cj,k2
for any finite I ⊆ J. Thus, Λ is a gf-Riesz basis if and only if {uj,k : j ∈
J, k ∈ Kj} is a Riesz basis.
Now, let (Wj, Λj) be a gf-orthonormal basis for H with respect to {vj}j∈J.
We get for any j1, j2 ∈ J, k1 ∈ Kj1 and k2 ∈ Kj2
huj1,k1 , uj2,k2i = hvj1πWj1 Λ∗j1 ej1,k1, vj2 πWj2 Λ∗j2ej2,k2i
= δj1,k1δj2,k2.
So, {uj,k : j ∈ J, k ∈ Kj} is an orthonormal sequence. Moreovere
kfk2 = Xj∈J
v2
jkΛj πWj fk2 = Xj∈J Xk∈Kj
hf, uj,ki2
for any f ∈ H. Hence {uj,k : j ∈ J, k ∈ Kj} is an orthonormal basis. For
the opposite implication, we need only to prove that (6) holds. We have by
(11) for each j1 6= j2 ∈ J, gj1 ∈ Hj1 and gj2 ∈ Hj2,
hvj1 πWj1 Λ∗
j1 gj1 , vj2 πWj2 Λ∗
hgj1 , ej1,k1iuj1,k1 , Xk2 ∈Kj2
j2 gj2i = (cid:10) Xk1∈Kj1
hgj2 , ej2,k2iuj2,k2(cid:11) = 0
and for all g1, g2 ∈ Hj
hvjπWj Λ∗
j g1, vjπWj Λ∗
j g2i = (cid:10) Xk1 ∈Kj
hg1, ej,k1iuj,k1 , Xk2 ∈Kj
hg2, ej,k2iuj,k2(cid:11) = hg1, g2i.
(II). By (10) and (11) we have for any f ∈ H
Xj∈J
v2
j πWj Λ∗j Λj πWj f = Xj∈J Xk∈Kj
hvjΛj πWj f, ej,kiuj,k
hf, uj,k′ej,k′i, ej,kEuj,k
= Xj∈J Xk∈KjD Xk′∈Kj
= Xj∈J Xk∈Kj
hf, uj,kiuj,k.
5. Perturbation of g-Fusion Frames
In this section, we present some perturbation of g-fusion frames and re-
view some results about them. First, we need the following which is proved
in [3].
(cid:3)
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 9
Lemma 1. Let U be a Linear operator on a Banach space X and assume
that there exist λ1, λ2 ∈ [0, 1) such that
kx − U xk ≤ λ1kxk + λ2kU xk
for all x ∈ X. Then U is bounded and invertible. Moreovere,
1 − λ1
1 + λ2kxk ≤ kU xk ≤
1 + λ1
1 − λ2kxk
and
1 − λ2
1 + λ1kxk ≤ kU−1xk ≤
1 + λ2
1 − λ1kxk
for all x ∈ X.
Theorem 5.1. Let Λ := (Wj , Λj , vj) be a g-fusion frame for H with bounds
A, B and {Θj ∈ B(H, Hj)}j∈J be a sequence of operators such that for any
finite subset I ⊆ J and for each f ∈ H
kXj∈I
j ΛjπWj f−πWj Θ∗
+ µkXj∈I
, µ} < 1, then Θ := (Wj , Θj, vj) is a g-fusion frame
j ΘjπWj f(cid:1)k ≤ λkXj∈I
j ΘjπWj fk + γ(Xj∈I
jkΛjπWj fk2)
v2
j(cid:0)πWj Λ∗
j ΛjπWj fk
where 0 ≤ max{λ + γ√A
for H with bounds
v2
j πWj Θ∗
v2
j πWj Λ∗
v2
1
2 ,
1 −(cid:0)λ + γ√A(cid:1)
1 + µ
A
and B
1 + λ + γ√B
1 − µ
Proof. Assume that I ⊆ J is a finite subset and f ∈ H. We have
kXj∈I
j ΘjπWj f(cid:1)k
j ΘjπWj fk ≤ kXj∈I
j(cid:0)πWj Λ∗
v2
j πWj Θ∗
v2
j ΛjπWj f − πWj Θ∗
+ kXj∈I
v2
j πWj Λ∗
v2
j πWj Λ∗
j ΛjπWj fk
j ΛjπWj fk + µkXj∈I
jkΛjπWj fk2)
v2
2 .
1
+ γ(Xj∈I
≤ (1 + λ)kXj∈I
v2
j πWj Θ∗
j ΘjπWj fk
Then
kXj∈I
v2
j πWj Θ∗
j ΘjπWj fk ≤
1 + λ
1 − µkXj∈I
v2
j πWj Λ∗
j ΛjπWj fk +
γ
1 − µ
(Xj∈I
jkΛjπwj fk2)
v2
1
2 .
Let SΛ be the g-fusion frame operator of Λ, then
j πWj Λ∗j Λj πWj fk ≤ kSΛfk ≤ Bkfk.
v2
kXj∈I
10
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
j πWj Θ∗j Θj πWj fk ≤ (
v2
Therefore, for all f ∈ H
kXj∈I
So, Pj∈J v2
√B +
1 + λ
1 − µ
γ
1 − µ
)(Xj∈I
v2
jkΛj πWj fk2)
1
2 < ∞.
j πWj Θ∗j ΘjπWj f is unconditionally convergent. Let
SΘ : H −→ H
SΘ(f ) = Xj∈J
j πWj Θ∗j Θj πWj f.
v2
SΘ is well defined and bounded operator with
kSΘk ≤
and for each f ∈ H, we have
1 + λ
1 − µ
B +
γ√B
1 − µ
jkΘj πWj fk2 = hSΘf, fi ≤ kSΘkkfk2.
v2
Xj∈J
It follows that Θ := (Wj , Θj , vj) is a g-fusion Bessel sequence for H. Thus,
we obtain by the hypothesis
kSΛf − SΘfk ≤ λkSΛfk + µkSΘfk + γ(Xj∈I
jkΛj πwj fk)2)
v2
1
2 .
Therefore, by (3)
kf − SΘS−1
Λ fk ≤ λkfk + µkSΘS−1
Λ fk + γ(cid:0)Xj∈J
√A(cid:17)kfk + µkSΘS−1
Λ fk.
γ
≤ (cid:16)λ +
JkΛj πWj S−1
v2
Λ fk2(cid:1)
1
2
Since 0 ≤ max{λ + γ√A
SΘ is invertible and we get
, µ} < 1, then by Lemma 1, SΘS−1
Λ and consequently
kS−1
Θ k ≤ kS−1
Λ kkSΛS−1
Θ k ≤
So, the proof is completed.
1 + µ
A(cid:0)1 − (λ + γ√A
)(cid:1)
.
(cid:3)
Corollary 5. The optimal lower and upper bounds of Θ which defined in
Theorem 5.1 are kS−1
Corollary 6. Let Λ be a g-fusion frame for H with bounds A, B and {Θj ∈
B(H, Hj)}j∈J be a sequence of operators. If there exists a constant 0 < R <
A such that
Θ k−1 and kSΘk, respectively.
jkπWj Λ∗j Λj πWj f − πWj Θ∗j ΘjπWj fk ≤ Rkfk
v2
Xj∈J
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 11
for all f ∈ H, then Θ := (Wj, Θj , vj) is a g-fusion frame for H with bounds
A − R and min{B + Rr B
A
, R + √B}.
Proof. It is easy to check that Pj∈J v2
f ∈ H. Thus, we obtain for each f ∈ H
j πWj Θ∗j ΘjπWj f is converges for any
Xj∈J
jkπWj Λ∗j Λj πWj f − πWj Θ∗j Θj πWj fk ≤ Rkfk
v2
≤
R
√A(cid:0)Xj∈J
v2
jkΛj πWj fk2(cid:1)
1
2
and also
v2
Xj∈J
jkπWj Θ∗j ΘjπWj fk ≤ Rkfk +Xj∈J
≤ (R + √B)kfk.
By using Theorem 5.1 with λ = µ = 0 and γ = R√A
jkπWj Λ∗j ΛjπWj fk
v2
, the proof is completed.
(cid:3)
The following is another version of perturbation of g-fusion frames.
Theorem 5.2. Let Λ be a g-fusion frame for H with bounds A, B and
{Θj ∈ B(H, Hj)}j∈J be a sequence of operators such that for any finite subset
I ⊆ J and for each {fj}j∈J ∈ H2,
kXj∈I
vj(cid:0)πWj Λ∗j fj − πWj Θ∗j fj(cid:1)k ≤ λkXj∈I
vj(πWj Λ∗j fj)k+
+ µkXj∈I
vj(πWj Θ∗j fj)k + γ(Xj∈I
kfjk2)
1
2 ,
where 0 ≤ max{λ + γ√A
for H with bounds
, µ} < 1, then Θ := (Wj, Θj , vj) is a g-fusion frame
A(cid:16)
1 −(cid:0)λ + γ√A(cid:1)2
1 + µ
Proof. Let {fj}j∈J ∈ H2, then
j fjk ≤ kXj∈I
kXj∈I
vjπWj Θ∗
(cid:17) and B(cid:16)
1 + λ + γ√B
1 − µ
.
(cid:17)2
vj(cid:0)πWj Λ∗
j fj − πWj Θ∗
j fj(cid:1)k + kXj∈I
vjπWj Λ∗
j fjk
≤ (1 + λ)kXj∈I
vjπWj Λ∗
j fjk+
+ µkXj∈I
vjπWj Θ∗
j fjk + γ(cid:0)Xj∈I
1
2 .
kfjk2(cid:1)
12
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
Hence,
kXj∈I
Let
vjπWj Θ∗j fjk ≤
1 + λ
1 − µkXj∈I
√B +
≤ (cid:16) 1 + λ
1 − µ
vj πWj Λ∗j fjk +
1 − µ(cid:17)(cid:16)Xj∈I
γ
γ
1 − µ(cid:0)Xj∈I
kfjk2(cid:17)
.
1
2
1
2
kfjk2(cid:1)
TΘ : H −→ H2
TΘ{fj}j∈J = Xj∈I
vj πWj Θ∗j fj.
Therefore, TΘ is a well-defined and bounded. Then, by Theorem 2.1, Θ is
a g-fusion Bessel sequence. Suppose that G := TΘT ∗ΛS−1
Λ . Then, we get by
the hypotesis and (9) for fj := ΛjπWj S−1
Λ f ,
kf − Gfk ≤ λkfk + µkGfk + γ(cid:0)Xj∈J
√A(cid:1)kfk + µkGfk.
≤ (cid:0)λ +
γ
jkΛj πWj S−1
v2
Λ fk2(cid:1)
1
2
Since 0 ≤ max{λ + γ√A
invertible and
, µ} < 1, by Lemma 1, G and consequently TΘT ∗Λ is
kG−1k ≤
1 + µ
1 − (λ + γ√A
)
.
Now, let f ∈ H and we have
kfk4 = hGG−1f, fi2
2
Λ (G−1f ), f(cid:11)(cid:12)(cid:12)
jkΘj πWj fk2
v2
j πWj Θ∗j ΛjπWj S−1
v2
= (cid:12)(cid:12)(cid:10)Xj∈J
≤ Xj∈J
kΛj πWj S−1
≤ A−1kG−1fk2Xj∈J
Λ (G−1f )k2.Xj∈J
jkΘj πWj fk2
v2
and this copmletes the proof.
(cid:3)
Theorem 5.3. Let Λ be a g-fusion frame for H with bounds A, B and
{Θj ∈ B(H, Hj)}j∈J be a sequence of operators. If there exists a constant
0 < R < A such that
for all f ∈ H, then Θ := (Wj, Θj , vj) is a g-fusion frame for H with bounds
√R)2
and
(√R + √B)2.
jkΛj πWj f − ΘjπWj f(cid:1)k2 ≤ Rkfk2
v2
Xj∈J
(√A −
WEIGHTED RIESZ BASES IN G-FUSION FRAMES AND THEIR PERTURBATION 13
Proof. Let f ∈ H. We can write
k{vj ΘjπWj f}j∈Jk2 ≤ k{vjΘj πWj f − vjΛjπWj f}j∈Jk2 + k{vjΛj πWj f}j∈Jk2
≤ (√R + √B)2kfk2
and also
k{vj ΘjπWj f}j∈Jk2 ≥ k{vjΛj πWj f}j∈Jk2 − k{vjΘjπWj f − vjΛj πWj f}j∈Jk2
≥ (√A −
√R)2kfk2.
Thus, these complete the proof.
(cid:3)
6. Conclusions
In this paper, we could transfer some common properties of g-frames
to g-fusion frames. First, we reviewed gf-Riesz and orthonormal bases
and showed that the weights have basic roles in their definitions. After-
wards, characterizations of these frames were presented in Section 4 by Sun's
method by using the operator (9). Finally, in Section 5, we introduced the
perturbation of g-fusion frames and reviewed some useful statements.
References
[1] H. Blocsli, H. F. Hlawatsch, and H. G. Fichtinger, Frame-Theoretic analysis of over-
sampled filter bank, IEEE Trans. Signal Processing 46 (12), 3256- 3268, 1998.
[2] E. J. Candes, and D. L. Donoho, New tight frames of curvelets and optimal represen-
tation of objects with piecwise C 2 singularities. Comm. Pure and App. Math. 57(2),
219-266, 2004.
[3] P. G. Casazza, and O. Christensen, Perturbation of Operators and Application to
Frame Theory. J. Fourier Anal. Appl. 3, 543-557, 1997.
[4] P. G. Casazza, and G. Kutyniok, Frames of Subspaces. Contemp. Math. 345, 87-114,
2004.
[5] P. G. Casazza, G. Kutyniok, and S. Li, Fusion Frames and distributed processing.
Appl. comput. Harmon. Anal. 25(1), 114-132, 2008.
[6] O. Christensen, Frames and Bases: An Introductory Course (Applied and Numerical
Harmonic Analysis). Birkhauser, 2008.
[7] J. Diestel, Sequences and series in Banach spaces. Springer-Verlag, New York, 1984.
[8] R.J. Duffin, and A. C. Schaeffer, A class of nonharmonik Fourier series. Trans. Amer.
Math. Soc. 72(1), 341-366, 1952.
[9] M. H. Faroughi, and R. Ahmadi, Some Properties of C-Frames of Subspaces. J. Non-
linear Sci. Appl. 1(3), 155-168, 2008.
[10] H. G. Feichtinger, and T. Werther, Atomic Systems for Subspaces. Proceedings
SampTA. Orlando, FL, 163-165, 2001.
[11] L. Gavruta, Frames for Operators. Appl. Comp. Harm. Annal. 32, 139-144, 2012.
[12] S. B. Heineken, P. M. Morillas, A. M. Benavente, and M. I. Zakowich, Dual Fusion
Frames. arXiv: 1308.4595v2 [math.CA] 29 Apr 2014.
[13] H. Heuser, Functional Analysis. John Wiley, New York, 1991.
[14] M. Khayyami, and A. Nazari, Construction of Continuous g-Frames and Continuous
Fusion Frames. SCMA 4(1), 43-55, 2016.
[15] A. Najati, M. H. Faroughi, and A. Rahimi, g-frames and stability of g-frames in
Hilbert spaces. Methods of Functional Analysis and Topology. 14(3), 305–324, 2008.
14
GH. RAHIMLOU, V. SADRI, AND R. AHMADI
[16] V. Sadri, Gh. Rahimlou, R. Ahmadi and R. Zarghami Farfar, Generalized Fusion
Frames in Hilbert Spaces. Submited 2018.
[17] W. Sun, G-Frames and G-Riesz bases, J. Math. Anal. Appl.326 , 437-452, 2006.
Department of Mathematics, Faculty of Tabriz Branch, Technical and Vo-
cational University (TUV), East Azarbaijan , Iran,
E-mail address: [email protected]
Department of Mathematics, Faculty of Tabriz Branch, Technical and Vo-
cational University (TUV), East Azarbaijan , Iran
E-mail address: [email protected]
Institute of Fundamental Sciences, University of Tabriz, , Iran,
E-mail address: [email protected]
|
1309.7469 | 1 | 1309 | 2013-09-28T15:55:19 | Operators on Banach Spaces of Bourgain-Delbaen Type | [
"math.FA"
] | We begin by giving a detailed exposition of the original Bourgain-Delbaen construction and the generalised construction due to Argyros and Haydon. We show how these two constructions are related, and as a corollary, are able to prove that there exists some $\delta > 0$ and an uncountable set of isometries on the original Bourgain-Delbaen spaces which are pairwise distance $\delta$ apart.
We subsequently extend these ideas to obtain our main results. We construct new Banach spaces of Bourgain-Delbaen type, all of which have $\ell_1$ dual. The first class of spaces are HI and possess few, but not very few operators. We thus have a negative solution to the Argyros-Haydon question. We remark that all these spaces have finite dimensional Calkin algebra, and we investigate the corollaries of this result. We also construct a space with $\ell_1$ Calkin algebra and show that whilst this space is still of Bourgain-Delbaen type with $\ell_1$ dual, it behaves somewhat differently to the first class of spaces.
Finally, we briefly consider shift-invariant $\ell_1$ preduals, and hint at how one might use the Bourgain-Delbaen construction to produce new, exotic examples. | math.FA | math | Operators on Banach Spaces of
Bourgain -- Delbaen Type
3
1
0
2
p
e
S
8
2
]
.
A
F
h
t
a
m
[
1
v
9
6
4
7
.
9
0
3
1
:
v
i
X
r
a
Matthew Tarbard
St John's College
University of Oxford
A thesis submitted for the degree of
Doctor of Philosophy
Michaelmas 2012
To my family, for all their love and support.
Acknowledgements
Firstly, I would like to thank my supervisor, Professor Richard Haydon, whose invaluable
suggestions, help and support made the work in this thesis possible.
I especially thank the hard work of my examiners, Professor Charles Batty and Dr Matthew
Daws, who provided a very intellectually stimulating viva. Their careful reading of the the-
sis also led to the correction of several errors.
I have been fortunate enough to present some of the results in this thesis at various semi-
nars and conferences. These opportunities, and the people I have met through these events,
helped shape my thoughts and ideas, which led to some of the results in this thesis.
I
would therefore like to thank Dr Bunyamin Sari for inviting me to the BIRS Banach Space
Theory Workshop 2012, where I learnt a great deal about Bourgain-Delbaen spaces and
other interesting problems. I would also like to thank Professor William B. Johnson, who
brought to my attention at the BIRS workshop a corollary of my result that I had not
originally spotted. I would also like to thank Professor Charles Batty, Dr Matthew Daws
and Dr Niels Laustsen for inviting me to speak at seminars at the Universities of Oxford,
Leeds and Lancaster respectively.
Special thanks also go to Dr Richard Earl, who has provided me with fantastic teaching
opportunities at Worcester College throughout my DPhil, as well as introducing me to a
number of great people, notably, Dr Brian King and Dr Martin Galpin. I am thankful to
Richard, Brian and Martin for introducing me to bridge, and for the considerable generosity
they have shown me (particularly at the Worcester College Bar)!
I am of course very grateful to all my friends and family, who have supported me through-
out my DPhil.
I am particularly thankful to Victoria Mason, who has provided endless
encouragement and support whilst writing this thesis, and contributed significantly to the
non-academic aspect of my life. Special thanks also go to Stephen Belding, David Hewings,
Tanya Gupta and Carly Leighton for always being around to socialise with and making
Oxford a more interesting place to live.
Finally, I'd like to acknowledge that my doctoral studies were funded by the EPSRC (En-
gineering and Physical Sciences Research Council).
Abstract
The research in this thesis was initially motivated by an outstanding problem posed by
Argyros and Haydon. They used a generalised version of the Bourgain-Delbaen construction
to construct a Banach space XAH for which the only bounded linear operators on XAH
are compact perturbations of (scalar multiples of) the identity; we say that a space with
this property has very few operators. The space XAH possesses a number of additional
interesting properties, most notably, it has (cid:96)1 dual. Since (cid:96)1 possesses the Schur property,
weakly compact and norm compact operators on XAH coincide. Combined with the other
properties of the Argyros-Haydon space, it is tempting to conjecture that such a space must
necessarily have very few operators. Curiously however, the proof that XAH has very few
operators made no use of the Schur property of (cid:96)1. We therefore arrive at the following
question (originally posed in [2]): must a HI, L∞, (cid:96)1 predual with few operators (every
operator is a strictly singular perturbation of λI) necessarily have very few operators?
We begin by giving a detailed exposition of the original Bourgain-Delbaen construction
and the generalised construction due to Argyros and Haydon. We show how these two
constructions are related, and as a corollary, are able to prove that there exists some δ > 0
and an uncountable set of isometries on the original Bourgain-Delbaen spaces which are
pairwise distance δ apart.
We subsequently extend these ideas to obtain our main results. We construct new Banach
spaces of Bourgain-Delbaen type, all of which have (cid:96)1 dual. The first class of spaces are
HI and possess few, but not very few operators. We thus have a negative solution to the
Argyros-Haydon question. We remark that all these spaces have finite dimensional Calkin
algebra, and we investigate the corollaries of this result. We also construct a space with
(cid:96)1 Calkin algebra and show that whilst this space is still of Bourgain-Delbaen type with (cid:96)1
dual, it behaves somewhat differently to the first class of spaces.
Finally, we briefly consider shift-invariant (cid:96)1 preduals, and hint at how one might use the
Bourgain-Delbaen construction to produce new, exotic examples.
Contents
1 Introduction
1.1 Historical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Overview of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Notation and Elementary Definitions . . . . . . . . . . . . . . . . . . . . . .
1.4 Preliminary Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Basic sequence techniques . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2
Separation Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.3 Hereditary Indecomposability . . . . . . . . . . . . . . . . . . . . . .
1.4.4 Elementary Results from Operator Theory . . . . . . . . . . . . . .
1.4.5 Complexification . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.6
Strictly Singular Operators
. . . . . . . . . . . . . . . . . . . . . . .
2 The Bourgain-Delbaen Construction
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 The generalised Bougain-Delbaen Construction . . . . . . . . . . . . . . . .
2.3 The Bourgain-Delbaen Construction . . . . . . . . . . . . . . . . . . . . . .
2.4 Connecting the two constructions . . . . . . . . . . . . . . . . . . . . . . . .
2.5 The operator algebras for the original Bourgain-Delbaen spaces . . . . . . .
2.5.1 Construction of the basic operator . . . . . . . . . . . . . . . . . . .
2.5.2 Non-separability of L(X)
. . . . . . . . . . . . . . . . . . . . . . . .
3 Spaces with few but not very few operators
3.1 The Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Corollaries of the Main Theorem . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . .
3.2.1 On the structure of the closed ideals in L(Xk).
3.2.2 The Calkin algebra L(Xk)/K(Xk).
3.2.3 Commutators in Banach Algebras
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
3.2.4
Invariant subspaces.
. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 The Basic Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
i
1
1
3
5
6
6
9
10
12
16
17
28
28
29
37
39
48
49
54
60
60
63
63
64
65
66
66
3.4 Rapidly Increasing Sequences and the operator S : Xk → Xk . . . . . . . . .
3.5 Operators on the Space Xk
. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Strict Singularity of S : Xk → Xk
3.7 The HI Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
4 A Banach space with (cid:96)1 Calkin algebra
4.1 The Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 The Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 L(X∞)/K(X∞) is isometric to (cid:96)1(N0) . . . . . . . . . . . . . . . . . . . . . .
5 Shift invariant (cid:96)1 preduals
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Connection to the BD construction . . . . . . . . . . . . . . . . . . . . . . .
5.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References
78
88
95
98
100
100
102
119
121
121
127
133
134
ii
Chapter 1
Introduction
1.1 Historical background
Knowledge about the types of bounded linear operator that exist from a Banach space into
itself can reveal much about the structure of the underlying Banach space. In particular, it is
possible to infer a great deal about the structure of the space X when its operator algebra,
L(X), is 'small'. The first substantial results in this direction are those of Gowers and
Maurey, presented in [16] and [15]. As a motivational example, we consider the space Xgm
constructed by Gowers and Maurey in [16]. Here it was shown that all (bounded linear)
operators defined on a subspace Y of Xgm (and mapping into Xgm) are strictly singular
perturbations of the inclusion operator iY : Y → Xgm. More precisely, every such operator
is expressible in the form λiY + S, where λ is a scalar and S : Y → Xgm is a strictly singular
operator.
We shall give a precise definition of a strictly singular operator in Section 1.4.6. For now,
it suffices to think of the strictly singular operators as those which are, in some sense, small.
Indeed, it is a well known result of Fredholm theory that strictly singular perturbations of
Fredholm operators are still Fredholm, with the same index.
The representation of operators on subspaces of Xgm just discussed allows us to infer
some remarkable structural properties of the space Xgm. We obtain the following:
1. Xgm is not decomposable, that is, it cannot be written as a (topological) direct sum
of two of its infinite dimensional subspaces. This is because a non-trivial projection
is not expressible as a strictly singular perturbation of the identity.
2. In fact, we conclude by the same argument that Xgm is hereditarily indecomposable,
It
that is to say, no closed infinite dimensional subspace of Xgm is decomposable.
follows that no subspace of Xgm has an unconditional basis, i.e. Xgm has no un-
conditional basic sequence. Indeed, if (ei)∞
i=1 were an unconditional basis for some
1
subspace Y ⊆ Xgm, then we could decompose Y as Y = [e2i]∞
i=1. (We
remark, and contrast this to, the well known fact that every Banach space contains a
i=1 ⊕ [e2i−1]∞
basic sequence.)
3. Xgm is not isomorphic to any of its proper subspaces.
Indeed, it follows from the
operator representation and elementary results from Fredholm theory that every op-
erator from Xgm to itself is either strictly singular (and thus not an isomorphism),
In the latter case, if Xgm were isomorphic to a proper
or Fredholm with index 0.
subspace Y , then there would be an isomorphic embedding T : Xgm → Xgm that maps
onto Y (we simply take an isomorphism from Xgm → Y and then compose with the
inclusion operator sending Y into Xgm). In particular T is not strictly singular and so
must be Fredholm with index 0. However, this is clearly not the case as T is injective
but not onto Xgm.
One could of course consider the relationship between a Banach space and its operator
algebra from a different perspective to that just described. Instead of assuming we have some
well behaved properties of the operator algebra and asking what consequence this has for the
structure of the underlying Banach space, we may choose to impose some kind of structural
conditions on a Banach space and see what affect this has on the associated operator algebra.
In 1980, Bourgain and Delbaen [7, 6] introduced two classes of separable Banach spaces
which have 'well-behaved' finite dimensional structure, specifically they are L∞-spaces (we
refer the reader to Definition 2.2.1 for more details). These spaces have many interesting
properties and recently, Argyros and Haydon (see [2]) have managed to modify the original
construction of Bourgain and Delbaen to construct a space, XAH, which solves the scalar-
plus-compact problem. More precisely XAH is a (hereditarily indecomposable) L∞ space
with a Schauder basis, (cid:96)1 dual and L(XAH) = RI ⊕ K(XAH) (where, as usual, K(XAH)
denotes the subspace of L(XAH) of compact operators). The proof that all operators are
compact perturbations of the identity made essential use of the finite dimensional structure
of the space and the specific structure of Bourgain-Delbaen spaces, which embeds some
very explicit finite dimensional (cid:96)∞−spaces into XAH. Interestingly, the Schur property of (cid:96)1
plays no role in the proof that XAH has the scalar-plus-compact property; this will be the
main subject of Chapter 3, and so we will defer any further discussion of this until then.
As well as solving the scalar-plus-compact problem, the space XAH is interesting for
many reasons. It is well known that the space of compact operators K(X) is a separable
subspace of L(X) whenever X is a Banach space which has the approximation property and
a separable dual space. It follows that the operator algebra L(XAH) is separable. Moreover,
2
it is elementary to show that the compact operators on XAH are a complemented subspace
of L(XAH).
Another interesting property of the space XAH is that every operator in L(XAH) has a
proper closed invariant subspace, the first space for which such a result is known. Indeed,
Aronszajn and Smith showed in [3] that every compact operator on a Banach space has a
proper closed invariant subspace. Clearly if a subspace is invariant under some operator,
then it is invariant under that operator perturbed by some scalar multiple of the identity.
In particular, every operator of the form λI + K, with K compact admits a proper, closed
invariant subspace and the claim about XAH follows. (One may also prove the claim by
appealing to a result of Lomonosov, [23], which states that every operator that commutes
with a non-zero compact operator must have a proper invariant subspace.)
1.2 Overview of the thesis
In this thesis, we will continue to investigate the interplay between the structure of a Banach
space and its associated operator algebra. The question originally motivating the research
that follows was whether or not the (cid:96)1 duality of the Argyros-Haydon space XAH forces the
scalar-plus-compact property. Stated slightly more precisely, must every strictly singular
operator on a space with (cid:96)1 dual necessarily be compact if the operator algebra of the space
is small? Standard facts about basic sequences and strictly singular operators will play a
prominent role in this work, so we begin by recalling in Section 1.4 some well known results
from Banach space theory and operator theory which will be used throughout the thesis.
Whilst we present nothing new in this chapter, many of the results and proofs are difficult
to find in the existing literature. For completeness, results we will frequently rely on are
stated and we either provide references to their proofs or outline them here.
Chapter 2: Given the success of Argyros and Haydon in modifying the Bourgain-Delbaen
construction to solve the scalar-plus-compact problem, the spaces of 'Bourgain-Delbaen
type' are a sensible starting place to look for examples of exotic Banach spaces having
(cid:96)1 dual and well behaved operator algebras. The construction employed by Argyros and
Haydon in [2] appears to be somewhat different to the original construction of Bourgain and
Delbaen in, for example, [7, 6]. We begin Chapter 2 by investigating the relationship between
the two constructions. Specifically, we show that the Argyros-Haydon construction in [2]
is essentially a generalised 'dualised version' of the original Bourgain-Delbaen construction.
Of course, this is certainly not a new result, however, to the best knowledge of the author,
the precise details of this connection cannot be found in any of the existing literature.
The Argyros-Haydon approach to the Bourgain-Delbaen construction enables us to see
how we might go about constructing interesting operators on spaces of Bourgain-Delbaen
3
type. Having understood how the original Bourgain-Delbaen construction fits into that of
the Argyros-Haydon construction, we present our first new result at the end of Chapter
2:
for any of the original Bourgain-Delbaen spaces, X, constructed in [6], there exists a
constant c > 0 and an uncountable set of isometries from X to itself, such that any two
(non-equal) isometries are separated by distance at least c with respect to the operator
norm. In particular, L(X) is non-separable for all of the original Bourgain-Delbaen spaces.
This answers a question of Beanland and Mitchell ([4]).
Chapter 3: In Chapter 3, we address a question left over from the work of Argyros and
Haydon in [2], namely, must the strictly singular and compact operators on a space with (cid:96)1
dual coincide if every operator on the space is a strictly singular perturbation of the identity?
The main result of this chapter is a negative solution to this question. In fact, we exhibit
for each natural number k, a Banach space Xk with (cid:96)1 dual such that the Calkin algebra
L(Xk)/K(Xk) is k-dimensional, with basis {I + K(Xk), S + K(Xk), . . . Sk−1 + K(Xk)}. Here
S is some strictly singular operator on Xk. The results are, in some sense, continuations
of the work of Gowers and Maurey and Argyros and Haydon already mentioned; indeed,
for each space Xk, we have a representation of a general operator as some polynomial of S,
possibly perturbed by a compact operator. We immediately obtain from this an interesting
corollary concerning the norm-closed ideal structure of L(Xk); it is a finite, totally ordered
chain of ideals. We also observe that the space X2 provides a counterexample for an open
conjecture of Johnson concerning the form of commutators in the operator algebra. The
material in this chapter is an expanded version of the published paper [28].
Chapter 4 : Having shown it is possible to construct Banach spaces with Calkin algebras
of any finite dimension, the natural generalisation is to consider what infinite dimensional
Calkin algebras can be obtained. In Chapter 4 we exhibit a Banach space, X∞, which has
Calkin algebra isometric (as a Banach algebra) to the algebra (cid:96)1(N0). This generalises a
result of Gowers ([15]) where a Banach space X was constructed with the property that the
quotient algebra L(X)/SS(X) is isomorphic as a Banach algebra to (cid:96)1(Z). (Here SS(X)
denotes the strictly singular operators on X.) The ideas used to construct X∞ are very
similar to those of Chapter 3, though the proof is a little harder, requiring a combination
of the arguments from Chapter 3 and the Gowers' paper [15].
Chapter 5 : The reader will notice that all the Banach spaces constructed in this thesis
have (cid:96)1 dual. We conclude the thesis by discussing the exoticness of (cid:96)1 preduals and some
open research problems. This leads us to discuss some recently published work by Daws,
Haydon, Schlumprecht and White on so-called 'shift-invariant preduals of (cid:96)1' (see [9] for
more details). We show that one of the spaces constructed in the aforementioned paper can
in fact be considered (in some sense) as being obtained from a specific Bourgain-Delbaen
4
construction. It is hoped by the author that this insight may eventually lead to new, exotic
(cid:96)1 preduals.
1.3 Notation and Elementary Definitions
Generally notation will be introduced as and when needed. Nevertheless, in this section we
remind the reader of some standard notation and state the conventions and definitions that
will be used throughout this thesis without further explanation.
• All Banach spaces are assumed to be over the real scalar field unless explicitly stated.
Given a subset of vectors of a Banach space X, {xi}i∈I , we denote by [xi]i∈I the
smallest closed subspace of X generated by the (xi)i∈I . In other words, [xi]i∈I is the
closed linear span of the (xi)i∈I .
• The continuous dual space of a Banach space X is denoted by X∗ as usual, unless
explicitly stated otherwise. (In Chapters 3 and 4 we introduce a 'star notation' in an
attempt to be consistent with this notation.)
• We occasionally make use of the 'angle bracket notation': for vectors x∗ ∈ X∗ and
x ∈ X,(cid:104)x∗, x(cid:105) = x∗(x), the evaluation of the functional x∗ at x.
• Given a closed subspace M of a Banach space X, the quotient space X/M is the
Banach space which as a set consists of the cosets {x + M}. We equip it with the
obvious vector space operations and use the norm (cid:107)x + M(cid:107) = dist(x, M ) = inf{(cid:107)x −
m(cid:107) : m ∈ M}.
• All operators between Banach spaces are assumed to be continuous and linear. For
an operator T : X → Y between Banach spaces (X,(cid:107)(cid:107)X ) and (Y,(cid:107)(cid:107)Y ), we denote by
(cid:107)T(cid:107) the usual operator norm of T , that is, (cid:107)T(cid:107) = sup{(cid:107)T x(cid:107)Y : (cid:107)x(cid:107)X ≤ 1}.
• With the same notation as above, T ∗ : Y ∗ → X∗ denotes the dual operator defined
by T ∗y∗(x) = y∗(T x) for all x ∈ X and y∗ ∈ Y ∗.
• An operator T : X → Y is said to be an isomorphism if there exist strictly positive
constants a, b > 0, such that
a(cid:107)x(cid:107)X ≤ (cid:107)T x(cid:107)Y ≤ b(cid:107)x(cid:107)X
Note we do not require an isomorphism to be onto. When we wish to be explicit
about an isomorphism that is not surjective, we use the term isomorphic embedding .
5
• An onto operator T : X → Y between normed spaces X and Y is said to be a quotient
operator if the (bounded) linear operator T : X/ Ker T → Y , defined by T (x+Ker T ) =
T x is an isomorphism, i.e.
it has a continuous inverse. It is easy to check that this
is equivalent to the condition 'there exists a constant M > 0 such that BY (0; 1) ⊂
M T (BX (0; 1))'.
1.4 Preliminary Results
1.4.1 Basic sequence techniques
It would be impossible to include all the known results about Schauder bases and basic
sequences. Nevertheless, the notion of a Schauder basis and selected results from basic
sequence techniques in Banach space theory feature prominently in this thesis.
In this
section, we state the results we will need throughout the rest of this work.
Definition 1.4.1. Let X be an infinite dimensional Banach space.
• A sequence (xn)∞
n=1 ⊂ X is said to be a Schauder basis for X if whenever x ∈ X,
n=1 αnxn (where
n=1 is a basic sequence in X if
there exists a unique sequence of scalars (αn)∞
the sum converges in the norm topology). We say (xn)∞
it is a Schauder basis for the closed linear span [xn : n ∈ N].
n=1 such that x =(cid:80)∞
• More generally, a sequence (Fn)∞
n=1 of non-trivial subspaces of X is said to form a
Schauder decomposition of X, written X = ⊕n∈NFn, if every x ∈ X can be written
uniquely as(cid:80)∞
n=1 xn with xn ∈ Fn for all n.
In a Banach space X we say that vectors yj are successive linear combinations, or that
(yj) is a block sequence of a basic sequence (xi) if there exist 0 = q1 < q2 < ··· such
that, for all j ≥ 1, yj is in the linear span [xi : qj−1 < i ≤ qj]. If we may arrange that
yj ∈ [xi : qj−1 < i < qj] we say that (yj) is a skipped block sequence. More generally, if X
n∈N Fn we say that (yj) is a block sequence (resp. a
skipped block sequence) with respect to (Fn) if there exist 0 = q0 < q1 < ··· such that yj is
Fn). A block subspace is the closed subspace generated
has a Schauder decomposition X =(cid:76)
in(cid:76)
Fn (resp. (cid:76)
qj−1<n≤qj
qj−1<n<qj
by a block sequence.
Finally, a basic sequence (xn) is called unconditional if every permutation of the sequence
is a Schauder basis for [xn : n ∈ N].
It is well known (see, for example, [1, Proposition 1.1.9]) that a sequence (xn)∞
n=1 of
non-zero vectors in a Banach space X is a basic sequence if, and only if, there exists a
6
i=1 are scalars,
(cid:107) m(cid:88)
constant M ≥ 1 such that whenever m < n are natural numbers and (λi)n
λixi(cid:107) ≤ M(cid:107) n(cid:88)
defined by Pk : [xn : n ∈ N] → [xn : n ≤ k], (cid:80)∞
n : [xn : n ∈ N] → K defined by (cid:80)∞
From this, it follows easily that there exist uniformly bounded linear projections (Pk)∞
k=1
n=1 anxn; the constant,
supk∈N (cid:107)Pk(cid:107), is known as the basis constant. We note that the co-ordinate functionals,
m=1 amxm (cid:55)→ an (also known in the literature as
x∗
biorthogonal functionals) are continuous and refer the reader to [25, Theorem 4.1.14] for
n=1 anxn (cid:55)→ (cid:80)k
λixi(cid:107).
i=1
i=1
further details of these facts.
When the Banach space has an unconditional basis, (xn)∞
n=1, it admits a number of
non-trivial projections and the associated operator algebra has a rich structure. Indeed, it
is known that for any (ξn)∞
n=1 ξnanxn
is a bounded operator. In particular, for every A ⊂ N, the linear projection PA : X → X
n∈A anxn is bounded; in fact, supA⊂N (cid:107)PA(cid:107) < ∞. More details
n=1 ∈ (cid:96)∞, T(ξn) : X → X defined by(cid:80)∞
n=1 anxn (cid:55)→(cid:80)∞
n=1 anxn (cid:55)→(cid:80)
defined by(cid:80)∞
and proofs of these facts can be found in [1, Proposition 3.1.3, Remark 3.1.5].
Frequently we will have some condition that needs to be checked for all closed infinite
dimensional subspaces of a Banach space. When the Banach space has a Schauder basis, it
is often possible to show that it is sufficient to check the condition only on block subspaces.
We will make use of this idea later in the thesis. The proofs of such results generally rely
on the following technical result.
Proposition 1.4.2. Let X be a Banach space with Schauder basis (en)∞
n=1 and suppose Y
is a closed, infinite dimensional subspace of X. Denote by K the basis constant of the basis
n=1 in Y with (cid:107)yn(cid:107) = 1 for all n
(en)∞
and a block sequence (xn)∞
n=1. For every 0 < θ < 1, we can find a sequence (yn)∞
n=1 in X satisfying
∞(cid:88)
n=1
2K
(cid:107)yn − xn(cid:107)
(cid:107)xn(cid:107) < θ
n=1,(cid:80)∞
n=1 is a basic sequence equivalent to (xn)∞
n=1 anyn converges if and only if(cid:80)∞
The sequence (yn)∞
scalars, (an)∞
Proof. Note that for any n ∈ N, there must exist a non-zero vector y ∈ Y with e∗
m(y) = 0
for all m ≤ n (here, the e∗
m are the biorthogonal functionals of the basis (en)). Indeed, if
not, the (bounded) projection Pn : Y → [em : m ≤ n] is injective; this contradicts the fact
that Y is infinite dimensional.
n=1, i.e. given a sequence of
n=1 anxn converges.
It follows from the above observation that we can pick a sequence (y(cid:48)
n(cid:107) = 1 for every n and e∗
n=1 ⊆ Y with
n)∞
n) = 0 for all m ≤ n. By the proof of the Bessaga-Pelczy´nski
m(y(cid:48)
(cid:107)y(cid:48)
7
Selection Principle given in [1, Proposition 1.3.10], there is a subsequence of the y(cid:48)
say, and a block basic sequence (xk) (with respect to the basis (en)) satisfying
n, (y(cid:48)
)
nk
∞(cid:88)
k=1
2K
(cid:107)y(cid:48)
nk
− xk(cid:107)
(cid:107)xk(cid:107)
< θ
The first part of the proposition is proved by simply setting yk = y(cid:48)
block basic sequence (xk)∞
k=1 as above.
nk
To see that (yn)∞
n=1, it is
enough, by standard results, to show that there is an isomorphism from X to X which
n=1 is a basic sequence equivalent to the basic sequence (xn)∞
and choosing the
sends xn to yn for every n. This is proved in [1, Theorem 1.3.9].
The Banach spaces constructed later in this thesis rely heavily on the following well-
known proposition:
Proposition 1.4.3. Suppose that (x∗
sociated to a basis (xn)∞
(with basis constant no bigger than that of (xn)∞
n=1).
n=1 of a Banach space X. Then (x∗
n)∞
n=1 is the sequence of biorthogonal functionals as-
n)∞
n=1 is a basic sequence in X∗
If we denote by H := [x∗
n : n ∈ N] the closed subspace of X∗ formed by taking the closed
linear span of the biorthogonal vectors, then X embeds isomorphically into H∗ under the
natural mapping j : X → H∗, x (cid:55)→ j(x)H (where j : X → X∗∗ is the canonical embedding
of X into its bidual).
Proof. See [1, Proposition 3.2.1, Lemma 3.2.3 and Remark 3.2.4].
We conclude this brief overview of basic sequences by recalling the notions and some
basic facts about shrinking and boundedly-complete bases.
Definition 1.4.4. Let X be a Banach space.
tionals (x∗
• A basis (xn)∞
n=1 of X is said to be shrinking if the sequence of its biorthogonal func-
n=1 is a basis for X∗.
n)∞
• A basis (xn)∞
n=1 of X is said to be boundedly-complete if whenever (an)∞
n=1 is a sequence
of scalars such that
the series(cid:80)∞
n=1 anxn converges.
(cid:107)
sup
N
anxn(cid:107) < ∞
We will use the following results repeatedly.
Proposition 1.4.5. A basis (xn)∞
bounded block basic sequence of (xn)∞
n=1 is weakly null.
∞(cid:88)
n=1
8
n=1 of a Banach space X is shrinking if and only if every
n=1 be a basis for a Banach space X with biorthogonal functionals
Proof. See [1, Proposition 3.2.7].
Theorem 1.4.6. Let (xn)∞
n)∞
(x∗
n=1. The following are equivalent.
(i) (xn)∞
n=1 is a boundedly-complete basis for X,
(ii) (x∗
n)∞
n=1 is a shrinking basis for H := [x∗
(iii) The map j : X → H∗ defined by j(x)(h) = h(x), for all x ∈ X and h ∈ H, is an onto
n : n ∈ N],
isomorphism.
Proof. See [1, Theorem 3.2.10].
1.4.2 Separation Theorems
There are a number of separation theorems in the existing literature. Whilst we do not feel
it is necessary to present the Hahn-Banach Theorem as found in almost any introductory
textbook on functional analysis, we find it convenient to state two separation theorems.
These results are sufficiently general to cover all cases we will need in this thesis.
Theorem 1.4.7 (Hahn, Banach). Let C be a closed convex set in a Banach space X. If
x0 /∈ C then there is f ∈ X∗ such that Re(f (x0)) > sup{Re(f (x)) : x ∈ C}.
Proof. See [11, Theorem 2.12].
Theorem 1.4.8. Let X be a topological vector space (TVS) and let C1 and C2 be non-empty
convex subsets of X such that C2 has non-empty interior. If C1 ∩ C◦
2 = ∅ then there is a
continuous linear functional x∗ ∈ X∗ and a real number s such that
(1) Re x∗x ≥ s for each x ∈ C1;
(2) Re x∗x ≤ s for each x ∈ C2;
(3) Re x∗x < s for each x ∈ C◦
2 .
Proof. See [25, Theorem 2.2.26].
9
1.4.3 Hereditary Indecomposability
We have already discussed hereditarily indecomposable (HI) spaces in the introduction.
However, since we will want to prove that the spaces constructed in Chapter 3 are HI, we
take this opportunity to formally state the HI condition.
A Banach space X is indecomposable if there do not exist infinite-dimensional closed
subspaces Y and Z of X such that X can be written as the topological direct sum X = Y ⊕Z;
that is to say the bounded operator
Y ⊕ext Z → Y ⊕ Z
(y, z) (cid:55)→ y + z
fails to have continuous inverse. X is hereditarily indecomposable (HI) if every closed
subspace is indecomposable.
It follows that X is HI if and only if, whenever Y and Z
are infinite dimensional, closed subspaces of X, there does not exist δ > 0 such that
(cid:107)y + z(cid:107) ≥ δ((cid:107)y(cid:107) + (cid:107)z(cid:107)) for all y ∈ Y, z ∈ Z.
δ > 0 to exist, it would certainly have to be the case that Y ∩ Z = {0}.)
Schauder basis, it is sufficient to check the condition on block subspaces. More precisely
(We note in particular that were such a
If X has a
Proposition 1.4.9. Let X have a Schauder basis. Then X is HI if and only if whenever
Y, Z are block subspaces, there does not exist δ > 0 such that (cid:107)y + z(cid:107) ≥ δ((cid:107)y(cid:107) + (cid:107)z(cid:107)) for all
y ∈ Y, z ∈ Z.
Proof. Clearly if X is HI then the condition on block subspaces holds. Conversely, suppose
the condition on block subspaces holds and for contradiction that there exist infinite dimen-
sional, closed subspaces Y , Z, with Y ∩ Z = {0} and such that Y ⊕ Z is a topological direct
sum. Consequently, there exists δ > 0 such that (cid:107)y + z(cid:107) ≥ δ((cid:107)y(cid:107) +(cid:107)z(cid:107)) for all y ∈ Y, z ∈ Z.
(Clearly for this to hold, we must have δ ≤ 1.)
We choose 0 < ε < 1/4 and such that
n=1 ⊂ Y , (zn)∞
n=1 ⊂ Z and block basic sequences (y(cid:48)
4ε
1−4ε < δ. By Proposition 1.4.2 we can find norm
n)∞
n=1 with the
n)∞
n=1, (z(cid:48)
1 sequences (yn)∞
property that
2K
and
∞(cid:88)
∞(cid:88)
n=1
(cid:107)yn − y(cid:48)
n(cid:107)
(cid:107)y(cid:48)
n(cid:107) < ε
(cid:107)zn − z(cid:48)
n(cid:107)
n(cid:107) < ε
where K is the basis constant of the Schauder basis of X.
(cid:107)z(cid:48)
2K
n=1
Let Y (cid:48) be the block subspace generated by (y(cid:48)
n)∞
n=1. Note the basis constant of the basic sequence (y(cid:48)
n)∞
n=1 and Z(cid:48) the block subspace generated
n) is at most K as it is a block
by (z(cid:48)
10
basic sequence. It follows that if y(cid:48) :=(cid:80)∞
Consequently, if y(cid:48) :=(cid:80)∞
(we recall that (yn) and (y(cid:48)
n=1 any(cid:48)
n ∈ Y (cid:48), then y :=(cid:80)∞
n ∈ Y (cid:48) then an ≤ 2K(cid:107)y(cid:48)
∞(cid:88)
n=1 any(cid:48)
∞(cid:88)
n) are equivalent) and that
n(cid:107)(cid:107)y(cid:48)(cid:107) for every n.
n=1 anyn is a well-defined vector in Y
(cid:107)y − y(cid:48)(cid:107) = (cid:107)
an(yn − y(cid:48)
n)(cid:107) ≤ 2K(cid:107)y(cid:48)(cid:107)
n=1
n=1
n(cid:107)
(cid:107)yn − y(cid:48)
n(cid:107) < ε(cid:107)y(cid:48)(cid:107).
(cid:107)y(cid:48)
An analogous result holds for z(cid:48) ∈ Z(cid:48); we get a corresponding z ∈ Z with (cid:107)z − z(cid:48)(cid:107) < ε(cid:107)z(cid:48)(cid:107).
We now suppose that y(cid:48) ∈ Y (cid:48) has norm 1, so that the corresponding vector y ∈ Y
satisfies (cid:107)y(cid:107) > 1 − ε. If z(cid:48) ∈ Z(cid:48) is such that (cid:107)z(cid:48)(cid:107) ≤ 1 + δ then
(cid:107)y(cid:48) − z(cid:48)(cid:107) = (cid:107)(y − z + y(cid:48) − y) − (z(cid:48) − z)(cid:107)
≥ (cid:107)(y − z) − (y − y(cid:48))(cid:107) − (cid:107)z(cid:48) − z(cid:107)
≥ (cid:107)y − z(cid:107) − (cid:107)y − y(cid:48)(cid:107) − (cid:107)z(cid:48) − z(cid:107)
≥ δ((cid:107)y(cid:107) + (cid:107)z(cid:107)) − ε(cid:107)y(cid:48)(cid:107) − ε(cid:107)z(cid:48)(cid:107)
≥ δ(cid:107)y(cid:107) − ε − ε(1 + δ)
≥ δ(1 − ε) − 2ε − εδ
>
δ
2
.
where the final inequality follows by choice of ε. On the other hand, if z(cid:48) ∈ Z(cid:48) is such that
(cid:107)z(cid:48)(cid:107) > 1 + δ, then (cid:107)y(cid:48) − z(cid:48)(cid:107) ≥ (cid:107)z(cid:48)(cid:107) − (cid:107)y(cid:48)(cid:107) > 1 + δ − 1 > δ
2 .
We have therefore shown that for any y(cid:48) ∈ Y (cid:48) with (cid:107)y(cid:48)(cid:107) = 1 and z(cid:48) ∈ Z(cid:48), (cid:107)y(cid:48) − z(cid:48)(cid:107) > δ
2 .
2(cid:107)y(cid:48)(cid:107). By symmetry of
2(cid:107)z(cid:48)(cid:107). Consequently,
4 ((cid:107)y(cid:48)(cid:107) + (cid:107)z(cid:48)(cid:107)); this contradictions
By scaling, this implies that whenever y(cid:48) ∈ Y (cid:48), z(cid:48) ∈ Z(cid:48), (cid:107)y(cid:48) + z(cid:48)(cid:107) ≥ δ
the argument, we also have that whenever y(cid:48) ∈ Y (cid:48), z(cid:48) ∈ Z(cid:48), (cid:107)y(cid:48) + z(cid:48)(cid:107) ≥ δ
for all y(cid:48) ∈ Y (cid:48), z(cid:48) ∈ Z(cid:48), (cid:107)y(cid:48) + z(cid:48)(cid:107) ≥ δ
the assumed property satisfied by block subspaces.
2 max((cid:107)y(cid:48)(cid:107),(cid:107)z(cid:48)(cid:107)) ≥ δ
Whilst there are a number of additional results known about HI spaces, the only other
result we will need for the purposes of this thesis is the following proposition (which can be
found in [2]).
Proposition 1.4.10. Let X be an infinite dimensional Banach space. Then X is HI if and
only if, for every pair Y, Z of closed, infinite-dimensional subspaces, and every > 0, there
exist y ∈ Y and z ∈ Z with (cid:107)y − z(cid:107) < ε(cid:107)y + z(cid:107). If X has a Schauder basis it is enough that
the previous condition should hold for block subspaces.
Proof. We claim that whenever Y and Z are infinite dimensional, closed subspaces, Y + Z
fails to be a topological direct sum if and only if ∀ε > 0,∃y ∈ Y, z ∈ Z with (cid:107)y−z(cid:107) < ε(cid:107)y+z(cid:107).
This clearly proves the first part of the proposition. Moreover, once we have proved this
11
claim, we can deduce that if the condition holds for block subspaces, then no two block
subspaces form a topological direct sum. The previous proposition then implies that X is
HI. Therefore, it only remains to prove the claim made at the beginning of the proof.
Suppose first that Y and Z are closed infinite dimensional subspaces for which Y + Z
fails to be a topological direct sum. We assume for contradiction that there exists ε > 0
such that (cid:107)y − z(cid:107) ≥ ε(cid:107)y + z(cid:107) whenever y ∈ Y, z ∈ Z. It follows that (cid:107)y + z(cid:107) ≥ ε(cid:107)y − z(cid:107)
for all y ∈ Y, z ∈ Z. Consequently, (cid:107)y(cid:107) = 1
2 ((cid:107)y + z(cid:107) + (cid:107)y − z(cid:107)) ≤
ε )(cid:107)y + z(cid:107).
2 ((cid:107)y + z(cid:107) + 1
ε )(cid:107)y + z(cid:107). A similar calculation yields (cid:107)z(cid:107) ≤ 1
2(cid:107)y + z + y − z(cid:107) ≤ 1
ε(cid:107)y + z(cid:107)) = 1
2 (1 + 1
2 (1 + 1
1
So
(cid:107)y + z(cid:107) ≥ 2(1 +
1
ε
)−1 max((cid:107)y(cid:107),(cid:107)z(cid:107)) ≥ (1 +
)−1((cid:107)y(cid:107) + (cid:107)z(cid:107))
1
ε
from which it follows that Y ⊕ Z is a direct sum, giving us the required contradiction.
Conversely, assume that Y and Z satisfy the condition that ∀ε > 0,∃y ∈ Y, z ∈ Z
with (cid:107)y − z(cid:107) < ε(cid:107)y + z(cid:107). We must see that Y + Z fails to be a topological direct sum.
We again argue by contradiction, assuming that Y + Z is topological. Consequently, there
exists δ > 0 such that (cid:107)y + z(cid:107) ≥ δ((cid:107)y(cid:107) + (cid:107)z(cid:107)) whenever y ∈ Y, z ∈ Z.
we note that this implies that (cid:107)y + z(cid:107) ≥ δ(cid:107)z(cid:107) for all y ∈ Y, z ∈ Z.
(cid:107)y − z(cid:107) = (cid:107)y + z − 2z(cid:107) ≤ (cid:107)y + z(cid:107) + 2(cid:107)z(cid:107) ≤ (1 + 2
the other hand, taking ε = 1
by the hypothesis (with z replaced by −z).
(cid:107)y − z(cid:107) ≤ (1 + 2
2(cid:107)y − z(cid:107). This only fails to be a contradiction
δ )(cid:107)y − z(cid:107) = 1
if (cid:107)y − z(cid:107) = 0, which, since it is assumed that Y and Z are topological, can only happen
if y = z = 0. But, recalling that y, z were chosen so as to satisfy (cid:107)y + z(cid:107) < ε(cid:107)y − z(cid:107), we
clearly cannot have y = z = 0 and we once again have the contradiction we seek.
δ )(cid:107)y + z(cid:107) for all y ∈ Y, z ∈ Z. On
δ )−1, there exist y ∈ Y, z ∈ Z with (cid:107)y + z(cid:107) < ε(cid:107)y − z(cid:107)
It follows that for this choice of y and z,
δ )(cid:107)y + z(cid:107) ≤ ε(1 + 2
It follows that
In particular,
2 (1 + 2
1.4.4 Elementary Results from Operator Theory
We will make repeated use of the following duality results when constructing operators in
later chapters of this thesis.
Lemma 1.4.11. Let X be a Banach space, Y a normed space, T : X → Y a bounded linear
operator. Then T is a quotient operator if, and only if, T ∗ : Y ∗ → X∗ is an isomorphic
embedding.
Proof. Suppose first that T is a quotient operator. Then there is some M > 0 such that
BY (0; 1) ⊆ M T(cid:0)BX (0; 1)(cid:1). It follows that
M(cid:107)T ∗y∗(cid:107) = sup
x∈BX (0;1)
y∗(M T x) ≥ sup
y∈BY (0;1)
y∗y = (cid:107)y∗(cid:107)
12
i.e. that T ∗ is an isomorphic embedding. Suppose conversely that T ∗ is an isomorphic
embedding and let M > 0 be such that M(cid:107)T ∗y∗(cid:107) ≥ (cid:107)y∗(cid:107). We claim that M T(cid:0)BX (0; 1)(cid:1) ⊇
BY (0; 1). Since X is a Banach space, this is sufficient to deduce that T : X → Y is a
quotient operator. Let's suppose by contradiction that
M T(cid:0)BX (0; 1)(cid:1) =
(cid:92)
M T(cid:0)BX (0; 1)(cid:1) + εB◦
ε>0
Y and ε > 0 such that y0 /∈ M T(cid:0)BX (0; 1) + εB◦
Y (0; 1)
So there is some y0 ∈ B◦
Separation Theorem 1.4.8, we can find a non-zero y∗ ∈ Y ∗ with
Y (0; 1) (cid:43) B◦
Y . Applying
M(cid:107)T ∗y∗(cid:107) + ε(cid:107)y∗(cid:107) = M sup
x∈BX (0;1)
Re y∗T x + ε sup
y∈B◦
Y
Re y∗y ≤ Re y∗(y0) ≤ (cid:107)y∗(cid:107)
But this implies that (1 + ε)(cid:107)y∗(cid:107) ≤ (cid:107)y∗(cid:107) which is clearly a contradiction.
Lemma 1.4.12. Let X and Y be Banach spaces and suppose T : X∗ → Y ∗ is a bounded
linear operator which is also weak∗ to weak∗ continuous. Then there is a bounded linear
operator S : Y → X such that S∗ = T .
Furthermore, if for each n ∈ N, Tn : X∗ → X∗ is a bounded linear operators which is
weak* to weak* continuous and T : X∗ → X∗ is a bounded linear operator with (cid:107)Tn−T(cid:107) → 0
(i.e. Tn → T in the operator norm topology), then T is also weak* to weak* continuous.
Proof. Let JX : X → X∗∗, JY : Y → Y ∗∗ denote the canonical embeddings. We first see that
T ∗JY maps Y into JX (X). It is enough to see that for every y ∈ Y , the map T ∗JY y : X∗ →
K is weak∗ continuous. To this end, we suppose (x∗
α) is a net converging in the weak∗
x∗). By weak∗-weak∗ continuity of T , we have
topology to some x∗ (we write x∗
T x∗
T x∗ so that
α (cid:42)w∗
α (cid:42)w∗
(T ∗JY y)x∗
α = JY y(T x∗
α) = T x∗
α(y) → T x∗(y) = (T ∗JY y)x∗
as required. We can thus define a bounded linear map S : Y → X by S := J−1
note that for all x∗ ∈ X∗, y ∈ Y
X T ∗JY . We
(S∗x∗)y = x∗(Sy) = x∗(J−1
X T ∗JY y) = T ∗JY y(x∗) = (T x∗)y
so that S∗ = T as required.
To prove the second part of the lemma, note that by the first part of the proof, each Tn
is the dual operator of an operator Sn : X → X. Precisely, Sn = J−1
n JX . Since Tn → T
with respect to the operator norm, T ∗
n=1 is a
Cauchy sequence in L(X), so there is some bounded linear operator S such that Sn → S. It
n → T ∗ in operator norm. It follows that (Sn)∞
X T ∗
13
n → S∗ in L(X∗). However, S∗
n = Tn for all n, and since Tn → T , uniqueness
follows that S∗
of limits gives that T = S∗. Since T is the dual of some operator, it is certainly weak*
continuous.
Lemma 1.4.13. Let X and Y be Banach spaces, T : X → Y a bounded linear operator. If
T ∗ is one to one and has closed range then im T = Y .
Proof. Since T ∗ is one to one and has closed range in the Banach space Y , T ∗ is an isomor-
phism onto its image and thus T ∗ : Y ∗ → X∗ is an isomorphic embedding. It follows from
Lemma 1.4.11 that T is a quotient operator. In particular T is onto.
Theorem 1.4.14 (Closed Range Theorem). Let X and Y be Banach spaces, T : X → Y a
bounded linear operator. Then T has closed range if and only if T ∗ has closed range.
Proof. The proof is essentially taken from [10].
1 : T [X]∗ → X∗ is an isomorphic embedding and consequently that im T ∗
Let us suppose first that T has closed range. Since Y is a Banach space and im T is
closed, T1 : X → T [X] defined by T1x = T x is a quotient operator. It follows by Lemma
1.4.11 that T ∗
1 is
closed. We claim that im T ∗
1 = im T ∗ and thus the image of T ∗ is closed as required. Indeed
suppose z∗ ∈ T [X]∗ and let y∗ ∈ Y ∗ be an extension of z∗ to Y (the existence of which is
of course guaranteed by the Hahn Banach Theorem). Then for x ∈ X,
1 z∗)x = z∗T1x = y∗T x = (T ∗y∗)x
(T ∗
1 ⊆ im T ∗. Conversely, if y∗ ∈ Y ∗, we let z∗ ∈ T [X]∗ be the
1 . It
1 z∗ and so im T ∗ ⊆ im T ∗
1 z∗ = T ∗y∗ and im T ∗
So T ∗
restriction of y∗ to T [X]. It is easy to see that T ∗y∗ = T ∗
follows that im T ∗
1 = im T ∗ as required.
1 is one-to-one. If x∗ ∈ X∗ is in the closure of T ∗
Conversely, suppose T ∗ has closed range. We let Z ⊆ Y be the Banach space T [X] and
consider the map T1 : X → Z defined by T1x = T x. Since T1 has dense range, it follows
that T ∗
n where
n ∈ Y ∗ of the
z∗
n are in Z∗. We apply the Hahn Banach Theorem obtaining extensions y∗
n and since T ∗ has closed range, x∗ = T ∗y∗ for some
z∗
n. It follows that x∗ = limn→∞ T ∗y∗
y∗ ∈ Y ∗. If z∗ is the restriction of y∗ to Z, then x∗ = T ∗
1 is one-to-one and
has closed range. It follows by Lemma 1.4.13 that im T1 = im T = Z and so T has closed
range as required.
1 (Z∗) then x∗ = limn→∞ T ∗
1 z∗
1 z∗. Hence T ∗
As well as the above duality results, we will make use of the following basic sequence
technique which provides a sufficient condition for an operator to be compact.
14
Proposition 1.4.15. Let X be a Banach space with a Schauder basis and T : X → X a
bounded linear operator. If T xk → 0 for all bounded block basic sequences (xk)∞
k=1, then T
is compact. If we demand that the basis of X is shrinking, then the converse is also true.
Proof. Denote by (en)∞
n=1 the Schauder basis of X. We noted in Section 1.4.1 that the
j=1 ajej are bounded. We will
show that (cid:107)T − T Pn(cid:107) → 0 (as n → ∞); this shows that T is a uniform limit of finite rank
operators and consequently that T is compact.
projections Pn : X → [ej : j ≤ n] defined by(cid:80)∞
Suppose for contradiction that (cid:107)T − T Pn(cid:107) (cid:88)→ 0. It follows that we can find δ > 0, a
j=1 ajej (cid:55)→(cid:80)n
j=1 of natural numbers, and a sequence of norm 1 vectors
strictly increasing sequence (Nj)∞
(xj)∞
j=1 such that
(cid:107)(T − T PNj )xj(cid:107) > δ
n=1 is a Schauder basis, we can find M1 > N1 such that (cid:107)x1 − PM1x1(cid:107) < δ/2(cid:107)T(cid:107).
Since (en)∞
We set y1 = (PM1 − PN1)x1 noting that (cid:107)y1(cid:107) ≤ 2K (where K is the basis constant of the
basis (en)∞
n=1). Moreover,
(cid:107)T y1(cid:107) = (cid:107)T x1 − T PN1x1 − (T x1 − T PM1x1)(cid:107)
≥ (cid:107)(T − T PN1)x1(cid:107) − (cid:107)T ◦ (I − PM1)x1(cid:107) >
δ
2
Now, we can find Nj2 > M1 and M2 > Nj2 such that (cid:107)xj2 − PM2xj2(cid:107) < δ/2(cid:107)T(cid:107). We set
)xj2. Estimating as before yields (cid:107)y2(cid:107) ≤ 2K and (cid:107)T y2(cid:107) > δ/2. Continuing
n=1 with (cid:107)T yn(cid:107) > δ/2 for all n.
n=1 is a bounded block basic
y2 = (PM2 −PNj2
in this way, we obtain a bounded block basic sequence (yn)∞
This contradicts the hypothesis that T xn → 0 whenever (xn)∞
sequence, completing the first part of the proof.
The converse of the theorem does not hold in general. However, in our statement of a
partial converse, we demand that the basis is shrinking. Let (xn)∞
n=1 be a bounded block
basic sequence. We are required to show that if T is compact, T xn → 0. Arguing by
contradiction we assume that, after choosing a subsequence and relabelling if necessary,
there is δ > 0 such that (cid:107)T xn(cid:107) ≥ δ for all n. We note that (xn)∞
Proposition 1.4.5. Consequently (T xn)∞
n=1 is weakly null by
n=1 converges weakly to 0.
Since T is compact, some subsequence (T xnj )∞
j=1 converges in norm to x ∈ X. It follows
that T xnj → x in the weak topology. By the above argument and uniqueness of weak limits
j=1 converges in norm to x, it follows that (cid:107)T xnj(cid:107) → 0,
we have that x = 0. Since (T xnj )∞
contradicting the fact that (cid:107)T xnj(cid:107) ≥ δ for all j. This completes the proof.
15
1.4.5 Complexification
We will require results from spectral theory. Of course, many of these results assume that
the Banach algebras in question are over the field of complex scalars. In order to pass to
the real case, we recall some basic complexification arguments. If X is a real Banach space,
the complexified space is defined as XC := X ⊕ X with (complex) scalar multiplication and
vector addition defined by
(x, y) + (u, v) = (x + u, y + v)
(α + iβ)(x, y) = (αx − βy, βx + αy) ∀α, β ∈ R, x, y ∈ X.
∀x, y, u, v ∈ X
Obviously we can identify X as a (real) subspace of the complexification under the linear
injection given by j : X → XC, x (cid:55)→ (x, 0). Noting that i(y, 0) = (0, y), we can write the
vector (x, y) ∈ XC as j(x) + ij(y) and it is obvious that this representation is unique. We
will often find it convenient to suppress the use of the embedding j and simply write z ∈ XC
by z = x + iy, where x, y ∈ X ⊆ XC. Consequently, we write XC = X ⊕ iX.
There are many ways to define a complex norm on XC. For the purposes of this thesis,
we shall work with the norm defined by (cid:107)x + iy(cid:107)XC := supt∈[0,2π] (cid:107)x cos t − y sin t(cid:107)X .
is a trivial exercise to check that this defines a norm and we only give the proof of the
homogeneity. Note that if cos θ + i sin θ ∈ C,
(cid:107)(cos θ + i sin θ)(x + iy)(cid:107)XC = sup
t∈[0,2π]
= sup
t∈[0,2π]
= sup
t∈[0,2π]
= (cid:107)x + iy(cid:107)XC
(cid:107)(x cos θ − y sin θ) cos t − (x sin θ + y cos θ) sin t(cid:107)X
(cid:107)(cos θ cos t − sin θ sin t)x − (sin θ cos t + cos θ sin t)y(cid:107)X
(cid:107) cos(θ + t)x − sin(θ + t)y(cid:107)X
It
so that λ(cid:107)z(cid:107)XC = (cid:107)λz(cid:107)XC for all λ ∈ C, z ∈ XC.
We remark that whenever x, y ∈ X, we have (cid:107)j(x)(cid:107)XC = (cid:107)x(cid:107)X and (cid:107)x + iy(cid:107)XC =
(cid:107)x − iy(cid:107)XC. Using these observations we obtain the following lemma.
Lemma 1.4.16. Let x, y ∈ X. Then (cid:107)x(cid:107)X = (cid:107)x(cid:107)XC ≤ (cid:107)x + iy(cid:107)XC and (cid:107)y(cid:107)X = (cid:107)y(cid:107)XC ≤
(cid:107)x + iy(cid:107)XC
Proof. Just note that 2(cid:107)x(cid:107)XC ≤ (cid:107)x + iy(cid:107)XC + (cid:107)x − iy(cid:107)XC = 2(cid:107)x + iy(cid:107)XC. The second part
of the lemma is similar.
If T is a (real) operator on X, then we can uniquely extend it to a (complex) operator
TC on XC by TC(x + iy) = T x + iT y. We have the following lemma:
16
Lemma 1.4.17. If T ∈ L(X) then TC ∈ L(XC) and (cid:107)T(cid:107)X→X = (cid:107)TC(cid:107)XC→XC. Moreover, if
T is compact, then so is TC.
Proof. It is easy to see that TC is a linear operator. Moreover, it is clear that (cid:107)TC(cid:107) ≥ (cid:107)T(cid:107)
since TC extends T (and (cid:107)j(x)(cid:107)XC = (cid:107)x(cid:107)X ). Note,
(cid:107)TC(x + iy)(cid:107)XC = sup
t∈[0,2π]
(cid:107)(T x) cos t − (T y) sin t(cid:107)X = sup
t∈[0,2π]
(cid:107)T (x cos t − y sin t)(cid:107)X
≤ (cid:107)T(cid:107) sup
t∈[0,2π]
(cid:107)x cos t − y sin t(cid:107)X = (cid:107)T(cid:107)(cid:107)x + iy(cid:107)XC
so that (cid:107)TC(cid:107) ≤ (cid:107)T(cid:107).
To see that TC is compact when T is, note that if (xn + iyn)∞
n=1 is a bounded sequence in
XC then the sequences (xn)∞
n=1 are bounded sequences in X by Lemma 1.4.16.
By compactness of T , we can choose subsequences (xnk )∞
k=1 such that both
(T xnk ) and (T ynk ) converge. It follows easily that TC(xnk + iynk ) converges in XC, so TC
is compact as required.
k=1 and (ynk )∞
n=1 and (yn)∞
1.4.6 Strictly Singular Operators
The majority of this thesis is concerned with the interplay between the compact and strictly
singular operators. We recall in this section the definition and some some of the elementary
results on strictly singular operators which will be used throughout the thesis.
We recall the following definition
Definition 1.4.18. Let X, Y be Banach spaces and T : X → Y a bounded linear operator.
T is strictly singular if whenever Z ⊆ X is a subspace such that there exists some δ > 0
with (cid:107)T z(cid:107) ≥ δ(cid:107)z(cid:107) for all z ∈ Z, then Z is finite dimensional.
In other words, a strictly singular operator is one which is not an isomorphic embedding
on any infinite dimensional (closed) subspace. The following characterisation of strictly
singular operators on a Banach space with a Schauder basis will be particularly useful to
us later on in this thesis.
n=1 and suppose T : X → X
Proposition 1.4.19. Let X be a Banach space with a basis (en)∞
is a bounded linear operator on X. Then T is strictly singular ⇐⇒ whenever [yn : n ∈ N]
is a block subspace of X, the restriction, T[yn] : [yn] → X is not an isomorphic embedding.
Proof. Clearly when T is strictly singular, we have the condition about block subspaces. To
complete the proof, we show that if T is an isomorphism on some closed infinite dimensional
subspace Y of X, then T must be an isomorphism on some block subspace of X.
17
Indeed, suppose for some infinite dimensional, closed subspace Y of X, there is δ > 0
such that for every y ∈ Y,(cid:107)T y(cid:107) ≥ δ(cid:107)y(cid:107). Choose any 0 < θ < 1
By Proposition 1.4.2, there exists a sequence (yk)∞
block basic sequence (xk) (with respect to the basis (en)) satisfying
2 such that (cid:107)T(cid:107)θ( 1
2 −θ)−1 ≤ δ.
k=1 ⊆ Y with (cid:107)yk(cid:107) = 1 for every k and a
∞(cid:88)
k=1
2K
(cid:107)yk − xk(cid:107)
(cid:107)xk(cid:107) < θ
where K is the basis constant of the basis (en)∞
n=1.
We recall that the sequences (xk) and (yk) are equivalent basic sequences. Moreover,
that T is an isomorphism on the block subspace [xk]. To see this, let(cid:80)
the sequence (xk) as a block sequence of the (en) has basis constant at most K. We claim
k akxk be a norm
1 vector in [xk], and observe that this implies in particular that ak ≤ 2K(cid:107)xk(cid:107) since (xk) is a
basic sequence with basis constant at most K. Then
akyk +
akyk
(cid:88)
ak(xk − yk)(cid:1)(cid:13)(cid:13)
(cid:1)(cid:13)(cid:13) −(cid:13)(cid:13)T(cid:0)(cid:88)
(cid:88)
k
k
ak(xk − yk)(cid:1)(cid:13)(cid:13)
(cid:107)xk − yk(cid:107)
(cid:107)xk(cid:107)
k
akyk(cid:107) − (cid:107)T(cid:107)2K
akyk(cid:107) − (cid:107)T(cid:107)θ.
ak(yk − xk)(cid:107) ≥ 1 − 2K
(cid:88)
k
(cid:107)yk − xk(cid:107)
(cid:107)xk(cid:107) ≥ 1 − θ.
(cid:107)T(cid:0)(cid:88)
k
akxk
k
k
(cid:1)(cid:107) =(cid:13)(cid:13)T(cid:0)(cid:88)
≥(cid:13)(cid:13)T(cid:0)(cid:88)
≥ δ(cid:107)(cid:88)
≥ δ(cid:107)(cid:88)
akxk(cid:107) − (cid:107)(cid:88)
(cid:1)(cid:107) ≥ δ(cid:107)(cid:88)
k
k
k
k akyk(cid:107) as follows:
(cid:107)(cid:88)
We can estimate (cid:107)(cid:80)
akyk(cid:107) ≥ (cid:107)(cid:88)
(cid:107)T(cid:0)(cid:88)
So,
k
k
akxk
k
k
akyk(cid:107) − (cid:107)T(cid:107)θ ≥ δ(1 − θ) − (cid:107)T(cid:107)θ ≥ δ
2
with the final inequality resulting in our choice of θ. We have thus seen that any norm 1
vector x ∈ [xk] has (cid:107)T x(cid:107) ≥ δ
2 > 0. Thus T is an isomorphism on [xk] as claimed.
There are two further observations about strictly singular operators that are important
in relation to the work of this thesis. The first is that the strictly singular operators on
a Banach space X form a closed ideal of the operator algebra L(X). The second is that
the the essential spectrum of a strictly singular operator on a complex Banach space is just
{0}. (We will actually give a more formal statement of this fact that also holds for real
Banach spaces.) In the remainder of this section, we provide proofs of these facts. We find
it convenient to introduce another class of operators.
18
Definition 1.4.20. Let T : X → X be a bounded linear operator on a Banach space X.
We say T is upper semi-Fredholm if the kernel of T is finite dimensional and T has closed
range.
Remark 1.4.21.
1. We recall that if T : X → X has finite dimensional kernel and finite co-dimensional
image, then T is said to be a Fredholm operator. The range of T is automatically closed
in this case, thus all Fredholm operators are, in particular, upper semi-Fredholm.
Moreover, for Fredholm operators, we can define the index of T by
ind T := dim Ker T − codim imT.
It is well known (see, e.g. [26]) that ind : F → Z is a continuous map (here F ⊆ L(X)
denotes the subset of Fredholm operators). For upper semi-Fredholm operators T , we
can define the generalised index by the same formula as above; this will give a finite
integer if T is in fact Fredholm. Otherwise it must be the case that the image of T
has infinite co-dimension and the generalised index is set equal to −∞.
2. Suppose T : X → X is upper semi-Fredholm. Since finite dimensional subspaces are
always complemented, we can write X as the topological direct sum X = Ker T ⊕ X(cid:48)
for some closed, finite co-dimensional subspace X(cid:48) ⊂ X. Since also Im T = T (X(cid:48))
is closed, it is easy to see that T : X(cid:48) → X is an isomorphic embedding. In other
words, we have seen that an upper semi-Fredholm operator is an isomorphism on
some finite co-dimensional, closed subspace. Conversely, if T is an isomorphism on
some closed, finite co-dimensional subspace of X, then T is upper semi-Fredholm.
Indeed, Let X1 be a finite co-dimensional subspace on which T is an isomorphism. It
follows that Ker T ∩ X1 = {0} and so we can write X as the (algebraic) direct sum
X = X1 ⊕ Ker T ⊕ F for some (necessarily finite dimensional) subspace F . Since X1
is finite co-dimensional, dim Ker T ≤ dim (Ker T ⊕ F ) = codim X1 < ∞. The image
of T is closed since T (X) = T (X1)⊕ T (F ) and the sum is easily seen to be topological
(T (X1) is closed since T is an isomorphism on X1 and T (F ) is finite dimensional).
We have therefore proved that an operator is upper semi-Fredholm if and only if it is
an isomorphism on some finite co-dimensional subspace. In some literature (e.g. [16],
[22]) one defines an operator to be finitely singular if it is an isomorphism on some
finite co-dimensional subspace. We have therefore proved that the class of finitely
singular operators is exactly the same as class of upper semi-Fredholm operators.
19
3. We also note that every strictly singular operator on an infinite dimensional Banach
space fails to be upper semi-Fredholm.
Later we will require the following proposition which extends the well known result that
the index of a Fredholm operator defines a continuous map into Z, to the case where the
operators are upper semi-Fredholm.
Proposition 1.4.22. If T : X → Y is an upper semi-Fredholm operator, there exists a
neighbourhood U of 0 ∈ L(X, Y ) such that T + U is upper semi-Fredholm for all U ∈ U and
moreover, the (generalised) index of T + U is equal to the (generalised) index of T .
Proof. When T is a Fredholm operator, the result follows from the well known result that the
index map, defined on Fredholm operators, is continuous. In the case where T is upper semi-
Fredholm but not Fredholm, the result follows easily from the following proposition.
(We remark that the following proposition can be found in [24, Proposition 3.1].)
Proposition 1.4.23. Let T : X → Y be an isomorphic embedding. Then there exists a
neighbourhood of 0 ∈ L(X, Y ) such that for every S is this neighbourhood, T + S is an
isomorphic embedding and codim(T + S)X = codim T X (finite or +∞).
To prove the proposition, we need:
Lemma 1.4.24. Let T ∈ L(X, Y ) be an isomorphic embedding and k ≥ 0 an integer.
1. If codim T X ≥ k, there exists c > 0 such that codim(T + S)X ≥ k whenever (cid:107)S(cid:107) < c.
2. If codim T X = k, there exists c > 0 such that codim(T + S)X = k whenever (cid:107)S(cid:107) < c.
Proof. Since T is an isomorphic embedding, there exists a δ > 0 such that (cid:107)T x(cid:107) ≥ δ(cid:107)x(cid:107) for
all x ∈ X.
Now, if codim T X ≥ k, there exists a subspace F ⊂ Y such that dim F = k and
T X ∩ F = {0}. Let πF : Y → Y /F denote the usual quotient map. We claim that πF ◦ T is
an isomorphic embedding from X into Y /F . Indeed, if not there exists a sequence (xn)∞
of vectors in X with (cid:107)xn(cid:107) = 1 for all n and (cid:107)T xn + F(cid:107) → 0. It follows that there exists a
n=1 in F such that (cid:107)T xn − fn(cid:107) → 0. It is clear that the sequence
sequence of vectors (fn)∞
(fn)∞
n=1 is bounded and since F is finite dimensional, an easy compactness argument yields
that (without loss of generality, after passing to a subsequence and relabelling if necessary)
fn → f ∈ F . It follows that T xn → f ∈ F . Since T is an isomorphic embedding, T (X) is
closed, so limn T xn = f ∈ T X ∩ F = {0}. On the other hand, (cid:107)f(cid:107) = limn (cid:107)T xn(cid:107) ≥ δ. So
πF ◦ T is an isomorphic embedding as claimed.
n=1
20
It is easy to see that the property of being an isomorphic embedding is closed under
small norm perturbation. Since we know πF ◦ T is an isomorphic embedding, it follows that
if we take c > 0 sufficiently small, πF ◦ (T + S) is also an isomorphic embedding if (cid:107)S(cid:107) < c.
This implies that F ∩ (T + S)X = {0}, hence codim(T + S)X ≥ k, proving the first claim.
The proof of the second case is similar; in this case we can select the subspace F such
that Y = T X ⊕ F and dim F = k. It follows (using the same argument as before) that
πF ◦ T is an onto isomorphism. Consequently, there exists c > 0 such that if (cid:107)S(cid:107) < c
then πF ◦ (T + S) is an onto isomorphism. In particular, F ∩ (T + S)X = {0}. Moreover,
since πF ◦ (T + S) is an onto isomorphism, for every y ∈ Y there exists x ∈ X such that
πF (y) = πF ((T + S)x). Consequently, y − (T + S)x ∈ F , showing that Y = F + (T + S)X.
So Y = F ⊕ (T + S)X and codim(T + S)X = k as required.
Proof of Proposition 1.4.23. Let c > 0 be such that T + S is an isomorphic embedding
whenever
S ∈ Bc := {U ∈ L(X, Y ) : (cid:107)U(cid:107) < c}
For each positive integer k ≥ 0, the set Dk = {U ∈ Bc : codim(T + U )X = k} is open in Bc
by the previous lemma. Moreover,
Dk = {U ∈ Bc : codim(T + U )X ≥ k + 1}C(cid:92)∩k−1
j=0{U ∈ Bc : codim(T + U )X = j}C,
thus Dk is closed in Bc as an intersection of closed sets (each set in the intersection is a
complement of an open set by the previous lemma). Since Bc is connected, each Dk is either
empty or equal to Bc. The result now follows in the case where codim T X = k for some
positive integer k ≥ 0. In the case where codim T X = +∞, we note that by the preceding
argument, Dk has to be empty for every k ≥ 0. Consequently, codim(T + U )X = +∞ for
every U ∈ Bc and the result is again proved.
Lemma 1.4.25. Let X be a Banach space (real or complex) and suppose T : X → X is a
bounded linear operator which fails to be upper semi-Fredholm. Then, for all ε > 0, there
exists an infinite dimensional (closed) subspace Yε ⊆ X with (cid:107)TYε(cid:107) < ε. The restriction of
T to Yε is a compact operator.
Proof. The proof given here follows closely that given in [22]. By Remark 1.4.21, we note
that T is not an isomorphism on any finite co-dimensional subspace of X.
We begin the proof by fixing a sequence of real numbers an > 0 such that(cid:81)∞
2. We will construct inductively a normalised, basic sequence (xn)∞
properties
n=1(1+an) ≤
n=1 with the following
21
1. If m < n are natural numbers, (λi)n
i=1 are scalars, then
(cid:107) m(cid:88)
λixi(cid:107) ≤
(cid:32) n(cid:89)
(cid:33)
(cid:107) n(cid:88)
(1 + ai)
λixi(cid:107)
i=1
i=m+1
i=1
2. (cid:107)T xn(cid:107) ≤ ε/2n+2 for all n.
at most(cid:81)∞
i=1(1 + ai) ≤ 2.
Note that as a consequence of the first property, (xn) is a basic sequence with basis constant
To begin the induction, since in particular T is not an isomorphism, we can choose
x1 ∈ X with (cid:107)x1(cid:107) = 1 and (cid:107)T x1(cid:107) ≤ ε/8.
Inductively, suppose we have constructed x1, . . . xn ∈ X with (cid:107)xi(cid:107) = 1 and satisfying
the two properties above. It is enough to show that we can find xn+1 ∈ X with (cid:107)xn+1(cid:107) = 1,
(cid:107)T xn+1(cid:107) ≤ ε/2n+3 and having the property that whenever y ∈ E := span{x1, . . . xn}, λ is
a scalar, (1 + an+1)(cid:107)y + λxn+1(cid:107) ≥ (cid:107)y(cid:107).
To this end, let y1, y2, . . . ym be a (finite) δ-net for the unit sphere SE of the finite
dimensional subspace E, where 0 < δ < 1 is such that 1 + an+1 ≥ 1/(1 − δ). By the Hahn-
Banach Theorem, we can choose norm 1 linear functionals y∗
i (yi) = 1
for all i = 1, . . . m. By hypothesis of the lemma, T is not an isomorphism when restricted
to the finite co -- dimensional subspace Z := ∩m
i , so we may choose a norm 1 vector,
xn+1 ∈ Z with (cid:107)T xn+1(cid:107) ≤ ε/2n+3. If y ∈ SE, λ a scalar, there exists some i ∈ {1, . . . m}
with (cid:107)y − yi(cid:107) ≤ δ, and
m ∈ SX∗ with y∗
i=1 Ker y∗
1, . . . y∗
(1 + an+1)(cid:107)y + λxn+1(cid:107) = (1 + an+1)(cid:107)yi + λxn+1 + y − yi(cid:107)
i (yi + λxn+1 + y − yi)
≥ (1 + an+1) y∗
≥ (1 + an+1) (1 − y∗
≥ (1 + an+1) (1 − (cid:107)y − yi(cid:107))
≥ (1 + an+1)(1 − δ) ≥ 1
i (y − yi))
We set Yε := [xn]∞
from which we easily conclude that (cid:107)y + λxn+1(cid:107) ≥ (cid:107)y(cid:107) for any y ∈ E and scalar λ as
required.
n=1, an infinite dimensional subspace of X with Schauder basis (xn)∞
If y = (cid:80)∞
(cid:107)(cid:80)∞
n=1 αnT xn(cid:107) ≤ 4ε(cid:80)∞
TN : Yε → X, defined by(cid:80)∞
n=1.
n=1 αnxn ∈ Yε with (cid:107)y(cid:107) = 1, it follows that αn ≤ 4 for all n and (cid:107)T y(cid:107) =
n=1 2−(n+2) = ε, giving the desired estimate for (cid:107)TYε(cid:107). It is easy
to see (by a similar estimation) that T is the uniform limit of the finite rank operators
n=1 αnT xn, and thus T : Yε → X is compact as
n=1 αnxn (cid:55)→(cid:80)N
claimed.
22
Corollary 1.4.26. An operator T : X → Y is strictly singular ⇐⇒ whenever Z ⊂ X is
an infinite dimensional subspace of X, there exists a further infinite dimensional subspace
Z(cid:48) ⊂ Z such that the restriction of T to Z(cid:48) is compact.
Proof. Suppose first T is strictly singular and let Z be an infinite dimensional subspace of
X. We consider the operator T1 : Z → X where T1 is simply the restriction of T to Z.
It is easily verified that T1 is strictly singular because T is, and therefore T1 fails to be
upper semi-Fredholm. By the previous lemma, we can find an infinite dimensional subspace
Z(cid:48) ⊂ Z such that the restriction of T1 to Z(cid:48) is compact. Obviously by definition of T1, this
implies that the restriction of T to Z(cid:48) is compact.
Conversely, suppose for contradiction that Z is an infinite dimensional subspace of X on
which T is an isomorphism. By the assumed property, there exists an infinite dimensional
subspace Z(cid:48) of Z on which T is both a compact operator and an isomorphism. However,
these two facts imply that Z(cid:48) has a compact unit ball, contradicting the fact that Z(cid:48) is infinite
It follows that T fails to be an isomorphism on every infinite dimensional
dimensional.
subspace, i.e. T is strictly singular.
Corollary 1.4.27. The set of strictly singular operators on a Banach space X forms a
closed ideal of L(X).
Proof. Suppose first that T, S : X → X are strictly singular operators on X. We will
use Corollary 1.4.26 to show T + S is strictly singular. To this end, let Z be an infinite
dimensional subspace of X. Since T is strictly singular, there exists an infinite dimensional
subspace Z(cid:48) ⊂ Z on which T is compact. We now apply Corollary 1.4.26 to S and the
subspace Z(cid:48), obtaining an infinite dimensional subspace Z(cid:48)(cid:48) ⊂ Z(cid:48) on which both S and T
are compact. It follows that T + S is compact when restricted to Z(cid:48)(cid:48), and consequently
T + S is strictly singular by the corollary.
To prove that the strictly singular operators are an ideal, it remains to prove that if P
is strictly singular and R, Q ∈ L(X) then RP Q is a strictly singular operator. Suppose Z
is an infinite dimensional subspace of X and fix ε > 0. We will exhibit a norm 1 vector,
z ∈ Z with (cid:107)RP Qz(cid:107) < ε. Note that if Q(Z) is finite dimensional, then since Z is infinite
dimensional, we must have ker Q ∩ Z (cid:54)= {0}. In this case, we just take z to be any norm 1
vector in ker Q∩ Z and the proof is trivial. So we assume that Q(Z) is infinite dimensional.
Since P is assumed strictly singular, P is not an isomorphism on the infinite dimensional
subspace Q(Z). Consequently, we can find a norm 1 vector x = Q(z0) (for some z0 ∈ Z)
with (cid:107)P x(cid:107) = (cid:107)P Qz0(cid:107) < ε/(cid:107)Q(cid:107)(cid:107)R(cid:107). We take z = z0/(cid:107)z0(cid:107) noting that 1/(cid:107)z0(cid:107) ≤ (cid:107)Q(cid:107) since
1 = (cid:107)x(cid:107) = (cid:107)Qz0(cid:107), so that (cid:107)P Qz(cid:107) < (cid:107)Q(cid:107)ε/(cid:107)Q(cid:107)(cid:107)R(cid:107). Therefore, (cid:107)RP Qz(cid:107) ≤ (cid:107)R(cid:107)(cid:107)P Qz(cid:107) < ε
as required.
23
Finally, note if Tn is a sequence of strictly singular operators on X with (cid:107)Tn − T(cid:107) → 0
then T is strictly singular, so that the ideal of strictly singular operators is closed. Indeed, if
Z is an infinite dimensional subspace of X, ε > 0, then we find N such that (cid:107)TN −T(cid:107) < ε/2.
Since TN is strictly singular, there exists z ∈ Z with (cid:107)z(cid:107) = 1 and (cid:107)TN z(cid:107) < ε/2. It follows
that (cid:107)T z(cid:107) ≤ (cid:107)TN z(cid:107) + (cid:107)(T − TN )z(cid:107) < ε.
We conclude this section by stating and proving a result about strictly singular operators
that we will need in Chapter 4.
Theorem 1.4.28. Let X be a Banach space (real or complex). If T ∈ L(X) is strictly
singular then r(T ) := limn→∞ (cid:107)T n + K(X)(cid:107) 1
n = 0.
In the case where X is a complex Banach space, the above result just says that the
essential spectrum of a strictly singular operator on a complex Banach space is {0}. To
prove the result in the real case, we will make use of a complexification argument and will
consequently need the following technical lemma; it is essentially a generalisation of Lemma
1.4.25, with the only real difficulty being to make sure we extract a real subspace of the
complexified space.
Lemma 1.4.29. Let T : X → X be a bounded linear operator on a real Banach space X,
and denote by XC = X ⊕ iX the complexification of X, TC the complex extension of T to
XC (as in Section 1.4.5). Suppose there exists λ ∈ C such that TC − λIdXC fails to be upper
semi-Fredholm. Then the operator Tλ on X defined by Tλ = T 2 − 2ReλT + λ2IdX has the
property that for all ε > 0, there exists a real infinite dimensional Yε ⊆ X with (cid:107)TλYε(cid:107) < ε.
Proof. The proof is very similar to that of Lemma 1.4.25 and we continue the notation
n=1 ⊆ X with
from that proof. We will construct inductively a normalised sequence (xn)∞
the following properties:
1. If m < n are natural numbers, (λi)n
i=1 are real scalars, then
(cid:107) m(cid:88)
λixi(cid:107) ≤
(cid:32) n(cid:89)
(cid:33)
(cid:107) n(cid:88)
(1 + ai)
λixi(cid:107)
i=1
i=m+1
i=1
2. (cid:107)Tλxn(cid:107) ≤ ε/2n+2 for all n.
a real basic sequence with basis constant at most (cid:81)∞
As in the proof of Lemma 1.4.25 , it is a consequence of the first property that (xn) is
i=1(1 + ai) ≤ 2. Given such a basic
n=1 ⊆ X to complete the proof (which is then identical to the
sequence, we take Yε := [xn]∞
proof of the previous lemma).
24
Observe that whenever x, y ∈ X, (cid:0)TC − λIdXC
(cid:1) (TC − λIdXC) (x + iy) = Tλx + iTλy =
(Tλ)C(x + iy). This, combined with the fact that TC − λIdC is not an isomorphism on
any finite co-dimensional subspace of XC (see Remark 1.4.21), implies that (Tλ)C is not
an isomorphism on any finite co-dimensional subspace of XC. In particular, we can choose
z1 ∈ XC with (cid:107)z1(cid:107) = 1 and (cid:107)(Tλ)C z1(cid:107) ≤ ε/16. Since (cid:107)z1(cid:107) = 1, writing z1 uniquely as
z1 = w1 + iy1 with w1, y1 ∈ X, we must have either (cid:107)w1(cid:107) ≥ 1
2 , in which case we define
x1 := w1/(cid:107)w1(cid:107), or (cid:107)y1(cid:107) ≥ 1
2 (in which case we define x1 = y1/(cid:107)y1(cid:107)).
In either case,
(cid:107)Tλx1(cid:107) ≤ ε/8. Indeed, considering the case where x1 = w1/(cid:107)w1(cid:107), we have
(cid:107)w1(cid:107)(cid:107)(Tλ)C z1(cid:107) ≤ ε
(cid:107)w1(cid:107)(cid:107)Tλw1 + iTλy1(cid:107) =
(cid:107)Tλx1(cid:107) ≤ 1
1
8
.
(The first inequality above follows from Lemma 1.4.16, and the final inequality from the
fact that (cid:107)w1(cid:107) ≥ 1
2 .) The other case is dealt with similarly.
Inductively, suppose we have constructed x1, . . . xn ∈ X with (cid:107)xi(cid:107) = 1 and satisfying
the two properties above. As in the proof of Lemma 1.4.25, it is enough to show that we
can find xn+1 ∈ X with (cid:107)xn+1(cid:107) = 1, (cid:107)Tλxn+1(cid:107) ≤ ε/2n+3 and having the property that
whenever y ∈ E := spanR{x1, . . . xn}, λ ∈ R, (1 + an+1)(cid:107)y + λxn+1(cid:107) ≥ (cid:107)y(cid:107).
1, . . . y∗
1)C, . . . (y∗
To this end, let y1, y2, . . . ym be a (finite) δ-net for the unit sphere SE of the finite
dimensional subspace E, where 0 < δ < 1 is such that 1 + an+1 ≥ 1/(1 − δ). By the
m ∈ SX∗ with
Hahn-Banach Theorem, we can choose norm 1 linear functionals y∗
y∗
i (yi) = 1 for all i = 1, . . . m. Then (y∗
m)C are norm 1 linear functionals on
XC with (y∗
i )C(yi) = 1 for all i. By our earlier observation, (Tλ)C is not an isomorphism
when restricted to the finite co -- dimensional subspace Z := ∩m
i )C ⊆ XC. It follows
that we can choose a norm 1 vector, zn+1 ∈ Z with (cid:107)(Tλ)C zn+1(cid:107) ≤ ε/2n+4. Writing
zn+1 = wn+1 + iyn+1 with wn+1, yn+1 ∈ X, we must have (for the same reason as earlier)
either (cid:107)wn+1(cid:107) ≥ 1
2 . As before, we set xn+1 to be either wn+1/(cid:107)wn+1(cid:107) or
yn+1/(cid:107)yn+1(cid:107) (depending on which of wn+1, yn+1 has norm at least 1
2 ). The same argument
as before yields (cid:107)Tλxn+1(cid:107) ≤ ε/2n+3. Note also that for j ∈ {1, . . . m}, 0 = (y∗
j )C zn+1 =
j (wn+1) + iy∗
y∗
j (xn+1) = 0 for
all j. Finally, if y ∈ SE, λ ∈ R there is some i ∈ {1, . . . m} with (cid:107)y − yi(cid:107) ≤ δ, and
j (yn+1). It follows that y∗
j (yn+1) and thus y∗
j (wn+1) = 0 = y∗
2 or (cid:107)yn+1(cid:107) ≥ 1
i=1 Ker(y∗
(1 + an+1)(cid:107)y + λxn+1(cid:107) = (1 + an+1)(cid:107)yi + λxn+1 + y − yi(cid:107)
i (yi + λxn+1 + y − yi)
≥ (1 + an+1) y∗
≥ (1 + an+1) (1 − y∗
≥ (1 + an+1) (1 − (cid:107)y − yi(cid:107))
≥ (1 + an+1)(1 − δ) ≥ 1
i (y − yi))
25
from which we easily conclude that (cid:107)y + λxn+1(cid:107) ≥ (cid:107)y(cid:107) for any y ∈ E, λ ∈ R as required.
n = limn→∞ (cid:107)T n + K(X)(cid:107) 1
Proof of Theorem 1.4.28. Note that the limit certainly exists (in fact, it is a classical result
that inf n (cid:107)T n + K(X)(cid:107) 1
is a complex Banach space and argue by contradiction, assuming T is a strictly singular
operator with r(T ) > 0. Note it is enough to show that there exists some non-zero λ ∈ C,
0 < ε < λ and an infinite dimensional subspace Yε ⊆ X with (cid:107)T y − λy(cid:107) ≤ ε(cid:107)y(cid:107) whenever
y ∈ Yε, since this contradicts the strict singularity of T . The argument is identical to that
used by Gowers and Maurey to show that every operator on a complex HI space is a strictly
n ). We consider first the case where X
singular perturbation of the identity. For completeness we include it here.
Denote by ΩT the set of all µ ∈ C such that T − µI is upper semi-Fredholm; we recall
that the generalised index of such an operator is equal to some finite integer or −∞.
We show that if r(T ) > 0, there is a non-zero λ ∈ C \ ΩT . It is a well known theorem
of Atkinson that an operator U ∈ L(X) is Fredholm if and only if the class [U ] of U in the
Calkin algebra is invertible. It follows that the essential spectrum of T, σess(T ) := {λ ∈ C :
T −λI is not Fredholm} is precisely the spectrum, σ([T ]), of the element [T ] ∈ L(X)/K(X).
It is well known that σ([T ]) (cid:54)= ∅ (since we are working over complex scalars) and moreover,
it follows from well known results about the spectral radius that there exists λ ∈ σ([T ])
with λ = maxµ∈σ([T ]) µ = r(T ) > 0. We claim that λ /∈ ΩT . Note that if µ ∈ C with
µ > λ then µ /∈ σ([T ]) and consequently T − µI is Fredholm. Moreover, by continuity of
the index, T − µI is Fredholm with index 0 for all µ > λ (because this is true when µ is
large enough to make T − µI invertible, i.e. when µ > (cid:107)T(cid:107)).
Now if λ ∈ ΩT , we know that T − λI must be upper semi-Fredholm with generalised
index −∞ (since T − λI is not Fredholm by choice of λ). It follows from Proposition 1.4.22
that the generalised index of (T − µI) = −∞ for all µ close to λ, but this contradicts the
fact that T − µI is Fredholm with index 0 for all µ > λ. Thus 0 (cid:54)= λ ∈ C\ ΩT as required.
Since λ /∈ ΩT , it follows by Lemma 1.4.25 that, given ε > 0, we can find an infinite
dimensional subspace Yε of X with (cid:107)T y − λy(cid:107) ≤ ε(cid:107)y(cid:107) for every y ∈ Yε. Taking any ε < λ
completes the proof of the lemma in the complex case.
To prove the real case, we use a complexification argument that appeared in [16], though
is attributed to Haydon (see [16, Lemma 20]). As in Section 1.4.5, we let XC = X ⊕ iX
denote the complexification of X with the norm (cid:107)x + iy(cid:107)XC := supt∈[0,2π] (cid:107)x cos t− y sin t(cid:107)X .
We recall that when L ∈ L(X), there is a unique complex linear extension of L, LC ∈ L(XC),
defined by LC(x+iy) = Lx+iLy, Moreover, (cid:107)LC(cid:107) = (cid:107)L(cid:107) and if K ∈ K(X), then the complex
linear extension KC ∈ K(XC). We refer the reader back to Lemma 1.4.17 for the details.
From these observations, it easily follows that (cid:107)L + K(X)(cid:107) = (cid:107)LC + K(XC)(cid:107).
26
Now assume as before that T is a strictly singular operator with r(T ) > 0. It follows by
n = limn→∞ (cid:107)T n + K(X)(cid:107) 1
the preceding remarks that limn→∞ (cid:107)(TC)n + K(XC)(cid:107) 1
n > 0. So
by the proof just given for the complex case, there exists a non-zero λ ∈ C such that the
operator TC−λIXC : XC → XC fails to be upper semi-Fredholm. It follows by Lemma 1.4.29
that for all ε > 0, there is an infinite dimensional subspace Yε ⊂ X such that (cid:107)TλYε(cid:107) < ε
where we recall that Tλ is the operator T 2 − 2ReλT + λ2I ∈ L(X). If we choose ε < λ2,
then it follows that T 2 − 2ReλT is an isomorphism on Yε. But, if T is strictly singular then
so is T 2 − 2ReλT by Corollary 1.4.27, so we once again have a contradiction.
27
Chapter 2
The Bourgain-Delbaen
Construction
Introduction
2.1
In 1980, Bourgain and Delbaen [7, 6] introduced two classes of separable L∞-spaces, X and
Y. We shall look at these classes in detail and give a precise definition of an L∞ space a
little later in this chapter. For now, we simply remark that these spaces were shown to
have a number of interesting properties, providing counterexamples for many previously
unanswered conjectures in Banach space theory. For instance, none of these spaces have a
subspace isomorphic to c0, giving the first set of examples of L∞ spaces with no c0 isomorph.
In fact, it was shown in [7] that if Y ∈ Y, then every infinite dimensional subspace of Y
contains a further infinite dimensional subspace which is reflexive.
More recently, Argyros and Haydon [2] have managed to modify the original construc-
tion to exhibit a space which solves the scalar-plus-compact problem. Although it is a
question remaining from Argryos' and Haydon's solution to this problem that really mo-
tivates our interest in the Bourgain-Delbaen spaces, we concern ourselves in this chapter
with investigating the original Bourgain-Delbaen construction and the newer generalisa-
tion used by Arygros and Haydon. We are a little relaxed with our terminology, referring
to a Banach space constructed using either of the methods just mentioned as a space of
Bourgain-Delbaen type. As is remarked in [2], it is interesting (and worth emphasising) that
Bourgain-Delbaen constructions are different from the majority of other Banach space con-
structions that occur in the literature. It is usual to start with the vector space of finitely
supported scalar sequences, c00, and complete with respect to some exotic norm. We shall
shortly see that spaces of Bourgain-Delbaen type are 'exotic subspaces' of (cid:96)∞, i.e. the norm
is just the usual (cid:107) · (cid:107)∞ norm.
We shall begin the chapter by looking at the Argyros-Haydon generalisation in Section
2.2. Here spaces of Bourgain-Delbaen type are subspaces of (cid:96)∞, obtained as the closed linear
28
span of the biorthogonal functionals of a special kind of Schauder basis for (cid:96)1. We will show
that we can make a further, minor modification to the Argyros-Haydon construction. The
new generalisation that we obtain will be an essential tool in proving the results of later
chapters and as such, we choose to look at it in detail.
Following this, we briefly recall the original Bourgain-Delbaen construction in Section
2.3, though we present it via the notation used by Haydon in [17]. This really is just
notational convenience; it is clear that if one enumerates the set Γ (defined in Section
2.3) in a natural way, we get an obvious isometry from the subspace of (cid:96)∞(Γ) appearing
in Section 2.3, and the subspace of (cid:96)∞(N) defined in the original work of Bourgain and
Delbaen in [7].
Having looked in detail at the two different constructions, we will then, in some sense,
unify them in Section 2.4 by showing that the Argyros-Haydon construction is a genuine
generalisation of that due to Bourgain and Delbaen. Precisely, we will show the original
Bourgain-Delbaen construction can be obtained as a special case of the Argyros-Haydon
construction. Of course, this is not a new result. However, it is, to the author's best
knowledge, the only documented exposition describing in detail the relationship between
the two constructions.
One of the advantages of the Arygros-Haydon generalisation is that it gives us a potential
way to construct interesting operators on a space of Bourgain-Delbaen type, B. The idea
is to define an operator on (cid:96)1, take the dual operator, and restrict to B. Of course, some
care needs to be taken to ensure that the restriction also maps into B.
Having shown how to describe the original spaces of Bourgain and Delbaen in the new
framework of the Argyros-Haydon construction, we use the preceding idea to conclude the
chapter (Section 2.5) by providing a partial answer to a question of Beanland and Mitchell,
[4], namely if Y ∈ Y, is L(Y ) separable? Precisely, we will show that for any X ∈ X ∪ Y,
there exists a constant CX > 0 and an uncountable collection of isometries on X which are
pairwise distance CX apart with respect to the operator norm. Clearly this shows L(X) is
non-separable for any X ∈ X ∪Y. We will note in Section 2.5 that there is in fact a shorter
argument to see that L(X) is non-separable for X ∈ X and also defer to this section a short
discussion of why the author finds this problem interesting.
2.2 The generalised Bougain-Delbaen Construction
Before continuing any further, we recall that what we want to be able to do is construct
interesting L∞ spaces. After all, it was seen in [2] that the finite-dimensional subspace
structure possessed by such spaces is what proves to be so fundamental in showing that the
Argyros-Haydon space solves the scalar-plus-compact problem. Moreover, we will exploit
29
similar techniques to prove results about the operator algebras of the new spaces we con-
struct in subsequent chapters of this thesis. It is about time we define precisely what we
mean by an L∞ space.
sequence (Fn)n∈N of finite dimensional subspaces of X such that the union(cid:83)
Definition 2.2.1. A separable Banach space X is an L∞,λ-space if there is an increasing
n∈N Fn is dense
. We say a Banach space is an L∞ space
in X and, for each n, Fn is λ-isomorphic to (cid:96)dim Fn∞
if it is a L∞,λ space for some λ.
The constructions of Bourgain and Delbaen, and the generalisation due to Argyros and
Haydon, provide a way of constructing interesting classes of L∞ spaces. Consequently,
we begin by looking in detail at these constructions, starting with the Argyros-Haydon
generalisation. The content of this section follows very closely the work in [2] and we make
no claim to originality unless explicitly stated.
As remarked in the introduction to this chapter, the idea of the Argyros-Haydon con-
struction is to construct a particular kind of Schauder basis for the space (cid:96)1 and to study
the subspace X of (cid:96)∞ spanned by the biorthogonal elements. It follows by Proposition 1.4.3
that (cid:96)1 naturally embeds into X∗. (Moreover, by Theorem 1.4.6, this embedding is onto
X∗ precisely when the Schauder basis of (cid:96)1 is boundedly complete.) Consequently, we will
think of elements of (cid:96)1 as functionals and, as in [2], denote them using a star notation, b∗,
c∗ and so on. In accordance with this notation, we denote the canonical basis of (cid:96)1(N) by
(e∗
n)∞
n=1.
It is perhaps easiest to understand the construction by working first with (cid:96)1(N) and this
is indeed what is done in [2]. The special kind of Schauder basis of (cid:96)1(N) is obtained by
considering a sequence (d∗
1 = 0
n ⊆ {1, 2, . . . , (n − 1)} for n ≥ 2. An easy induction argument yields that the
and supp c∗
n=1 are linearly independent and that moreover, the linear span [d∗
2, . . . , d∗
n)∞
vectors (d∗
n]
is the same as [e∗
n=1 ⊆ (cid:96)1(N) where each d∗
n)∞
n has the form d∗
n with c∗
n], so the closed linear span, [d∗
The clever part of the construction is to choose the c∗
2, . . . , e∗
1, e∗
n form
a Schauder basis for (cid:96)1. In fact, things are a little more subtle than this, in the sense that
the sequence (d∗
n)∞
n=1 will actually be a basic sequence equivalent to the canonical basis of
(cid:96)1 if the c∗
n have sufficiently small norm. This case certainly won't be useful in producing
any interesting Banach spaces, though we defer further discussion of this to Lemma 2.2.4
n = e∗
n−c∗
1, d∗
n : n ∈ N] is the whole of (cid:96)1.
n in such a way that the d∗
so as not to detract from the point at hand.
Argyros and Haydon showed that the d∗
n will always form a Schauder basis for (cid:96)1 if
the c∗
n assume a certain form (we refer the reader to [2] for more details). For this reason,
it turns out that it is more convenient to work with the space (cid:96)1(Γ) for some suitably
30
defined countable set Γ; the elements γ ∈ Γ can then be used to code the form of the
γ ∈ (cid:96)1(Γ). We choose to immediately state (and prove) the main result of this
vector c∗
section, working with this more convenient notation, and refer the reader to [2] for a more
comprehensive introduction.
(cid:83)
1≤p≤q ∆p, Γ = (cid:83)
Theorem 2.2.2. Let (∆q)q∈N be a disjoint sequence of non-empty finite sets; write Γq =
2 and a mapping τ defined on
p∈N ∆p. Assume that there exists θ < 1
Γ \ ∆1, assigning to each γ ∈ ∆q+1 a tuple of one of the following forms:
0. (ε, α, ξ) with ε = ±1, 0 ≤ α ≤ 1 and ξ ∈ Γq;
1. (p, β, b∗) with 0 ≤ p < q, 0 < β ≤ θ and b∗ ∈ ball (cid:96)1 (Γq \ Γp);
2. (ε, α, ξ, p, β, b∗) with ε = ±1, 0 < α ≤ 1, 1 ≤ p < q, ξ ∈ Γp, 0 < β ≤ θ and
γ − c∗
γ ∈ (cid:96)1(Γ) and projections P ∗
(0,q] on (cid:96)1(Γ) uniquely determined
b∗ ∈ ball (cid:96)1 (Γq \ Γp).
γ = e∗
Then there exist d∗
by the following properties:
(A) P ∗
(0,q]d∗
γ =
(B)
c∗
γ =
d∗
0
γ
(cid:40)
if γ ∈ Γq
if γ ∈ Γ \ Γq
if γ ∈ ∆1
if τ (γ) = (ε, α, ξ)
if τ (γ) = (p, β, b∗)
if τ (γ) = (ε, α, ξ, p, β, b∗).
ξ
0
εαe∗
β(I − P ∗
εαe∗
(0,p])b∗
ξ + β(I − P ∗
(0,p])b∗
γ)γ∈Γ is a basis for (cid:96)1(Γ) with basis constant at most M = (1 − 2θ)−1. The
The family (d∗
norm of each projection P ∗
(0,q] is at most M . The biorthogonal vectors dγ generate a L∞,M -
subspace X(Γ, τ ) of (cid:96)∞(Γ); if the basis (dγ)γ∈Γ of X(Γ, τ ) is shrinking, then X∗ is naturally
isomorphic to (cid:96)1(Γ). For each q and each u ∈ (cid:96)∞(Γq), there is a unique vector in [dγ :
γ ∈ Γq] whose restriction to Γq is u; therefore there exists a unique extension operator
iq : (cid:96)∞(Γq) → X(Γ, τ )∩ [dγ : γ ∈ Γq] and this operator has norm at most M . The subspaces
Mq = [dγ : γ ∈ ∆q] = iq[(cid:96)∞(∆q)] form a finite-dimensional decomposition (FDD) for X.
Remark. Strictly speaking, a Schauder basis is, in particular, a sequence indexed by the
natural numbers, so there is a natural ordering of the vectors in the sequence. Therefore
when we talk of (d∗
γ)γ∈Γ being a basis of (cid:96)1(Γ), or (dγ)γ∈Γ being a basis of X(Γ, τ ), we
really need to enumerate Γ in some way. The enumeration we have in mind will always be
the one that is described in the proof of this theorem, and we will henceforth speak of (for
example) (dγ) being a basis without mentioning the explicit enumeration.
31
We observe that the statement here is almost identical to that of Theorem 3.5 in [2].
We have added an extra degree of freedom to tuples of form (0) and (2) appearing above
via the addition of the parameter ε which takes values in {±1}. This means, for example,
we can have vectors c∗ of the form −αe∗
(0,q])b∗. This change will only be useful
for when we come to examine the connection between this construction and the original
ξ + β(I − P ∗
Bourgain-Delbaen construction. In later chapters, we will always set ε = 1.
The more important change we have made is that we do not demand the set ∆1 have
only one element. This extra freedom will allow us to construct some non-trivial operators
on the spaces we construct in later chapters. Since this theorem is so essential to this thesis,
we include the proof, though all we are really doing is translating the arguments given in
[2] into the notation used throughout this thesis.
Proof of Theorem 2.2.2. The construction is recursive and to begin with, we work only
γ = e∗
γ for γ ∈ ∆1.
Recursively, assume that c∗
algebraically, not worrying about the continuity of the projection maps we define. We start
by setting d∗
γ and d∗
γ have been defined for all γ ∈ Γn in accordance
with the theorem. It follows from the recursive construction that the vectors (d∗
γ)γ∈Γn are
linearly independent and that their span is all of (cid:96)1(Γn). Consequently, the (restricted)
(0,q](cid:96)1(Γn) for q ≤ n are uniquely determined on (cid:96)1(Γn) by the formula given
projections P ∗
in (A). For γ ∈ Γn+1, we can therefore define vectors c∗
γ − c∗
γ as in (B) and set d∗
γ.
γ ⊆ ∆n+1, we see that the vectors
γ has support in Γn and γ ∈ ∆n+1 so that supp e∗
Since c∗
(d∗
γ(cid:48))γ(cid:48)∈Γn+1 are linearly independent and that their span is (cid:96)1(Γn+1). We can thus extend
the projections P ∗
(0,n+1]
on (cid:96)1(Γn+1) in the same way. This completes the recursive construction; we obtain vectors
(0,q] defined on c00(Γ) ⊆ (cid:96)1(Γ), satisfying
γ and d∗
c∗
the conclusions of the theorem.
(0,q] for q ≤ n to all of (cid:96)1(Γn+1) in accordance with (A) and define P ∗
γ for each γ ∈ Γ and projections P ∗
γ = e∗
γ = e∗
γ−c∗
We now show that (d∗
γ)γ∈Γ is a Schauder basis for (cid:96)1(Γ). In the process, we shall see
that the projections P ∗
(0,q] defined on c00(Γ) are (uniformly) bounded by M , and thus extend
to projections on all of (cid:96)1(Γ). Let us be a little more precise; for p ≥ 1 let kp = #Γp, set
k0 = 0 and let n (cid:55)→ γ(n) : N → Γ be a bijection with the property that for each q ≥ 1,
∆q = {γ(n) : kq−1 < n ≤ kq}. We will show (d∗
n=1 is a Schauder basis. It is clear from
the recursive construction that this sequence has dense linear span, so it is enough to see
that the (densely defined) basis projections, P ∗
γ(n))∞
(cid:40)
m defined by
if j ≤ m
otherwise
d∗
0
γ(j)
γ(j) (cid:55)→
d∗
32
are uniformly bounded. We show (cid:107)P ∗
immediately get (cid:107)P ∗
for all j ≤ m and 0 otherwise since span{d∗
obvious from the recursive construction.
(0,q] = P ∗
(0,q](cid:107) ≤ M as claimed in the theorem. We also note that P ∗
m(cid:107) ≤ M for all m. Since P ∗
kq
for all q, we
γ(j) = e∗
me∗
γ(j)
γ(j) : j ≤ m} = span{e∗
γ(j) : j ≤ m}; this is again
me∗
me∗
γ(j) = e∗
Note that for j ≤ k1, d∗
γ(j). Consequently, for j ≤ k1, P ∗
every m and n in N. We shall prove by induction on n that (cid:107)P ∗
Since we are working on the space (cid:96)1(Γ) it is enough to show that (cid:107)P ∗
γ(n)(cid:107) ≤ M for
γ(j)(cid:107) ≤ M for all m, j ≤ n.
γ(j) is either equal to
e∗
γ(j) or 0. In particular, there is nothing to prove for our inductive statement in the case
when 1 ≤ n ≤ k1. Suppose inductively that (cid:107)P ∗
γ(j)(cid:107) ≤ M holds for all m, j ≤ n (w.l.o.g.
n ≥ k1). We must see that this holds for all m, j ≤ n + 1. If j ≤ m, then, as observed in
the previous paragraph, P ∗
γ(j) so there is nothing to prove. We therefore assume
j > m. If, in addition, m < j ≤ n we are done by the inductive hypothesis. So we just need
to see that (cid:107)P ∗
γ(n+1)(cid:107) ≤ M for all m ≤ n. Since we assume n ≥ k1, there exists a q ≥ 1
with γ(n + 1) ∈ ∆q+1. Note that this implies
γ(j) = e∗
me∗
me∗
me∗
me∗
Γq ⊆ {γ(1), γ(2), . . . , γ(n)}.
(2.1)
We now make use of the fact that
γ(n+1) = d∗
e∗
γ(n+1) + c∗
γ(n+1),
where supp c∗
described in the theorem. We consider only the case where
γ(n+1) ⊆ Γq ⊆ {γ(1), γ(2), . . . , γ(n)} and moreover c∗
γ(n+1) has one of the forms
c∗
γ(n+1) = εαe∗
ξ + β(I − P ∗
(0,p])b∗,
where ε = ±1, 1 ≤ p < q, ξ ∈ Γp, b∗ ∈ ball (cid:96)1(Γq \ Γp) and α, β are as in the theorem, with
β ≤ θ by our hypothesis. The other cases are similar.
Now, because n + 1 > m we have P ∗
γ(n+1) = 0. We can also write ξ = γ(j) where
j ≤ kp < kq ≤ n (this follows from the fact that ξ ∈ Γp with p < q and Equation 2.1). As
observered earlier P ∗
. Combining these observations, we get
md∗
kp
(0,p] = P ∗
P ∗
me∗
γ(n+1) = εαP ∗
me∗
γ(j) + β(P ∗
m − P ∗
min{m,kp})b∗.
If kp ≥ m the second term vanishes so that
γ(n+1)(cid:107) = α(cid:107)P ∗
me∗
me∗
(cid:107)P ∗
γ(j)(cid:107) ≤ (cid:107)P ∗
me∗
γ(j)(cid:107),
which is at most M by our inductive hypothesis.
33
If, on the other hand, kp < m, we certainly have j < m (since recall j ≤ kp) so that
γ(j) = e∗
me∗
P ∗
γ(j), leading to the estimate
γ(n+1)(cid:107) ≤ α(cid:107)e∗
γ(j)(cid:107) + β(cid:107)P ∗
Now b∗ is a convex combination of functionals ±e∗
hypothesis is applicable to all of these. We thus obtain
me∗
(cid:107)P ∗
kpb∗(cid:107).
mb∗(cid:107) + β(cid:107)P ∗
γ(l) with l ≤ kq ≤ n, and our inductive
(cid:107)P ∗
me∗
γ(n+1)(cid:107) ≤ α + 2M β ≤ 1 + 2M θ = M,
by the definition of M = 1/(1 − 2θ) and the assumption that 0 ≤ β ≤ θ.
We conclude that (d∗
γ(n))∞
n=1 is a Schauder basis for (cid:96)1(Γ) as claimed. It follows from
Proposition 1.4.3 that the biorthogonal functionals (dγ)γ∈Γ form a Schauder basis for the
Banach space X := X(Γ, τ ) = [dγ : γ ∈ Γ] ⊆ (cid:96)∞(Γ), or more precisely that the sequence
(dγ(n))∞
n=1 is a Schauder basis for X. We easily deduce from this that the finite dimensional
subspaces Mq := [dγ : γ ∈ ∆q] form a finite dimensional decomposition of X. Moreover,
standard results in Banach space theory yield that X∗ is naturally isomorphic to (cid:96)1 in the
case that the basis (dγ)γ∈Γ is shrinking (see Theorem 1.4.6).
We must also show that X is a L∞,M space. To see this, we consider the projections
P ∗
(0,n] defined in the theorem. We have just shown these are well-defined and moreover
(0,n](cid:107) ≤ M . If we modify P ∗
(cid:107)P ∗
(0,n] = (cid:96)1(Γn),
rather than the whole of (cid:96)1(Γ), what we have is a quotient operator, which we shall denote
qn, of norm at most M . The dual of this quotient operator is an isomorphic embedding
in : (cid:96)∞(Γn) → (cid:96)∞(Γ), also of norm at most M . Of course, this is immediate by Lemma
1.4.11. However, we can also show it explicity; if u ∈ (cid:96)∞(Γn), γ ∈ Γn, we have
(0,n] by taking the codomain to be the image im P ∗
(inu)(γ) = (cid:104)inu, e∗
γ(cid:105) = (cid:104)u, e∗
So in is an extension operator (cid:96)∞(Γn) → (cid:96)∞(Γ) and we have
(cid:107)u(cid:107)∞ ≤ (cid:107)inu(cid:107)∞ ≤ M(cid:107)u(cid:107)∞
γ(cid:105) = (cid:104)u, qne∗
γ(cid:105) = u(γ).
for all u ∈ (cid:96)∞(Γn). We claim that the image of in is precisely [dγ : γ ∈ Γn] and that for
each γ ∈ ∆n, ineγ = dγ so that in((cid:96)∞(∆n)) = Mn as claimed. We prove these facts in the
following lemma (Lemma 2.2.3) as we shall use them again later.
Now, since the image of in is M -isomorphic to (cid:96)∞(Γn) ≡ (cid:96)kn∞ and ∪n∈N[dγ : γ ∈ Γn] is
dense in X, it follows that X is a L∞,M -space.
To complete the proof, it remains to prove that for each n ∈ N and u ∈ (cid:96)∞(Γn) there
exists a unique vector in [dγ : γ ∈ Γn] which when restricted to Γn is precisely the vector u.
In fact, we have already shown the existence of this vector; in(u) satisfies this property by
34
the above argument. So we only need to prove uniqueness. Suppose v(u) =(cid:80)
γ ⊂ Γn. Thus the αγ are uniquely determined by u and so v(u) is unique.
is such a vector. It follows that αγ = v(u)(d∗
because supp d∗
Lemma 2.2.3. The image of the map in : (cid:96)∞(Γn) → (cid:96)∞(Γ) defined in the proof of the above
theorem is precisely [dγ : γ ∈ Γn]. Moreover, for each γ ∈ ∆n, ineγ = dγ.
γ) = u(d∗
αγdγ
γ); the final equality must be true
γ∈Γn
Proof. We saw that the operator in is an isomorphic embedding, so its image is kn-dimensional,
where as in the previous proof, kn = #Γn. Since the vectors (dγ)γ∈Γn ⊆ (cid:96)∞(Γ) are linearly
independent, it is enough to show that for each γ ∈ Γn there is a u ∈ (cid:96)∞(Γn) such that
θ : θ ∈ Γ] = (cid:96)1(Γ), it would be enough to see there is a u ∈ (cid:96)∞(Γn) such
inu = dγ. Since [d∗
that (cid:104)inu, d∗
θ(cid:105) = δθ,γ. This ends up being an easy exercise in linear algebra.
θ(cid:105) = (cid:104)dγ, d∗
We choose to give an alternative proof. Fix γ ∈ Γn. We will find u ∈ (cid:96)∞(Γn) such that
inu(θ) = dγ(θ) for all θ ∈ Γ. Since in is an extension operator, if such a u exists, we must
have
∈ (cid:96)∞(Γn). All that remains to be
u(θ) = inu(θ) = dγ(θ)
whenever θ ∈ Γn. Consequently we set u = (dγ(θ))θ∈Γn
seen is that inu(θ) = dγ(θ) for all θ ∈ Γ \ Γn. To this end we note that
(cid:88)
dγ(ξ)[P ∗
(0,n]e∗
θ(cid:105)
= (cid:104)dγ, P ∗
(0,n]e∗
θ(cid:105)
= (cid:104)dγ, e∗
= dγ(θ).
inu(θ) = (cid:104)inu, e∗
θ(cid:105) = (cid:104)u, P ∗
(0,n]e∗
θ(cid:105) =
ξ∈Γn
θ](ξ)
The penultimate equality follows from the fact that dγ(x∗) = dγ(P ∗
and x∗ ∈ (cid:96)1(Γ). This is a consequence of the biorthogonalilty of the sequences (d∗
and (dγ) ⊆ (cid:96)∞(Γ).
(0,n]x∗) for all γ ∈ Γn
γ) ⊆ (cid:96)1(Γ)
Finally, note that if γ ∈ ∆n and θ ∈ Γn then
θ(cid:105) = (cid:104)dγ, d∗
dγ(θ) = (cid:104)dγ, e∗
θ(cid:105) + (cid:104)dγ, c∗
θ(cid:105) = (cid:104)dγ, d∗
θ(cid:105)
θ ∈ (cid:96)1(Γn−1) so (cid:104)dγ, c∗
since c∗
dγ(θ) = 1. In other words, when γ ∈ ∆n, u = eγ, proving that ineγ = dγ as claimed.
θ(cid:105) = 0. Therefore dγ(θ) (cid:54)= 0 only when γ = θ, in which case
Having proved the main theorem of this section, we return to a remark that we noted
γ)γ∈Γ of (cid:96)1(Γ) obtained in the Bourgain-Delbaen con-
γ)γ∈Γ. To obtain any interesting
γ's has to be
earlier. One can think of the basis (d∗
struction as a perturbation (by c∗
Banach spaces, we note that the norms of the perturbing vectors, i.e. the c∗
γ) of the canonical basis (e∗
35
n ∈ (cid:96)1(N) with d∗
'large'. If not, then the basis (d∗
γ)γ∈Γ will be equivalent to the canonical basis of (cid:96)1 and the
corresponding Bourgain-Delbaen space will simply be an isomorph of c0. More precisely,
we have the following lemma.
Lemma 2.2.4. Let (d∗
each n ∈ N there exists a vector c∗
exists θ < 1 such that (cid:107)c∗
basis (e∗
[dn : n ∈ N], is isomorphic to c0.
n)∞
n=1 be a sequence of vectors in (cid:96)1(N) with the property that for
n − c∗
n = e∗
n. Suppose further that there
n(cid:107)(cid:96)1 ≤ θ for all n. Then (d∗
n)∞
n=1 is equivalent to the canonical
n)∞
n=1 of (cid:96)1 and the closed subspace of (cid:96)∞ generated by the biorthogonal vectors,
Proof. It is easily checked that T : (cid:96)1 → (cid:96)1, T ((cid:80)∞
it is enough to show that (cid:107)I − T(cid:107) < 1. This is easy, since (cid:107)(I − T )(cid:80)∞
(cid:107)(cid:80)∞
j defines a bounded
operator on (cid:96)1 (with norm at most 1 + θ). To see the sequence (d∗
n) is equivalent to the
canonical basis, it is enough to see that T has continuous inverse. By standard results,
j(cid:107)(cid:96)1 =
j(cid:107) ≤ θ(cid:107)(aj)(cid:107)(cid:96)1 by the assumed norm condition on the c∗
j )(cid:107) = (cid:107)(cid:80)∞
j ) =(cid:80)∞
j=1 aj(e∗
j=1 ajd∗
j=1 aje∗
j=1 aje∗
j=1 ajc∗
j − d∗
n
vectors. It follows that (cid:107)I − T(cid:107) ≤ θ < 1.
The fact that X := [dn : n ∈ N] is isomorphic to c0 follows by taking the dual of T and
restricting to X. It is straightforward to check that T ∗ maps the vector dn ∈ (cid:96)∞ to the
vector en ∈ (cid:96)∞. Moreover, since T is an isomorphism, so is T ∗, and consequently T ∗(X)
is closed. The previous two observations imply that T ∗ restricted to X is an isomorphism
onto T ∗(X) = [en : n ∈ N] = c0.
We conclude this section by proving that the basis, (dγ)γ∈Γ, of any Bourgain-Delbaen
space, X(Γ, τ ), is normalised.
Lemma 2.2.5. Continuing with the notation as above, if X(Γ, τ ) is a Bourgain-Delbaen
space as in Theorem 2.2.2, then the basis (dγ)γ∈Γ is normalised.
Proof. Fix ν ∈ Γ. We prove by induction on n that if γ ∈ ∆n then dν(e∗
γ ∈ ∆1, e∗
some k and choose γ ∈ ∆k+1. We make use of the decomposition e∗
that c∗
the other possible forms of c∗
γ) ≤ 1. When
γ and the proof is obvious. So assume the estimate holds for all γ(cid:48) ∈ Γk for
γ and assume
2 , ξ ∈ Γl and b∗ ∈ ball (cid:96)1(Γk \ Γl);
(l,k]b∗ for some l < k, α ≤ 1, β ≤ θ < 1
γ = εαe∗
ξ + βP ∗
γ = d∗
γ = d∗
γ + c∗
γ) = (cid:104)dν, d∗
We consider two possible cases. If dν(d∗
ν ∈ ∆k+1 and dν(e∗
holds.
γ) = dν(d∗
γ can be treated similarly and are in any case easier. Now,
dν(e∗
γ + εαe∗
γ) (cid:54)= 0, then we must have ν = γ. In particular
γ ⊆ (cid:96)1(Γk). Consequently the inequality
γ) = 1 since supp c∗
ξ + βP ∗
(l,k]b∗(cid:105)
36
The remaining case is when dν(d∗
If, in addition, ν ∈ Γl then it follows that dν(e∗
αdν(e∗
ξ) ≤ 1 by the inductive hypothesis and the fact that α ≤ 1.
γ) = 0. We divide this case into two further possibilities.
γ) ≤
ξ) and consequently, dν(e∗
γ) = dν(εαe∗
Finally, if dν(d∗
γ) = 0 and ν ∈ Γ \ Γl, dν(e∗
(l,k]b∗(cid:105) = β(cid:104)P(l,k]dν, b∗(cid:105). This
last expression is either 0 (if ν ∈ Γ \ Γk) or β(cid:104)dν, b∗(cid:105) otherwise. In either case, it is again
easily seen using the inductive hypothesis (and that b∗ ∈ ball (cid:96)1(Γk)) that dν(e∗
γ) ≤ 1 as
required.
γ) = (cid:104)dν, βP ∗
It follows that (cid:107)dν(cid:107) ≤ 1. On the other hand, it is easily seen that dν(e∗
ν) = (cid:104)dν, d∗
ν +c∗
ν(cid:105) =
1 so that (cid:107)dν(cid:107) ≥ 1. It follows that (cid:107)dν(cid:107) = 1 as required.
2.3 The Bourgain-Delbaen Construction
We will return to the construction just described shortly. For now, we move on to look
at a few details of the original Bourgain-Delbaen construction so that we can explain the
relationship between the two constructions in the following section. The notation that
follows is that used by Haydon in [17].
We begin by setting ∆1 to be the set consisting of just one element. Inductively, if we
have sets ∆k defined for k = 1, 2, . . . , n, we set Γn = ∪n
∆n+1 = {n + 1} × (cid:91)
1≤k<n
k=1∆k and define
{k} × Γk × Γn × {±1} × {±1}.
We let Γ = ∪n≥1Γn. We make two observations. First, ∆2 = ∅ so that Γ1 = Γ2. Second,
an element γ ∈ ∆n+1 is a 6-tuple of the form γ = (n + 1, k, ξ, η, ε, ε(cid:48)). We shall use the
terminology introduced by Haydon and say that such a γ has rank n + 1.
For fixed real numbers a, b > 0, we define inductively, linear maps un : (cid:96)∞(Γn) →
(cid:96)∞(∆n+1) (for n ≥ 2) and im,n : (cid:96)∞(Γm) → (cid:96)∞(Γn) (for all n > m, n, m ∈ N). We
denote by in the map in,n+1. We set i1 = i1,2 := id(cid:96)∞(Γ1). Suppose, inductively, that
ik has been defined for all k < n. Let ik,n = in−1 ◦ ··· ◦ ik : (cid:96)∞(Γk) → (cid:96)∞(Γn). For
x ∈ (cid:96)∞(Γ) we denote by πkx the obvious restriction of x onto coordinates in Γk. We define
un : (cid:96)∞(Γn) → (cid:96)∞(∆n+1) by
(unx)(n + 1, k, ξ, η, ε, ε(cid:48)) = εax(ξ) + ε(cid:48)b[x(η) − (ik,nπkx)(η)]
and in : (cid:96)∞(Γn) → (cid:96)∞(Γn+1) by
x(γ)
(unx)(γ) γ ∈ ∆n+1.
It is easily seen that for m < n < p and x ∈ (cid:96)∞(Γm) we have
inx(γ) =
γ ∈ Γn
(cid:40)
(im,px)Γn = im,nx
37
and so we can well-define a linear mapping jm : (cid:96)∞(Γm) → RΓ by setting
(jmx)(δ) = (im,nx)(δ) whenever δ ∈ Γn.
Bourgain and Delbaen supposed the numbers a and b are chosen subject to one of the
following conditions:
(1) 0 < b < 1
2 < a < 1 and a + b > 1 (in which case we define λ := a
1−2b )
(2) a = 1 and 0 < b < 1
2 such that 1 + 2bλ ≤ λ for some λ > 1.
In either of these cases, we obtain the following lemma:
Lemma 2.3.1. The maps jm take values in (cid:96)∞(Γ). Moreover, for each m, jm is a linear
isomorphic embedding of (cid:96)∞(Γm) into (cid:96)∞(Γ) satisfying (cid:107)x(cid:107) ≤ (cid:107)jm(x)(cid:107) ≤ λ(cid:107)x(cid:107) for all
x ∈ (cid:96)∞(Γm) (where λ is as above). If m < n then im jm ⊆ im jn.
Proof. It is clear from the definition that jm is an extension operator from (cid:96)∞(Γm) to
(cid:96)∞(Γ) and consequently (cid:107)x(cid:107) ≤ (cid:107)jm(x)(cid:107) for all x ∈ (cid:96)∞(Γm). To complete the first part of
the lemma, it is clearly enough to see that (cid:107)im,n(cid:107) ≤ λ for all m < n. The proof was given in
[7] and is an easy induction argument. Indeed, since i1,2 = id(cid:96)∞(Γ1), we have (cid:107)i1,2(cid:107) = 1 ≤ λ.
Suppose inductively (cid:107)ik,m(cid:107) ≤ λ for all k < m ≤ n. We consider an x ∈ (cid:96)∞(Γn) and show
first that (cid:107)in,n+1(cid:107) ≤ λ. It is enough to show that un(x)(γ) ≤ λ(cid:107)x(cid:107) for γ ∈ ∆n+1. To this
end, write γ = (n + 1, k, ξ, η, ε, ε(cid:48)). We have
unx(γ) = aεx(ξ) + bε(cid:48)[x(η) − ik,nπkx(η)]
≤ a(cid:107)x(cid:107) + b((cid:107)x(cid:107) + λ(cid:107)x(cid:107))
≤ (a + 2bλ)(cid:107)x(cid:107)
where we have again made use of the fact that λ ≥ 1 in the final inequality. We are now
done, since in either of the cases we have a + 2bλ ≤ λ. It remains to consider (cid:107)im,n+1(cid:107)
when m < n. Let x ∈ (cid:96)∞(Γm). If γ ∈ Γn, we have im,n+1x(γ) = in,n+1 ◦ im,nx(γ) =
im,nx(γ) ≤ λ(cid:107)x(cid:107) by the inductive hypothesis. So it once again remains to consider those
γ = (n + 1, k, ξ, η, ε, ε(cid:48)) in ∆n+1. By a similar argument we get
im,n+1x(γ) = aεim,nx(ξ) + bε(cid:48)[im,nx(η) − ik,nπkim,nx(η)].
We consider two separate cases. Firstly, if k ≤ m, since ξ ∈ Γk, we have im,nx(ξ) = x(ξ)
and also, from construction of the map im,n we have πkim,nx = πkx. We therefore have
im,n+1x(γ) = aεx(ξ) + bε(cid:48)[im,nx(η) − ik,nπkx(η)] and estimating as before yields this is
at most (a + 2bλ)(cid:107)x(cid:107) ≤ λ(cid:107)x(cid:107).
38
Otherwise n > k > m and it is easily verified that πkim,nx = im,kx. So ik,nπkim,nx =
ik,n ◦ im,kx = im,nx. Consequently im,n+1x(γ) = aεim,nx(ξ) ≤ (cid:107)im,nx(cid:107) ≤ λ(cid:107)x(cid:107). Here we
made use of the fact that a ≤ 1 and the inductive hypothesis. This completes the induction
and the first part of the proof.
It remains to see that if m < n, im jm ⊆ im jn. We will show that for x ∈ (cid:96)∞(Γm),
jmx = jn(πnjmx). It is easy to see that πnjmx = im,nx so it suffices to show that jmx =
jn(im,nx). To see this, note that if δ ∈ Γ, we can choose p > n > m such that δ ∈ Γp. We
then have
jn(im,nx)(δ) = in,p(im,nx)(δ) = im,px(δ) = jmx(δ)
as required.
It follows from the lemma that
X = Xa,b =
(cid:91)
n≥1
im(jn)
is a closed subspace of (cid:96)∞(Γ). Moreover, the lemma shows that the image, im(jn) is λ-
isomorphic to (cid:96)∞(Γn) for every n so that X is a L∞,λ space. In Bourgain's notation, the
spaces of class X are the spaces Xa,b for which a = 1 and 0 < b < 1
2 , whilst those of class
Y are the spaces Xa,b with 0 < b < 1
2 < a < 1 and a + b > 1.
Remark 2.3.2. We conclude this section with a remark.
It is shown in [17] that each
of the spaces, Xa,b, described above has a natural Schauder basis, and finite dimensional
decomposition. More precisely, for γ ∈ Γ, let n = rank (γ) and eγ be the usual unit vector
(i.e. eγ(δ) = 1 if δ = γ and 0 otherwise). Define dγ := jn(eγ). The sequence (dγ)γ∈Γ (under
a suitable enumeration) is a Schauder basis for X. In fact, it is observed in [17] that this
basis is normalized, i.e. (cid:107)dγ(cid:107) = 1 for all γ. Moreover, if we let Mn := span{dγ : γ ∈ ∆n},
then (Mn)∞
n=1 is a FDD for X. We will see in the next section that once we have shown
that we can describe the original Bourgain-Delbaen construction in the Argyros-Haydon
framework, these facts are immediate.
2.4 Connecting the two constructions
As explained earlier, we wish to formulate the original Bourgain-Delbaen construction just
discussed within the framework of the generalised construction discussed in Section 2.2. We
continue with the same notation from the previous section; in particular, we work in this
section with the set Γ just described in the previous section. We continue to denote vectors
of (cid:96)1 with a star notation. This notation allows us to distinguish between eγ (the standard
unit vector in (cid:96)∞(Γ)) and e∗
γ (the standard unit vector in (cid:96)1(Γ)) for example. Our aim is to
39
construct a suitable τ mapping on Γ as in Theorem 2.2.2 so that we can apply this theorem,
obtaining a space of Bourgain-Delbaen type, which is precisely the same space obtained in
the previous section.
We note that the extension operator jn : (cid:96)∞(Γn) → (cid:96)∞(Γ) defined in the previous section
could really be thought of as an isomorphic embedding from (cid:96)1(Γn)∗ to (cid:96)1(Γ)∗ under the
canonical identifications, (cid:96)1(Γn)∗ ≡ (cid:96)∞(Γn) and (cid:96)1(Γ)∗ ≡ (cid:96)∞(Γ). Thought of as a map on
these dual spaces, it is easy to see that jn is weak*-norm, and thus weak*-weak* continuous.
Indeed, this follows since the weak* and norm topologies coincide on finite dimensional
spaces, and we saw in the previous section that jn is norm-norm continuos. It follows by
Lemma 1.4.11 that jn can be obtained as the dual of some operator from (cid:96)1(Γ) to (cid:96)1(Γn).
We will show that under the correct choice of τ mapping on Γ, this dual operator is exactly
the operator P ∗
(0,n] appearing in Theorem 2.2.2.
If we are attempting to find the operator which has dual jn, it is sensible to consider the
dual of jn since, roughly speaking, jn and j∗∗
n can be considered 'the same' operator. More
precisely, it is well known that for any operator T : X → Y , the operator T ∗∗JX maps into
JY (Y ) and moreover, T = J−1
Y T ∗∗JX . Thinking about how the operators jn are defined, to
compute j∗
n, we must first compute the dual of the operators im,n.
We begin by computing the dual operator of in, where we assume n ≥ 2, since i1 =
Id(cid:96)∞(Γ1). Throughout this section, we will make the usual identifications of (cid:96)1(Γn)∗ with
(cid:96)∞(Γn) and (cid:96)∞(Γn)∗ with (cid:96)1(Γn). Under these identifications, we compute i∗
n, obtaining
n : (cid:96)1(Γn+1) → (cid:96)1(Γn). For x∗ ∈ (cid:96)1(Γn+1), γ ∈ Γn,
i∗
(cid:88)
n(x∗)](γ) = x∗(in(eγ))
[i∗
= x∗(γ) +
x∗(ξ)[in(eγ)](ξ).
ξ∈∆n+1
nx∗ = x∗ and we see that i∗
We observe that if supp x∗ ⊆ Γn then all the terms appearing in the above sum are zero
n : (cid:96)1(Γn+1) →
except the x∗(γ) term. Thus, for such vectors, i∗
(cid:96)1(Γn) is a projection onto (cid:96)1(Γn). Furthermore, we note that this is not the usual projection
θ to itself when θ ∈ Γn and 0 otherwise). To see this, we
on (cid:96)1(Γn) (which would send e∗
exhibit a γ ∈ Γ with rank γ = n + 1 but i∗
γ (cid:54)= 0. We take γ = (n + 1, k, ξ, η, 1, 1) with
ne∗
k < n and both ξ and η belonging to Γk. Setting x∗ = e∗
γ in the previous formula we see
γ)(ξ) = in(eξ)(γ) = aeξ(ξ) + b[eξ(η) − (ik,nπkeξ)(η)] = a (cid:54)= 0. (Here, we made use
that i∗
of the fact that ik,nπkeξ(η) = eξ(η) since ξ and η are in Γk.)
n(e∗
Given a γ with rank γ = n + 1, we can obviously decompose e∗
projection, i∗
ne∗
γ, of e∗
γ, and a vector in ker i∗
n. Precisely, we have
γ into a sum of the
e∗
γ = i∗
n(e∗
γ) + [e∗
γ − i∗
n(e∗
γ)].
40
Recalling that we are attempting to obtain the original Bourgain-Delbaen spaces via the
n(e∗
γ := 0.
γ := i∗
γ) and d∗
generalised construction of Argyros-Haydon and comparing this formula with Theorem
γ (note we are still assuming n ≥ 2). For
2.2.2, we define c∗
γ ∈ Γ1 define c∗
Remark 2.4.1. Note the vectors d∗
γ just defined should not be confused with the
vectors appearing in Thorem 2.2.2; in any case, we haven't even defined a mapping τ on Γ yet
γ and c∗
γ = e∗
γ − c∗
for the theorem to be applied in any sensible way. The vectors are defined by the formulae
just stated and we work with these vectors throughout this section unless explicitly stated
otherwise. The reason for this (perhaps slightly confusing) notation is the following. We
appearing in Theorem 2.2.2. We can thus apply the theorem, and obtain vectors (cid:98)c∗
will show that for each γ ∈ Γ, it is possible to define a tuple, τ (γ), having one of the forms
(cid:99)d∗
γ and
γ = e∗
γ. Moreover, it
will turn out that the space of Bourgain-Delbaen type constructed via Theorem 2.2.2 under
γ. We will eventually show that (cid:98)c∗
γ and consequently(cid:99)d∗
γ −(cid:98)c∗
γ = d∗
γ = c∗
this τ mapping is precisely the original Bourgain-Delbaen space of the previous section (for
a given pair of real numbers a, b satisfying the conditions already discussed).
It remains to see how to construct such a τ map. To this end, we write γ = (n +
1, k, ϕ, η, ε, ε(cid:48)) and for γ ∈ Γn, we look at
[i∗
n(e∗
γ)](γ) = e∗
γ(γ) +
(cid:88)
ξ∈∆n+1
e∗
γ(ξ)[in(eγ)](ξ)
= ineγ(γ)
= uneγ(γ)
= εaeγ(ϕ) + ε(cid:48)b[eγ − ik,nπkeγ](η).
The second term appearing in the above sum is clearly the harder term to deal with so
we look at this first. There is one situation in which this term is easy to compute, namely
if η ∈ Γk. In this case ik,nπkeγ(η) = πkeγ(η), so [eγ − ik,nπkeγ](η) = [eγ − πkeγ](η). We
then consider cases:
(1) γ (cid:54)= η. In this case, we clearly have that eγ(η) = 0. Since the only possible vectors that
πkeγ can be are 0 and eγ, we see that πkeγ(η) = 0 also. Thus [eγ − ik,nπkeγ](η) = 0.
(2) γ = η ∈ Γk. In this case, πkeγ = eγ and we again find [eγ − ik,nπkeγ](η) = 0.
To obtain a useful expression for [eγ − ik,nπkeγ](η) when η ∈ Γn \ Γk, it turns out to
m,n : (cid:96)1(Γn) →
be useful to look at the dual operator of im,n : (cid:96)∞(Γm) → (cid:96)∞(Γn), giving i∗
41
(cid:96)1(Γm). For x∗ ∈ (cid:96)1(Γn), γ ∈ Γm
[i∗
m,n(x∗)](γ) = x∗(im,neγ)
= x∗(γ) +
x∗(ξ)[im,neγ](ξ).
(cid:88)
ξ∈Γn\Γm
We take this opportunity to document a special case of this formula which will be
particularly useful to us later. We note that for η ∈ Γn \ Γm, γ ∈ Γm, we have
[i∗
m,n(e∗
As before, it is easy to see that each i∗
η)](γ) = im,neγ(η)
(2.2)
m,n is a projection onto (cid:96)1(Γm). Since ∪n(cid:96)1(Γn) =
(0,k] : (cid:96)1(Γ) →
c00(Γ) is a dense subspace of (cid:96)1(Γ), it is natural to attempt to define projections P ∗
(cid:96)1(Γk) by
(0,k](x) = i∗
P ∗
k,n(x) whenever x ∈ (cid:96)1(Γn), (n > k).
Assuming this is well-defined, the formula defines (linear) projections P ∗
(0,k] from the
dense subspace c00(Γ) ⊆ (cid:96)1(Γ), onto (cid:96)1(Γk). Moreover, it follows from the proof of Lemma
k,n(cid:107) ≤ λ for all k < n. Consequently
2.3.1 (and standard facts about dual operators) that (cid:107)i∗
P ∗
(0,k] is bounded (with norm at most λ) and thus extends to a bounded projection defined
on (cid:96)1(Γ) as required. So it remains to check that we can well define P ∗
(0,k] by the above
formula. Suppose x∗ ∈ (cid:96)1(Γn) ⊂ (cid:96)1(Γn(cid:48)) (n(cid:48) > n > k). We saw earlier that for γ ∈ Γk,
k,n(cid:48)(x∗)](γ) = x∗(γ) +
[i∗
= x∗(γ) +
x∗(ξ)[ik,n(cid:48)eγ](ξ)
x∗(ξ)[ik,n(cid:48)eγ](ξ)
(cid:88)
(cid:88)
ξ∈Γn(cid:48)\Γk
ξ∈Γn\Γk
where the final equality is because x∗ ∈ (cid:96)1(Γn) so x∗(ξ) = 0 for ξ ∈ Γn(cid:48) \ Γn.
It is
immediate from the definition of the operators im,n that ik,n(cid:48) = in,n(cid:48) ◦ ik,n. So for ξ ∈ Γn,
ik,n(cid:48)eγ(ξ) = in,n(cid:48) ◦ ik,neγ(ξ) = ik,neγ(ξ). Combining these observations, we find that
[i∗
k,n(cid:48)(x∗)](γ) = x∗(γ) +
(cid:88)
(cid:88)
k,n(x∗)](γ)
ξ∈Γn\Γk
ξ∈Γn\Γk
= x∗(γ) +
= [i∗
x∗(ξ)[ik,n(cid:48)eγ](ξ)
x∗(ξ)[ik,neγ](ξ)
so that P ∗
(0,k] is well defined.
We recall what we set out to do. For γ = (n + 1, k, ϕ, η, ε, ε(cid:48)) ∈ ∆n+1 and γ ∈ Γn, we
γ)](γ). We have seen already that
γ(γ) := [i∗
n(e∗
were attempting to find an expression for c∗
γ(γ) = εaeγ(ϕ) + ε(cid:48)b[eγ − ik,nπkeγ](η).
c∗
We consider 2 separate cases:
42
(1) η ∈ Γk. We have seen that the second term in the above expression is just 0. We thus
see that c∗
γ(γ) = εaeγ(ϕ) = εae∗
ϕ(γ). In other words, c∗
γ = εae∗
ϕ.
(2) η ∈ Γn \ Γk. If in addition, rank (γ) ≤ k, it follows that πkeγ = eγ. Combining this
observation with equation (2.2) and the definition of P ∗
(0,k], we get
(0,k](e∗
[P ∗
η)](γ) = [i∗
k,n(e∗
η)](γ) = ik,neγ(η) = ik,nπkeγ(η).
Thus
γ(γ) = εaeγ(ϕ) + ε(cid:48)b[eγ − ik,nπkeγ](η)
c∗
η)](γ)
(cid:110)
η − [P ∗
ϕ(γ) + ε(cid:48)b[e∗
[(I − P ∗
ϕ(γ) + ε(cid:48)b
(0,k](e∗
(0,k])(e∗
= εae∗
= εae∗
η)](γ)
(cid:111)
.
If on the other hand rank (γ) > k, πkeγ = 0 so
γ(γ) = εaeγϕ + ε(cid:48)b[eγ − ik,nπkeγ](η)
c∗
= εae∗
= εae∗
(cid:110)
ϕ(γ) + ε(cid:48)be∗
ϕ(γ) + ε(cid:48)b
η(γ)
[(I − P ∗
(0,k])(e∗
η)](γ)
where the final equality follows since rank (γ) > k =⇒ [P ∗
x∗ ∈ (cid:96)1(Γ). We thus conclude that
(cid:111)
(0,k](x∗)](γ) = 0 for any
c∗
γ = εae∗
ϕ + ε(cid:48)b(I − P ∗
(0,k])(e∗
η).
We make one more observation before coming to the main result of this section.
Remark 2.4.2. An easy induction argument yields that span{d∗
γ : rank γ ≤ n} = span{e∗
γ :
rank γ ≤ n} for every n. Indeed, the base case is immediate since by definition, d∗
γ = e∗
γ for
rank γ ≤ 2 (recalling that ∆2 = ∅ so Γ1 = Γ2). The inductive step follows from the relation
e∗
γ = d∗
γ ⊆ Γn when rank γ = n + 1.
γ for all γ and supp c∗
γ + c∗
If follows that [d∗
γ : γ ∈ Γ] = (cid:96)1(Γ). Consequently the operators P ∗
are completely determined by their action on the d∗
(0,k] previously defined
(cid:40)
d∗
0
γ
γ. In fact
rank (γ) ≤ k
otherwise.
P ∗
(0,k](d∗
γ) =
To see this, we note the formula certainly holds when rank (γ) ≤ k since we have already
observed that the maps P ∗
γ) =
0 when rank (γ) > k. We consider the case where k ≥ 2. Suppose first rank (γ) = k + 1.
Then P ∗
γ = 0. (We have of course
(0,k] are projections onto (cid:96)1(Γk). It remains to see that P ∗
γ) = i∗
γ) = i∗
γ) = i∗
γ) − c∗
(0,k](d∗
γ − c∗
γ = c∗
γ − c∗
(0,k](d∗
k(d∗
k(e∗
k(e∗
43
γ as being equal to i∗
ke∗
made use of the definition of the vector c∗
γ since rank γ = k + 1). If
γ) = i∗
n−1,n(d∗
γ) = i∗
k,n(d∗
rank (γ) = n > k + 1, P ∗
k,n−1(0) = 0 where the
penultimate equality follows from the previous argument with k replaced by n − 1. Note
we also made use of the following easily verified fact: if k < m < n then i∗
m,n.
The proof for the case k = 1 is similar.
k,n−1 ◦ i∗
k,n = i∗
k,m ◦ i∗
(0,k](d∗
γ) = i∗
We are finally ready to present the main result of this section.
(cid:40)
τ (γ) =
Proposition 2.4.3. Let (∆q)∞
q=1, Γn and Γ be the sets defined as in the original Bourgain-
Delbaen construction and let a, b be real numbers satisfying either of the two conditions
assumed by Bourgain and Delbaen (see the previous section). We define τ on Γ \ Γ1 as
follows. Given γ ∈ ∆n+1, write γ = (n + 1, l, ξ, η, ε, ε(cid:48)) and set
if η ∈ Γl
if η ∈ Γn \ Γl.
(ε, a, ξ)
(ε, a, ξ, l, b, ε(cid:48)e∗
η)
(We clarify that the slightly clumsy notation ε(cid:48)e∗
interpreted as the scalar ε(cid:48) multiplied by the vector e∗
BD setups, a ≤ 1 and b < 1
projections (cid:91)P ∗
2 , Theorem 2.2.2 yields that there exist(cid:99)d∗
η appearing in the final tuple above is to be
η). Noting that in either of the original
γ ∈ (cid:96)1(Γ) and
γ −(cid:98)c∗
(0,q](cid:99)d∗
(cid:98)c∗
(cid:16)(cid:99)d∗
γ = e∗
(0,q] on (cid:96)1(Γ) uniquely determined by the following properties:
(cid:40)(cid:99)d∗
0
elements dγ of the(cid:99)d∗
With the notation that has been used throughout this section, we have (cid:98)c∗
γ generate a L∞,λ subpace of (cid:96)∞(Γ), X(Γ) := X(Γ, τ ).
the (cid:99)d∗
(cid:16)
(cid:17)∗
γ) coincide with the c∗
The family
ous section). Moreover, the norm of each projection (cid:91)P ∗
γ ∈ ∆1
γ = (n + 1, l, ξ, η, ε, ε(cid:48)) and η ∈ Γl
γ = (n + 1, l, ξ, η, ε, ε(cid:48)) and η ∈ Γn \ Γl.
γ (and hence
(0,k]. Moreover, the mappings
: (cid:96)∞(Γk) → (cid:96)∞(Γ) coincide with the mappings jk of the original Bourgain-Delbaen
γ (respectively d∗
γ) and (cid:91)P ∗
is a basis for (cid:96)1(Γ) with basis constant at most λ (as defined in the previ-
(0,q] is at most λ and the biorthogonal
ξ + ε(cid:48)b(I − (cid:91)P ∗
(0,l])e∗
γ ∈ Γq
γ ∈ Γ \ Γq
(0,k] = P ∗
εae∗
εae∗
ξ
γ
γ∈Γ
(cid:91)P ∗
γ =
γ =
(cid:17)
γ
0
η
P ∗
construction.
(0,k]
Before giving the proof we make a few observations that follow from the proposition.
Note we have the following important corollary.
Corollary 2.4.4. The subspace X(Γ) of the above proposition is precisely the original
Bourgain-Delbaen space Xa,b.
44
Proof. It follows from the proof of Theorem 2.2.2 and Lemma 2.2.3 that
So, by the proposition,
X(Γ) =
X(Γ) =
(cid:17)∗
.
(cid:17)∗
(0,n]
(cid:16) (cid:91)P ∗
(cid:16)
P ∗
(0,n]
im
im
(cid:91)
n∈N
(cid:91)
(cid:91)
n∈N
im jn
=
n∈N
= Xa,b.
Thus we can think of the original Bourgain-Delbaen construction in terms of the Argyros-
Haydon construction as we wanted. Moreover, we see from the generalised construction,
i.e. by Theorem 2.2.2, that the biorthogonal elements dγ of the d∗
γ are a basis of
X(Γ) = Xa,b. It follows by Lemma 2.2.3 and the proposition that for γ ∈ ∆n, dγ = jneγ.
The fact this basis is normalised follows from Lemma 2.2.5. Thus we obtain all the results
γ = (cid:99)d∗
stated in Remark 2.3.2 as immediate corollaries.
γ for all
It is
γ. Thus by Remark 2.4.2 and
γ. So
(0,k]d∗
(0,k](cid:99)d∗
It remains for us to prove the proposition.
Proof of Proposition 2.4.3. We prove by induction that the statement " (cid:99)d∗
γ, (cid:99)d∗
We recall that the vectors (cid:98)c∗
(0,k](cid:96)1(Γn)" holds for every n ∈ N.
immediate from the definitions that when γ ∈ Γ1,(cid:99)d∗
γ and operators (cid:91)P ∗
γ = d∗
γ ∈ Γn and for all k ∈ N, P ∗
(0,k](cid:96)1(Γn) = (cid:91)P ∗
(0,k] are defined inductively.
γ = e∗
γ = d∗
(0,k](cid:96)1(Γ1)
all γ ∈ Γn and P ∗
(0,k](cid:96)1(Γ1) = (cid:91)P ∗
P ∗
the way in which the (cid:91)P ∗
∀k ∈ N
(0,k] are constructed, we have that for γ ∈ Γ1, P ∗
γ = (cid:91)P ∗
and the statement is true for n = 1. Suppose inductively that for some n ≥ 1,(cid:99)d∗
(0,k](cid:96)1(Γn) = (cid:91)P ∗
(0,k](cid:96)1(Γn). Then for γ = (n + 1, l, ξ, η, ε, ε(cid:48)) ∈ ∆n+1,
(cid:40)
εae∗
(cid:40)
εae∗
εae∗
=
εae∗
= c∗
γ.
η ∈ Γl
η ∈ Γn \ Γl
η ∈ Γl
η ∈ Γn \ Γl
ξ + ε(cid:48)b(I − (cid:91)P ∗
ξ + ε(cid:48)b(I − P ∗
(0,l])e∗
η
(0,l])e∗
η
(cid:98)c∗
γ =
ξ
ξ
The first of the above equalities holds due to the inductive construction of the (cid:98)c∗
γ, the second
by the inductive hypothesis and the final equality is a consequence of the very definition of
γ = d∗
γ for
45
γ − (cid:98)c∗
(0,k](cid:96)1(Γn+1) = (cid:91)P ∗
(0,k] and all our previous calculations. It follows that (cid:99)d∗
the P ∗
γ. We again
(0,k](cid:96)1(Γn+1) for all k ∈ N as a consequence of Remark 2.4.2 and
get that P ∗
the way in which the (cid:91)P ∗
(0,k] are constructed. The first part of the proof is thus complete
(cid:17)∗
by induction and an elementary density argument. As a consequence we now drop the 'hat
(cid:104)(cid:16)
= jk : (cid:96)∞(Γk) → (cid:96)∞(Γ). For x ∈ (cid:96)∞(Γk), γ ∈ Γ
It remains to prove
γ = d∗
γ = e∗
notation'.
(cid:17)∗
(cid:16)
P ∗
(cid:105)
(0,k]
P ∗
(0,k]
(x)
(γ) = (cid:104)x, P ∗
γ)(cid:105)
(0,k](e∗
where we are, of course, thinking of x as an element of (cid:96)1(Γk)∗ (under the canonical iden-
tification) in the RHS of the equality. If rank (γ) ≤ k, P ∗
γ from which it follows
that
(γ) = x(γ) = [jk(x)](γ) as required. Obviously we must see that this
γ = e∗
(0,k]e∗
(cid:104)(cid:16)
(cid:17)∗
P ∗
(cid:105)
(x)
(0,k]
equality also holds when rank γ > k. We will need the following lemma.
Lemma 2.4.5. If x ∈ (cid:96)∞(Γn) (which we identify in the canonical way with (cid:96)1(Γn)∗) and
γ ∈ Γ with rank (γ) = n + 1 then
x(c∗
γ) = [jnx](γ).
In order not to detract from the proof at hand, we defer the proof of this lemma to the
end.
Suppose now that rank (γ) = k + 1. Then P ∗
ke∗
γ = c∗
γ = i∗
γ. So
(cid:104)(cid:16)
(cid:17)∗
(cid:105)
(0,k]e∗
(γ) = x(c∗
P ∗
(0,k]
(x)
γ) = [jkx](γ)
where we recall x ∈ (cid:96)∞(Γk) and appeal to Lemma 2.4.5 to obtain the final equality.
Finally, we assume rank (γ) = n > k + 1. We have
(cid:104)(cid:16)
(cid:17)∗
(cid:105)
P ∗
(0,k]
(x)
(γ) = (cid:104)x, P ∗
= (cid:104)x, i∗
= (cid:104)x, i∗
= (cid:104)x ◦ i∗
= (cid:104)x ◦ i∗
γ)(cid:105)
(0,k](e∗
γ)(cid:105)
k,n(e∗
γ)(cid:105)
n−1,n(e∗
k,n−1i∗
γ(cid:105)
k,n−1, i∗
n−1,ne∗
γ(cid:105).
k,n−1, c∗
We now think of x ◦ i∗
again to Lemma 2.4.5 we see
k,n−1 ∈ (cid:96)1(Γn−1)∗ as an element of (cid:96)∞(Γn−1). Appealing once
(cid:105)
(cid:17)∗
(cid:104)(cid:16)
P ∗
γ) = [jn−1(x ◦ i∗
So to complete the proof it is enough to see that [jn−1(x ◦ i∗
(γ) = x ◦ i∗
k,n−1(c∗
(x)
(0,k]
k,n−1)](γ) = [jk(x)](γ). But
we recall from the original B.D construction that for any x ∈ (cid:96)∞(Γk), jk(x) = jn−1(ik,n−1x).
k,n−1)](γ).
46
(We saw this in the proof of Lemma 2.3.1.) So in fact, all we need to see is that when ik,n−1x
is thought of as an element of (cid:96)1(Γn−1)∗ under the canonical identification, it is precisely
the element x ◦ i∗
k,n−1 ∈ (cid:96)1(Γn−1)∗. To this end, let y∗ be an element of (cid:96)1(Γn−1). So
x ◦ i∗
k,n−1(y∗) =
x ◦ i∗
k,n−1(e∗
θ)y∗(θ)
(2.3)
where we are, of course, still thinking of x as an element of (cid:96)1(Γk)∗. Really x ∈ (cid:96)∞(Γk), so
we can compute
x ◦ i∗
k,n−1(e∗
θ) =
x(ν)[i∗
k,n−1(e∗
θ)](ν).
(2.4)
Fortunately, we computed earlier that for ν ∈ Γk,
[i∗
k,n−1(e∗
θ)](ν) = e∗
θ(ν) +
e∗
θ(ε)[ik,n−1eν](ε).
(cid:88)
θ∈Γn−1
(cid:88)
ν∈Γk
(cid:88)
ε∈Γn−1\Γk
If θ ∈ Γn−1 \ Γk then this is just [ik,n−1eν](θ), whilst for θ ∈ Γk, we obtain e∗
So by Equation (2.4) we find
θ(ν) = eν(θ).
x ◦ i∗
k,n−1(e∗
θ) =
x(ν)eν(θ) = x(θ)
x(ν)[ik,n−1eν](θ) = ik,n−1x(θ)
if θ ∈ Γk
if θ ∈ Γn−1 \ Γk.
ν∈Γk
ν∈Γk
(cid:40)(cid:80)
(cid:80)
Substituting this back into Equation (2.3) we get
(cid:88)
(cid:88)
θ∈Γk
x ◦ i∗
k,n−1(y∗) =
=
(cid:88)
x(θ)y∗(θ) +
[ik,n−1x](θ)y∗(θ)
θ∈Γn−1\Γk
[ik,n−1x](θ)y∗(θ)
θ∈Γn−1
= [ik,n−1x](y∗)
where we think of ik,n−1x as an element of (cid:96)1(Γn−1)∗. The second equality above comes
from observing that when θ ∈ Γk, [ik,n−1x](θ) = x(θ).
Proof of Lemma 2.4.5. We use the same notation as that used for the statement of the
lemma. We consider only the more complex case where γ = (n+1, l, ξ, η, ε, ε(cid:48)) and η ∈ Γn\Γl
since the other case is similar (and easier). So,
(cid:88)
(cid:88)
θ∈Γn
θ∈Γn
x(c∗
γ) =
=
x(θ)c∗
γ(θ)
(cid:88)
θ∈Γn
x(θ)εae∗
ξ(θ) +
x(θ)ε(cid:48)b[(I − P ∗
(0,l])e∗
η](θ)
= εax(ξ) +
x(θ)ε(cid:48)b[(I − P ∗
(0,l])e∗
η](θ)
(cid:88)
θ∈Γn
47
ξ(θ) can only be non-zero when θ = ξ and since ξ ∈ Γl, l < n, this happens precisely
since e∗
once. If rank (θ) ≤ l, i.e. θ ∈ Γl, since we also assume η ∈ Γn \ Γl, the definition of P ∗
given earlier and Equation (2.2) yields
l,n(e∗
(cid:40)
where the final equality holds since it is assumed rank (θ) ≤ l. It is then easily seen that
η)](θ) = [il,neθ](η) = [il,nπleθ](η)
η](θ) = [i∗
[P ∗
(0,l]e∗
(0,l]
[(I − P ∗
(0,l])e∗
η](θ) =
e∗
η(θ) = eθ(η)
[eθ − il,nπleθ](η)
if rank (θ) > l
if rank (θ) ≤ l
However, it is clear that [eθ − il,nπleθ](η) = eθ(η) if rank (θ) > l. So in fact, we have shown
[(I − P ∗
η](θ) = [eθ − il,nπleθ](η). Consequently,
(cid:88)
η](θ) = ε(cid:48)b
x(θ)[(I − P ∗
(0,l])e∗
ε(cid:48)b
(0,l])e∗
(cid:88)
since clearly x =(cid:80)
θ∈Γn
x(θ)[eθ − il,nπleθ](η)
θ∈Γn
= ε(cid:48)b{(x − il,nπlx)(η)}
θ∈Γn
x(c∗
x(θ)eθ and all the maps are linear. So, we finally have
γ) = εax(ξ) + ε(cid:48)b{(x − il,nπlx)(η)} = [jn(x)](γ)
as claimed.
2.5 The operator algebras for the original Bourgain-Delbaen
spaces
We conclude this chapter by exhibiting an uncountable set of isometries on the original
Bourgain-Delbaen spaces which are pairwise distance C apart (with respect to the usual
operator norm) for some constant C > 0. In particular, this establishes that the operator
algebra L(X) is non-separable for any of the original Bourgain-Delbaen spaces X.
We take this opportunity to note that it is in fact easy to show K(X) is non-separable
when X is a Bourgain-Delbaen space of class X , so certainly L(X) is non-separable. Indeed,
it was shown in [7, Theorem 4.4] that in this case, every infinite dimensional subspace of X
has a subspace isomorphic to (cid:96)1. In particular, (cid:96)1 embeds into X. Since the dual operator of
an isomorphic embedding is a quotient operator, it follows that there is a quotient operator
Q : X∗ → (cid:96)∞. Consequently, X∗ cannot be separable and therefore neither is K(X) since
X∗ always embeds isometrically into K(X).
The non-separability is interesting because, as commented by Beanland and Mitchell in
[4], Emmanule showed that (cid:96)∞ does not embed into L(X) when X ∈ X , thus giving an
example of a non-separable Banach space containing no isomorphic copy of (cid:96)∞. Similar
48
questions about the separability of L(Y ) when Y ∈ Y remained open however. This section
addresses this problem; we remark that at the time of writing, it remains unknown if (cid:96)∞
embeds into L(Y ).
2.5.1 Construction of the basic operator
For the remainder of this chapter, we fix real numbers a, b satisfying either of the conditions
of Bourgain and Delbaen (see Section 2.3) and work with the Bourgain-Delbaen space
X = Xa,b = X(Γ, τ ) (for some suitable τ mapping, as shown in the previous section). We
begin by inductively constructing an isometry S∗ : (cid:96)1(Γ) → (cid:96)1(Γ). We will show that the
dual of this operator restricts to give an isometry S : X → X. In the following section, we
will then take appropriate compositions of operators of this type to obtain the uncountable
set of isometries we seek. We continue with the notation used in the preceding section.
To construct S∗, we fix a γ = (n + 1, k, ξ, η, ε, ε(cid:48)) ∈ ∆n+1 (where n ≥ 2) and aim to
construct (inductively) a bijective mapping F : Γ → Γ with the following properties
(1) F (γ) = (n + 1, k, ξ, η,−ε,−ε(cid:48)) =: γ
The operator S∗ : (cid:96)1(Γ) → (cid:96)1(Γ) will map d∗
τ to d∗
F (τ ) for all τ /∈ {γ, γ} and d∗
τ to −d∗
(2) F (γ) = γ
(3) F (τ ) = τ if rank (τ ) ≤ rank (γ) and τ /∈ {γ, γ}
(4) rank (F (τ )) = rank (τ ) ∀ τ ∈ Γ. We shall say F is 'rank-preserving'.
Note that we must assume n ≥ 2 in the above since we recall that ∆2 = ∅ and Γ1 = Γ2 is
simply a singleton set.
F (τ )
for τ ∈ {γ, γ}. (The reason for this sign change in the special case where τ ∈ {γ, γ} will
shortly become apparent.) The important point here is that we know how S∗ acts on the
d∗
τ vectors; it is this property that will ensure the dual map restricts to give a map from
X into X. Of course, the property just described only defines a linear map on the dense
subspace sp{d∗
τ : τ ∈ Γ} of (cid:96)1(Γ). We also need to show that this map is bounded, so that
it can be extended to a bounded operator on (cid:96)1(Γ). Note that since we are working with
the (cid:96)1 norm, in order to prove S∗ is bounded, we only need to be able to control (cid:107)S∗(e∗
τ )(cid:107).
More precisely, if there is some M ≥ 0 such that (cid:107)S∗(e∗
τ )(cid:107) ≤ M for every τ ∈ Γ, then it is
elementary to check that S∗ is bounded with norm at most M .
49
To begin the inductive constructions of S∗ and F we define F : Γn+1 → Γn+1 and
S∗ : (cid:96)1(Γn+1) → (cid:96)1(Γn+1) by
γ
(cid:40)
F (τ ) =
S∗(d∗
τ ) =
if τ = γ
if τ = γ
τ /∈ {γ, γ}
γ
τ
d∗
−d∗
θ : θ ∈ Γn} = sp{d∗
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}.
F (τ )
τ
Recall that if rank τ ≤ n + 1, c∗
the linearity of S∗, it is easily seen that for all τ ∈ Γ with rank τ ≤ n + 1, S∗c∗
Therefore, S∗e∗
τ ∈ sp{e∗
τ = S∗(c∗
θ : θ ∈ Γn}. Consequently, using
τ = c∗
τ .
τ for all τ ∈ Γn+1 \ {γ, γ}. We also have
γ − d∗
γ − d∗
γ + c∗
γ)
γ + (c∗
γ
τ ) + S∗(d∗
S∗(e∗
τ ) = c∗
γ) = S∗(c∗
τ + d∗
γ + d∗
τ = e∗
γ) = c∗
= −c∗
= −e∗
γ
since if we look at the possible values of c∗
(We remark that this computation works due to the 'sign change', S∗d∗
earlier.) In a similar way we obtain S∗(e∗
summarised by the following formulae which are valid for all τ ∈ Γn+1.
γ obtained in Section 2.4 we see c∗
γ = −d∗
γ + c∗
γ = 0.
γ, noted
γ. The above discussion can be succinctly
γ) = −e∗
γ and c∗
(cid:40)
(cid:40)
S∗(e∗
τ ) =
e∗
τ = e∗
−e∗
F (τ )
F (τ )
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}.
Since also c∗
τ = e∗
τ − d∗
τ for any τ ∈ Γ it is easily checked that
c∗
τ = c∗
τ = −c∗
c∗
(We once again made use of the fact that c∗
S∗(c∗
τ ) =
F (τ )
F (τ )
γ + c∗
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}.
γ = 0 in the above formula).
It is now clear how the inductive construction proceeds. We assume inductively that for
some k ∈ N we have defined F : Γn+k → Γn+k and a linear map S∗ : (cid:96)1(Γn+k) → (cid:96)1(Γn+k)
satisfying:
(1) F : Γn+k → Γn+k is a bijection, satisfying the four desired properties of F given at the
beginning of this section.
(2) For each τ ∈ Γn+k,
(i)
S∗(e∗
τ ) =
(cid:40)
e∗
−e∗
F (τ )
F (τ )
50
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}
(ii)
(iii)
S∗(d∗
τ ) =
S∗(c∗
τ ) =
(cid:40)
(cid:40)
d∗
−d∗
F (τ )
F (τ )
c∗
−c∗
F (τ )
F (τ )
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}.
noting that the previous discussion gives us the case k = 1. We will show that we can
extend the maps F and S∗ to F : Γn+k+1 → Γn+k+1 and S∗ : (cid:96)1(Γn+k+1) → (cid:96)1(Γn+k+1)
such that the above properties are still satisfied. Let τ = (n + k + 1, l, θ, ϕ, , (cid:48)). Recalling
that c∗
In fact, by
Proposition 2.4.3,
τ ) is already defined by the inductive hypothesis.
τ ∈ (cid:96)1(Γn+k), S∗(c∗
(cid:40)aS∗(e∗
θ)
aS∗(e∗
θ) + (cid:48)bS∗(cid:16)
S∗(c∗
τ ) =
(cid:17)
if ϕ ∈ Γl
if ϕ ∈ Γn+k \ Γl.
(I − P ∗
(0,l])e∗
ϕ
ϕ) are given by the inductive hypothesis and property (2i) of S∗. The
ϕ) where ϕ ∈ Γn+k \ Γl. We observe that since
(0,l]e∗
θ) and S∗(e∗
S∗(e∗
slightly harder term to deal with is S∗(P ∗
ϕ ∈ (cid:96)1(Γn+k) we can write
e∗
e∗
ϕ =
(cid:88)
By the inductive hypothesis, we get
ω∈Γn+k\{γ,γ}
(cid:88)
αωd∗
ω + αγd∗
γ + αγd∗
γ
S∗(e∗
ϕ) =
αωd∗
F (ω) − αγd∗
γ − αγd∗
γ
ω∈Γn+k\{γ,γ}
and, using the fact that F is rank preserving, we find that
P ∗
(0,l](S∗(e∗
ϕ)) =
In either case, it is now easily seen by a similar computation that P ∗
From this computation and property 2(i) of the inductive hypothesis, we find that
(0,l]e∗
ϕ).
F (ω) − αγd∗
γ − αγd∗
γ
if l < n + 1
if n + 1 ≤ l < n + k.
ϕ)) = S∗(P ∗
(0,l](S∗(e∗
F (ω)
αωd∗
(cid:40)(cid:80)
(cid:80)
ω∈Γl
ω∈Γl\{γ,γ} αωd∗
(cid:40)
S∗P ∗
(0,l]e∗
ϕ =
P ∗
(0,l](e∗
F (ϕ))
(0,l](−e∗
P ∗
F (ϕ))
if ϕ /∈ {γ, γ}
if ϕ ∈ {γ, γ}.
Note that since θ ∈ Γl, it not possible to have ϕ ∈ Γn+k \ Γl and both θ and ϕ in {γ, γ}.
So when ϕ ∈ Γn+k \ Γl, it is apparent from our computations that there are precisely 3
51
possibilities to consider when computing S∗(c∗
(0,l])e∗
S∗(c∗
ϕ
τ ) = S∗(cid:16)
=
(cid:110)
θ + (cid:48)b(I − P ∗
ae∗
(cid:110)
F (θ) + (cid:48)b
(cid:110)
F (θ) + (cid:48)b
F (θ) + (cid:48)b
(I − P ∗
(I − P ∗
(I − P ∗
ae∗
−ae∗
ae∗
In the easier case where ϕ ∈ Γl, we have
τ ) = aS∗(e∗
S∗(c∗
θ) =
τ ). We obtain
(cid:111)
(cid:17)
(0,l])(e∗
F (ϕ))
(0,l])(e∗
(0,l])(−e∗
(cid:40)
ae∗
−ae∗
F (θ)
F (θ)
F (ϕ))
F (ϕ))
(cid:111)
(cid:111)
θ, ϕ /∈ {γ, γ}
θ ∈ {γ, γ}, ϕ /∈ {γ, γ}
θ /∈ {γ, γ}, φ ∈ {γ, γ}.
if θ /∈ {γ, γ}
if θ /∈ {γ, γ}.
Finally, we note that since F is rank-preserving, all of the above formulae are equal to
Ξ for some suitably chosen Ξ ∈ ∆n+k+1. We can therefore define F on ∆n+k+1 so that
c∗
S∗(c∗
F (τ ). Our above calculations show that in order to achieve this, the correct
definition of F is F ((n + k + 1, l, θ, ϕ, , (cid:48))) =
τ ) = c∗
(n + k + 1, l, F (θ), ϕ, , (cid:48))
(n + k + 1, l, F (θ), ϕ,−, (cid:48))
(n + k + 1, l, F (θ), F (ϕ), , (cid:48))
(n + k + 1, l, F (θ), F (ϕ),−, (cid:48))
(n + k + 1, l, F (θ), F (ϕ), ,−(cid:48))
if ϕ ∈ Γl and θ /∈ {γ, γ}
if ϕ ∈ Γl and θ ∈ {γ, γ}
if ϕ ∈ Γn+k \ Γl and θ, ϕ /∈ {γ, γ}
if ϕ ∈ Γn+k \ Γl and θ ∈ {γ, γ}, ϕ /∈ {γ, γ}
if ϕ ∈ Γn+k \ Γl and θ /∈ {γ, γ}, ϕ ∈ {γ, γ}.
F : Γn+k+1 → Γn+k+1 is easily seen to be a bijection. We now extend the definition of S∗ by
defining S∗(d∗
for any τ ∈ ∆n+k+1. Thus we have extended F and S∗ and both maps have all the desired
properties that were stated earlier.
F (τ ) for τ ∈ ∆n+k+1. Since e∗
τ it is clear that S∗(e∗
τ ) := d∗
τ ) = e∗
τ = c∗
τ + d∗
By induction, we have thus succeeded in defining a linear map S∗ : c00(Γ) → (cid:96)1(Γ). It
F (τ ) for all τ ∈ Γ (with a - sign only when
is bounded since, by construction, S∗(e∗
τ ∈ {γ, γ}). So (cid:107)S∗(e∗
τ )(cid:107) = 1 for all τ ∈ Γ and therefore S∗ is bounded by our earlier
remark. It follows that S∗ extends to a bounded linear map on (cid:96)1(Γ). In fact, S∗ is an
F (τ ) for all τ ∈ Γ, S∗
isometry onto (cid:96)1(Γ). Indeed, since F is a bijection, and S∗(e∗
τ ∈ (cid:96)1(Γ), then
τ ) = ±e∗
τ ) = ±e∗
F (τ )
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
aτ e∗
F (τ ) − aγe∗
γ − aγe∗
γ
aτ + − aγ + − aγ
aτ e∗
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)S∗
(cid:32)(cid:88)
is certainly onto. Moreover, suppose(cid:80)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =
τ∈Γ aτ e∗
(cid:88)
(cid:88)
(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
τ∈Γ\{γ,γ}
aτ
τ∈Γ
τ∈Γ
=
=
=
τ
aτ e∗
τ
τ∈Γ\{γ,γ}
τ∈Γ
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
52
where we have repeatedly made use of the fact that F : Γ → Γ is a bijection, so e∗
F (τ(cid:48))
when τ (cid:54)= τ(cid:48). It follows that the dual operator (S∗)∗ : (cid:96)1(Γ)∗ → (cid:96)1(Γ)∗ is also an isometry.
For convenience, we drop the brackets and just write S∗∗ for (S∗)∗.
F (τ ) (cid:54)= e∗
Recall that dτ ∈ (cid:96)1(Γ)∗ denotes the usual biorthogonal vector (as in Section 2.3) of the
basis (d∗
θ)θ∈Γ of (cid:96)1(Γ). Using the fact that
d∗
−d∗
S∗(d∗
τ ) =
F (τ )
F (τ )
(cid:40)
(cid:40)
it is elementary to check that
S∗∗(dω) =
dF −1(ω)
−d∗
F −1(ω)
if τ /∈ {γ, γ}
if τ ∈ {γ, γ}
if ω /∈ {γ, γ}
if ω ∈ {γ, γ}
Recalling from Section 2.3 that the Bourgain-Delbaen space is precisely X = [dγ : γ ∈ Γ],
we see that S∗∗ restricts to give an onto isometry S := S∗∗ : X → X. We shall call S a
basic operator. By the previous observation, we get
Lemma 2.5.1. Let(cid:80)
ω αωdω ∈ X. Then
S
where
(cid:32)(cid:88)
ω
αω =
αωdω
(cid:40)
αω
−αω
(cid:33)
(cid:88)
ω
=
αωdF −1(ω)
if ω /∈ {γ, γ}
if ω ∈ {γ, γ}
Of course, the definition of S∗, and consequently S, depends on the initial choice of γ ∈
∆n+1. Moreover, the construction of S∗ is very much dependent on the bijection F : Γ → Γ,
which also depends on our initial choice of γ ∈ ∆n+1. We define some terminology that will
be useful later.
Definition 2.5.2. Suppose S is constructed as above by fixing γ ∈ ∆n+1. We shall say
that S has rank n and that γ is the base of S. We write base (S) = γ and rank (S) = n.
We shall say the bijection F : Γ → Γ, used in the construction and definition of S∗, is the
underlying function on Γ.
We will usually denote a rank n basic operator by Sn and its underlying function on Γ
by Fn.
Lemma 2.5.3. Let m < n and suppose Sm, Sn : X → X are basic operators of ranks
θ)θ∈Γ of (cid:96)1(Γ) such that
m and n respectively.
X = [dθ : θ ∈ Γ], then
If λ is the basis constant of the basis (d∗
(cid:107)Sn − Sm(cid:107) ≥ 1
2λ
53
Proof. Let γ = base (Sm) (so in particular rank (γ) = m + 1) and let Fn : Γ → Γ be the
the underlying function on Γ for Sn. So by Lemma 2.5.1, Sn(dγ) = dF −1
n (γ) = dγ since
FnΓn = IdΓn, and Sm(dγ) = −dγ. We recall that (cid:107)dτ(cid:107) = 1 for all τ ∈ Γ (see Lemma 2.2.5
and Remark 2.3). (This was also proved using a different method in [17].) Using this fact,
we can estimate as follows.
(cid:107)Sn − Sm(cid:107) ≥ (cid:107)Sn(dγ) − Sm(dγ)(cid:107)
= (cid:107)dγ + dγ(cid:107)
≥ dγ((cid:107)d∗
= (cid:107)d∗
γ(cid:107)−1d∗
γ(cid:107)−1 ≥ 1
2λ
γ(cid:107)−1d∗
γ)
γ) + dγ((cid:107)d∗
.
To obtain the final inequality, we have made use of the fact that for all τ ∈ Γ, (cid:107)d∗
(cid:107)e∗
τ(cid:107) ≤ 1 + λ ≤ 2λ.
τ − P ∗
(0,rankτ )e∗
τ(cid:107) =
We therefore have at least a countably infinite number of operators on X, all of which
are pairwise distance at least 1 apart. To establish the non-separability of L(X) we need
to work a bit harder.
2.5.2 Non-separability of L(X)
In the final section of this chapter, we establish the non-separability of L(X). Throughout
this section, λ is the same constant as appearing in the previous lemma. The idea is to
construct an uncountable set of operators on X, all of which are pairwise distance 1/2λ
apart by taking suitable compositions of basic operators of increasing ranks. For each
n ∈ N, n ≥ 2, we fix a γn ∈ ∆n+1 and let Sn be the basic operator (as constructed in
Section 2.5.1) with base (Sn) = γn. We work with these operators for the remainder of the
section. The underlying function on Γ for Sn will be denoted by Fn. We recall that each
n : (cid:96)1(Γ) → (cid:96)1(Γ) and restricting it to
Sn is obtained by taking the dual of some operator S∗
X.
Before progressing any further, we need
Proposition 2.5.4. Let (nj)∞
n1 ≥ 2. For p ≤ m, (p, m ∈ N) denote by T ∗
nm ◦ ··· ◦ S∗
S∗
Proof. It is enough to show (T ∗
np+1 ◦ S∗
1,nx∗)∞
j=1 be a strictly increasing sequence of natural numbers with
p,m :=
p,m the composition of basic operators T ∗
np. For any x∗ ∈ (cid:96)1(Γ), the sequence (T ∗
1,nx∗)∞
n=1 converges in (cid:96)1(Γ).
n=1 is a Cauchy sequence. We note that for k > l,
(cid:13)(cid:13)T ∗
1,lx∗(cid:13)(cid:13) =(cid:13)(cid:13)T ∗
1,kx∗ − T ∗
1,lx∗(cid:13)(cid:13)
l+1,k(T ∗
1,lx∗) − T ∗
54
We recall that for any p ∈ N, S∗
is strictly increasing, it is easily seen that T ∗
(Schauder) basis for (cid:96)1(Γ), we can write
T ∗
1,lx∗ =
So
(cid:13)(cid:13)T ∗
1,lx∗(cid:13)(cid:13) =
1,kx∗ − T ∗
p(cid:96)1(Γp) = Id(cid:96)1(Γp). Therefore, since the sequence (nj)∞
j=1
ω)ω∈Γ is a
ω
ω∈Γ
ω d∗
α(l)
l+1,k(cid:96)1(Γnl+1 ) = Id(cid:96)1(Γnl+1 ). Since (d∗
(cid:88)
(cid:17) −(cid:88)
(cid:16) (cid:88)
(cid:13)(cid:13)(cid:13)
(cid:17) −(cid:88)
ω∈Γ\Γnl+1
ω d∗
α(l)
ω d∗
α(l)
(cid:13)(cid:13)(cid:13)
ω d∗
α(l)
ω∈Γ
ω∈Γ
l+1,k
ω
ω
ω
(cid:13)(cid:13)(cid:13)
l+1,k
(cid:13)(cid:13)(cid:13)T ∗
(cid:16)(cid:88)
(cid:13)(cid:13)(cid:13) (cid:88)
(cid:13)(cid:13)(cid:13) (cid:88)
ω∈Γnl+1
ω∈Γ\Γnl+1
=
≤ 2
ω
ω d∗
α(l)
ω + T ∗
ω∈Γ
ω d∗
α(l)
ω d∗
α(l)
ω
1,l − P ∗
(0,nl+1]T ∗
where the final inequality is obtained from the triangle inequality and the fact that T ∗
l+1,k
is a composition of isometries, therefore has norm 1. To prove the proposition, it is enough
to show (cid:107)(T ∗
To see this, we consider z∗ = (cid:80)
1,l)x∗(cid:107) → 0 as l → ∞.
ω∈Γ δωd∗
ω to be any vector in (cid:96)1(Γ). For any j, m ∈
N, since Fm preserves the rank of all elements in Γ, an easy computation yields that
P ∗
mz∗ = S∗
(0,j]S∗
m operators, the same equality
l,k. for any l, k ∈ N. Using this observation, it is now easy to
holds with S∗
see that (cid:107)(T ∗
(0,j]z∗. Consequently, as a composition of S∗
1,l)x∗(cid:107) → 0 as l → ∞. Indeed,
1,l)x∗(cid:107) = (cid:107)T ∗
1,l(I − P ∗
(0,nl+1])x∗(cid:107) = (cid:107)(I − P ∗
(0,nl+1])x∗(cid:107).
mP ∗
m replaced by T ∗
1,l − P ∗
1,l − P ∗
(cid:107)(T ∗
(0,nl+1]T ∗
(0,nl+1]T ∗
We have made use of the fact that the operators T ∗
1,l are isometries on (cid:96)1(Γ) to obtain the
final equality. The right-hand-side of the above expression clearly converges to 0 as l → ∞
since the sequence (nl)∞
l=1 → ∞ and the P ∗
(0,nl+1] operators are basis-projections.
Using the above observation, it is easy to prove the following corollary:
Corollary 2.5.5. Let (nj)∞
n1 ≥ 2, and let T ∗
linear operator T ∗
j=1 be a strictly increasing sequence of natural numbers with
1,n (n ∈ N) be defined as in Proposition 2.5.4. Then we may well-define a
(nj )∞
j=1
: (cid:96)1(Γ) → (cid:96)1(Γ) by
T ∗
(x∗) = lim
(nj )∞
j=1
j→∞ T ∗
1,j(x∗)
(x∗ ∈ (cid:96)1(Γ))
Moreover, T ∗
(nj )∞
j=1
: (cid:96)1(Γ) → (cid:96)1(Γ) is an onto isometry.
We are finally ready to prove the main result of this chapter.
55
Theorem 2.5.6. Given a strictly increasing sequence of natural numbers (nj)∞
2, the dual operator of the operator T ∗
X. Moreover, if (nj)∞
j=1 with n1 ≥
: X →
(nj )∞
j=1 are two different strictly increasing sequences, with
restricts to give an onto isometry, T(nj )∞
j=1
j=1
j=1 and ((cid:102)nj)∞
nk =(cid:102)nk for k = 1, 2 and n1 ≥ 2 then(cid:13)(cid:13)(cid:13)T(nj )∞
j=1
(cid:13)(cid:13)(cid:13) ≥ 1
2λ
− T((cid:102)nj )∞
j=1
Since the set of all strictly increasing sequences of natural numbers with the first two terms
being fixed is uncountable, it follows that L(X) is non-separable.
Proof. The dual operator of T ∗
is an onto isometry of (cid:96)∞(Γ) by standard duality
arguments. So for the first part of the proof, we must only see that it maps X into (and
onto) X. We fix a θ ∈ Γ and show (T ∗
(nj ))∗(dθ)](d∗
(nj ))∗(dθ) ∈ [dγ : γ ∈ Γ]. For τ ∈ Γ, we have
(cid:0)T ∗
(nj )(d∗
τ ) = dθ
[(T ∗
(nj )∞
j=1
τ )(cid:1)
(cid:0)T ∗
(cid:0)S∗
τ )(cid:1)
1,j(d∗
nj ◦ ··· ◦ S∗
n1(d∗
τ )(cid:1)
= lim
j→∞ dθ
j→∞ dθ
= lim
We let m be mininal such that τ ∈ Γnm (i.e. τ /∈ Γnm−1, so in particular rank (τ ) > nm−1)
nj ◦···◦ S∗
and consider 3 possibilities:
n1(d∗
(1) m = 1. In this case S∗
(2) m = 2. Now we have that ∀k ≥ 1, S∗
n1d∗
τ )
Fn1 (τ ))
(cid:40)dθ(d∗
(nj ))∗(dθ)](d∗
τ ) = dθ(S∗
[(T ∗
τ ) = d∗
nk
=
−dθ(d∗
Fn1 (τ ))
τ so it follows that [(T ∗
◦ ··· ◦ S∗
n1(d∗
τ ) = S∗
n1(d∗
(nj ))∗(dθ)](d∗
τ ) = dθ(d∗
τ ).
τ ). It follows that
if τ (cid:54)= base S∗
otherwise.
n1 or Fn1(base S∗
n1)
(3) m > 2. It is seen that ∀k ≥ m − 1, S∗
(nj ))∗(dθ)](d∗
[(T ∗
nk
τ ) = dθ
(cid:16)
◦ ··· ◦ S∗
n1(d∗
τ ) = S∗
nm−1 ◦ ··· ◦ S∗
S∗
nm−1 ◦ ··· ◦ S∗
n1(d∗
τ )
.
n1(d∗
τ ). So
τ ) = d∗
Since rank (τ ) > nm−1 > nm−2 ··· > n1 and we are assuming m > 2, S∗
S∗
n1(d∗
We consider two sub-cases:
(a) Either base (S∗
nm−1) = Fnm−2 ◦ ··· ◦ Fn1(τ ) or Fnm−1
Fnm−2◦···◦Fn1 (τ ) and so dθ
nm−1◦···◦S∗
nm−1(d∗
n1(d∗
nm−1
nm−2 ◦ ··· ◦
Fnm−2◦···◦Fn1 (τ ))(cid:1).
(cid:1) = Fnm−2 ◦ ··· ◦
(cid:0)S∗
(cid:17)
(cid:0)S∗
τ )(cid:1) = dθ
(cid:0)base S∗
Fn1(τ ). We note that in particular this implies rank (τ ) = nm−1 + 1. In this case,
nm−1(d∗
(nj ))∗(dθ)](d∗
S∗
−dθ(d∗
Fnm−1◦Fnm−2◦···◦Fn1 (τ ). It follows that [(T ∗
Fnm−2◦···◦Fn1 (τ )) = −d∗
Fnm−1◦Fnm−2◦···◦Fn1 (τ )).
τ ) =
56
(b) Both base Sn∗
m−1
and Fnm−1(base S∗
nm−1) are not equal to Fnm−2◦···◦Fn1(τ ). An ar-
gument similar to the previous case shows [(T ∗
(nj ))∗(dθ)](d∗
We note that in all possible cases, [(T ∗
ω ∈ (cid:96)1(Γ) we have
given x∗ =(cid:80)
ω∈Γ αωd∗
τ ) = dθ(d∗
(nj ))∗(dθ)](d∗
τ ) (cid:54)= 0 only if rank (θ) = rank (τ ). So
Fnm−1◦Fnm−2◦···◦Fn1 (τ ))
[(T ∗
(nj ))∗(dθ)](x∗) = [(T ∗
(nj ))∗(dθ)](cid:0)(cid:88)
ω∈Γ
αωd∗
ω
(cid:1) =
=
(cid:88)
(cid:88)
ω∈Γ
rank (ω)=rank (θ)
ω∈Γ
rank (ω)=rank (θ)
αω[(T ∗
(nj ))∗(dθ)](d∗
ω)
λωdω(x∗)
(nj ))∗(dθ)](d∗
(nj ))∗(dθ) ∈ X as it is a finite linear combination of the dγ. Thus (T ∗
ω) and we have made use of the obvious fact that αω = dω(x∗).
(nj ))∗
where λω := [(T ∗
It follows that (T ∗
does indeed restrict to give an isometry from X into X. One can do similar computations
to verify that (T ∗
in the theorem, is an onto isometry as required.
(nj ))∗ maps X onto X. It follows that the operator T(nj ) : X → X, defined
the same first two terms and such that n1 ≥ 2. We choose k minimal s.t. nk (cid:54)= (cid:102)nk and
w.lo.g. assume nk >(cid:102)nk. We note that by the minimality of k we have nj =(cid:102)nj whenever
j=1 are two strictly increasing sequences which have
j=1 and ((cid:102)nj)∞
We suppose now (nj)∞
j < k and that by assumptions on the sequences, k > 2. We note that for any θ, τ ∈ Γ
(cid:13)(cid:13)(cid:13)T(nj ) − T((cid:102)nj )
(cid:13)(cid:13)(cid:13) ≥(cid:12)(cid:12)(cid:12)[T(nj )(dθ)]((cid:107)d∗
τ(cid:107)−1d∗
τ ) − [T((cid:102)nj )(dθ)]((cid:107)d∗
τ(cid:107)−1d∗
τ )
(cid:12)(cid:12)(cid:12).
We will find specific θ, τ ∈ Γ such that the right hand side of the above inequality is equal
to (cid:107)d∗
τ(cid:107)−1, which will complete the proof. Our previous calculations show that
[T(nj )(dθ)](d∗
τ ) = dθ(±d∗
Fnm−1◦···◦Fn1 (τ ))
where m ∈ N is minimal such that τ ∈ Γnm (and it is assumed that τ is chosen s.t.
τ is chosen in Γ \ Γn2). Moreover we have a minus sign precisely when
m > 2, i.e.
base Snm−1 = Fnm−2 ◦ ··· ◦ Fn1(τ ) or Fnm−1
(cid:1) = Fnm−2 ◦ ··· ◦ Fn1(τ ). Similarly,
(cid:0)base Snm−1
[T((cid:102)nj )(dθ)](d∗
τ ) = dθ(±d∗
◦···◦F(cid:102)n1
(τ ))
F(cid:94)np−1
(cid:1) = F(cid:94)np−2
2). Moreover we have a minus sign precisely when base S(cid:94)np−1
F(cid:94)np−1
where p ∈ N is minimal such that τ ∈ Γ(cid:102)np (and it is assumed that τ is chosen s.t. p >
(cid:0)base S(cid:94)np−1
◦ ··· ◦ F(cid:102)n1(τ ) or
We will in fact choose τ such that rank (τ ) =(cid:102)nk + 1. It is easily seen that with a choice
above must in fact be equal to k since(cid:102)nk < nk =⇒ nk ≥(cid:102)nk + 1 and thus τ ∈ Γ(cid:102)nk+1 ⊆ Γnk ,
of τ like this, the 'p' above must be equal to k + 1. On the other hand, the choice of m
◦ ··· ◦ F(cid:102)n1(τ ).
= F(cid:94)np−2
57
whilst τ /∈ Γnk−1 = Γ(cid:94)nk−1
. Since we have observed k > 2 our above calculations remain
valid, and we find that for such a choice of τ we have
[T((cid:102)nj )(dθ)](d∗
[T(nj )(dθ)](d∗
τ ) = dθ(±d∗
F(cid:102)nk
τ ) = dθ(±d∗
Fnk−1◦···◦Fn1 (τ ))
◦···◦F(cid:102)n1
(τ ))
(cid:0) − d∗
where we have a minus sign precisely in the cases described above. Since all the Fm's are
rank preserving, a perfectly good choice of τ fulfulling the condition that rank (τ ) =(cid:102)nk + 1
is τ = F −1(cid:102)n1
(base S(cid:102)nk ). If we now choose θ = F(cid:102)nk (base S(cid:102)nk ) we see that
[T((cid:102)nj )(dθ)](d∗
[T(nj )(dθ)](d∗
τ ) = dθ
τ ) = dθ(±d∗
(cid:1) = −1
◦ ··· ◦ F −1
(cid:94)nk−1
where to establish the final equalities we made use of the facts that(cid:102)nj = nj for all j < k
and θ = F(cid:102)nk (base S(cid:102)nk ) (cid:54)= base S(cid:102)nk by the very construction of F(cid:102)nk . So, for these choices of
θ, τ we finally get(cid:13)(cid:13)(cid:13)T(nj ) − T((cid:102)nj )
(cid:13)(cid:13)(cid:13) ≥(cid:12)(cid:12)(cid:12)[T(nj )(dθ)]((cid:107)d∗
τ ) − [T((cid:102)nj )(dθ)]((cid:107)d∗
τ(cid:107)−1d∗
τ )
τ(cid:107)−1d∗
F(cid:102)nk
base S(cid:102)nk
(cid:12)(cid:12)(cid:12)
(base S(cid:102)nk
)
) = 0
τ(cid:107)−1
= (cid:107)d∗
≥ 1
2λ
as required.
We conclude this chapter by observing that we can use the non-separability result just
obtained to provide a short proof that no Bourgain-Delbaen space of class Y can have the
scalar-plus-compact property.
Corollary 2.5.7. There exists no X ∈ Y for which every bounded operator on X is a scalar
multiple of the Identity plus a compact operator.
Proof. It was shown in [7] that for any Bourgain-Delbaen space X ∈ Y, X∗ is isomorphic
to (cid:96)1. Moreover, we have seen that all spaces of Bourgain-Delbaen type have a Schauder
basis, and consequently have the approximation property. For any Banach space Y with
separable dual and approximation property, the space of compact operators on Y , K(Y ) is
separable, and thus so also is the set of all operators of the form λIY +K (λ ∈ R, K ∈ K(Y )).
Therefore if X ∈ Y had the scalar-plus-compact property, its operator algebra would be
separable. This contradicts Theorem 2.5.6.
58
Whilst our non-separability result allows us to give a nice proof that the spaces in Y
do not have the scalar-plus-compact property, we remark that the ideas in this section can
be suitably modified to produce much stronger results. Indeed, one can use similar ideas
to construct non-trivial projections on any of the Bourgain-Delbaen spaces X in X ∪ Y.
(A non-trivial projection is one which has both infinite dimensional kernel and image.) It
follows that X does not have the property that every bounded linear operator on X is a
strictly singular perturbation of a scalar operator. Obviously this implies that the previous
corollary holds for any of the original Bourgain-Delbaen spaces, not just those of class Y.
59
Chapter 3
Spaces with few but not very few
operators
3.1 The Main Theorem
Having looked in detail at the Bourgain-Delbaen construction in the previous chapter, we
move on to consider a problem posed by Argyros and Haydon in [2]. We begin by introducing
the following definition.
Definition 3.1.1. Let X be a Banach space. We will say X has few operators if every
operator from X to itself is a strictly singular perturbation of the identity, that is, expressible
as λI + S for some strictly singular operator S : X → X. We say X has very few operators
if every operator from X to itself is a compact perturbation of the identity.
The fact that Banach spaces with few operators exist was first established in 1993 by
Gowers and Maurey in [16]. The existence of a Banach space with very few operators was
shown much more recently in [2]. In addition to their obvious intrinsic interest, such Banach
spaces exhibit remarkable Banach space structure, as discussed in the introduction to this
thesis. The motivation behind this chapter comes from a question that naturally arises from
the work of Argyros and Haydon in [2].
We recall once again that the Banach space XAH, constructed by Argyros and Haydon
in [2], is done so using the generalised Bourgain-Delbaen construction, as discussed in the
previous chapter. Consequently, it has a Schauder basis and is a separable L∞ space. It
was seen in [2] that the control over the finite dimensional subspace structure provided by
the L∞ property was essential in proving that XAH has very few operators and we will
use the same type of argument later in this chapter. Moreover, one can exploit the L∞
structure to show that the dual space of XAH is (cid:96)1.
In light of the previous discussion, it is natural to conjecture that the ideals of strictly
singular and compact operators on a separable L∞ space with (cid:96)1 dual coincide. Clearly this
60
must be true for the space XAH. Moreover, if one considers the space c0, the most obvious
example of a separable L∞ space with (cid:96)1 dual, it is once again found that the conjecture
holds. Indeed, it is well known (see, e.g. [1, Theorem 2.4.10]) that a bounded linear operator
defined on c0 is compact ⇐⇒ it is weakly compact ⇐⇒ it is strictly singular. More
generally, for spaces X with (cid:96)1 dual, the weakly compact and norm compact operators from
X to itself must coincide as a consequence of (cid:96)1 having the Schur property; this provides
further hope that the conjecture holds if we demand (cid:96)1 duality.
On the other hand, if we instead look at operators defined on X, a separable L∞ space
with (cid:96)1 dual, but mapping into a different target space, it is certainly possible for us to
construct strictly singular operators that are not weakly compact (and hence not compact).
To see this we need the following lemma.
Lemma 3.1.2. Let X be a separable Banach space and suppose (cid:96)1 embeds isomorphically
into X∗. Then there is a quotient operator Q : X → c0.
Proof. Since (cid:96)1 embeds into X∗, we can find a sequence (y∗
n) in the unit ball, BX∗, of
X∗ equivalent to the canonical basis (en) of (cid:96)1. Since X is separable, the weak* topology
restricted to the BX∗ is metrizable. It follows from this (and weak∗ compactness of BX∗)
that we may assume (passing to a subsequence of the (y∗
n) if necessary) that the sequence
2n−1 is weak∗
n) is weak∗ convergent. Now the sequence (x∗
(y∗
n) ∼ (en) ⊆ (cid:96)1 (recall every seminormalized block basic sequence of
null and moreover, (x∗
(cid:96)1 is equivalent to the canonical basis of (cid:96)1, see, e.g. [1, Lemma 2.1.1, Remark 2.1.2]). Let
0 ≡ (cid:96)1 → X∗ be the isomorphic embedding which maps the basis vector en of (cid:96)1 to
T : c∗
n. We claim that T is weak∗-weak∗ continuous. Indeed, suppose ((aβ
x∗
n=1)β∈I is a weak*
(cid:80)∞
nξn →
convergent net in c∗
n=1 aβ
n=1 αnξn. Now, for any x ∈ X, the sequence (x∗
nx) ∈ c0 since the sequence (x∗
n) is weak∗
∞(cid:88)
n=1 ∈ c0, (cid:80)∞
n) = (αn), i.e. whenever (ξn)∞
n) ⊆ X∗ where x∗
0 with limβ(aβ
null. It follows that
n := y∗
2n − y∗
nx = T(cid:0)(αn)(cid:1)x
αnx∗
n)∞
∞(cid:88)
T(cid:0)(aβ
n)(cid:1)x =
T(cid:0)(αn)(cid:1) as required.
n=1
nx∗
aβ
nx →
n=1
i.e. T(cid:0)(aβ
n)(cid:1) (cid:42)w∗
0 → X∗
is the dual of some operator Q : X → c0. Moreover, by Lemma 1.4.11, Q is a quotient
operator.
It follows by Lemma 1.4.12 that T : c∗
Corollary 3.1.3. Suppose X is an L∞ space of Bourgain-Delbaen type, i.e. obtained from
one of the constructions discussed in the previous chapter. If X has no subspace isomorphic
to c0 then there exists a strictly singular operator Q : X → c0 which fails to be weakly
compact.
61
Proof. We have seen in the previous chapter that all the L∞ spaces X of Bourgain-Delbaen
type are separable (they have a Schauder basis) and moreover, that (cid:96)1 embeds into the dual
space. Consequently, by the above lemma, there is a quotient operator Q : X → c0.
Suppose for contradiction that Q is weakly compact. Since Q is a quotient operator, it
X ) ⊆
so that Bc0 is weakly compact (it is a
). This contradicts the fact that
follows from the Open Mapping Theorem, that there is an M such that B◦
M Q(BX ). Therefore, Bc0 = B◦
weakly closed subset of the weakly compact set M Q(BX )
c0 is not reflexive, so Q cannot be weakly compact.
c0 ⊆ M Q(B◦
w
w ⊆ M Q(BX )
w
c0
Since X has no subspace isomorphic to c0 it follows that Q is strictly singular. To see
this, suppose there is an infinite dimensional subspace Y on which Q is an isomorphism.
Then, by [1, Proposition 2.2.1] the subspace Q(Y ) contains a subspace Z isomorphic to
c0. The image (QY →Q(Y ))−1(Z) is then a subspace of X isomorphic to c0, giving us a
contradiction.
We remark that there do exist L∞ spaces of Bourgain-Delbaen type which have no
subspace isomorphic to c0 and therefore satisfy the hypotheses of the previous corollary.
Indeed, it is immediate from the HI property that the Arygros-Haydon space is one such
example. In fact, it was shown in [7], that if X ∈ X , Y ∈ Y, where X ,Y are the original
classes of Bourgain-Delbaen spaces, then every infinite dimensional subspace of X contains
a subspace isomorphic to (cid:96)1 and every infinite dimensional subspace of Y contains an infinite
dimensional reflexive subspace. It follows from these facts and elementary results in Banach
space theory that the original Bourgain-Delbaen spaces also have no subspace isomorphic
to c0.
Of course, the previous corollary does not give us a counterexample to the conjecture
posed at the beginning of this chapter; the proof relies on exploiting properties of the
space c0 which features as the co-domain of the operator exhibited. We are interested in
operators from a space to itself. The purpose of this chapter is to exhibit a class of Banach
spaces which are genuine counterexamples to the previously stated conjecture. In fact, we
provide counterexamples to the following, stronger version of the original conjecture; 'must
a separable L∞ space with few operators and (cid:96)1 dual necessarily have very few operators?'.
This question was originally posed by Argyros and Haydon in [2, Problem 10.7]. Our main
result is the following:
Theorem 3.1.4. Given any k ∈ N, k ≥ 2, there is a separable L∞ space, Xk, with the
following properties:-
1. Xk is hereditarily indecomposable (HI) and X∗
k = (cid:96)1.
62
2. There is a non-compact bounded linear operator S : Xk → Xk. S is nilpotent of degree
k, i.e. Sj (cid:54)= 0 for 1 ≤ j < k and, Sk = 0.
3. Moreover, Sj (0 ≤ j ≤ k − 1) is not a compact perturbation of any linear combination
of the operators Sl, l (cid:54)= j.
4. The operator S : Xk → Xk is strictly singular (and consequently Sj is strictly singular
for all j ≥ 1).
5. Every operator T on Xk can be uniquely represented as T =(cid:80)k−1
λi ∈ R, (0 ≤ i ≤ k − 1) and K is a compact operator on Xk.
i=0 λiSi + K, where
It is immediate from the above properties (and the fact that the strictly singular oper-
ators are a closed ideal in the operator algebra - see Corollary 1.4.27) that the spaces Xk
have few operators but not very few operators. In other words, RI + K(Xk) (cid:36) L(Xk) ⊆
RI + SS(Xk) (where SS(Xk) is the space of strictly singular operators on Xk ). We thus
have a negative solution to Problem 10.7 of Argyros and Haydon ([2]).
In addition to solving the previously discussed problem, it turns out that there are a
number of other interesting consequences of these constructions, specifically concerning the
Calkin algebras and the structure of closed ideals in L(Xk). Since the details of the proof
of the main result are fairly long and technical, we choose to present these corollaries first.
3.2 Corollaries of the Main Theorem
3.2.1 On the structure of the closed ideals in L(Xk).
The structure of norm-closed ideals in the algebra L(X) of all bounded linear operators
on an infinite dimensional Banach space X is generally not well understood. For example,
classifying all the norm-closed ideals in L((cid:96)p ⊕ (cid:96)q) with p (cid:54)= q remains an open problem
(though some progress has been made in [27]). What is known, is that for the (cid:96)p spaces,
1 ≤ p < ∞, and c0, there is only one non-trivial closed ideal in L(X), namely the ideal of
compact operators. This was proved by Calkin, [8], for (cid:96)2 and then extended to (cid:96)p and c0
by Gohberg et al., [14]. More recently, the complete structure of closed ideals in L(X) was
described in [20] for X = (⊕∞
2 )(cid:96)1. In both cases, there
are exactly two nested proper closed ideals. Until the space constructed by Argyros and
2 )c0 and in [21] for X = (⊕∞
n=1(cid:96)n
n=1(cid:96)n
Haydon, [2], these were the only known separable, infinite dimensional Banach spaces for
which the norm-closed ideal structure of the operator algebra is completely known.
Clearly the space XAH of Argyros and Haydon provides another example of a separable
Banach space for which the ideal of compact operators is the only (proper) closed ideal in
the operator algebra. The spaces constructed in this chapter allow us to add to the list
63
of spaces for which the ideal structure of L(X) is completely known. In fact, we see that
we can construct Banach spaces for which the ideals of the operator algebra form a finite,
totally ordered lattice of arbitrary length. More precisely, the following is an immediate
consequence of our main theorem.
Theorem 3.2.1. There are exactly k norm-closed, proper ideals in L(Xk). The lattice of
closed ideals is given by
K(Xk) (cid:40) (cid:104)Sk−1(cid:105) (cid:40) (cid:104)Sk−2(cid:105) . . .(cid:104)S(cid:105) (cid:40) L(Xk).
Here, if T is an operator on Xk, (cid:104)T(cid:105) is the smallest norm-closed ideal in L(Xk) containing
T.
3.2.2 The Calkin algebra L(Xk)/K(Xk).
We note that as a consequence of properties (3) - (5) of the main theorem, the Calkin algebra
L(Xk)/K(Xk) is k dimensional with basis {I +K(Xk), S +K(Xk), . . . , Sk−1 +K(Xk)}. More
precisely, it is isomorphic as an algebra to the subalgebra A of k × k upper-triangular-
Toeplitz matrices, i.e. A is the subalgebra of Mat(k × k) generated by
0 1
0
j
1
. . .
. . .
. . . 1
0
: 0 ≤ j ≤ k − 1
An explicit isomorphism is given by ψ : L(Xk)/K(Xk) → A ∼= R[X]/(cid:104)xk(cid:105),
k−1(cid:88)
j=0
λjSj + K(Xk) (cid:55)→
λ0 λ1 λ2
0
···
λ0 λ1 λ2
0
...
...
0
0
...
...
0
λ0 λ1
. . .
0
...
0
···
··· λk−1
··· λk−2
...
. . .
...
. . .
...
. . .
λ0
0
We remark that the quotient algebra L(X)/SS(X) (where SS(X) denotes the strictly
singular operators on X) was studied by Ferenczi in [12] for X a HI, or more generally, a HDn
space. Here it was shown that for a complex HDn space, dim(L(X)/SS(X)) ≤ n2. The
above remarks show that the Calkin algebra of a HI space can behave somewhat differently
however; indeed any finite dimension for the Calkin algebra can be achieved. In the next
chapter, we will extend these results further still, and show that it is possible to obtain a
Banach space X∞ which has Calkin algebra isometric as a Banach algebra to (cid:96)1(N0).
64
3.2.3 Commutators in Banach Algebras
Rather unexpectedly, the space X2 also provides a counterexample to a conjecture of Pro-
fessor William B. Johnson; the author is thankful to Professor Johnson for bringing this
to his attention. To motivate the conjecture we recall the following classical theorem of
Wintner.
Theorem 3.2.2 (Wintner). The identity in a unital Banach algebra, A, is not a commu-
tator, that is, there do not exist x, y ∈ A such that 1 = [x, y] := xy − yx.
Proof. Suppose for contradiction the theorem is false and choose x, y ∈ A such that 1 =
xy − yx. We claim that xyn − ynx = nyn−1 for all n ∈ N. The proof is a simple induction;
clearly it is true for n = 1. Suppose that xyn − ynx = nyn−1 holds for some n ∈ N. Then
xyn+1 − yn+1x = (xyn − ynx)y + yn(xy − yx) = nyn−1 · y + yn · 1 = (n + 1)yn
completing the induction.
Observe that yn (cid:54)= 0 for any n, since otherwise there is a smallest value of n with
yn = 0. However, we now find that 0 = xyn − ynx = nyn−1, contradicting the minimality
of n. Finally, we have for any n
n(cid:107)yn−1(cid:107) = (cid:107)xyn − ynx(cid:107) ≤ 2(cid:107)x(cid:107)(cid:107)y(cid:107)(cid:107)yn−1(cid:107)
and since (cid:107)yn−1(cid:107) (cid:54)= 0 for all n, we have n ≤ 2(cid:107)x(cid:107)(cid:107)y(cid:107) for all n. This obvious contradiction
completes the proof.
If X is a Banach space and M is a proper norm-closed ideal of L(X), applying the above
theorem to the unital Banach algebra L(X)/M, it is easily seen that no operator of the
form λI + M with λ (cid:54)= 0 and M ∈ M is a commutator in the algebra L(X). In fact, as is
commented in [18], this is in general the only known obstruction for an operator in L(X) to
fail to be a commutator when X is any infinite dimensional Banach space. Consequently the
authors of [18] define a Wintner space as a Banach space X such that every non commutator
in L(X) is of the form λI + M where λ (cid:54)= 0 and M lies in a proper norm-closed ideal.
Johnson conjectured that every infinite dimensional Banach space is a Wintner space.
Until the existence of the spaces constructed in this thesis, all infinite dimensional Banach
spaces X for which the commutators in L(X) could be classified satisfied this conjecture.
However, we can easily see that X2 fails to be a Wintner space.
Lemma 3.2.3. X2 is not a Wintner space.
65
Proof. We note that the operator S ∈ L(X2) fails to be of the form λI + M with λ (cid:54)= 0 and
M lying in a proper norm-closed ideal of L(X2). Indeed, if S has such a form, we can write
S = λI + µS + K where λ (cid:54)= 0, since by Theorem 3.2.1 the only proper norm-closed ideals
in X2 are K(X2) and (cid:104)S(cid:105) = {µS + K : µ ∈ R, K ∈ K(X2)}. However, this contradicts the
fact that the vectors I + K(X2) and S + K(X2) are linearly independent in L(X2)/K(X2)
(see property (3) of the main theorem).
To complete the proof, it suffices to see that S is not a commutator; this is easy using
the operator representation, property (5), of the main theorem. Indeed suppose for con-
tradiction that there exist T1, T2 ∈ L(X2) with S = [T1, T2]. For i = 1, 2, we can write
Ti = λiI + µiS + Ki where λi, µi are scalars and Ki are compact operators. We then have
S = T1T2 − T2T1
= (λ1I + µ1S + K1)(λ2I + µ2S + K2) − (λ2I + µ2S + K2)(λ1I + µ1S + K1)
= K(cid:48)
for some compact operator K(cid:48). However, we know by property (2) of the main theorem
that S is not compact.Therefore S cannot be a commutator in L(X2).
3.2.4 Invariant subspaces.
Finally, we remark that it was shown in [2] that all operators on the Argyros-Haydon space
admit non-trivial, closed, invariant subspaces. The same is true for all the spaces Xk.
Indeed, by a result of Lomonosov (see [23]), if an operator T commutes with a non-zero
compact operator, then T has a proper, closed, invariant subspace. In particular, if there
is some polynomial of T which is compact and non-zero, then certainly T has a proper,
closed, invariant subspace.
For an operator T : Xk → Xk, T =(cid:80)k−1
j=0 λjSj + K, we consider the polynomial of T ,
given by P(T ) := (T − λ0I)k. It follows (by the ring isomorphism of the Calkin algebra
with the ring R[X]/(cid:104)xk(cid:105)) that P(T ) is a compact operator. So if P(T ) (cid:54)= 0 then we are
done by the result of Lomonosov. Otherwise it is clear that λ0 is an eigenvalue of T , so it
has a one dimensional invariant subspace.
3.3 The Basic Construction
The rest of this chapter is devoted to proving the main theorem. Before continuing any
further, we define the notation and terminology that shall be used throughout this chapter.
We construct our spaces Xk by modifying the ideas used by Argyros and Haydon in their
66
construction of the space XAH. Consequently, it is convenient for us to work with the same
notation and terminology that they introduced in [2].
We will be working with two strictly increasing sequences of natural numbers (mj) and
(nj) which satisfy the following assumptions.
Assumption 3.3.1. We assume that (mj, nj)j∈N satisfy the following:
1. m1 ≥ 4;
2. mj+1 = m2
j ;
3. n1 ≥ m2
1;
4. nj+1 ≥ (16nj)log2 mj+1 = m2
j+1(4nj)log2 mj+1.
We note that these are almost the same assumptions as in [2]; the only difference is that
j for all j ∈ N. (In [2], it was only assumed
we demand the stronger condition that mj+1 = m2
that mj+1 ≥ m2
j for all j.) Consequently, we have no problems drawing upon results from
[2]. In fact, for the purposes of this chapter, asking that mj+1 ≥ m2
j for all j ∈ N would
have been sufficient, but we need the slightly stronger condition in the following chapter.
Like the Argyros-Haydon space, our spaces Xk will be obtained from the generalised
Bourgain-Delbaen construction, described in the previous chapter (Theorem 2.2.2). Con-
I ⊆ N we define the projection PI : X →(cid:76)
tinuing with the same notation used in Theorem 2.2.2, we recall that the subspaces Mn =
[dγ : γ ∈ ∆n] form a finite-dimensional decomposition for X = X(Γ, τ ). For each interval
n∈I Mn in the natural way; this is consistent
with our use of P ∗
(0,n] in Theorem 2.2.2. As in [2], many of the arguments will involve
sequences of vectors that are block sequences with respect to this FDD. It will therefore be
useful to make the following definition; for x ∈ X, we define the range of x, denoted ran x,
to be the smallest interval I ⊆ N such that x ∈(cid:76)
n∈I Mn.
We recall from Theorem 2.2.2 that there is an 'ε parameter' appearing in the admissible
tuples of τ (γ) in the cases (0) and (2). Whilst this extra degree of freedom was required in
the previous chapter in order for us to unite the original and generalised Bourgain-Delbaen
constructions, we will not require it for the rest of this thesis; ε shall always be set equal
to 1. Consequently, we choose to simplify our notation and simply omit the ε appearing in
these tuples. This is also consistent with the notation from [2]. With this new notation,
if γ ∈ ∆n+1, we say the corresponding vector c∗
τ (γ) = (α, ξ) (in the notation of Theorem 2.2.2 this would have been τ (γ) = (1, α, ξ)).
Similarly, c∗
γ is called a Type 1 BD-functional if τ (γ) = (p, β, b∗) or a Type 2 BD-functional
if τ (γ) = (α, ξ, p, β, b∗) (see Theorem 2.2.2). We will call γ a Type 0, 1 or 2 element in each
of the respective cases.
γ ∈ (cid:96)1(Γ) is a Type 0 BD-functional
if
67
Our space will be a specific Bourgain-Delbaen space very similar to the space XAH
constructed in [2]. We adopt the same notation used in [2], in which elements γ of ∆n+1
automatically code the corresponding BD-functionals. Consequently, we can write X(Γ)
rather than X(Γ, τ ) for the resulting L∞ space. To be more precise, an element γ of ∆n+1
will be a tuple of one of the following forms:
1. γ = (n + 1, p, β, b∗),
2. γ = (n + 1, ξ, β, b∗) in which case τ (γ) = (1, ξ, rank ξ, β, b∗).
in which case τ (γ) = (p, β, b∗);
In each case, the first co-ordinate of γ tells us what the rank of γ is, that is to say to which set
∆n+1 it belongs, while the remaining co-ordinates specify the corresponding BD-functional.
We observe that BD-functionals of Type 0 do not arise in this construction. Moreover,
in the definition of a Type 2 functional, the scalar α that occurs is always 1 and p is
always equal to rank ξ. As in the Argyros-Haydon construction, we shall make the further
restriction that the β parameter must be of the form m−1
(nj) satisfy Assumption 3.3.1. We shall say that the element γ has weight m−1
of a Type 2 element γ = (n + 1, ξ, m−1
m−1
, where the sequences (mj) and
. In the case
, b∗) we shall insist that ξ be of the same weight,
, as γ.
j
j
j
j
To ensure that the sets ∆n+1 are finite we shall admit into ∆n+1 only elements of weight
mj with j ≤ n + 1. A further restriction involves the recursively defined function called
"age" (also defined in [2]). For a Type 1 element, γ = (n + 1, p, β, b∗), we define age γ = 1.
For a Type 2 element, γ = (n + 1, ξ, m−1
, b∗), we define age γ = 1 + age ξ, and further
restrict the elements of ∆n+1 by insisting that the age of an element of weight m−1
j may not
exceed nj. Finally, we shall restrict the functionals b∗ that occur in an element of ∆n+1 by
requiring them to lie in some finite subset Bn of (cid:96)1(Γn). It is convenient to fix an increasing
sequence of natural numbers (Nn) and take Bp,n to be the set of all linear combinations
η aη ≤ 1 and each aη is a rational number with denominator
dividing Nn!. We may suppose the Nn are chosen in such a way that Bp,n is a 2−n-net in
the unit ball of (cid:96)1(Γn \ Γp). The above restrictions may be summarized as follows.
η, where(cid:80)
b∗ =(cid:80)
η∈Γn\Γp
aηe∗
j
, b∗) : 0 ≤ p < n, b∗ ∈ Bp,n
(cid:111)
(n + 1, p, m−1
j
Assumption 3.3.2.
∆n+1 ⊆ n+1(cid:91)
∪ n−1(cid:91)
j=1
(cid:110)
p(cid:91)
(cid:110)
(n + 1, ξ, m−1
j
, b∗) : ξ ∈ ∆p, weight ξ = m−1
j
, age ξ < nj, b∗ ∈ Bp,n
p=1
j=1
As in [2] we shall also assume that ∆n+1 contains a rich supply of elements of "even
weight", more exactly of weight m−1
j with j even.
68
(cid:111)
∆n+1 ⊇
(n + 1, p, m−1
2j , b∗) : 0 ≤ p < n, b∗ ∈ Bp,n
(cid:111)
Assumption 3.3.3.
(cid:110)
(cid:98)(n+1)/2(cid:99)(cid:91)
(cid:110)
(cid:98)p/2(cid:99)(cid:91)
∪ n−1(cid:91)
j=1
n+1(cid:91)
∪ n−1(cid:91)
j=1
(cid:110)
p(cid:91)
(cid:110)
(cid:111)
(cid:111)
.
(n + 1, ξ, m−1
2j , b∗) : ξ ∈ ∆p, weight ξ = m−1
2j , age ξ < n2j, b∗ ∈ Bp,n
p=1
j=1
For the main construction, there will be additional restrictions on the elements with
"odd weight" m−1
2j−1, though we will come to these later. To begin with, we shall work
with a different class of Bourgain-Delbaen spaces, and therefore choose to slightly change
our notation from the above, reserving the above notation for the main construction. In
j=1Λj.
what follows, Υ will take the role of Γ above, and Λn the role of the ∆n's. In particular,
Υ = ∪∞
n=1Λn and Υn = ∪n
Now, we fix a k ∈ N, k ≥ 2 and define the space X(Υ) to be the Bourgain-Delbaen
n=1Λn and the Λn are finite sets defined by
space given as in Theorem 2.2.2, where Υ = ∪∞
recursion:
Definition 3.3.4. We define Υ by the recursion Λ1 = {0, 1, . . . (k − 1)},
Λn+1 =
(n + 1, p, m−1
, b∗) : 0 ≤ p < n, b∗ ∈ Bp,n
j
(cid:111)
(n + 1, ξ, m−1
j
, b∗) : ξ ∈ Λp, weight ξ = m−1
j
, age ξ < nj, b∗ ∈ Bp,n
p=1
j=1
Remark 3.3.5.
1. Note that the cardinality of Λ1 depends on the choice of k ∈ N. In
this way, Υ (and consequently the BD space X(Υ)) really depend on the chosen k ∈ N.
In an attempt to avoid even more complicated notation, we consider k ∈ N, k ≥ 2 to
be fixed for the remainder of the chapter.
2. Later on, we will want to take a suitable subset of Υ. To avoid any ambiguity in nota-
the set of all linear combinations b∗ =(cid:80)
tion, in the above definition, and throughout the rest of the chapter, Bp,n will denote
η aη ≤ 1
η, where, as before,(cid:80)
aηe∗
η∈Υn\Υp
and each aη is a rational number with denominator dividing Nn!.
Eventually, we want to have a non-compact, bounded linear operator S on our space.
The ideas used to construct this operator are very similar, but more involved, than those
used in Section 2.5. We will make use of a single element set which is disjoint from Υ, and
label the element 'undefined'. We obtain the following theorem.
69
Theorem 3.3.6. There is a map G : Υ → Υ ∪ {undefined} (we say G(γ) is undefined
if G(γ) = undefined, otherwise we say G(γ) is defined) and a norm 1, linear mapping
R∗ : (cid:96)1(Υ) → (cid:96)1(Υ) satisying:
1. G(j) = j − 1 for 1 ≤ j ≤ k − 1 and G(0) is undefined (recall Λ1 = {0, 1, . . . k − 1}).
2. For elements γ ∈ Υ\ Λ1 such that G(γ) is defined, rank γ = rank G(γ) and weight γ =
weight G(γ) (i.e. G preserves weight and rank). Moreover, age G(γ) ≤ age γ (G
doesn't increase age).
(cid:40)
(cid:40)
e∗
G(γ)
0
d∗
0
G(γ)
if G(γ) is defined
otherwise.
if G(γ) is defined
otherwise.
3.
4.
R∗(e∗
γ) =
R∗(d∗
γ) =
Proof. We will construct the maps G and R∗ inductively. We note that since R∗ will be
a linear operator on (cid:96)1(Υ), in order to ensure it is bounded, we only need to be able to
control (cid:107)R∗(e∗
γ)(cid:107) ≤ M
for every γ ∈ Υ, then it is elementary to check that R∗ is bounded with norm at most M .
In particular, if property (3) of Theorem 3.3.6 holds, it follows that (cid:107)R∗(cid:107) = 1.
γ)(cid:107) (for γ ∈ Υ). More precisely, if there is some M ≥ 0 s.t. (cid:107)R∗(e∗
To begin the inductive constructions of R∗ and G we define G : Λ1 → Λ1 by setting
γ = d∗
j−1 for j ≥ 1.
G(j) = j − 1 for 1 ≤ j ≤ k − 1 and declaring that G(0) is undefined. Noting that e∗
for γ ∈ Λ1, we define R∗(e∗
j−1 = d∗
We observe that this definition is consistent with the properties (1) - (4) above.
0) = 0 and R∗(e∗
j ) = R∗(d∗
0) = R∗(d∗
j ) = e∗
Suppose that we have defined G : Υn → Υn and R∗ : (cid:96)1(Υn) → (cid:96)1(Υn) satisfying
properties (1) - (4) above. We must extend G to Υn+1 and R∗ to a map on (cid:96)1(Υn+1).
We consider a γ ∈ Λn+1 and recall that (see Theorem 2.2.2) e∗
γ. We wish to
define R∗d∗
γ and R∗e∗
γ however we like.
γ ∈ (cid:96)1(Υn)), and we want R∗ to be linear, once we
However, since R∗c∗
have defined R∗d∗
γ. By linear independence, we are free to define R∗d∗
γ is already defined (c∗
γ, in order to have linearity we must have R∗e∗
γ + R∗d∗
γ.
γ = R∗c∗
γ = c∗
γ + d∗
γ
Let us consider R∗c∗
γ. We suppose first that age γ = 1 so that we can write γ =
(n + 1, p, β, b∗) where b∗ ∈ (cid:96)1(Υn \ Υp). Consequently
(cid:16)
(0,p]b∗(cid:17)
c∗
γ = β
I − P ∗
= βP ∗
(p,∞)b∗.
70
αδd∗
We claim that R∗P ∗
unique choice of αδ). It follows from property (4) and the inductive hypothesis that
(p,∞)b∗ = P ∗
δ ∈ Υn
δ (for a
(p,∞)R∗b∗. Indeed, we can write b∗ = (cid:80)
(p,∞)b∗ = R∗(cid:0) (cid:88)
R∗P ∗
(cid:1) =
αδd∗
δ
δ∈Υn\Υp
(cid:88)
αδd∗
G(δ)
δ∈Υn\Υp∩
{η ∈ Υn : G(η) is defined}
and it is easily checked (by a similar calculation) that we obtain the same expression for
P ∗
(p,∞)R∗b∗. It follows that
R∗c∗
γ = βP ∗
(p,∞)R∗b∗.
We define G(γ) by
G(γ) =
(cid:40)
undefined
(n + 1, p, β, R∗b∗)
(p,∞)R∗b∗ = 0
if P ∗
otherwise
where it is a simple consequence of the facts that R∗ : (cid:96)1(Υn) → (cid:96)1(Υn) has norm 1 and
satisfies property (3) that the element (n+1, p, β, R∗b∗) ∈ Υ. In the case where P ∗
(p,∞)R∗b∗ (cid:54)=
0, G(γ) is defined (with the definition as above) and it is evident that R∗c∗
G(γ). We
γ = c∗
can define R∗d∗
γ) that we
have R∗e∗
γ = e∗
γ = 0, so we can set
R∗d∗
γ = d∗
G(γ) as required. Otherwise, when P ∗
γ = 0.
G(γ) and it follows by linearity (and the fact that e∗
(p,∞)R∗b∗ = 0, R∗c∗
γ = 0 and again by linearity we get that R∗e∗
Now if γ has age > 1, we can write γ = (n + 1, ξ, β, b∗) and c∗
(rank ξ,∞)b∗.
Making use of the inductive hypothesis yet again and the preceding argument, we have that
γ = c∗
γ + d∗
ξ + βP ∗
γ = e∗
R∗c∗
(rank ξ,∞)R∗b∗
(rank ξ,∞)R∗b∗
(cid:40)
ξ) + βP ∗
γ = R∗(e∗
G(ξ) + βP ∗
e∗
βP ∗
(rank ξ,∞)R∗b∗
It follows that if G(ξ) is undefined and P ∗
declare G(γ) to be undefined. Otherwise, there are two remaining possiblities
(i) G(ξ) is undefined but P ∗
(rank ξ,∞)R∗b∗ = 0 then R∗c∗
if G(ξ) is defined
otherwise
=
element G(γ) := (n + 1, rank ξ, β, R∗b∗) ∈ Υ.
(rank ξ,∞)R∗b∗ (cid:54)= 0. In this case, it is easily verified that the
γ = 0. In this case we
(ii) G(ξ) is defined. It is again easily checked that the element G(γ) := (n+1, G(ξ), β, R∗b∗) ∈
Υ (here we note that in addition to the above arguments, we also need the inductive
hypothesis that G does not increase the age of an element).
In either of these cases, we see that R∗c∗
before, we then necessarily have R∗e∗
γ = e∗
G(γ). We can define R∗d∗
γ = c∗
G(γ) (in order that R∗ be linear).
γ to be d∗
We have thus succeeded in extending the maps G and R∗. By induction, we therefore
obtain maps G : Υ → Υ ∪ {undefined} and R∗ : (c00(Υ),(cid:107) · (cid:107)1) → (c00(Υ),(cid:107) · (cid:107)1) satisfying
G(γ) and as
71
the four properties above (here (c00(Υ),(cid:107) · (cid:107)1) is the dense subspace of (cid:96)1(Υ) consisting of
all finitely supported vectors). It follows from property (3), and the argument above, that
R∗ is continuous, with (cid:107)R∗(cid:107) = 1. It therefore extends (uniquely) to a bounded linear map
R∗ : (cid:96)1(Υ) → (cid:96)1(Υ) also having norm 1. This completes the proof.
We make some important observations about the mappings G and R∗.
Lemma 3.3.7. The dual operator of R∗, which we denote by (R∗)(cid:48) : (cid:96)1(Υ)∗ → (cid:96)1(Υ)∗
restricts to give a bounded linear operator R := (R∗)(cid:48)X(Υ) : X(Υ) → X(Υ) of norm at most
1.
Proof. It is a standard result that the dual operator (R∗)(cid:48) is bounded with the same norm as
R∗. It follows that the restriction of the domain to X(Υ) is a bounded, linear operator into
(cid:96)1(Υ)∗ with norm at most 1. It only remains to see that this restricted mapping actually
maps into X(Υ). Since the family (dγ)γ∈Υ is a basis for X(Υ), it is enough to see that the
image of dγ under (R∗)(cid:48) lies in X(Υ). It is therefore sufficient for us to prove that
(cid:88)
(R∗)(cid:48)dδ =
dγ
γ∈G−1({δ})
where of course G−1({δ}) denotes the pre-image of {δ} under G; we remark that G fails to
be injective and so is certainly not invertible. Since (d∗
γ)γ∈Υ is a basis for (cid:96)1(Υ) it is enough
to show that for every θ ∈ Υ(cid:0)(R∗)(cid:48)dδ
θ =(cid:0) (cid:88)
(cid:1)d∗
(cid:1)d∗
θ.
dγ
γ∈G−1({δ})
The right hand side of this expression is easy to evaluate; it is only non-zero if G(θ) = δ,
in which case it is equal to 1. In particular, if G(θ) is undefined, then the right hand side of
the expression is certainly 0, as is(cid:0)(R∗)(cid:48)dδ
(cid:1)d∗
θ = dδ(R∗d∗
θ) = dδ(0). If G(θ) is defined, the
G(θ)) which is clearly 1 if G(θ) = δ and 0 otherwise.
left hand side of the expression is dδ(d∗
So the expressions are indeed equal, as required.
Lemma 3.3.8. For every γ ∈ Υ, there is a unique l, 1 ≤ l ≤ k such that Gj(γ) is defined
whenever 1 ≤ j < l but Gl(γ) is undefined.
Proof. The uniqueness is easy; if G(γ) is defined, l is the maximal j ∈ N such that Gj−1(γ)
is defined. Otherwise we must have l = 1. So we only have to prove existence of such l.
We prove by induction on n that if rank γ = n there is some 1 ≤ l ≤ k such that Gl(γ) is
undefined, but Gj(γ) is defined if j < l. The case where n = 1 is clear from the construction
of the map G. So, inductively, we assume the statement holds whenever rank γ = m ≤ n.
Let γ ∈ Λn+1 and consider 2 cases:
72
(i) age γ = 1. We write γ = (n + 1, p, β, b∗). Now b∗ ∈ (cid:96)1(Υn) and by the inductive
hypothesis, for every θ ∈ Υ with rank θ ≤ n, there is some l ≤ k such that Gl(θ) is
undefined. It follows that we must have (R∗)lb∗ = 0 for some 1 ≤ l ≤ k. So it is
(p,∞)(R∗)lb∗ = 0 for some 1 ≤ l ≤ k. It follows by construction of
certainly true that P ∗
the map G that the 'l' we seek is the minimal l (1 ≤ l ≤ k) such that P ∗
(p,∞)(R∗)lb∗ = 0.
(ii) age γ > 1. We write γ = (n + 1, ξ, β, b∗). If G(γ) is undefined we are done; we must
have l = 1. Otherwise, G(γ) ∈ Υ and it follows from the construction of the map
G (see the proof of Theorem 3.3.6) that there are two possibilities to consider; either
G(γ) = (n + 1, rank ξ, β, R∗b∗) or G(γ) = (n + 1, G(ξ), β, R∗b∗). In the first of these
two possibilities, the same argument as in the previous case shows that the l we seek
is the minimal l (2 ≤ l ≤ k) such that P ∗
(p,∞)(R∗)lb∗ = 0. In the latter case, G(ξ)
is defined. But, since rank ξ < n, we know by the inductive hypothesis that there
exists some l0 ∈ N with 2 ≤ l0 ≤ k and Gl0(ξ) undefined but Gj(ξ) defined for j < l0.
Now, if P ∗
(rank ξ,∞)(R∗)l0b∗ = 0, then it follows from construction of G that Gl0(γ)
is undefined and Gj(γ) is defined for j < l0 so we are done. Otherwise, it follows
from an argument above that l0 < k, and there is some (minimal) l, l0 < l ≤ k with
P ∗
(rank ξ,∞)(R∗)lb∗ = 0. Once again, this is the desired l.
Corollary 3.3.9. The maps R∗ : (cid:96)1(Υ) → (cid:96)1(Υ) and R : X(Υ) → X(Υ) satisfy (R∗)k = 0
and Rk = 0.
Proof. It is clear from Lemma 3.3.8 and construction that the restriction of (R∗)k to c00(Υ)
is the zero map. It follows by density and continuity that (R∗)k = 0. The other claims are
immediate from the definition of R as the restriction of the dual operator of R∗.
To obtain the extra constraints that we place on "odd-weight" elements we need to
introduce a 'coding function' σ : Υ → N. This is similar to the coding function used in the
Argyros-Hadyon construction, [2], and analogous to the coding function used in the Gowers-
Maurey construction, [16]. We shall demand that σ satisfies the following properties:
(1) σ is injective
(2) σ(γ) > rank γ ∀ γ ∈ Υ
(3) for γ ∈ Λn+1 (i.e. rank γ = n + 1), σ(γ) > max{σ(ξ) : ξ ∈ Υn}.
73
Such a σ can be constructed recursively as Υ is constructed. Now, for each γ ∈ Υ we can
well-define a finite set Σ(γ) by
Σ(γ) := {σ(γ)} ∪ k−1(cid:91)
(cid:91)
{σ(δ)}
j=1
δ∈G−j (γ)
where G−j(γ) := {θ ∈ Υ : Gj(θ) = γ}. (In particular, if θ ∈ G−j(γ) then Gl(θ) ∈ Υ for
every l ≤ j.)
Before giving our main construction, we document some important observations.
Lemma 3.3.10. If γ ∈ Υ is such that G(γ) is defined then Σ(γ) ⊆ Σ(G(γ)).
Proof. Certainly γ ∈ G−1(G(γ)) so σ(γ) ∈ Σ(G(γ)). Suppose σ(δ) ∈ Σ(γ), δ (cid:54)= γ. So, there
is some 1 ≤ j ≤ k− 1 with δ ∈ G−j(γ), i.e. there is some δ such that Gj(δ) = γ. Since G(γ)
is defined, we must have Gj+1(δ) = G(γ) ∈ Υ. In particular, by Lemma 3.3.8, we must in
fact have had j < k − 1 so that j + 1 ≤ k − 1. Thus δ ∈ G−(j+1)(G(γ)) and σ(δ) ∈ Σ(G(γ)),
as required.
Lemma 3.3.11. If γ, γ(cid:48) ∈ Υ, rank γ > rank γ(cid:48) then Σ(γ) > Σ(γ(cid:48)), i.e. max{k : k ∈
Σ(γ(cid:48))} < min{k : k ∈ Σ(γ)}.
Proof. This follows immediately from the definition of the sets Σ(γ) and Σ(γ(cid:48)), the fact that
G is rank preserving and the assumption that for γ ∈ Λn+1, σ(γ) > max{σ(ξ) : ξ ∈ Υn}.
Lemma 3.3.12. Suppose γ, δ ∈ Υ. If σ(γ) ∈ Σ(δ) then either γ = δ or there is some
1 ≤ j ≤ k − 1 such that Gj(γ) = δ.
Proof. Since σ(γ) ∈ Σ(δ) there are two possibilities. Either σ(γ) = σ(δ) or there is some
1 ≤ j ≤ k − 1 and θ ∈ Υ with Gj(θ) = δ and σ(γ) = σ(θ). By injectivity of σ, this implies
that either γ = δ, or that θ = γ and Gj(γ) = δ as required.
We are finally in a position to describe the main construction. We will take a subset Γ ⊂
Υ by placing some restrictions on the elements of odd weight we permit. As a consequence of
imposing these additional odd weight restrictions, we are also forced to (roughly speaking)
remove those elements (n + 1, p, β, b∗) and (n + 1, ξ, β, b∗) of Υ for which the support of
b∗ is not contained in Γ, in order that we can apply the Bourgain-Delbaen construction to
obtain a space X(Γ). Note that the subset Γ will also be constructed inductively, so it will
be well defined and consistent with the Bourgain-Delbaen method previously described.
We will denote by ∆n the set of all elements in Γ having rank n, and denote by Γn the
union Γn = ∪j≤n∆j. The permissible elements of odd weight will be as follows. For an age
1 element of odd weight, γ = (n + 1, p, m−1
2j−1, b∗), we insist that either b∗ = 0 or b∗ = e∗
η
74
4i < n−2
2j−1, b∗), we insist that either b∗ = 0 or b∗ = e∗
where η ∈ Γn \ Γp and weight η = m−1
γ = (n + 1, ξ, m−1
weight η = m−1
Definition 3.3.13. We define recursively sets ∆n ⊆ Λn. Then Γn := ∪j≤n∆j and Γ :=
∪n∈N∆n ⊆ Υ. To begin the recursion, we set ∆1 = Λ1. Then
2j−1. For an odd weight element of age > 1,
η where η ∈ Γn \ Γrank ξ and
2j−1, k ∈ Σ(ξ). Let us be more precise:
4k < n−2
∆n+1 =
(n + 1, p, m−1
2j , b∗) : 0 ≤ p < n, b∗ ∈ Bp,n ∩ (cid:96)1(Γn)
(cid:111)
j=1
(cid:110)
(cid:98)(n+1)/2(cid:99)(cid:91)
(cid:110)
∪ n−1(cid:91)
(cid:98)p/2(cid:99)(cid:91)
(cid:110)
(cid:98)(n+2)/2(cid:99)(cid:91)
p=1
j=1
∪
j=1
∪ n−1(cid:91)
(cid:98)(p+1)/2(cid:99)(cid:91)
p=1
j=1
(cid:110)
(n + 1, ξ, m−1
2j , b∗) : ξ ∈ ∆p, w(ξ) = m−1
2j , age ξ < n2j, b∗ ∈ Bp,n ∩ (cid:96)1(Γn)
(n + 1, p, m−1
2j−1, b∗) : 0 ≤ p < n, b∗ = 0 or b∗ = e∗
η, η ∈ Γn \ Γp
(cid:111)
4i < n−2
2j−1
and w(η) = m−1
(n + 1, ξ, m−1
2j−1, b∗) : ξ ∈ ∆p, w(ξ) = m−1
2j−1, age ξ < n2j−1 ,
b∗ = 0 or b∗ = e∗
η with η ∈ Γn \ Γp, w(η) = m−1
4k < n−2
2j−1, k ∈ Σ(ξ)
(cid:111)
(cid:111)
.
Here the Bp,n are defined as in Remark 3.3.5, and, for brevity, we temporarily write w(ξ)
for weight ξ. We define Xk to be the Bourgain-Delbaen space X(Γ) where Γ is the subset
of Υ just defined.
For the rest of this chapter we will work with the space Xk. As in [2], the structure of
the space Xk is most easily understood in terms of the basis (dγ)γ∈Γ and the biorthogonal
functionals d∗
γ in order
to estimate norms. To this end, we have the following proposition.
γ. However, we will need to work with the evaluation functionals e∗
Proposition 3.3.14. Let n be a positive integer and let γ be an element of ∆n+1 of weight
m−1
j and age a ≤ nj. Then there exist natural numbers p0 < p1 < ··· < pa = n+1, elements
ξ1, . . . , ξa = γ of weight m−1
that
j with ξr ∈ ∆pr and functionals b∗
(cid:0)Γpr−1 \ Γpr−1
(cid:1) such
r ∈ ball (cid:96)1
a(cid:88)
a(cid:88)
r=1
e∗
γ =
=
a(cid:88)
a(cid:88)
r=1
ξr + m−1
d∗
j
ξr + m−1
d∗
j
P ∗
(pr−1,∞)b∗
r
P ∗
(pr−1,pr)b∗
r.
r=1
r=1
75
If 1 ≤ t < a we have
a(cid:88)
ξr + m−1
d∗
j
a(cid:88)
r=t+1
r=t+1
e∗
γ = e∗
ξt +
(pr−1,∞)b∗
P ∗
r .
Proof. The proof is an easy induction on the age a of γ. We omit the details because the
argument is the same as in [2] except our p0 is not necessarily 0.
Remark 3.3.15. As in [2], we shall refer to any of the above identities as the evaluation
analysis of the element γ, and the data (p0, (pr, b∗
r, ξr)1≤r≤a) as the analysis of γ. We will
omit the p0 when p0 = 0.
We will now construct the operator S : Xk → Xk. We need:
Proposition 3.3.16. Γ is invariant under G. More precisely, if γ ∈ Γ ⊆ Υ and G(γ) is
defined, then G(γ) ∈ Γ. It follows that the map G : Υ → Υ∪{undefined} defined in Theorem
3.3.6 restricts to give F : Γ → Γ ∪ {undefined}. Consequently the map R∗ : (cid:96)1(Υ) → (cid:96)1(Υ)
(also defined in 3.3.6) can be restricted to the subspace (cid:96)1(Γ) ⊆ (cid:96)1(Υ) giving S∗ : (cid:96)1(Γ) →
(cid:96)1(Γ). S∗ is a bounded linear map on (cid:96)1(Γ) of norm 1 which satisfies
(cid:40)
S∗e∗
γ =
0
e∗
F (γ)
if F (γ) is undefined
otherwise
for every γ ∈ Γ. Moreover, the dual operator of S∗ restricts to Xk to give a bounded linear
operator S : Xk → Xk of norm at most 1, satisfying Sj (cid:54)= 0 for 1 ≤ j ≤ k − 1, Sk = 0.
Proof. It is enough to show that when γ ∈ Γ and G(γ) is defined, then G(γ) ∈ Γ. The
claims about the operator S∗ follow immediately from the definition of S∗ as the restriction
of R∗ and the definition of R∗. The fact that S : Xk → Xk is well defined follows by the
same argument as in 3.3.7; indeed it is seen that for δ ∈ Γ,
(cid:88)
Sdδ =
dγ.
γ∈F −1(δ)
Moreover, by Corollary 3.3.9, we see that (S∗)k = 0 and therefore that Sk = 0. That Sj (cid:54)= 0
for 1 ≤ j ≤ k − 1 is clear from the above formula and consideration of the elements dm for
m ∈ ∆1.
We use induction on the rank of γ to prove that if γ ∈ Γ and G(γ) is defined, then
G(γ) ∈ Γ. This is certainly true when rank γ = 1. Suppose by induction, that whenever
γ ∈ Γ, rank γ ≤ n and G(γ) is defined, G(γ) ∈ Γ and consider a γ ∈ Γ, of rank n+1 such that
G(γ) is defined. Let us suppose first that this γ has age 1. We can write γ = (n + 1, p, β, b∗)
η. Since G(γ) is defined, we
where supp b∗ ⊆ Γn \ Γp, and we write b∗ = (cid:80)
η∈Γn\Γp
aηe∗
76
have G(γ) = (n + 1, p, β, S∗b∗). We consider further sub-cases.
(i.e. γ is an "even weight" element), we see that the only way G(γ) fails to be in Γ is if
supp S∗b∗ (cid:42) Γn \ Γp. But
If β = m−1
for some j
2j
(cid:88)
S∗b∗ =
aηe∗
F (η)
η∈Γn\Γp ∩
{η : F (η) is defined}
and by the inductive hypothesis all the F (η) in this sum are in Γn\Γp. So supp S∗b∗ ⊆ Γn\Γp
and G(γ) = F (γ) ∈ Γ as required. In the case where β = m−1
2j−1 (i.e. γ is an "odd-weight"
element) we must also check that the odd weight element G(γ) is of a permissible form.
(p,∞)S∗b∗ (cid:54)= 0. This of course
Since G(γ) is defined, in particular we must have that P ∗
η for some η ∈ Γn \ Γp where
implies that S∗b∗ (cid:54)= 0 and b∗ (cid:54)= 0. As γ ∈ Γ, b∗ = e∗
weight η = m−1
2j−1. Since S∗b∗ (cid:54)= 0, we must have S∗e∗
η = e∗
F (η) where in particular,
F (η) is defined and lies in Γ by the inductive hypothesis. Since G is weight preserving, we
have weight F (η) = weight η = m−1
2j−1. Moreover, since G preserves rank, we can
now conclude that F (γ) = G(γ) = (n + 1, p, m−1
F (η)) ∈ Γ as
required.
2j−1, S∗b∗) = (n + 1, p, m−1
4i < n−2
4i < n−2
2j−1, e∗
When age γ > 1, we can write γ = (n + 1, ξ, β, b∗) where (in particular) supp b∗ ⊆ Γn
and ξ ∈ Γp (p < n).
If this γ is of even weight, it follows easily from the inductive
hyptohesis and arguments similar to the 'age 1' case that G(γ) = F (γ) ∈ Γ (when G(γ)
is defined). So we consider only the case when weight γ = β = m−1
2j−1 (for some j) and
G(γ) is defined.
If G(ξ) = F (ξ) is undefined, then since G(γ) is defined, we must have
G(γ) = (n + 1, rank ξ, m−1
S∗b∗ (cid:54)= 0. Again by the restrictions on elements of odd weight, we must have b∗ = e∗
weight η = m−1
and S∗b∗ = S∗e∗
hypothesis, that F (η) ∈ Γn \ Γrank ξ. Furthermore, weight F (η) = weight η = m−1
We conclude that G(γ) = F (γ) ∈ Γ as required.
(rank ξ,∞)S∗b∗ (cid:54)= 0. So in particular, b∗ (cid:54)= 0 and
η where
2j−1, some k ∈ Σ(ξ). Since S∗b∗ (cid:54)= 0, we must have G(η) = F (η) defined
F (η). Since G preserves rank, we see as a consequence of the inductive
2j−1.
2j−1, S∗b∗) and P ∗
4k < n−2
η = e∗
4k < n−2
In the case where F (ξ) is defined, G(γ) = (n + 1, F (ξ), m−1
If S∗b∗ = 0
then by the inductive hypothesis, we certainly have G(γ) = F (γ) ∈ Γ and we are done.
2j−1, some k ∈ Σ(ξ) and
Otherwise, we again must have b∗ = e∗
2j−1 and k ∈ Σ(ξ) ⊆ Σ(G(ξ)) by
S∗b∗ = e∗
Lemma 3.3.10.
F (η). Now, weight F (η) = weight η = m−1
η where weight η = m−1
2j−1, S∗b∗).
4k < n−2
4k < n−2
Later, we will need the following lemma about the elements of odd weight in Γ.
i )(cid:1)
Lemma 3.3.17. Let γ ∈ Γ be an element of odd weight and age γ > 1. Let(cid:0)p0, (pi, ξi, b∗
be the analysis of γ, where we know each b∗
are i, j, 1 ≤ i < j ≤ age γ := a with b∗
i = e∗
i is either 0 or e∗
ηi and b∗
j = e∗
ηi for some suitable ηi. If there
ηj then weight ηj (cid:12) weight ηi.
77
1 = e∗
Proof. We consider the cases when i = 1 and i > 1 separately. Suppose b∗
weight η1 = m−1
m−1
j where 1 ≤ j ≤ p. So weight η1 = m−1
monotonicity of the sequence mj, we get that m−1
η1 where
4i . By construction, elements of rank p are only allowed to have weights
4i =⇒ 4i ≤ rank η1 < rank ξ1 = p1. By the strict
j (cid:54)= 0, b∗
Now for any j > 1, if b∗
4k for some k ∈ Σ(ξj−1).
Since the mapping F preserves rank of elements, and σ(θ) > rank θ ∀θ ∈ Γ (Assumption
(2) of the σ mapping), it is immediate from the definition of Σ(ξj−1) that k ∈ Σ(ξj−1) =⇒
k > rank ξj−1 ≥ rank ξ1. Now
ηj where weight ηj = m−1
4i = weight η1.
j = e∗
< m−1
rank ξ1
weight ηj = m−1
4k < m−1
4rank ξj−1
≤ m−1
4rank ξ1
≤ weight η1
i = e∗
thus concluding the proof for the case i = 1. The case when i > 1 is easy. In this case
for some l ∈ Σ(ξi−1). Now j > i, so (by Lemma
we suppose b∗
3.3.11) Σ(ξj−1) > Σ(ξi−1) so that l < k and therefore weight ηj = m−1
4l = weight ηi,
completing the proof.
ηi, weight ηi = m−1
4k < m−1
4l
3.4 Rapidly Increasing Sequences and the operator S : Xk →
Xk
We recall from [2] that special classes of block sequences, namely the rapidly increasing
sequences admit good upper estimates. This class of block sequences will also be useful in
our construction. We recall the definition:
Definition 3.4.1. Let I be an interval in N and let (xk)k∈I be a block sequence (with
respect to the FDD (Mn)). We say that (xk) is a rapidly increasing sequence, or RIS, if
there exists a constant C such that the following hold:
1. (cid:107)xk(cid:107) ≤ C for all k ∈ I,
and there is an increasing sequence (jk) such that, for all k,
2. jk+1 > max ran xk ,
3. xk(γ) ≤ Cm−1
i whenever weight γ = m−1
i
and i < jk .
If we need to be specific about the constant, we shall refer to a sequence satisfying the
above conditions as a C-RIS.
Remark 3.4.2. We make the following important observation. If (xi)i∈N is a C-RIS, then
so also is the sequence (Sxi). We omit the very easy proof.
78
We also note that the estimates of Lemmas and Propositions 5.2 - 5.6 and 5.8 of [2] all
still hold. The same proofs go through, with only minor modifications to take account of
the fact that p0 doesn't need to be 0 in the evaluation analysis of an element γ in our Γ (see
Proposition 3.3.14). For convenience, we state Proposition 5.6 of [2] as we shall be making
repeated use of it throughout this chapter:-
Proposition 3.4.3. Let (xk)
nj0
k=1 be a C-RIS. Then
1. For every γ ∈ Γ with weight γ = m−1
nj0(cid:88)
n−1
j0
xk(γ) ≤
(cid:40)
h we have
16Cm−1
4Cn−1
j0
j0
m−1
+ 6Cm−1
h
h
k=1
n−1
j0
In particular,
if h > j0 and
if h < j0
if h ≥ j0
,
xk(γ) ≤ 10Cm−2
j0
nj0(cid:88)
nj0(cid:88)
k=1
k=1
nj0(cid:88)
(cid:107)n−1
j0
xk(cid:107) ≤ 10Cm−1
j0
.
(cid:88)
k∈J
λkxk(γ) ≤ C max
k∈J
λk,
(cid:107)n−1
j0
λkxk(cid:107) ≤ 10Cm−2
j0
.
2. If λk (1 ≤ k ≤ nj0) are scalars with λk ≤ 1 and having the property that
for every γ of weight m−1
j0
and every interval J ⊆ {1, 2, . . . , nj0}, then
k=1
Another result of particular importance to us will be the following proposition of [2]:
Proposition 3.4.4. Let Y be any Banach space and T : Xk → Y be a bounded linear
operator. If (cid:107)T (xk)(cid:107) → 0 for every RIS (xk)k∈N in Xk then (cid:107)T (xk)(cid:107) → 0 for every bounded
block sequence in Xk.
The proof exploits the explicit finite dimensional subspace structure provided by the
L∞ spaces of Bourgain-Delbaen type; exactly the same argument used by Argyros and
Haydon in [2] works for the spaces Xk and we refer the reader to the cited paper for further
details. This proposition will be particularly important when we come to prove the operator
representation property of the main theorem.
We can also make use of Propositions 3.4.3 and 3.4.4 to prove the (cid:96)1 duality of the space
Xk.
79
Proposition 3.4.5. The dual of Xk is (cid:96)1(Γ). More precisely the map
defined by
ϕ : (cid:96)1(Γ) → X∗
k
ϕ(x∗)x := (cid:104)x, x∗(cid:105)
(where x ∈ Xk ⊆ (cid:96)∞(Γ) = (cid:96)∗
1(Γ), x∗ ∈ (cid:96)1(Γ)) is an isomorphism.
Proof. The argument here can be found in [2], though we write out the details more ex-
plicitly. We note that by Theorem 1.4.6 it is enough to show that the basis (dγ)γ∈Γ of Xk
is shrinking. Using Proposition 1.4.5 it is enough to show that every bounded block basic
sequence with respect to this basis is weakly null. By Proposition 3.4.4, it is sufficient to
see that every RIS (xi)i∈N is weakly null. To this end, suppose (xi)i∈N is a C-RIS which
fails to be weakly null; without loss of generality we may assume that there exists δ > 0
and an x∗ ∈ X∗
k such that (cid:107)x∗(cid:107) = 1 and x∗(xi) ≥ δ for all i. It follows that
nj0(cid:88)
i=1
1
nj0
x∗(xi) ≥ δ
for any j0 ∈ N. On the other hand, we can choose j0 ∈ N such that 10Cm−1
from Proposition 3.4.3 that for this choice of j0
j0
< δ. It follows
nj0(cid:88)
i=1
nj0(cid:88)
i=1
1
nj0
x∗(xi) ≤ (cid:107) 1
nj0
xi(cid:107) ≤ 10Cm−1
j0
< δ.
This contradiction completes the proof.
This canonical identification of X∗
k with (cid:96)1(Γ) allows us to prove some important prop-
erties of the operator S : Xk → Xk.
Lemma 3.4.6. The dual of the operator S : Xk → Xk is precisely the operator S∗ : (cid:96)1(Γ) →
(cid:96)1(Γ) under the canonical identification, ϕ, of X∗
k with (cid:96)1(Γ). Moreover, Sj has closed range
for every 1 ≤ j ≤ k − 1, and consequently, Sj : Xk → im Sj is a quotient operator.
Proof. We will temporarily denote the dual map of S by S(cid:48) : X∗
k so that we don't
confuse it with the S∗ mapping on (cid:96)1(Γ). By continuity and linearity, the maps ϕ−1S(cid:48)ϕ
γ for γ ∈ Γ. For x ∈ Xk,
and S∗ are completely determined by their action on the vectors e∗
k → X∗
(cid:40)
80
(cid:0)S(cid:48)ϕ(e∗
γ)(cid:1) (x) = ϕ(e∗
γ)Sx = (cid:104)Sx, e∗
γ(cid:105) = (cid:104)x, S∗e∗
γ(cid:105) =
x(F (γ)
0
if F (γ) is defined
otherwise
It follows from this that
(cid:40)
e∗
F (γ)
0
ϕ−1S(cid:48)ϕ(e∗
γ) =
= S∗(e∗
γ)
if F (γ) is defined
otherwise
as required.
To see that the image of Sj is closed, we make use of Banach's Closed Range Theorem,
Theorem 1.4.14; it follows that it is enough to show that the image of (S∗)j is closed. This
is easy, as it is clear that the image of (S∗)j is just (cid:96)1(Γ ∩ im F j) ⊆ (cid:96)1(Γ) (with the obvious
embedding), so certainly (S∗)j has closed image.
Since im Sj is closed, an easy application of the Inverse Mapping Theorem yields that
Sj : Xk → im Sj is a quotient operator.
Corollary 3.4.7. For 1 ≤ j ≤ k − 1there are aj, bj ∈ R, aj, bj > 0 such that whenever
x ∈ Xk,
aj(cid:107)Sjx(cid:107) ≤ dist(x, Ker Sj) ≤ bj(cid:107)Sjx(cid:107)
Proof. This is immediate from the fact that Sj : Xk → im Sj is a quotient operator.
Corollary 3.4.8. Suppose λi ∈ R (0 ≤ i ≤ k − 1) are such that (cid:80)k−1
i=0 λiSi is compact.
Then λi = 0 for every i. Consequently, there does not exist j (1 ≤ j ≤ k − 1) such
that Sj is a compact perturbation of some linear combination of the operators Sl, l (cid:54)= j.
(Equivalently, {I + K(Xk), Sj + K(Xk) : 1 ≤ j ≤ k − 1} is a linearly independent set of
vectors in L(Xk)/K(Xk).)
Proof. It follows by a standard result that (cid:80)k−1
operator is compact. But the dual operator is just T :=(cid:80)k−1
i=0 λiSi is compact if, and only if, the dual
i=0 λi(S∗)i : (cid:96)1(Γ) → (cid:96)1(Γ) and
it is now easily seen that this is not compact unless all the λi are 0. Indeed, suppose T is
0) ∈ Γ.
compact and consider first the sequence (e∗
Then, for m (cid:54)= n (observing that F (γ0
n=1 ⊆ B(cid:96)1(Γ) where γ0
)∞
n = (n+1, m−1
2 , 0, e∗
γ0
n
(cid:107)T e∗
γ0
n
− T e∗
γ0
m
n) is undefined for every n) we have
(cid:107)1 = λ0(cid:107)e∗
(cid:107)1 = 2λ0
− e∗
γ0
n
γ0
m
We must therefore have that λ0 = 0 in order that the sequence (T e∗
subsequence. Then, considering in turn the sequences (e∗
γj
1, m−1
n
be 0 as claimed.
n = (n +
j ) ∈ Γ (for 1 ≤ j ≤ k − 1), we see by the same arguments that all the λi must
n=1 ⊆ B(cid:96)1(Γ) where γj
)∞
) has a convergent
2 , 0, e∗
γ0
n
81
expressible in the form T =(cid:80)k−1
To complete the proof of the main theorem, it still remains for us to see that Xk is HI, the
operator S is strictly singular, and that every bounded linear operator T ∈ L(Xk) is uniquely
i=0 λiSi + K for some compact operator K : Xk → Xk. We
note that the claimed uniqueness of this operator representation follows immediately from
the above corollary. We focus now on proving the existence of the representation and aim
to prove:
Theorem 3.4.9. Let T : Xk → Xk be a bounded linear operator on Xk and (xi)i∈N a RIS
in Xk. Then dist(T xi,(cid:104)xi, Sxi . . . , Sk−1xi(cid:105)R) → 0 as i → ∞.
The proof is similar to that given in [2]. We will need slight modifications to the
definitions of exact pairs and dependent sequences. We find it convenient to define both the
0 (δ = 0 in the definitions that follow) and 1 (δ = 1 in the definitions that follow) exact pairs
and dependent sequences below. However, initially, we will only be concerned with the 0
exact pairs and dependent sequences. The 1 exact pairs and dependent sequences will only
be needed to establish that the space Xk is hereditarily indecomposable. We also introduce
the new, but related notions, of 'weak exact pairs' and 'weak dependent sequences' which
will be useful to us later in establishing strict singularity of S.
Definition 3.4.10. Let C > 0, δ ∈ {0, 1}. A pair (x, η) ∈ Xk × Γ is said to be a (C, j, δ)-
special exact pair if
j
j
for all ξ ∈ Γ
1. (cid:107)x(cid:107) ≤ C
ξ, x(cid:105) ≤ Cm−1
2. (cid:104)d∗
3. weight η = m−1
4. x(η) = δ and Slx(η) = 0 for every 1 ≤ l ≤ k − 1
5. for every element η(cid:48) of Γ with weight η(cid:48) = m−1
Cm−1
Cm−1
x(η(cid:48)) ≤
(cid:40)
i
(cid:54)= m−1
j
, we have
i
j
if i < j
if i > j.
Given also an ε > 0 we will say a pair (x, η) ∈ Xk × Γ is a (C, j, 0, ε)-weak exact pair if
condition (4) is replaced by the following (weaker) condition:
(4(cid:48)) Slx(η) ≤ Cε for 0 ≤ l ≤ k − 1.
We will say a pair (x, η) ∈ Xk × Γ is a (C, j, 1, ε)-weak exact pair if condition (4) is replaced
by condition:
82
(4(cid:48)(cid:48)) x(η) = 1 and Slx(η) ≤ Cε for 1 ≤ l ≤ k − 1.
We note that a (C, j, δ) special exact pair is a (C, j, δ, ε) weak exact pair for any ε > 0.
Moreover, the definition of a (C, j, δ)-special exact pair is the same as the definition of a
(C, j, δ) exact pair given in [2] but with the additional requirement that Sjx(η) = 0 for all
j ≥ 1. The remark made in [2] (following the definition of exact pairs) is therefore still
valid. In fact it is easily verified that the same remark in fact holds for weak exact pairs.
For convenience, we state the remark again as it will be useful to us later:
Remark. A (C, j, δ, ε) weak exact pair also satisfies the estimates
(cid:104)e∗
η(cid:48), P(s,∞)x(cid:105) ≤
6Cm−1
6Cm−1
i
j
if i < j
if i > j
(cid:40)
n2j(cid:88)
k=1
n2j(cid:88)
for elements η(cid:48) of Γ with weight η(cid:48) = m−1
i
(cid:54)= m−1
j
.
We will need the following method for constructing special exact pairs.
2j
(cid:80)n2j
k=1 xk. For each k let b∗
k, xk(cid:105) = (cid:104)b∗
k=1 be a skipped-block C-RIS, and let q0 < q1 < q2 < ··· < qn2j
Lemma 3.4.11. Let (xk)n2j
be natural numbers such that ran xk ⊆ (qk−1, qk) for all k. Let z denote the weighted sum
z = m2jn−1
k, xk(cid:105) = 0 and
k, Slxk(cid:105) = 0 for all l.Then there exist ζi ∈ ∆qi (1 ≤ i ≤ n2j) such that
(cid:104)(S∗)lb∗
the element η = ζn2j has analysis (qi, b∗
i , ζi)1≤i≤n2j and (z, η) is a (16C, 2j, 0)-special exact
pair.
Proof. The proof is the same as in [2]. We only need to show that Slz(η) = 0 for 1 ≤ l ≤
k − 1. This is easy.
k be an element of Bqk−1,qk−1 with (cid:104)b∗
Slz(η) = (cid:104)Slz, e∗
η(cid:105) = (cid:104)Slz,
d∗
ζk
+ m−1
2j P ∗
k(cid:105)
(qk−1,qk)b∗
It is clear from the definition of S that ran xk ⊆ (qk−1, qk) =⇒ ran Slxk ⊆ (qk−1, qk) and
since rank ζk = qk for every k, it follows that (cid:104)Slz,(cid:80)n2j
(cid:105) = 0. We thus see that
k=1 d∗
ζk
Slz(η) = n−1
2j
(cid:104)Slxk, b∗
k(cid:105) = 0
as required.
k=1
Definition 3.4.12. Consider the space Xk. We shall say that a sequence (xi)i≤n2j0−1 is a
(C, 2j0 − 1, δ)-special dependent sequence if there exist 0 = p0 < p1 < p2 < ··· < pn2j0−1,
together with ηi ∈ Γpi−1 \ Γpi−1 and ξi ∈ ∆pi (1 ≤ i ≤ n2j0−1) such that
1. for each k, ran xk ⊆ (pk−1, pk)
83
2. the element ξ = ξn2j0−1 of ∆pn2j0−1
has weight m−1
2j0−1 and analysis (pi, e∗
ηi, ξi)
n2j0−1
i=1
3. (x1, η1) is a (C, 4j1, δ)-special exact pair
4. for each 2 ≤ i ≤ n2j0−1, (xi, ηi) is a (C, 4σ(ξi−1), δ)-special exact pair.
If we instead ask that
(3(cid:48)) (x1, η1) is a (C, 4j1, δ, n−1
(4(cid:48)) (xi, ηi) is a (C, 4σ(ξi−1), δ, n−1
n2j0−1
i=0
we shall say the sequence (xi)
2j0−1) weak exact pair
2j0−1) weak exact pair for 2 ≤ i ≤ n2j0−1
is a weak (C, 2j0 − 1, δ) dependent sequence.
In either case, we notice that, because of the special odd-weight conditions, we necessar-
2j0−1, by Lemma 3.3.17
2j0−1, and weight ηi+1 = m−1
= weight η1 < n−2
4σ(ξi) < n−2
ily have m−1
for 1 ≤ i < n2j0−1.
4j1
We also observe that a (C, 2j0 − 1, δ) special dependent sequence is certainly a weak
(C, 2j0 − 1, δ) dependent sequence.
Lemma 3.4.13. Let (xi)i≤n2j0−1 be a weak (C, 2j0 − 1, 0)-dependent sequence in Xk and let
J be a sub-interval of [1, n2j0−1]. For any γ(cid:48) ∈ Γ of weight m2j0−1 we have
(cid:88)
ξn2j0−1, an element of weight m2j0−1. Let(cid:0)p(cid:48)
i∈J
each b(cid:48)∗
i
is either 0 or e∗
η(cid:48)
The proof is easy if all the b(cid:48)∗
i
0, (p(cid:48)
for some suitable η(cid:48)
i.
r are 0, or if
xi(γ(cid:48)) ≤ 7C.
Proof. Let ξi, ηi, pi, j1 be as in the definition of a dependent sequence and let γ denote
i)1≤i≤a(cid:48)(cid:1) be the analysis of γ(cid:48) where
i, b(cid:48)∗
i , ξ(cid:48)
{weight η(cid:48)
r : 1 ≤ r ≤ a(cid:48), b(cid:48)∗
r = e∗
η(cid:48)
r
So we may suppose that there is some 1 ≤ r ≤ a(cid:48) s.t. b(cid:48)∗
for some i. We choose l maximal such that there exists i with b(cid:48)∗
weight ηl. Clearly we can estimate as follows
} ∩ {weight ηi : 1 ≤ i ≤ n2j0−1} = ∅.
with weight η(cid:48)
i = e∗
η(cid:48)
r
i
r = weight ηi
i =
and weight η(cid:48)
(cid:88)
k∈J
xk(γ(cid:48)) ≤ (cid:88)
k∈J, k<l
xk(γ(cid:48)) + xl(γ(cid:48)) +
xk(γ(cid:48)).
r = e∗
η(cid:48)
(cid:88)
k∈J, k>l
We now estimate the three terms on the right hand side of the inequality separately.
xl(γ(cid:48)) ≤ (cid:107)xl(cid:107) ≤ C. Also
84
(cid:88)
k∈J, k>l
xk(γ(cid:48)) =
(cid:88)
(cid:88)
i≤a(cid:48)
l<k∈J, i≤a(cid:48) (cid:104)d∗
ξ(cid:48)
(cid:104)d∗
ξ(cid:48)
i
≤ n2
i
k∈J, k>l
2j0−1 max
where weight η(cid:48)
i = 0, then
Now each b(cid:48)∗
i
is 0 or e∗
η(cid:48)
contradict maximality of l). If b(cid:48)∗
i
, xk(cid:105) + m−1
2j0−1(cid:104)b(cid:48)∗
, xk(cid:105) + m−1
i , P(p(cid:48)
2j0−1(cid:104)b(cid:48)∗
i−1,∞)xk(cid:105)
i , P(p(cid:48)
i−1,∞)xk(cid:105).
i (cid:54)= weight ηk for any k > l (or else we would
(cid:104)d∗
ξ(cid:48)
i
, xk(cid:105) + m−1
2j0−1(cid:104)b(cid:48)∗
i , P(p(cid:48)
i−1,∞)xk(cid:105) = (cid:104)d∗
ξ(cid:48)
i
, xk(cid:105) ≤ Cweight ηk ≤ Cn−2
2j0−1
where the penultimate inequality follows from the definition of a (weak) exact pair, and
the final inequality follows from Lemma 3.3.17. Otherwise b(cid:48)∗
(by restrictions on elements of odd weight) weight η(cid:48)
exact pair and the remark following it we have
where in particular
2j0−1. By the definition of (weak)
i = e∗
η(cid:48)
i < n−2
i
(cid:104)d∗
ξ(cid:48)
i
, xk(cid:105) + m−1
2j0−1P(pi−1,∞)xk(η(cid:48)
i) ≤ Cweight ηk + 6Cm−1
2j0−1 max{weight η(cid:48)
i, weight ηk}
Finally we consider (cid:80)
≤ 3Cn−2
2j0−1.
is proved. So we can suppose l > 1. By definition of l, there exists some i such that b(cid:48)∗
and weight ηl = weight η(cid:48)
k∈J, k<l xk(γ(cid:48)). Obviously if l = 1 this sum is zero, and the lemma
i = e∗
η(cid:48)
i. Now either i = 1 or i > 1. We consider the 2 cases separately.
i
Suppose first that i = 1.
xk(γ(cid:48)) = (cid:104) (cid:88)
(cid:88)
k∈J, k<l
a(cid:48)(cid:88)
k∈J, k<l
2j0−1 max
≤ n2
r (cid:105)
r−1,∞)b(cid:48)∗
2j0−1P ∗
(p(cid:48)
(cid:105) + m−1
2j0−1(cid:104) (cid:88)
k∈J, k<l
(cid:88)
r≤a(cid:48)s.t.
r =e∗
b(cid:48)∗
η(cid:48)
r
xk,
r−1,∞)e∗
P ∗
η(cid:48)
(p(cid:48)
r
(cid:105)
r
r=1
xk,
d∗
ξ(cid:48)
+ m−1
J(cid:51)k<l, r≤a(cid:48) (cid:104)xk, d∗
ξ(cid:48)
2j0−1(cid:104) (cid:88)
xk,
r
k∈J, k<l
(cid:88)
r≤a(cid:48)s.t.
r =e∗
b(cid:48)∗
η(cid:48)
r
≤ C + m−1
r−1,∞)e∗
P ∗
η(cid:48)
(p(cid:48)
r
(cid:105)
where the last inequality follows once again from the definition of exact pair. Suppose that
for some k ∈ J, k < l, there is an r in {1, 2, . . . a(cid:48)} with b(cid:48)∗
By Lemma 3.3.17, we get weight η(cid:48)
weight η(cid:48)
to γ(cid:48). Thus there does not exist r in {1, 2, . . . a(cid:48)} with b(cid:48)∗
for some k ∈ J, k < l. Using an argument similar to the above, we finally deduce that
r = weight ηk.
r >
1. But since γ(cid:48) also has odd weight, this clearly contradicts Lemma 3.3.17 applied
r = weight ηk
r = weight ηk > weight ηl = weight η(cid:48)
1, i.e. weight η(cid:48)
and weight η(cid:48)
and weight η(cid:48)
r = e∗
η(cid:48)
r = e∗
η(cid:48)
r
r
2j0−1(cid:104) (cid:88)
m−1
xk,
k∈J, k<l
(cid:88)
r≤a(cid:48)s.t.
b(cid:48)∗
r =e∗
η(cid:48)
r
85
r−1,∞)e∗
P ∗
η(cid:48)
(p(cid:48)
r
(cid:105) ≤ 2C
and so we get the required result.
Finally it remains to consider what happens when i > 1. Recall we are also assuming
l > 1 and weight η(cid:48)
m−1
4σ(ξl−1), and by restrictions on elements of odd weights, weight η(cid:48)
By strict monotonicity of the sequence mj, we deduce that ω = σ(ξl−1) ∈ Σ(ξ(cid:48)
Lemma 3.3.12 there are now two possibilites. Either ξl−1 = ξ(cid:48)
1 ≤ j ≤ k− 1, such that F j(ξl−1) = ξ(cid:48)
this implies pl−1 = p(cid:48)
γ(cid:48) as
i = weight ηl. But by definition of a special exact pair, we have weight ηl =
i−1).
i−1). By
i−1 or, if not, there is some j,
i−1. In either of these cases, we note that in particular
i−1 since F preserves rank and we can write the evaluation analysis of
4ω with ω ∈ Σ(ξ(cid:48)
i = m−1
e∗
γ(cid:48) = (S∗)j(e∗
ξl−1
) +
d∗
ξ(cid:48)
r
+ m−1
2j0−1P ∗
(p(cid:48)
r−1,p(cid:48)
r)b(cid:48)∗
r
a(cid:48)(cid:88)
r=i
for some 0 ≤ j ≤ k − 1. Now, for k < l, since ran xk ⊆ (pk−1, pk) ⊆ (0, pl−1) = (0, p(cid:48)
see that
i−1), we
(cid:104)xk, e∗
γ(cid:48)(cid:105) = (cid:104)xk, (S∗)je∗
l−1(cid:88)
ξl−1
ξr + m−1
d∗
(cid:105)
= (cid:104)Sjxk,
= m−1
≤ Cn−1
r=1
2j0−1(cid:104)Sjxk, e∗
2j0−1
ηk
(cid:105)
2j0−1P ∗
(pr−1,pr)e∗
ηr(cid:105)
by definition of a weak exact pair. It follows that
(cid:88)
k∈J, k<l
xk(γ(cid:48)) ≤ n2j0−1 max
k∈J, k<l
xk(γ(cid:48)) ≤ C.
This completes the proof.
As a consequence of the above lemma and Proposition 3.4.3 we obtain the following
upper norm estimate for the averages of weak special dependent sequences.
Proposition 3.4.14. Let (xi)i≤n2j0−1 be a weak (C, 2j0 − 1, 0) dependent sequence in Xk.
Then
(cid:107)n−1
2j0−1
xi(cid:107) ≤ 70Cm−2
2j0−1
n2j0−1(cid:88)
i=1
Proof. We apply the second part of Proposition 3.4.3, with λi = 1 and 2j0−1 playing the role
of j0. Lemma 3.4.13 shows that the extra hypothesis of the second part of Proposition 3.4.3
is satisfied, provided we replace C by 7C. We deduce that (cid:107)n−1
2j0−1
as claimed.
(cid:80)n2j0−1
xi(cid:107) ≤ 70Cm−2
2j0−1
i=1
86
We will need the following lemma from elementary linear algebra in the proof of Theorem
3.4.9.
Lemma 3.4.15. Suppose for j = 1, . . . , k, uj ∈ Qn ⊆ Rn (where k ≤ n) and moreover, the
uj are linearly independent over R. Let V = {v ∈ Rn : (cid:104)v, uj(cid:105) = 0 for j = 1, . . . , k}. Then
V ∩ Qn is dense in V .
Proof. Consider the linear map L : Qn → Qk, defined by Qn (cid:51) v (cid:55)→ ((cid:104)v, u1(cid:105),(cid:104)v, u2(cid:105), . . . ,(cid:104)v, uk(cid:105)).
Thinking of Qn and Qk as vector spaces over Q, and by the assumed linear independence
of the uj, we can apply the Rank-Nullity Theorem to L to conclude that V ∩ Qn = ker L is
a Q-vector space of dimension n − k.
Now, V ∩ Qn (where the closure is taken in Rn), is easily seen to be a real vector space.
Moreover, V ∩ Qn ⊆ V = V , and it is easily seen (e.g. by a similar rank-nullity argument
as before) that V is an n− k dimensional real vector space. So V ∩ Qn is a real vector space
of dimension at most n − k.
On the other hand, if q1, . . . ql (l ≤ n) are vectors in Qn which are linearly independent
over Q, they must in fact be linearly independent over R. Indeed, without loss of generality,
we may assume l = n; if not, working in the Q-vector space, Qn, we may extend the (qi)i≤l
to a basis, (qi)1≤i≤n, of Qn and our claim we will be proved if we show this extended
set (which is obviously linearly independent over Q) is really linearly independent over R.
Now, since q1, . . . , qn is a basis of Qn, the standard basis of Rn, (ej)n
j=1 is in the real (in
fact, rational) span of q1, . . . , qn. Consequently, Rn is the real span of the q1, . . . qn, and so
by basic linear algebra, the q1, . . . qn must in fact be linearly independent over R which is
what we wanted.
Since we have seen that V ∩ Qn is an n − k dimensional Q-vector space, there are
n − k, Q-linearly independent vectors in V ∩ Qn ⊆ V ∩ Qn. By the preceding paragraph,
these n − k vectors are linearly independent over R and so V ∩ Qn is a real vector space of
dimension at least n − k. Combined with our earlier observation, we conclude that V ∩ Qn
is a real vector space of dimension exactly n− k. Now V ∩ Qn ⊆ V and both are real vector
spaces of the same dimension, so they must in fact be equal, as required.
We can now prove the following analogous result to Lemma 7.1 of [2].
Lemma 3.4.16. Let m < n be natural numbers and let x ∈ Xk ∩ QΓ, y ∈ Xk be such that
ran x, ran y are both contained in the interval (m, n]. Suppose that dist(y,(cid:104)x, Sx . . . , Sk−1x(cid:105)R) >
δ. Then there exists b∗ ∈ ball (cid:96)1(Γn \ Γm), with rational coordinates, such that (cid:104)b∗, Sjx(cid:105) = 0
for j = 0, 1, . . . k − 1 and b∗(y) > 1
2 δ.
87
(cid:107)y −(cid:80)k−1
j=0 λjSjx(cid:107) ≤ (cid:107)in(cid:107)(cid:107)v −(cid:80)k−1
Proof. It is easily seen (from the construction of S) that since x ∈ Xk ∩ QΓ, with ran x ⊆
(m, n], we have Sjx ∈ Xk ∩ QΓ and ran Sjx ⊆ (m, n] for every j = 0, 1, . . . k − 1. For
each j, let uj ∈ (cid:96)∞(Γn \ Γm) be the restriction of Sjx and let v ∈ (cid:96)∞(Γn \ Γm) be the
restriction of y. Then Sjx = inuj for every j, y = inv and so, for any scalars (λj)k−1
j=0 ,
2 δ and so,
by the Hahn -- Banach Theorem in the finite dimensional space (cid:96)∞(Γn \ Γm), there exists
a∗ ∈ ball (cid:96)1(Γn \ Γm) with a∗(uj) = 0 for every j and a∗(v) > 1
2 δ. Since Sjx has rational
coordinates for every j, our vectors uj are in QΓn\Γm. It follows from Lemma 3.4.15 that
we can approximate a∗ arbitrarily well with b∗ ∈ QΓn\Γm retaining the condition b∗(uj) = 0
for every j.
j=0 λjuj(cid:107). Hence dist(v,(cid:104)u0, u1, . . . , uk−1(cid:105)R) > 1
The proof of Theorem 3.4.9 is now easy. We obtain the following minor variation of [2,
Lemma 7.2]. We omit the proof since exactly the same argument as in [2] works:
Lemma 3.4.17. Let T be a bounded linear operator on Xk, let (xi) be a C-RIS in Xk ∩ QΓ
and assume that dist(T xi,(cid:104)xi, Sxi . . . , Sk−1xi(cid:105)R) > δ > 0 for all i. Then, for all j, p ∈ N,
there exist z ∈ [xi : i ∈ N], q > p and η ∈ ∆q such that
1. (z, η) is a (16C, 2j, 0)-special exact pair with ran z ⊂ (p, q);
2. (T z)(η) > 7
3. (cid:107)(I − P(p,q))T z(cid:107) < m−1
2j δ;
4. (cid:104)P ∗
η, T z(cid:105) > 3
8 δ.
16 δ;
(p,q]e∗
Theorem 3.4.9 is now proved in exactly the same way as that of [2, Proposition 7.3].
3.5 Operators on the Space Xk
In this section, we see that all operators on the space Xk are expressible as(cid:80)k−1
j=0 λjSj +K for
suitable scalars λj and some compact operator K on Xk. We choose to give an elementary
proof of this fact in this chapter. However, in the following chapter, we will construct
a space X∞ which has a similar operator representation; S is no longer nilpotent in this
j=0 λjSj, where the
scalars (λj)∞
j=0 lie in (cid:96)1(N0). The proof of the operator representation in Chapter 4 will be
much more technical, relying on a fixed-point-theorem due to Kakutani. We remark that
case, so we obtain an infinite sum in the powers of S of the form (cid:80)∞
the proof of the operator representation in Chapter 4 would also work (with only obvious
changes) to prove the operator representation we require in this section. With that said,
we now proceed to give a more elementary proof. We will need the following lemmas.
88
(cid:48). Certainly then ran yi = ran xi for every
(cid:13)(cid:13) ≤ 5(cid:13)(cid:13)(cid:101)xi
(cid:48) − xi
(cid:13)(cid:13) → 0.
(cid:48) − xi
Proof. Let ran xi = (pi, qi) and set yi = P(pi,qi)(cid:101)xi
i. Note that(cid:0)I − P(pi,qi)
(cid:1) xi = 0 and consequently
(cid:13)(cid:13)(cid:0)I − P(pi,qi)
(cid:1)(cid:101)xi
It follows that
(cid:48)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:0)I − P(pi,qi)
(cid:48) − xi)(cid:13)(cid:13) ≤(cid:13)(cid:13)I − P(pi,qi)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:101)xi
(cid:1) ((cid:101)xi
(cid:13)(cid:13)
(cid:107)yi − xi(cid:107) =(cid:13)(cid:13)(cid:101)xi
(cid:48) −(cid:0)(I − P(pi,qi))(cid:101)xi
(cid:48)(cid:1) − xi
(cid:48)(cid:13)(cid:13) → 0.
(cid:13)(cid:13) +(cid:13)(cid:13)(cid:0)I − P(pi,qi)
≤(cid:13)(cid:13)(cid:101)xi
(cid:1)(cid:101)xi
γ(cid:105) = (cid:104)(cid:101)xi
γ(cid:105) = (cid:104)SjP(pi,qi)(cid:101)xi
(cid:48), e∗
= (cid:104)(cid:101)xi
γ(cid:105) = (cid:104)Sj(cid:101)xi
(pi,qi)e∗
(cid:48), (S∗)jP ∗
(cid:48) − xi
(cid:48), P ∗
k satisfying
(cid:48) − xi(cid:107) → 0 as i → ∞ and Sj(cid:101)xi
(cid:48)
(cid:48) = 0 for every i. Then there is a subsequence
Lemma 3.5.1. Let 1 ≤ j ≤ k − 1 and (xi)i∈N be a C-RIS in Xk. Suppose there are (cid:101)xi
such that (cid:107)(cid:101)xi
(xik )k∈N of (xi) and vectors x(cid:48)
(1) ran xik = ran x(cid:48)
(2) Sjx(cid:48)
(3) (cid:107)x(cid:48)
(4) (x(cid:48)
We note that in particular, if (xi)i∈N is a C-RIS with Sjxi → 0, then the above hypothesis
are satisfied as a consequence of Corollary 3.4.7.
k = 0 for every k
k − xik(cid:107) → 0
k)k∈N is a 2C-RIS.
k
Note also that Sjyi = 0 for every i. Indeed, for γ ∈ Γ,
Sjyi(γ) = (cid:104)Sjyi, e∗
γ(cid:105)
(pi,qi)(S∗)je∗
(cid:48), P ∗
γ(cid:105) = 0
(pi,qi)e∗
(cid:48) = 0 for every i. (We also made use of the fact that P ∗
(pi,qi)(S∗)j = (S∗)jP(pi,qi).
This follows from the construction of the operator S∗ and the fact that F is rank-preserving.
A more detailed argument of this fact is presented in the proof of Theorem 3.3.6.)
since Sj(cid:101)xi
So far we have managed to achieve (1) - (3) of the above. We show we can extract a
subsequence of the yi, (yik )k∈N say, such that (yik )k∈N is a 2C-RIS. The proof will then be
complete if we set x(cid:48)
k = yik and take the subsequence (xik ) of the xi.
Since (cid:107)xi(cid:107) ≤ C for every i and (cid:107)yi − xi(cid:107) → 0, we can certainly assume (by ignoring
some finite number of terms at the beginning of the sequence) that (cid:107)yi(cid:107) ≤ 2C for every i.
Let (jk) be the increasing sequence corresponding to the C-RIS (xi), i.e.
(i) jk+1 > max ran xk
(ii) xk(γ) ≤ Cm−1
i when weight γ = m−1
i
and i < jk
89
Set l1 = j1. We can certainly find an i1 ≥ 1 such that (cid:107)yi1 − xi1(cid:107) ≤ Cm−1
weight γ = m−1
w with w < l1, then certainly w < l1 = j1 ≤ ji1 so
yi1(γ) ≤ (yi1 − xi1)(γ) + xi1(γ) ≤ Cm−1
l1
+ Cm−1
w ≤ 2Cm−1
w
l1
. So if γ ∈ Γ,
Now set l2 = ji1+1. So l2 > max ran xi1 = max ran yi1.
Inductively, suppose we have defined natural numbers l1 ≤ l2 ≤ ··· ≤ ln and i1 < i2 <
··· < in such that
(i') lk+1 > max ran yik for all 1 ≤ k ≤ n − 1, and
(ii') for all 1 ≤ k ≤ n, yik (γ) ≤ 2Cm−1
w whenever γ ∈ Γ, weight γ = m−1
w with w < lk
It is easily seen from the inductive construction that ln+1 ≥ ln and
Set ln+1 = jin+1.
moreover (by choice of jk), ln+1 > max ran xin = max ran yin. Now we can certainly find
in+1 > in such that (cid:107)yin+1 − xin+1(cid:107) ≤ Cm−1
w with
w < ln+1 In particular w < jin+1 ≤ jin+1, so by choice of in+1 and the fact that (xi) is a
RIS, we see that
. So suppose γ ∈ Γ, weight γ = m−1
ln+1
yin+1(γ) ≤ (yin+1 − xin+1)(γ) + xin+1(γ) ≤ Cm−1
ln+1
+ Cm−1
w ≤ 2Cm−1
w .
Inductively we obtain a subsequence (yik )k∈N which is evidently a 2C-RIS (with the sequence
(lk)k∈N satisfying the RIS definition), as required.
We will also need the following technical lemma.
Lemma 3.5.2. Suppose (xi)i∈N is a normalised sequence in Xk and λj ∈ R (0 ≤ j ≤ k − 1)
are scalars such that(cid:80)k−1
j=0 λjSjxi → 0. If Sk−1xi(cid:88)→ 0 then λj = 0 for every j. Otherwise,
there is 1 ≤ m ≤ k − 1 such that Smxi → 0 but Sjxi(cid:88)→ 0 if j < m, in which case, we must
have that λj = 0 for all j < m.
Proof. We consider first the case where Sk−1xi → 0 and choose m ∈ {1, . . . k − 1} minimal
such that Smxi → 0 (noting such an m obviously exists). We must observe that λj = 0 for
j=0 λjSjxi → 0. If m = 1,
this of course implies that λ0 = 0 since the sequence (xi) is normalised and we are done.
all j < m. Since Sjxi → 0 for all j ≥ m we in fact know that(cid:80)m−1
Otherwise, we apply the operator Sm−1 to the previous limit. Once again making use
Sm−1xi(cid:88)→ 0, we must again have λ0 = 0, and moreover,(cid:80)m−1
of the fact that Sjxi → 0 when j ≥ m, we find that λ0Sm−1xi → 0. Since, by choice of m,
j=1 λjSjxi → 0. If m = 2, then
this implies λ1Sxi → 0 which implies that λ1 = 0 (since Sxi (cid:88)→ 0). Otherwise, we apply
the operator Sm−2. A similar argument concludes once again that we must have λ1 = 0.
Continuing in this way, we get that λj = 0 for all j < m as required.
90
required.
Theorem 3.5.3. Let T : Xk → Xk be a bounded linear operator on Xk. Then there are
j=0 λjSj + K.
Proof. We will show that there exist scalars λj such that whenever (xi)i∈N is a RIS, T xi −
j=0 λjSjxi → 0
j=0 λjSj is
compact. We note that it is enough to show that there are λj ∈ R such that whenever
In the case where Sk−1xi (cid:88)→ 0, we notice that in particular this implies Sjxi (cid:88)→ 0 for
(cid:80)k−1
every 1 ≤ j ≤ k − 1. Applying the operators Sk−1, Sk−2, . . . S sequentially to the limit,
j=0 λjSjxi → 0, first yields that λ0 = 0, then λ1 = 0 etc. So λj = 0 for every j as
λj ∈ R (0 ≤ j ≤ k− 1) and a compact operator K : Xk → Xk such that T =(cid:80)k−1
j=0 λjSjxi → 0 as i → ∞. By Proposition 3.4.4, this implies T xi −(cid:80)k−1
(cid:80)k−1
for every block sequence (xi) which, by Proposition 1.4.15, implies that T −(cid:80)k−1
(xi)i∈N is a RIS, we can find some subsequence (xil) with T xil −(cid:80)k−1
Claim 1. Suppose (xi)i∈N is a normalised RIS and that Sk−1xi(cid:88)→ 0 (noting in particular
(cid:88)→ 0 for every 1 ≤ j ≤ k − 1). Then there are λj ∈ R
(0 ≤ j ≤ k − 1) and a subsequence (xil) of (xi) such that T xil −(cid:80)k−1
i ∈ R such that (cid:107)T xi −(cid:80)k−1
i Sjxi(cid:107) → 0. We first show that the λ0
that this implies that Sjxi
j=0 λjSjxil → 0.
j=0 λjSjxil → 0.
j=0 λj
By passing to a subsequence, we may assume that there is some ε > 0 such that
(cid:107)Sjxi(cid:107) ≥ ε for every i ∈ N and every 1 ≤ j ≤ k − 1. By Theorem 3.4.9, there are
λj
i must converge. The
argument is similar to that of [2]. If not, by passing to a subsequence, we may assume that
2i+1 − λ0
λ0
2i ≥ δ > 0 for some δ. Since yi := x2i−1 + x2i is a RIS, we deduce from Theorem
3.4.9 that there are µj
(cid:16)
≤ (cid:107) k−1(cid:88)
i ∈ R with (cid:107)T yi −(cid:80)k−1
(cid:17)
2i−1Sjx2i−1 − T x2i−1(cid:107) + (cid:107)T yi − k−1(cid:88)
(cid:17)
2iSjx2i − T x2i(cid:107) + (cid:107) k−1(cid:88)
i Sjyi(cid:107) → 0. Now
(cid:107) k−1(cid:88)
2i−1 − µj
λj
Sjx2i−1(cid:107)
2i − µj
λj
i Sjyi(cid:107)
µj
k−1(cid:88)
Sjx2i +
j=0 µj
(cid:16)
i
λj
i
λj
j=0
j=0
j=0
j=0
j=0
and so we deduce that both sides of the inequality converge to 0. Since the sequence (xi) is
a block sequence, there exist lk such that P(0,lk]yk = x2k−1 and P(lk,∞)yk = x2k. Recalling
that if x ∈ Xk has ran x = (p, q] then ran Sjx ⊆ (p, q], we consequently have
(cid:16)
(cid:107) k−1(cid:88)
2i − µj
λj
i
j=0
and similarly
(cid:17)
(cid:17)
(cid:16)
Sjx2i(cid:107) ≤ (cid:107)P(li,∞)(cid:107)(cid:107) k−1(cid:88)
(cid:16)
(cid:107) k−1(cid:88)
j=0
i
2i − µj
λj
(cid:17)
2i−1 − µj
λj
i
Sjx2i−1(cid:107) → 0.
k−1(cid:88)
(cid:16)
j=0
(cid:17)
Sjx2i +
2i−1 − µj
λj
i
Sjx2i−1(cid:107) → 0
j=0
91
By continuity of S and the fact that Sk = 0, applying Sk−1 to both limits above (and
recalling that (cid:107)Sjxi(cid:107) ≥ ε for every 1 ≤ j ≤ k − 1) we obtain
(cid:107)(cid:0)λ0
2i − µ0
i
(cid:1) Sk−1x2i(cid:107) ≤ 1
(cid:107)Sk−1(cid:107)(cid:107) k−1(cid:88)
(cid:16)
ε
j=0
λ0
2i − µ0
i ≤ 1
ε
(cid:17)
2i − µj
λj
i
Sjx2i(cid:107) → 0
and similarly we find that λ0
our assumption. So the λ0
2i−1 → 0 contrary to
i converge to some λ0 as claimed. It follows that (cid:107)T xi − λ0xi −
i Sjxi(cid:107) → 0. We observe that, since (xi) is normalised and (cid:107)Sk−1xi(cid:107) ≥ ε, applying
i → 0. It follows that λ0
(cid:80)k−1
2i−1 − µ0
2i − λ0
j=1 λj
Sk−2 to the previous limit, we see that
λ1
i ≤ 1
ε
(cid:107)λ1
i Sk−1xi(cid:107) ≤ 1
ε
(cid:0)(cid:107)Sk−2(cid:107)(cid:107)T xi − λ0xi − k−1(cid:88)
i Sjxi(cid:107) + (cid:107)Sk−2T − λ0Sk−2(cid:107)(cid:1)
λj
j=1
so that in particular the λ1
quence λ1
il
(limit λ1) of the λ1
i are bounded. Consequently there is some convergent subse-
i . It follows that the corresponding subsequence (xil) satisfies
T xil − λ0xil − λ1Sxil − k−1(cid:88)
Sjxil → 0
λj
il
j=2
Now, if k = 2 we are done (the last sum is empty). Otherwise, we can apply Sk−3 to the
)∞
previous limit and use the same argument to conclude that (λ2
l=1 is a bounded sequence
il
of scalars. Continuing in this way, we eventually find (after passing to further subsequences
which we relabel as xil) that there are λj with (T −(cid:80)k−1
subsequence (xil) of (xi) such that (T −(cid:80)m0−1
Claim 2. Suppose (xi)i∈N is a normalised C-RIS and that Smxi → 0 for some 1 ≤ m ≤ k−1.
Let m0 ≥ 1 be minimal such that Sm0xi → 0. Then there are λj ∈ R (0 ≤ j < m0) and a
j=0 λjSj)xil → 0 as required.
j=0 λjSj)xil → 0.
By minimality of m0, we can assume (by passing to a subsequence if necessary) that
(cid:107)Sjxi(cid:107) ≥ ε for all i ∈ N and all j < m0. By Lemma 3.5.1, (with j = m0) there is a 2C-RIS
l ∈ Ker Sm0 ⊆ Ker Sj for j ≥ m0 and
(x(cid:48)
l)l∈N and some subsequence (xil) of (xi) such that x(cid:48)
every l. Moreover, (cid:107)xil − x(cid:48)
l(cid:107) → 0 (as l → ∞). By Theorem 3.4.9, there are λj
l s.t
l − k−1(cid:88)
(cid:107)T x(cid:48)
l − m0−1(cid:88)
l Sjx(cid:48)
λj
l(cid:107) = (cid:107)T x(cid:48)
l Sjx(cid:48)
λj
l(cid:107) → 0.
j=0
j=0
We claim the λ0
l must converge to some λ0. The argument is the same as that used in Claim
1, except now we obtain
(cid:107) m0−1(cid:88)
(cid:16)
(cid:17)
2l−1(cid:107) → 0 and (cid:107) m0−1(cid:88)
Sjx(cid:48)
(cid:16)
2l − µj
λj
l
(cid:17)
Sjx(cid:48)
2l(cid:107) → 0.
2l−1 − µj
λj
l
j=0
j=0
92
2k−1 + x(cid:48)
2j−1 → 0 and consequently it follows that λ0
2k also lies in Ker Sm0). We apply Sm0−1, noting that Sjx(cid:48)
(We note there are no terms of the form 'Sjy' when j ≥ m0 since the RIS (yk) defined by
yk := x(cid:48)
l = 0 for every
l and j ≥ m0. We reach the same contradiction as in the previous argument, i.e. that
λ0
2j − λ0
l must converge to some λ0 as claimed.
l → 0. We use the same argument as above
k − λ0x(cid:48)
It easily follows that T x(cid:48)
l=1 for j = 1, 2 . . . , m0 − 1 all have convergent subsequences
l )∞
to show that the sequences (λj
→ 0.
and consequently that we can find some subsequence x(cid:48)
It follows that
l −(cid:80)m0−1
j=0 λjSj)x(cid:48)
j=1 λj
l Sjx(cid:48)
lr
− m0−1(cid:88)
j=0
(cid:107)T xilr
λjSjxilr
(cid:107) ≤ (cid:107)(cid:0)T − m0−1(cid:88)
≤ (cid:107)T − m0−1(cid:88)
j=0
λjSj(cid:1)(xilr
− x(cid:48)
λjSj(cid:107)(cid:107)xilr
− x(cid:48)
lr
with (T −(cid:80)m0−1
lr )(cid:107) + (cid:107)(cid:0)T − m0−1(cid:88)
lr(cid:107) + (cid:107)(cid:0)T − m0−1(cid:88)
λjSj(cid:1)(x(cid:48)
λjSj(cid:1)(x(cid:48)
j=0
lr )(cid:107)
lr )(cid:107) → 0.
j=0
j=0
A priori, the λj found in Claims 1 and 2 may depend on the RIS. We see now that this is
not the case.
Claim 3. There are λj ∈ R (0 ≤ j ≤ k − 1) such that whenever (xi)i∈N is a RIS, there is a
subsequence (xil) of (xi) such that T xil −(cid:80)k−1
j=0 λjSjxil → 0.
Note that if (xi)i∈N is a RIS with some some subsequence converging to 0 then any λj
can be chosen satisfying the conclusion of the claim. So it is sufficient to only consider
normalised RIS. We begin by choosing a normalised RIS, (xi)i∈N, with Sk−1xi(cid:88)→ 0. Note
that such a RIS must exist. Indeed, if not, then Sk−1xi → 0 whenever (xi)i∈N is a RIS and
by the argument in the first paragraph of the proof, this would imply that Sk−1 is compact,
contradicting Corollary 3.4.8. It follows from Claim 1, after passing to a subsequence and
relabelling if necessary, that there are λj ∈ R with
λjSjxi → 0.
(1)
T xi − k−1(cid:88)
j=0
j ∈ R with
i − k−1(cid:88)
j=0
T x(cid:48)
Now suppose (x(cid:48)
)l∈N, of (x(cid:48)
subsequence, (x(cid:48)
proving the claim.
relabelling if necessary, that there are λ(cid:48)
il
i)i∈N is any normalised RIS in Xk. We will show that there is some
, thus
i) such that Equation (1) holds when xi is replaced by x(cid:48)
il
It follows from Claims 1 and 2, after passing to a subsequence and
λ(cid:48)
jSjx(cid:48)
i → 0.
(2)
We pick natural numbers i1 < i2 < . . . and j1 < j2 . . . such that max ran xik < min ran x(cid:48)
max ran x(cid:48)
< min ran xik+1 for every k and such that the sequence (xik + x(cid:48)
≤
)k∈N is again
jk
jk
jk
93
a RIS. For notational convenience, we (once again) relabel the subsequences (xik ), (x(cid:48)
(xi) and (x(cid:48)
numbers li such that P(0,li](xi + x(cid:48)
i) = xi and P(li,∞)(xi + x(cid:48)
Claims 1 and 2 that there are µj and a subsequence (xim + x(cid:48)
) by
i)i∈N is a RIS and there are natural
i. It follows again from
i). So (by choice of the subsequences) (xi + x(cid:48)
i) = x(cid:48)
) such that
jk
im
(3)
j=0
im) − k−1(cid:88)
i − k−1(cid:88)
(cid:0)T x(cid:48)
i − k−1(cid:88)
(cid:0)T x(cid:48)
j=0
j=0
im) → 0.
µjSj(xim + x(cid:48)
k−1(cid:88)
(cid:1) +
(cid:1) → 0
λ(cid:48)
jSjx(cid:48)
λ(cid:48)
jSjx(cid:48)
j=0
i
i
T (xim + x(cid:48)
= P(0,li]
P(0,li]T x(cid:48)
i = P(0,li]
λ(cid:48)
jP(0,li]Sjx(cid:48)
i
We note also that
and similarly, P(li,∞)T xi → 0. Passing to the appropriate subsequences of Equations (1)
and (2) and substracting them from Equation (3) we see that
T x(cid:48)
im − k−1(cid:88)
T xim − k−1(cid:88)
j=0
(µj − λj)Sjxim − k−1(cid:88)
im − k−1(cid:88)
j)Sjx(cid:48)
(µj − λ(cid:48)
j=0
j=0
j=0
µjSjx(cid:48)
im → 0
µjSjxim → 0.
(4)
(5)
and
Finally we apply the projections P(0,lim ] and P(lim ,∞) to Equations (4) and (5) respectively
to obtain (using the above observations) that
k−1(cid:88)
k−1(cid:88)
(µj − λj)Sjxim → 0 and
(µj − λ(cid:48)
j)Sjx(cid:48)
im → 0.
j=0
j=0
j = µj = λj for every j. So, we can replace λ(cid:48)
limits implies that we must have λj = µj for all j. We now consider two cases:
(i) Sk−1x(cid:48)
Since (xi)i∈N was chosen such that Sk−1xi (cid:88)→ 0, Lemma 3.5.2 and the first of the above
im (cid:88)→ 0. By Lemma 3.5.2 and the second of the above limits, we see that we
im (cid:88)→ 0 for any j < r.
i −(cid:80)k−1
j = µj = λj for all j < r. So, replacing λ(cid:48)
i → 0.
jSjx(cid:48)
→ 0, so that
must have λ(cid:48)
we see that Equation (1) does indeed hold with (xi) replaced by (x(cid:48)
i).
Again, by Lemma 3.5.2, we must have λ(cid:48)
by λj for j < r in Equation (2), we find that T x(cid:48)
Since Slx(cid:48)
Equation (1) holds with xi replaced by x(cid:48)
(ii) There is some 1 ≤ r ≤ k − 1 such that Srx(cid:48)
→ 0 for all l ≥ r, it is now clear that T x(cid:48)
i −(cid:80)r−1
j by λj in Equation (2) and
→ 0, but Sjx(cid:48)
j=0 λjSjx(cid:48)
−(cid:80)k−1
j=0 λjSjx(cid:48)
j=r λ(cid:48)
im
im
im
im
.
j
im
This completes the proof of the claim and thus (as noted earlier), the proof.
94
3.6 Strict Singularity of S : Xk → Xk
In this section we will prove that S is strictly singular. By Proposition 1.4.19, it is enough to
see that S is not an isomorphism when restricted to any infinite dimensional block subspace
Z of Xk. We begin by stating a result taken from the paper of Argyros and Haydon, [2,
Corollary 8.5]. The reader can check that the same proofs as given in [2] will also work for
the spaces Xk constructed here.
Lemma 3.6.1. Let Z be a block subspace of Xk, and let C > 2 be a real number. Then Z
contains a normalized C-RIS.
We will need a variation of Lemma 3.4.11 to be able to construct weak dependent
sequences. We first observe that the lower norm estimate for skipped block sequences given
in Proposition 4.8 of [2] also holds in the space Xk and exactly the same proof works. We
state it for here for convenience:
Lemma 3.6.2. Let (xr)a
that a ≤ n2j and 2j < min ran x2, then there exists an element γ of weight m−1
r=1 be a skipped block sequence in Xk. If j is a positive integer such
2j satisfying
a(cid:88)
xr(γ) ≥ 1
2 m−1
2j
a(cid:88)
(cid:107)xr(cid:107).
(cid:107)xr(cid:107).
Hence
r=1
(cid:107) a(cid:88)
r=1
xr(cid:107) ≥ 1
2 m−1
2j
r=1
a(cid:88)
r=1
Lemma 3.6.3. Let ε > 0, j be a positive integer and let (xi)n2j
such that min ran x2 > 2j. Suppose further that one of the following hypotheses holds:
(i) (cid:107)Sk−1xi(cid:107) ≥ δ for all i (some δ > 0)
(ii) There is some 2 ≤ m ≤ k − 1 (where we are, of course, assuming here that k > 2)
i=1 be a skipped-block C-RIS,
such that (cid:107)Sm−1xi(cid:107) ≥ δ for all i (some δ > 0) and (cid:107)Smxi(cid:107) ≤ Cm−1
2j ε.
Then there exists η ∈ Γ such that x(η) ≥ δ
2 where x is the weighted sum
n2j(cid:88)
x = m2jn−1
2j
xi.
Moreover, if hypothesis (i) above holds, the pair (Sx, η) is a (16C, 2j, 0)-special exact pair.
Otherwise, hypothesis (ii) holds and (Sx, η) is a (16C, 2j, 0, ε) weak exact pair.
i=1
95
Proof. Let us consider first the case where hypothesis (i) holds. Since (Sk−1xi)n2j
i=1 is a
skipped block sequence, it follows from Lemma 3.6.2 that there is an element γ ∈ Γ of
weight m−1
2j satisfying
n2j(cid:88)
n2j(cid:88)
m2jn−1
2j
Sk−1xi(γ) ≥ 1
2 n−1
2j
i=1
i=1
(cid:107)Sk−1xi(cid:107) ≥ δ
2
Consequently, we must have F k−1(γ) being defined, and x(F k−1(γ)) ≥ δ
2 . We set η =
η(cid:105) = 0 for any j ≥ 0 (since, by Lemma
F k−1(γ) ∈ Γ. Certainly SjSx(η) = (cid:104)x, (S∗)j+1e∗
3.3.8, we must have F (η) = F k(γ) being undefined). So conditions 3 and 4 are satisfied
for (Sx, η) to be a (16C, 2j, 0)-special exact pair. The other conditions are satisfied since a
careful examination of the corresponding argument in [2], i.e. the proof of [2, Lemma 6.2],
reveals that the remaining conditions will hold for the weighted sum x, regardless of the
specific element η of weight m−1
θ ∈ Γ
2j chosen. We then simply make use of the fact that for any
and similarly
(cid:104)Sx, e∗
θ(cid:105) =
(cid:104)Sx, d∗
θ(cid:105) =
(cid:40)(cid:104)x, e∗
(cid:40)(cid:104)x, d∗
0
F (θ)(cid:105)
F (θ)(cid:105)
if F (θ) is defined
otherwise
if F (θ) is defined
0
otherwise
In the case where hypothesis (ii) holds, we find by the same argument as above that
there is a γ ∈ Γ of weight m−1
2j with F m−1(γ) being defined and x(F m−1(γ)) ≥ δ
2 .
We now set η = F m−1(γ). Now, for any 0 ≤ j ≤ k − 1, either SjSx(η) = 0 (if
(cid:80)n2j
F j+1(η) = F j+m(γ) is undefined) or SjSx(η) = x(F j+m(γ)) = Sj+mx(γ) ≤ (cid:107)Sj+mx(cid:107) ≤
m2jn−1
i=1 (cid:107)Sj+mxi(cid:107) ≤ Cε. The final inequality here is obtained using the fact that
(cid:107)Sj+mxi(cid:107) ≤ Cm−1
2j ε for all i and j ≥ 0; indeed, when j = 0, this is the hypothesis in the
lemma, and when j > 0, we simply use the fact that S has norm at most 1. We have
shown conditions (3) and (4(cid:48)) hold for (Sx, η) to be a (16C, 2j, 0, ε) weak exact pair. The
remaining conditions hold once again because they hold for x; the argument is the same as
2j
before.
Theorem 3.6.4. The operator S : Xk → Xk is strictly singular.
Proof. We suppose by contradiction that S is not strictly singular. It follows that there is
some infinite dimensional block subspace Y of Xk on which S is an isomorphism, i.e. there
is some 0 < δ ≤ 1 such that whenever y ∈ Y,(cid:107)Sy(cid:107) ≥ δ(cid:107)y(cid:107). By Lemma 3.6.1, Y contains a
normalised skipped block 3-RIS, (xi)i∈N ⊆ Y . We note that certainly Sxi(cid:88)→ 0 and consider
two possibilities. Either Sk−1xi → 0 or it does not.
In the latter of these possibilities,
passing to a subsequence, we can assume without loss of generality that (cid:107)Sk−1xi(cid:107) ≥ ν > 0
96
for every i (and some ν). Thus, we see by Lemma 3.6.3 that we can construct (48, 2j, 0)
special exact pairs (Sx, η) for any j ∈ N, with min ran Sx arbitrarily large and x(η) ≥ ν
2 .
Otherwise, we must have k > 2 and there is an m ∈ {2, . . . , (k − 1)} with Sm−1xi(cid:88)→ 0
but Smxi → 0. By passing to a subsequence, we can assume that (cid:107)Sm−1xi(cid:107) ≥ ν for all i.
Moreover, for a fixed j0 ∈ N, since Smxi → 0, given any j ∈ N, we can find an Nj ∈ N such
that (cid:107)Smxi(cid:107) ≤ Cm−1
2j0−1 for every i ≥ Nj. So by Lemma 3.6.3, we can construct weak
(48, 2j, 0, n−1
2j0−1) exact pairs (Sx, η) for any j ∈ N, with min ran Sx arbitrarily large and
x(η) ≥ ν
2 .
2j n−1
Now, we choose j0, j1 with m2j0−1 > 6720δ−1ν−1 and m4j1 > n2
and the argument above, there is a y1 ∈ Y, η1 ∈ Γ such that (Sy1, η1) is a (48, 4j1, 0, n−1
weak exact pair and y1(η1) ≥ ν
be (p1, 0, m2j0−1, e∗
2j0−1. By Lemma 3.6.3,
2j0−1)-
2 . We let p1 > rank η1 ∨ max ran y1 and define ξ1 ∈ ∆p1 to
Now set j2 = σ(ξ1). Again by Lemma 3.6.3 and the argument above, there is y2 ∈
2j0−1)-weak exact
η2),
Y, η2 ∈ Γ with min ran y2 > p1, y2(η2) ≥ ν
pair. We pick p2 > rank η2 ∨ max ran y2 and take ξ2 to be the element (p2, ξ1, m2j0−1, e∗
noting that this tuple is indeed in ∆p2.
2 and (Sy2, η2) a (48, 4j2, 0, n−1
η1).
Continuing in this way, we obtain a (48, 2j0 − 1, 0, n−1
2j0−1)-weak dependent sequence
(Syi). By Proposition 3.4.14 we see that
with the final inequality following by the choice of j0. On the other hand,
n2j0−1(cid:88)
i=1
2j0−1
(cid:107)m2j0−1n−1
n2j0−1(cid:88)
n2j0−1(cid:88)
Syi(cid:107) ≤ 70 × 48m−1
2j0−1 <
δν
2
yi(ξn2j0−1) = m−1
2j0−1
i=1
i=1
yi(ηi) ≥ m−1
2j0−1n2j0−1
ν
2
So,
(cid:107)m2j0−1n−1
2j0−1
n2j0−1(cid:88)
i=1
n2j0−1(cid:88)
n2j0−1(cid:88)
i=1
Syi(cid:107) ≥ δ(cid:107)m2j0−1n−1
2j0−1
yi(cid:107)
≥ δm2j0−1n−1
2j0−1
yi(ξn2j0−1) ≥ δν
2
This contradiction completes the proof.
i=1
The previous theorem combined with Theorem 3.5.3 and Corollary 1.4.27 immediately
yield the following corollary.
Corollary 3.6.5. The operators Sj : Xk → Xk (j ≥ 1) are strictly singular. The spaces Xk
have few but not very few operators.
97
3.7 The HI Property
It only remains to see that the spaces Xk are hereditarily indecomposable. The proof is
sufficiently close to the corresponding proof of [2] that we will omit most of the details. We
first observe we have the following generalisations of Lemmas 8.8 and 8.9 of [2].
Lemma 3.7.1. Let (xi)i≤n2j0−1 be a (C, 2j0 − 1, 1)−weak dependent sequence in Xk and let
J be a sub-interval of [1, n2j0−1]. For any γ(cid:48) ∈ Γ of weight m2j0−1 we have
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88)
i∈J
(−1)ixi(γ(cid:48))
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 7C.
n2j0−1(cid:88)
i=1
It follows that
(cid:107)n−1
2j0−1
n2j0−1(cid:88)
i=1
xi(cid:107) ≥ m−1
2j0−1
but
(cid:107)n−1
2j0−1
(−1)ixi(cid:107) ≤ 70Cm−2
2j0−1.
Proof. The proof of the first claim is sufficiently close to the proof of Lemma 3.4.13 that
we omit any more details. The upper bound
n2j0−1(cid:88)
(cid:107)n−1
2j0−1
(−1)ixi(cid:107) ≤ 70Cm−2
2j0−1
then follows using exactly the same argument as in Proposition 3.4.14. The lower bound is
proved in the same way as Lemma 8.9 of [2]; one simply observes that using the notation
i=1
d∗
ξi
+ m−1
2j0−1
n2j0−1(cid:88)
i=1
ηi(cid:105)
(pi−1,pi)e∗
P ∗
of Definition 3.4.12,
n2j0−1(cid:88)
xi(ξn2j0−1) = (cid:104)
i=1
i=1
= m2j0−1
n2j0−1(cid:88)
n2j0−1(cid:88)
n2j0−1(cid:88)
i=1
xi,
i=1
We immediately obtain
(cid:107)n−1
2j0−1
n2j0−1(cid:88)
i=1
2j0−1.
xi(ηi) = n2j0−1m−1
n2j0−1(cid:88)
xi(cid:107) ≥ n−1
2j0−1
xi(ξn2j0−1) = m−1
2j0−1.
i=1
To see the spaces Xk are HI, we claim it will be enough to see that we have the following
lemma:
Lemma 3.7.2. Let Y be a block subspace of Xk. There exists δ > 0 such that whenever j, p ∈
N, ε > 0, there exists q ∈ N, x ∈ Y, η ∈ Γ with ran x ⊆ (p, q) and (x, η) a (96δ−1, 2j, 1, ε)
weak exact pair.
98
We omit the proof of Lemma 3.7.2. It is essentially the same as Lemma 3.6.3 combined
with the proof of [2, Lemma 8.6].
Proposition 3.7.3. Xk is hereditarily indecomposable.
Proof. The argument is the same as in [2]; by Proposition 1.4.10 it is enough to show that
if Y and Z are block subspaces of Xk, then for each ε > 0 there exists a y ∈ Y and z ∈ Z
with (cid:107)y − z(cid:107) < ε(cid:107)y + z(cid:107).
By Lemma 3.7.2, given two block subspaces Y and Z of Xk there exists some δ > 0
such that for all j0 ∈ N, we can construct (96δ−1, n2j0−1, 1)−weak dependent sequences,
(xi)i≤n2j0−1 with xi ∈ Y when i is odd and xi ∈ Z when i is even. We choose j0 ∈ N such
that m2j0−1 > 6720δ−1ε−1 and obtain a (96δ−1, n2j0−1, 1)−weak dependent sequence as in
the preceding sentence. We define
(cid:88)
(cid:88)
xi
y =
xi and z =
i odd
i even
n2j0−1(cid:88)
i=1
xi(cid:107) ≥ n2j0−1m−1
2j0−1
(−1)ixi(cid:107) ≤ 70 · 96δ−1n2j0−1m−2
2j0−1
and observe that by Lemma 3.7.1
(cid:107)y + z(cid:107) = (cid:107)
while
n2j0−1(cid:88)
(cid:107)y − z(cid:107) = (cid:107)
so that (cid:107)y − z(cid:107) < ε(cid:107)y + z(cid:107) as required.
i=1
99
Chapter 4
A Banach space with (cid:96)1 Calkin
algebra
4.1 The Main Theorem
In this chapter, we push the results of the preceding chapter a little further. We construct
another (cid:96)1 predual, X∞, which shares some of the properties of the previously constructed
spaces Xk; our new space is still an L∞ space of Bourgain-Delbaen type and we will once
again obtain a representation formula for all bounded linear operators on X∞. However,
the space X∞ also has some radically different properties from the spaces in the previous
chapter - the Calkin algebra is isomorphic as a Banach algebra to (cid:96)1(N0), the space fails
to have few operators and moreover, the space is indecomposable, but is not hereditarily
indecomposable. Our main result is the following:
Theorem 4.1.1. There exists a separable L∞ space with a Schauder basis, X∞, that pos-
sesses the following properties:
1. X∗∞ = (cid:96)1.
2. There exists a non-trivial, non-compact operator 'S' on X∞. (By 'non-trivial', we
simply mean that the operator S is not a scalar multiple of the identity.) The sequences
(Sj)∞
equivalent to the canonical basis of (cid:96)1(N0).
j=0 ⊂ L(X∞)/K(X∞) are basic sequences isometrically
3. If T ∈ L(X∞) then there are unique scalars (λi)∞
i=0 and a compact operator K ∈
j=1 ⊂ L(X∞) and (cid:0)[Sj](cid:1)∞
L(X∞) with(cid:80)∞
i=0 λi < ∞ and
∞(cid:88)
i=0
T =
λiSi + K.
The operator S appearing in the above sum is the same operator on X∞ described in
Property 2 above.
100
4. The Calkin algebra L(X∞)/K(X∞) is isometric as a Banach algebra to the convolution
algebra (cid:96)1(N0).
For clarity, we remark that when we talk of (cid:96)1(N0) as a Banach algebra, we are thinking
of the Banach space (cid:96)1(N0) equipped with the usual multiplication coming from convolution.
That is, if (an)∞
n=0 ∈ (cid:96)1(N0) then the product (an)∞
n=0 ∗ (bn)∞
n=0 is defined by
n=0, (bn)∞
(cid:0)(an) ∗ (bn)(cid:1)(m) :=
m(cid:88)
ajbm−j.
j=0
It turns out that possessing (cid:96)1 Calkin algebra is a rather remarkable property. Before
proving the main theorem, we note some corollaries that demonstrate this.
Proposition 4.1.2. If X is a Banach space for which the Calkin algebra is isomorphic (as
a Banach algebra) to (cid:96)1(N0) then the strictly singular and compact operators on X coincide.
A particular consequence of the proposition is, of course, that the strictly singular and
compact operators on X∞ coincide. Consequently X∞ fails to have few operators; the
operator S cannot be written as a compact perturbation of the identity by property (2) of
the theorem, and since the compact and strictly singular operators coincide, S cannot be
written as a strictly singular perturbation of the identity.
Before giving the proof of Proposition 4.1.2, we document the following easy lemma:
j=0 and b = (bj)∞
Lemma 4.1.3. Let a = (aj)∞
j=0 be non-zero vectors in (cid:96)1(N0) and choose
k, l ≥ 0 minimal such that ak (cid:54)= 0 and bl (cid:54)= 0. Then a∗ b(j) is equal to 0 whenever j < k + l.
When j = k + l, a ∗ b(j) = akbl. In particular, (cid:107)a ∗ b(cid:107)(cid:96)1(N0) ≥ akbl and consequently
(cid:107)an(cid:107)(cid:96)1(N0) ≥ akn.
Proof. The proof is an easy computation using the definition of the convolution product.
Proof of 4.1.2. If T is strictly singular then by Theorem 1.4.28
n→∞(cid:107)[T ]n(cid:107) 1
lim
n = lim
n→∞(cid:107)T n + K(X)(cid:107) 1
n = 0
n=0 ∈ (cid:96)1(N0) and the above observation yields that limn→∞ (cid:107)an(cid:107) 1
Since the Calkin algebra is isomorphic as a Banach algebra to (cid:96)1(N0), we can identify [T ] with
an element a = (an)∞
n = 0.
By the previous lemma, if a (cid:54)= 0, then letting k be minimal such that ak (cid:54)= 0, (cid:107)an(cid:107)(cid:96)1 ≥ akn,
so that limn→∞ (cid:107)an(cid:107) 1
n ≥ ak > 0. So we must have that a = 0, and therefore that [T ] = 0.
This is the same as saying that T is compact.
We have therefore shown that every strictly singular operator is compact. The converse
is always true.
101
We can also easily deduce from the previous proposition that the space X∞ fails to be
hereditarily indecomposable.
Proposition 4.1.4. The space X∞ is not hereditarily indecomposable.
Proof. It was shown in [12] that for a real HI space X, the quotient algebra L(X)/SS(X)
has dimension at most 4 (where, as usual, SS(X) denotes the ideal of strictly singular
operators on X). By the previous proposition, SS(X∞) = K(X∞), and so we see that
L(X∞)/SS(X∞) is isomorphic to (cid:96)1(N0) by the final property of the main theorem. Com-
bining these observations, it is clear that X∞ is not HI.
On the other hand, the fact that the Calkin algebra is (cid:96)1(N0) forces the space to be
indecomposable.
Proposition 4.1.5. If X is a Banach space for which the Calkin algebra is isomorphic (as
a Banach algebra) to (cid:96)1(N0) then X is indecomposable.
i=0 ∈ (cid:96)1(N0) It follows that (ai)∞
Proof. If P is a projection on X, then [P ] is an idempotent in the Calkin algebra. We
identify [P ] with (ai)∞
i=0 is an idempotent in the Banach
algebra (cid:96)1(N0). An easy computation yields that the only idempotents in (cid:96)1(N0) are 0 and
(1, 0, 0, . . . ). Thus either [P ] = 0 or [I − P ] = 0, i.e. either P or I − P is compact. In either
case, P is certainly a trivial projection.
4.2 The Construction
The construction of X∞ is almost identical to that used in the previous chapter to obtain the
spaces Xk, though the proof of the operator representation for the new space is somewhat
more involved. We continue to work with the fast increasing sequences of natural numbers
(mj)∞
3.3.1 for the precise assumptions on these sequences.
j=1 from the previous chapter and refer the reader back to Assumption
j=1 and (nj)∞
We begin by working with a Bourgain-Delbaen space generated by a countable set
Υ = ∪nΛn which is essentially the same as in the construction of the previous chapter
except for the addition of some new elements which will have degenerate corresponding
BD functionals. Continuing with the terminology from the previous chapter, we will define
these additional elements as having age and weight 0 and refer to them as 'trivial elements'
since their corresponding BD functionals will just be 0. Let us be more precise.
102
Definition 4.2.1. We define Υ by the recursion Λ1 = {0},
Λn+1 =
(n + 1, p, m−1
, b∗) : 0 ≤ p < n, b∗ ∈ Bp,n
j
n+1(cid:91)
∪ n−1(cid:91)
j=1
(cid:110)
p(cid:91)
(cid:110)
(cid:111)
(cid:111)
(n + 1, ξ, m−1
j
, b∗) : ξ ∈ Λp, weight ξ = m−1
j
, age ξ < nj, b∗ ∈ Bp,n
p=1
j=1
∪ {(n + 1, j) : 0 ≤ j ≤ n}
The notation is exactly as before, whereby elements of Υ code the corresponding BD
functional. For a trivial element, γ = (n + 1, j) ∈ Λn+1, we define the corresponding c∗
γ to
be 0 and so d∗
γ. It is readily checked that Theorem 2.2.2 still holds with the addition
of this degenerate case (indeed, within the notation of the theorem, we could just as well
γ = e∗
consider τ (γ) to be the Type 0 tuple, τ (γ) = (1, 0, 0), where the first coordinate sets ε = 1,
the second coordinate denotes the 0 scalar, and the third coordinate denotes the element
0 ∈ Λ1). Consequently we obtain a Bourgain-Delbaen space X(Υ).
As in the previous chapter, we bring to the attention of the reader the same remark
that was made in the previous chapter (see below). We once again ask the reader to consult
Chapter 3 for an explanation of the notation being used.
Remark 4.2.2.
Later on, we will want to take a suitable subset of Υ. To avoid any ambiguity in notation,
all linear combinations b∗ =(cid:80)
in the above definition, and throughout the rest of the chapter, Bp,n will denote the set of
η aη ≤ 1 and each aη is a
aηe∗
η∈Υn\Υp
η, where, as before,(cid:80)
rational number with denominator dividing Nn!.
We have the following theorem, which is the analogue of Theorem 3.3.6 and Lemma
3.3.7.
Theorem 4.2.3. There is a map G : Υ → Υ ∪ {undefined} (we say G(γ) is undefined
if G(γ) = undefined, otherwise we say G(γ) is defined) and a norm 1, linear mapping
R∗ : (cid:96)1(Υ) → (cid:96)1(Υ) satisfying:
1. G(0) is undefined (where we recall Λ1 = {0}).
2. G ((n + 1, j)) = (n + 1, j − 1) whenever 1 ≤ j ≤ n and G ((n + 1, 0)) is undefined.
3. For non-trivial elements γ ∈ Υ \ Λ1, if G(γ) is defined, rank γ = rank G(γ) and
weight γ = weight G(γ) (i.e. G preserves weight and rank). Moreover, age G(γ) ≤
age γ (G doesn't increase age).
103
4.
5.
R∗(e∗
γ) =
R∗(d∗
γ) =
(cid:40)
(cid:40)
e∗
G(γ)
0
d∗
0
G(γ)
if G(γ) is defined
otherwise.
if G(γ) is defined
otherwise.
Moreover, the dual operator of R∗ restricts to give a bounded linear operator R : X(Υ) →
X(Υ) of norm at most 1.
We omit the proof, since it is essentially the same as the proof given for the corresponding
results in the previous chapter. The following observation about the operator R∗ will be
essential later in the proof.
Lemma 4.2.4. For every γ ∈ Υ, there is a unique l ∈ N, 1 ≤ l ≤ rank γ such that Gj(γ)
is defined whenever 1 ≤ j < l but Gl(γ) undefined. Consequently, (R∗)je∗
γ = 0
whenever j ∈ N, j ≥ rank γ.
γ = (R∗)jd∗
Proof. The proof is an easy modification of the proof of Lemma 3.3.8.
As with the constructions in [2] and the previous chapter, for our main construction, we
will place some restrictions on the elements of odd-weight via the use of a coding function
σ : Υ → N. We make the same assumptions on this coding function as in the previous
chapter; we refer the reader back to Section 3.3 for the details. As before, the coding
function can be constructed recursively as Υ is constructed. For each γ ∈ Υ, we define a
finite set Σ(γ) by
Σ(γ) := {σ(γ)} ∪
{σ(δ)}
∞(cid:91)
(cid:91)
j=1
δ∈G−j (γ)
where G−j(γ) := {θ ∈ Υ : Gj(θ) = γ}. (In particular, if θ ∈ G−j(γ) then Gl(θ) ∈ Υ for
every l ≤ j.)
It is easily checked that Lemmas 3.3.10 and 3.3.11 still hold with (essentially) the same
proofs and that we have the following modification of Lemma 3.3.12.
Lemma 4.2.5. Suppose γ, δ ∈ Υ.
1 ≤ j < ∞ such that Gj(γ) = δ.
If σ(γ) ∈ Σ(δ) then either γ = δ or there is some
We are now in a position to describe the main construction. The ideas are exactly the
same as in Chapter 3. We take a subset Γ ⊂ Υ by placing restrictions on the elements of odd
weight we permit. As a consequence of imposing these additional odd weight restrictions, we
are also forced to (roughly speaking) remove those elements (n+1, p, β, b∗) and (n+1, ξ, β, b∗)
104
of Υ for which the support of b∗ is not contained in Γ, in order that we can apply the
Bourgain-Delbaen construction to obtain a space X(Γ). Exactly as was the case in Chapter
3, the subset Γ is also constructed inductively, so it will be well defined and consistent with
the Bourgain-Delbaen method.
We will denote by ∆n the set of all elements in Γ having rank n, and denote by Γn the
union Γn = ∪j≤n∆j. The permissible elements of odd weight will be as follows. For an age
1 element of odd weight, γ = (n + 1, p, m−1
2j−1, b∗) we insist that either b∗ = 0 or b∗ = e∗
η
where η ∈ Γn \ Γp and weight η = m−1
2j−1. For an odd weight element of age > 1,
γ = (n + 1, m−1
η where η ∈ Γn \ Γrank ξ and
weight η = m−1
Definition 4.2.6. We define recursively sets ∆n ⊆ Λn. Then Γn := ∪j≤n∆j and Γ :=
∪n∈N∆n ⊆ Υ. To begin the recursion, we set ∆1 = Λ1. Then
2j−1, ξ, b∗) we insist that either b∗ = 0 or b∗ = e∗
4k < n−2
2j−1, k ∈ Σ(ξ). Let us be more precise:
4i < n−2
∆n+1 =
(n + 1, p, m−1
2j , b∗) : 0 ≤ p < n, b∗ ∈ Bp,n ∩ (cid:96)1(Γn)
(n + 1, ξ, m−1
2j , b∗) : ξ ∈ ∆p, w(ξ) = m−1
(cid:111)
2j , age ξ < n2j, b∗ ∈ Bp,n ∩ (cid:96)1(Γn)
(cid:111)
j=1
(cid:110)
(cid:98)(n+1)/2(cid:99)(cid:91)
(cid:110)
(cid:98)p/2(cid:99)(cid:91)
∪ n−1(cid:91)
(cid:110)
(cid:98)(n+2)/2(cid:99)(cid:91)
p=1
j=1
∪
j=1
∪ n−1(cid:91)
(cid:98)(p+1)/2(cid:99)(cid:91)
p=1
j=1
(cid:110)
(n + 1, p, m−1
2j−1, b∗) : 0 ≤ p < n, b∗ = 0 or b∗ = e∗
η, η ∈ Γn \ Γp
(cid:111)
4i < n−2
2j−1
and w(η) = m−1
(n + 1, ξ, m−1
2j−1, b∗) : ξ ∈ ∆p, w(ξ) = m−1
b∗ = 0 or b∗ = e∗
η with η ∈ Γn \ Γp, w(η) = m−1
2j−1, k ∈ Σ(ξ)
2j−1, age ξ < n2j−1 ,
4k < n−2
(cid:111)
∪ {(n + 1, j) : 0 ≤ j ≤ n}.
Here the Bp,n are defined as in Remark 4.2.2, and, for brevity, we temporarily write w(ξ)
for weight ξ. We define X∞ to be the Bourgain-Delbaen space X(Γ) where Γ is the subset
of Υ just defined.
Once again, to compute norms we need to work with the evaluation functionals e∗
γ. We
note that this is no harder than in the previous constructions.
Proposition 4.2.7. For a γ ∈ Γ, either it is a trivial, age 0 element (γ = (n + 1, j)) in
γ, or age γ ≥ 1 in which case we once again have the
which case we simply have e∗
evaluation analysis given by Proposition 3.3.14.
γ = d∗
105
(cid:40)
S∗e∗
γ =
0
e∗
F (γ)
if F (γ) is undefined
otherwise
Proof. If γ is a trivial, age 0 element, there is nothing to prove; the statement follows from
our earlier definitions. In the case where γ ∈ Γ and age γ ≥ 1, the same proof as before
goes through.
We still need to show that there exists an operator 'S' on X∞ having all the properties
described in the main theorem. We first define the operator. It is constructed in exactly
the same way as in Proposition 3.3.16.
Proposition 4.2.8. Γ is invariant under G. More precisely, if γ ∈ Γ ⊆ Υ and G(γ) is
defined, then G(γ) ∈ Γ. It follows that the map G : Υ → Υ∪{undefined} defined in Theorem
4.2.3 restricts to give F : Γ → Γ ∪ {undefined}. Consequently the map R∗ : (cid:96)1(Υ) → (cid:96)1(Υ)
(also defined in 4.2.3) can be restricted to the subspace (cid:96)1(Γ) ⊆ (cid:96)1(Υ) giving S∗ : (cid:96)1(Γ) →
(cid:96)1(Γ). S∗ is a bounded linear map on (cid:96)1(Γ) of norm 1 which satisfies
for every γ ∈ Γ. Moreover, the dual operator of S∗ restricts to X∞ to give a bounded linear
operator S : X∞ → X∞ of norm at most 1, satisfying Sj (cid:54)= 0 for all j ∈ N.
Proof. We once again omit the details, as the proof is only a very minor modification of
the proof of the original result. Note that Sjd(n+1,0) = d(n+1,j) whenever 0 ≤ j ≤ n so that
Sj (cid:54)= 0 for any j ∈ N.
The reader can easily verify that Lemma 3.3.17 of the previous chapter still holds for the
odd-weight elements of our new construction. Moreover, we can define a C-RIS in the space
X∞ using exactly the same definition as in Chapter 3; we refer the reader to Definition 3.4.1.
It is then an easy exercise to check that Lemmas and Propositions 3.4.3, 3.4.4 and 3.4.5 all
still hold, with only minor modifications required to the original proofs to take into account
the new, weight and age 0 elements. In particular, we note that the basis (dγ)γ∈Γ of the
space X∞ is shrinking and X∗∞ is naturally isomorphic to (cid:96)1. Moreover, having established
that X∗∞ is naturally isomorphic to (cid:96)1, the same proof as in Lemma 3.4.6 works to show
that the dual operator of S : X∞ → X∞ is precisely the operator S∗ : (cid:96)1(Γ) → (cid:96)1(Γ) of
Proposition 4.2.8.
The strategy used to prove our main result will be to follow the arguments of Gowers
and Maurey in [15]. We will thus be interested in the seminorm (cid:107) · (cid:107) on L(X∞), defined
by
(cid:107)T(cid:107) := sup
(xn)n∈N∈L
where L denotes the set of all 1-RIS, (xn)n∈N.
(cid:107)T xn(cid:107)
lim sup
n
We first document some observations about this seminorm.
106
Lemma 4.2.9. Given T ∈ L(X∞), T is compact ⇐⇒ (cid:107)T(cid:107) = 0.
Proof. Suppose first that T is compact. Since the basis (dγ)γ∈Γ is shrinking, it follows by
Proposition 1.4.15 that if (xn)n∈N ∈ L, (cid:107)T xn(cid:107) → 0. It follows immediately that (cid:107)T(cid:107) = 0.
Conversely, suppose (cid:107)T(cid:107) = 0. If (xn)n∈N is a C-RIS, then ( xn
C )n∈N ∈ L so (cid:107)T xn(cid:107) =
C(cid:107) T xn
C (cid:107) → 0. It follows from Proposition 3.4.4 (and our earlier observation that this result
still holds for the space X∞), that whenever (yn)n∈N is a block sequence in X∞, (cid:107)T yn(cid:107) → 0.
This implies that T is compact, as we saw in Proposition 1.4.15.
j=0 λjSj. Then (cid:107)T(cid:107) = (cid:80)N
Lemma 4.2.10. Let T = (cid:80)N
(cid:80)N
j=0 λj.
j=0 λj. Moreover, it is an elementary fact that (cid:107)T(cid:107) = (cid:107)T ∗(cid:107) = (cid:107)(cid:80)N
(cid:80)N
Proof. We already know (cid:107)S(cid:107) ≤ 1;
it follows from the triangle inequality that (cid:107)T(cid:107) ≤
j=0 λj(S∗)j(cid:107). Since
j=0 λj and moreover (cid:107)T(cid:107) ≥
1
6
(N + 1, N ) ∈ Γ, we thus have,
(cid:107)T(cid:107) = (cid:107) N(cid:88)
j=0
λj(S∗)je∗
(N +1,N )(cid:107)(cid:96)1
λje∗
(N +1,N−j)(cid:107)(cid:96)1 =
N(cid:88)
j=0
λj
j=0
λj(S∗)j(cid:107) ≥ (cid:107) N(cid:88)
= (cid:107) N(cid:88)
(cid:16)(cid:80)
j=0
and we obtain the first equality stated in the lemma.
εγ ∈ {1,−1} (for γ ∈ A), (cid:80)
Lemma 2.2.3. Consequently, (cid:107)(cid:80)
To obtain the final inequality stated in the lemma, note that whenever A ⊆ ∆q, and
; this follows immediately from
γ∈A εγdγ(cid:107) ≤ 2. Note that (N + n, N − j) ∈ Γ whenever
γ∈A εγdγ = iq
γ∈A εγeγ
(cid:17)
n ≥ 1 and 0 ≤ j ≤ N . We consider the sequence (xn)∞
n=1 defined by
N(cid:88)
j=0
xn =
1
6
sign λj d(N +n,N−j)
3 ≤ 1 for all n. Moreover, we see
We claim (xn) ∈ L. By our first observation, (cid:107)xn(cid:107) ≤ 1
that the local support (see [2, Definition 5.7]) of any xn contains no element of weight > 0.
Thus by Lemma 5.8 of [2], whenever γ ∈ Γ and weight γ = m−1
for some h, xn(γ) ≤ m−1
h .
It follows that (xn)n∈N is a 1-RIS and is therefore in L. (For the sequence (jk) appearing
in the definition of a RIS, we take j1 = 1 and jk+1 = max ran xk + 1 for k ≥ 1.)
h
107
Finally, we note that
(cid:107)T xn(cid:107) ≥ (cid:104)e∗
(N +n,N ),
N(cid:88)
j=0
λjSjxn(cid:105) = (cid:104)d∗
(N +n,N ),
N(cid:88)
j=0
λjSjxn(cid:105) = (cid:104) N(cid:88)
j=0
λj(S∗)jd∗
(N +n,N ), xn(cid:105)
λjsign λk (cid:104)d∗
(N +n,N−j), d(N +n,N−k)(cid:105)
N(cid:88)
k=0
λj
N(cid:88)
N(cid:88)
j=0
j=0
=
=
1
6
1
6
from which it follows that (cid:107)T(cid:107) ≥ lim supn (cid:107)T xn(cid:107) ≥ 1
6
We aim to prove the following theorem:
(cid:80)N
j=0 λj as required.
Theorem 4.2.11. Every operator in L(X∞) is in the (cid:107) · (cid:107)-closure of the set of operators
in the algebra A, generated by the Identity and S.
As the proof is long and technical, we first see that our main result, Theorem 4.1.1,
almost completely follows from Theorem 4.2.11 and the above lemmas.
Proof of Theorem 4.1.1. We have already observed that X∗∞ is isomorphic to (cid:96)1; indeed, the
proof is almost identical to the proof for the spaces Xk.
We note that the first part of Lemma 4.2.10 is equivalent to saying that (Sj)∞
j=0 is a basic
sequence in L(X∞), 1-equivalent to the canonical basis of (cid:96)1 = (cid:96)1(N0). Moreover, it follows
from the second part of Lemma 4.2.10 and Lemma 4.2.9 that K(X∞) ⊕ [Sj : j ∈ N0] is an
algebraic direct sum in L(X∞). We first show that this direct sum is the entire operator
algebra. The proof of this is identical to the arguments used in [15] but for completeness
we include it here.
We write G for the Banach space obtained by taking the (cid:107) · (cid:107)-completion of A, noting
that by the second part of Lemma 4.2.10, (cid:107) · (cid:107) actually defines a norm on A. (Here A
is defined as in Theorem 4.2.11.) Given T ∈ L(X∞), by Theorem 4.2.11 we can find a
(cid:107)·(cid:107)-Cauchy sequence (Tn)∞
n=1 of operators in A such that (cid:107)T − Tn(cid:107) → 0. Let φ(T ) be the
n=1 in G. It is easily checked that, in this way, we get a well-defined, bounded
limit of (Tn)∞
linear map φ : L(X∞) → G (of norm at most 1). Note that the restriction of φ to A is
the identity (or more accurately, the embdedding of A into G). It is also easily seen that
the kernel of φ is the set of T such that (cid:107)T(cid:107) = 0, which by Lemma 4.2.9 is precisely the
compact operators, K(X∞).
Now, we note that we can define a linear isomorphic embedding I : G → L(X∞), with
I(G) = [Sj : j ∈ N0]. Precisely, we define I by defining it on the dense subspace of vectors
j=0 λjSj] is the vector
j=0 λjSj], where it is to be understood that [(cid:80)N
in G of the form [(cid:80)N
108
j=0 λjSj] (cid:55)→(cid:80)N
in G obtained under the natural embedding A (cid:44)→ G of the vector(cid:80)N
i.e. [(cid:80)N
j=0 λjSj (recalling that
G is simply the (cid:107) · (cid:107)-completion of A). We can therefore well-define I in the obvious way,
j=0 λjSj. By the second part of Lemma 4.2.10, this is a continuous
linear map, so we may take the unique continuous linear extension to G. Since (cid:107)·(cid:107) ≤ (cid:107)·(cid:107),
it is easily seen that this map is an isomorphic embedding onto the closed linear span
[Sj : j ∈ N0], as required.
We observe that I ◦ φ(T ) − T is in the kernel of φ and is therefore compact. Indeed, it
is clear from the definitions of φ and I that φ ◦ I is the identity on the dense subspace of
vectors in G of the form [(cid:80)N
j=0 λjSj]. Using this observation and continuity, we see that
φ (I ◦ φ(T ) − T ) = φ ◦ I (φ(T )) − φ(T )
= φ(T ) − φ(T ) = 0
as required. Consequently, if T ∈ L(X∞) there exists a compact operator K such that
T = I ◦ φ(T ) + K and we have
L(X∞) = [Sj : j ∈ N0] ⊕ K(X∞)
as claimed.
Property (3) of Theorem 4.1.1 now follows immediately from the above and the fact
j=0 is a basic sequence in L(X∞). Moreover, if we let e : (cid:96)1(N0) → L(X∞) denote
that [Sj]∞
j=0 λjSj and let π : L(X∞) → L(X∞)/K(X∞) denote
the isometric embedding (λj)∞
the canonical projection, then the map I : (cid:96)1(N0) → L(X∞)/K(X∞) defined by I = π ◦ e is
now easily seen to be a continuous bijection; it is therefore an isomorphism by the Inverse
j=0 (cid:55)→(cid:80)∞
Mapping Theorem. It is readily checked that this map is a Banach algebra isomorphism.
j=0 ⊂ L(X∞)/K(X∞) is a basic sequence also follows from
The fact that the sequence ([Sj])∞
the fact that I is an isomorphism.
We have very nearly completed the proof. However, it is claimed in property (4) that
L(X∞)/K(X∞) is actually isometric to (cid:96)1(N0). We will show in the next section that the
sequence [Sj]∞
j=0 is really 1-equivalent to the canonical basis of (cid:96)1, from which it follows
that I is really an isometry.
It remains for us to prove Theorem 4.2.11, i.e. that every operator in L(X∞) is in the
(cid:107) · (cid:107)−closure of the set of operators in the algebra A generated by the identity operator
In order to do so, we will need to make further modifications to the definition
and S.
of a (C, j, 0, ε) weak exact pair (see Definition 3.4.10) and also to the definition of weak
dependent sequences (see Definition 3.4.12). Since these definitions are quite complicated,
we will include all the details in our new definitions, although most of the conditions we
109
demand are as before. Moreover, unlike in Chapter 3, we won't need 1-exact pairs and
dependent sequences in what follows. Consequently we make a very small change in notation
to the previous chapter.
Definition 4.2.12. Let C, ε > 0 and ω ∈ N ∪ {∞}. A pair (x, η) ∈ X∞ × Γ is said to be a
(C, j, ε, ω)-weak exact pair of order ω if
1. (cid:107)x(cid:107) ≤ C
ξ, x(cid:105) ≤ Cm−1
2. (cid:104)d∗
3. weight η = m−1
4. Slx(η) ≤ Cε for every 0 ≤ l ≤ ω if ω < ∞ and Slx(η) ≤ Cε for every 0 ≤ l < ω if
for all ξ ∈ Γ
j
j
ω = ∞
5. for every element η(cid:48) of Γ with weight η(cid:48) = m−1
Cm−1
Cm−1
x(η(cid:48)) ≤
(cid:40)
i
i
j
(cid:54)= m−1
j
, we have
if i < j
if i > j.
As before, we observe that we have the following estimates for weak exact pairs (for
exactly the same reasons as in [2] and as in the previous chapter).
Remark. A (C, j, ε, ω) weak exact pair also satisfies the estimates
(cid:40)
(cid:104)e∗
η(cid:48), P(s,∞)x(cid:105) ≤
6Cm−1
6Cm−1
i
j
if i < j
if i > j
for elements η(cid:48) of Γ with weight η(cid:48) = m−1
i
(cid:54)= m−1
j
.
We will need the following method for constructing 0 weak exact pairs of order N .
Lemma 4.2.13. Let ω ∈ N ∪ {∞}, (xk)n2j
k=1 be a skipped-block C-RIS, and let q0 < q1 <
q2 < ··· < qn2j be natural numbers such that q1 ≥ 2j and ran xk ⊆ (qk−1, qk) for all k.
Let z denote the weighted sum z = m2jn−1
k be an element of
Bqk−1,qk−1 with (cid:104)(S∗)lb∗
k, Slxk(cid:105) ≤ Cε for all 0 ≤ l ≤ ω (or 0 ≤ l < ∞ if
ω = ∞).Then there exist ζi ∈ ∆qi (1 ≤ i ≤ n2j) such that the element η = ζn2j has analysis
(qi, b∗
(cid:80)n2j
k=1 xk. For each k let b∗
i , ζi)1≤i≤n2j and (z, η) is a (16C, 2j, ε, ω)-special exact pair.
k, xk(cid:105) = (cid:104)b∗
2j
The proof of the lemma is sufficiently close to the proof of Lemma 3.4.11 that we omit
it.
110
Definition 4.2.14. We shall say that a sequence (xi)i≤n2j0−1 in X∞ is a (C, 2j0− 1, 0)-weak
dependent sequence (of finite order) if there exist sequences of natural numbers 0 = p0 <
p1 < p2 < ··· < pn2j0−1 and 1 = N0 < N1 < N2 < ··· < Nn2j0−1 < ∞ together with
ηi ∈ Γpi−1 \ Γpi−1 and ξi ∈ ∆pi (1 ≤ i ≤ n2j0−1) such that
1. for each k, ran xk ⊆ (pk−1, pk)
2. for each k, Sjxk(ηk) = 0 whenever j ≥ Nk
has weight m−1
2j0−1 and analysis (pi, e∗
ηi, ξi)
n2j0−1
i=1
3. the element ξ = ξn2j0−1 of ∆pn2j0−1
4. (x1, η1) is a (C, 4j1, n−1
5. for each 2 ≤ i ≤ n2j0−1, (xi, ηi) is a (C, 4σ(ξi−1), n−1
2j0−1, N0)-weak exact pair
2j0−1, Ni−1)-weak exact pair.
In certain applications, we can remove the more complicated conditions involving the se-
quence (Nj)
dependent sequence (of infinite order) if we remove condition (2) above, and ask that all
Specifically, we will say the sequence (xi)i≤n2j0−1 is a (C, 2j0 − 1, 0)-weak
n2j0−1
j=0
the weak exact pairs occurring in (4) and (5) have infinite order.
In either case, we notice that, because of the special odd-weight conditions, we necessar-
2j0−1, by Lemma 3.3.17
2j0−1, and weight ηi+1 = m−1
= weight η1 < n−2
4σ(ξi) < n−2
ily have m−1
for 1 ≤ i < n2j0−1.
4j1
The following two lemmas are fundamental to proving our main result.
Lemma 4.2.15. Let (xi)i≤n2j0−1 be a weak (C, 2j0 − 1, 0)-dependent sequence (of finite or
infinite order) in X∞ and let J be a sub-interval of [1, n2j0−1]. For any γ(cid:48) ∈ Γ of weight
m2j0−1 we have
(cid:88)
i∈J
xi(γ(cid:48)) ≤ 7C.
Proof. The proof is almost the same as the proof of Lemma 3.4.13. We recall the notation
used; we let ξi, ηi, pi, Ni (if we are considering a dependent sequence of finite order), j1 be
as in the definition of a dependent sequence and let γ denote ξn2j0−1, an element of weight
is either 0 or
i)1≤i≤a(cid:48)(cid:1) the analysis of γ(cid:48) where each b(cid:48)∗
m2j0−1. We denote by(cid:0)p(cid:48)
0, (p(cid:48)
i, b(cid:48)∗
i , ξ(cid:48)
i
e∗
η(cid:48)
i
for some suitable η(cid:48)
i.
We recall that we can estimate as follows
(cid:88)
k∈J
xk(γ(cid:48)) ≤ (cid:88)
k∈J, k<l
xk(γ(cid:48)) + xl(γ(cid:48)) +
(cid:88)
k∈J, k>l
xk(γ(cid:48)),
where we are assuming we have chosen l maximal such that there exists i with b(cid:48)∗
and weight η(cid:48)
i = e∗
η(cid:48)
i = weight ηl. (We refer the reader back to the proof of Lemma 3.4.13 for the
i
111
details on reducing to this case.) The terms xl(γ(cid:48)) and(cid:80)
The argument for estimating the remaining term, (cid:80)
in exactly the same way as in the proof of Lemma 3.4.13.
k∈J, k<l xk(γ(cid:48)), is a little more
involved. Obviously if l = 1, the sum we are trying to estimate is just zero, and the lemma
is proved. So we can suppose l > 1. By definition of l, there exists some i such that b(cid:48)∗
i = e∗
η(cid:48)
and weight ηl = weight η(cid:48)
i. Now either i = 1 or i > 1. When i = 1, we can use exactly the
same argument as used for the corresponding case in the proof of Lemma 3.4.13 to conclude
i
k∈J, k>l xk(γ(cid:48)) can be estimated
that (cid:80)
k∈J, k<l xk(γ(cid:48)) ≤ 3C and we are done.
i = weight ηl. But by definition of an exact pair, we have weight ηl = m−1
It only remains to consider what happens when i > 1. Recall we are also assuming l > 1
4σ(ξl−1),
i−1). By
i−1). By Lemma
i−1. We note that in particular
i−1 since F preserves rank and we can write the evaluation analysis of
and weight η(cid:48)
and by restrictions on elements of odd weights, weight η(cid:48)
strict monotonicity of the sequence mj, we deduce that ω = σ(ξl−1) ∈ Σ(ξ(cid:48)
4.2.5, there is some j, 0 ≤ j < ∞, such that F j(ξl−1) = ξ(cid:48)
this implies pl−1 = p(cid:48)
γ(cid:48) as
4ω with ω ∈ Σ(ξ(cid:48)
i = m−1
2j0−1P ∗
(p(cid:48)
r−1,p(cid:48)
Now, for k < l, since ran xk ⊆ (pk−1, pk) ⊆ (0, pl−1) = (0, p(cid:48)
γ(cid:48) = (S∗)j(e∗
e∗
+ m−1
d∗
ξ(cid:48)
) +
ξl−1
r=i
r
r)b(cid:48)∗
r .
i−1), we see that
a(cid:48)(cid:88)
(cid:104)xk, e∗
γ(cid:48)(cid:105) = (cid:104)xk, (S∗)je∗
l−1(cid:88)
ξl−1
ξr + m−1
d∗
(cid:105)
= (cid:104)Sjxk,
= m−1
r=1
2j0−1(cid:104)Sjxk, e∗
ηk
2j0−1P ∗
(cid:105) ≤ (cid:104)Sjxk, e∗
ηr(cid:105)
(pr−1,pr)e∗
(cid:105).
ηk
Now, if (xi)i≤n2j0−1 is a dependent sequence of infinite order, we have Slxk(ηk) ≤ Cn−1
for every l ∈ N0, so certainly (cid:104)Sjxk, e∗
(cid:105) ≤ Cn−1
2j0−1 and consequently,
ηk
2j0−1
(cid:88)
k∈J, k<l
xk(γ(cid:48)) ≤ n2j0−1 max
k∈J, k<l
xk(γ(cid:48)) ≤ C.
Otherwise (xi)i≤n2j0−1 is a dependent sequence of finite order, and we estimate by consider-
ing three possibilities. If j < N1 ≤ Nk−1 for all k ≥ 2, then by definition of weak dependent
sequence, (cid:104)Sjxk, η∗
2j0−1 for k ≥ 2. So,
k(cid:105) ≤ Cn−1
xk(γ(cid:48)) ≤ x1(γ(cid:48)) + n2j0−1 max
xk(γ(cid:48)) ≤ 2C.
k∈J, 2≤k<l
(cid:88)
(cid:105) = 0 and so (cid:80)
k∈J, k<l
112
If j > Nn2j0−1 ≥ Nk for all k, then condition (2) of weak dependent sequence implies that
(cid:104)Sjxk, e∗
k∈J, k<l xk(γ(cid:48)) = 0.
ηk
The only remaining possibility is that there is some r ∈ {1, 2, . . . , n2j0−1 − 1} with
Nr ≤ j ≤ Nr+1. Now if k ≤ r, it follows that j ≥ Nr ≥ Nk and so Sjxk(ηk) = 0
If k ≥ r + 2, j ≤ Nr+1 ≤ Nk−1 and so
by condition (2) of weak dependent sequence.
(cid:104)Sjxk, e∗
(cid:105) ≤ Cn−1
2j0−1. It follows that
xk(γ(cid:48)) ≤ xr+1(γ(cid:48)) + (cid:88)
xk(γ(cid:48)) + (cid:88)
xk(γ(cid:48))
ηk
(cid:88)
k∈J, k<l
k∈J, k<l, k≤r
≤ (cid:107)xr+1(cid:107) + 0 + n2j0−1
max
k∈J, k<l, k≥r+2
k∈J, k<l, k≥r+2
(cid:104)Sjxk, e∗
(cid:105) ≤ 2C.
ηk
This completes the proof.
Lemma 4.2.16. If (xi)1≤i≤n2j0−1 is a (C, 2j0 − 1, 0) weak dependent sequence (of finite or
infinite order), then
(cid:107)n−1
2j0−1
xk(cid:107) ≤ 70Cm−2
2j0−1.
n2j0−1(cid:88)
k=1
and moreover, if γ ∈ Γ, weight (γ) = m−1
28Cn−1
28Cn−1
n2j0−1(cid:88)
xk(γ) ≤
n−1
2j0−1
(cid:40)
h , then
2j0−1 + 84Cm−2
2j0−1 + 42Cm−1
h
k=1
2j0−1m−1
h
if h < 2j0 − 1
if h > 2j0 − 1.
Proof. The proof follows from Lemma 4.2.15 (readjusting the RIS constant to 7C), and by
the proof of [2, Proposition 5.6].
In the spirit of [15], we aim to prove the following proposition.
Proposition 4.2.17. Let T : X∞ → X∞ be a bounded linear operator. Then for every
δ > 0, there exists an l ∈ N such that whenever x ∈ X∞ is a block vector satisfying
(cid:107)x(cid:107) ≤ 1, min ran x > l and x(γ) ≤ m−1
i whenever weight γ = m−1
i with i < l,
dist(T x, l conv{λSjx : j ∈ {0, 1, . . . , l},λ = 1}) ≤ δ.
In order to prove the above proposition, we will need the following lemmas, which are
analogous to Lemma 7.1 of [2] and Lemma 3.4.17.
Lemma 4.2.18. Let m, n and l be natural numbers with m < n and let x, y ∈ X∞ be such
that ran x, ran y are both contained in the interval (m, n]. Suppose that dist(y, l conv{λSjx :
j ∈ {0, 1, . . . , l},λ = 1}) > δ. Then there exists b∗ ∈ ball (cid:96)1(Γn \ Γm), with rational
coordinates, such that (cid:104)Sjx, b∗(cid:105) ≤ (cid:107)y(cid:107)
for every j ∈ {0, 1, . . . , l} and b∗(y) > 1
2 δ.
l
113
Proof. For 0 ≤ j ≤ l, let uj ∈ (cid:96)∞(Γn\Γm) be the restriction of Sjx, and let v ∈ (cid:96)∞(Γn\Γm)
be the restriction of y. Then Sjx = inuj for 0 ≤ j ≤ l, y = inv and for any scalars (µj)l
j=0,
j=0 µj ≤ l, δ < (cid:107)y −(cid:80)l
j=0 µjSjx(cid:107) ≤ (cid:107)in(cid:107)(cid:107)v −(cid:80)l
j=0 µjuj(cid:107). Hence
with(cid:80)l
dist(v, l conv{λuj : j ∈ {0, 1, . . . , l},λ = 1}) >
1
2
δ.
Let Cl = conv{λuj : j ∈ {0, 1, . . . , l},λ = 1}. By the Hahn -- Banach Theorem (see Theorem
1.4.7) in the finite dimensional space (cid:96)∞(Γn \ Γm), there exists a∗ ∈ sphere (cid:96)1(Γn \ Γm) with
(cid:26)
(cid:27)
sup
(cid:104)a∗, z(cid:105) : z ∈ Cl +
1
2
δB((cid:96)∞(Γn \ Γm)
< (cid:104)a∗, v(cid:105).
It follows that (cid:104)a∗, v(cid:105) > 1
2 δ < (cid:104)a∗, v(cid:105) ≤ (cid:107)v(cid:107) ≤ (cid:107)y(cid:107).
Consequently, (cid:104)a∗, uj(cid:105) <
for every 0 ≤ j ≤ l. To complete the proof, we note that
by an elementary continuity argument, we may approximate a∗ arbitrarily well by b∗ ∈
ball (cid:96)1(Γn \ Γm) ∩ QΓn\Γm retaining the desired conditions.
2 δ and supa∗(Cl) < supa∗(Cl) + 1
(cid:107)y(cid:107)
l
The proof of Proposition 4.2.17 will be easy once we have proved the following lemma.
Lemma 4.2.19. Let T be a bounded linear operator on X∞. Suppose for this T , there exists
some δ > 0 for which the conclusion of Proposition 4.2.17 fails. Then, given any ε > 0 and
natural numbers, j, p, N , there exists a block vector z ∈ X∞, a natural number q > p, and
an η ∈ ∆q such that ran z ⊆ (p, q) and
(1) (z, η) is a (16, 2j, ε, N )-exact pair;
16 δ;
(2) (T z)(η) > 7
(3) (cid:107)(I − P(p,q))T z(cid:107) < m−1
2j δ;
(4) (cid:104)P ∗
η, T z(cid:105) > 3
8 δ.
(p,q]e∗
Proof. Note that under the hypothesis of the lemma, we can obtain a skipped block 1-RIS
(xk)k∈N with the property that for every k
dist(T xk, k conv{λSjxk : j ∈ {0, 1, . . . , k},λ = 1} > δ.
We now fix L ∈ N such that L ≥ N and L > 4(cid:107)T(cid:107)/ε. Observe that if m < n and x ∈ X∞,
then m conv{λSjx : j ∈ {0, 1, . . . , m},λ = 1} ⊆ n conv{λSjx : j ∈ {0, 1, . . . , n},λ = 1}
so, for k ≥ L,
dist(T xk, L conv{λSjxk : j ∈ {0, 1, . . . , L},λ = 1})
≥ dist(T xk, k conv{λSjxk : j ∈ {0, 1, . . . , k},λ = 1} > δ.
114
We pass to the subsequence (xk)k≥L so that the above inequality holds. Note also that the
sequence (T xk) is weakly null (since the basis (dγ)γ∈Γ, is shrinking). Consequently, passing
to a further subsequence if necessary, we can assume that there exist p < q0 < q1 < q2 < . . .
such that, for all k ≥ 1, ran xk ⊆ (qk−1, qk) and (cid:107)(I − P(qk−1,qk))T xk(cid:107) < 1
2j δ ≤
1280 δ. It follows from this that dist(P(qk−1,qk)T xk, L conv{λSjxk : j ∈ {0, 1, . . . , L},λ =
1}) > 1279
k ∈ ball (cid:96)1(Γqk−1 \ Γqk−1), with
rational coordinates, satisfying
1280 δ. We may apply Lemma 4.2.18 to obtain b∗
2j δ ≤ 1
80 m−1
5 m−2
1
k, Sjxk(cid:105) ≤ (cid:107)P(qk−1,qk)T xk(cid:107)
(cid:104)b∗
L
≤ 4(cid:107)T(cid:107)
L
< ε
whenever 0 ≤ j ≤ L, and
(cid:104)b∗
k, P(qk−1,qk)T xk(cid:105) > 1279
2560 δ.
k have denominators dividing Nqk−1!, so that b∗
Taking a further subsequence if necessary (and redefining the qk), we may assume that
k ∈ Bqk−1,qk−1, and we
the coordinates of b∗
may also assume that q1 ≥ 2j. We are thus in a position to apply Lemma 4.2.13, getting
elements ξk of weight m−1
has evaluation
analysis
in ∆qk such that the element η = ξn2j of ∆qn2j
2j
n2j(cid:88)
e∗
η =
d∗
ξi
+ m−1
2j
P ∗
(qi−1,qi)b∗
i .
i=1
i=1
n2j(cid:88)
n2j(cid:88)
z = m2jn−1
2j
xi.
and such that (z, η) is a (16, 2j, ε, L)-weak exact pair, where z denotes the weighted average
Since we also chose L ≥ N , (z, η) is certainly a (16, 2j, ε, N )-weak exact pair as required.
We set q = qn2j ; the rest of the proof is now exactly the same as in [2, Lemma 7.2].
i=1
Proof of Proposition 4.2.17. We suppose for contradiction that there exists δ > 0, such that
the conclusion of the Proposition 4.2.17 fails.
The argument is the same as Proposition 7.3 of [2]. We construct a (16, 2j0 − 1, 0)-
weak dependent sequence of finite order (for a suitably chosen j0 ∈ N) by making repeated
applications of Lemma 4.2.19.
We begin by choosing j0 such that m2j0−1 > 4480(cid:107)T(cid:107)δ−1 and j1 such that m4j1 > n2
Taking p = p0 = 0 and j = 2j1, N = N0 = 1, ε = n−1
q1 and a (16, 4j1, n−1
(cid:107)(I − P(0,q1])(T z1)(cid:107) < m−1
of ∆p1 given by ξ1 = (p1, m−1
Slz1(η1) = 0 whenever l ≥ N1.
2j0−1.
2j0−1 in Lemma 4.2.19, we can find
8 δ and
4j1−2δ. Let p1 = q1 + 1 and let ξ1 be the special Type 1 element
η1). By Lemma 4.2.4, we can find an N1 > N0 such that
2j0−1, N0)-exact pair (z1, η1) with rank η1 = q1, (T z1)(η1) > 3
2j0−1, e∗
115
Now, recursively for 2 ≤ i ≤ n2j0−1, define ji = σ(ξi−1), and use Lemma 4.2.19 again to
choose qi and a (16, 4ji, n−1
2j0−1, Ni−1)-exact pair (zi, ηi) with rank ηi = qi, ran zi ⊆ (pi−1, qi),
(cid:104)P ∗
δ. We now define pi = qi + 1, choose
Ni such that Slxi(ηi) = 0 whenever l ≥ Ni (which is again possible by 4.2.4), and let ξi be
the Type 2 element (pi, ξi−1, m−1
8 δ and (cid:107)(I − P(pi,qi])(T zi)(cid:107) < m−1
ηi, T zi(cid:105) > 3
(pi−1,qi]e∗
4ji
ηi) of ∆pi.
2j0−1, e∗
It is clear that we have constructed a (16, 2j0−1, 0)-weak dependent sequence (zi)1≤i≤n2j0−1.
We set z = n−1
other hand, the same argument as in the proof of Proposition 7.3 of [2] yields that
zi and note that (cid:107)z(cid:107) ≤ 70 · 16m−2
2j0−1 by Lemma 4.2.16. On the
2j0−1
(cid:80)n2j0−1
i=1
(cid:107)T z(cid:107) >
m−1
2j0−1δ
4
≥ δm2j0−1(cid:107)z(cid:107)
4 · 70 · 16
which implies that (cid:107)T(cid:107) >
δm2j0−1
4480
, contradicting m2j0−1 > 4480(cid:107)T(cid:107)δ−1.
Proposition 4.2.20. Let T be a bounded linear operator on X∞, δ > 0. For this δ, choose
l ∈ N as given by Proposition 4.2.17 and let Al = l conv{λSj : j ∈ {0, 1, . . . , l},λ = 1}.
Then, there exists U ∈ Al with (cid:107)T − U(cid:107) ≤ 28, 674δ.
For the proof, we will need the following fixed-point theorem due to Kakutani. We refer
the reader to [19] for the proof and more details.
Theorem 4.2.21 (Kakutani Fixed Point Theorem). Let S be an r-dimensional closed
simplex and denote by R(S) the set of all non-empty, closed, convex subsets of S.
If
Φ : S → R(S) is upper semi-continuous, then there exists an x0 ∈ S such that x0 ∈ Φ(x0).
We remind the reader of the following definition from [19]:
Definition 4.2.22. With the notation as in the preceding theorem, a point-to-set mapping
Φ : S → R(S) is called upper semi-continuous if whenever (xn)∞
n=1 is a sequence in S
converging to x0 ∈ S, yn ∈ Φ(xn) and yn → y0 ∈ S, then y0 ∈ Φ(x0).
Proof of 4.2.20. The proof follows closely that of [15, Lemma 9]. We find it convenient to
introduce the following piece of notation. Given a sequence of block vectors (xi)i∈N, we
write x1 < x2 < . . . if there exist natural numbers q0 < q1 < . . . with ran xi ⊆ (qi−1, qi) for
every i (with similar notation if the sequence has only finite length).
It is sufficient to show that if T is an operator on X∞, with matrix representation with
respect to the basis (dγ)γ∈Γ possessing the property that there are only finitely many non-
zero entries in each row and column, then there is U ∈ Al with (cid:107)T − U(cid:107) ≤ 28, 673δ. (We
temporarily call this property P so that we can refer back to this statement.) Indeed, since
the basis (dγ)γ∈Γ is shrinking, T can be perturbed by a distance at most δ in operator norm
116
to an operator of this form, and the seminorm, (cid:107) · (cid:107), is bounded above by the operator
norm.
Having performed the described perturbation, it follows that we may assume T sends
block vectors (with respect to the FDD) to block vectors. Note we can also assume (without
loss of generality) that if N ∈ N is given, whenever x ∈ X∞ is a block vector with min ran x
sufficiently large, min ran T x > N . Indeed, if this were not the case, an elementary argument
involving the fact that the matrix of T is assumed to have only finitely many non-zero entries
in each row and column would imply that T is in fact a finite-rank, hence compact, operator.
We could thus take U = 0 and the proof would be complete by Lemma 4.2.9.
Now, if property P is false, then for every U ∈ Al, there exists a 1-RIS (xU (i))i∈N with
lim supi (cid:107)(T − U )xU (i)(cid:107) > 28, 673δ. Since subsequences of 1-RIS are again 1-RIS, passing
to a subsequence if necessary, we see that for every U ∈ Al, there exists a skipped block
1-RIS (xU (i))i∈N such that (cid:107)(T − U )xU (i)(cid:107) > 28, 673δ for all i.
Note that Al is compact.
L(X∞), defined by (λ0, λ2, . . . , λl) (cid:55)→(cid:80)l
Indeed, consider the (continuous) linear map h : (cid:96)l+1
j=0 λjSj. Then the restriction of h to l B((cid:96)l+1
1 →
) is
j=1 of Al by open sets of
j=1 be a partition of unity on Al with φj supported inside Uj
1
a homeomorphism onto Al. We can thus choose a covering (Uj)k
diameter less than δ. Let (φj)k
for each j.
For every j = 1, . . . , k, let Uj ∈ Uj and let (xj,i)i∈N be a skipped block 1-RIS with
the above property with U = Uj. By the condition on the diameter of Uj, we have (cid:107)(T −
U )xj,i(cid:107) > 28, 672δ for all i ∈ N and U ∈ Uj. Moreover, by the remarks made earlier (and
the fact that ran x ⊂ (p, q) =⇒ ran Sx ⊆ (p, q)), passing to a subsequence of (xj,i)i∈N if
necessary, we may also assume that ((T − U )xj,i)i∈N is a skipped block sequence of successive
vectors for every U ∈ Al.
For the rest of the proof, we work with j0 ∈ N chosen large enough that
(1) 2j0 − 1 > l
(2) m2j0n−1
(3) kn2j0−1 ≤ n2j0.
2j0−1 ≤ m−1
l
Note that it is certainly possible to choose such a j0. Indeed, by Assumptions (2) and (4) on
the sequences (mj) and (nj) (see Assumption 3.3.1), m2jn−1
2j−2 → 0
as j → ∞ and nj+1n−1
j+1 → ∞ as j → ∞.
Now, suppose we are given ε > 0, p, r ∈ N. We can select a skipped block 1-RIS of length
n2r, (xj,i)M +n2r
i=M +1 (M suitably chosen) from the 1-RIS (xj,i)i∈N, such that there are natural
numbers p < q0 < q1 < ··· < qn2r with q1 ≥ 2r and ran xj,M +i ⊆ (qi−1, qi). Now, setting
j ≥ m2
2j−1n−1
2j−1 ≤ n−1
2j−1 = m2
117
i = 0 ∈ Bqi−1,qi for all i, we can apply Lemma 4.2.13 to see that there exists q = qn2r and
b∗
a (16, 2r, ε,∞)-weak exact pair (z, η) where z is the weighted sum
z = m2rn−1
2r
xj,M +i
n2r(cid:88)
i=1
and η ∈ ∆q. Note also that ran z ⊆ (p, q) and whenever U ∈ Uj,
(cid:107)(T − U )z(cid:107) = m2rn−1
(T − U )xj,M +i(cid:107)
2r (cid:107) n2r(cid:88)
i=1
≥ 1
2
m−1
2r m2rn−1
2r
n2r(cid:88)
i=1
(cid:107)(T − U )xj,M +i(cid:107) > 14, 336δ
where the penultimate inequality follows from Lemma 3.6.2 and the assumption we made
earlier about ((T − U )xj,i)i∈N being a skipped block sequence of successive vectors for every
U ∈ Al.
Using this observation, for each j = 1, . . . , k, we may inductively construct a (16, 2j0 −
, with min ran zj,1 > l and (cid:107)(T −
1, 0) weak dependent sequence of infinite order, (zj,i)
U )zj,i(cid:107) > 14, 336δ whenever U ∈ Uj, 1 ≤ i ≤ n2j0−1. Moreover, once again making use of
the assumptions on T discussed at the beginning of the proof, we may arrange that
n2j0−1
i=1
(T − U )zj,1 < (T − U )zj,2,··· < (T − U )zj,n2j0−1
for every j and U ∈ Al and moreover that
(T − U )zj,n2j0−1 < (T − U )zj+1,1
1792 m2j0n−1
for every 1 ≤ j < k, and U ∈ Al. Now, for j = 1, . . . , k, we set yj = 1
Observe that min ran yj > l. Moreover, by Lemma 4.2.16, we have (cid:107)yj(cid:107) ≤ 1
16m−2
2j0−1 = 1120
1792 and if γ ∈ Γ, weight γ = m−1
m2j0n−1
2j0−1 +
m−1
1344
1792
yj(γ) ≤ 28 × 16
m−1
1792
≤ 448
1792
l +
h
h with h < 2j0 − 1,
m2j0m−2
84 × 16
1792
2j0−1m−1
h
(cid:80)n2j0−1
2j0−1
1792 m2j0 × 70 ×
i=1
zj,i.
For U ∈ Al, let y(U ) =(cid:80)k
by the choice of j0. Since j0 was chosen so that l < 2j0 − 1, if h < l < 2j0 − 1, m−1
and we have yj(γ) ≤ m−1
h .
j=1 φj(U )yj. We have min ran y(U ) > l,(cid:107)y(U )(cid:107) ≤ 1120
1792 ≤ 1
and y(U )(γ) ≤ m−1
h with h < l (since we have just
observed these facts are true when y(U ) is replaced by any yj). We show that y(U ) is also
a 'bad vector' for U , by showing that (cid:107)(T − U )y(U )(cid:107) > 4δ.
h whenever γ ∈ Γ and weight γ = m−1
l ≤ m−1
h
118
14, 336δ whenever j ∈ J and 1 ≤ i ≤ n2j0−1. Observe that (T − U )y(U ) =(cid:80)
To see this, let U ∈ Al be fixed, and let J = {j : φj(U ) > 0}, noting that (cid:107)(T −U )zj,i(cid:107) >
j∈J φj(U )(T −
φj(U )(T −U )zj,i. By the way in which we constructed
U )yj = 1
the zj,i, the final sum is a sum of at most kn2j0−1 ≤ n2j0 skipped, successive vectors. It
follows that
(cid:80)n2j0−1
(cid:80)
j∈J
i=1
2j0−1
1792 m2j0n−1
n2j0−1(cid:88)
(cid:107)(cid:88)
j∈J
i=1
φj(U )(T − U )zj,i(cid:107) ≥ 1
2
m−1
2j0
(cid:107)φj(U )(T − U )zj,i(cid:107)
(cid:88)
j∈J
φj(U )
(cid:88)
n2j0−1(cid:88)
j∈J
i=1
> m−1
2j0
= m−1
2j0
n2j0−17168δ
n2j0−17168δ
and consequently, (cid:107)(T − U )y(U )(cid:107) > 4δ as claimed.
The proof of the proposition now follows by the Kakutani Fixed Point Theorem, The-
Indeed, for each U ∈ Al, let Φ(U ) be the set of V ∈ Al such that (cid:107)(T −
orem 4.2.21.
V )y(U )(cid:107) ≤ 4δ. Clearly Φ(U ) is a closed, convex subset of Al, Moreover, by Proposition
4.2.17, and our earlier observations about y(U ), Φ(U ) is non-empty for every U ∈ Al.
Clearly, the function Al → X∞ defined by U (cid:55)→ y(U ) is continuous, and this ensures that
Φ is upper semi-continuous. Indeed, suppose Un ∈ Al, Un → U , Vn ∈ Φ(Un) and Vn → V .
Then
(cid:107)(T − V )y(U )(cid:107) ≤ (cid:107)(T − Vn)(cid:0)y(U ) − y(Un)(cid:1)(cid:107) + (cid:107)(T − Vn)y(Un)(cid:107) + (cid:107)(Vn − V )y(U )(cid:107).
The first term on the right-hand-side of the inequality converges to 0 by continuity of
U (cid:55)→ y(U ) and the fact that Un → U . The third term converges to 0 since Vn → V . The
second term is at most 4δ since Vn ∈ Φ(Un). It follows that (cid:107)(T − V )y(U )(cid:107) ≤ 4δ so that
V ∈ Φ(U ) as required. Consequently, by the fixed point theorem, there exists some U ∈ Al
with U ∈ Φ(U ); this contradicts the fact that y(U ) is a 'bad' vector for U , as was shown
earlier.
We conclude this section by making the obvious observation that the proof of Theorem
4.2.11 is now an immediate corollary of Proposition 4.2.20.
4.3 L(X∞)/K(X∞) is isometric to (cid:96)1(N0)
We have already seen that the Calkin algebra, L(X∞)/K(X∞), is isomorphic to (cid:96)1(N0). In
the final section of this chapter, we complete the proof of the main theorem and show that
the algebras are in fact isometric. As observed in the proof of Theorem 4.1.1, it will be
enough for us to show that the sequence ([Sj])∞
(cid:96)1.
j=0 is 1-equivalent to the canonical basis of
119
Lemma 4.3.1. For any x ∈ X∞, (cid:107)Snx(cid:107) → 0 as n → ∞.
Proof. Observe that if x is a block vector with ran x ⊆ (p, q) then Snx = 0 whenever n ≥ q.
Indeed, since the vectors (d∗
γ)γ∈Γ form a basis for the dual of X∞, it is enough to show
(cid:104)d∗
γ, Snx(cid:105) = 0 whenever γ ∈ Γ and n ≥ q. Since ran x ⊆ (p, q) =⇒ ran Snx ⊆ (p, q) for
γ, Snx(cid:105) = 0 for every n ∈ N and γ ∈ Γ with rank γ ≥ q. For
every n, we certainly have (cid:104)d∗
γ ∈ Γ with rank γ < q, we have (S∗)nd∗
γ = 0 whenever n ≥ q by Lemma 4.2.4, and so
(cid:104)d∗
γ, x(cid:105) = 0 as required.
γ, Snx(cid:105) = (cid:104)(S∗)nd∗
Now for a general x ∈ X∞, fix ε > 0 and choose q ∈ N such that (cid:107)(I − P(0,q))x(cid:107) ≤ ε.
Then for n ≥ q,
(cid:107)Snx(cid:107) ≤ (cid:107)SnP(0,q)x(cid:107) + (cid:107)Sn(I − P(0,q))x(cid:107) ≤ (cid:107)(I − P(0,q))x(cid:107) ≤ ε
so (cid:107)Snx(cid:107) → 0 as claimed.
Lemma 4.3.2. Let K ∈ K(X∞). Then (cid:107)SnK(cid:107) → 0 as n → ∞.
2 -net, x1, . . . xm of K(BX∞). By Lemma 4.3.1, there
2 ε whenever n ≥ N, j = 1, . . . , m. Now suppose
2 ε. For n ≥ N we have
2 ε ≤ ε. Thus (cid:107)SnK(cid:107) → 0 as n → ∞
Proof. Fix ε > 0 and choose a finite ε
exists N ∈ N such that (cid:107)Snxj(cid:107) ≤ 1
x ∈ BX∞ and let j ∈ {1, . . . , m} be such that (cid:107)Kx − xj(cid:107) ≤ 1
(cid:107)SnKx(cid:107) ≤ (cid:107)Sn(Kx − xj)(cid:107) + (cid:107)Snxj(cid:107) ≤ (cid:107)Kx − xj(cid:107) + 1
as required.
Proposition 4.3.3. Let (λj)∞
quently, the sequence ([Sj])∞
j=0 ∈ (cid:96)1. Then (cid:107)(cid:80)∞
j=0 λjSj +K(X∞)(cid:107) ≤ (cid:107)(cid:80)∞
Proof. Certainly (cid:107)(cid:80)∞
j=0 λj by Lemma 4.2.10. To
prove the converse inequality, fix ε > 0 and suppose K ∈ K(X∞). By Lemma 4.3.2, we can
find m ∈ N with (cid:107)SmK(cid:107) ≤ ε. It follows that
j=0 is 1-equivalent to the canonical basis of (cid:96)1.
j=0 λj. Conse-
j=0 λjSj + K(X∞)(cid:107) =(cid:80)∞
j=0 λjSj(cid:107) =(cid:80)∞
∞(cid:88)
λjSj+m(cid:107)
j=0
λjSj+m(cid:107) − (cid:107)SmK(cid:107)
∞(cid:88)
j=0
(cid:107)K +
λjSj(cid:107) ≥ (cid:107)SmK +
≥ (cid:107)
∞(cid:88)
∞(cid:88)
j=0 λjSj + K(X∞)(cid:107) ≥ (cid:80)∞
≥
j=0
j=0
λj − ε.
It follows that (cid:107)(cid:80)∞
(cid:107)(cid:80)∞
j=0 λjSj + K(X∞)(cid:107) ≥(cid:80)∞
j=0 λj.
120
j=0 λj − ε, and since ε > 0 was arbitrary,
Chapter 5
Shift invariant (cid:96)1 preduals
5.1 Introduction
We have seen in Chapters 3 and 4 that there exist a multitude of non-isomorphic Banach
spaces with (cid:96)1 dual. Moreover, these spaces all possess some interesting Banach space
structure; the spaces in Chapter 3 are easily seen to be saturated with infinite dimensional
reflexive subspaces and have the property that their operator algebras and Calkin algebras
can be considered 'small', whereas the interesting property of the space in Chapter 4 is the
fact it has (cid:96)1 Calkin algebra.
In addition to the results in this thesis, we also note that there are a number of other
exotic (cid:96)1 preduals discussed in the existing literature. For example, it was shown in [5] that
there exist isometric preduals of (cid:96)1 which are not isomorphic to a complemented subspace
of any C(K) space and, more recently, it has been shown ([13]) that any Banach space with
separable dual can be embedded isomorphically into an (cid:96)1 predual.
In this chapter, we briefly look at a problem considered by Haydon, Daws, Schlumprecht
and White in [9], where (cid:96)1 preduals satisfying an additional 'shift-invariant' condition are
studied. For completeness, we begin by discussing the motivation behind this property and
define precisely what we mean by a shift-invariant predual of (cid:96)1. We then show that one of
the spaces defined in [9] (which was shown to be isomorphic to c0) can in fact be considered
(in some sense) as being obtained from a specific Bourgain-Delbaen construction. Thus,
whilst none of the results in this chapter are new, our approach is somewhat different to
that taken by the authors of [9]. Given the success of constructing exotic Banach spaces via
the Bourgain-Delbaen method, the author had hoped that this new approach might lead
to new shift invariant preduals of (cid:96)1 possessing some interesting Banach space properties.
Unfortunately, at the time of writing, this remains an open problem and the author has
been unable to make any further progress.
121
The motivation behind [9] comes from the study of dual Banach algebras. If a Banach
algebra, A, is also a dual space of some Banach space, E, i.e. there exists a (Banach space)
isomorphism T : A → E∗, we can induce a weak* topology on A, simply by demanding that
T should be a weak*-weak* homeomorphism. This is, of course, equivalent to saying that
a net (xα)α∈I converges weak* to x in A if and only if T xα converges weak* to T x in E∗.
We then have the following definition:
Definition 5.1.1. A Banach algebra A is a dual Banach algebra if it is a dual space of
some Banach space, E, and the product is separately weak*-continuous (with respect to
the induced weak* topology just described).
We note that the induced weak* topology, and hence the above definition, depends on
the choice of isomorphism T : A → E∗. Therefore, one should be more precise and define
a predual of A as a Banach space E together with an isomorphism T : A → E∗. We shall
see shortly that we can forget about this technicality for the purposes of this chapter as we
will be working with so-called concrete preduals of (cid:96)1 (see Definition 5.1.2).
As noted in [9], it is well known that a C∗-algebra, M , which is isometric to a dual
space is a von Neumann algebra. In this case, the product and involution are automatically
weak* continuous, so that M is a dual Banach algebra. Moreover, the (isometric) predual
is unique. However, we observe that isomorphic preduals need not be unique, even in the
case of C∗ algebras. Indeed, Pelczy´nski proved that the Banach spaces (cid:96)∞ and L∞[0, 1] are
isomorphic. It follows that both (cid:96)1 and L1[0, 1] are isomorphic preduals of the C∗ algebra,
(cid:96)∞, while of course (cid:96)1 is not isomorphic to L1[0, 1].
The authors of [9] are specifically concerned with the Banach algebra (cid:96)1(Z) equipped
with the natural convolution product,
(f ∗ g)(n) =
f (k)g(n − k),
f, g ∈ (cid:96)1(Z), n ∈ Z
(cid:88)
k∈Z
and the isomorphic preduals of (cid:96)1(Z) with respect to which (cid:96)1(Z) is a dual Banach algebra.
We should really be a little more precise. Observe that if we were to equip (cid:96)1(Z) with the
zero product, any isomorphic predual of (cid:96)1(Z) makes (cid:96)1(Z) equipped with the zero product
into a dual Banach algebra. We are only interested in a predual E of (cid:96)1(Z) with respect to
which (cid:96)1(Z) equipped with the convolution product (defined above) is a dual Banach algebra.
Consequently, in what follows, we are always thinking of (cid:96)1(Z) as the 'usual' Banach algebra
equipped with the multiplication being the convolution just defined. We also remark that
the canonical isometry of c0(Z)∗ with (cid:96)1(Z) makes (cid:96)1(Z) into a dual Banach algebra, so the
sorts of preduals we are interested in certainly exist. Whilst this happens to be an isometric
predual, we emphasise that we are not restricting our attention to just isometric preduals.
122
In [9], the authors are looking for preduals E for which E has some interesting Banach
space structure or for which the isomorphism between E and (cid:96)1(Z) induces an exotic weak*
topology.
Before continuing any further we remark that we shall frequently be making the canoni-
cal identification of (cid:96)1(Z)∗ with (cid:96)∞(Z) and find it convenient to denote by θ : (cid:96)1(Z)∗ → (cid:96)∞(Z)
the usual isometry.
Observe that every predual, E, of (cid:96)1(Z) can be canonically embedded (isomorphically)
into (cid:96)∞(Z). Indeed, if JE denotes the canonical isometric embedding of E into its second
dual and T : (cid:96)1(Z) → E∗ is an isomorphism, the composition θ ◦ T ∗ ◦ JE : E → (cid:96)∞(Z) is
an isomorphic embedding of E into (cid:96)∞(Z). It turns out (see Lemmas 5.1.3 and 5.1.4) that
the preduals of (cid:96)1(Z) which make the multiplication separately weak* continuous, i.e. the
preduals with respect to which (cid:96)1(Z) is a dual Banach algebra, are precisely those which,
when thought of as a subspace of (cid:96)∞(Z) under this embedding, are invariant under the
bilateral shift on (cid:96)∞(Z). Consequently, we refer to such preduals as shift-invariant preduals
of (cid:96)1(Z).
If was shown in [9] that, when considering shift-invariant preduals of (cid:96)1(Z), we lose
no generality by considering so called concrete preduals. As well as being a considerable
simplification, the (cid:96)1 duality arising from concrete preduals is precisely that which arises
naturally from the embedding of (cid:96)1(Z) into the dual of a space of Bourgain-Delbaen type.
Consequently, concrete preduals are fundamental, both to the work of the authors in [9]
and to enable us to connect the work of this thesis with the results in the aforementioned
paper. For this reason, we chose to include the details of concrete preduals now; the proofs
and results largely follow the work of the authors in [9].
Given a closed subspace F ⊆ (cid:96)∞(Z), it is well known that the dual space F ∗ is canon-
ically isometric to (cid:96)∞(Z)∗/F ⊥, where F ⊥ = {Φ ∈ (cid:96)∞(Z)∗ : (cid:104)Φ, x(cid:105) = 0 ∀x ∈ F}. Let
I : (cid:96)∞(Z)∗/F ⊥ → F ∗ denote this identification and π : (cid:96)∞(Z)∗ → (cid:96)∞(Z)∗/F ⊥ be the obvi-
ous projection.
Definition 5.1.2. Given a closed subspace F ⊆ (cid:96)∞(Z), we define the map ιF : (cid:96)1(Z) → F ∗
to be the composition ιF = Iπ(θ−1)∗J(cid:96)1(Z) and say F is a concrete predual for (cid:96)1(Z) if the
map ιF is an isomorphism.
It is an easy computation to see that if a ∈ (cid:96)1(Z) and x ∈ F ⊆ (cid:96)∞(Z), then (cid:104)ιF (a), x(cid:105) =
(cid:104)x, a(cid:105) (where we are of course thinking of x as an element of (cid:96)1(Z)∗ in the right-hand-side of
the equality). Consequently, if F is a concrete predual, the identification of (cid:96)1 with F ∗ under
ιF is the same type of identification of (cid:96)1 as a subspace of X∗ when X is a Bourgain-Delbaen
space (see, for example, Theorem 2.2.2, Proposition 3.4.5).
123
As stated earlier, we lose no generality in working with concrete preduals for (cid:96)1(Z) and
this is the contents of the next lemma (which is proved in [9, Lemma 2.1]). We remark that
the lemma states both the Banach space structure and induced weak* topologies on (cid:96)1(Z)
are preserved when changing to the setup of a concrete predual. This is fundamental given
we are concerned with those preduals which turn (cid:96)1(Z) into a dual Banach algebra.
Lemma 5.1.3. Let E be a Banach space and T : (cid:96)1(Z) → E∗ be an isomorphism. Then the
map θT ∗JE : E → (cid:96)∞(Z) is an isomorphism onto its range, say F ⊆ (cid:96)∞(Z). Furthermore,
ιF is an isomorphism so that F is a concrete predual for (cid:96)1(Z) and the weak* topologies
induced by T : (cid:96)1(Z) → E∗ and ιF : (cid:96)1(Z) → F ∗ agree. That is, given a net (aα) in (cid:96)1(Z), we
have that limα(cid:104)T (aα), x(cid:105) = 0 for all x ∈ E if and only if limα(cid:104)ιF (aα), y(cid:105) = limα(cid:104)y, aα(cid:105) = 0
for all y ∈ F .
Proof. The first statement is obvious. We prove that ιF is an isomorphism and the desired
properties of the weak* topologies. Let R denote the isomorphism obtained by thinking
of the operator θT ∗JE as a map onto its image, that is, R = θT ∗JE : E → F (where
F ⊆ (cid:96)∞(Z) is defined as in the lemma). Note that for x ∈ E, a ∈ (cid:96)1(Z),
(cid:104)R∗ιF (a), x(cid:105) = (cid:104)ιF (a), Rx(cid:105) = (cid:104)θ−1Rx, a(cid:105)
= (cid:104)T ∗JEx, a(cid:105) = (cid:104)T a, x(cid:105).
In other words, R∗ιF = T . It follows that ιF = (R∗)−1T and thus ιF is an isomorphism
from (cid:96)1(Z) to F ∗ as required.
Next, suppose (aα) is a net in (cid:96)1(Z). We have (aα) weak* convergent to 0 with respect
to the topology induced by the isomorphism ιF if and only if (cid:104)ιF (aα), y(cid:105) → 0 for all y ∈ F .
Since ιF = (R−1)∗T , this happens if and only (cid:104)(R−1)∗T (aα), y(cid:105) = (cid:104)T aα, R−1y(cid:105) → 0 for all
y ∈ F . Since R : E → F is an isomorphism, this happens if and only if (cid:104)T aα, x(cid:105) → 0 for
all x ∈ E, that is, if and only if (aα) is weak* convergent to 0 with respect to the weak*
topology induced by the isomorphism T : (cid:96)1(Z) → E∗ as required.
For completeness, we state and prove two more lemmas. The results and proofs are
taken from [9] and the author of this thesis makes no claim to originality of the two results
that follow. The first lemma proves that the shift invariant condition discussed at the
beginning of this chapter is indeed equivalent to requiring the predual make (cid:96)1 with its
natural convolution product into a dual Banach algebra. Of course, since we now know
that we lose no generality in working with concrete preduals, we choose to work with these
in the lemma. The second lemma provides further evidence that concrete preduals do indeed
provide a simplification to the study of the problem originally motivating the authors of
124
[9].
Indeed, we see that concrete preduals allow one to detect if two concrete preduals
F1, F2 ⊆ (cid:96)∞(Z) induce the same weak* topology on (cid:96)1(Z); this happens if and only if
F1 = F2.
Lemma 5.1.4. Let F ⊆ (cid:96)∞(Z) be a concrete predual for (cid:96)1(Z). The following are equivalent:
(1) The bilateral shift on (cid:96)1(Z) is weak*-continuous with respect to F .
(2) The subspace F is invariant under the bilateral shift on (cid:96)∞(Z).
(3) (cid:96)1(Z) is a dual Banach algebra with respect to F .
Proof. We show that (1) and (3) are equivalent and (2) and (3) are equivalent. We denote
by en the usual vector in (cid:96)1(Z) (i.e. en = (. . . , 0, 0, 1, 0, 0 . . . ) ∈ (cid:96)1(Z) where the 1 occurs in
the n'th coordinate) and, for y ∈ (cid:96)1(Z), we denote by ∗y : (cid:96)1(Z) → (cid:96)1(Z) the linear operator
defined by x (cid:55)→ x ∗ y.
If (3) holds then, in particular, the maps ∗e1 and ∗e−1 on (cid:96)1(Z) are weak* continuous
with respect to F . Note that if n ∈ Z,
x ∗ e1(n) =
(cid:88)
x(k)e1(n − k) = x(n − 1)
k∈Z
so the map ∗e1 is just the right shift on (cid:96)1(Z). It is just as easy to see that ∗e−1 is the left
shift operator. So we conclude that the bilateral shifts on (cid:96)1(Z) are weak* continuous with
respect to F .
Conversely, suppose (1) holds, i.e. the right and left shifts on (cid:96)1(Z) are weak* continuous.
We must see that for a fixed y ∈ (cid:96)1(Z) the map ∗y on (cid:96)1(Z) is weak* continuous with respect
to F .
A similar computation to the one above shows that for j ∈ N, the map ∗ej
the right shift operator on (cid:96)1(Z) applied j times and the map ∗e−j
(cid:96)1(Z) applied j times. Thus if there is some N ∈ N such that y = (cid:80)j≤N ajej then
∗y =(cid:80)j≤N aj∗ej . For an arbitrary y ∈ (cid:96)1(Z), we can write y = limN→∞(cid:80)j≤N ajej. We
write PN y =(cid:80)j≤N ajej. Observe that
the map ∗y is weak* continuous since it is just the finite sum of weak* continuous maps;
is the left shift on
is just
(cid:107)x ∗ y − x ∗ PN y(cid:107) = (cid:107)x ∗ (y − PN y)(cid:107) ≤ (cid:107)x(cid:107)(cid:107)y − PN y(cid:107)
so that ∗y is the the limit (in the operator norm topology) of the (weak* continuous)
operators ∗PN y. It follows by the second part of Lemma 1.4.12 that ∗y is weak* continuous,
as required.
125
Now suppose F is invariant under the left shift operator on (cid:96)∞(Z). We claim the right
x,
shift, σ, on (cid:96)1(Z) is weak*-continuous. Indeed, let (xα) be a net in (cid:96)1(Z) with xα (cid:42)w∗
i.e. (cid:104)θ−1f, xα(cid:105) → (cid:104)θ−1f, x(cid:105) for all f ∈ F ⊆ (cid:96)∞(Z). We are required to see σxα (cid:42)w∗
σx.
To see this, let f ∈ F ⊆ (cid:96)∞(Z) and note that
(cid:104)θ−1f, σxα(cid:105) = (cid:104)σ∗θ−1f, xα(cid:105)
= (cid:104)θ−1 ◦ (θσ∗θ−1)f, xα(cid:105).
An easy computation shows that the operator θσ∗θ−1 : (cid:96)∞(Z) → (cid:96)∞(Z) is just the left shift
on (cid:96)∞(Z). Indeed, if (am)m∈Z is in (cid:96)∞(Z) and em denotes the m'th standard basis unit
vector of (cid:96)1(Z) as before, we have
(cid:0)θσ∗θ−1 ((am)m∈Z)(cid:1) (n) = (cid:104)σ∗(cid:0)θ−1(am)(cid:1) , en(cid:105) = (cid:104)θ−1(am), σen(cid:105) = an+1
so θσ∗θ−1 is the left shift on (cid:96)∞ as required.
Now, since we are assuming that F is invariant under the left shift, f(cid:48) := θσ∗θ−1f ∈ F .
Thus, from our previous calculation, and that xα (cid:42)w∗
x we have,
(cid:104)θ−1f, σxα(cid:105) = (cid:104)θ−1f(cid:48), xα(cid:105) → (cid:104)θ−1f(cid:48), x(cid:105)
= (cid:104)σ∗θ−1f, x(cid:105)
= (cid:104)θ−1f, σx(cid:105)
as required. A similar argument yields that if F is invariant under the right shift operator
on (cid:96)∞(Z), then the left shift on (cid:96)1(Z) is weak* continuous. We have thus proved that (2)
=⇒ (3).
Conversely, suppose we know the right shift σ on (cid:96)1(Z) is weak* continuous. We will
show F is left shift invariant. We continue with the notation used previously, recalling that
θσ∗θ−1 : (cid:96)∞(Z) → (cid:96)∞(Z) is the left shift on (cid:96)∞(Z).
Now, if our claim is false, there exists an element f(cid:48) ∈ F such that it's image under the
left shift, θσ∗θ−1f(cid:48) ∈ (cid:96)∞(Z)\ F . By the Hahn-Banach Theorem, we may find a γ∗ ∈ (cid:96)∞(Z)∗
such that
(i) (cid:104)γ∗, θσ∗θ−1f(cid:48)(cid:105) = 1
(ii) γ∗(F ) = {0} or, equivalently, (cid:104)γ∗, f(cid:105) = 0 for all f ∈ F .
Note that θ∗γ∗ ∈ (cid:96)1(Z)∗∗. We can rewrite (i) above as (cid:104)(θ−1)∗θ∗γ∗, θσ∗θ−1f(cid:48)(cid:105) =
(cid:104)θ∗γ∗, σ∗θ−1f(cid:48)(cid:105) = 1, and similarly, we can rewrite (ii) as (cid:104)θ∗γ∗, θ−1f(cid:105) = 0 for all f ∈ F .
By Goldstine's Theorem, there is a net (xα)α∈I in (cid:96)1(Z) with (cid:107)xα(cid:107)(cid:96)1 ≤ (cid:107)θ∗γ∗(cid:107) for every
126
α and J(cid:96)1xα (cid:42)w∗
limα(cid:104)θ−1f, xα(cid:105) = 0 for all f ∈ F .
θ∗γ∗. We conclude that limα(cid:104)σ∗θ−1f(cid:48), xα(cid:105) = limα(cid:104)θ−1f(cid:48), σxα(cid:105) = 1 and
Now (xα) is a net in some closed ball of (cid:96)1(Z) (cid:39) F ∗. By the weak* compactness of
closed balls in F ∗, (xα) has a subnet (xg(β))β which is weak* convergent to some x ∈ (cid:96)1(Z).
Since (cid:104)θ−1f(cid:48), σxg(β)(cid:105) is a subnet of (cid:104)θ−1f(cid:48), σxα(cid:105), the latter of which we know converges to
1, we have limβ(cid:104)θ−1f(cid:48), σxg(β)(cid:105) = 1. Moreover, by the assumed weak* continuity of σ, we
have limβ(cid:104)θ−1f(cid:48), σxg(β)(cid:105) = (cid:104)θ−1f(cid:48), σx(cid:105) and therefore, (cid:104)θ−1f(cid:48), σx(cid:105) = 1.
On the other hand, by the very definition of the weak* topology induced on (cid:96)1(Z), we
have (cid:104)θ−1f, x(cid:105) = limβ(cid:104)θ−1f, xg(β)(cid:105) = limα(cid:104)θ−1f, xα(cid:105) = 0 for all f ∈ F . This is equivalent
to saying (cid:104)ιF (x), f(cid:105) = 0 for all f ∈ F , which of course tells us that ιF (x) = 0, and hence
x = 0 (ιF is an isomorphism). This clearly contradicts (cid:104)θ−1f(cid:48), σx(cid:105) = 1. Again, one can
argue similarly to prove that if the left shift on (cid:96)1(Z) is weak* continuous, then F must be
right shift invariant. This completes the proof.
Lemma 5.1.5. Let E1 and E2 be preduals of (cid:96)1(Z), and use these to induce concrete
preduals F1, F2 ⊆ (cid:96)∞(Z) as described in Lemma 5.1.3. Then E1 and E2 induce the same
weak* topology on (cid:96)1(Z) if and only if F1 = F2. In particular, two concrete preduals F1 and
F2 induce the same weak* topology on (cid:96)1(Z) if and only if F1 = F2.
Proof. If F1 = F2 then it follows immediately from Lemma 5.1.3 that E1 and E2 induce
the same weak* topology on (cid:96)1(Z). Conversely, for i=1,2, let Ti : (cid:96)1(Z) → E∗
i be an isomor-
phism, and suppose that these induce the same weak* topology on (cid:96)1(Z). Suppose for a
contradiction that there exists an x ∈ F2 \ F1. By the Hahn-Banach Theorem, there exists
Λ ∈ (cid:96)∞(Z)∗ with (cid:104)Λ, x(cid:105) = 1 and (cid:104)Λ, y(cid:105) = 0 for all y ∈ F1. By Goldstine's Theorem, there is
θ∗Λ. It follows that, for all y ∈ F1, we have
a bounded net (aα) in (cid:96)1(Z) with J(cid:96)1(Z)aα (cid:42)w∗
(cid:104)θ−1y, aα(cid:105) = lim
0 = (cid:104)Λ, y(cid:105) = (cid:104)θ∗Λ, θ−1y(cid:105) = lim
(cid:104)ιF1(aα), y(cid:105)
α
α
so that (T1(aα)) is weak* null in E∗
is weak* null in E∗
2 , but this contradicts that
1 by Lemma 5.1.3. By assumption, it follows that (T2(aα))
1 = (cid:104)Λ, x(cid:105) = lim
α
(cid:104)θ−1x, aα(cid:105) = lim
(cid:104)ιF2(aα), x(cid:105).
α
This shows that F2 ⊆ F1, and an identical argument shows F1 ⊆ F2, as required.
5.2 Connection to the BD construction
Having described the motivation behind the work in [9] and understood the importance of
concrete preduals, we now move on to discuss a specific class of shift invariant preduals that
127
were introduced in [9]. We follow the same notation used in [9]; for n ≥ 0 in Z, we denote
by b(n) the number of ones in the binary expansion of n. (So b(0) = 0, b(1) = 1, b(2) =
1, b(3) = 2 and so on.)
Definition 5.2.1. Let λ ∈ R,λ > 1. Define the vector xλ
when n ≥ 0 and 0 otherwise. Thus xλ
0 is given by
0 ∈ (cid:96)∞(Z) by xλ
0 (n) = λ−b(n)
x0 = (. . . 0, 0, 1, λ−1, λ−1, λ−2, λ−1, λ−2, λ−2, λ−3, λ−1, . . . ).
Fλ ⊆ (cid:96)∞(Z) is defined to be the closed, shift-invariant subspace of (cid:96)∞(Z) generated by the
vector xλ
0 , i.e. Fλ is the closed linear span of the bilateral shifts of xλ
0 .
It was shown in [9] that the collection of subspaces (Fλ)λ>1 of (cid:96)∞(Z) are all concrete
preduals of (cid:96)1(Z). Clearly, by construction, they are shift invariant and so by Lemma 5.1.4,
they are preduals of (cid:96)1(Z) that make (cid:96)1(Z) with the convolution product a dual Banach
algebra. The authors of [9] used a somewhat technical argument involving the Szlenk index
to show that as Banach spaces, the spaces Fλ are all in fact isomorphic to c0. The interest
in the spaces Fλ is therefore due to the weak* topologies that they induce on (cid:96)1(Z). Since
the Fλ are pairwise distinct subspaces of (cid:96)∞(Z), and distinct from c0, it is immediate from
Lemma 5.1.5 that these spaces induce an uncountable family of distinct weak*-topologies
making (cid:96)1(Z) into a dual Banach algebra.
The final result of this thesis, is an observation about the space F2. We will show that
we can obtain a 'one-sided' version of the space F2 using the Bourgain Delbaen construction.
When viewed in this framework, it is very easy to see that the one-sided version of F2 is
isomorphic as a Banach space to c0 and that its dual space is naturally isomorphic to (cid:96)1.
In what follows N0 denotes the set of natural numbers including 0. We first set up the
notation.
We define the vector y0 ∈ (cid:96)∞(N0) by y0(n) = 2−b(n) where, as before, b(n) denotes the
number of ones in the binary expansion of n. Thus y0 is given by
(1, 2−1, 2−1, 2−2, 2−1, . . . )
We define R2 to be the closed, right-shift-invariant subspace of (cid:96)∞(N0) generated by y0,
that is, R2 is the closed linear span of the vector y0 and all its right shifts.
We also use Theorem 2.2.2 to construct a space of Bourgain-Delbaen type. Comparing
with the notation from Theorem 2.2.2, we take Γ = N0 and define finite subsets ∆n ⊂ N0
as follows. We set ∆0 = {0}, and for n ≥ 1, we set ∆n = {2n−1 + j : 0 ≤ j < 2n−1}. For
each k ≥ 1, if n ∈ ∆k, we can uniquely write n = 2k−1 + m where m < 2k−1 and m, k ∈ N0.
Again comparing with the notation from Theorem 2.2.2 and making using of the previous
128
observation, we define a τ mapping by setting τ (n) to be the Type 0 tuple (1, 1
2 , m). With
this setup, it is easily seen that the Bourgain-Delbaen Theorem (Theorem 2.2.2) applies.
n ∈ (cid:96)1(N0). For clarity,
m ∈ (cid:96)1(N0) and d∗
We get corresponding vectors c∗
2 e∗
n = 1
the first few d∗ vectors are as follows:
n = e∗
n − c∗
d∗
0 = (1, 0, 0, 0, 0, 0, . . . )
1 = (−1/2, 1, 0, 0, 0, 0, . . . )
d∗
2 = (−1/2, 0, 1, 0, 0, 0, . . . )
d∗
3 = (0,−1/2, 0, 1, 0, 0, . . . ).
d∗
We obtain a L∞ space of Bourgain-Delbaen-type, RBD := X(N0, τ ). We recall that RBD is
nothing other than the closed linear span of the biorthogonal vectors [dn : n ∈ N0] ⊆ (cid:96)∞(N0).
The main observation of this chapter is the following lemma:
Lemma 5.2.2. The spaces R2 and RBD defined above are the same subspace of (cid:96)∞(N0).
The vectors (d∗
n)∞
n=1 are in fact a basic sequence in (cid:96)1(N0) equivalent to the canonical basis
of (cid:96)1(N0). Consequently, RBD is a right-shift-invariant subspace of (cid:96)∞ which is isomorphic
to c0 and has dual space naturally isomorphic to (cid:96)1(N0).
Proof. We first show that R2 and RBD are the same subspace of (cid:96)∞(N0). The proof involves
nothing other than a number of lengthy computations so we proceed by proving a number
of smaller claims. It is convenient for us to re-introduce the τ operator that featured in [9].
Precisely, τ : (cid:96)∞(N0) → (cid:96)∞(N0) is the bounded linear map defined by
τ (x)(n) =
x( n
0
2 ) n even
n odd.
We denote by σ : (cid:96)∞(N0) → (cid:96)∞(N0) the right shift, i.e. σx(n) = x(n − 1). It was shown in
[9, Lemma 3.1] that τ k(y0) ∈ R2 for all k ∈ N0 and moreover, τ σ = σ2τ .
We also define a bounded operator β : (cid:96)∞(N0) → (cid:96)∞(N0) by
(cid:40)
(cid:40)
β(x)(n) =
We note that στ = β so that β(y0) ∈ R2.
Step 1. d0 = y0.
x( n−1
0
2 ) n odd
n even.
We use induction on n ∈ N0 to prove that d0(n) = (cid:104)d0, e∗
0(cid:105) = (cid:104)d0, d∗
0(cid:105) = 1 = 2−b(0) as required. Inductively, assume that (cid:104)d0, e∗
n(cid:105) = 2−b(n) = y0(n). Clearly
j(cid:105) = 2−b(j) for all
m(cid:105) where n + 1 = 2k + m with m < 2k.
(cid:104)d0, e∗
j ≤ n. Then (cid:104)d0, e∗
0, e∗
n+1(cid:105) = (cid:104)d0, d∗
n+1 + c∗
n+1(cid:105) = 1
2(cid:104)d∗
129
In particular, m ≤ n, so by the inductive hypothesis, (cid:104)d0, e∗
m < 2k, we can write n + 1 = 2k +(cid:80)k−1
2 2−b(m). Note that since
j=0 εj2j where εj ∈ {0, 1} for all j. It follows that
n+1(cid:105) = 1
b(n + 1) = b(m) + 1 and so (cid:104)d0, e∗
Step 2. For all j ∈ N0, d2j+1 = βdj = στ dj.
n+1(cid:105) = 2−b(n+1), completing the induction.
We use induction on n ∈ N0 to prove that d2j+1(n) = βdj(n). It is easily seen that
for n < 2j + 1, d2j+1(n) = 0. Moreover, for such n, βdj(n) = 0 if n is even. When n is
odd and n < 2j + 1, (n − 1)/2 < j so that βdj(n) = dj((n − 1)/2) = 0. Consequently
the inductive statement holds when n < 2j + 1. When n = 2j + 1 it is easily verified that
d2j+1(n) = βdj(n) = 1.
We now suppose that the inductive statement holds for all k ≤ n and must show that
d2j+1(n + 1) = βdj(n + 1). By the preceding argument, we may as well assume that
n ≥ 2j + 1, so n + 1 > 2j + 1 ≥ 1. We write n + 1 = 2k + m where 0 ≤ m < 2k and since
n + 1 ≥ 2 we must have k ≥ 1. Now d2j+1(n + 1) = (cid:104)d2j+1, c∗
n+1(cid:105) = 1
m(cid:105). Certainly
m ≤ n, so we may apply the apply inductive hypothesis which yields
2(cid:104)d2j+1, e∗
(cid:40) 1
2 dj( m−1
d2j+1(n + 1) =
1
2
βdj(m) =
2 ) m odd
0
otherwise.
Finally, we consider cases where n + 1 is even or odd separately, noting that n + 1 is even
if and only if m is even, since we have seen that k ≥ 1. Consequently, if n + 1 is even, then
our above calculation yields that d2j+1(n + 1) = 0 = βdj(n + 1). It remains only to consider
2 dj((m− 1)/2).
the possibility that n + 1 is odd, in which case, m is odd and d2j+1(n + 1) = 1
We need to see that this is equal to βdj(n + 1) = dj(n/2). Since we assumed n ≥ 2j + 1,
we have n/2 ≥ j + 1/2 and since n/2 ∈ Z, in fact, n/2 ≥ j + 1 > j. Consequently,
dj(n/2) = (cid:104)dj, c∗
n/2(cid:105). Recalling that n + 1 = 2k + m, where m must now be odd, we see that
n/2 = 2k−1 + m−1
Step 3. For all j ≥ 1, d2j = τ dj.
and consequently (cid:104)dj, c∗
2 ) as required.
2 dj( m−1
n/2(cid:105) = 1
2(cid:104)dj, e∗
(cid:105) = 1
m−1
2
2
Again, we use induction on n and prove that d2j(n) = τ dj(n) for all n ∈ N0. The
calculations are very similar to the previous step. When n < 2j it's easy to see that
d2j(n) = 0 = τ dj(n) and when n = 2j, d2j(n) = 1 = τ dj(n). So we suppose that the
inductive statement holds for all k ≤ n and prove the statement must hold for n + 1. By
the previous argument, we can assume n ≥ 2j ≥ 2. Then n + 1 > 2j and d2j(n + 1) =
(cid:104)d2j, c∗
m(cid:105) where n + 1 = 2k + m and we must have k ≥ 1 since n + 1 ≥ 3. By
the inductive hypothesis we have d2j(n + 1) = 1
2 τ dj(m). Now, if n + 1 is odd, m must also
be odd, so d2j(n + 1) = 1
2 τ dj(m) = 0 = τ dj(n + 1). So we assume now that n + 1 is even,
(cid:105)
so m is even and d2j(n + 1) = 1
2 dj(m/2). In this case, τ dj(n + 1) = dj( n+1
2 ) = (cid:104)dj, c∗
n+1(cid:105) = 1
2(cid:104)d2j, e∗
n+1
2
130
since we recall that n + 1 > 2j. Also n + 1 = 2k + m so that (n + 1)/2 = 2k−1 + m/2. It
follows that τ dj(n + 1) = (cid:104)dj, c∗
Step 4. d2j +m = σmd2j whenever j ≥ 1, 1 ≤ m < 2j.
2 dj(m/2) = d2j(n + 1) as required.
(cid:105) = 1
n+1
2
To see this, we will show by induction on j ∈ N that d2j +m = σmd2j whenever 1 ≤
m < 2j. The base case, j = 1 follows easily from the previous two steps. Indeed, the only
permissible value of m is 1 and d21+1 = στ d1 = σd21 as required.
So, suppose by induction, that for some j > 1, d2j−1+m = σmd2j−1 whenever 1 ≤ m <
2j−1. We must show that d2j +m = σmd2j whenever 1 ≤ m < 2j. We will do this by
an induction on m. The base case, m = 1, is again easy. By Steps 2 and 3, we have
d2j +1 = στ d2j−1 = σd2j as needed. So suppose that d2j +m = σmd2j and that m + 1 < 2j.
If m = 2k, we have d2j +m+1 = d2(2j−1+k)+1 = στ d2j−1+k = σd2j +2k = σd2j +m = σm+1d2j .
Here we have appealed to the previous two steps and the inductive hypothesis on m.
It remains to consider the case m = 2k + 1. Again, we will make repeated use of the
previous two steps. We have d2j +m+1 = d2j +2k+2 = d2(2j−1+k+1) = τ d2j−1+k+1. Moreover,
m + 1 = 2(k + 1) < 2j so that k + 1 < 2j−1. Thus by our inductive hypothesis on j,
d2j−1+k+1 = σk+1d2j−1 and so d2j +m+1 = τ σk+1d2j−1 = σ2k+2τ d2j−1 = σ2k+2d2j = σm+1d2j .
(Here we made use of the fact that τ σ = σ2τ .) This completes both inductive steps.
Step 5. σdn = dn+1 except if n is of the form n = 2j − 1, with j ≥ 0.
This is almost immediate from the previous step. Indeed, suppose j ≥ 1 and 1 ≤ m <
2j − 1. Then if n = 2j + m, σdn = σd2j +m = σm+1d2j = d2j +m+1 = dn+1. We made
use of the fact that m + 1 < 2j so that we could apply the previous step to obtain the
penultimate equality. If n = 2j with j ≥ 1 then dn+1 = d2(2j−1)+1 = στ d2j−1 by Step 2, and
στ d2j−1 = σd2j = σdn by Step 3.
We therefore have that σdn = dn+1 whenever n can be written in the form 2j + m with
j ≥ 1 and 0 ≤ m < 2j − 1. The only values of n which aren't of the form n = 2j + m with
j ≥ 1 and 0 ≤ m < 2j − 1 are precisely those n of the form 2k − 1 for some k ∈ N0.
Step 6. For j ≥ 0,
(cid:40)
d2j (n) =
d0(l)
0
if n = 2j(2l + 1) (for some l ∈ N0)
otherwise.
In particular the vectors {d2j : j ∈ N0} have disjoint supports.
We use induction on j ∈ N0. When j = 0 we have d1(n) = στ d0(n) by Step 2. So
d1(n) = τ d0(n − 1) =
d0( n−1
2 )
0
if n − 1 even
otherwise.
(cid:40)
131
It follows that d1(n) = 0 unless n = 2l + 1 in which case d1(n) = d0(l) as required.
Now assume the claim holds for some j ≥ 0. We have d2j+1(n) = d2·2j (n) = τ d2j (n) by
Step 3. So
d2j+1(n) =
d2j ( n
2 )
0
if n even
otherwise.
(cid:40)
From this, and the inductive hypothesis, we see that d2j+1(n) = 0 unless n/2 = 2j(2l + 1)
for some l ∈ N0, i.e.
if n = 2j+1(2l + 1), in which case we have d2j+1(n) = d0(l). The
statement about disjoint supports is obvious.
Step 7. For all j ≥ 1,
d2j−1(n) =
d0(k)
0
if there is k ∈ N0 such that n = 2j(k + 1) − 1
otherwise.
(cid:40)
We proceed by induction on j. When j = 1, d1(n) = στ d0(n) by Step 2. So d1(n) =
τ d0(n − 1); this equals d0(k) if n = 2k + 1 = 2(k + 1) − 1 for some k ∈ N0 and 0 otherwise
by the definition of the τ operator.
Assuming the claim holds for some j ≥ 1, d2j+1−1(n) = d2(2j−1)+1(n) = στ d2j−1(n) by
Step 2. So
d2j+1−1(n) = τ d2j−1(n − 1) =
(cid:40)
d2j−1( n−1
2 )
0
if n − 1 even
otherwise.
0 = d∗
If m = 0, then σd0(n) = σd0(2l + 1) = d0(2l) = d0(l).
k=0 2−kd2k (n) =(cid:80)∞
From this, and the inductive hypothesis, we see that d2j+1−1(n) (cid:54)= 0 if and only if (n−1)/2 =
Step 8. For all j ≥ 0, σd2j−1 =(cid:80)∞
2j(k + 1) − 1 for some k ∈ N0 in which case d2j+1−1(n) = d0(k) as required.
k=0 2−kd2k+j .
(cid:80)∞
We first treat the case j = 0. Since e∗
0 we easily see that σd0(0) = 0 =
k=0 2−kd2k (0). Now, for every n ∈ N, we can write n = 2m(2l + 1) for unique m, l ∈ N0.
here and the fact that b(2l) = b(l).) By Step 6,(cid:80)∞
(We have made use of Step 1
k=0 2−kd2k (2l + 1) =
n = 2m(2l + 1) and m ≥ 1. In this case, (cid:80)∞
d1(2l + 1) = d0(l) = σd0(n) where the penultimate equality follows by Step 7. Otherwise,
other hand, we can write n =(cid:80)N
k=0 2−kd2k (n) = 2−md0(l) by Step 6. On the
and εm = 1. It follows that n− 1 = 2m − 1 +(cid:80)N
k=m εk2k for some N ≥ m, where εk ∈ {0, 1} for k > m
k=m+1 εk2k. We
see from this that b(n − 1) = b(n) + m − 1. Consequently, σd0(n) = d0(n − 1) = 2−b(n−1)
2−b(l)−m = 2−md0(l) =(cid:80)∞
(by Step 1) = 2−b(n)−m+1. Finally we note that b(n) = b(2l + 1) = b(l) + 1 so, σd0(n) =
that σd2j−1(0) = 0 = (cid:80)∞
The calculation for when j ≥ 1 is only slightly more involved. It is again easily seen
If n ∈ N, we again write n = 2m(2l + 1) and
σd2j−1(n) = d2j−1(n − 1) = d2j−1(2m(2l + 1) − 1). By Step 7, this will be non-zero if and
k=m+1 εk2k =(cid:80)m−1
k=0 2k +(cid:80)N
k=0 2−kd2k (n) as required.
k=0 2−kd2k+j (0).
132
a=0
a=m−j+1 εa2a =(cid:80)m−j−1
only if m ≥ j. Moreover, if m = j, i.e. n = 2m(2l + 1), then we see from the previous
calculation and Step 7 that σd2j−1(n) = d0(2l). By Step 1, this equals d0(l). If m > j, then
that 2m−j(2l + 1) =(cid:80)N
n = 2j(2m−j(2l + 1)) and by Step 7, we have σd2j−1(n) = d0(2m−j(2l + 1) − 1). We note
εm−j = 1. It follows that 2m−j(2l + 1) − 1 = 2m−j − 1 +(cid:80)N
a=m−j εa2a for some N ≥ m− j, where εa ∈ {0, 1} for a > m− j and
(cid:80)N
2a +
a=m−j+1 εa2a. This implies that b(2m−j(2l + 1) − 1) = b(2m−j(2l + 1)) + m − j − 1.
So, σd2j−1(n) = d0(2m−j(2l + 1) − 1) = 2−b(2m−j (2l+1)−1) = 2−m+j+12−b(2m−j (2l+1)) =
It remains to compute (cid:80)∞
2−m+j+12−b(2l+1) = 2−m+j2−b(l) = 2−m+jd0(l).
get (cid:80)∞
we have (cid:80)∞
k=0 2−kd2k+j (n) and see that the formulae we obtain agree
with those just found for σd2j−1(n). We note that if n = 2m(2l + 1) then d2k+j (n) will
only be non-zero if k + j = m by Step 6. In particular, if n is such that m < j then we
k=0 2−kd2k+j (n) = 0 = σd2j−1(n). Otherwise, if n = 2m(2l + 1) is such that m ≥ j
k=0 2−kd2k+j (n) = 2−kd0(l) where k is such that k + j = m, by Step 6. The
latter expression is therefore equal to 2−m+jd0(l) = σd2j−1(n), as found in the previous
paragraph.
We are finally in a position to show that R2 = RBD. By Step 5 and Step 8, we clearly
(cid:1) = σ(cid:0)[dn : n ∈ N0](cid:1) ⊆ RBD. Since, by Step 1, d0 = y0, we conclude that
have that σ(cid:0)RBD
RBD is a closed, right-shift invariant subspace of (cid:96)∞(N0) containing y0; we must therefore
have R2 ⊆ RBD.
On the other hand, since τ k(y0) = τ k(d0) ∈ R2 for all k, and σ2τ = τ σ, it is easily
seen by Steps 2 and 3 that for each n ∈ N there are l, m ∈ N0 with dn = σmτ ld0 ∈ R2.
Consequently RBD ⊆ R2 and so the two subspaces are indeed equal.
n(cid:107) ≤ 1
Finally, we observe that in this construction, (cid:107)c∗
2 for all n ∈ N0. Appealing to
Lemma 2.2.4, we see that RBD is isomorphic to c0. Moreover, the sequence of vectors
(d∗
n)∞
n) in (cid:96)1(N0) is equivalent to (e∗
n=1 is a boundedly complete basis
n). It follows by Theorem 1.4.6 that (dn) is a shrinking basis for RBD.
for (cid:96)1 so also is (d∗
Consequently by the Bourgain-Delbaen Theorem (Theorem 2.2.2), the dual space of RBD
is naturally isomorphic to (cid:96)1, i.e. the (cid:96)1 duality obtained here is the same as that which
defines the ιF mapping discussed earlier.
n) ⊆ (cid:96)1(N0). Since (e∗
5.3 Concluding Remarks
We have obtained a right-shift invariant (cid:96)1 predual using the Bourgain-Delbaen construction
which has a striking resemblance to the space F2 constructed in [9]. However, the space
R2 in the above lemma does not quite do what we want it to. Tracing through the proof
of Lemma 5.1.4, we see that for the natural convolution on (cid:96)1(N0) to be weak* continuous
133
with respect to the topology induced by the natural isomorphism of (cid:96)1 with R∗
need R2 to be left shift invariant. Alas, our shift is in the wrong direction!
2, we would
At the time of writing, it is unclear to the author whether or not it is possible to modify
the construction in some way so as to obtain a one-sided version of F2 that is left-shift
invariant as a subspace of (cid:96)∞(N0). Less clear still, is whether or not it is possible to obtain
2-sided (bilateral) versions of these spaces via the Bourgain-Delbaen construction; indeed,
the recursive nature of the Bourgain-Delbaen construction makes the one-sided object a
much more natural object to obtain.
We make one final remark on shift-invariant preduals in relation to the Bourgain-Delbaen
construction. We have seen in Chapters 3 and 4 that the Bourgain-Delbaen construction is
much more naturally understood with respect to a countable set Γ whereby elements of Γ
code the form of the corresponding c∗ vector. Of course, when working with shift invariant
preduals of (cid:96)1, one needs to make some enumeration of Γ so that the shift operator can
be defined. This provides yet another obstruction to finding exotic (cid:96)1 preduals via the
Bourgain-Delbaen construction.
Whilst the work contained in this thesis has allowed us to gain an insight into the
structure of Calkin algebras, we note that there are a number of open problems that arise
from the work here. Indeed, it is unclear exactly what Banach algebras can be realised as a
Calkin algebra. If it is too much to hope to solve this problem, a useful partial result might
be to obtain some kind of obstruction that would prevent a Banach algebra being obtained
as a Calkin algebra.
134
References
[1] F. Albiac and N. J. Kalton, Topics in banach space theory, Springer, 2006.
[2] S.A Argyros and R.G Haydon, A Hereditarily Indecomposable L∞ space that solves the
scalar-plus-compact problem, Acta Math. 206 (2011), no. 1, 1 -- 54.
[3] N. Aronszajn and K.T. Smith, Invariant subspaces of completely continuous operators,
Ann. of Math. 60 (1954), no. 2, 345 -- 350.
[4] K. Beanland and L. Mitchell, Operators on the L∞-spaces of Bourgain and Delbaen,
Quaestiones Mathematicae 33 (2010), 443 -- 448.
[5] Y. Benyamini and J. Lindenstrauss, A predual of (cid:96)1 which is not isomorphic to a C(K)
space, Israel J. Math. 13 (1972), 246 -- 254.
[6] J. Bourgain, New classes of Lp-spaces, Lecture Notes in Math. 889, Springer, 1981.
[7] J. Bourgain and F. Delbaen, A class of special L∞-spaces, Acta Math. 145 (1980),
155 -- 176.
[8] J. W. Calkin, Two-sided ideals and congruences in the ring of bounded operators in
hilbert space, Ann. of Math. 42 (1941), no. 2, 839 -- 873.
[9] M. Daws, R. Haydon, Th. Schlumprecht, and S. White, Shift invariant preduals of
(cid:96)1(Z), preprint, arXiv:1101.5696.
[10] N. Dunford and J.T. Schwartz, Linear Operators. Part 1: General Theory, Wiley-
Blackwell, 1988.
[11] M. Fabian, P. Habala, P. Hajek, V.M. Santalucia, J. Pelant, and V. Zizler, Functional
Analysis and Infinite-Dimensional Geometry, Springer, 2001.
[12] V. Ferenczi, Hereditarily finitely decomposable Banach spaces, Studia. Math. 123
(1997), no. 2, 135 -- 149.
135
[13] D. Freeman, E. Odell, and Th. Schlumprecht, The universality of (cid:96)1 as a dual space,
Mathematische Annalen 351 (2011), no. 1, 149 -- 186.
[14] C. Gohberg, A.S. Markus, and I.A. Fel'dman, Normally solvable operators and ideas
associated with them, Bul. Akad. Stiince RSS Moldoven. 76 (1960), no. 10, 51 -- 70,
English translation: Amer. Math. Soc. Transl. 61 (1967), 63-84.
[15] W. Gowers and B. Maurey, Banach Spaces with Small Spaces of Operators, Mathema-
tische Annalen 307 (1997), no. 4, 543 -- 568.
[16] W.T. Gowers and B. Maurey, The unconditional basic sequence problem, J. Amer.
Math. Soc. 6 (1993), 851 -- 874.
[17] R. Haydon, Subspaces of the Bourgain Delbaen space, Studia Math. 139 (2000), 275 --
293.
[18] W.B. Johnson, D. Chen, and B. Zheng, Commutators on ((cid:80) (cid:96)q)p, Studia Math., to
appear.
[19] S. Kakutani, A generalisation of Brouwer's fixed point theorem, Duke Math. J. 7 (1941),
457 -- 459.
[20] N. J. Laustsen, R.J. Loy, and C.J. Read, The lattice of closed ideals in the Banach
algebra of operators on certain Banach spaces, J. Funct. Anal. 214 (2004), 106 -- 131.
[21] N.J. Laustsen, Th. Schlumprecht, and A. Zs´ak, The lattice of closed ideals in the
Banach algebra of operators on a certain dual Banach space, J. Operator Theory 56
(2006), 391 -- 402.
[22] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I and II, Springer, 1996.
[23] V.I. Lomonosov, Invariant subspaces of the family of operators that commute with a
completely continuous operator, Akademija Nauk SSSR. Funkcional' nyi Analiz iego
Prilozenija 7 (1973), 55 -- 56.
[24] B. Maurey, Operator Theory and Eexotic Banach Spaces, Online Lecture Notes,
http://www.math.jussieu.fr/ maurey/articles/csp.pdf, 1994.
[25] R. E. Megginson, An Introduction to Banach Space Theory, Springer, 1998.
[26] G.J Murphy, C∗-Algebras and Operator Theory, Academic Press, 1990.
[27] Th. Schlumprecht, N. Tomczak-Jaegermann, and V.G. Troitsky, On norm closed ideals
in L((cid:96)p, (cid:96)q), Studia. Math. 179 (2007), no. 3, 239 -- 262.
136
[28] M. Tarbard, Hereditarily Indecomposable, separable L∞ Banach spaces with (cid:96)1 dual
having few but not very few operators, J. London. Math. Soc 85 (2012), no. 3, 737 -- 764.
137
|
1605.02873 | 1 | 1605 | 2016-05-10T07:10:28 | Recent Progress in Shearlet Theory: Systematic Construction of Shearlet Dilation Groups, Characterization of Wavefront Sets, and New Embeddings | [
"math.FA"
] | The class of generalized shearlet dilation groups has recently been developed to allow the unified treatment of various shearlet groups and associated shearlet transforms that had previously been studied on a case-by-case basis. We consider several aspects of these groups: First, their systematic construction from associative algebras, secondly, their suitability for the characterization of wavefront sets, and finally, the question of constructing embeddings into the symplectic group in a way that intertwines the quasi-regular representation with the metaplectic one. For all questions, it is possible to treat the full class of generalized shearlet groups in a comprehensive and unified way, thus generalizing known results to an infinity of new cases. Our presentation emphasizes the interplay between the algebraic structure underlying the construction of the shearlet dilation groups, the geometric properties of the dual action, and the analytic properties of the associated shearlet transforms. | math.FA | math |
RECENT PROGRESS IN SHEARLET THEORY: SYSTEMATIC
CONSTRUCTION OF SHEARLET DILATION GROUPS,
CHARACTERIZATION OF WAVEFRONT SETS, AND NEW
EMBEDDINGS
GIOVANNI S. ALBERTI, STEPHAN DAHLKE, FILIPPO DE MARI, ERNESTO DE VITO,
AND HARTMUT F UHR
Abstract. The class of generalized shearlet dilation groups has recently been
developed to allow the unified treatment of various shearlet groups and asso-
ciated shearlet transforms that had previously been studied on a case-by-case
basis. We consider several aspects of these groups: First, their systematic con-
struction from associative algebras, secondly, their suitability for the character-
ization of wavefront sets, and finally, the question of constructing embeddings
into the symplectic group in a way that intertwines the quasi-regular represen-
tation with the metaplectic one. For all questions, it is possible to treat the
full class of generalized shearlet groups in a comprehensive and unified way,
thus generalizing known results to an infinity of new cases. Our presentation
emphasizes the interplay between the algebraic structure underlying the con-
struction of the shearlet dilation groups, the geometric properties of the dual
action, and the analytic properties of the associated shearlet transforms.
1. Introduction
This chapter is concerned with several important aspects of modern signal anal-
ysis. Usually, signals are modeled as elements of function spaces such as L2 or
Sobolev spaces. To analyze such a signal and to extract the information of interest
from it, the first step is always to decompose the signal into suitable building blocks.
This is performed by transformation, i.e., the signal is mapped into function spaces
on an underlying parameter set, and then the signal is processed and analyzed by
studying and modifying the resulting coefficients. By now, a whole zoo of suitable
transforms have been introduced and analyzed such as the Fourier transform, the
Gabor transform, or the wavelet transform, just to name a few. Some of them have
already been very succesful, e.g., the Fourier transform works excellently for sig-
nals that are well-localized in the frequency domain, whereas wavelets are often the
method of choice for the analysis of piecewise smooth signals with well-localized sin-
gularities such as edges in an image. Which transform to choose obviously depends
on the application, i.e., on the type of information one wants to detect from the
signal. However, in recent years, it has turned out that a serious bottleneck still has
to be removed. Most of the classical transforms such as the wavelet transform per-
form suboptimally when it comes to the detection of directional information. The
reason is very simple: most of these transforms are essentially isotropic, whereas
directional information is of anisotropic nature. This observation triggered many
Date: May 9, 2016.
2010 Mathematics Subject Classification. 42C15, 42C40, 46F12, 22D10.
1
2
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
innovative studies how to design new building blocks that are particularly tuned to
this problem, such as curvelets [4], contourlets [16], ridgelets [3] and many others.
In this chapter, we are in particular interested in one specific contribution to this
problem, i.e., the shearlet approach. Shearlets are new affine representation sys-
tems that are based on translations, shearings, and anisotropic dilations. We refer
to the monograph [37] for an overview. Among all the new approaches, the shearlet
transform stands out for the following reason: the continuous shearlet transform
can be derived from a square-integrable representation of a specific group, the full
shearlet group [9, 10, 13]. This property is not only of academic interest but has the
important consequence that the whole powerful machinery derived in the realm of
square-integrable group representations such as reproducing kernels, inversion for-
mulas etc. can directly be employed. This feature of the shearlet transform clearly
has strengthened the interest in the shearlet theory, and many important results
concerning the group-theoretical background have been derived so far. It is the aim
of this chapter to push forward, to clarify and to unify this theory with respect to
several important aspects. Our main objectives can be described as follows.
After the full shearlet group has been discovered, the question arose if other
suitable concepts of shearlet groups could be constructed. A first example was the
shearlet Toeplitz group [14], where the shearing part of the group has a Toeplitz
structure. As we will see later in Subsection 3.3 of this chapter, the full shearlet
group and the shearlet Toeplitz group are in a certain sense the "extreme" cases of
a general construction principle. In this view, the full shearlet group corresponds to
the nilpotency class n = 2, whereas the Toeplitz case corresponds to the nilpotency
class n = d, where d denotes the space dimension. Therefore, one would conjecture
that there should be a lot of examples "in between". Indeed, in [30], a positive
answer has been given, and a first classification of low-dimensional shearlet groups
has been derived. In this chapter, we further extend these results and present an
approach to the systematic construction of suitable shearlet groups. The starting
point is a general class of shearlet groups introduced in [30]. We say that a dilation
group H is a shearlet group if every h ∈ H can be written as h = ±ds, d ∈ D, s ∈ S
where D is a diagonal scaling subgroup and S denotes a connected, closed abelian
matrix group, the shearing subgroup. The key to understanding and constructing
shearing subgroups lies in the realization that their associated Lie algebras carry a
very useful associative structure. This associative structure also greatly facilitates
the task of identifying the diagonal scaling groups compatible with a given shearing
subgroup. Through the notion of Jordan-Holder bases the problem of characterizing
all suitable scaling group generators translates to a rather simple linear system of
equations.
It turns out that all examples known so far are special cases of this
general construction.
In recent studies, it has also been observed that shearlets provide a very powerful
tool in microlocal analysis [32], e.g., to determine the local regularity of a function.
In the one-dimensional case, pointwise smoothness can very efficiently be detected
by examining the decay of the continuous wavelet transform as the scale parameter a
tends to zero [34]. In the multivariate setting, pointwise smoothness does not cover
all the geometic information one might be interested in. E.g., if the function under
consideration exhibits singularities, one usually wants to know in which direction
the function is singular. This can be described by the so-called wavefront set of
a distribution.
It has turned out that the continuous shearlet transform can be
RECENT PROGRESS IN SHEARLET THEORY
3
employed to detect this wavefront set, once again by studying its decay as the
scaling parameter tends to zero. This property has been first observed in [36],
In [22], this concept has been generalized
we also refer to [32] for an overview.
to much more general classes of dilation groups.
It has been shown that under
natural assumptions, a wavefront set can again be detected by the decay of the voice
transform. Essentially, two fundamental conditions are needed, that are related
with the dual action of the dilation group H: the dual action must be microlocally
admissible in direction ξ and it must satisfy the V -cone approximation property at
ξ, see Section 4.1 for the precise definitions. If these properties hold for one point
ξ0 in the open dual orbit, a characterization of wavefront sets is possible. In this
chapter, we show that both properties are satisfied for our general construction of
shearlet dilation groups, provided that the infinitesimal generator Y of the scaling
subgroup satisfies Y = diag(1, λ2, . . . , λd), 0 < λi < 1, 2 ≤ i ≤ d. Consequently,
characterizations of wavefront sets are possible for a huge subclass of our general
construction. It is worth mentioning that anisotropic dilations are necessary for the
detection of wavefront sets, in particular the classical (isotropic) continuous wavelet
transform would not do the job.
A third important issue we will be concerned with in this chapter is the relations
of our general shearlet groups to other classical groups, in particular to the symplec-
tic groups Sp(d, R). The symplectic groups are one of the most important classical
groups, because they play a prominent role in classical mechanics. We therefore
investigate to which extent our shearlet dilation groups can be embedded into sym-
plectic groups, in a way that intertwines the quasi-regular representation with the
metaplectic representation. For the full shearlet groups and the shearlet Toeplitz
groups, these issues have been studied in [6], see also [35]. Their connected com-
ponents can indeed be embedded into the symplectic groups, which yields group
isomorphisms of the positive parts of shearlet groups with the so -- called TDS(d)
subgroups that have already been studied in [15]. In this chapter, we generalize
this result to dilation groups of the form G = Rd ⋊ H, where H is a subgroup of
T (d, R)+ = {h ∈ GL(d, R) : h1,1 > 0 and hi,j = 0 for every i > j}. We show that
for any such group there exists a group embedding φ : G → Sp(d, R), and that
its quasi-regular representation is unitarily equivalent to µ ◦ φ, where µ denotes
the metaplectic representation of Sp(d, R). Since the positive part of any shearlet
group falls into this general category, the desired embeddings for shearlet groups
follow from this result. Let us also mention the following very interesting fact: for
the full shearlet dilation groups, such embeddings are never possible. Indeed, in [6]
it has been shown that already for the two-dimensional full shearlet group there
does not exist an injective continuous homomorphism into Sp(2, R) or into any of
its coverings.
Let us also mention a nice by-product of our construction.
In recent studies
[8, 10, 11, 13], an interesting relation of the shearlet approach to the coorbit theory
derived by Feichtinger and Grochenig [17, 18, 19, 20] has been established. Based on
a square integrable group representation, coorbit space theory gives rise to canonical
associated smoothness spaces, where smoothness is measured by the decay of the
underlying voice transform.
In [8, 10, 11, 13], it has been shown that all the
conditions needed in the coorbit setting can be satisfied for the full shearlet and
the shearlet Toeplitz groups. In [27], the coorbit approach has been extended to
much more general classes of dilation groups, and it turns out that the analysis
4
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
from [27] also carries over to the construction presented in this chapter, so that
we obtain many new examples of coorbit spaces. In particular, we refer to [30] for
explicit criteria for compactly supported functions that can serve as atoms in the
coorbit scheme.
This chapter is organized as follows. In Sections 2 and 3, we present our construc-
tion of generalized shearlet dilation groups. After discussing the basic notations and
definitions in the Subsections 2.1 and 2.2, in Subsection 2.3 we start with the sys-
tematic investigation of the Lie algebras of shearing subgroups. One of the main
results is Lemma 6 which provides a complete description of a shearing subgroup
in terms of the canonical basis of its Lie algebra. This fact can be used to derive
linear systems whose nonzero solutions determine the anisotropic scaling subgroups
that are compatible with S (Lemma 9). These relationships are then used in Sec-
tion 3 to derive a systematic construction principle. The canonical basis can be
directly computed from the structure constants of a Jordan-Holder basis (Lemma
13). The power of this approach is demonstrated by several examples. In Section 4,
we study the suitability of shearlet dilation groups for the characterization of wave-
front sets. Here the main result is Theorem 28 which shows that shearlet groups
with anisotropic dilations and suitable infinitesimal generators for the scaling sub-
groups do the job. The proof is performed by verifying the basic conditions from
[22]. The last section is concerned with the embeddings of shearing dilation groups
into symplectic groups. The main result of this section is Theorem 33 which shows
that the huge class of semidirect products of the form G = Rd ⋊ H, where H is a
subgroup of T (d, R)+ can be embedded into Sp(d, R).
2. Generalities on shearlet dilation groups
2.1. Basic notations and definitions. This chapter is concerned with the con-
struction and analysis of large classes of generalized shearlet transforms. These
transforms are constructed by fixing a suitable matrix group, the so-called shearlet
dilation group. By construction, these groups have a naturally associated isometric
continuous wavelet transform, which will be the generalized shearlet transform. In
this subsection, we summarize the necessary notation related to general continu-
ous wavelet transforms in higher dimensions. We let GL(d, R) denote the group of
invertible d × d-matrices. We use Id to denote the d × d identity matrix. The Lie
algebra of GL(d, R) is denoted by gl(d, R), which is the space of all d × d matrices,
endowed with the Lie bracket [X, Y ] = XY − Y X. Given h ∈ gl(d, R) its (operator)
norm is denoted by
khk = sup
x≤1hx.
We let exp : gl(d, R) → GL(d, R) denote the exponential map, defined by
exp(X) =
X k
k!
∞Xk=0
known to converge absolutely for every matrix X. Given a closed subgroup
H < GL(d, R), the associated Lie algebra of H is denoted by h, and it is de-
fined as tangent space of H at Id, or, equivalently, as the set of all matrices X with
exp(RX) ⊂ H. It is a Lie subalgebra of gl(d, R), i.e., it is closed under taking Lie
brackets.
RECENT PROGRESS IN SHEARLET THEORY
5
A matrix group of particular importance for the following is the group T (d, R)
of upper triangular matrices with ones on the diagonal. Elements of T (d, R) are
called unipotent. Its Lie algebra is the subspace t(d, R) ⊂ gl(d, R) of all strictly
It is well-known that exp : t(d, R) → T (d, R) is a
upper triangular matrices.
homeomorphism [33]. In particular, whenever s ⊂ t(d, R) is a Lie subalgebra, the
exponential image exp(s) is a closed, simply connected and connected matrix group
with Lie algebra given by s. Conversely, any connected Lie subgroup S of T (d, R)
is closed, simply connected and S = exp(s) where s ⊂ t(d, R) is the corresponding
Lie algebra, see Theorem 3.6.2 of [39].
For the definition of generalized wavelet transforms, we fix a closed matrix group
H < GL(d, R), the so-called dilation group, and let G = Rd ⋊ H. This is the
group of affine mappings generated by H and all translations. Elements of G
are denoted by pairs (x, h) ∈ Rd × H, and the product of two group elements
is given by (x, h)(y, g) = (x + hy, hg). The left Haar measure of G is given by
dµG(x, h) = det(h)−1dxdh, where dx and dh are the Lebesgue measure and the
(left) Haar measure of Rd and H, respectively.
The group G acts unitarily on L2(Rd) by the quasi-regular representation defined
by
(1)
[π(x, h)f ](y) = det(h)−1/2f(cid:0)h−1(y − x)(cid:1) .
We assume that H is chosen irreducibly admissible, i.e. such that π is an (irre-
ducible) square-integrable representation. Recall that a representation is irreducible
if the only invariant closed subspaces of the representation space are the trivial
ones. Square-integrability of the representation means that there exists at least one
nonzero admissible vector ψ ∈ L2(Rd) such that the matrix coefficient
is in L2(G), which is the L2-space associated to the left Haar measure dµG. In this
case the associated wavelet transform
(x, h) 7→ hψ, π(x, h)ψi
(2)
Wψ : L2(Rd) ∋ f 7→ ((x, h) 7→ hf, π(x, h)ψi) ∈ L2(G)
is a scalar multiple of an isometry, which gives rise to the wavelet inversion formula
(3)
f =
Wψf (x, h)π(x, h)ψ dµG(x, h) ,
1
cψZG
where the integral is in the weak sense.
We note that the definition of Wψf also makes sense for tempered distributions
f , as soon as the wavelet ψ is chosen as a Schwartz function and the L2-scalar
product is properly extended to a sesquilinear map S′ × S → C. Analogs of the
wavelet inversion formula are not readily available in this general setting, but it will
be seen below that the transform has its uses, for example in the characterization
of wavefront sets.
Most relevant properties of the wavelet transform are in some way or another
connected to the dual action, i.e., the (right) linear action Rd × H ∋ (ξ, h) 7→ hT ξ.
For example, H is irreducibly admissible if and only if the dual action has a single
open orbit O = {hT ξ0 : h ∈ H} ⊂ Rd of full measure (for some ξ0 ∈ O), such
that in addition the stabilizer group Hξ0 = {h ∈ H : hT ξ0 = ξ0} is compact [28].
This condition does of course not depend on the precise choice of ξ0 ∈ O. The dual
action will also be of central importance to this chapter.
6
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
2.2. Shearlet dilation groups. The original shearlet dilation group was intro-
duced in [9, 10], as
H =(cid:26)±(cid:18) a
0 a1/2 (cid:19) : a > 0, b ∈ R(cid:27) .
b
The rationale behind this choice was that the anisotropic scaling, as prescribed
by the exponents 1, 1/2 on the diagonal, combines with the shearing (controlled
by the parameter b) to provide a system of generalized wavelets that are able to
swiftly adapt to edges of all orientations (except one). A mathematically rigourous
formulation of this property is the result, due to Kutyniok and Labate, that the
continuous shearlet transform characterizes the wavefront set [36]. Approximation-
theoretic properties of a different, more global kind were the subject of the chapter
[10] describing the so-called coorbit spaces defined in terms of weighted integrability
conditions on the wavelet coefficients.
The original shearlet dilation group has since been generalized to higher dimen-
sions. Here, the initial construction was introduced in [13], and further studied,
e.g., in [11, 5]. It is a matrix group in dimension d ≥ 3 defined by
S =
±
a
s1
aλ2
. . .
sd−1
. . .
aλd
: a > 0, s1, . . . , sd−1 ∈ R
.
Here λ2, . . . , λd are positive exponents, often chosen as λ2 = . . . = λd = 1/2. It
should, however, be noted that they can be chosen essentially arbitrarily (even
negative), without affecting the wavelet inversion formula. Coorbit space theory is
applicable to all these groups as well [11, 27]. Furthermore, it was recently shown
that the associated shearlet transform also characterizes the wavefront set [22], as
long as the exponents λ2, . . . , λd are strictly between zero and one.
A second, fundamentally different class of shearlet groups are the Toeplitz shear-
let groups introduced in [14] and further studied in [8]. These groups are given
by
(5)
(4)
: a > 0, s1, . . . , sd−1 ∈ R
.
H =
±
a
s1
a
s2
s1
. . .
. . .
s2
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
sd−1
sd−2
...
...
s2
s1
a
Coorbit space theory can be applied to these groups as well [8, 27]. By [22, Lemma
4.10], the fact that H contains nontrivial multiples of the identity implies that H
does not characterize the wavefront set. However, it will be shown below that by
properly adjusting the diagonal entries, it is possible to construct a closely related
group H ′ that does lend itself to the characterization of the wavefront set.
A closer inspection of the two higher-dimensional families of shearlet group re-
veals several common traits: fix one of the above-listed groups H. Then each h ∈ H
RECENT PROGRESS IN SHEARLET THEORY
7
factors as
h = ±diag(a, aλ2, . . . , aλd ) · u
where the first factor denotes the diagonal matrix with the same diagonal entries
as h, and the second factor u is unipotent. In fact, this factorization is necessarily
unique. Furthermore, denoting by D the set of all diagonal matrices occurring in
such factorizations, and by S the set of all unipotent ones that arise, it is easy to
see that D (and consequently S) are closed subgroups of H. Finally, one readily
verifies that the groups S that occur in the examples are in fact commutative. We
will now use these properties to define a general class of shearlet dilation groups,
that we will study in this chapter:
Definition 1. Let H < GL(d, R) denote an irreducibly admissible dilation group.
H is called generalized shearlet dilation group, if there exist two closed subgroups
S, D < H with the following properties:
matrix;
(i) S is a connected abelian Lie subgroup of T (d, R);
(ii) D = {exp(rY ) : r ∈ R} is a one-parameter group, where Y is a diagonal
(iii) Every h ∈ H can be written uniquely as h = ±ds, with d ∈ D and s ∈ S.
S is called the shearing subgroup of H, and D is called the diagonal complement or
scaling subgroup of H.
Remark 2. As noted in Subsection 2.1, S is closed, simply connected and the
exponential map is a diffeomorphism from its Lie algebra s onto S.
Remark 3. The class of shearlet dilation groups was initially defined in [30], and
for some of the following results and observations, more detailed proofs can be
found in that paper. In particular, it was shown there that coorbit space theory
applies to all generalized shearlet dilation groups. In fact, it is possible to construct
wavelet frames with compactly supported atoms, with frame expansions that, de-
pending on the provenance of the signal, converge in a variety of coorbit space
norms simultaneously.
As will be seen below, shearlet dilation groups can be constructed systematically.
The natural order in finding the constituent subgroups S, D is to first pick a can-
didate for S, and then determine the infinitesimal generators of the one-parameter
group D that are compatible with S. The details of this programme are given in
the next subsections.
2.3. Shearlet dilation groups and their Lie algebras. It is the aim of this sub-
section to give an overview of the most important structural properties of shearlet
dilation groups. The following proposition gives a first characterization of these
groups, see [30, Proposition 4.3].
Proposition 4. Let S denote a connected abelian subgroup of T (d, R). Then the
following are equivalent:
(i) S is the shearing subgroup of a generalized shearlet dilation group;
(ii) There is ξ ∈ Rd such that S acts freely on ST ξ via the dual action, and in
(iii) The matrix group A = {rs : s ∈ S, r ∈ R×} is an abelian irreducibly
addition, dim(S) = d − 1;
admissible dilation group. It is also a shearlet dilation group.
8
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
The fundamental observation made in [26, Remark 9] is that if A is abelian and
admissible, as in part (iii) of the above proposition, then its Lie algebra a is in
fact an associative subalgebra containing the identity element, hence it is closed
under matrix multiplication. This associative structure is in many ways decisive.
To begin with, one has the relations
a = span(A) , A = a×
i.e., A consists precisely of the multiplicatively invertible elements of the associative
algebra a. We will see in Subsection 3.1 below that this connection to associative
algebras can be used for the systematic -- even exhaustive -- construction of shearing
subgroups.
There is however a second ingredient, that is more directly related to the prop-
erties of the dual action. It is described in the following lemma, see [30, Corollary
4.7]. We use e1, . . . , ed for the canonical basis of Rd.
Lemma 5. Let S denote a connected abelian subgroup of T (d, R) of dimension
d − 1, with Lie algebra s. Then the following are equivalent:
(i) S is a shearing subgroup;
(ii) There exists a unique basis X2, . . . , Xd of s with X T
i e1 = ei, for all i =
2, . . . , d.
We call the basis from part (ii) the canonical basis of s.
The canonical basis plays a special role for the description of shearing subgroups.
As a first indication of its usefulness, we note that all off-diagonal entries of the
elements of shearing groups depend linearly on the entries in the first row.
Lemma 6. Let S denote a shearing subgroup with Lie algebra s, and canonical
basis X2, . . . , Xd of s. Then the following holds:
(a) S = {Id + X : X ∈ s}.
(b) Let h ∈ S be written as
h =
Then
1 h1,2
0
1
0
0
0
0
0
0
0
0
. . .
h2,3
. . .
0
0
0
. . .
. . .
. . .
. . .
0
0
. . .
. . .
...
. . .
1
0
h1,d
h2,d
...
...
hd−1,d
1
h = Id +
h1,iXi .
dXi=2
.
Proof. For part (a), denote the right-hand side by S1. Since s is an associative
subalgebra consisting of nilpotent matrices, S1 consists of invertible matrices, and
it is closed under multiplication. Furthermore, the inverse of any element of S1 can
be computed by a Neumann series that breaks off after at most d terms:
(Id + X)−1 = Id +
d−1Xk=2
(−1)k−1X k ,
RECENT PROGRESS IN SHEARLET THEORY
9
and the result is again in S1. Hence S1 is a matrix group. It is obviously closed and
connected, with tangent space of S1 at the identity matrix given by s. It follows
that S1 is a Lie subgroup of T (d, R) and, hence, it is simply connected. Thus S
and S1 are closed, connected, and simply connected subgroups sharing the same
Lie algebra, hence they are equal. Now part (b) directly follows from (a) and the
properties of the canonical basis.
(cid:3)
We now turn to the question of characterizing the scaling subgroups D that are
compatible with a given shearing subgroup S. It is convenient to describe D in terms
of its Lie algebra as well. Since D is one-dimensional, we have D = exp(RY ), with
a diagonal matrix Y = diag(λ1, λ2, . . . , λd). We then have the following criterion
[30, Proposition 4.5]:
Proposition 7. Let S < GL(d, R) denote a shearing subgroup. Let Y denote
a nonzero diagonal matrix, and let D := exp(RY ) the associated one-parameter
group with infinitesimal generator Y . Then the following are equivalent:
(i) H = DS ∪ (−DS) is a shearlet dilation group;
(ii) For all X ∈ s we have [X, Y ] = XY − Y X ∈ s, and in addition the first
diagonal entry of Y is nonzero.
Remark 8. The above proposition states that H = S ⋊ R×, so that H is solvable
group with two connected components, and each of them is simply connected.
Since Y and rY , for nonzero r ∈ R, determine the same one-parameter subgroup,
part (ii) of the proposition allows to fix λ1 = 1. Note that part (ii) is trivially
fulfilled by isotropic scaling, which corresponds to taking 1 = λ1 = λ2 = . . . = λd.
In what follows, we will be particularly interested in anisotropic solutions; our
interest in these groups is mainly prompted by the crucial role of anisotropic scaling
for wavefront set characterization.
It turns out that the relation [Y, s] ⊂ s translates to a fairly transparent system
of linear equations. Once again, the canonical basis X2, . . . , Xd of s proves to be
particularly useful: As the following lemma shows, the adjoint action s ∋ X 7→
[Y, X] maps s into itself if and only if the Xi are eigenvectors of that map. The
lemma uses the notation Ei,j for the matrix having entry one at row i and column
j, and zeros everywhere else.
Lemma 9. Let s denote the Lie algebra of a shearing subgroup, and let X2, . . . , Xd
denote the canonical basis of s, given by
(6)
Xi = E1,i +
dXj=2
dXk=j+1
di,j,kEj,k
with suitable coefficients di,j,k. Let Y = diag(1, λ2, . . . , λd) be given. Then [Y, s] ⊂ s
if and only if
for all i = 2, . . . , d : λi = 1 + µi ,
and the vector (µ2, . . . , µd) is a solution of the system of linear equations given by
for all (i, j, k) ∈ {2, . . . , d}3 with di,j,k 6= 0 : µi + µj = µk .
(7)
In particular, (µ2, . . . , µd) 7→ (1, 1 + µ2, . . . , 1 + µd) sets up a bijection between the
nonzero solutions of (7) on the one hand and the anisotropic scaling subgroups D
compatible with S on the other.
10
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
Remark 10. Note that (6) shows that di,j,k = (Xi)jk.
Proof. We first note that the Ej,k are eigenvectors under the adjoint action of any
diagonal matrix:
(8)
[Y, Ej,k] = (λj − λk)Ej,k .
As a consequence, given any matrix X, the support of the matrix [Y, X] (i.e., the
set of indices of its nonzero entries) is contained in the support of X.
Note that Y normalizes s if and only if [Y, Xi] ∈ s for i = 2, . . . , d. Now the
calculation
(9) [Y, Xi] = [Y, E1,i] +X(j,k)
di,j,k[Y, Ej,k] = (1 − λi)E1,i +X(j,k)
di,j,k(λj − λk)Ej,k
shows that the only (potentially) nonzero entry in the first row of [Y, Xi] occurs at
the ith column, hence [Y, Xi] is in s if and only if it is a scalar multiple of Xi. In
view of (9) and the linear independence of the Ej,k, this holds precisely when
for all (i, j, k) ∈ {2, . . . , d}3 with di,j,k 6= 0 : 1 − λi = λj − λk .
(10)
Rewriting this system for µi = λi − 1, for i = 2, . . . , d, yields (7).
(cid:3)
Finally, let us return to properties of the associated shearlet transforms. In view
of the central role of the dual action, it is important to compute the associated
open dual orbit. Here we have the following, see [30, Proposition 4.5]:
Proposition 11. Let S be a shearing subgroup, and D any diagonal complement
of S. Then H = DS ∪ −DS acts freely on the unique open dual orbit given by
O = R× × Rd−1.
Note that the dual orbit is the same for all shearing groups. Somewhat surpris-
ingly, the same can be said of the admissibility condition [30, Theorem 4.12]:
Theorem 12. Let H < GL(Rd) denote a generalized shearlet dilation group. Then
ψ ∈ L2(Rd) is admissible iff
ZRd
bψ(ξ)2
ξ1d dξ < ∞ .
3. A construction method for shearlet dilation groups
3.1. Constructing shearing subgroups. In this subsection we want to describe
a general method for the systematic construction of shearing subgroups. Recall that
given a shearing subgroup S with Lie algebra s, taking the Lie algebra a = RId ⊕ s
and its associated closed matrix group A results in an abelian irreducibly admissible
matrix group. Following [26], this entails that a is an associative matrix algebra.
Furthermore, note that s consists of strictly upper triangular matrices, which entails
that any product of d elements of s vanishes.
These features of a can be described in general algebraic terms. Given a finite-
dimensional, associative commutative algebra A, we call an element a ∈ A nilpotent
if there exists n ∈ N such that an = 0. The set of all nilpotent elements in A is
called the nilradical of A, denoted by N . We call A nilpotent if every element of A
is nilpotent. N is an ideal in A, i.e., given a ∈ N and an arbitrary b ∈ A, one has
(ab)n = anbn = 0 for sufficiently large n, i.e. ab is again in the nilradical. We call
the algebra A irreducible (over R) if it has a unit element 1A satisfying 1Ab = b for
RECENT PROGRESS IN SHEARLET THEORY
11
all b ∈ A, and such that A = R · 1A ⊕ N holds. Note that N determines A in this
case, and we will freely switch between A and N in the following.
Now the above considerations show that a is an irreducible associative commu-
tative algebra. In the remainder of this subsection, we will be concerned with a
converse to this statement, i.e., with the construction of shearing subgroups from
an abstractly given irreducible associative algebra. Assume that A is an irreducible
commutative associative algebra of dimension d, and denote its nilradical by N .
We let
n(A) = min{k ∈ N : ak = 0, ∀a ∈ N} ,
which is called the nilpotency class of A. Letting
N k = {a1 . . . ak : ai ∈ N},
for k ≥ 1, and N 0 = N , one can prove that
n(A) = min{k ∈ N : N k = {0}} ≤ d .
By definition of the nilpotency class, we obtain that N n(A)−1 6= {0}, and for all
a ∈ N n(A)−1 and b ∈ N , it follows that ab = 0.
Hence, choosing a nonzero ad ∈ N n(A)−1, we find that Id := R · ad is an ideal in
N ; in fact, we get NId = {0}. Applying the same reasoning to the algebra N /Id
(and choosing any representative modulo Id) produces a second element ad−1 with
the property that Id−1 = span(ad−1, ad) fulfills NId−1 ⊂ Id. Further repetitions of
this argument finally yield a basis a2, . . . , ad of N , that we supplement by a1 = 1A
to obtain a basis of A with the property
(11)
NIk ⊂ Ik+1 for 1 ≤ k < d ,
and I2 = N . We call a basis a2, . . . , ad of N satisfying condition (11) a Jordan-
Holder basis of N .
The existence of a Jordan-Holder basis can be also proved by referring to a
general result about nilpotent representations of nilpotent algebras, see Theorem
3.5.3 of [39]. Indeed, regard N as nilpotent algebra and A as a vector space. It is
easy to check that the (regular) representation ρ of the Lie algebra N acting on A
as ρ(a)b = ab is nilpotent, so that there exists a basis {a1, . . . , ad} of A such that for
each a ∈ N , the endomorphism ρ(a) is represented by a strictly upper triangular
matrix Ψ(a) ∈ gl(d, R) according to the canonical isomorphism
ρ(a)aj =
dXk=1
Ψ(a)j,k ak
j = 1, . . . , d .
Since ρ(a)1A = a, it is always possible to choose a1 = 1A and, by construction, for
all a ∈ N and for i = 1, . . . , d − 1
ρ(a)span{ai, . . . , ad} ⊂ span{ai+1, . . . , ad}
ρ(a)ad = 0 .
These bases provide accesss to an explicit construction of an associated shearing
subgroup, explained in detail in the next lemma. Recall the notation Ei,j for the
matrix possessing entry one in row i, column j, and zeros elsewhere. Note that the
map Ψ : A → gl(d, R) in the following lemma coincides with the identically denoted
map that we just introduced.
12
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
Lemma 13. Let A denote an irreducible commutative associative algebra of di-
mension d with nilradical N possessing the Jordan-Holder basis
Let a1 = 1A, and let ψ : Rd → A denote the induced linear isomorphism
a2, . . . , ad ∈ A .
ψ((x1, . . . , xd)T ) =
xiai .
dXi=1
Let Ψ : A → gl(d, R) denote the associated linear map satisfying for all a ∈ A and
for all x ∈ Rd:
ψ−1(a · ψ(x)) = Ψ(a) · x .
(a) The set
is a shearing subgroup, with associated Lie algebra given by
S = {Id + Ψ(a)T : a ∈ N}
s = {Ψ(a)T : a ∈ N} .
(b) Defining Xi = Ψ(ai)T , for i = 1, . . . , d, we get that X1 is the identity
matrix, and X2, . . . , Xd is the canonical basis of s in the sense of Lemma
5.
(c) Let (di,j,k)1≤i,j,k≤d denote the structure constants associated to the basis,
defined by the equations
(12)
(13)
Then
for all 1 ≤ i, j ≤ d : aiaj =
di,j,kak .
dXk=1
Xi =
di,1,1
...
di,d,1
di,1,2
...
di,d,2
. . . di,1,d
...
...
. . . di,d,d
.
(d) We note the following nontrivial properties of the di,j,k, valid for all 1 ≤
i, j, k ≤ d:
di,j,k = dj,i,k , d1,j,k = δj,k , di,j,k = 0 whenever k ≤ max(i, j) .
In particular, we get for 2 ≤ j ≤ d
(14)
Xi = E1,i +
di,j,kEj,k .
dXj=2
dXk=j+1
Proof. We start with part (c). Since multiplication with a1 = 1A is the identity
operator, the statement about X1 is clear. Let 1 ≤ i, j ≤ d. By definition of ψ, we
have ψ(ej) = aj, and hence by definition of Ψ
Ψ(ai)ej = ψ−1(ai · ψ(ej)) = ψ−1(aiaj)
dXk=1
= ψ−1 dXk=1
di,j,kak! =
di,j,kek .
Hence the jth column of Ψ(ai) is the vector (di,j,1, . . . , di,j,d)T , and its transpose
is the jth row of Ψ(ai)T . This shows (13).
RECENT PROGRESS IN SHEARLET THEORY
13
Now, with (c) established, the equation
a1aj = aj
for i = 2, . . . , d, yields that d1,j,k = δj,k, which also takes care of part (b). Fur-
thermore, the fact that aiaj = ajai ensures that di,j,k = dj,i,k. Finally, recall
that AIi ⊂ Ii+1 by (11), which entails aiaj ∈ Ii+1, and thus di,j,k = 0 whenever
k ≤ i. Since di,j,k = dj,i,k, we then obtain more generally that k ≤ max(i, j) entails
di,j,k = 0. Now equation (14) is clear, and (d) is shown.
In order to prove (a), we first note that Ψ is a homomorphism of associative
algebras, hence s = Ψ(N )T is a commutative associative matrix algebra. In par-
ticular, it is also an abelian Lie-subalgebra. Furthermore, the relation di,j,k = 0
whenever k ≤ max(i, j) ensures that the basis X2, . . . , Xd consists of strictly upper
triangular matrices. In addition, d1,j,k = δj,k entails that X2, . . . , Xd is indeed a
canonical basis, and thus Lemma 5 gives that the associated Lie group is a shearing
subgroup. Now part (a) of Lemma 6 yields (a) of the current lemma, and (b) is
also shown.
(cid:3)
Remark 14. It is natural to ask whether the construction of shearing subgroups S
from irreducible commutative associative algebras A, as described in Lemma 13, is
exhaustive. The answer is yes. To see this, consider the Lie algebra s of a shearing
subgroup S. Let X2, . . . , Xd be the canonical basis of s. Since Xj is strictly upper
triangular, and the first row of Xi equals eT
i , it follows that the first i entries of the
first row of XiXj vanish. This product is again in the span of the Xk, hence
XiXj =Xk>i
di,j,kXk ,
with suitable coefficients di,j,k. But the fact that the sum on the right-hand side
starts with k = i + 1 shows that the basis X2, . . . , Xd is a Jordan-Holder basis of
the nilpotent associative matrix algebra s. If one now applies the procedure from
Lemma 13 (with ai = Xi), direct calculation allows to verify that Ψ(X)T = X for
all X ∈ s. Hence every shearing subgroup arises from the construction in Lemma
13.
In particular, the observations concerning the structure constants di,j,k made in
part (d) of Lemma 13 also apply to the di,j,k in Lemma 9.
Remark 15. A further benefit of the above construction of shearing groups via
associative algebras is that it settles the question of conjugacy as a byproduct.
By Theorem 13 in [26] and the remarks prior to that result, one sees that two
shearing subgroups S1 and S2 are conjugate iff their Lie algebras are isomorphic as
associative algebras.
In particular, following the observation made in [26, Theorem 15], in dimension
d ≥ 7 there exist uncountably many nonconjugate shearing subgroups.
3.2. An inductive approach to shearlet dilation groups. For possible use in
inductive proof strategies, we note a further consequence of the block structure:
Proposition 16. Let H = ±DS < Gl(d, R) denote a shearlet dilation group, with
d ≥ 3, and let
H1 =(cid:26)h′ ∈ GL(d − 1, R) : ∃ h ∈ H, z ∈ Rd−1, s ∈ R\ {0} with h =(cid:18) h′
0
z
s (cid:19)(cid:27) .
Then H1 is a shearlet dilation group as well.
14
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
Conversely, the elements of H can be described in terms of H1 as follows: There
exists a map y : H1 → Rd−1 such that we can write each h ∈ H uniquely as
h(h1, r) =
h1
0 . . .
. . . 0
r
y1(h1)
y2(h1)
...
yd−2(h1)
yd−1(h1)
,
with h1 ∈ H1, r ∈ R.
3.3. Examples. As a result of the previous subsections, we obtain the following
general procedure for the systematic construction of shearlet dilation groups:
(1) Fix a nilpotent associative algebra N .
(2) Pick a Jordan-Holder basis a2, . . . , ad of N , and compute the canonical basis
X2, . . . , Xd of the Lie algebra s of the associated shearing subgroup. Note
that this amounts to determining the structure constants (di,j,k)1≤i,j,k≤d.
The shearing subgroup is then determined as S = Id + s.
(3) In order to determine the diagonal scaling groups that are compatible with
S, set up and solve the linear system (7) induced by the nonvanishing di,j,k.
We will now go through this procedure for several examples or classes of exam-
ples.
Example 17. We start out with the simplest case of a nilpotent algebra N of
dimension d−1, namely that of nilpotency class 2. Here one has ab = 0 for any a, b ∈
N , and it is clear that for two such algebras, any linear isomorphism is an algebra
isomorphism as well. Picking any basis a2, . . . , an of N , we obtain Xi = E1,i. In
particular, the linear system (7) is trivial. Hence any one-parameter diagonal group
can be used as scaling subgroup. We thus recover the groups described in (4).
Example 18. Another extreme class of nilpotent algebras of dimension d is that
of nilpotency class d. Here there exists b ∈ N with bd−1 6= 0. This implies that
b, . . . , bd−1 are linearly independent, and then it is easily seen that ai = bi−1, for
i = 2, . . . , d, defines a Jordan-Holder basis of N .
In this example, the defining
relations read
(15)
aiaj = ai+j−1, 2 ≤ i, j, i + j − 1 ≤ d,
and the resulting canonical Lie algebra basis is then determined as
X2 =
0
1
0
1
. . .
. . .
. . .
. . .
0
1
0
, X3 =
0 0
0
1
0
. . .
1
. . .
0
. . .
0
0
1
0
0
, . . . ,
RECENT PROGRESS IN SHEARLET THEORY
15
Xd =
0 . . .
. . .
. . .
0
0 1
0
...
...
...
0
.
Thus we see that the resulting shearing subgroup is that of the Toeplitz shearlet
group from (5). The linear system (7) becomes
It is easy to see that all solutions of this system are given by
µi + µj = µi+j−1, for 2 ≤ i, j, i + j − 1 ≤ d.
µj = (j − 1)δ, j = 2, . . . , d
with δ an arbitrary real parameter. Thus the scaling subgroups compatible with
the Toeplitz dilation group are precisely given by
exp(Rdiag(1, 1 + δ, . . . , 1 + (d − 1)δ)) ,
with δ ∈ R arbitrary.
Remark 19. For d = 3, the two above listed cases are all possible examples of
shearing subgroups, and not even just up to conjugacy. In particular, we find that
all shearing subgroups in dimension 3 are compatible with anisotropic dilations.
We now turn to the shearing subgroups in dimension 4, with focus on the groups
not covered by (4) and (5).
Example 20. Since the nilpotency classes n = 2, 4 are already covered by the
previous examples, the remaining 4-dimensional cases of irreducible algebras A all
have nilpotency class 3. It is shown in [26] that A ∼= R[Y1, Y2]/(Y 3
1 , Y1Y2),
with α ∈ {−1, 0, 1}. Here, R[Y1, Y2] denotes the algebra of polynomials with real
coefficients and indeterminates Y1, Y2, and J = (Y 3
2 − αY 2
1 , Y1Y2) denotes the
ideal generated by the three polynomials. Then the nilradical N is generated by
Y1 +J , Y2 + J . We choose the basis a2 = Y1 +J , a3 = Y2 +J , a4 = Y 2
1 +J , and
obtain as the only nonzero relations
2 −αY 2
1 , Y 2
1 , Y 2
a2
2 = a4 , a2
3 = αa4 .
This allows to conclude that a2, a3, a4 is indeed a Jordan-Holder basis. Follow-
ing Lemma 13 (c), we can read off the canonical basis of the associated shearing
subgroup as
X2 =
0
0
0
0
1 0
0 0
0 0
0 0
0
1
0
0
, X3 =
0
0
0
0
0
0 1
0 0
0
0 0 α
0 0
0
, X4 =
0 0
0 0
0 0
0 0
0 1
0 0
0 0
0 0
.
We next determine the compatible scaling subgroups. In the case α 6= 0, we obtain
the system of equations
2µ2 = µ4, 2µ3 = µ4.
Thus the infinitesimal generators of scaling subgroups are of the form Y = diag(1, 1+
δ, 1 + δ, 1 + 2δ), with δ ∈ R arbitrary.
16
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
In the case α = 0, we only get one equation, namely
2µ2 = µ4,
showing that here the compatible infinitesimal generators are of the form Y =
diag(1, 1 + δ1, 1 + δ2, 1 + 2δ1), with δ1, δ2 ∈ R arbitrary.
Finally, we give an example of a shearing subgroup which is only compatible
with isotropic scaling. It is based on the same algebra as Example 18 (with d = 4),
and as a result the associated shearing subgroups are conjugate. Recall that the
groups in Example 18 are compatible with anisotropic scaling. This illustrates an
important, somewhat subtle point: While the precise choice of Jordan-Holder basis
in the procedure described in Lemma 13 is immaterial if one is just interested in
guaranteeing the shearing subgroup property, it may have a crucial influence on the
availability of compatible anisotropic scaling subgroups.
Example 21. Let A = R[X]/(X 4). We use the Jordan-Holder algebra a2 =
X + X 2 + (X 4), a3 = X 2 + (X 4), a4 = X 3 + (X 4). This leads to the following
nonzero relations
a2
2 = a3 + 2a4, a2a3 = a4,
which gives rise to the basis
X2 =
0
0
0
0
1 0
0 1
0 0
0 0
0
2
1
0
, X3 =
0
0
0
0
0 1
0 0
0 0
0 0
0
1
0
0
, X4 =
0 0
0 0
0 0
0 0
0 1
0 0
0 0
0 0
.
Now the nonzero entries in the matrix X2 imply that the linear system (7) contains
the equations
2µ2 = µ3, 2µ2 = µ4, µ2 + µ3 = µ4.
The first two equations imply µ3 = 2µ2 = µ4, and then the third equation yields
µ2 = 0. Hence this shearing subgroup is only compatible with isotropic scaling.
4. Anisotropic scaling and wavefront set characterizations
In this section we investigate the suitability of the various groups for microlocal
analysis. The idea is to verify the criteria derived in [22] that allow to establish
the suitability of a dilation group for the characterization of the wavefront set via
wavelet coefficient decay. As it will be seen, this property only depends on the
scaling subgroup.
4.1. Criteria for wavefront set characterization. Throughout this subsection
H is an irreducibly admissible matrix group, i.e. its dual action has a single open
orbit O ⊂ Rd, with associated compact fixed groups. We use V ⋐ O to denote that
the closure of V inside O is compact.
Given R > 0 and x ∈ Rd, we let BR(x) and BR (x) denote the open/closed
ball with radius R and center x, respectively. We let Sd−1 ⊂ Rd denote the unit
sphere. By a neighborhood of ξ ∈ Sd−1, we will always mean a relatively open set
W ⊂ Sd−1 with ξ ∈ W . Given R > 0 and an open set W ⊂ Sd−1, we let
C(W ) := {rξ′ : ξ′ ∈ W, r > 0} =(cid:26)ξ ∈ Rd \ {0} :
ξ
ξ ∈ W(cid:27) ,
C(W, R) := C(W ) \ BR(0).
RECENT PROGRESS IN SHEARLET THEORY
17
Both sets are clearly open subsets of Rd \ {0} and thus of Rd.
Given a tempered distribution u, we call (x, ξ) ∈ Rd × Sd−1 a regular directed
point of u if there exists ϕ ∈ C∞
c (Rd), identically one in a neighborhood of x, as
well as a ξ-neighborhood W ⊂ Sd−1 such that for all N ∈ N there exists a constant
CN > 0 with
(16)
for all ξ′ ∈ C (W ) :
cϕu(ξ′) ≤ CN (1 + ξ′)−N .
We next formally define the sets Ki and Ko which will allow to associate group
elements to directions.
Definition 22. Let ∅ 6= W ⊂ Sd−1 be open with W ⊂ O (which implies C (W ) ⊂
O). Furthermore, let ∅ 6= V ⋐ O and R > 0. We define
as well as
Ki(W, V, R) :=(cid:8)h ∈ H : h−T V ⊂ C(W, R)(cid:9)
Ko(W, V, R) :=(cid:8)h ∈ H : h−T V ∩ C(W, R) 6= ∅(cid:9) .
If the parameters are provided by the context, we will simply write Ki and Ko.
Here, the subscripts i/o stand for "inner/outer".
We now define what we mean by dilation groups characterizing the wavefront
set. We first extend the continuous wavelet transform to the space of tempered dis-
tributions. I.e., we use Wψu, for a Schwartz wavelet ψ and a tempered distribution
u.
Definition 23. The dilation group H characterizes the wavefront set if there exists
a nonempty open subset V ⋐ O with the following property: For all 0 6= ψ ∈ S(Rd)
following statements are equivalent:
with supp(bψ) ⊂ V , for every u ∈ S′(Rd) and all (x, ξ) ∈ Rd × (O ∩ Sd−1), the
(a) (x, ξ) is a regular directed point of u.
(b) There exists a neighborhood U of x, some R > 0 and a ξ-neighborhood
W ⊂ Sd−1 such that for all N ∈ N there exists a constant CN > 0 such
that for all y ∈ U , and for all h ∈ Ko(W, V, R) the following estimate holds:
Wψu(y, h) ≤ CNkhkN .
Note that the definition excludes a set of directions ξ from the analysis of the
wavefront set, namely the directions not contained in O ∩ Sd−1. These directions
always constitute a set of measure zero. Recall from Proposition 11 that in the case
of shearlet dilation groups, this exceptional set is given by ({0} × Rd−1) ∩ Sd−1.
We next recall the sufficient conditions for dilation groups that characterize the
wavefront set, as established in [22]. The first one is related to the problem that
one would like to interpret the norm as a scale parameter.
Definition 24. Let ξ ∈ O ∩ Sd−1 and ∅ 6= V ⋐ O. The dual action is called V -
microlocally admissible in direction ξ if there exists a ξ-neighborhood W0 ⊂ Sd−1∩O
and some R0 > 0 such that the following hold:
(1) There exist α1 > 0 and C > 0 such that
kh−1k ≤ C · khk−α1
holds for all h ∈ Ko(W0, V, R0).
18
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
(2) There exists α2 > 0 such that
ZKo(W0,V,R0) khkα2 dh < ∞.
The dual action is called microlocally admissible in direction ξ if it is V -microlocally
admissible in direction ξ for some ∅ 6= V ⋐ O.
The second important condition is contained in the following definition. It can
be understood as formalizing the ability of the associated wavelet systems to be able
to make increasingly fine distinctions between different directions, as the scales go
to zero.
Definition 25. Let ξ ∈ O∩ Sd−1 and ∅ 6= V ⋐ O. The dual action has the V -cone
approximation property at ξ if for all ξ-neighborhoods W ⊂ Sd−1 and all R > 0
there are R′ > 0 and a ξ-neighborhood W ′ ⊂ Sd−1 such that
Ko(W ′, V, R′) ⊂ Ki(W, V, R).
We now have the following [22, Corollary 4.9]:
Theorem 26. Assume that the dual action is V -microlocally admissible at some
ξ0 ∈ O and has the V -cone approximation property at ξ0, for some nonempty open
subset V ⊂ O. Then H characterizes the wavefront set.
Remark 27. The property of characterizing the wavefront set is linked to anisotropic
scaling, in the following sense: If H characterizes the wavefront set, then
H ∩ R+ · Id = {Id} ,
by [22, Lemma 4.10]. Hence if H is a shearlet dilation group characterizing the
wavefront set, its shearing subgroup must admit at least one anisotropic compatible
scaling subgroup. This excludes the shearing group constructed in Example 21.
Theorem 26 therefore implies that every group failing the anisotropy criterion
H ∩ R+ · Id = {Id} must necessarily fail either the microlocal admissibility or the
cone approximation property. It is in fact the latter that breaks down, as noted in
[22, Lemma 4.4].
These considerations highlight the importance of understanding when a given
shearing groups admits anisotropic scaling.
4.2. Characterization of the wavefront set for shearlet dilation groups.
We can now state a very general theorem concerning the ability of shearlet groups
to characterize the wavefront set. Note that there are no conditions on the shearing
subgroups.
Theorem 28. Let H be a shearlet dilation group and let Y = diag(1, λ2, . . . , λd)
denote the infinitesimal generator of the scaling subgroup. If 0 < λi < 1 holds, for
all 2 ≤ i ≤ d, then H characterizes the wavefront set.
Remark 29. We can now quickly go through the examples of shearing subgroups in
Subsection 3.3 and show that for most cases, there exists a compatible anisotropic
scaling subgroup that allows to characterize the wavefront set. Writing λi = 1 + µi
as in Lemma 5, the condition from Theorem 28 translates to −1 < µi < 0, for
2 ≤ i ≤ d. Apart from the group in Example 21, which was specifically constructed
to not allow any anisotropic scaling, all other shearing groups can be combined
with a compatible scaling group in such a way that the resulting shearlet transform
RECENT PROGRESS IN SHEARLET THEORY
19
fulfills the conditions of Theorem 28, and therefore characterizes the wavefront set.
Note that this was previously known only for the original shearlet group [36, 22].
In particular, we may combine the Toeplitz shearing subgroup with the scaling
subgroup with exponents (1, 1− δ, . . . , 1− (d− 1)δ), and choosing δ ∈ (0, 1/(d− 1))
guarantees that the Toeplitz shearlet transform characterizes the wavefront set.
The proof of the Theorem amounts to verifying the cone approximation prop-
erty and microlocal admissibility of the dual action, and this will be carried out in
the following two propositions. For the remainder of this section, we fix a shear-
let dilation group H with infinitesimal generator diag(1, λ2, . . . , λd) of the scaling
subgroup. We let λmax = maxi≥2 λi, and λmin = mini≥2 λi.
Proposition 30. If λmax < 1, there exists an open subset ∅ 6= V ⋐ O such that
the dual action of H on the orbit O has the V -cone approximation property at
(1, 0, . . . , 0)T ∈ Sd−1 ∩ O.
Proof. We will employ the structural properties of shearing subgroups derived in
Section 2.2. We let S and D denote the shearing and scaling subgroups of H,
respectively. The infinitesimal generator of D is a diagonal matrix with the entries
1, λ2, . . . , λd. We let X2, . . . , Xd denote the canonical basis of s, consisting of strictly
upper triangular matrices Xi. By Lemma 6, each h ∈ S is uniquely described by
h = h(t, 1) = Id +
tiXi ,
dXi=2
where t = (t2, . . . , td)T denotes the vector of first row entries of h(t, 1). For H, we
thus obtain the global chart
h(t, a) = Id +
tiXi! sgn(a)diag(a,aλ2 , . . . ,aλd ) ∈ GL(d, R),
dXi=2
with (t, a) ∈ Rd−1 × R×.
For the purpose of the following computations, it is possible and beneficial to
slightly modify this construction and replace h(t, 1) by its inverse. Thus, every
h ∈ H can be written (uniquely) as h = ±h(t, 1)−1h(0, a) with t ∈ Rd−1 and
a ∈ (0, +∞). The dual action is then given by
(h−1)T = ±(h(0, a)−1h(t, 1))T = ± Id +
i ! h(0, a−1),
tiX T
(17)
dXi=2
where by construction
(18)
Id +
tiX T
i =(cid:18)1
t
0T
Id−1 + A(t)T(cid:19) ,
dXi=2
with A(t) being a (d − 1) × (d − 1) strictly lower-triangular matrix satisfying
(19)
kA(t)k ≤ Ct
with a constant C depending only on H.
We now parametrise the open orbit O by the global chart provided by affine
coordinates
Ω : R× × Rd−1 → O
Ω(τ, v) = τ (1, vT )T ,
20
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
and Sd−1 ∩ O by the corresponding diffeomorphism to its image
.
(1, vT )T
ω : Rd−1 → Sd−1 ∩ O,
ω(v) =
p1 + v2
Wǫ = {v ∈ Rd−1 : v < ǫ} = Bǫ(0),
Given ǫ > 0, we set
since {Wǫ : ǫ > 0} is a neighbourhood basis of the origin in Rd−1 and {ω(Wǫ) : ǫ >
0} is a neighbourhood basis of ξ0 = (1, 0, . . . , 0) ∈ Sd−1 ∩ O.
Furthermore, for fixed 0 < τ1 < τ2 and ǫ0 > 0 the set
V = Ω( (τ1, τ2) × Wǫ0 )
is an open subset with V ⋐ O.
v ∈ Wǫ0 , and we get
Given h ∈ H, as in (17), and ξ ∈ V , then ξ = Ω(τ, v) with τ1 < τ < τ2 and
(h−1)T ξ = ±τ(cid:18)1
t
0T
Id−1 + A(t)T(cid:19)(cid:18)a−1
v′ (cid:19) = ±a−1τ(cid:18)
t + (Id−1 + A(t)T )v′′(cid:19)
1
where v′, v′′ ∈ Rd−1 have components given by v′
all i = 2, . . . , d. Hence
i = a−λivi and v′′
i = a1−λivi for
ǫ0 )(cid:1)
(h−1)T (V ) = Ω(cid:0) (±a−1τ1,±a−1τ2) × (t + (Id−1 + A(t)T )W a
i = a1−λi vi, v < ǫ0}.
ǫ0 = {v′′ ∈ Rd−1 : v′′
where W a
Fix now R > 0 and a neighborhood W ⊂ Sd−1 ∩ O of ξ0. Without loss of
generality we can assume that W = ω(Wǫ) for some ǫ > 0. Furthermore, since
(R, +∞) × Wǫ ⊂ Ω−1(C(ω(Wǫ), R)) ⊂ (
, +∞) × Wǫ,
where the last inclusion on the right holds if ǫ ≤ 1, then the V -cone approximation
property holds true if there exist R′ > 0 and 0 < ǫ′ ≤ 1 such that for all h ∈ H
satisfying
, +∞) × Wǫ ⊂ (
√1 + ǫ2
R
2
R
(20b)
Take R′ > 0 and 0 < ǫ′ < √3, which we will fix later on as functions of R and ǫ,
and h ∈ H as in (17). If h = −(h(0, a)−1h(t, 1))T then
so that (20a) implies that h = +(h(0, a)−1h(t, 1))T and
R′
2
(cid:18) (
(cid:18) (
, +∞) × Wǫ′(cid:19) ∩(cid:0) (−a−1τ2,−a−1τ1) × (t + (Id−1 + A(t)T )W a
, +∞) × Wǫ′(cid:19) ∩(cid:0) (a−1τ1, a−1τ2) × (t + (Id−1 + A(t)T )W a
ǫ0(cid:1) 6= ∅.
Wǫ′ ∩(cid:0)t + (Idd−1 +A(t)T )W a
R′ < 2a−1τ2,
ǫ0 )(cid:1) = ∅,
ǫ0 )(cid:1) 6= ∅.
R′
2
If we choose R′ > 2τ2, the first inequality gives
Hence
(21)
a <
2τ2
R′ < 1,
(20a)
it holds that
R′
2
(h−1)T (V ) ∩ Ω(cid:18) (
, +∞) × Wǫ′(cid:19) 6= ∅,
(h−1)T (V ) ⊂ Ω ( (R, +∞) × Wǫ ) .
RECENT PROGRESS IN SHEARLET THEORY
21
and, since a < 1, setting λmax = max{λ2, . . . , λd}, clearly
(22)
W a
ǫ0 ⊂ Wǫ0a1−λmax .
By the above inclusion, since Wǫ′ ∩(cid:0)t + (Id−1 + A(t)T )W a
ξ ∈ Wǫ0a1−λmax such that t + ξ + A(t)T ξ < ǫ′. Hence, triangle inequality, (19)
and (21) give
ǫ0(cid:1) 6= ∅, then there exists
t < ǫ′ + (1 + kA(t)Tk)ξ ≤ ǫ′ + (1 + Ct)a1−λmax ǫ0
≤ ǫ′ + (
2τ2
R′ )1−λmax (1 + Ct)ǫ0 ≤ 2ǫ′ +
1
2t,
where the last inequality holds true provided that
(23)
R′ > 2τ2 max{1, (
ǫ0
ǫ′ )
1
1−λmax , (2Cǫ0)
1
1−λmax }.
Hence, if (20a) holds true with R′ satisfying (23), then
(24a)
(24b)
(24c)
(
2τ2
R′ < 1
a <
t < 4ǫ′
2τ2
R′ )1−λmax ǫ0 < min{ǫ′,
1
2C }.
The condition (20b) is equivalent to
(a−1τ1, a−1τ2) × (t + Id−1 + A(t)T )W a
ǫ0 ) ⊂ (R, +∞) × Wǫ,
which is ensured by a−1τ1 > R and, recalling (22), by t+(Id−1+A(t)T )Wǫ0a1−λmax ⊂
Wǫ.
R′ . Taking into account (23),
By (24a) the first condition is satisfied if τ1/R > 2τ2
it is sufficient to assume that
(25)
R′ > 2τ2 max{1, (
ǫ0
ǫ′ )
1
1−λmax , (2Cǫ0)
1
1−λmax ,
R
τ1 }.
To ensure that t+(Id−1 +A(t)T )Wǫ0a1−λmax ⊂ Wǫ, note that, for all ξ ∈ Wǫ0a1−λmax ,
conditions (19), (24a), and (24b) give
t + (Id−1 + A(t)T )ξ ≤ t + (1 + Ct)ξ ≤ t + (1 + Ct)a1−λmax ǫ0
< 4ǫ′ + (1 + C4ǫ′)(
≤ 4ǫ′ + ǫ′ + 2ǫ′ = 7ǫ′,
2τ2
R′ )1−λmax ǫ0
where the last inequality follows from (24c). Hence, with the choice ǫ′ = min{1, ǫ/7}
and R′ satisfying (25) for all ξ ∈ Wǫ0a1−λmax ,
t + (Id−1 + A(t)T )ξ < ǫ,
so that (20b) holds true for all h ∈ H satisfying (20a).
Remark 31. The proof does not make use of the fact that the shearlet group S is
abelian. The proof is based only on the following two properties of S
(cid:3)
(a) a global smooth chart t 7→ s(t) from Rd−1 onto S;
22
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
(b) for all t ∈ Rd the dual action of s(t) is of the form
(s(t)−1)T =(cid:18)1
t B(t)(cid:19)
0T
where kB(t)k ≤ C1 + C2t for a suitable choice of C1 and C2.
With the cone approximation property already established, the remaining con-
dition is quite easy to check.
Proposition 32. If 0 < λmin ≤ λmax < 1, there exists an open subset ∅ 6= V ⋐
O such that the dual action of H on the orbit O is V -microlocally admissible in
direction (1, 0, . . . , 0) ∈ Sd−1 ∩ O.
Proof. We retain the notations from the previous proof, as well as the open set
V = Ω( (τ1, τ2) × Wǫ0 ) ,
with τ1 < 1 < τ2. Since we assume λmax < 1, the cone approximation property
holds, and then condition (2) of Definition 24 follows from condition (1) by [22,
Lemma 4.7]. In addition, the cone approximation property allows to replace Ko in
that condition by the smaller set Ki. In short, it remains to prove the existence of
α > 0 and C′′ > 0 such that
kh−1k ≤ C′′khk−α
holds for all h ∈ Ki(ω(Wǫ), V, R), for suitable ǫ, R > 0.
In the following com-
putations, we let ǫ = 1 and R > 2. Now assume that h = ±h(t, 1)−1h(0, a) ∈
Ki(ω(Wǫ), V, R), which means that h−T V ⊂ C(ω(Wǫ), R). This implies in partic-
ular that
h−T
1
0
...
0
= ±h(t, 1)T
a−1
0
...
0
1
0
...
0
= ±h(t, 1)T h(0, a)−1
= ±a−1(cid:18) 1
t (cid:19) ∈ C(ω(Wǫ), R) .
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
h−T
2a−1 ≥
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
0
...
0
> R > 2 ,
This implies that the sign is in fact positive. Furthermore, we have t ≤ ǫ = 1, and
then
which implies a < 1. By using the fact kh(t, 1)k ≤ C(1 + t) ≤ 2C ≤ √2(1 + (1 +
C)t) ≤ √2(2 + C), where C was the constant from (19), we can now estimate
where we used a < 1 and λmax ≤ 1 to estimate the norm of h(0, a)−1. In addition,
khk = kh(t, 1)−1h(0, a)k ≤ kh(t, 1)−1kkh(0, a)k ≤ C′aλmin .
kh−1k = kh(0, a)−1h(t, 1)k ≤ kh(0, a)−1kkh(t, 1)k ≤
√2(2 + C)a−1,
RECENT PROGRESS IN SHEARLET THEORY
23
Here we used that the set {h(t, 1) : t ≤ 1} ⊂ H is compact to uniformly estimate
the norm of the inverses by a suitable C′, and a < 1 to estimate the norm of h(0, a).
But these estimates combined yield
kh−1k ≤
√2(2 + C)a−1 ≤
√2(2 + C)(C′)1/λminkhk−1/λmin .
Since we assume that λmin > 0, the proof is finished.
(cid:3)
5. Embeddings into the symplectic group
From the analytical point of view, we saw that shearlet dilation groups are a
useful tool for the characterization of the wavefront set of distributions. On the
other hand, from the algebraic and geometrical points of view, these groups and the
associated generalized wavelet representation exhibit an interesting link with the
symplectic group and the metaplectic representation. More precisely, in this section
we show that the positive part DS of any shearlet dilation group DS ∪ (−DS) may
be imbedded into the symplectic group. Note that the full group DS ∪ (−DS)
cannot be expected to be imbedded into Sp(d, R) [6, Theorem 3.5]. Moreover,
we prove that the wavelet representation is unitarily equivalent to the metaplectic
representation, provided that they are restricted to a suitable subspace of L2(Rd).
In fact, a much more general class of groups is allowed, see Theorem 33.
The relevance of the symplectic group and of the metaplectic representation in
this context has already been shown in several works [15, 1, 2, 6]. In particular,
the argument given here generalizes [6].
Let T (d, R)+ denote the subgroup of GL(d, R) consisting of the upper triangular
matrices with positive entry in position (1, 1), namely
T (d, R)+ = {h ∈ GL(d, R) : h1,1 > 0 and hi,j = 0 for every i > j}.
We consider the following subspace of L2(Rd):
H = {f ∈ L2(Rd) : supp f ⊆ ΘL}, where ΘL = {ξ ∈ Rd : ξ1 ≤ 0}.
The main result of this section reads as follows.
Theorem 33. Take H < T (d, R)+. The group G = Rd ⋊ H may be embedded
into the symplectic group, namely there exists a group embedding φ : G → Sp(d, R).
Moreover, the restriction to H of the quasi-regular representation π defined in (1) is
unitarily equivalent to µ◦φ restricted to H, where µ is the metaplectic representation
of Sp(d, R).
The rest of this section is devoted to the proof of this theorem. The embedding φ,
the subgroup φ(G), as well as the intertwining operator between the quasi-regular
representation and the metaplectic representation will be explicitly constructed.
First, we construct the subgroup φ(G) < Sp(d, R) and the map φ. The vectorial
part of G = Rd ⋊ H will correspond to the subspace of the d-dimensional symmetric
matrices given by
We shall need the following preliminary result concerning the map
(26)
ρ : T (d, R)+ → GL(d, R),
Σ := {σb :=
b1
b2/2
...
bd/2
b2/2 ···
0
: b ∈ Rd}.
bd/2
h 7→ph1,1 h−T .
24
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
Lemma 34. The map ρ is a group homomorphism and for all b ∈ Rd and h ∈
T (d, R)+ there holds
(27)
ρ(h)−T σbρ(h)−1 = σhb.
Proof. The first part is trivial, since the matrices in H are upper triangular with
h1,1 > 0. The second part can be proven as follows. Fix b ∈ Rd and h ∈ T (d, R)+.
The assertion is equivalent to
hσb hT = h1,1σhb.
Write for i = 2, . . . , d
h1
h2
...
hd
,
h1,1 h′
1
h′
0
2
...
...
h′
0
d
(cid:18) b1
b′/2
We have
h =
hσb =
hσb hT =
=
whence
h1,1b1 + h′
2b′/2
h′
...
db′/2
h′
Therefore, since b′T h′T
h1,1h′
2b′/2
...
db′/2
h1,1h′
i = h′
h1 = [h1,1 h′
1],
hi = [0 h′
i],
b′ =
b2
...
bd
.
h1,1b1 + h′
2b′/2
h′
...
h′
db′/2
b′T /2
0 (cid:19) =
(cid:18)h1,1
h′T
1
...
0
0
1b′/2
h1,1b′T /2
1b′/2
h1,1b′T /2
,
0
h′T
2
···
···
0
h′T
d (cid:19)
0
0
...
0
.
(cid:3)
h1,1(h1,1b1 + h′
1b′/2) + h1,1b′T h′T
1 /2
h1,1b′T h′T
2 /2 ···
h1,1b′T h′T
d /2
ib′ for every i and h′
ib′ = hib for i ≥ 2, we obtain
2b′/2 ···
h′
db′/2
h′
1b′
h1b1
h2b/2
h1,1b1 + h′
2b′/2
h′
...
db′/2
h′
hσb hT = h1,1
= h1,1
σh h−T(cid:19) ∈ Sp(d, R),
hdb/2
...
whence hσb hT = h1,1σhb, as desired.
We use the notation
g(σ, h) =(cid:18) h
h2b/2 ···
0
0
hdb/2
,
σ ∈ Sym(d, R), h ∈ GL(d, R).
RECENT PROGRESS IN SHEARLET THEORY
25
The product law is
(28)
g(σ1, h1)g(σ2, h2) = g(σ1 + h−T
1 σ2h−1
1 , h1h2).
In the following result we show that G = Rd ⋊ H is isomorphic to the subgroup of
Sp(d, R) given by Σ⋊ρ(H) := g(Σ, ρ(H)). This proves the first part of Theorem 33.
Proposition 35. Take H < T (d, R)+. Then the map
φ : Rd ⋊ H → g(Σ, ρ(H)) < Sp(d, R),
is a group isomorphism.
(b, h) 7→ g(σb, ρ(h))
It is worth mentioning that Lemma 2.3 in [6] immediately follows from this result.
Proof. Recall that the product in Rd ⋊ H is defined by
(b1, h1)(b2, h2) = (b1 + h1b2, h1h2),
By definition of φ and using (28) there holds
bi ∈ Rd, hi ∈ H.
φ(b1, h1)φ(b2, h2) = g(σb1 , ρ(h1))g(σb2 , ρ(h2))
= g(σb1 + ρ(h1)−T σb2 ρ(h1)−1, ρ(h1)ρ(h2)).
Therefore, Lemma 34 gives
φ(b1, h1)φ(b2, h2) = g(σb1 + σh1b2 , ρ(h1h2))
= g(σb1+h1b2 , ρ(h1h2))
= φ((b1, h1)(b2, h2)),
as desired. Note that the fact that g(Σ, ρ(H)) is a subgroup follows a posteriori. (cid:3)
Intertwining the quasi-regular representation π, given in (1), with the Fourier
transform F : H → L2(ΘL) we obtain the representation π(b, h) := F π(b, h)F −1
on L2(ΘL) given by
f ∈ L2(ΘL).
The metaplectic representation restricted to Σ ⋊ ρ(H) takes the form
π(b, h) f (ξ) = det h1/2e−2πihb,ξi f (hT ξ),
(29)
µ(φ(b, h)) f (ξ) = det ρ(h)−1/2eπihσbξ,ξi f (ρ(h)−1ξ),
f ∈ L2(ΘL).
We now show that π and µ are unitarily equivalent, which concludes the proof of
Theorem 33. The intertwining operator is given by
Ψ : L2(ΘL) → L2(ΘL), Ψ f (ξ) = det JQ−1 (ξ)1/2 f (Q−1(ξ)),
where Q : ΘL → ΘL is defined by Q(ξ) = − 1
Proposition 36. Let φ be the group isomorphism given by Proposition 35. For
every (b, h) ∈ Rd ⋊ H there holds
2 ξ1ξ.
Ψµ(φ(b, h))Ψ−1 = F π(b, h)F −1 = π(b, h).
Proof. We start by giving a few identities without proof [6]:
(30)
(31)
(32)
(33)
2 −1ξ1− d
2 ,
d
det JQ(ξ) = 21−dξ1d,
det JQ−1(ξ) = 2
hσbξ, ξi = −2hb, Q(ξ)i,
Q−1(ξ) = √2ξ/p−ξ1.
26
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
By (32) and (27) there holds
−2hb, Q(hT ξ)i = hσb hT ξ, hT ξi = hhσb hT ξ, ξi = h1,1hσhbξ, ξi.
Therefore, using again (32) we obtain
whence
(34)
−2hb, Q(hT ξ)i = −2h1,1hhb, Q(ξ)i = −2hb, h1,1 hT Q(ξ)i,
Q(hT ξ) = h1,1 hT Q(ξ).
By using the definition of Ψ, (29), (32), (26) and once again the definition of Ψ, we
can now compute for f ∈ L2(ΘL), b ∈ Rd and h ∈ H
(35)
Ψµ(φ(b, h))Ψ−1 f (ξ) = det JQ−1 (ξ)1/2(µ(φ(b, h))Ψ−1 f )(Q−1(ξ))
= det JQ−1 (ξ)1/2 det ρ(h)−1/2eπihσbQ−1(ξ),Q−1(ξ)iΨ−1 f (ρ(h)−1Q−1(ξ))
= det JQ−1 (ξ)1/2 det ρ(h)−1/2e−2πihb,ξiΨ−1 f (ρ(h)−1Q−1(ξ))
− 1
1,1 hT Q−1(ξ))
= det JQ−1 (ξ)1/2h
= det JQ−1 (ξ)1/2h
1,1 det h1/2e−2πihb,ξiΨ−1 f (h
1,1 det h1/2e−2πihb,ξi
· det JQ(h
1,1 hT Q−1(ξ))1/2 f (Q(h
− d
− d
− 1
4
2
4
2
Now note that by (34) and by the fact that Q is quadratic there holds
(36)
Q(h
2
− 1
1,1 hT Q−1(ξ)) = h1,1 hT Q(h
2
− 1
1,1 Q−1(ξ)) = hT Q(Q−1(ξ)) = hT ξ.
2
− 1
1,1 hT Q−1(ξ))).
Moreover we have
det JQ(h
− 1
2
1,1 hT Q−1(ξ)) = 21−d(h
= 21−dh
− 1
2
2
− d
1,1 hT Q−1(ξ))1d
1,1 ( hT Q−1(ξ))1d
− d
1,1 hd
1,1Q−1(ξ)1d
2 ξ1
1,12
2 ,
d
2
d
d
2
= 21−dh
= 21−dh
where the first equality follows from (30), the third one from the fact that hT is
lower triangular and the forth one from (33). Therefore by (31)
(37)
det JQ−1 (ξ)1/2 det JQ(h
1,1 hT Q−1(ξ))1/2 = 2
2ξ1− d
4 2
4 ξ1
4 = h
1,12
1,1.
2 − d
4 − 1
2 h
− 1
d
4
d
4
d
d
d
1
2
Finally, inserting (36) and (37) into (35) we obtain
Ψµ(φ(b, h))Ψ−1 f (ξ) = det h1/2e−2πihb,ξi f ( hT ξ) = π(b, h) f (ξ),
as desired.
(cid:3)
Acknowledgements G. S. Alberti was partially supported by the ERC Advanced
Grant Project MULTIMOD-267184.
S. Dahlke was supported by Deutsche
Forschungsgemeinschaft (DFG), Grant DA 360/19 -- 1. He also acknowledges the
support of the Hausdorff Research Institute for Mathematics during the Special
Trimester "Mathematics of Signal Processing". F. De Mari and E. De Vito were
partially supported by Progetto PRIN 2010-2011 "Variet`a reali e complesse: geome-
tria, topologia e analisi armonica". They are members of the Gruppo Nazionale
per l'Analisi Matematica, la Probabilit`a e le loro Applicazioni (GNAMPA) of the
RECENT PROGRESS IN SHEARLET THEORY
27
Istituto Nazionale di Alta Matematica (INdAM). H. Fuhr acknowledges support
from DFG through the grant Fu 402/5-1.
Part of the work on this paper was carried out during visits of S. Dahlke and H.
Fuhr to Genova, and they thank the Universit`a di Genova for its hospitality.
References
[1] Alberti, G.S., Balletti, L., De Mari, F., De Vito, E.: Reproducing subgroups of Sp(2, R).
Part I: Algebraic classification. J. Fourier Anal. Appl. 19(4), 651 -- 682 (2013)
[2] Alberti, G.S., De Mari, F., De Vito, E., Mantovani, L.: Reproducing subgroups of Sp(2, R)
Part II: admissible vectors. Monatsh. Math. 173(3), 261 -- 307 (2014)
[3] Cand`es, E.J., Donoho, D.L.: Ridgelets: A key to higher-dimensional intermittency? Phil.
Trans. R. Soc. 357(1760), 2495 -- 2509 (1999)
[4] Cand`es, E.J., Donoho, D.L.: New tight frames of curvelets and optimal representations of
objects with piecewise C 2 singularities. Comm. Pure Appl. Math. 57(2), 219 -- 266 (2004)
[5] Czaja, W., King, E.J.: Isotropic shearlet analogs for L2(Rk) and localization operators.
Numer. Funct. Anal. Optim. 33(7-9), 872 -- 905 (2012)
[6] Dahlke, S., De Mari, F., De Vito, E., Hauser, S., Steidl, G., Teschke, G.: Different faces of
the shearlet group. J. Geom. Anal. , 1 -- 37 (2015), DOI:10.1007/s12220-015-9605-7
[7] Dahlke, S., Hauser, S., Steidl, G., Teschke, G.: Shearlet coorbit spaces: traces and embed-
dings in higher dimensions. Monatsh. Math. 169(1), 15 -- 32 (2013)
[8] Dahlke, S., Hauser, S., Teschke, G.: Coorbit space theory for the Toeplitz shearlet transform.
Int. J. Wavelets Multiresolut. Inf. Process. 10(4), (2012)
[9] Dahlke, S., Kutyniok, G., Maass, P., Sagiv, C., Stark, H.-G., Teschke, G.: The uncertainty
principle associated with the continuous shearlet transform. Int. J. Wavelets Multiresolut.
Inf. Process. 6(2) , 157 -- 181 (2008)
[10] Dahlke, S., Kutyniok, G., Steidl, G., Teschke, G.: Shearlet coorbit spaces and associated
banach frames. Appl. Comput. Harmon. Anal. 27(2), 195 -- 214 (2009)
[11] Dahlke, S., Steidl, G., Teschke, G.: Multivariate shearlet transform, shearlet coorbit spaces
and their structural properties. In: G. Kutyniok, D. Labate (eds.) Shearlets, pp. 105 -- 144.
Birkhauser/Springer, New York (2012)
[12] Dahlke, S., Steidl, G., Teschke, G.: Shearlet coorbit spaces: compactly supported analyzing
shearlets, traces and embeddings. J. Fourier Anal. Appl. 17(6), 1232 -- 1255 (2011)
[13] Dahlke, S., Steidl, G., Teschke, G.: The continuous shearlet transform in arbitrary space
dimensions. J. Fourier Anal. Appl. 16(3), 340 -- 364 (2010)
[14] Dahlke, S., Teschke, G.: The continuous shearlet transform in higher dimensions: Variations
of a theme. In: Group Theory: Classes, Representations and Connections, and Applications,
pp. 165-175. Nova Science Publishers, (2010)
[15] De Mari, F., De Vito, E.: Admissible vectors for mock metaplectic representations. Appl.
Comput. Harmon. Anal. 34(2), 163 -- 200 (2013)
[16] Do, M.N., Vetterli, M.: Contourlets: a directional multiresolution image representation. In:
Proceedings of the International Conference on Image Processing, vol. 1, pp.357 -- 360 (2002)
[17] Feichtinger, H.G., Grochenig, K.: A unified approach to atomic decompositions via integrable
group representations. In: Proc. Conf. Lund 1986 'Function spaces and applications', Lect.
Notes in Math. 1302 (1988), pp 52 -- 73
[18] Feichtinger, H.G., Grochenig, K.: Banach spaces related to integrable group representations
and their atomic decompositions. Part I. J. Funct. Anal. 86(2), 307 -- 340 (1989)
[19] Feichtinger, H.G., Grochenig, K.: Banach spaces related to integrable group representations
and their atomic decompositions. Part II. Monatsh. Math. 108(2-3), 129 -- 148 (1989)
[20] Feichtinger, H.G., Grochenig, K.: Non-orthogonal wavelet and Gabor expansions and group
representations. In: Ruskai, M.B., et al. (eds.) Wavelets and Their Applications, pp.353-376.
Jones and Bartlett, Boston (1992)
[21] Feichtinger, H.G., Sun, W., Zhou, X.: Two Banach spaces of atoms for stable wavelet frame
expansions. J. Approx. Theory 146(1), 28 -- 70 (2007)
[22] Fell, J., Fuhr, H., Voigtlaender, F.: Resolution of the wavefront set using general continuous
wavelet transforms. J. Fourier Anal. Appl., 1 -- 62 (2015)
[23] Folland, G.B.: A course in abstract harmonic analysis. CRC Press, Boca Raton, FL(1995)
28
G. S. ALBERTI, S. DAHLKE, F. DE MARI, E. DE VITO, AND H. F UHR
[24] Frazier, M., Jawerth, B.: Decomposition of Besov spaces. Indiana Univ. Math. J. 34(4),
777 -- 799 (1985)
[25] Frazier, M., Jawerth, B., Weiss, G.: Littlewood-Paley theory and the study of function
spaces. (1991)
[26] Fuhr, H.: Continuous wavelet transforms with abelian dilation groups. J. Math. Phys. 39(8),
3974 -- 3986 (1998)
[27] Fuhr, H.: Coorbit spaces and wavelet coefficient decay over general dilation groups. Trans.
Amer. Math. Soc. 367(10), 7373 -- 7401 (2015)
[28] Fuhr, H.: Generalized Calder´on conditions and regular orbit spaces. Colloq. Math. 120(1),
103 -- 126 (2010)
[29] Fuhr, H.: Vanishing moment conditions for wavelet atoms in higher dimensions. Adv. Com-
put. Math. 42(1), 127-153 (2016)
[30] Fuhr, H., Raisi-Tousi, R.: Simplified vanishing moment criteria for wavelets over general
dilation groups, with applications to abelian and shearlet dilation groups. To appear in
Appl. Comp. Harm. Anal. DOI:10.1016/j.acha.2016.03.003
Preprint available under http://arxiv.org/abs/1407.0824 (2014)
[31] Grochenig, K.: Describing functions: atomic decompositions versus frames. Monatsh. Math.
112(1),1 -- 42 (1991)
[32] Grohs, P.: Shearlet and microlocal analysis. In: Kutyniok, G., Labate, D. (eds.) Shearlets:
Multiscale Analysis for Multivariate Data, pp. 39-67. Birkhauser, Boston (2012)
[33] Hilgert, J., Neeb, K.-H.: Structure and geometry of Lie groups. Springer Monographs in
Mathematics. Springer, New York (2012)
[34] Jaffard, S., Meyer, Y.: Wavelet Methods for Pointwise Regularity and Local Oscillations of
Functions, Mem. Amer. Math. Soc. 123(587), (1996)
[35] King, E.J.: Wavelet and frame theory: frame bound gaps, generalized shearlets, Grassman-
nian fusion frames, and p-adic wavelets. Dissertation, University of Maryland, College Park
(2009)
[36] Kutyniok, G., Labate, D.: Resolution of the wavefront set using continuous shearlets. Trans.
Amer. Math. Soc. 361(5), 2719 -- 2754 (2009)
[37] Kutyniok, G., Labate, D. (eds.): Shearlets: Multiscale Analysis for Multivariate Data.
Birkhauser/Springer, New York (2012)
[38] Triebel, H.: Characterizations of Besov-Hardy-Sobolev spaces: a unified approach. J. Ap-
prox. Theory 52(2), 162 -- 203 (1988)
[39] Varadarajan, V. S.: Lie groups, Lie algebras, and their representations. Springer-Verlag,
New York (1984)
Department of Mathematics, ETH Zurich, Ramistrasse 101, 8092 Zurich, Switzerland
E-mail address: [email protected]
FB12 Mathematik und Informatik, Philipps-Universitat Marburg, Hans-Meerwein-
Strasse, Lahnberge, 35032 Marburg, Germany
E-mail address: [email protected]
Dipartimento di Matematica, Universit`a di Genova, Via Dodecaneso 35, Genova, Italy
E-mail address: [email protected]
Dipartimento di Matematica, Universit`a di Genova, Via Dodecaneso 35, Genova, Italy
E-mail address: [email protected]
Lehrstuhl A fur Mathematik, RWTH Aachen University, 52056 Aachen, Germany
E-mail address: [email protected]
|
1811.09183 | 1 | 1811 | 2018-11-22T14:13:13 | On the Maurey--Pisier and Dvoretzky--Rogers theorems | [
"math.FA"
] | A famous theorem due to Maurey and Pisier asserts that for an infinite dimensional Banach space $E$, the infumum of the $q$ such that the identity map $id_{E}$ is absolutely $\left( q,1\right) $-summing is precisely $\cot E$. In the same direction, the Dvoretzky--Rogers Theorem asserts $id_{E}$ fails to be absolutely $\left( p,p\right) $-summing, for all $p\geq1$. In this note, among other results, we unify both theorems by charactering the parameters $q$ and $p$ for which the identity map is absolutely $\left( q,p\right)$-summing. We also provide a result that we call \textit{strings of coincidences} that characterize a family of coincidences between classes of summing operators. We illustrate the usefulness of this result by extending classical result of Diestel, Jarchow and Tonge and the coincidence result of Kwapie\'{n}. | math.FA | math |
ON THE MAUREY -- PISIER AND DVORETZKY -- ROGERS THEOREMS
G. ARA ´UJO AND J. SANTOS
Abstract. A famous theorem due to Maurey and Pisier asserts that for an infinite dimensional
Banach space E, the infumum of the q such that the identity map idE is absolutely (q, 1)-
summing is precisely cot E. In the same direction, the Dvoretzky -- Rogers Theorem asserts idE
fails to be absolutely (p, p)-summing, for all p ≥ 1. In this note, among other results, we unify
both theorems by charactering the parameters q and p for which the identity map is absolutely
(q, p)-summing. We also provide a result that we call strings of coincidences that characterize
a family of coincidences between classes of summing operators. We illustrate the usefulness
of this result by extending classical result of Diestel, Jarchow and Tonge and the coincidence
result of Kwapie´n.
1. Introduction and background
Let 2 ≤ q < ∞. A Banach space E has cotype q (see [5, page 218]) if there is a constant
C > 0 such that, no matter how we select finitely many vectors x1, . . . , xn ∈ E,
(1.1)
1
q
kxkkq!
n
Xk=1
≤ C
Z[0,1](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n
Xk=1
rk(t)xk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2
1/2
dt
,
where rk denotes the k-th Rademacher function. In this context we also define
cot E := inf{q : E has cotype q}.
j=1 ∈ ℓq(F ) whenever (xj)∞
j=1 ∈ ℓw
Recall that if 1 ≤ p ≤ q ≤ ∞, for Banach spaces E, F , a linear operator u : E → F is
absolutely (q, p)-summing if (u(xj))∞
p (E)
is the linear space of the sequences (xj)∞
j=1 ∈ ℓp for every continuous
linear functional ϕ : E → K. The class of all absolutely (q, p)-summing operators from E to
F will be denoted by Π(q,p)(E; F ). We denote Π(p,p)(E; F ) by Πp(E; F ). From now on, for any
p > 1 the symbol p∗ denotes the conjugate of p, i.e., p∗ = p/ (p − 1) . Cotype and absolutely
summing operators are closely related by the famous Maurey -- Pisier Theorem (see also [4] for
further relations between cotype and absolutely summing operators):
j=1 in E such that (ϕ (xj))∞
p (E); we recall that ℓw
Theorem 1.1 (Maurey -- Pisier). For every infinite dimensional Banach space E, we have
cot E = inf{a : idE is absolutely (a, 1)-summing}.
In the same direction, the Dvoretzky -- Rogers Theorem [6] tells us that idE is not absolutely
(p, p)-summing, regardless of the p ≥ 1. In a more general version, as stated in [5, Theorem
10.5], it reads as follows:
Theorem 1.2 (Dvoretzky -- Rogers). If E is an infinite dimensional Banach space, the identity
map idE is not absolutely (q, p)-summing whenever 1
p − 1
q < 1
2.
Our first main result revisits and unifies both theorems. Our second main result provides
a family of coincidences for the classes of absolutely summing operators that encompasses the
following classical results:
J. Santos was supported by Conselho Nacional de Desenvolvimento Cient´ıfico e Tecnol´ogico -- CNPq.
2010 Mathematics Subject Classification: 46A32, 47H60.
Key words: Absolutely summing operators, Maurey -- Pisier theorem, Dvoretzky -- Rogers theorem.
1
2
ARA ´UJO AND SANTOS
Theorem 1.3 (Diestel -- Jarchow -- Tonge). Let E and F be Banach spaces
(a) If E has cotype 2, then Π2 (E; F ) = Π1 (E; F ).
(b) If E has cotype 2 < q < ∞, then Πr (E; F ) = Π1 (E; F ) for all 1 < r < q∗.
Theorem 1.4 ([7], Kwapie´n). Let 1 ≤ p ≤ ∞. Then
Π(r(p),1) (ℓ1; ℓp) = L (ℓ1; ℓp) where
Moreover, if r < r(p), then Π(r,1) (ℓ1; ℓp) 6= L (ℓ1; ℓp).
1
r(p)
−
1
p
= 1 −(cid:12)(cid:12)(cid:12)(cid:12)
.
1
2(cid:12)(cid:12)(cid:12)(cid:12)
2. Main results
The next definition was introduced by M.C. Matos in [8] and, independently, by D. P´erez-
Garc´ıa in [9].
Definition 2.1. Let 1 ≤ p ≤ q ≤ ∞. A multilinear operator T : E1 × · · · × Em → F is multiple
(q, p)-summing if there exist a constant C > 0 such that
∞
Xj1,...,jm=1(cid:13)(cid:13)(cid:13)
j )∞
j=1 ∈ ℓw
for all (x(k)
summing operators by Πm
Πm
(q,p) (mE; F ).
T (x(1)
j1 , . . . , x(m)
1
q
≤ C
q!
jm )(cid:13)(cid:13)(cid:13)
m
Yk=1(cid:13)(cid:13)(cid:13)
(x(k)
j )∞
j=1(cid:13)(cid:13)(cid:13)w,p
p (Ek), with k = 1, ..., m. We represent the class of all multiple (q, p)-
(q,p) (E1, . . . , Em; F ). When E1 = · · · = Em = E, we denote simply by
(zj)∞
Following [1], when we write (xj)∞
j=1 ∈ ℓw
The following result of Arregui and Blasco [1, Lemma 3 and Proposition 6] is crucial for us:
s (E) such that xj = ajzj for all j.
s (E), it means that there are (aj)∞
j=1 ∈ ℓr and
j=1 ∈ ℓrℓω
Proposition 2.2. Let 1 < r < ∞. Then ℓω
In particular,
1 (E) = ℓrℓω
r∗(E) if and only if L (c0; E) = Πr (c0; E).
(a) If E has cotype 2, then ℓω
(b) If E has cotype q > 2, then ℓω
1 (E) = ℓ2ℓω
2 (E).
1 (E) = ℓrℓω
r∗(E) for any r > q.
ℓrℓω
Note that from the previous result it is immediate that if E has cotype 2 then ℓω
r∗(E) for any r ≥ 2.
Let us to present now an inclusion theorem for multiple summing multilinear operators that,
when restricted to the linear case, will be important in the proof of our main result. Before that,
let us to recall the classical inclusion theorem for absolutely summing operators [5, Theorem
10.4], which will be very useful to our propose.
1 (E) =
Theorem 2.3. If 1 ≤ aj ≤ bj for j = 1, 2 and 1
b1
(a1, b1)-summing operator is also absolutely (a2, b2)-summing.
− 1
a1
≤ 1
b2
− 1
a2
, then every absolutely
Lemma 2.4. Let E1, ..., Em and F be infinite dimensional Banach spaces.
(a) If E1, ..., Em has cotype 2 and r ≤ 2, then
Π(a,r) (E1, ..., Em; F ) ⊂ Π(s,1) (E1, ..., Em; F ) with
1
s
(b) If Ei has cotype qi > 2, q = max{qi : i = 1, ..., m} and 1 < r < q∗, then
1
s
Π(a,r) (E1, ..., Em; F ) ⊂ Π(s,1) (E1, ..., Em; F ) with
≤ 1 −
≤ 1 −
1
a
1
a
1
r
1
r
−
−
.
.
Proof. (a) If r ≤ 2, then r∗ ≥ 2 and, from Theorem 2.3, Π2(E; F ) ⊆ Πr∗(E; F ) for all Banach
spaces E and F . Thus, since Ei has cotype 2, it follows that L(c0; Ei) = Π2(c0; Ei) ⊆ Πr∗(c0; Ei),
ON THE MAUREY -- PISIER AND DVORETZKY -- ROGERS THEOREMS
3
that is, L(c0; Ei) = Πr∗(c0; Ei). From Proposition 2.2 we thus have ℓw
if (x(i)
j )∞
j=1 ∈ ℓw
1 (Ei),
1 (Ei) = ℓr∗ℓw
r (Ei). Then,
x(i)
j = a(i)
j z(i)
j
with (a(i)
j )j ∈ ℓr∗ and (z(i)
j )j ∈ ℓw
r (Ei). Since
from Holder's inequality we have
1
s
≤
1
r∗ +
1
a
,
∞
Xj1,...,jm=1(cid:13)(cid:13)(cid:13)
T (cid:16)x(1)
= ∞
Xj1,...,jm=1(cid:12)(cid:12)(cid:12)
≤ ∞
Xj1,...,jm=1(cid:12)(cid:12)(cid:12)
< ∞.
j1 ...a(m)
a(1)
a(1)
j1 ...a(m)
j1 , ..., z(m)
1
a
a!
jm (cid:17)(cid:13)(cid:13)(cid:13)
j1 , ..., x(m)
1
s
1
s
j1 , ..., z(m)
s!
jm (cid:17)(cid:13)(cid:13)(cid:13)
s!
s(cid:13)(cid:13)(cid:13)
jm (cid:17)(cid:13)(cid:13)(cid:13)
jm (cid:12)(cid:12)(cid:12)
T (cid:16)z(1)
r∗!
r∗ ∞
jm (cid:12)(cid:12)(cid:12)
Xj1,...,jm=1(cid:13)(cid:13)(cid:13)
T (cid:16)z(1)
1
(b) Follows as in (a) using Proposition 2.2(b).
(cid:3)
Remark 2.5. This theorem is a generalization of [10, Theorem 10].
Theorem 2.6. Let a, b ∈ [1, ∞) and E be an infinite dimensional Banach space.
(i) If b ≥ (cot E)∗, then
inf {a : idE is absolutely (a, b) -summing} = ∞.
(ii) (ii) If b < (cot E)∗, then
inf {a : idE is absolutely (a, b) -summing} =
b cot E
b + cot E − b cot E
.
Proof. (i) Let us to prove that idE fails to be (a, (cot E)∗)-summing; this result seems to have
been overlooked in the literature; it appears, in a more general form, in the preprint [2] and
we sketch the proof below. A famous result of Maurey and Pisier asserts that ℓcot E is finitely
representable in E, i.e., for all n ≥ 1, there exists En ⊂ E and an isomorphism Sn : ℓn
cot E → En
such that kSnk · kS −1
n k ≤ 2. There is no loss of generality in assuming kSnk ≤ 1. For 1 ≤ i ≤ n,
let
yi = Sn(ei).
Note that
1 = keikℓn
cot E
=(cid:13)(cid:13)S −1
n (yi)(cid:13)(cid:13)ℓn
cot E
≤(cid:13)(cid:13)S −1
n (cid:13)(cid:13) kyik ≤ 2 kyik ,
4
ARA ´UJO AND SANTOS
and thus kyik ≥ 1/2. Moreover
sup
ϕ∈BE ∗ n
Xi=1
hϕ, yii(cot E)∗!
1
(cot E)∗
n
= sup
ϕ∈BE ∗
sup
α∈Bℓn
cot E
αihϕ, yii
αiyii
n
hϕ,
Xi=1
Xi=1
αiyi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)E
αiSn(ei)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)E
= sup
α∈Bℓn
cot E
sup
ϕ∈BE ∗
= sup
α∈Bℓn
= sup
α∈Bℓn
= sup
α∈Bℓn
n
n
cot E(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
cot E(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
cot E
kSn (α)kE
We thus conclude that for any a ≥ 1 we have
≤ kSnk
sup
α∈Bℓn
cot E
kαkℓn
cot E
≤ 1.
n
2a ≤
Xj=1
n
n
kidE(yj)ka and sup
ϕ∈BE ∗
Xi=1
hϕ, yii(cot E)∗
≤ 1
and this means that idE is not absolutely (a, (cot E)∗)-summing. The proof of (i) is done.
Now let us prove (ii). Consider
If r < λ, then
So
λ =
b cot E
b + cot E − b cot E
.
1
r
>
1
b
− 1 +
1
cot E
.
cot E −
rb
rb + b − r
> 0.
Thus, there is 0 < ε < cot E − rb
rb+b−r such that
1
b
−
1
r
< 1 −
1
cot E − ε
.
From Lemma 2.4, with m = 1 and s = cot E − ε, we conclude that
Π(r,b) (E; E) ⊂ Π(cot E−ε,1) (E; E) 6= L (E; E)
and thus idE fails to be (r, b)-summing. So
λ ≤ inf {a : idE is absolutely (a, b) -summing} .
Now, if r > λ then
1
r
Therefore, there exist ε > 0 such that
<
1
b
− 1 +
1
cot E
.
Thus r > cot E + ε and
1
r
≤
1
b
− 1 +
1
cot E + ε
.
1 −
1
cot E + ε
≤
1
b
−
1
r
.
ON THE MAUREY -- PISIER AND DVORETZKY -- ROGERS THEOREMS
5
The inclusion theorem (Theorem 2.3) asserts that every absolutely (cot E + ε, 1)-summing oper-
ator is also absolutely (r, b)-summing. Since idE is absolutely (cot E + ε, 1)-summing, it follows
that idE is also absolutely (r, b)-summing. Consequently,
λ = inf {a : idE is absolutely (a, b) -summing} ,
and the proof is done.
(cid:3)
a
cot E
2
1
1
1
(cot E)∗
b − 1
a = 1
(cot E)∗
2
b
idE is not (a, b)-summing
idE is (a, b)-summing
Figure 1. Graphical overview of Theorem 2.6.
Remark 2.7. Note that this result completes the information of the Dvoretzky -- Rogers Theorem
(Theorem 1.2) and, when b = 1, we recover the classical result of Maurey and Pisier (Theorem
1.1).
In the previous theorem we prove that for b < (cot E)∗ we can fully characterize the classes
of (a, b)-summing operators that coincide. The Figure 1 illustrates this result and motivates us
to name the next result as "strings of coincidences", which generalizes Theorem 1.3.
Theorem 2.8 (Strings of coincidences). Let E and F be infinite dimensional Banach spaces.
(a) If E has cotype 2 and a ≤ a1 and r ≤ r1 ≤ 2, then
1
r
Π(a,r) (E; F ) = Π(a1,r1) (E; F ) with
−
(b) If a ≤ a1 and 1 ≤ r ≤ r1 < (cot E)∗, then
Π(a,r) (E; F ) = Π(a1,r1) (E; F ) with
1
r
−
1
a
1
a
=
=
1
r1
1
r1
−
1
a1
.
−
1
a1
.
Proof. Note that
(2.1)
with
Π(a,r) (E; F ) = Π(s,1) (E; F ) = Π(a1,r1) (E; F ) ,
1 −
1
s
=
1
r
−
1
a
=
1
r1
−
1
a1
.
In fact, the two equalities in (2.1) are immediate consequences of the previous theorem and the
inclusion theorem for absolutely summing operators (Theorem 2.3).
(cid:3)
6
ARA ´UJO AND SANTOS
a
2
1
1
r
2
Figure 2. Strings of coincidence for (a, r)-summability when cot E = 2.
Remark 2.9. The strings of coincidence can also be obtained as a consequence of [3, Theorem
2.1].
We finish this section by illustrating how the strings of coincidence can be useful to extend
the classical result of Kwapie´n (Theorem 1.4).
Theorem 2.10. Let 1 ≤ p ≤ ∞ and a, b ≥ 1. If b ≤ 2, then Π(a,b) (ℓ1; ℓp) = L (ℓ1; ℓp) if, and
only if, a ≥
br(p)
r(p)+b−br(p) where 1
1
p − 1
r(p) = 1 −(cid:12)(cid:12)(cid:12)
.
2(cid:12)(cid:12)(cid:12)
Proof. For b ≤ 2 the string of coincidence associated to (r(p), 1), where
composed by the pairs (a, b) such that
1
r(p) = 1 −(cid:12)(cid:12)(cid:12)
1
p − 1
, is
2(cid:12)(cid:12)(cid:12)
Thus
1
b
−
1
a
= 1 −
1
r(p)
.
a =
br(p)
r(p) + b − br(p)
.
In view of the optimality of
of a is sharp.
br(p)
r(p)+b−br(p) when b = 1, we can conclude that the above estimate
(cid:3)
References
[1] J.L. Arregui and O. Blasco, (p, q)-summing sequences, J. Math. Anal. Appl. 274 (2002), 812 -- 827.
[2] F. Bayart, D. Pellegrino and P. Rueda, On coincidence results for summing multilinear operators: interpo-
lation, ℓ1-spaces and cotype, arXiv:1805.12500v1 (2018).
[3] A.T. Bernardino, On cotype and a Grothendieck-type Theorem for absolutely summing multilinear opera-
tors, Quaestiones Math. 34 (2011), 1 -- 7.
[4] G. Botelho, D. Pellegrino and P. Rueda, Cotype and absolutely summing linear operators, Math. Z., 267
(2001), 1 -- 7.
[5] J. Diestel, H. Jarchow and A. Tonge, Absolutely summing operators, Cambridge University Press, 1995.
[6] A. Dvoretzky and C.A. Rogers, Absolute and unconditional convergence in normed spaces, Proc. Nat. Acad.
Sci. USA 36 (1950), 192 -- 197.
[7] S. Kwapie´n, Some remarks on (p, q)-absolutely summing operators in ℓp-space, Studia Math. 29 (1968),
327 -- 337.
[8] M.C. Matos, Fully absolutely summing and Hilbert-Schmidt multilinear mapping, Collect. Math. 54 (2003),
111 -- 136.
[9] D. P´erez-Garc´ıa, Operadores multilineales absolutamente sumantes, PhD Thesis, Universidad Complutense
de Madrid, 2003.
[10] D. Popa, Reverse inclusions for multiple summing operators, J. Math. Anal. Appl. 350 (2009), 360 -- 368.
ON THE MAUREY -- PISIER AND DVORETZKY -- ROGERS THEOREMS
7
(G. Ara´ujo) Departamento de Matem´atica, Universidade Estadual da Para´ıba, 58.429-600 -
Campina Grande, Brazil
E-mail address: [email protected]
(J. Santos) Departamento de Matem´atica, Universidade Federal da Para´ıba,
58.051-900 -
Joao Pessoa, Brazil
E-mail address: [email protected] or [email protected]
|
1907.05241 | 1 | 1907 | 2019-07-10T13:59:28 | Metrization of probabilistic metric spaces. Applications to fixed point theory and Arzela-Ascoli type theorem | [
"math.FA"
] | Schweizer, Sklar and Thorp proved in 1960 that a Menger space $(G,D,T)$ under a continuous $t$-norm $T$, induce a natural topology $\tau$ wich is metrizable. We extend this result to any probabilistic metric space $(G,D,\star)$ provided that the triangle function $\star$ is continuous. We prove in this case, that the topological space $(G,\tau)$ is uniformly homeomorphic to a (deterministic) metric space $(G,\sigma_D)$ for some canonical metric $\sigma_D$ on $G$. As applications, we extend the fixed point theorem of Hicks to probabilistic metric spaces which are not necessarily Menger spaces and we prove a probabilistic Arzela-Ascoli type theorem. | math.FA | math |
METRIZATION OF PROBABILISTIC METRIC SPACES.
APPLICATIONS TO FIXED POINT THEORY AND
ARZELA-ASCOLI TYPE THEOREM
MOHAMMED BACHIR, BRUNO NAZARET
Abstract. Schweizer, Sklar and Thorp proved in 1960 that a Menger space
(G, D, T ) under a continuous t-norm T , induce a natural topology τ wich is
metrizable. We extend this result to any probabilistic metric space (G, D, ⋆)
provided that the triangle function ⋆ is continuous. We prove in this case, that
the topological space (G, τ ) is uniformly homeomorphic to a (deterministic)
metric space (G, σD) for some canonical metric σD on G. As applications, we
extend the fixed point theorem of Hicks to probabilistic metric spaces which
are not necessarily Menger spaces and we prove a probabilistic Arzela-Ascoli
type theorem.
Keywords: Metrization of probabilistic metric space; Probabilistic 1-Lipschitz
map; Probabilistic Arzela-Ascoli type Theorem; Probabilistic fixed point theorem.
msc: 54E70, 46S50.
1. Introduction
Let (G, D, T ) be a Menger space equipped with a probabilistic metric D and
a t-norm T (the definitions and notation reminders will be given in the details
in Section 2). Schweizer and Sklsar [13] defined for ε, λ > 0 and each x ∈ G a
neighborhood Nx(ε, λ) as follows
Nx(ε, λ) = {y ∈ G : D(x, y)(ε) > 1 − λ}.
Schweizer, Sklar and Thorp proved in [14] that, given a t-norm T of a Menger
space (G, D, T ) satisfying 1 = supx<1 T (x, x) (in particular if T is continuous),
the collection {Nx(ε, λ) : x ∈ G} taken as a neighborhood base at x gives rise to a
metrizable topology. In [11] Morrel and Nagata proved the following two extensions:
(1) The class of topological Menger spaces coincides with that of semi-metrizable
topological spaces.
(2) No condition on T weaker than 1 = supx<1 T (x, x) can guarantee that a
Menger space, under T , is topological.
The aim of the present paper is to prove that, in a general probabilistic metric
space (G, D, ⋆), not necessarily being a Menger space, the collection {Nx(ε, λ) : x ∈
X} taken as a neighborhood base at x gives rise to a topology which is uniformly
homeomorphic to a metric space, provided that the triangle function ⋆ is continuous
(necessarily uniformly continuous by Sibley's result in [16] on the compactness of
Date: 05/06/2019.
1991 Mathematics Subject Classification. 54E70, 46S50.
1
2
MOHAMMED BACHIR, BRUNO NAZARET
(∆+, dL), where dL denotes the modified L´evy distance and ∆+ denotes the set of
all nondecreasing and left-continuous distributions that vanish at 0).
We get an even more precise result : if w⋆ : [0, +∞] → [0, +∞] is a modulus of
uniform continuity for the triangle function ⋆, then the (deterministic) metric σD
on G defined canonically from the probabilistic metric D by
∀x, y ∈ G, σD(x, y) := sup
z∈G
dL(D(x, z), D(z, y)),
satisfies the following inequalities
∀x, y ∈ G, dL(D(x, y), H0) ≤ σD(x, y) ≤ w⋆(dL(D(x, y), H0)).
(1)
Note from [12] that y ∈ Nx(t, t) if and only if dL(D(x, y), H0) < t, for all t > 0.
If moreover we assume that ⋆ is k-Lipschitz given some positive real number k
(necessarily k ≥ 1), then we can take w⋆(t) = kt for all t ≥ 0 (see Proposition 4 for
examples of such functions). As an immediate consequence of (1), the semi-metric
α(x, y) := dL(D(x, y), H0) define a topology on (G, D, ⋆) which is uniformly (resp.
Lipschitz) homeomorphic to the metric space (G, σD), whenever ⋆ is continuous
(resp. Lipschitz continuous). This result is an extension to non necessarily Menger
spaces of the works established in Menger spaces by Schweizer, Sklar and Thorp
in [14]. In particular, the formula (1) allows us to transfer several known results
from metric space theory to the probabilistic metric theory. For instance, using (1)
and the Ekeland variational principle we give some extensions of the fixed point
theorem of Hicks (see [4]), or using again (1) we give an Arzela-Ascoli type theorem
for the the space of probabilistic 1-Lipschtz maps introduced recently in [1]. Notice
that other results such as Baire theorem and all its variants/consequences can be
transfered, thanks to our result, to the probabilistic metric framework.
This paper is organized as follows. In Section 2, we recall some classical notions
related to probabilistic metric space. In Section 3, we treat the metrization of prob-
abilistic metric space and prove Theorem 1. We also give some new properties. In
Section 4, we establish fixed point theorems (Theorem 2 and Theorem 3) extending
a result of Hicks (see [4]).
In Section 5, we prove Theorem 5, showing that the
set of probabilistic 1-Lipschitz maps introduced in [1] is a compact space for the
uniform convergence, giving a probabilistic Arzela-Ascoli theorem.
2. Definitions and notation
In this section, we recall some known facts about probabilistic metric spaces,
the modified L´evy distance and the weak convergence. All these notions can be
found in [12], [5] and [6]. We also recall the notion of probabilistic 1-Lipschitz map
introduced in [1], which shall play an important role in the sequel.
2.1. Probabilistic metric space and triangle function. By ∆+ we denote
the set of all (cumulative) distribution functions F : [−∞, +∞] −→ [0, 1], nonde-
creasing and left-continuous with F (−∞) = 0; F (+∞) = 1 and F (0) = 0. For
a ∈ [0, +∞[, we denote Ha(t) = 0 if t ≤ a and Ha(t) = 1, if t > a.
In the sequel, we shall write F ≤ G for
which defines an ordering relation on ∆+.
∀t ∈ R, F (t) ≤ G(t),
METRIZATION OF PROBABILISTIC METRIC SPACES
3
Definition 1. ([12, 4, 5, 6]) A binary operation ⋆ on ∆+ is called a triangle function
if and only if it is commutative, associative, non-decreasing in each place, and has
H0 as neutral element. In other words:
(i) F ⋆ L ∈ ∆+ for all F, L ∈ ∆+.
(ii) F ⋆ L = L ⋆ F for all F, L ∈ ∆+.
(iii) F ⋆ (L ⋆ K) = (F ⋆ L) ⋆ K, for all F, L, K ∈ ∆+.
(iv) F ⋆ H0 = F for all F ∈ ∆+.
(v) F ≤ L =⇒ F ⋆ K ≤ L ⋆ K for all F, L, K ∈ ∆+.
Definition 2. A t-norm is a function T : [0, 1] × [0, 1] → [0, 1], usually called a
triangular norm (see [12, 4, 5, 6]), satisfying
• T (x, y) = T (y, x) ( commutativity);
• T (x, T (y, z)) = T (T (x, y), z) (associativity);
• T (x, y) ≤ T (x, z) whenever y ≤ z (monotonicity );
• T (x, 1) = x (boundary condition).
Definition 3. A probabilistic metric space (G, D, ⋆) (an PM-space) is a set G
together with a triangle function ⋆ and a function D : G × G → ∆+ satisfying:
(i) D(x, y) = H0 iff x = y.
(ii) D(x, y) = D(y, x) for all x, y ∈ G
(iii) D(x, y) ⋆ D(y, z) ≤ D(x, z) for all x, y, z ∈ G
Usually, D(x, y) is denoted by Fx,y in the literature. A probabilistic metric space
(G, D, ⋆) is called a Menger space and denoted by (G, D, T ), iff the triangle function
⋆ := ⋆T is defined from a t-norm T as follows: for all F, L ∈ ∆+ and for all t ∈ R,
(F ⋆T L)(t)
:= sup
u+v=t
T (F (u), L(v))
(2)
=
sup
T (F (u), L(v))
u,v≤0:u+v=t
2.2. L´evy distance and weak convergence.
Definition 4. Let F and G be in ∆+. For any h > 0 we set
Ah
F,G = {t ≥ 0 st. G(t) ≤ F (t + h) + h} .
The modified L´evy distance is the map dL defined on ∆+ × ∆+ as
G,F(cid:9) .
dL(F, G) = inf(cid:8)h > 0 st. [0, h−1[⊂ Ah
F,G ∩ Ah
Notice that, for all F , G ∈ ∆+,
(i) if F ≤ G then AG,F = [0, +∞[, hence
dL(F, G) = inf (cid:8)h > 0 st. [0, h−1[⊂ Ah
F,G = Ah
F,G(cid:9) .
(ii) if h ≥ 1, Ah
(iii) The usual Levy distance between general cumulative distribution functions
G,F = [0, +∞[, hence dL(F, G) ≤ 1.
can be expressed as
inf (cid:8)h > 0 st. Ah
F,G = Ah
It is invariant under the action of translations which, as we shall see later,
is not the case for the modified version since it somehow does not see the
behaviour at infinity.
G,F = [0, +∞[(cid:9) .
4
MOHAMMED BACHIR, BRUNO NAZARET
Definition 5. Let ⋆ be a triangle function on ∆+.
(1) A sequence (Fn) of distributions in ∆+ converges weakly to a function F in
∆+ if (Fn(t)) converges to F (t) at each point t of continuity of F . In this case, we
write indifferently Fn
w−→ F or limn Fn = F .
(2) We say that the law ⋆ is continuous at (F, L) ∈ ∆+ × ∆+ if we have Fn ⋆
w−→ F ⋆ L, whenever Fn
w−→ F and Ln
w−→ L.
We recall the following results due to D. Sibley in [16, Theorem 1. and Theorem
Ln
2].
Lemma 1. ([16, 12]) The function dL is a metric on ∆+ and (∆+, dL) is compact.
Lemma 2. ([16, 12]) Let (Fn) be a sequence of functions in ∆+, and let F be an
element of ∆+. Then (Fn) converges weakly to F if and only if dL(Fn, F ) −→ 0,
when n −→ +∞.
Remark 1. Thanks to Lemma 2, we shall indifferently use the notations Fn
or dL(Fn, F ) −→ 0 to say that (Fn) converges weakly to F .
w−→ F
Definition 6. Let (G, D, ⋆) be a probabilistic metric space. For x ∈ G and t > 0,
the strong t-neighborhood of x is the set
and the strong neighborhood system for G is {Nx(t); x ∈ G, t > 0}.
Nx(t) = {y ∈ G : D(x, y)(t) > 1 − t},
Lemma 3. ([12, Lemme 4.3.3]) Let t > 0 and x, y ∈ G. Then we have y ∈ Nx(t)
if and only if dL(D(x, y), H0) < t.
2.3. Probabilistic 1-Lipschitz map.
Definition 7. Let (G, D, ⋆) be a probabilistic metric space and let f be a function
f : (G, D, ⋆) −→ (∆+, dL). We say that f is a probabilistic 1-Lipschitz map if :
∀x, y ∈ G, D(x, y) ⋆ f (y) ≤ f (x).
We can also define probabilistic k-Lipschitz maps for any nonegative real number
k ≥ 0 as the maps f satisfying
∀x, y ∈ G, Dk(x, y) ⋆ f (y) ≤ f (x),
where, for all x, y ∈ G and all t ∈ R, Dk(x, y)(t) = D(x, y)( t
k ) if k > 0 and
D0(x, y)(t) = H0(t) if k = 0. For sake of simplicity, when we use the notion in
Definition 7, we shall only treat in this paper the case of probabilistic 1-Lipschitz
maps, but our main result result could be easily extended to this more general
setting.
Examples 1. Let (G, d) be a metric space. Assume that ⋆ is a triangle function on
∆+ satisfying Ha ⋆ Hb = Ha+b for all a, b ∈ R+ (for example if ⋆ = ⋆T where T is
a lef-continuous triangular norm). Let (G, D, ⋆) be the probabilistic metric space
defined with the probabilistic metric
Let L : (G, d) −→ R+ be a real-valued map. Then, L is a non-negative 1-Lipschitz
map if and only if f : (G, D, ⋆) −→ ∆+ defined for all x ∈ G by
D(p, q) = Hd(p,q).
f (x) := HL(x)
METRIZATION OF PROBABILISTIC METRIC SPACES
5
is a probabilistic 1-Lipschitz map. This example shows that the framework of
probabilistic 1-Lipschitz maps encompasses the classical determinist case.
By Lip1
⋆(G, ∆+) we denote the space of all probabilistic 1-Lipschitz maps
Lip1
⋆(G, ∆+) := {f : G −→ ∆+/D(x, y) ⋆ f (y) ≤ f (x); ∀x, y ∈ G}.
For all x ∈ G, by δx we denote the map
δx : G −→ ∆+
y
7→ D(y, x).
It follows from the properties of the probabilistic metric D that δx is a prob-
abilistic 1-Lipschitz for every x ∈ G. We set G(G) := {δx, x ∈ G} and by δ, we
denote the operator
δ : G −→ G(G) ⊂ Lip1
⋆(G, ∆+)
x
7→ δx.
2.4. Modulus of uniform continuity of a triangle function on ∆+. Let
⋆ : ∆+ × ∆+ → ∆+ be a continuous triangle function (with respect to the modified
L´evy distance dL). Since (∆+, dL) is a compact metric space (see Lemma 1) and
⋆ is continuous, then ⋆ is uniformly continuous from ∆+ × ∆+ into ∆+. Let
ω⋆ : [0, +∞] → [0, +∞] be a modulus of uniform continuity for ⋆ (limt→0 ω⋆(t) =
ω⋆(0) = 0), that is for all (F, L), (F ′, L′) ∈ ∆+ × ∆+
dL(F ⋆ L, F ′ ⋆ L′) ≤ ω⋆(dL(F, F ′) + dL(L, L′)).
In particular for all (F, L) ∈ ∆+ × ∆+
dL(F ⋆ L, L) ≤ ω⋆(dL(F, H0)).
(3)
If moreover the operation ⋆ is k-Lipschitz (with respect to dL) for some positive
number k then ω⋆(t) = kt for all t ≥ 0 (necessarily k ≥ 1, by using 3 with L = H0) is
a modulus of uniform continuity. We give in Proposition 4 examples of k-Lipschitz
triangle function using k-Lipschitz t-norms.
3. Metrization of Probabilistic Metric space.
We give below the main result of this section, that is a metrization of probabilistic
metric space extending the result of Schweizer, Sklar and Thorp in [14].
Let (G, D, ⋆) be a probabilistic metric space. We define canonically the metric
σD on G using the probabilistic metric D as follows: for all x, y ∈ G
σD(x, y) := sup
z∈K
dL(D(x, z), D(y, z)) := sup
z∈K
dL(δx(z), δy(z)) := d∞(δx, δy)
It is easy to see that σD is a metric on G and that for all x, y ∈ G
dL(D(x, y), H0) ≤ σD(x, y).
Theorem 1. Let (G, D, ⋆) be a probabilistic metric space such that ⋆ is continuous
(resp. k-lipschitz). Let ω⋆ be a modulus of uniform continuity of ⋆ on ∆+. Then,
the metric σD satisfies: for all x, y ∈ G
dL(D(x, y), H0) ≤ σD(x, y) ≤ ω⋆(dL(D(x, y), H0)).
(resp. dL(D(x, y), H0) ≤ σD(x, y) ≤ kdL(D(x, y), H0)).
6
MOHAMMED BACHIR, BRUNO NAZARET
In particular, the identity map i : (G, τ ) → (G, σD) is an uniform homeomorphism,
where τ is the topology induced by the strong neighborhood system {Nx(t); x ∈ G, t >
0} (see Definition 6).
This theorem is a mere consequence of the following lemma, that we will also
use for proving Theorem 5.
Lemma 4. Let (G, D, ⋆) be a probabilistic metric space such that ⋆ is continuous.
Let ω⋆ be a modulus of uniform continuity of ⋆ on ∆+. Then, the set Lip1
⋆(G, ∆+)
is uniformly equicontinuous. More precisely, we have ∀x, y ∈ G :
sup
⋆(G,∆+)
f ∈Lip1
dL(f (x), f (y)) ≤ ω⋆(dL(D(x, y), H0)).
Proof. From the formula (3) about the modulus of uniform continuity of ⋆, we have
that ∀L ∈ ∆+, ∀x, y ∈ G :
dL(D(x, y) ⋆ L, L) ≤ ω⋆(dL(D(x, y), H0)).
In particular, we have for all f ∈ Lip1
⋆(G, ∆+) and all x, y ∈ G,
max[dL(D(x, y) ⋆ f (x), f (x)), dL(D(x, y) ⋆ f (y), f (y))] ≤ ω⋆(dL(D(x, y), H0)),
hence, it is enough to prove that
dL(f (x), f (y)) ≤ max[dL(D(x, y) ⋆ f (x), f (x)), dL(D(x, y) ⋆ f (y), f (y))].
Let h1, h2 > 0 such that,
[0, h−1
1 [⊂ Ah1
D(x,y)⋆f (x),f (x) ∩ Ah1
f (x),D(x,y)⋆f (x).
that is, for all t ∈]0, h−1
[0, h−1
D(x,y)⋆f (y),f (y) ∩ Ah2
2 [⊂ Ah2
1 [ and all t′ ∈]0, h−1
2 [, we have
f (y),D(x,y)⋆f (y).
(4)
(5)
0 ≤ D(x, y) ⋆ f (x)(t) ≤ f (x)(t + h1) + h1
0 ≤ f (x)(t) ≤ D(x, y) ⋆ f (x)(t + h1) + h1
0 ≤ D(x, y) ⋆ f (y)(t′) ≤ f (y)(t′ + h2) + h2
0 ≤ f (y)(t′) ≤ D(x, y) ⋆ f (y)(t′ + h2) + h2.
From the second, the fourth inequalities and the fact that f is 1-Lipschitz, we get
that for all t ∈]0, h−1
1 [ and all t′ ∈]0, h−1
2 [
0 ≤ f (x)(t) ≤ f (y)(t + h1) + h1
0 ≤ f (y)(t′) ≤ f (x)(t′ + h2) + h2.
It follows that for all s ∈]0, max(h1, h2)−1[ (a subset of ]0, min(h−1
1 , h−1
2 )[)
0 ≤ f (x)(s) ≤ f (y)(s + max(h1, h2)) + max(h1, h2)
0 ≤ f (y)(s) ≤ f (x)(s + max(h1, h2)) + max(h1, h2).
Thus, we have that dL(f (x), f (y)) ≤ max(h1, h2) for all h1, h2 > 0 satisfying (4)
and (5). This implies that
dL(f (x), f (y)) ≤ max(dL(D(x, y) ⋆ f (x), f (x)), dL(D(x, y) ⋆ f (y), f (y))),
and the conclusion.
Let us now prove Theorem 1.
(cid:3)
METRIZATION OF PROBABILISTIC METRIC SPACES
7
Proof of Theorem 1. The inequality at the left is a direct consequence of the defi-
nition of σD. To prove the inequality at the right, we use Lemma 4 noticing that
G(G) := {δx/x ∈ G} ⊂ Lip1
⋆(G, ∆+).
The second part of the theorem follows from Lemma 3 since, y ∈ Nx(t) if and only
if dL(D(x, y), H0) < t for each t > 0 and x, y ∈ G.
(cid:3)
The notion of probabilistic distance naturally leads to associated metric concepts,
such as Cauchy sequence, completeness, separability, density and compatness.
Definition 8. A complete probabilistic metric space (K, D, ⋆) is called compact if
for all t > 0, the open cover {Nx(t) : x ∈ K} has a finite subcover.
Definition 9. In a probabilistic metric space (G, D, ⋆), a sequence (zn) ⊂ G is said
to be a Cauchy sequence if for all t ∈ R,
lim
n,p−→+∞
D(zn, zp)(t) = H0(t).
(Equivalently, if D(zn, zp) w−→ H0 or dL(D(zn, zp), H0) → 0, when n, p −→ +∞). A
probabilistic metric space (G, D, ⋆) is said to be complete if every Cauchy sequence
(zn) ⊂ G weakly converges to some z∞ ∈ G, that is limn→+∞ D(zn, z∞)(t) = H0(t)
for all t ∈ R, we will briefly note limn D(zn, z∞) = H0.
Corollary 1. Let (G, D, ⋆) be a probabilistic metric space such that ⋆ is continuous.
Then, the following assertions hold.
(1) (G, D, ⋆) is a probabilistic complete metric space iff (G, σD) is a complete
metric space.
(2) (G, D, ⋆) is compact as probabilistic metric space iff (G, σD) is a compact
metric space.
(3) (G, D, ⋆) is separable as probabilistic metric space iff (G, σD) is separable
metric space.
Proof. It is a direct consequence of Theorem 1 using Lemma 3.
(cid:3)
Notice that several results in the litterature proved for probabilistic metric spaces
could be easily deduced from Corollary 1 and Theorem 1. For instance, recall that
a Baire space is a topological space such that every intersection of a countable
collection of open dense sets is also dense.
In [15], H. Sherwood proved that a
complete Menger space under a continuous t-norm, equipped with the topology
τ induced by the strong neighborhood system {Nx(t); x ∈ G, t > 0} is a Baire
space. Now, Theorem 1 expressing the fact that as soon as the triangle function
is continuous then the induced topology is metrizable, we immediately obtain the
following result.
Proposition 1. Let (G, D, ⋆) be a probabilistic complete metric space such that ⋆ is
continuous. Let τ be the topology induced by strong neighborhood system {Nx(t); x ∈
G, t > 0} (see Theorem 1). Then, (G, τ ) is a Baire space.
In the same spirit, we also easily recover the following proposition already proven
by other means in [10, Theorem 2.2, Theorem 2.3].
Proposition 2. Let (K, D, ⋆) be a probabilistic metric space. Suppose that the
triangle function ⋆ is continuous. Then,
8
MOHAMMED BACHIR, BRUNO NAZARET
(1) (K, D, ⋆) is compact as probabilistic metric space iff every sequence of K has
a convergent subsequence.
(2) If (K, D, ⋆) is compact as probabilistic metric space, then it is separable.
We end the section by showing that the metric σD is canonical in the following
sens. We know that every (complete) metric space induce a probabilistic (complete)
metric space. Indeed, if d is a (complete) metric on G and ⋆ is a triangle function
on ∆+ satisfying Ha ⋆ Hb = Ha+b for all a, b ∈ R+ (see references [12] and [4]),
then (G, D, ⋆) is a probabilistic (complete) metric space, where
D(p, q) = Hd(p,q), ∀p, q ∈ G.
Using Proposition 3 below, we get that
dL(D(p, q), H0) ≤ σD(p, q) := sup
z∈G
dL(D(p, z), D(z, q)) = sup
z∈G
≤ sup
z∈G
dL(Hd(p,z), Hd(z,q))
min(1, d(p, z) − d(z, q))
= min(1, d(p, q))
= dL(Hd(p,q), H0)
= dL(D(p, q), H0)
Thus, we have the equality
dL(D(p, q), H0) = σD(p, q) = min(1, d(p, q)).
It follows that σD(p, q) = d(p, q), for all p, q ∈ G such that d(p, q) ≤ 1. In particular,
σD and d coincides if (G, d) is of diameter less than 1.
Proposition 3. Let a, b ≥ 0. Then,
dL(Ha, Hb) = min(cid:18)1, b − a,
1
min(a, b)(cid:19) ,
and, in particular,
dL(Ha, Hb) ≤ min (1, b − a) = dL(Hb−a, H0).
(6)
Notice that the inequality (6) expresses the more general fact that, for all λ > 0
and for all F , G ∈ ∆+,
which is a consequence of the following property,
dL(τλF, τλG) ≤ dL(F, G),
∀λ > 0, (cid:8)h > 0, [0, h−1[⊂ Ah
F,G(cid:9) ⊂ (cid:8)h > 0, [0, h−1[⊂ Ah
τλF,τλG(cid:9) ,
where τλF (t) = F (t − λ). This contraction property is an equality for the standard
Levy metric while Proposition (3) shows that it is not true for the modified version
dL.
Proof. In this proof, we will assume without loss of generality that a < b and use
the shortened notation
Ah
a,b := Ah
Ha,Hb = {t ≥ 0, Ha(t) ≤ Hb(t + h) + h} ,
since in this case we have Ha ≥ Hb. Notice that the inequality
Ha(t) ≤ Hb(t + h) + h
METRIZATION OF PROBABILISTIC METRIC SPACES
9
is immediate for t ∈ [0, a], while if t > a and since h < 1, it is equivalent to
that is t + h > b. As a consequence,
Hb(t + h) ≥ 1 − h > 0,
Ah
a,b = [0, a] ∪ (]a, +∞[∩]b − h, +∞[) = [0, a]∪] max(a, b − h), +∞[.
We then have 2 cases :
• If a > 1, then for all h ≥ a−1, [0, h−1[⊂ [0, a] ⊂ Ah
In addition, if
a,b if and only if b − h ≤ a, that is h ≥ b − a.
a,b.
h < a−1, then [0, h−1[⊂ Ah
This leads to
(cid:8)h > 0, [0, h−1[⊂ Ah
a,b(cid:9) = [a−1, +∞[∪[b − a, +∞[= [min(a−1, b − a), +∞[,
hence in this case, dL(Ha, Hb) = min(a−1, b − a) = min(1, a−1, b − a).
• If a ≤ 1, we have h1 > a for all h ∈]0, 1[, hence [0, h−1[⊂ Ah
a,b if and only
if b − h ≤ a, that is if h ≥ b − a. It follows that
(cid:8)h > 0, [0, h−1[⊂ Ah
a,b(cid:9) = [1, +∞]∪(]0, 1[∩[min(1, b − a), +∞[) = [min(1, b−a), +∞[,
hence, in this case, dL(Ha, Hb) = min(1, b − a) = min(1, a−1, b − a).
This concludes the proof.
(cid:3)
4. Fixed point and contraction
This section is divided on two subsections. In Subsection 4.1, we give two new
fixed point theorems and in Subsection 4.2, we give some general examples of k-
Lipschitz triangle functions constructed canonically from k-Lipschitz t-norms.
4.1. Fixed point theorem. Let us start from the following probabilistic notion
of contraction introduced by Hicks (see, [4]).
Definition 10. Let (G, D, ⋆) be a probabilistic metric space. A map f : G → G is
said to be a C-contraction if there exists q ∈ (0, 1) such that for every x, y ∈ G and
every t > 0
D(x, y)(t) > 1 − t =⇒ D(f (x), f (y))(qt) > 1 − qt.
Lemma 5. A map f : G → G is a C-contraction with constant q iff for all x, y ∈ G,
dL(D(f (x), f (y)), H0) ≤ qdL(D(x, y), H0).
Proof. From Lemma 3, we have that for every x, y ∈ G, D(x, y)(t) > 1 − t if and
only if dL(D(x, y), H0) < t. For every ε > 0, set tε = dL(D(x, y), H0) + ε > 0.
Then, D(x, y)(tε) > 1 − tε. Suppose that f is a C-contraction, then we have that
D(f (x), f (y))(qtε) > 1 − qtε which is equivalent to
dL(D(f (x), f (y), H0) ≤ qtε = q(dL(D(x, y), H0) + ε).
Sending ε to 0, we get dL(D(f (x), f (y), H0) ≤ qdL(D(x, y), H0). The converse is
straightforward.
(cid:3)
Hicks proved that a C-contraction map in Menger space under the minimum
t-norm TM (a, b) = min(a, b) has a unique fixed point. We can find a extension of
this result for generalised C-contraction in Menger space in [4]. We introduce the
following new definition of contraction.
10
MOHAMMED BACHIR, BRUNO NAZARET
Definition 11. Let (G, D, ⋆) be a probabilistic metric space. Suppose that ⋆ is a
continuous triangle function (hence uniformly continuous) and let ω⋆ be a modulus
of uniform continuity of ⋆. A map f : G → G is said to be a ω⋆-contraction if there
exists q ∈ (0, 1) such that for every x, y ∈ G
ω⋆[dL(D(f (x), f (y)), H0)] ≤ qω⋆[dL(D(x, y), H0)].
Remark 2. Using Lemma 5, the notion of ω⋆-contraction concides with the C-
contraction, when the triangle function ⋆ is k-Lipschitz since in this case ω⋆(t) = kt
for all t ≥ 0 is a modulus of uniform continuity. Examples of k-Lipschitz triangle
functions are given in Proposition 4. The original result of Hicks is a particular
case corresponding to the 1-Lipschitz triangle function ⋆TM .
Using Theorem 1 and the Ekeland variational principle, we give below an exten-
sion of the result of Hicks in probabilistic metric spaces which are not necesarily
Menger spaces, where the triangle function ⋆ is continuous. Notice that this result
seems to be new even in the non probabilistic setting.
Theorem 2. Let (G, D, ⋆) be a probabilistic complete metric space, where ⋆ is
continuous triangle function with modulus of uniform continuity ω⋆. Let f : G → G
be a ω⋆-contraction with a constant of contraction q ∈ (0, 1). Then, f has a unique
fixed point x∗ ∈ G.
Proof. By assumption, we have for all x, y ∈ G
ω⋆(dL(D(f (x), f (y)), H0)) ≤ qω⋆(dL(D(x, y), H0)).
Let us consider the function φ : (G, σD) → R defined by φ(x) = ω⋆(dL(D(x, f (x)), H0))
and prove that φ is continuous. Indeed, for x, y ∈ G, from the triangle inequality
for dL, the definition of σD and Theorem 1 we have
dL(D(x, f (x)), H0) − dL(D(y, f (y)), H0) ≤ dL(D(x, f (x)), D(f (x), y))
+dL(D(f (x), y), D(y, f (y)))
≤ σD(x, y) + σD(f (x), f (y))
≤ σD(x, y) + ω⋆(dL(D(f (x), f (y)), H0))
≤ σD(x, y) + qω⋆(dL(D(x, y), H0)).
From the continuity of ω⋆, we have that
dL(D(x, y), H0) ≤ σD(x, y) → 0 =⇒ ω⋆(dL(D(x, y), H0)) → 0.
Thus, from the above inequalities, the function x 7→ dL(D(x, f (x)), H0) is con-
tinuous from (G, σD) into R, and by composing it with the uniformly continuous
function ω⋆, we get that φ is continuous.
Now, by the Ekeland variational principle [3] (since (G, σD) is a complete metric
space by Corollary 1), let ε > 0 and u ∈ G such that φ(u) ≤ inf G φ + ε. Then, for
all λ > 0 there exists v ∈ G :
(i) φ(v) ≤ φ(u) ;
(ii) σD(u, v) ≤ λ;
(iii) for all x ∈ G x 6= v, φ(v) < φ(x) + ε
Now, let us choose ε < 1 − q and set λ = 1. Using Theorem 1 and (iii), we have
λ σD(x, v).
that
φ(v) ≤ φ(x) + εω⋆(dL(D(x, v), H0)), for all x ∈ G.
(7)
METRIZATION OF PROBABILISTIC METRIC SPACES
11
We claim that x∗ := f (v) is the unique fixed point of f . Indeed, we have
ω⋆(dL(D(x∗, f (x∗)), H0)) = ω⋆(dL(D(f (v), f (x∗)), H0)) ≤ qω⋆(dL(D(v, x∗)).
(8)
By (7) with x∗ and v, we have for all x ∈ G
ω⋆(dL(D(v, x∗), H0)) = ω⋆(dL(D(v, f (v)), H0))
≤ ω⋆(dL(D(x∗, f (x∗)), H0)) + εω⋆(dL(D(x∗, v), H0)).(9)
Combining (8) and (9), we get
ω⋆(dL(D(v, x∗), H0)) ≤ (q + ε)ω⋆(dL(D(v, x∗), H0)).
Since q+ε < 1, we obtain that ω⋆(dL(D(v, x∗), H0)) = 0, which implies by Theorem
1 that σD(v, x∗) = 0, that is x∗ = v. Thus, f (x∗) = f (v) =: x∗. The unicity of the
fixed point x∗ is immediate from q < 1.
(cid:3)
Applying the Banach fixed point we give the following extenstion of Hicks's result
with an estimation of convergence of sequences xn+1 = f (xn). Note that we recover
the Hicks's result with ⋆ = ⋆TM which is k-Lipschitz, with k = 1.
Theorem 3. Let (G, D, ⋆) be a probabilistic complete metric space, where ⋆ is
k-Lipschitz triangle function (k ≥ 1). Let f : G → G be a C-contraction with
a constant of contraction q ∈]0, 1
k [. Then, f has a unique fixed point x∗ ∈ G.
Moreover, every sequence (xn) of G such that xn+1 = f (xn), satisfies: for every
t > 0
D(x1, x0)(t) > 1 − t =⇒ D(xn, x∗)(
k(kq)n
1 − kq
t) > 1 −
k(kq)n
1 − kq
t,
or equivalently,
dL(D(xn, x∗), H0) ≤
k(kq)n
1 − kq
dL(D(x1, x0), H0).
In particular, dL(D(xn, x∗), H0) → 0, when n → +∞.
Proof. By Lemma 5, we have that dL(D(f (x), f (y), H0) ≤ qdL(D(x, y), H0), for all
x, y ∈ G. Using Theorem 1, we get
σD(f (x), f (y)) ≤ qkσD(x, y).
Since qk < 1, we can apply the Banach fixed point theorem in the complete metric
space (G, σD). Thus, we obtain a unique fixed point x∗ such that, for all n ∈ N
σD(xn, x∗) ≤
(kq)n
1 − kq
σD(x1, x0).
Using again Theorem 1 (the k-Lipschitz part) we give
dL(D(xn, x∗), H0) ≤
k(kq)n
1 − kq
dL(D(x1, x0), H0),
which is equivalent by Lemma 5 to: for all t > 0
D(x1, x0)(t) > 1 − t =⇒ D(xn, x∗)(
k(kq)n
1 − kq
t) > 1 −
k(kq)n
1 − kq
t.
(cid:3)
12
MOHAMMED BACHIR, BRUNO NAZARET
4.2. k-Lipschitz triangle function. One of the standard way to construct trian-
gle function goes in the following way. We refer to [4] for more details.
Definition 12. We denote by L the set of all binary operators L on [0, +∞[ which
satisfy the following conditions:
(i) L maps [0, +∞[2 to [0, +∞[
(ii) L is non-deceasing in both coordinate
(iii) L is continuous on [0, +∞[2.
For a t-norm T , we define the operation ⋆T,L from ∆+ × ∆+ to ∆+ as follows:
for every F, G ∈ ∆+ and every t ≥ 0
(F ⋆T,L G)(t) = sup
T (F (u), G(v)).
L(u,v)=t
In the speciale case where L(u, v) = u + v we obtain ⋆T,L = ⋆T .
Theorem 4.
commutative, associative, has 0 as identity and satisfy the condition
([4, Theorem 2.15]) if T is a left-continuous t-norm and L ∈ L is
if u1 < u2 and v1 < v2 then L(u1, v1) < L(u2, v2),
then, ⋆T,L is a triangle function.
The above theorem works for example with L(u, v) := L+(u, v) = u + v or
L(u, v) := LM (u, v) = max(u, v).
Another way to construct a triangle function from a t-norm T is the use the
t-conorm T ∗(u, v) = 1 − T (1 − u, 1 − v) as follows : for every F, G ∈ ∆+ and for
every s > 0
(F ⋆T ∗ G)(s) = inf
u+v=s
T ∗(F (u), G(v)).
Recall that a t-norm T : [0, 1]×[0, 1] → [0, 1] is k-Lipschitz if there exists k ∈ [0, +∞[
such that, for all a, b, c, d ∈ [0, 1], we have
T (a, b) − T (c, d) ≤ k(a − c + b − d).
Since T (x, 1) = x, we necessarily have that k ≥ 1. Note also that the minimum
t-norm TM (a, b) := min(a, b) is 1-Lipschitz. Other examples of k-Lipschitz t-norms
are studied in [7, 8, 9]. In order to give examples of k-Lipschitz triangle functions
in Proposition 4, we need the following lemma.
Lemma 6. Let T : [0, 1] × [0, 1] → [0, 1] be a k-Lipschitz t-norm. Then, for every
(a, b), (c, d) ∈ [0, 1]×[0, 1] and every h ∈ [0, +∞[ such that a ≤ c+h1 and b ≤ d+h2
we have that
T (a, b) − T (c, d) ≤ k(h1 + h2)
T ∗(a, b) − T ∗(c, d) ≤ k(h1 + h2).
Proof. Four cases are discussed.
case 1. If a ≤ c and b ≤ d. In this case, since T is a t-norm, then
T (a, b) − T (c, d) ≤ 0 ≤ k(h1 + h2).
case 2. If a ≤ c and b ≥ d. In this case, since T is a t-norm, then T (a, b) ≤ T (c, b)
and so since it is k-Lipschitz we have that
T (a, b) − T (c, d) ≤ T (c, b) − T (c, d) ≤ kb − d = k(b − d) ≤ kh1 ≤ k(h1 + h2).
METRIZATION OF PROBABILISTIC METRIC SPACES
13
case 3. If a ≥ c and b ≤ d. This case is similar to case 2.
case 4. If a ≥ c and b ≥ d. In this case, since T is k-Lipschitz we have that
T (a, b) − T (c, d) ≤ k(a − c + b − d) = k(a − c + b − d) ≤ k(h1 + h2).
The case of T ∗ comes easily from the case of T .
(cid:3)
In the following proposition, we consider the cases where L(u, v) := L+(u, v) =
u + v and L(u, v) := LM (u, v) = max(u, v).
Proposition 4. Let T : [0, 1] × [0, 1] → [0, 1] be a k-Lipschitz t-norm (k ≥ 1).
Then, the triangle functions ⋆T ; ⋆T,LM ; ⋆T ∗ and T are k-Lipschitz , where for all
F, G ∈ ∆+ and all s > 0
(F ⋆T G)(s) = sup
u+v=s
T (F (u), G(v)),
(F ⋆T ∗ G)(s) = inf
u+v=s
T ∗(F (u), G(v)),
(F ⋆T,LM G)(s) =
sup
T (F (u), G(v)),
max(u,v)=s
T(F, G)(s) = T (F (s), G(s)).
Proof. We give the prove for ⋆T , the technique is similar for the other triangle
functions. Let F, F ′, G, G′ ∈ ∆+. Let h1, h2 > 0 be such that
[0, h−1
1 [⊂ Ah1
F,F ′ ∩ Ah1
meanning that for all t ∈]0, h−1
F ′,F and [0, h−1
1 [ and all t′ ∈]0, h−1
2 [⊂ Ah2
2 [ we have:
G,G′ ∩ Ah2
G′,G,
(10)
0 ≤ F (t) ≤ F ′(t + h1) + h1
0 ≤ F ′(t) ≤ F (t + h1) + h1
0 ≤ G(t′) ≤ G′(t′ + h2) + h2
0 ≤ G′(t′) ≤ G(t′ + h2) + h2.
Thus, combining the first and the third (resp.
the second and the fourth) in-
equalities, and using Lemma 6, we have that for every u, v ∈]0, max(h1, h2)−1[(⊂
]0, min(h−1
1 , h−1
0 ≤ T (F (u), G(v)) ≤ T (F ′(u + h1), G′(v + h2)) + k(h1 + h2)
0 ≤ T (F ′(u), G′(v)) ≤ T (F (u + h1), G(v + h2)) + k(h1 + h2)
2 )[)
Let s ∈]0, max(h1, h2)−1[, taking the supremum over 0 ≤ u, v such that u + v = s
in the above inequalities with the fact that k ≥ 1, we get
0 ≤ (F ⋆T G)(s) ≤ (F ′ ⋆T G′)(s + (h1 + h2)) + k(h1 + h2)
≤ (F ′ ⋆T G′)(s + k(h1 + h2)) + k(h1 + h2)
0 ≤ (F ′ ⋆T G′)(s) ≤ (F ⋆T G)(s + (h1 + h2)) + k(h1 + h2)
≤ (F ⋆T G)(s + k(h1 + h2)) + k(h1 + h2)
This shows that for all h1, h2 > 0 satisfying (10), we have
dL(F ⋆T G, F ′ ⋆T G′) ≤ k(h1 + h2).
Thus, taking the infinimum over h1 and h2, we get
dL(F ⋆T G, F ′ ⋆T G′) ≤ k(dL(F, F ′) + dL(G, G′)).
14
MOHAMMED BACHIR, BRUNO NAZARET
(cid:3)
5. Probabilistic Arzela-Ascoli type theorem
This section is divided on two subsections. In subsection 5.1 we give some general
definitions of probabilistic function spaces and in subsection 5.2, we give the main
result of this section, a probabilistic Arzela-Ascoli type theorem.
5.1. The space of continuous and functions. We are going to define continuity
of functions defined from a probabilistic metric space (G, D, ⋆) to a (deterministic)
metric space (F, dF ).
Definition 13. Let (G, D, ⋆) be a probabilistic metric space, (F, dF ) be a met-
ric space and let f be a function f : (G, D, ⋆) −→ (F, dF ). We say that f is
(probabilistic) continuous at z ∈ G if dF (f (zn) f (z)) → 0 whenever D(zn, z) w−→ H0
(equivalently dL(D(zn, z), H0) → 0). We say that f is continuous if f is continuous
at each point z ∈ G.
By C⋆(G, F ) we denote the space of all (probabilistic) continuous functions
f : (G, D, ⋆) −→ (F, dF ). By C(G, F ) we denote the space of all (deterministic)
continuous functions f : (G, σD) −→ (F, dF ). We both equip the spaces C⋆(G, F )
and C(G, F ) with the uniform metric
d∞(f, g) := sup
x∈G
dF (f (x), g(x))
As in the standard case, the completeness of C⋆(G, F ) only relies on the complete-
ness of the arrival space.
Proposition 5. Let (G, D, ⋆) be a probabilistic metric space (here ⋆ is not as-
sumed to be continuous) and (F, dF ) be a complete metric space. Then, the space
(C⋆(G, F ), d∞) is a complete metric space.
Proof. Let (fn) be a Cauchy sequence in (C⋆(G, F ), d∞). In particular, for each
x ∈ G, (fn(x)) is Cauchy in (F, dF ) which is complete. Thus, there exists a function
f : G −→ F such that the sequence (fn) pointwise converges to f on G. It is easy
to see that in fact (fn) uniformly converges to f , since it is Cauchy sequence in
(C⋆(G, F ), d∞). We need to prove that f is a continuous function from (G, D, ⋆)
into (F, dF ). Let x ∈ G and (xk) be a sequence such that dL(D(xk, x), H0) −→ 0,
when k −→ +∞. For all ε > 0, there exists Nε ∈ N such that
n ≥ Nε =⇒ d∞(fn, f ) := sup
x∈G
dF (fn(x), f (x)) ≤ ε
Using the continuity of fNε, we have that there exists η(ε) > 0 such that
dL(D(xk, x), H0) ≤ η(ε) =⇒ dF (fNε(xk), fNε(x)) ≤ ε
Using (11) and (12), we have that
(11)
(12)
dF (f (xk), f (x)) ≤ dF (f (xk), fNε(xk)) + dF (fNε(xk), fNε(x)) + dF (fNε(x), f (x))
≤ 3ε
This shows that f is continuous on G. Finally, we proved that every Cauchy
sequence (fn) uniformly converges to a continuous function f . In other words, the
space (C⋆(G, F ), d∞) is complete.
(cid:3)
METRIZATION OF PROBABILISTIC METRIC SPACES
15
Since, for all x, y ∈ G,
dL(D(x, y), H0) ≤ σD(x, y),
We have in general that C⋆(G, F ) ⊂ C(G, F ). Assuming the continuity of the
triangle function ⋆, we obtain the equality.
Proposition 6. Let (G, D, ⋆) be a probabilistic metric space and (F, d) be a metric
space. Suppose that ⋆ is continuous. Then, we have C⋆(G, F ) = C(G, F ). In the
case where (F, dF ) = (∆+, dL), we also have that Lip1
⋆(G, ∆+) ⊂ C⋆(G, ∆+).
Proof. Thanks to Theorem 1, we have, for all x, y ∈ G,
σD(x, y) ≤ ω⋆ (dL(D(x, y)), H0) ,
providing the equality between C⋆(G, F ) and C(G, F ). For the second part of the
statement, we use Lemma 4 to ensure that, for any f ∈ Lip1
⋆(G, ∆+) and for any
x, y ∈ G,
dL(f (x), f (y)) ≤
sup
⋆(G,∆+)
g∈Lip1
[dL(g(x), g(y))] ≤ ω⋆ (dL(D(x, y)), H0) ,
which gives the conclusion.
(cid:3)
Remark 3. We do not know if Lip1
⋆(G, ∆+) ⊂ C⋆(G, ∆+) when ⋆ is not continuous.
⋆(K, ∆+). The following
5.2. Arzela-Ascoli type theorem for the space Lip1
proposition gives a canonical way to build probabilistic 1-Lipschitz maps from
(G, D, ⋆) into ∆+.
Definition 14. A triangle function ⋆ is said to be sup-continuous (see for instance
[2]) if for all nonempty set I and all familly (Fi)i∈I of distributions in ∆+ and all
L ∈ ∆+, we have
sup
i∈I
(Fi ⋆ L) = sup
i∈I
(Fi) ⋆ L.
Proposition 7. Let (G, D, ⋆) be a probabilistic metric space such that ⋆ is sup-
continuous. Let f : (G, D, ⋆) −→ ∆+ be any map and A be any no-empty subset of
G. Then, the map fA(x) := supy∈A[f (y) ⋆ D(x, y)], for all x ∈ G is a probabilistic
1-Lipschitz map and we have fA(x) ≥ f (x), for all x ∈ A.
Proof. The proof is similar to the standard inf-convolution construction. The fact
that fA(x) ≥ f (x) for all x ∈ A is immediate from the definition of fA. Let us now
prove that it is probabilistic 1-Lipschiptz. Let x, y ∈ G. Then, for all z ∈ A, we
have
fA(y) = sup
z∈A
[f (z) ⋆ D(y, z)] ≥ f (z) ⋆ D(y, z)
≥ f (z) ⋆ (D(y, x) ⋆ D(x, z)) = (f (z) ⋆ D(x, z)) ⋆ D(y, x).
We get the conclusion by taking the supremum with respect to z ∈ A and using
the sup-continuity of ⋆.
(cid:3)
Let us now recall the following result from [1].
Proposition 8. ([1, Proposition 3.5]) Let (Fn), (Ln), (Kn) ⊂ (∆+, ⋆). Suppose
that
(a) the triangle function ⋆ is continuous,
(b) Fn
w−→ L and Kn
w−→ F , Ln
w−→ K.
16
MOHAMMED BACHIR, BRUNO NAZARET
(c) for all n ∈ N, Fn ⋆ Ln ≤ Kn.
Then, F ⋆ L ≤ K.
Lemma 7. Let (K, D, ⋆) be a probabilistic compact metric space and (fn) be a
sequence of probabilistic 1-Lipschitz maps. Suppose that there exists a function
f defined from K into ∆+ such that, for all x ∈ K, dL(fn(x), f (x)) −→ 0, as
n → +∞. Then, f is (probabilistic) 1-Lipschitz on K and (fn) converges uniformly
to f , that is, d∞(fn, f ) −→ 0, as n → +∞.
Proof. Since each fn is a 1-Lipschitz map, we have for all x, y ∈ L and for all n ∈ N:
Using Proposition 8, we get that for all x, y ∈ L
D(x, y) ⋆ fn(x) ≤ fn(y)
D(x, y) ⋆ f (x) ≤ f (y)
In other words, f is 1-Lipschitz maps on L (Note that up to now, we have not
needed to use the compactness of K).
Now, let ε > 0 and, using Lemma 4, let η(ε) be the uniform modulus of equicon-
⋆(K, ∆+). Since (K, D, ⋆) is compact, there exists a finite set
tinuity for the set Lip1
A such that K = ∪a∈ANa(η(ε)). Since dL(fn(a), f (a)) −→ 0, as n → +∞ for all
a ∈ A. Then, for each a ∈ A, there exists Pa ∈ N such that
n ≥ Pa =⇒ dL(fn(a), f (a)) ≤ ε
Since A is finite, we have that
n ≥ max
a∈A
Pa =⇒ sup
a∈A
dL(fn(a), f (a)) ≤ ε
Thus, for all x ∈ K = ∪a∈ANa(η(ε)), there exists a ∈ A such that x ∈ Na(η(ε))
and so we have that for all n ≥ maxa∈A Pa :
dL(fn(x), f (x)) ≤ dL(fn(x), fn(a)) + dL(fn(a), f (a)) + dL(f (a), f (x))
≤ 3ε.
In other words,
n ≥ max
a∈A
Pa =⇒ d∞(fn, f ) := sup
x∈K
dL(fn(x), f (x)) ≤ 3ε
(cid:3)
We give now our main result of this section. For the classical Arzela-Ascoli
theorem we refer to the book of L. Schwartz, Analyse I, "Thorie des ensembles et
Topologie", page 346.
Theorem 5. Let (K, D, ⋆) be a probabilistic complete metric space such that ⋆ is
continuous. Then, the following assertions are equivalent.
(1) (K, D, ⋆) is compact.
(2) The metric space (Lip1
is a compact subset of (C⋆(K, ∆+), d∞) = (C(K, ∆+), d∞)).
⋆(K, ∆+), d∞) is compact (or equivalently, Lip1
⋆(K, ∆+)
Proof. • (1) =⇒ (2) Suppose that (K, D, ⋆) is compact, equivalently (K, σD) is
compact by Corollary 1. Using Lemma 4 and Theorem 1, the set Lip1
⋆(G, ∆+) is
uniformly equicontinuous with respect to the metric σD. Moreover, (∆+, dL) is
METRIZATION OF PROBABILISTIC METRIC SPACES
17
compact, hence Lip1
Ascoli theorem. On the other hand, by Lemma 7, the set Lip1
(C(G, ∆+), d∞). Hence it is compact.
⋆(G, ∆+) is relatively compact in (C(K, ∆+), d∞) by Arzela-
⋆(G, ∆+) is closed in
• (2) =⇒ (1) Suppose that (Lip1
⋆(K, ∆+), d∞) is compact. Let (xn) be a sequence
of K. We need to prove that (xn) has a convergent subsequence. Consider the
sequence (δxn) of 1-Lipschitz maps, defined by δxn : x 7→ D(xn, x) for each n ∈ N.
By assumption, there exists a subsequence (δxϕ(n)) that converges uniformly to
some 1-Lipschitz map, in particular it is a Cauchy sequence. In other words, we
have
lim
p,q−→+∞
sup
x∈K
dL(δxϕ(p) (x), δxϕ(q) (x)) = 0.
In particular we have
or equivalently,
lim
p,q−→+∞
dL(δxϕ(p) (xϕ(q)), H0) = 0,
lim
p,q−→+∞
dL(D(xϕ(p), xϕ(q)), H0) = 0.
This shows that the sequence (xϕ(n)) is Cauchy in (K, σD) (see Theorem 1). Thus,
the sequence (xϕ(n)) converges to some point x ∈ K for the metric σD, since (K, σD)
is complete. Hence, (K, σD) is compact, equivalently, (K, D, ⋆) is compact. This
ends the proof.
(cid:3)
References
1. M. Bachir, The Space of Probabilistic 1-Lipschitz map, Aequationes Math. (2019) 1-29.
2. S. Cobzas, Completeness with respect to the probabilistic Pompeiy-Hausdorff metric Studia
Univ. "Babes-Bolyai", Mathematica, Volume LII, Number 3, (2007) 43-65.
3. I. Ekeland On the variational principle, J. Math. Anal. Appl. 47 (1974), 324-353.
4. O. Hadzi´c and E. Pap Fixed Point Theory in Probabilistic Metric Space, vol. 536 of Math-
ematics and Its Applications, Kluwer Academic Publishers, Dordrecht, The Netherlands,
2001.
5. E. P. Klement, R. Mesiar, E. Pap, Triangular norms I: Basic analytical and algebraic prop-
erties, Fuzzy Sets and Systems 143 (2004), 5-26.
6. E. P. Klement, R. Mesiar, E. Pap, Triangular Norms. Kluwer, Dordrecht (2000).
7. A. Mesiarov´a: Triangular norms and k-Lipschitz property. In: Proc. EUSFLAT-LFA Con-
ference, Barcelona 2005, pp. 922-926.
8. A. Mesiarov´a, k-lp-Lipschitz t-norms, International Journal of Approximate Reasoning 46
(2007) 596-604.
9. A. Mesiarov´a, Lipschitz continuity of triangular norms, in: B. Reusch (Ed.), Computational
Intelligence, Theory and Applications (Proc. 9th Fuzzy Days in Dortmund), Springer, Berlin,
2006, pp. 309-321.
10. A. Mbarki, A. Ouahab, R. Naciri, On Compactness of Probabilistic Metric Space, Applied
Mathematical Sciences Volume(8), (2014) 1703-1710. No. 124, Paris 1954.
11. B. Morrel, J. Nagata: Statistical metric spaces as related to topological spaces, Gen. Top.
Appl. , 9 , (1978 233-237 .
12. B. Schweizer and A. Sklar, Probabilistic metric spaces, North-Holland Series in Probability
and Applied Mathematics, North-Holland Publishing Co., New York, 1983. 247 (1958), 2092-
2094.
13. B. Schweizer, A. Sklar Statistical metric spaces, Pacific. J. Math. 10, (1960) 313-334. Math.
Soc. 38 (1963), 401-406.
14. B. Schweizer, A. Sklar and E. Thorp, The metrization of statistical metric svaces, Pacific J.
Math. 10 (1960), 673-675.
15. H. Sherwood, Complete probabilistic metric spaces and random variables generated spaces,
Ph.D. Thesis, University of Arizona (1965). Math. Soc. s1-44, (1969) 441-448 Gebiete 20,
(1971) 117-128.
18
MOHAMMED BACHIR, BRUNO NAZARET
16. D. A. Sibley A metric for weak convergence of distribution functions, Rocky Mountain J.
Math. l (l971) 427-430.
Laboratoire SAMM 4543, Universit´e Paris 1 Panth´eon-Sorbonne, Centre P.M.F. 90
rue Tolbiac, 75634 Paris cedex 13, France
E-mail address: [email protected]
E-mail address: [email protected]
|
1704.06722 | 1 | 1704 | 2017-04-21T23:09:43 | On a theorem of Terzio\u{g}lu | [
"math.FA"
] | The theory of compact linear operators acting on a Banach space has such a classical core and is familiar to many. Perhaps lesser known is the characterization theorem of Terzio\u{g}lu for compact maps. In this paper we consider Terzio\u{g}lu's theorem and its consequences. We also give a similar characterization theorem in case where there is an approximation scheme on the Banach space. | math.FA | math |
ON A THEOREM OF TERZIO GLU
ASUMAN G UVEN AKSOY
Abstract. The theory of compact linear operators acting on a Banach space
has such a classical core and is familiar to many. Perhaps lesser known is
the characterization theorem of Terzioglu for compact maps.
In this paper
we consider Terzioglu's theorem and its consequences. We also give a similar
characterization theorem in case where there is an approximation scheme on
the Banach space.
1. Introduction
Let X and Y be Banach spaces and T : X → Y be an operator. We say T
is compact if and only if it maps closed unit ball BX of X into a pre-compact
subset of Y . In other words, T is compact if and only if for every norm bounded
sequence {xn} of X, the sequence {T xn} has a norm convergent subsequence in
Y . Equivalently, T is compact if and only if for every ǫ > 0, there exists elements
y1, y2, . . . , yn ∈ Y such that
T (BX) ⊆
{yk + ǫBY }
n[k=1
where by BX and BY we mean the closed unit balls of X and Y respectively.
Every compact linear operator is bounded, hence continuous, but clearly not
every bounded linear map is compact since one can take the identity operator on
an infinite dimensional space X. Compact operators are natural generalizations
of finite rank operators and thus dealing with compact operators provides us with
the closest analogy to the usual theorems of finite dimensional spaces. Recall that
L(X, Y ) denotes the normed vector space of all continuous operators from X to
Y and L(X) stands for L(X, X) and K(X, Y ) is the collection of all compact
operators from X to Y . It is well known that if Y is a Hilbert space then any
compact T : X → Y is a limit of finite rank operators, in other words if F (X, Y )
denotes the class of finite rank maps then,
K(X, Y ) = F (X, Y )
where the closure is taken in the operator norm. However, the situation is quite
different for Banach spaces, not every operator between Banach spaces is a uni-
form limit of finite rank maps. For further information we refer the reader to a
well known example due to P. Enflo [7], in which Enflo constructs a Banach space
without the approximation property. The following classical results on compact
operators will be used for our discussion later.
2010 Mathematics Subject Classification. Primary 47B07; Secondary 47B06.
Key words and phrases. Compact Operators, Approximation Schemes.
1
2
A. AKSOY
Theorem 1.1. For Banach spaces X, Y and Z, we have the following:
(1) K(X, Y ) is a norm closed vector subspace of L(X, Y ) .
(2) If X
T−−→ Z are continuous operators and either S or T is compact,
S−−→ Y
then T S is likewise compact.
If one consider the continuos operators on a Banach space X, the above theorem
asserts the fact that compact operators on X form a two sided ideal in L(X).
The following theorem of Schauder simply states that an operator is compact if
and only if its adjoint is compact.
Theorem 1.2 (Schauder). A norm bounded operator T : X → Y between Banach
spaces is compact if and only if its adjoint T ∗ : Y ∗ → X ∗ is compact.
The main idea in proving Schauder's theorem lies in the fact that
PnT − T → 0 implies
T ∗Pn − T ∗ → 0
where Pn : X → span{e1, . . . , en}. A well known proof of Schauder's theorem
may be found in Yosida [[15], p.282].
For our discussion below we also need the following characterization of the
compact sets in a Banach space; in some sense, it is a comment on the smallness
of compact sets.
Theorem 1.3 (Grothendieck). A subset of a Banach space is compact if and only
if it is included in the closed convex hull of a sequence that converges in norm to
zero.
In other words, if we have K a compact subset of a Banach space X, then we
can find a sequence {xn} in X such that
For a proof we refer the reader to[ [6] , p.3]
xn → 0 and K ⊆ co{xn}.
2. Terzioglu's Theorem
Theorem 2.1 (Terzioglu [13]). An operator T : X → Y between two Banach
spaces is compact if and only if there exists a sequence {un} of linear functionals
in X ∗ with un → 0 such that the inequality
holds for every x ∈ X.
T x ≤ sup
< un, x >
n
Proof. Suppose T : X → Y is compact, then by Schauder's theorem T ∗ : Y ∗ →
X ∗ is compact; thus by definition, if V denotes the closed unit ball of Y ∗, T ∗(V )
is a norm totally bounded subset of X ∗. Now applying Grothendieck's result, we
have a sequence {un} of elements of X ∗ with un → 0 and T ∗(V ) ⊆ co{un}. In
other words, each element of T ∗(V ) can be written of the form
αnun with
∞Xn=1
αn ≤ 1.
∞Xn=1
TERZIOGLU THM
3
Thus, for each x ∈ X we have
T x = sup
v≤1
< T ∗v, x > ≤
∞Xn=1
αn sup
< un, x > .
n
Suppose T satisfies the inequality T x ≤ supn < un, x > for some sequence
{un} ∈ X ∗. For ǫ > 0 choose N such that un < ǫ for n > N and set
Mǫ = {x ∈ X : < ui, x >= 0
for i = 1, 2, . . . N},
then one can have
T ∗(V ) ⊂ ǫU + M ⊥
ǫ
where U denotes the unit ball of X and for each linear subspace M of X, the
polar of M denoted by M is a linear subspace of X ∗ defined as:
M := {a ∈ X ∗ :
< x, a > ≤ 1
for x ∈ M} ,
this shows that T ∗ is compact and hence T is compact.
(cid:3)
An application of the Theorem 2.1 yields that every compact mapping of a
Banach space into a Pλ-space is ∞-nuclear.
Definition 2.2. We say X is a Pλ space, (λ ≥ 1) if every bounded linear operator
T from a Banach space Y to X and every Z ⊃ Y there is a linear extension T of
Z to X with
As illustrated in the following diagram:
T ≤ λT .
Z
Y
∼
T
X
T
If T = T in the above definition, we call X extendible. This property is
related to the existence of a global Hahn-Banach type an extension. J. Linden-
strauss in [9] examines the problem when is the extension T is compact if T itself
is compact and the author's results are diverse and numerous and touches upon
many related topics. Next, we define infinite nuclear mappings, this concept was
first introduced in [11].
Definition 2.3. Let X and Y be Banach spaces and T : X → Y a linear
operator. Then T is said to be infinite-nuclear, if there are sequences {un} ⊂ X ∗
and {yn} ⊂ Y such that lim
n
un = 0,
v≤1( ∞Xn=1
sup
v(yn) : v ∈ Y ∗) < +∞
and a sequence {yn} ∈ Y with sup
v≤1
form
T x =
< un, x > yn.
∞Xn=1
∞Xn=1
4
and
for x ∈ X.
A. AKSOY
T x =
∞Xn=1
< un, x > yn
As an application to Terzioglu's Theorem, under the condition that T : X → Y
where Y is a Pλ-space, Terzioglu also obtains a precise expression for T x, which
we state in the following:
Theorem 2.4 ([13]). Let T be a compact mapping of a Banach space X into a
Pλ space Y .Then for every ǫ > 0 there exists sequence {un} in X ∗ with
lim
n
un = 0
and
sup
un ≤ T + ǫ
n
< v, yn > ≤ λ such that T has the
The complete details of the proof can be found in [13]. However, it is worth
pointing out that idea of the proof provides a factorization of a compact map
through the space c0 as follows:
Use Theorem 2.1, choose the sequence {un} in X ∗ satisfying
lim
n
un = 0 and sup
n
Define linear mapping
un ≤ T + ǫ and T x ≤ sup < un, x > .
S : X → c0
by Sx = {< un, x >},
and observe that S is compact, then define a linear mapping
R0 : S(X) → Y by R0(Sx) = T x,
the inequality
R0(Sx) = T x ≤ sup < un, x > = Sx
implies that R0 ≤ 1.
c0
S(E)
∼
R
R0
F
Since Y is a Pλ space, there exists an extension R of R0 from R :c0 → F with
R ≤ λR0 = λ and
E
S−−→ c0
eR−−→ F,
evidently T = eRS.
TERZIOGLU THM
5
By considering {en} ∈ c0 and setting yn = eR(en) we obtain
< v, yn > ≤ λ, and T x =
< un, x > yn.
sup
v≤1
∞Xn=1
∞Xn=1
Using all of the above results of Terzioglu one can find the following conclusions
in [14].
Corollary 2.5.
(1) Every Pλ space has the approximation property.
(2) Every compact linear operator of an L∞ space into a Banach space is
infinite-nuclear.
(3) Let T be a compact linear map of an infinite-dimensional space X into a
Banach space Y . Then there exists an infinite dimensional closed subspace
M of X such that TM : M → T (M) is infinite nuclear.
3. Compactness with Approximation Scheme
Approximation schemes were introduced by Butzer and Scherer for Banach
spaces in 1968 [5] and later by Brudnij and Krugljak [4]. These concepts find
its best application in a paper by Pietsch [12], where he defined approximation
spaces, proved embedding, reiteration and representation results and established
connection to interpolation spaces.
Let X be a Banach space and {An} be a sequence of subsets of X satisfying:
(1) A1 ⊆ A2 ⊆ · · · ⊆ X
(2) λAn ⊆ An for all scalars λ and n = 1, 2, . . . .
(3) Am + An ⊆ Am+n for m, n = 1, 2, . . . .
For example, for 1 ≤ p < ∞ if we consider the space X = Lp[0, 1], then the
collection of sets {An} = {Lp+ 1
Pietsch's approximation spaces X ρ
considering the n-th approximation number αn(f, X), where
n
} form an approximation scheme like above.
µ (0 < ρ < ∞, 0 < µ ≤ ∞) is defined by
and
αn(f, X) := inf{f − a : a ∈ An−1}
X ρ
µ = {f ∈ X : {nρ− 1
µ αn(f, X)} ∈ ℓµ}.
In the same paper [12], embeddings, composition and commutation as well as
representation interpolation of such spaces are studied and applications to the
distribution of Fourier coefficients and eigenvalues of integral operators are given.
In the following we consider for each n ∈ N a family of subsets Qn of X
satisfying the very same three conditions stated above. For example for Qn could
be the set of all at most n-dimensional subspaces of any Banach space X, or if
our Banach space X = L(E), namely the set of all bounded linear operators on
another Banach space E, then we can take Qn = Nn(E) the set of all n-nuclear
maps on E.
Compactness relative to an approximation scheme for bounded sets and linear
operators can be studied by using Kolmogorov diameters as follows. Let D ⊂
X be a bounded subset and UX denote the closed unit ball of X. Suppose
Q = (Qn(X)n∈N) be an approximation scheme on X, then the nth Kolmogorov
6
A. AKSOY
diameter of D with respect to this scheme Q is denoted by δn(D, Q) and defined
as
δn(D, Q) = inf{r > 0 : D ⊂ rUX + A for some A ∈ Qn(X)}.
Let Y be another Banach space and T ∈ L(Y, X), then the nth Kolmogorov
diameter of T with respect to this scheme Q is denoted by δn(T, Q) and defined
as
Definition 3.1. We say D is Q-compact set if
δn(T, Q) = δn(T (UX ), Q).
and similarly T ∈ L(Y, X) is a Q-compact map, if
lim
n
δn(D, Q) = 0
lim
n
δn(T, Q) = 0.
The following example illustrates that not every Q-compact operator is com-
pact.
Example 3.2. Let {rn(t)} be the space spanned by the Rademacher functions.
It can be seen from the Khinchin inequality [10]that
ℓ2 ≈ {rn(t)} ⊂ Lp[0, 1] for all 1 ≤ p ≤ ∞.
We define an approximation scheme An on Lp[0, 1] as follows:
An = Lp+ 1
n
.
(3.1)
(3.2)
n
n+1
⊂ Lp+ 1
gives us An ⊂ An+1. for n = 1, 2, . . . , and it is easily seen that
Lp+ 1
An + Am ⊂ An+m for n, m = 1, 2, . . . , and that λAn ⊂ An for all λ. Thus {An}
is an approximation scheme.
Next, we claim that for p ≥ 2 the projection P : Lp[0, 1] → Rp is a Q-compact
map, but not compact, where Rp denotes the closure of the span of {rn(t)} in
Lp[0, 1].
Lp
P ↓
Rp
i−→ L2
↓ P2
←− R2
j
We know that for p ≥ 2, Lp[0, 1] ⊂ L2[0, 1] and R2 is a closed subspace of
L2[0, 1] and
P = j ◦ P2 ◦ i
where i, j are isomorphisms shown in the above figure. P is not a compact
operator, because dimRp = ∞, on the other hand it is a Q-compact operator
because, if we let URp, ULp denote the closed unit balls of Rp and Lp respectively,
it is easily seen that P (ULp) ⊂ kP kURp. But URp ⊂ CUR
where C is a
constant follows from the Khinchin inequality. Therefore,
P + 1
n
P (ULp) ⊂ Lp+ 1
n
,
which gives
δn(P, Q) → 0.
Next we give a characterization of Q-compact sets as subsets of the closed
convex hull of certain uniform null-sequences.
TERZIOGLU THM
7
Definition 3.3. Suppose X is a Banach space with an approximation scheme
Qn. A sequence {xn,k}k in X is called an order c0-sequence if
(1) ∀n = 1, 2, . . . there exists An ∈ Qn and a sequence {xn,k}k ⊂ An
(2) xn,k → 0 as n → ∞ uniformly in k.
Theorem 3.4. Let X be a Banach space with an approximation scheme with sets
An ∈ Qn satisfy the condition λAn ⊂ An for λ ≤ 1. A bounded subset D of X
is Q-compact if and only if there is an order c0-sequence {xn,k}k ⊂ An such that
D ⊂( ∞Xn=1
λnxn,k(n) :
xn,k(n) ∈ (xn,k)
λn ≤ 1) .
∞Xn=1
Proof of the above theorem can be obtained from the one given for p-Banach
spaces in [2]. Clearly this is an analogue of Grothendieck's theorem given above
in Theorem 1.3 for Q-compact sets.
4. Terzioglu's Theorem for Q-compact Maps
Terzioglu's characterization of compact maps relies on both Grothendieck's
and Schauder's theorems. Above Theorem 3.2 is Grothendieck's theorem for Q-
compact sets, therefore we turn our attention to the relationship between T being
Q-compact and its transpose T ∗ being Q-compact. The relationship between the
approximation numbers of T and T ∗ was studied by several authors, it is shown
in [8] that for T ∈ L(X), we have
dist(T, F ) ≤ 3 dist(T ∗, F ∗)
where F and F ∗ denote the class of all finite rank operators on X and X ∗ respec-
tively. Central to the proof of such result is the assumption of local reflexivity
possessed by all Banach spaces, (see [10]). It is not hard to show that if we assume
that our space X with approximation scheme Qn satisfies slight modification of
this property, called extended local reflexivity principle, then we have
αn(T, Q) ≤ 3 αn(T ∗, Q∗).
Where, by αn(T, Q) we mean the nth approximation number defined as
αn(T, Q) = inf {T − B : B ∈ L(X), B(X) ∈ Qn(X)} .
However, we do not have a proof of the Schauder's theorem for Q-compact
maps. In the following we present a result analogous to Terzioglu's Theorem for
Q-compact maps under the assumption that both T and T ∗ are Q-compact.
Theorem 4.1. Let E and F be Banach spaces , T ∈ L(E, F ) and assume that
both T and T ∗ are Q-compact maps. Then there exists sequence {un,k} ∈ Qn with
un,k → 0 for n → ∞ for all k, such that the inequality
T x ≤ sup < un,k(n), x >
holds for every x ∈ E. Here Qn is a "special" class of subsets of E∗ with the
property that un,k(n) ∈ {un,k}.
8
A. AKSOY
Proof. Since T ∗ : F ∗ → E∗ is Q-compact, thus by the Theorem 3.4, T ∗(UF ∗)
is a Q-compact set, thus there exists a sequence {un,k}k ⊂ An ∈ Qn such that
un,k → 0 as n → ∞ uniformly in k and
T ∗(UF ∗) ⊂( ∞Xn=1
Then for each x ∈ E, we have:
λnun,k(n) :
un,k(n) ∈ (un,k)
λn ≤ 1) .
∞Xn=1
λnun,k(n), x >
∞Xn=1
T x = sup
v∈UF ∗
< v, T x > = sup
v∈UF ∗
< T ∗v, x > = sup
n
<
and thus
T x ≤
∞Xn=1
λn sup
n
< un,k(n), x > ≤ sup
n
< un,k(n), x > .
(cid:3)
Remark 4.2. We say a map T ∈ L(E, F ) is a Q-compact map, if limn δn(T, Q) = 0.
To obtain "Schauder's Theorem" for Q-compact maps, one seeks a relationship
between δn(T ) and δn(T ∗). K. Astala in [3] proved that under the assumption
that the Banach space E has the lifting property and the Banach space F has
the extension property, for a map T ∈ L(E, F ), one has γ(T ) = γ(T ∗), where
γ(T ) denotes the measure of non-compactness. Since
lim
n→∞
δn(T ) = γ(T ),
by imposing extension and lifting properties on E and F respectively and keeping
tract of approximation schemes on these spaces one might obtain Schauder's type
of a theorem in this special case.
References
1. A. G. Aksoy, Q-Compact sets and Q-compact maps, Math. Japon. 36 (1991), no. 1, 1-7.
2. A. G. Aksoy and M. Nakamura, The approximation numbers γn(T ) and Q-compactness,
Math. Japon. 31 (1986), no. 6, 827-840.
3. K. Astala, On measures of non-compactness and ideal variations in Banach spaces, Ann.
Acad. Sci. Fenn. Ser. AI Math. Dissertations 29, (1980), 1-42.
4. Y. Brudnij and N. Krugljak, On a family of approximation spaces, Investigation in function
theory of several real variables, Yaroslavl State Univ., Yaroslavl, (1978), 15-42.
5. P.L. Butzer and K. Scherer, Approximationsprozesse und Interpolations methoden Bili-
ographisches Inst. Mannheim, 1968.
6. J. Diestel, Sequences and Series in Banach Spaces, Spriger-Verlag, Berlin, Heidelberg, New
York, 1984.
7. P. Enflo, A counter example to the approximation property in Banach spaces. Acta Math
130, (1973), 309-317.
8. C. V. Hutton On approximation numbers and its adjoint. Math. Ann. 210 (1974), 277-280.
9. J. Lindenstrauss, Extension of compact operators, Mem. Amer. Math. Soc. No. 48 (1964),
112 pp.
10. J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I, Sequence Spaces Springer-
Verlag, Berlin, Heidelberg, New York, 1977.
TERZIOGLU THM
9
11. A. Persson and P. Pietsch, p-nucleare une p-integrale Abbildungen in Banachraumen, Studia
Math. 33, (1969), 19-62.
12. P. Pietsch Approximation spaces, J. Approx. Theory, 32 (1981), no. 2, 115-134.
13. T. Terzioglu, A characterization of compact linear mappings, Arch. Math. 22 (1971), 76-78.
14. T. Terzioglu, Remarks on Compact and infinite-nuclear mappings Mathematica Balkanica,
2, (1972), 251-255.
15. K. Yosida, Functional analysis Spriger-Verlag, Berlin, Heidelberg, New York, 1965.
∗Department of Mathematics, Claremont McKenna College, 850 Columbia Av-
enue, Claremont, CA 91711, USA.
E-mail address: [email protected]
|
1201.2117 | 1 | 1201 | 2012-01-10T17:19:04 | Traceability of positive integral operators in the absence of a metric | [
"math.FA"
] | We investigate the traceability of positive integral operators on $L^2(X,\mu)$ when $X$ is a Hausdorff locally compact second countable space and $\mu$ is a non-degenerate, $\sigma$-finite and locally finite Borel measure. This setting includes other cases proved in the literature, for instance the one in which $X$ is a compact metric space and $\mu$ is a special finite measure. The results apply to spheres, tori and other relevant subsets of the usual space $\mathbb{R}^m$. | math.FA | math |
Traceability of positive integral operators in
the absence of a metric
Mario H. Castro, Valdir A. Menegatto and Ana P. Peron
12
Abstract
We investigate the traceability of positive integral operators on L2(X, µ) when
X is a Hausdorff locally compact second countable space and µ is a non-degenerate,
σ-finite and locally finite Borel measure. This setting includes other cases proved in
the literature, for instance the one in which X is a compact metric space and µ is a
special finite measure. The results apply to spheres, tori and other relevant subsets
of the usual space Rm.
Mathematics Subject Classification (2010).
15A18, 60G46.
42A82, 45P05, 47B34, 47B65,
Keywords. integral operators, positive definite kernels, trace-class, averaging, mar-
tingales.
1
Introduction
Let X be a Hausdorff locally compact and second countable topological space endowed
with a non-degenerate, σ-finite and locally finite Borel measure µ. In this paper, we shall
investigate the traceability of integral operators K : L2(X, µ) → L2(X, µ) generated by a
suitable kernel K : X × X → C from L2(X × X, µ × µ). The title of the paper refers to the
fact that the space X carries a topological structure rather than a metric one. The setting
just described allows the space L2(X, µ) to have a countable complete orthonormal subset
([7, p.92]) while the operator K, which is given by the formula
K(f ) :=ZX
K(·, y)f (y) dµ(y),
f ∈ L2(X, µ),
(1.1)
becomes compact. As so, the spectral theorem for compact operators is applicable and K
can be represented in the form
K(f ) =
∞
Xn=1
λnhf, fnifn,
f ∈ L2(X, µ),
(1.2)
in which {λn} is a sequence of real numbers (possibly finite) converging to 0 and {fn} is a
complete orthonormal sequence in L2(X, µ). The symbol h·, ·i will stand for the usual inner
product of L2(X, µ).
1First author partially supported by CAPES
2 Second author partially supported by FAPESP # 2010/19734-6
1
The basic requirement on the kernel K will be its positive definiteness. A kernel K from
L2(X × X, µ × µ) is L2(X, µ)-positive definite when the corresponding integral operator K,
is positive:
hK(f ), f i ≥ 0,
f ∈ L2(X, µ).
(1.3)
Fubini's theorem is all that is need in order to show that a L2(X, µ)-positive definite kernel
is hermitian µ × µ-a.e.. As so, the integral operator K is automatically self-adjoint with
respect to h·, ·i. In particular, the sequence {λn} mentioned in the previous paragraph needs
to be entirely composed of nonnegative numbers. In the present paper, we shall assume they
are listed in a decreasing order, with repetitions to account for multiplicities.
Under the conditions established above, the specific aim of this paper is to establish
additional conditions on K in order that K be trace-class, that is,
hK∗K(f ), f i1/2 < ∞
Xf ∈B
(1.4)
for every orthonormal basis B of L2(X, µ). In the formula above, K∗ is the adjoint of K.
We refer the reader to [5, 10, 11] for more information on trace-class operators.
The main result in this paper can be seen as a generalization of another one originally
proved in [11] for the case X = [a, b]. The proof there used in a key manner the so-called
Steklov's smoothing operator to construct an averaging process to generate a convenient
approximation to K. The upgrade to the case in which X is a subspace of Rn was discussed
in [8] and references therein. By assuming that the Lebesgue measure of nonempty inter-
sections of X with open balls of Rn was positive and using auxiliary approximation integral
operators generated by an averaging process constructed via the Hardy-Littlewood theory,
the main result in [8] described necessary and sufficient conditions for the traceability of the
integral operator, under the assumption of positive definiteness of the kernel. The process
used in [8] and other references as well provides a way to deal with the generating kernel
on the diagonal of X × X and it is convenient when the kernel is not continuous. Despite
using a similar average process, another achievement in the present paper is the inclusion
of a setting in which the measure does not need to be finite.
Since our spaces are no longer metric, the Hardy-Littlewood theory in the average
arguments need to be replaced or adapted. We will use techniques involving the construction
of auxiliary integral operators based on martingales constructed from special partitions of
X, following very closely the development of Brislawn in [2]. A similar construction have
appeared in [6] in an attempt to generalize Brislawn results to Lp spaces. The main difference
between the construction to be delineated here and those in [2] and [6] is that, in the present
one, we need to guarantee that the elements in the partitions belong to the topology of X.
This is the exact point where the assumption of local compactness will play an important
role.
For the sake of completeness we mention references [1, 14] where other characterizations
for traceability were obtained.
An outline of the paper is as follows. Section 2 contains the basic information on martin-
gales used in the paper, along with the key construction we will need in order to introduce
2
approximating auxiliary operators in Section 3. There, the main technical results are estab-
lished and proved. Section 4 contains the main results of the paper, including a convenient
equivalence for traceability.
2 A special martingale
This section contains several results involving a special martingale on X. Some of them are
just refined versions of results described in Section 2 of [2]. However, the reader is advised
that the basic references we used for the concepts and results either quoted or used here
are [4, 15].
Let (X, M, σ) denote a σ-finite measure space and F a sub-σ-algebra of M for which
(X, F , σ) is a σ-finite measure space too. If f : X → C is M-measurable, Radon-Nikodyn's
theorem asserts that we can find a unique F -measurable function g : X → C so that
ZA
f dσ =ZA
g dσ, A ∈ F .
(2.1)
The function g is called the conditional expectation of f relative to F and is written g =
E(f F ). If {Fn} is a family of sub-σ-algebras of M, a sequence {fn} of M-measurable
functions on X is a martingale if every fn is Fn-measurable and E(fnFm) = fm, m < n.
Next, we remind the reader about the basic setting we are assuming in the paper: X
is a Hausdorff, locally compact and second countable topological space endowed with a
non-degenerate, locally finite and σ-finite Borel measure µ. In addition to that, we will
write BX to denote the Borel σ-algebra of X.
Invoking the first countability axiom, we may infer that every point of X possesses an
open neighborhood. Since X is Hausdorff and locally compact, these neighborhoods can be
assumed to be the interior of a compact set. Thus, due to the local finiteness of (X, µ), we
can assume, in addition, that the open neighborhoods of elements of X have finite measure.
We intend to construct a special sequence of partitions of X from an open covering
{Ax}x∈X of it, composed of neighborhoods of the type just described, and use them to define
a particular martingale. If such a covering has been fixed, Lindeloff's theorem ([13, p.191])
implies that we can extract from it a countable sub-collection {An}, still covering X. Such
sub-collection can be used in the construction of a first stage partition P0 of X, following
these steps: the first two elements in the partition are A0 and its frontier ∂A0. Observing
that {An\A0} is an open and countable covering of X\A0, we pick A1\A0 and ∂A1\A0
to include in the partition. The family {An\A0 ∪ A1} is an open and countable covering
of X\A0 ∪ A1. We proceed, including its elements A2\A0 ∪ A1 and ∂A2\A0 ∪ A1 in the
partition. Proceeding inductively, we complete the construction of P0, which is countable
and entirely composed of Borel sets of finite measure. Since Theorem 7.8 in [9] implies that
µ is regular, all the sets of the form ∂An\A0 ∪ · · · ∪ An−1 in P0 have measure zero.
In the next step, we construct a sequence {Pn} of partitions of X from P0, using as we
can, a countable basis {Un} for the topology of X. For n = 0, 1, . . ., we put
Pn+1 = {Un ∩ A : A ∈ Pn} ∪ {(X \ Un) ∩ A : A ∈ Pn} ∪ {∂Un ∩ A : A ∈ Pn}.
3
Clearly, Pn+1 refines Pn and the sequence {Fn} of the corresponding σ-algebras generated
by those partitions increases to BX . In addition, every (X, Fn, µ) is σ-finite.
It is easy to see that for each x ∈ X and each positive n, there exists a unique set
On(x) ∈ Pn such that x ∈ On(x). We denote by N the subset of X containing all x ∈ X
for which µ(Om(x)) = 0, for some m ≥ 0. Since the sequence {On(x)} is telescoping, the
equality µ(Om(x)) = 0 implies µ(On(x)) = 0, n ≥ m. Being each Pn countable, it is easily
seen that µ(N) = 0.
The very same arguments used in [15, p.89] show that for every x ∈ X \ N and every
positive n, the conditional expectation En(f ) of f relative to Fn is given by the formula
En(f )(x) =
1
µ(On(x))ZOn(x)
f dµ.
(2.2)
The sequence {En(f )} defines a martingale generated by just one (measurable) function,
the martingale associated with f . Examples related to constructions similar to the one above
can be found in [15, p.88].
The section will be completed with a list of results involving the previous formula and
the maximal function M f of the martingale associated with f . Such martingale is defined
by the formula
M f (x) := sup{En(f )(x) : n = 1, 2, . . .},
x ∈ X.
(2.3)
Since the results are quite general and are not attached to the particular setting introduced
above, we will include sketches of the proofs for the convenience of the reader.
A classical result concerning the maximal function ([15, p.91]) implies that if p ∈ (0, ∞)
then
kM f kp ≤ cpkf kp,
f ∈ Lp(X, µ),
(2.4)
where cp is a constant depending on p only and k · kp denotes the usual norm of Lp(X, µ).
As for the conditional expectation, it transforms convergence in the mean into convergence
µ-a.e. Another basic result ([4, p.53] and [2, p.232]), commonly called Doob's martingale
convergence theorem, states that En(f ) converges to f µ-a.e., as long as f ∈ Lp(X, µ) and
p ∈ [1, ∞].
Moving forward, the inequalities
En(f )(x) ≤ M f (x),
x ∈ X \ N, n ≥ 1,
and
f (x) ≤ f (x) − En(f )(x) + M f (x),
x ∈ X \ N, n ≥ 1,
(2.5)
(2.6)
are easily deducted. Combining the last one with Doob's martingale convergence theorem,
we are led to the inequality f ≤ M f , µ-a.e.. As for the conditional expectation, we have the
following result found in [15, p.90]: if p ∈ [1, ∞] and f ∈ Lp(X, µ) then kEn(f )kp ≤ kf kp,
n ≥ 1. As a consequence, the following theorem holds.
Theorem 2.1 If p ∈ [1, ∞] then the linear map En : Lp(X, µ) → Lp(X, µ) is bounded. If
p = 2, then the previous map is a self-adjoint operator.
4
We close the section with a result for convergence in the mean of the conditional expec-
tation.
Theorem 2.2 If f ∈ L2(X, µ) then Enf converges to f in the mean.
Proof.
Now, inequality (2.5) leads to
If gn := f − En(f )2, n ≥ 1, the previous theorem yields that {gn} ⊂ L1(X, µ).
gn(x) ≤ 2(f (x)2 + En(f )(x)2) ≤ 4M f (x)2,
x ∈ X \ N, n ≥ 1.
(2.7)
Clearly, M f ∈ L2(X, µ) while Doob's convergence theorem gives us gn → 0 µ-a.e.. The
dominated convergence theorem connects the final arguments.
3 Approximating kernels
This section is entirely composed of technical results involving a family of operators con-
structed from the martingale defined in Section 2.
Under the notation in Section 2, Theorem 7.20 in [9] informs that the product measure
µ × µ is a regular Borel measure on X × X and the sequence {Pn × Pn} of partitions
of X × X increases to the Borel σ-algebra BX×X of (X × X, µ × µ). In particular, if
K ∈ L1
loc(X ×X, µ×µ), the conditional expectation with respect to the σ-algebra generated
by the partition Pn × Pn of X × X can be defined by the formula
En(K)(u, v) :=
1
σ(On(u))σ(On(v))ZOn(u)ZOn(v)
K(x, y) dµ(y)dµ(x).
(3.1)
Lemma 3.1 below provides information about a limit property regarding the open sets
On(x) previously defined. We will use the symbol χA to denote the characteristic function
of the subset A of X. We remind the reader that given x ∈ X and n ≥ 1, the construction
introduced in the previous section shows that there exists a unique On(x) ⊂ Pn so that
x ∈ On(x).
Lemma 3.1 If x ∈ X and n ≥ 1 then
lim
u→u0
χOn(u)(x) = χOn(u0)(x),
u0 ∈ X \ N.
(3.2)
Proof. Fix x ∈ X and n ≥ 1. If u ∈ X then x ∈ On(u) if and only if u ∈ On(x). Since
χOn(x)(u) = χOn(u)(x), we can write
χOn(u)(x) − χOn(u0)(x) = χOn(x)(u) − χOn(x)(u0),
u0 ∈ X.
(3.3)
Next, if u0 ∈ X \ N, the fact that On(u0) is open, leaves us with two cases: if x ∈ On(u0)
then u0 ∈ On(x) and, at the limit, we can assume u ∈ On(u0) = On(x) so that
lim
u→u0
χOn(x)(u) − χOn(x)(u0) = 1 − 1 = 0.
(3.4)
5
If x 6∈ On(u0) then u0 6∈ On(x), and assuming u ∈ On(u0) as we can, we conclude that
lim
u→u0
χOn(x)(u) − χOn(x)(u0) = 0 − 0 = 0.
(3.5)
The proof is complete.
It is now reasonable that the following result holds.
Lemma 3.2 If u0 ∈ X \ N then limu→u0 µ(On(u)) = µ(On(u0)), n = 1, 2, . . . .
Proof. Since
it follows that
µ(On(u)) =ZX
µ(On(u)) − µ(On(u0)) ≤ZX
χOn(u)(x) dµ(x),
u ∈ X,
(3.6)
χOn(u)(x) − χOn(u0)(x) dµ(x),
u ∈ X.
(3.7)
As so, the assertion of the lemma will be proved if we can show that
lim
u→u0ZX
χOn(u)(x) − χOn(u0)(x) dµ(x) = 0,
u0 ∈ X \ N.
(3.8)
Hence, in view of the previous lemma, it suffices to show that the integral and the limit in
the previous equation commute. The family {gu} defined by
gu(x) = χOn(u)(x) − χOn(u0)(x),
u, x ∈ X,
(3.9)
and the function g = χOn(u0) belong to L1(X, µ). Since gu ≤ g, µ-a.e., when u → u0, the
desired commuting property follows from the dominated convergence theorem.
We now turn to kernels of the form
Dn(u, x) =
1
µ(On(u))
χOn(u)(x),
u, x ∈ X, n = 1, 2, . . . .
(3.10)
and the corresponding integral operators Dn generated by Dn. For use ahead, we mention
the immediate formula
En(χOn(u) f ) = Dn(f ),
u ∈ X \ N,
f ∈ L2(X, µ).
(3.11)
Initially, we will use the above kernels to prove the following result.
Theorem 3.3 If K ∈ L2(X × X, µ × µ) and n ≥ 1 then En(K) is continuous µ × µ-a.e..
Proof.
u0, v0 ∈ X \ N. It is not hard to see that
It suffices to show that En(K) is continuous in the set (X \ N) × (X \ N). Let
En(K)(u, v) =ZXZX
Dn(u, x)K(x, y)Dn(v, y) dµ(y)dµ(x), u, v ∈ X,
(3.12)
6
and that we can use Lemma 3.1 and Lemma 3.2 to deduce that
lim
(u,v)→(u0,v0)
Dn(u, x)K(x, y)Dn(v, y) = Dn(u0, x)K(x, y)Dn(v0, y),
(3.13)
for x, y ∈ X a.e.. If (u, v) ∈ On(u0) × On(v0), we have
Dn(u, x)K(x, y)Dn(v, y) ≤
1
µ(On(u0))µ(On(v0))
K(x, y),
(3.14)
for x, y ∈ X a.e.. So, the continuity at (u0, v0) now follows from the dominated convergence
theorem.
Next, we will state and prove a list of technical results that will lead to the following
conclusion: DnKDn coincides with the integral operator generated by En(K).
The sequence of partitions {Pn} was constructed in such a way that each one of them
has the following feature: every element of {Ai} is a subset of at most finitely many On(x).
That been said, if n and i are fixed, we can write
(3.15)
Ai ⊂
m(n,i)
[j=1
On(xj)
[ N(n, i),
in which µ(N(n, i)) = 0 and 0 < µ(On(xj)) < ∞, j = 1, 2, . . . , m(n, i). The set N(n, i) is
nothing but the union of all elements of Pn for which the intersection with Ai has measure
zero.
In the next results, we will deal with a continuous function f : X → C with compact
support Xf . Since Xf can be covered by finitely many Ai, after re-ordering if necessary, we
can find an index l so that
l
Xf ⊂
[k=1
m(n,k)
[j=1
On(xj)
[ l
[k=1
with µ(∪l
k=1
N(n, k)) = 0. In that case, we will write
Yf =
l
m(n,k)
[k=1
[j=1
On(xj).
N(n, k)! ,
(3.16)
(3.17)
Lemma 3.4 Let f : X → C be a function with compact support Xf and K an element of
L1
loc(X × X, µ × µ). Then
ZX×XZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) d(µ × µ)(x, y)
=ZXZX×X
Dn(u, x)K(x, y)Dn(y, z)f (z) d(µ × µ)(x, y) dµ(z).
7
Proof. Pick M > 0 so that f (x) ≤ M, x ∈ X. We have
ZX×XZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) d(µ × µ)(x, y)
≤
M
µ(On(u))ZOn(u)×X
K(x, y)
µ(On(y))ZXf
χOn(y)(z) dµ(z) d(µ × µ)(x, y).
If y 6∈ Yf then y 6∈ Xf and, consequently, On(y) ∩ Xf = ∅. Thus χOn(y) = 0 in Xf and we
can take the above integral on On(u) × Yf . Now, if u ∈ X \ N, the local integrability of K
implies that
ZOn(u)×Yf
K(x, y)
µ(On(y))ZXf
χOn(y)(z) dµ(z) d(µ × µ)(x, y)
=ZOn(u)×Yf
≤ZOn(u)×Yf
K(x, y)
µ(On(y))
µ(On(y) ∩ Xf ) d(µ × µ)(x, y)
K(x, y) d(µ × µ)(x, y) < ∞.
Fubini's theorem ([12, p.386]) completes the proof.
Lemma 3.5 Let f : X → C be a continuous function with compact support Xf and K an
element of L1
loc(X × X, µ × µ). If u ∈ X \ N then
ZOn(u)×Yf
K(x, y)Dn(y, z) d(µ × µ)(x, y) =ZOn(u)ZYf
=ZYf ZOn(u)
K(x, y)Dn(y, z) dµ(y) dµ(x)
K(x, y)Dn(y, z) dµ(x) dµ(y),
with Yf as defined in (3.17).
Proof.
If u ∈ X \ N and z ∈ X then
ZOn(u)ZYf
K(x, y)Dn(y, z) dµ(y) dµ(x) =ZOn(u)ZYf (cid:12)(cid:12)(cid:12)(cid:12)
≤ZOn(u)ZYf
K(x, y)
µ(On(y))
K(x, y)
dµ(y) dµ(x)
χOn(y)(z)
µ(On(y))(cid:12)(cid:12)(cid:12)(cid:12)
dµ(y) dµ(x).
Introducing the decomposition (3.17) in the last expression above and recalling the unique-
ness property of the On(x), we deduce that
ZOn(u)ZYf
K(x, y)
µ(On(y))
K(x, y)
µ(On(y))
dµ(y) dµ(x)
m(n,k)
l
l
Xj=1 ZOn(yj )
dµ(y) dµ(x) =ZOn(u)
=ZOn(u)
Xj=1
1 ≤ j ≤ m(n, k) (cid:26)
Xk=1
Xk=1
max
m(n,k)
1 ≤ k ≤ l
≤
K(x, y) dµ(y) dµ(x)
1
µ(On(yj))ZOn(yj )
µ(On(yj))(cid:27) kKkL1(On(u)×Yf ) < ∞.
1
8
Once again, Fubini's theorem leads to the concluding statement.
Lemma 3.6 Let f : X → C be a continuous function with compact support Xf and K an
element of L1
loc(X × X, µ × µ). Then
ZXZXZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) dµ(y) dµ(x)
=ZX×XZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) d(µ × µ)(y, x).
Proof.
If u ∈ X \ N and M > 0 is a bound for f in X then it is easily seen that
ZXZX
ZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) dµ(y) dµ(x)
≤
≤
M
µ(On(u))ZOn(u)ZYf
µ(On(u))ZOn(u)ZYf
M
K(x, y)
µ(On(y))
µ(On(y) ∩ Xf ) dµ(x) dµ(y)
K(x, y) dµ(x) dµ(y) < ∞.
So, the result follows from Fubini's theorem once again.
The proof of the next lemma is analogous and will be omitted.
Lemma 3.7 Let f : X → C be a continuous function with compact support and K and
element in L1
loc(X × X, µ × µ). Then
ZXZXZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) dµ(y) dµ(x)
=ZXZXZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(x) dµ(y) dµ(z).
Recalling that if x, y ∈ X then z ∈ On(y) if and only if y ∈ On(z), the following lemma
becomes obvious.
Lemma 3.8 If n ≥ 1 then Dn(y, z) = Dn(z, y), y, z ∈ X \ N.
Below, E n
K will denote the integral operator generated by En(K).
Theorem 3.9 If K ∈ L2(X × X, µ × µ) then DnKDn(f ) = E n
K(f ), f ∈ L2(X, µ).
Proof. Clearly L2(X × X, µ × µ) ⊂ L1
f : X → C is continuous with compact support then the previous lemmas imply that
loc(X × X, µ × µ). If K ∈ L2(X × X, µ × µ) and
DnKDn(f )(u) =ZXZXZX
Dn(u, x)K(x, y)Dn(y, z)f (z) dµ(z) dµ(y) dµ(x)
Dn(u, x)K(x, y)Dn(y, z)f (z)dµ(z)dµ(y)dµ(x)
=ZOn(u)ZYf ZXf
=ZXZXZX
= E n
K(f )(u),
u ∈ X \ N.
9
Dn(u, x)K(x, y)Dn(z, y)f (z) dµ(x) dµ(y) dµ(z)
Hence, the result in the statement of the theorem follows from the equality µ(X \ N) = 0
and from a basis approximation theorem from measure theory ([12, p.197]).
The last result of the section refers to the positive definiteness of En(K).
Theorem 3.10 If K is L2(X, µ)-positive definite then so is En(K).
If K is L2(X, µ)-positive definite then both, K and En(K), belong to the space
Proof.
L2(X × X, µ × µ). On the other hand, Theorem 2.1 and (3.11) imply that Dn(L2(X, µ)) ⊂
L2(X, µ). Thus, an application of Theorem 3.9 leads to
hE n
K(f ), f i2 = hKDn(f ), Dn(f )i2 ≥ 0,
f ∈ L2(X, µ).
(3.18)
The proof is complete.
4 Traceability
This section contains the main results of the paper. They can be interpreted as generaliza-
tions of results obtained in [8] and other references quoted here. The traceability results de-
scribed here will be obtained via several known results on trace-class operators and singular
values of operators. We will quote some of them and just mention others. The construction
developed in Section 2 reveals that the diagonal of X is, up to a set of measure zero, a
subset of (X \ N) × (X \ N). This remark justify why some of the integrals appearing below
are not identically zero. Given K ∈ L2(X × X, µ × µ), we will consider E n
K acting like an
operator on L2(X, µ). All other operators mentioned here are to be understood acting in
the same way.
The following lemma is an adapted to our purposes version of Theorem 4.1 in [3].
Lemma 4.1 If K is a continuous (µ × µ-a.e.) L2(X, µ)-positive definite kernel and x ∈
X → K(x, x) is integrable then K is trace-class and
tr (K) =ZX
K(x, x) dµ(x).
Lemma 4.2 Let K be L2(X, µ)-positive definite. If
lim sup
n→∞ ZX
En(K)(x, x) dµ(x) < ∞,
then lim supn→∞ tr (E n
K) < ∞.
(4.1)
(4.2)
Proof. If (4.2) holds then there exists n0 ∈ N such that x ∈ X → En(K)(x, x) is integrable
for n ≥ n0. Theorem 3.10 implies that En(K) is L2(X, µ)-positive definite while Theorem
3.3 shows that En(K) is continuous µ × µ-a.e.. Applying Lemma 4.1 we see that
tr (E n
K) =ZX
En(K)(x, x) dµ(x), n ≥ n0.
(4.3)
10
The result follows.
Next, we recall some facts involving singular values of an operator. If T is a compact
operator on a Hilbert space, a singular value of T is an eigenvalue of (T ∗T )1/2. We shall
enumerate the nonzero singular values of T in decreasing order, taking multiplicities into
account: s1(T ) ≥ s2(T ) ≥ . . .. If the rank ρ of (T ∗T )1/2 is finite, obviously sj(T ) = 0,
j ≥ ρ + 1. If the eigenvalues of T are ordered like l1(T ) ≥ l2(T ) ≥ . . ., then a classical
result from operator theory states that sj(T ) = lj(T ), j = 1, 2, . . ., as long as T is either
hermitian or normal. If S is another compact operator of same type as T , and assuming the
same ordering on the singular values of S, the following inequality holds: sn(T ) − sn(S) ≤
kT − Sk, n = 1, 2 . . .. All of these results can be found with proofs in [10, 11].
In Theorem 4.4, a complement of Lemma 4.2, we also use the following nontrivial result
on convergence of operators ([8]).
Lemma 4.3 Let {Tn} be a countable set of bounded linear operators on a Hilbert space H
such that limn→∞ kTn(f ) − f kH = 0, f ∈ H. If every Tn is self-adjoint and T is a bounded
compact operator on H then limn→∞ kTnT Tn − T k = 0.
Theorem 4.4 Let K be L2(X, µ)-positive definite. If
then K is trace-class.
lim sup
n→∞ ZX
En(K)(x, x) dµ(x) < ∞,
(4.4)
Proof. Since {sj(E n
K)} ⊂ (0, ∞), it is quite clear that
sj(E n
K) ≤ tr (E n
K),
k = 1, 2, . . . .
(4.5)
k
Xj=1
Theorem 3.9 and the inequality mentioned before Lemma 4.3 imply that
sj(E n
K) − sj(K) ≤ kE n
K − Kk = kDnKDn − Kk,
j = 1, 2, . . . .
(4.6)
Since each Dn is self-adjoint, K is compact and
lim
n→∞
kDn(f ) − f k2 = 0,
f ∈ L2(X, µ),
we are authorized to apply Lemma 4.3 to conclude, from (4.6), that
lim
n→∞
sj(E n
K) = sj(K),
j = 1, 2, . . . .
(4.7)
(4.8)
It is now clear that
k
Xj=1
sj(K) = lim sup
n→∞
k
Xj=1
and that concludes the proof.
sj(E n
K) ≤ lim sup
n→∞
tr (E n
K),
k = 1, 2, . . . ,
(4.9)
In order to deal with the converse of the previous result, we will need the following
result ([10, p.51]): if S1, S2 and T are bounded linear operators on a Hilbert space and T is
compact then so is the composition S1T S2 and sj(S1T S2) ≤ kS1ksj(T )kS2k, j = 1, 2, . . . .
11
Lemma 4.5 Let p ∈ [1, ∞) and K ∈ Lp(X × X, µ × µ). If x ∈ X → K(x, x) is integrable
and µ(X) < ∞ then there is a positive integer n0 for which x ∈ X → En(K)(x, x) is
integrable when n ≥ n0.
Proof. Since
En(K)(u, u) ≤ En(K)(u, u) − K(u, u) + K(u, u),
u ∈ X \ N,
(4.10)
we can use Doob's convergence theorem to select a positive integer n0 so that
En(K)(u, u) ≤ 1 + K(u, u),
u ∈ X \ N, n ≥ n0.
(4.11)
Our assumptions on X and x ∈ X → K(x, x) imply the result.
Theorem 4.6 Let K be L2(X, µ)-positive definite. If x ∈ X → K(x, x) is integrable and
µ(X) < ∞ then there is n0 ∈ N so that E n
K ∈ B1(L2(X)) and
tr (E n
K) =ZX
En(K)(x, x) dµ(x), n ≥ n0.
(4.12)
Proof. The previous lemma reveals that x ∈ X → En(K)(x, x) is integrable for n large.
As so, the result follows from Theorem 3.3 and Lemma 4.1.
Theorem 4.7 Let K ∈ L2(X × X, µ × µ). If K is trace-class then so is every E n
number tr (K) is an upper bound for the sequence {tr (E n
K. The
K)}.
Proof. Assume K is trace-class. Since each Dn is bounded, Theorem 3.10 and the comments
preceding Lemma 4.5 imply that
sj(E n
K) = sj(DnKDn) ≤ kDnksj(K)kDnk, n = 1, 2, . . . .
Hence,
∞
and the two assertions of the lemma follow.
sj(E n
K) ≤ kDnk2
Xj=1
sj(K),
∞
Xj=1
The following result is very close to a converse of Theorem 4.4.
Theorem 4.8 Let K ∈ L2(X × X, µ × µ). If K is trace-class then
Proof. A basic inequality for the trace ([10, p.54]) implies that
∞
lim
n→∞
tr (E n
K) = tr (K).
(4.13)
(4.14)
tr (E n
K) − tr (K) ≤
sj(E n
K − K), n = 1, 2, . . . ,
(4.15)
Xj=1
as long as K is trace-class. On the other hand, since (see [10, p.89])
lim
n→∞
Theorem 3.9 completes the proof.
∞
Xj=1
sj(DnKDn − K) = 0,
(4.16)
Next, we move to a proof of the converse of Theorem 4.3 in the case when µ(X) < ∞.
12
Theorem 4.9 Let K be L2(X, µ)- positive definite. If K is trace-class and µ(X) < ∞ then
lim
n→∞ZX
En(K)(x, x) dµ(x) < ∞.
(4.17)
Proof. Assume K is trace-class. Since the function x ∈ X → K(x, x) is integrable already,
if µ(X) < ∞, we can use Theorem 4.6 to find a positive integer n0 such that
tr(E n
K) =ZX
En(K)(x, x) dµ(x), n ≥ n0.
(4.18)
An application of Theorem 4.8 finishes the proof.
At this point, it is very important to remind the reader that the results we have obtained
includes the case in which X is either a sphere or a torus.
Next, we intend to consider cases in which X has no finite measure. In order to handle
that, we use the cover {Am} of X constructed before to define a sequence of subsets
of X that increases to X. Precisely, defining Xj = ∪j
m=1Am, j ≥ 1, we immediately
have the following two properties: X = ∪∞
j=1Xj and if x ∈ X then there exists j0 ≥ 0
such that x ∈ Xj, j ≥ j0. Using the sequence just defined, we now take linear operators
Pj : L2(X, µ) → L2(X, µ) defined by the formula Pj(f ) = f χXj , f ∈ L2(X, µ). They are self-
adjoint and the uniform boundedness principle shows that the sequence {Pj} is bounded
in the space of bounded linear operators on L2(X, µ). Also, the dominated convergence
theorem implies that {Pj} converges pointwise to the identity operator on L2(X, µ). The
following technical lemma contains a critical information on the sequence {Pj}.
Lemma 4.10 If T : L2(X, µ) → L2(X, µ) is trace-class then each PjT Pj is so and the
limit formula limj→∞ tr (PjT Pj) = tr (T ) holds.
Proof. The first assertion is a consequence of the remark preceding Lemma 4.5. As for
the other, it follows from Theorem 11.3 in [10] .
The converse of Theorem 4.4 reads as follows.
Theorem 4.11 Let K be L2(X, µ)-positive definite. If K is trace-class then the limit
lim
n→∞ZX
En(K)(x, x) dµ(x)
(4.19)
exists and is finite.
Proof. The proof requires the double-indexed operator Qn
j given by the formula
Qn
j (f )(x) =ZXj
En(K)(x, y)f (y) dµ(y), x ∈ Xj,
f ∈ L2(Xj, µ).
(4.20)
13
If f ∈ L2(Xj), let us write f to denote a function on X that coincides with f on Xj and is
zero in X \ Xj. It is now clear that
Qn
ZXj
j (f )(x)f (x) dµ(x) = ZXZX
= ZXZX
= ZX
En(K)(x, y) f (y) f (x) dµ(y) dµ(x)
En(K)(x, y) f (y) dµ(y) f(x) dµ(x)
K( f )(x) f (x) dµ(x),
E n
f ∈ L2(Xj, µ).
j is L2(Xj, µ)-positive
Since K is L2(X, µ)-positive definite, Theorem 3.10 guarantees that Qn
definite. Also, the fact that K is trace-class implies that x ∈ X → K(x, x) is integrable.
Hence, due to Lemma 4.5, there exists n0 ≥ 0 such that x ∈ Xj → En(K)(x, x) whenever
n ≥ n0. Recalling Theorem 3.3 and applying Lemma 4.1, we deduce that Qn
j is trace-class
and
En(K)(x, x) dµ(x),
(4.21)
tr (Qn
j ) =ZXj
as long as n ≥ n0. Let us keep the previous condition on n in force. If Vj is the closed
subspace L2(X, µ) encompassing the functions on X which are zero in X \ Xj and Rn
:
j
Vj → Vj is the operator given by
Rn
j (f )(x) = χXj (x)ZX
En(K)(x, y)χXj (y)f (y) dµ(y),
(4.22)
with x ∈ X, f ∈ Vj, then Rn
previous lemma,
j and Qn
j possess the same eigenvalues. Having in mind the
(PjE n
KPj)(f )(x) =ZX
En(K)(x, y)χXj ×Xj (x, y)f (y) dµ(y),
(4.23)
for x ∈ X, f ∈ L2(X), and we can conclude now that Rn
eigenvalues. Therefore,
j and PjE n
KPj have the same
tr (PjE n
KPj) = tr (Rn
j ) = tr (Qn
j ) =ZXj
En(K)(x, x) dµ(x).
(4.24)
The monotone convergence theorem leads to
tr (E n
K) =ZX
En(K)(x, x) dµ(x).
(4.25)
Finally, (4.24) and the observation made before Theorem 4.3 lead to the assertion of the
theorem.
14
References
[1] Brislawn, C., Kernels of trace class operators. Proc. Amer. Math. Soc. 104 (1988), no.
4, 1181 -- 1190.
[2] Brislawn, C., Traceable integral kernels on countably generated measure spaces. Pacific
J. Math. 150 (1991), no. 2, 229 -- 240.
[3] Castro, M. H.; Menegatto, V. A.; Peron, A. P., Integral operators generated by Mercer-
like kernels: non-metric results, preprint.
[4] Chatterji, S. D., Les martingales et leurs applications analytiques. (French) ´Ecole d'´Et´e
de Probabilit´es: Processus Stochastiques (Saint Flour, 1971), pp. 27 -- 164. Lecture
Notes in Math., Vol. 307, Springer, Berlin, 1973.
[5] Conway, J. B., A course in operator theory. Graduate Studies in Mathematics, 21.
American Mathematical Society, Providence, RI, 2000.
[6] Delgado, J., The trace of nuclear operators on Lp(µ) for σ-finite Borel measures on
second countable spaces. Integral Equations Operator Theory 68 (2010), no. 1, 61-74.
[7] Doob, J. L., Measure theory. Graduate Texts in Mathematics, 143. Springer-Verlag,
New York, 1994.
[8] Ferreira, J. C.; Menegatto, V. A.; Oliveira, C. P., On the nuclearity of integral opera-
tors, Positivity, 13 (2009), no. 3, 519-541.
[9] Folland, G. B., Real analysis. Modern techniques and their applications. Second edi-
tion. Pure and Applied Mathematics (New York). A Wiley-Interscience Publication.
John Wiley & Sons, Inc., New York, 1999.
[10] Gohberg, I.; Goldberg, S.; Krupnik, N., Traces and determinants of linear operators.
Operator Theory: Advances and Applications, 116. Birkhauser Verlag, Basel, 2000.
[11] Gohberg, I. C.; Kreın, M. G., Introduction to the theory of linear nonselfadjoint op-
erators. Translated from the Russian by A. Feinstein. Translations of Mathematical
Monographs, Vol. 18 American Mathematical Society, Providence, R.I. 1969.
[12] Hewitt, E.; Stromberg, K., Real and abstract analysis. A modern treatment of the
theory of functions of a real variable. Third printing. Graduate Texts in Mathematics,
No. 25. Springer-Verlag, New York-Heidelberg, 1975.
[13] Munkres, J. R., Topology. 2nd ed. (English) Upper Saddle River, NJ: Prentice Hall.
(2000).
[14] Sato, H., Nuclearity of a nonnegative definite integral kernel on a separable metric
space. J. Theoret. Probab. 5 (1992), no. 2, 349 -- 353.
15
[15] Stein, E. M., Topics in harmonic analysis related to the Littlewood-Paley theory. An-
nals of Mathematics Studies, No. 63 Princeton University Press, Princeton, N.J.; Uni-
versity of Tokyo Press, Tokyo 1970.
Departamento de Matem´atica,
ICMC-USP - Sao Carlos, Caixa Postal 668,
13560-970 Sao Carlos SP, Brasil.
[email protected]
[email protected]
FAMAT-UFU
Caixa Postal 593
38400-902 Uberlandia-MG, Brasil
[email protected]
16
|
1302.2483 | 1 | 1302 | 2013-02-11T14:16:06 | On numerically hypercyclic operators | [
"math.FA",
"math.DS"
] | According to Kim, Peris and Song, a continuous linear operator $T$ on a complex Banach space $X$ is called {\it numerically hypercyclic} if the numerical orbit $\{f(T^nx):n\in\N\}$ is dense in $\C$ for some $x\in X$ and $f\in X^*$ satisfying $\|x\|=\|f\|=f(x)=1$. They have characterized numerically hypercyclic weighted shifts and provided an example of a numerically hypercyclic operator on $\C^2$.
We answer two questions of Kim, Peris and Song. Namely, we construct a numerically hypercyclic operator, whose square is not numerically hypercyclic as well as an operator which is not numerically hypercyclic but has two numerical orbits whose union is dense in $\C$. We characterize numerically hypercyclic operators on $\C^2$ as well as the operators similar to a numerically hypercyclic one and those operators whose conjugacy class consists entirely of numerically hypercyclic operators. We describe in spectral terms the operator norm closure of the set of numerically hypercyclic operators on a reflexive Banach space. Finally, we provide criteria for numeric hypercyclicity and decide upon the numerical hypercyclicity of operators from various classes. | math.FA | math |
On numerically hypercyclic operators
Stanislav Shkarin
Abstract
According to Kim, Peris and Song, a continuous linear operator T on a complex Banach space X is
called numerically hypercyclic if the numerical orbit tfpT nxq : n P Nu is dense in C for some x P X and
f P X satisfying }x} " }f} " fpxq " 1. They have characterized numerically hypercyclic weighted
shifts and provided an example of a numerically hypercyclic operator on C2.
We answer two questions of Kim, Peris and Song. Namely, we construct a numerically hypercyclic
operator, whose square is not numerically hypercyclic as well as an operator which is not numerically
hypercyclic but has two numerical orbits whose union is dense in C. We characterize numerically
hypercyclic operators on C2 as well as the operators similar to a numerically hypercyclic one and those
operators whose conjugacy class consists entirely of numerically hypercyclic operators. We describe in
spectral terms the operator norm closure of the set of numerically hypercyclic operators on a reflexive
Banach space. Finally, we provide criteria for numeric hypercyclicity and decide upon the numerical
hypercyclicity of operators from various classes.
MSC: 47A16, 37A25
Keywords: Numerically hypercyclic operators, numerical orbit, numerical range
1
Introduction
. . . zmk
1
1 , . . . , zn
Throughout this article X stands for a Banach space over the field C of complex numbers, while SpXq "
tx P X : }x} " 1u and BpXq " tx P X : }x} ď 1u. As usual, R is the field of real numbers, R` is
the set of non-negative real numbers, D " tz P C : z ă 1u, T " tz P C : z " 1u, Z` is the set of
non-negative integers and N is the set of positive integers. If M is a finite or countable set, the symbol
`pMq stands for the set of sequences in ℓ1pMq whose entries are all non-negative real numbers. We say
ℓ1
that z1, . . . , zk P T are independent if zm1
k ‰ 1 for every non-zero vector m P Zk. Equivalently,
zj " eiθj with θj P R are independent if and only if π, θ1, . . . , θk are linearly independent over the field
Q of rational numbers. It is well-known and easy to prove that z1, . . . , zk P T are independent if and
kq : n P Z`u is dense in Tn. Recall that a sequence tenunPN in X is called a
only if the set tpzn
Schauder basis in X if every x P X can be uniquely written as x "
cnxn, where cn P C and the
series is norm-convergent. The functionals en : X Ñ C, enpxq " cn are automatically continuous and
are called the coordinate functionals of the Schauder basis tenunPN. Obviously, enpekq " δn,k for every
n, k P N. A sequence tenunPN in X is called a Schauder basic sequence if tenunPN is a Schauder basis in
the closed linear span of the set ten : n P Nu. The symbol LpXq stands for the space of bounded linear
operators T : X Ñ X and X is the space of continuous linear functionals f : X Ñ C. For T P LpXq,
the dual operator T P LpXq acts according to the formula T fpxq " fpT xq. For T P LpXq, σpTq
stands for the spectrum of T , while σppTq denotes the point spectrum of T . Recall that T P LpXq is
called semi-Fredholm if TpXq is closed in X and either X{TpXq or ker T (or both, in which case T is
called Fredholm) is finite dimensional. The number ipTq " dim ker T ´ dim X{TpXq P Z Y t´8,`8u is
called the index of T . Note that the set of semi-Fredholm operators is norm-open in LpXq and that the
index function i is locally constant. Recall that λ P C is called a normal eigenvalue of T P LpXq if λ is
an isolated point of σpTq and the spectral projection corresponding to the clopen subset tλu of σpTq has
finite rank. This rank is called the multiplicity of the normal eigenvalue λ.
Recall that x P X is called a hypercyclic vector for T P LpXq if the orbit
8řn"1
OpT, xq " tT nx : n P Z`u
1
is dense in X. Similarly, x P X is called a weakly hypercyclic vector for T P LpXq if OpT, xq is dense in X
with respect to the weak topology. An operator T , which has a hypercyclic vector is called hypercyclic,
while an operator T , which has a weakly hypercyclic vector, is called weakly hypercyclic. For more
information on hypercyclicity see the books [2, 7] and references therein. Recall also that T P LpXq is
called power bounded if supt}T n} : n P Nu ă 8. Unless stated otherwise, we assume that Cn carries
the standard Euclidean norm. We shall routinely identify an operator T P LpCnq with its matrix. For
instance, a diagonal operator on Cn is an operator whose matrix is diagonal. The same agreement stands
for operators on classical sequence spaces.
1.1 The main concept
For T P LpXq and px, fq P X X, the numerical orbit of px, fq is the set
OpT, x, fq " fpOpT, xqq " tfpT nxq : n P Z`u Ă C.
Note that numerical orbits are also known as weak orbits, see [1] and references therein. The following
definition, motivated by the concept of the numerical range of an operator, was introduced by Kim, Peris
and Song [10], see also [11] for the generalizations to polynomials on Banach spaces.
Definition 1.1. Let
ΠpXq " tpx, fq P X X : }x} " }f} " fpxq " 1u.
We say that px, fq P ΠpXq is a numerically hypercyclic vector for T P LpXq if the numerical orbit
OpT, x, fq is dense in C. An operator T is called numerically hypercyclic if it has a numerically hypercyclic
vector. We use the symbol NHpXq to denote the set of all numerically hypercyclic operators T P LpXq.
The following proposition collects some elementary observations made in [10].
Proposition 1.2. Let T P LpXq.
• If T is power bounded, then it is not numerically hypercyclic.
• If T is weakly hypercyclic, then it is numerically hypercyclic.
• If the restriction of T to a closed invariant subspace is numerically hypercyclic, then so is T .
• If T is numerically hypercyclic, then T is numerically hypercyclic. In particular, if X is reflexive,
then T is numerically hypercyclic if and only if T is.
• If X is a Hilbert space, then T is numerically hypercyclic if and only if there is x P SpXq such that
txT nx, xy : n P Z`u is dense in C, where x,y is the inner product of X.
The last observation is due to the fact that if H is a Hilbert space and x P SpHq, then the only
functional f P SpHq satisfying fpxq " 1 is given by fpyq " xy, xy. Thus in this case the numerical orbits
featuring in the definition of numeric hypercyclicity are naturally labelled by the elements of SpHq. If H
is a Hilbert space and T P LpHq, we use the following notation:
NOpT, xq " txT nx, xy : n P Z`u for x P SpHq.
Recall that forward Fw and backward Bw weighted shifts on ℓppZ`q act according to the rule
px0, x1, x2, . . . q FwÞÑp0, w1x0, w2x1, . . . q and px0, x1, x2, . . . q BwÞÑpw1x1, w2x2, . . . q, where w " twnunPN is a
bounded sequence of non-zero scalars. A bilateral weighted shift Tw on ℓppZq is defined by pTwxqn "
wn`1xn`1, where again w " twnunPZ is a bounded sequence of non-zero scalars. The main results of [10]
can be summarized as follows.
Theorem KPS1. There is a diagonal numerically hypercyclic operator on C2. Consequently, there is a
numerically hypercyclic operator on Cn for every n ě 2.
Theorem KPS2. Let 1 ă p ă 8 and T be a backward or forward weighted shift on ℓppNq or a bilateral
weighted shift on ℓppZq. Then T is numerically hypercyclic if and only if T is not power bounded.
2
Remark 1.3. Theorem KPS2 is not exactly what is stated in [10]. Namely, the corresponding results
in [10] provide a condition on the weight sequence equivalent to the numeric hypercyclicity. For bilateral
weighted shifts the authors notice that this condition is equivalent to the operator being not power
bounded, while saying nothing for unilateral shifts. However upon reflection, one can readily see that
this condition is again equivalent to the operator being not power bounded. Furthermore in the case of
unilateral backward shifts the same condition is equivalent to hypercyclicity. One has just co compare
it to the characterization of Salas [15] of hypercyclic backward shifts. On the other hand, there are no
hypercyclic forward shifts and there are non-hypercyclic bilateral shifts, which are not power bounded.
Recall that T P LpXq and S P LpY q are called similar if there is an isomorphism (in the category
of topological vector spaces) J : X Ñ Y such that J´1SJ " T . Note that hypercyclicity and weak
hypercyclicity are similarity invariants, while the numeric hypercyclicity is not (below we provide plenty
of examples substantiating the last claim). This prompts us to consider the following concepts.
Definition 1.4. We say that T P LpXq is weakly numerically hypercyclic if T is similar to a numerically
hypercyclic operator. We say that T P LpXq is strongly numerically hypercyclic if every operator similar
to T is numerically hypercyclic. We denote the sets of weakly numerically hypercyclic operators and of
strongly numerically hypercyclic operators on X by the symbols WNHpXq and SNHpXq respectively.
The following elementary proposition characterizes weak numeric hypercyclicity.
Proposition 1.5. Let T P LpXq. Then T is weakly numerically hypercyclic if and only if there exist
x P X and f P X such that OpT, x, fq is dense in C.
Proof. The 'only if' part is obvious. Let x P X and f P X be such that OpT, x, fq is dense in C. It
remains to show that T P WNHpXq. Replacing x by T nx for an appropriate n P N and normalizing, we
can without loss of generality assume that }x} " fpxq " 1. Let D " tx P X : }x} ď 1, fpxq ď 1u and
}}1 be the Minkowski functional of D. Since D is bounded and contains a neighborhood of 0 in X, }}1
is a norm on X equivalent to the original norm } }. It is easy to see that }x}1 " }f}1 " fpxq " 1, where
}f}1 is the norm of f as a functional on the normed space X1 " pX,} }1q. Thus px, fq P ΠpX1q and
therefore px, fq is a numerically hypercyclic vector for T acting on X1. Since }} and }}1 are equivalent,
T acting on X is similar to T acting on X1. Thus T is weakly numerically hypercyclic.
According to Feldman [4], T P LpXq is called 1-weakly hypercyclic if there is x P X such that
fpOpT, xqq is dense in C for each non-zero f P X. Of course, every weakly hypercyclic operator is 1-
weakly hypercyclic. We collect a number of straightforward consequences of Definition 1.4 in the following
proposition. The proof is elementary and is left for the reader as an exercise.
Proposition 1.6. Let T P LpXq.
• If T is power bounded, then T R WNHpXq. If T is 1-weakly hypercyclic, then T P SNHpXq.
• If a restriction of T to an invariant closed linear subspace is weakly numerically hypercyclic pres-
pectively, strongly numerically hypercyclicq, then T is weakly numerically hypercyclic prespectively,
strongly numerically hypercyclicq.
• If T P WNHpXq, then T P WNHpXq. If additionally X is reflexive, then T P WNHpXq ðñ
T P WNHpXq and T P SNHpXq ðñ T P SNHpXq.
1.2 Finite dimensional results
The following two questions are raised in [10].
Question KPS3. Let T P NHpXq. Is T n numerically hypercyclic for every n P N?
Question KPS4. Let T P LpXq and there exist pxj , fjq P ΠpXq for 1 ď j ď n for which
is dense in C. Is T numerically hypercyclic?
nŤj"1
OpT, xj , fjq
We answer both questions negatively.
3
Theorem 1.7. There exists a diagonal T P NHpC2q such that T 2 R NHpC2q.
Theorem 1.8. There exist a diagonal T P LpC4q and x1, x2 P SpC4q such that T R NHpC4q, while
NOpT, x1q Y NOpT, x2q is dense in C.
The following two results are motivated by Theorem KPS1. They provide sufficient conditions of
weak/strong numeric hypercyclicity in terms of the restriction to a finite dimensional invariant subspace.
Theorem 1.9. Each of the following conditions is sufficient for the weak numeric hypercyclicity of
T P LpXq.
(1.9.1) There exist λ1, λ2 P σppTq such that λ1 " λ2 ą 1 and λ1
λ1
(1.9.2) There are independent λ1, λ2 P T such that kerpT ´ λj Iq2 ‰ kerpT ´ λj Iq for j P t1, 2u.
(1.9.3) There exist λ1, λ2, λ3 P σppTq such that λ1 " λ2 ą λ3 ą 1, λ1
has infinite order in the group T
are independent in T.
, λ2
λ2
λ2
and λ1
λ1
, λ3
λ3
are independent.
(1.9.4) There exist λ1, λ2 P σppTq and λ3 P T such that kerpT ´ λ3Iq2 ‰ kerpT ´ λ3Iq, λ1 " λ2 ą 1,
λ1
λ2
has infinite order in T and λ1
λ1
, λ3 are independent.
Theorem 1.10. Each of the following conditions is sufficient for the strong numeric hypercyclicity of
T P LpXq.
(1.10.1) There exist λ1, . . . , λn P σppTq and c1, . . . , cn P R` such that ! nřj"1
(1.10.2) There exist λ1, λ2, λ3 P σppTq such that λ1 " λ2 " λ3 ą 1 and λ1
λ1
cj λk
j : k P Z`) is dense in C.
, λ2
λ2
, λ3
λ3
are independent.
In the Hilbert space situation we can refine the first two parts of Theorem 1.9.
Proposition 1.11. Let H be a Hilbert space and T P LpHq. Assume that there exist λ1, λ2 P σppTq such
that λ1 " λ2 ą 1, the eigenspaces kerpT ´ λ1Iq and kerpT ´ λ2Iq are non-orthogonal and λ1
, λ2
λ1
λ2
are independent in T. Then T P NHpHq.
Proposition 1.12. Let H be a Hilbert space and T P LpHq. Assume that there exist λ1, λ2 P C such that
kerpT ´ λjIq2 ‰ kerpT ´ λj Iq for j P t1, 2u, λ1 " λ2 ě 1 and λ1
are independent in T. Then
λ1
T P NHpHq.
, λ2
λ2
We characterize NHpC2q, SNHpC2q, WNHpC2q and WNHpC3q.
, λ2
λ2
1 ` λk
Theorem 1.13. Let T P LpC2q. Then T P WNHpC2q if and only if σpTq " tλ1, λ2u, where λ1 "
λ2 ą 1 and λ1
are independent in T. Furthermore, T P SNHpC2q if and only if σpTq " tλ1, λ2u
λ1
with tλk
2 : k P Z`u being dense in C. Finally, T P NHpC2q if and only if either T P SNHpC2q or
T P WNHpC2q and T is not unitarily equivalent to a diagonal operator.
Theorem 1.14. Let T P LpC3q. Then T is weakly numerically hypercyclic if and only if either there are
λ1, λ2 P σpTq such that λ1 " λ2 ą 1 and λ1
are independent in T or σpTq " tλ1, λ2, λ3u, where
λ1
λ1 " λ2 ą λ3 ą 1, λ1
Remark 1.15. The above results provide easy explicit examples of numerically hypercyclic operators
on Cn. For instance, applying Theorems 1.13, Theorem 1.10 and Proposition 1.12 respectively, we have
has infinite order in the group T and λ1
λ1
are independent in T.
, λ2
λ2
, λ3
λ3
λ2
0
2eiπ log 2
and
1
2eiπ log 3 P NHpC2q,
eiπ log 2
1
0
0
0
eiπ log 2
0
0
0
0
eiπ log 3
0
0
1
0
eiπ log 3
2eiπ log 2
0
0
0
2eiπ log 3
0
‹‹‚P NHpC4q.
4
0
0
2eiπ log 5
‚P SNHpC3q Ă NHpC3q
In relation to Theorem 1.13 it makes sense to mention the following fact.
Proposition 1.16. For every R ą 1, pz, wq P T2 : tRnpzn ` wnq : n P Z`u is dense in C( is a dense
Gδ-subset of T2 of Lebesgue measure 0.
Remark 1.17. Let tx, yu be a fixed linear basis in C2 and R ą 1. For each pz, wq P T2, consider
T " Tz,w P LpC2q given by T x " Rzx and T y " Rwy. By Theorem 1.13 and Proposition 1.16, for
almost all pz, wq P T2 in the Lebesgue measure sense, Tz,w P WNHpC2qzNHpC2q if xx, yy " 0 and
Tz,w P NHpC2qzSNHpC2q if xx, yy ‰ 0. In particular, there are numerically hypercyclic operators on C2
which are not strongly numerically hypercyclic and there are weakly numerically hypercyclic operators
on C2 that are not numerically hypercyclic.
Due to Le´on-Saavedra and Muller [12], if T P LpXq is hypercyclic or weakly hypercyclic, then so is zT
for each z P T. Moreover, Theorem KPS2 ensures that if a weighted shift T is numerically hypercyclic,
then so is zT for each z P C satisfying z ě 1. In general, the scalar multiples of numerically hypercyclic
operators do not exhibit this nice behavior even in the friendly realm of diagonal operators on C2.
Proposition 1.18. For every T P LpC2q, there is w P T such that wT R WNHpC2q.
Proposition 1.19. For each z, w P T, Mpz, wq " "r ą 1 : rz 0
0 rw P NHpC2q* is a Gδ-subset of R,
being either infinite or empty. There exist w, z P T such that Mpz, wq is a dense Gδ-subset of p1,8q and
therefore is uncountable. For every w, z P T, Mpz, wq has zero Lebesgue measure.
By Proposition 1.19, not a single diagonal T P LpC2q satisfies rT P NHpC2q for each r ą 1.
1.3
Infinite dimensional results
The following two theorems provide further sufficient conditions for weak/strong numeric hypercyclicity.
Theorem 1.20. Each of the following conditions is sufficient for the weak numeric hypercyclicity of
T P LpXq.
(1.20.1) There is a sequence tznunPN in σppTq such that 1 ă z1 ă z2 ă . . .
(1.20.2) There is r ą 1 such that σppTq X rT is infinite.
(1.20.3) There is a cyclic vector x for T satisfying lim inf
nÑ8
(1.20.4) There are x P X and an infinite set A Ď N such that }T nx} Ñ 8 as n Ñ 8, n P A and the sequence
1`}T nx}
}T n} " 0.
is weakly convergent but is not norm convergent.
}T nx}(nPA
T nx
The following examples show how to apply Theorem 1.20.
Example 1.21. I ` V P WNHpL2r0, 1sq, where V fpxq "şx
Proof. It is easy to verify (see, for instance, [3]) that }T n} Ñ 8 and }T ng} " op}T n}q, where T " I ` V
and gpxq " x for x P r0, 1s. Since g is a cyclic vector for T , (1.20.3) is satisfied and T P WNHpL2r0, 1sq
by Theorem 1.20.
0 fptq dt is the classical Volterra operator.
Example 1.22. Let a P Cr0, 1s be non-constant and }a} ą 1. Then M P WNHpCr0, 1sq, where M is the
multiplication by a operator: M f " af for f P Cr0, 1s.
Proof. Let K " tt P r0, 1s : aptq " }a}u. If K ‰ r0, 1s, we can pick f P Cr0, 1s such that K " tt P
r0, 1s : fptq " 0u and let X be the closed linear span of OpM, fq. It is easy to see that }M nX} " }a}n
for every n P Z` and }T nf} " op}a}nq as n Ñ 8. Thus (1.20.3) is satisfied for MX and therefore
MX P WNHpXq. By Proposition 1.2, M P WNHpCr0, 1sq. It remains to consider the case K " r0, 1s.
Let R " }a}. Since a is non-constant and a " R, Proposition 1.16 ensures that we can pick w, z P T such
that Rz, Rw P apr0, 1sq and tRnpzn` wnq : n P Z`u is dense in C. Choose s, t P r0, 1s such that apsq " Rz
and aptq " Rw and let 1 be the constant 1 function on r0, 1s. Set ϕ P Cr0, 1s to be ϕpfq " 1
2pfpsq`fptqq.
Clearly, p1, ϕq P ΠpCr0, 1sq and ϕpM n1q " 1
2 Rnpzn ` wnq for n P Z`. Thus OpM, 1, ϕq is dense in C and
therefore M P NHpCr0, 1sq Ă WNHpCr0, 1sq.
5
Theorem 1.23. Each of the following conditions is sufficient for the strong numeric hypercyclicity of
T P LpXq.
(1.23.1) X is reflexive and there are a Schauder basic sequence tenunPN in X and c P ℓ1
`pNq such that
T en " λnen with λn P C for each n P N and ! 8řj"1
cj λk
j : k P Z`) is dense in C.
(1.23.2) X is reflexive and there is a Schauder basic sequence tenunPN in X such that T en " λnen for each
n P N, where λn P C are such that λ1 ą 1, the sequence tλnunPN is pmaybe non-strictlyq increasing
and the numbers λn
λn
are pairwise distinct elements of T.
(1.23.3) There is λ P C such that λ ě 1 and T ´ λI is semi-Fredholm of positive index.
(1.23.4) X is reflexive and there is λ P C such that λ ě 1 and T ´ λI is semi-Fredholm of negative index.
The reflexivity condition in the above theorem can not be removed entirely as illustrated by the
following examples.
cnzk
` such that f acts according to the formula fpyq " řnPA
Example 1.24. Let tznunPN be a sequence of finite order elements of T, r ą 1 and T P Lpc0pNqq be the
diagonal operator with the numbers rzn on the diagonal. Then T R NHpc0pNqq.
Proof. Let px, fq P Πpc0pNqq and A " tn P N : xn " 1u. Clearly A is non-empty and finite and
there is c P RA
cnxnyn for each y P c0pNq. Hence
fpT kxq " rk řnPA
n for each k P Z`. Since A is finite and each zn has finite order, OpT, x, fq is contained
in the union of finitely many lines in C through the origin. Hence OpT, x, fq is nowhere dense in C and
T R NHpc0pNqq.
Example 1.25. Let r ą 1 and T R NHpc0pNqq be given by pT xq1 " 0 and pT xqn " rxn´1 if n ą 1. Then
T ´ I is Fredholm of index ´1 and T R NHpc0pNqq.
Proof. Let px, fq P Πpc0pNqq and the finite set A Ă N and c P RA
` be as in the proof of Example 1.24.
It is easy to see that fpT kxq " 0 if k ě maxpAq. Thus OpT, x, fq is a finite set. Hence T R NHpc0pNqq.
On the other hand, T ´ I is injective and pT ´ Iqpc0pNqq is exactly the kernel of ϕ P c0pNq given by
ϕpxq "
r´nxn. Thus T ´ I is Fredholm of index ´1.
8řn"1
Obviously, a self-adjoint operator T on a Hilbert space H can not be numerically hypercyclic. Indeed,
NOpT, xq Ă R for each x P SpHq. The following theorem nearly (but not quite) characterizes weakly
numerically hypercyclic normal operators. It also allows a funny characterisation of weakly numerically
hypercyclic self-adjoint operators.
Theorem 1.26. Let H be a Hilbert space, T P LpHq and k P N be such that T k is normal. Then the
following statements are true.
(1.26.1) T P WNHpHq if there is a sequence tλnunPN in σpTq such that 1 ă λ1 ă λ2 ă . . .
(1.26.2) T P NHpHq if there is a sequence tλnunPN in σpTq such that 1 ă λ1 ă λ2 ă . . . and the numbers
λj
λj
are pairwise distinct.
(1.26.3) If T k is self-adjoint, then T P WNHpHq if and only if the set t´z : z P σpTq,
(1.26.4) T P NHpHq if there exist r ą 1 and a T k-invariant subspace K such that 1
well-ordered by the natural ordering of R.
operator with infinite spectrum.
z ą 1u is not
r T kK P LpKq is a unitary
Corollary 1.27. A self-adjoint operator T is weakly numerically hypercyclic if and only if the set
t´z : z P σpTq, z ą 1u is not well-ordered by the natural ordering of R.
Corollary 1.28. Let r ą 1 and T be a unitary operator with infinite spectrum. Then rT is numerically
hypercyclic.
6
Herrero [8, 9] described the norm closure of the set of hypercyclic operators on ℓ2pNq in terms of the
spectrum. Prajitura [14] demonstrated that the set of weakly hypercyclic operators on ℓ2pNq has the
same closure. We provide a spectral description of the closures of SNHpXq, WNHpXq and NHpXq for
an arbitrary reflexive Banach space X.
Theorem 1.29. Let X be reflexive and T P LpXq. Then the following statements are equivalent:
(1.29.1) T does not belong to the operator norm closure of SNHpXq;
(1.29.2) T does not belong to the operator norm closure of NHpXq;
(1.29.3) T does not belong to the operator norm closure of WNHpXq;
(1.29.4) the set tλ P σpTq : λ ě 1u consists of finitely many normal eigenvalues of multiplicity 1 with
pairwise distinct absolute values.
1.4 The structure
In Section 2 we prove Theorems 1.7 and 1.8 by means of explicit examples.
In the same section we
show that the existence of a numerical orbit NOpT, xq dense in a non-empty open subset of C does not
guarantee the numeric hypercyclicity of T P LpC2q. In Section 3 we prove Proposition 1.16 and build up a
toolbox used later on in order to prove the main results. Section 4 is devoted to the proof of Theorems 1.9
and 1.10, which provide sufficient conditions of weak and of strong numeric hypercyclicity in terms of a
restriction to a finite dimensional invariant subspace. In Section 5, we prove Propositions 1.11 and 1.12,
which deal with the similar issues in the Hilbert space setting. Theorem 1.20, providing genuinely
infinite dimensional sufficient conditions of weak numeric hypercyclicity, is proved in Section 6. We prove
Theorems 1.13, describing the numerically hypercyclic operators on C2, and Theorem 1.14, describing
the weakly numerically hypercyclic operators on C3, in Section 7. Theorem 1.23, which provides infinite
dimensional sufficient conditions for the strong numeric hypercyclicity, is proved in Section 8. In Section 9,
we prove Theorem 1.29, which describes the operator norm closure of the set of numerically hypercyclic
operators on a reflexive Banach space. Theorem 1.26, dealing with numeric hypercyclicity of normal
operators, is proved in Section10. In Section 11, we prove Propositions 1.18 and 1.19 on scalar multiples
of numerically hypercyclic operators on C2. In Section 12, we make some further remarks, construct few
examples and raise several questions on the numeric hypercyclicity.
2 Proof of Theorems 1.7 and 1.8
Our construction is similar to the proof of Theorem KPS1 in [10]. The main difference is that we need
and achieve better control of the numbers in the numerical orbit. Note also that there are natural
generalizations and some estimates below are far from optimal.
2.1 Funny numbers
As usual, for m, n P N, we write m n if m is a divisor of n. Let p be an odd prime number and νk " νkppq
for k P N be powers of p defined recurrently by the formula
ν1 " p and νk`1 " νkpk`νk for k P N.
Let also w " twkukPN be a fixed sequence of complex numbers satisfying
1
k ď wk ď k for each k P N.
We define a sequence tmk " mkpp, wqukPN of positive integers according to the following rule:
m1 " 1 and mk " mintm P N : m ą wk´1pk´1
π
, 2 m, p ffl mu for k ě 2.
(2.1)
(2.2)
(2.3)
7
Since the gaps between consecutive members of tm P N : 2m, p ffl mu do not exceed 4, we have
0 ď mk ´ wk´1pk´1
π
ă 4 for every k ě 2.
According to (2.2) and (2.4),
By (2.1) and (2.5), the series ř mk
νk
mk " Opkpkq as k Ñ 8.
converges, which allows us to define
τ " τpp, wq "
8řk"1
mk
νk P R and z " zpp, wq " eπτ i P T.
Lemma 2.1. In the above notation the following asymptotic relations hold:
lim
kÑ8`pνk1 ` zνk ´ wk " 0;
nRtνk :kPNu 1 ` zn1{n ě p´1{3.
lim inf
nÑ8
Proof. Note that for each k P N,
(2.4)
(2.5)
(2.6)
(2.7)
(2.8)
νkτ " Nk `
Since eπNki " ´1, we have zνk " ´ exp´πi
1 ` zνk " ´πi mk`1νk
8řs"k`1
νs
νs
msνk
msνk
, where Nk "
8řs"k`1
křs"1
νs ¯. Hence
νk`1 ` O´ mk`1νk
νk`1 ¯2
8řs"k`2
msνk
`
is an odd integer.
(2.9)
msνk
νs as k Ñ 8.
Multiplying by pνk and using (2.1), we get
pνk1 ` zνk " πmk`1
pk ` O m2
k`1νk
νk`1pk `
msνkpνk
νs as k Ñ 8.
8řs"k`2
By (2.1) and (2.5), the sequence in the O sign converges to 0 as k Ñ 8. Hence pνk1 ` zνk ´ πmk`1
as k Ñ 8. The inequality (2.4) yields πmk`1
pk ´ wk Ñ 0 as k Ñ 8 and (2.7) follows.
We prove (2.8) in two steps. Denote Λ " !νkm : k P N, m P 2N ` 1, m ă b νk`1
pk Ñ 0
νk ). First, we shall
verify that
If n P N is large enough, there is a unique k " kpnq ě 2 such that ?νk´1νk ă n ď ?νkνk`1. By (2.9),
lim
nÑ8
nRΛ 1 ` zn1{n " 1.
(2.10)
8řs"k`1
Using the estimate n ď ?νkνk`1, (2.1) and (2.5), we obtain
k¯.
νs " O´ 1
nτ " nNk
νk `
msn
ν 2
νs
msn
8řs"k`1
.
If n R Λ, n
νk
is not an odd integer. Since p ffl Nk, nNk
νk
nNk
νk
to the nearest odd integer is at least 1
νk
. Then by the last two displays,
can not be an odd integer. Hence the distance from
1 ` zn " 1 ` eπnτ i ě 1
2νk
if n R Λ is sufficiently large.
8
Since ?νk ă n, 2 ě 1 ` zn ě 1
In order to complete the proof of (2.8) and that of the lemma, it now suffices to show that
2n2 for all sufficiently large n R Λ, which immediately implies (2.10).
1 ` zn1{n ě p´1{3.
kÑ8`pνk1 ` zνk ´ wk " 0. Combining it with (2.2), we
k " Op1 ` zνkq and 1 ` zνk " Opkp´νkq as k Ñ 8. Hence, we can find a sequence rk of real
By the already verified equality (2.7),
lim inf
nÑ8
nPΛ
(2.11)
lim
have p´νk
numbers such that
zνk " ´eirk
for every k P N,
k " Oprkq and rk " Opkp´νkq as k Ñ 8.
p´νk
The last relation yields rk?νk`1νk Ñ 0 as k Ñ 8. Thus nrk ă π for n " νkm P Λ for all sufficiently
large k. Using the inequality eit ´ 1 ě 2
πt if t ď π, we see that if nrk ă π and n " νkm P Λ is
sufficiently large, then
1 ` zn " 1 ´ eirkm ě 2
πrkm ě c mp´νk
for some constant c ą 0 independent on n (c does exist because p´νk
k ě cp´νk
k " Oprkq). Hence
1 ` zn1{n ě pc{nq1{n`p´νk1{νkm " pc{nq1{np´1{m ě pc{nq1{np´1{3 Ñ p´1{3
n
since m ě 3. Thus the lower limit of 1 ` zn1{n is at least p´1{3 as n Ñ 8, n P Λ. This completes the
proof of (2.8) and that of the lemma.
Keeping the above notation, for each k P N, we choose qk P t0, 1, . . . , pk ´ 1u such that
wk1`zνk
wkp1`zνkq ´ exp´ 2πqki
wkp1`zνkq ´ exp´ 2πqi
The inequality 0 ď qk ă pk and (2.1) ensure that the series ř qk
pk ¯ " min" wk1`zνk
pkνk
pk ¯ : q P t0, 1, . . . , pk ´ 1u*.
converges. Thus we can define
θ " θpp, wq "
Lemma 2.2. In the above notation,
8řk"1
qk
pkνk P R, and u " upp, wq " e2πθi P T.
(2.12)
(2.13)
(2.14)
Proof. By (2.12), wk1`zνk
relations and the estimate wk ď k imply that
lim
kÑ8`uνk pνkp1 ` zνkq ´ wk " 0.
pk ¯ ď π
wkp1`zνkq ´ exp´ 2πqki
exp´ 2πqki
pk ¯pνkp1 ` zνkq ´ wk Ñ 0 as k Ñ 8.
uνk exp´´ 2πqki
νsps ¯ " 1 ` O´ 8řs"k`1
pk ¯ " exp´ 8řs"k`1
2πνkqki
By (2.13),
pk . By Lemma 2.1, `pνk1 ` zνk ´ wk Ñ 0. These two
(2.15)
νkqk
νsps¯.
8řs"k`1
νkqk
νsps " Opνk{νk`1q " Opp´k´νkq and pνk1 ` zνk " Opwkq " Opkq, we have
Since
puνk ´ expp2πqki{pkqqpνkp1 ` zqνk Ñ 0, which together with (2.15) implies (2.14).
9
2.2 T P LpC2q with a prescribed numerical orbit
The following result shows that the somewhere dense orbit theorem fails miserably for numerical orbits.
Proposition 2.3. Let a non-empty open subset U of C be the interior of a closed set. Then there exist
a diagonal T P LpC2q and x0 P SpC2q such that U is the interior of NOpT, x0q and for every x P SpC2q
either NOpT, xq " NOpT, x0q or NOpT, xq is nowhere dense in C.
Proof. Let p be an odd prime number. Choose a sequence w " twkukPN in C such that 1
k ď wk ď k
for every k P N and twk : k ě k0u is a dense subset of 2U for some k0 P N. By Lemmas 2.1 and 2.2,
(2.14) and (2.8) hold for νk " νkppq, z " zpp, wq and u " upp, wq defined in (2.1), (2.6) and (2.13)
respectively. Now let T P LpC2q be the diagonal operator with the numbers pu and puz on the diagonal.
If x " pa, bq P SpC2q, then
xT nx, xy " pnunpa2 ` b2znq for every n P Z`.
If a ‰ b, the above display yields xT nx, xy ě pna2 ´ b2 Ñ 8 and NOpT, xq is nowhere dense in
C. If a " b, then a " b " 1?2
and the above display reads
2xT nx, xy " pnunp1 ` znq for every n P Z`.
By (2.14), xT νk x, xy ´ wk
2 Ñ 0 as k Ñ 8. Since twk{2 : k ě k0u is a dense subset of U and U is the
interior of a closed set, U is the interior of the closure of txT νkx, xy : k P Nu. By the above display
and (2.8), xT nx, xy Ñ 8 as n Ñ 8, n R tνk : k P Nu. Hence U is the interior of the closure of
txT nx, xy : n P Nu " NOpT, xq. It remains to note that if a " b, then NOpT, xq does not depend on
x.
2.3 Proof of Theorem 1.7
Let p be an odd prime number. Choose a sequence w " twkukPN in C such that 1
k ď wk ď k for every
k P N and twk : k P Nu is dense in C. By Lemmas 2.1 and 2.2, (2.14) and (2.8) hold for νk " νkppq,
z " zpp, wq and u " upp, wq defined in (2.1), (2.6) and (2.13) respectively. Now let T P LpC2q be the
diagonal operator with the numbers pu and puz on the main diagonal.
, 1?2 P SpC2q, then 2xT nx0, x0y " pnunp1 ` znq for every n P Z`. By (2.14),
First, if x0 " ` 1?2
xT νkx0, x0y ´ wk
2 Ñ 0 as k Ñ 8. Since twk : k P Nu is dense in C, txT νk x0, x0y : k P Nu is dense in C.
Hence NOpT, xq is dense in C and T P NHpC2q.
If a ‰ b, as in the proof of Proposition 2.3, we have xT nx, xy ě
Let x " pa, bq P SpC2q.
pna2´b2 Ñ 8 and NOpT, xq is nowhere dense in C. Hence NOpT 2, xq, being a subset of NOpT, xq,
is nowhere dense in C. If a " b, then 2xpT 2qnx, xy " p2nu2np1 ` z2nq for n P Z`. Since each νk is odd,
nÑ8 1`z2n1{2n ě p2{3 ą 1. Hence xpT 2qnx, xy Ñ 8
(2.8) guarantees that lim inf
and NOpT 2, xq is nowhere dense in C. Thus T 2 R NHpC2q. The proof of Theorem 1.7 is complete.
2.4 Proof of Theorem 1.8
Let p and q be two odd prime numbers such that p ą q3{2, νk " νkppq and ν1k " νkpqq be as defined
in (2.1). Let also w " twkukPN and w1 " tw1kukPN be two sequences of complex numbers such that
k ď wk,w1k ď k for every k P N, twk : k P Nu is a dense subset of the half-plane tt P C : Re t ą 0u and
tw1k : k P Nu is a dense subset of the half-plane tt P C : Re t ă 0u. Let also z " zpp, wq, z1 " zpq, w1q,
u " upp, wq and u1 " upq, w1q be as defined in (2.6) and (2.13) and let T be the diagonal operator on C4
with the numbers pu, puz, qu1 and qu1z1 on the diagonal.
First, observe that if x, y P SpC4q are given by x "` 1?2
nÑ8 xpT 2qnx, xy1{2n " p lim inf
, 0, 0 and y "`0, 0, 1?2
, 1?2, then
, 1?2
1
2xT nx, xy " pnunp1 ` znq and 2xT ny, yy " qnpu1qnp1 ` pz1qnq for every n P Z`.
By Lemma 2.2, xT νkx, xy ´ wk
2 : k P Nu is dense
in C, NOpT, xq Y NOpT, yq is dense in C. Hence the union of 2 numerical orbits of T corresponding
2 Ñ 0. Since t wk
2 Ñ 0 and xT ν1
2 : k P Nu Y t w1
ky, yy ´ w1
k
k
10
to elements of ΠpC4q is dense in C. The proof will be complete if we prove that T is not numerically
hypercyclic. In order to do that, let x " pa, b, c, dq be an arbitrary element of SpC4q. It remains to show
that NOpT, xq is not dense in C. A direct computation yields
xT nx, xy " pnunpa2 ` b2znq ` qnpu1qnpc2 ` d2pz1qnq for every n P Z`.
(2.16)
Case 1: a ‰ b. By (2.16), xT nx, xy ě a2 ´ b2pnp1 ` Opqn{pnqq Ñ 8 as n Ñ 8 since q ă p.
Case 2: a " b ‰ 0 and c ` d ‰ 0. Since p ą q3{2, we can pick r P pq, p2{3q. By (2.8), there is
Hence NOpT, xq is non-dense in C.
c ą 0 such that pn1 ` zn ě crn for every n P Nztνk : k P Nu. According to (2.16), we have
xT nx, xy ě a2crn ´ 2qn Ñ 8 as n Ñ 8, n R tνk : k P Nu.
By (2.7), pνk1 ` zνk ´ wk Ñ 0. Since wk ď k, pνk1 ` zνk " Opkq as k Ñ 8. Pick s P p1, q2{3q.
If c " d ‰ 0, the obvious observation that ν1k ‰ νm for every k, m P N and (2.8) imply that there is
c1 ą 0 such that qνkpu1qνkpc2 ` d2pz1qνkq " qνkc21 ` pz1qνkq ě c1sνk for every k P N. If c ‰ d,
then qnpu1qnpc2 `d2pz1qnq ě c2 ´d2qn and therefore qνkpu1qνkpc2 `d2pz1qνkq ě c2 ´ d2qνk ě
c2 ´ d2sνk . Thus in any case there is c2 ą 0 such that qνkpu1qνkpc2 ` d2pz1qνkq ě c2sνk for every
k P N. Combining these estimates with (2.16), we get xT νk x, xy ě c2sνk ` Opkq Ñ 8 as k Ñ 8. Then
by the above display xT nx, xy Ñ 8 as n Ñ 8 and NOpT, xq is non-dense in C.
Case 3: a " b " 0 and c ‰ d. By (2.16), xT nx, xy ě c2 ´ d2qn Ñ 8 as n Ñ 8. Hence
NOpT, xq is non-dense in C.
Case 4: a " b " 0 and c " d. By (2.16), 2xT nx, xy " qnpu1qnp1 ` pz1qnq for n P N. Then by (2.8),
the lower limit of xT nx, xy1{n as n Ñ 8, n R tν1k : k P Nu is at least q2{3 ą 1. Hence xT nx, xy Ñ 8 as
2 ¯ " 0 . Since
n Ñ 8, n R tν1k : k P Nu. On the other hand, according to Lemma 2.2,
each w1k is in the left half-plane, NOpT, xq has no accumulation points in the open right half-plane. Thus
NOpT, xq is non-dense in C.
a " b and c " d " 0. By (2.16), 2xT nx, xy " pnunp1 ` znq for every n P N. Then
according to (2.8), the lower limit of xT nx, xy1{n as n Ñ 8, n R tνk : k P Nu is at least p2{3 ą 1. Hence
2 ¯ " 0.
xT nx, xy Ñ 8 as n Ñ 8, n R tνk : k P Nu. On the other hand, by Lemma 2.2, lim
Since each wk is in the right half-plane, NOpT, xq has no accumulation points in the open left half-plane.
Hence NOpT, xq is non-dense in C. The proof of Theorem 1.8 is complete.
kÑ8´xT νkx, xy´ wk
k x, xy ´ w1
Case 5:
kÑ8´xT ν1
lim
k
3 Dense sums of powers
We need a number of technical results. First, we state the Universality Theorem, which will be repeatedly
used in this and in the subsequent sections. A topological space X is called Baire if the intersection of
countably many dense open subsets of X is always dense in X. Recall that a Polish space is a separable
topological space X, whose topology can be defined by a metric d such that the metric space pX, dq is
complete. By the Baire Theorem, every Polish space is Baire. The following result is a particular case of
the Universality Theorem [6, p. 348 -- 349], see also [2, 7].
Theorem U. Let X be a Baire topological space, Y be a Polish space and F " tfa : a P Au be a collection
of continuous maps fa : X Ñ Y such that the set tpx, fapxqq : x P X, a P Au is dense in X Y . Then
x P X : tfapxq : a P Au is dense in Y( is a dense Gδ-subset of X. In particular, there is x P X such
that tfapxq : a P Au is dense in Y .
3.1 Finite sums
Lemma 3.1. Let z, w P T be independent and tRnunPN be a sequence of positive numbers such that
Rn Ñ 8. Then
pa, bq P T2 : tRnpazn ` bwnq : n P Z`u is dense in C(
is a dense Gδ-subset of T2.
11
Proof. Obviously, fnpa, bq " Rnpazn ` bwnq is a continuous map from T2 to C. By Theorem U, the proof
will be complete if we verify that
Λ " tpa, b, Rnpazn ` bwnqq : pa, bq P T2, n P Z`u is a dense subset of T T C.
and tn " v2Rn
nk P T, ak " sz´nk yk and bk " sw´nk y´1
Pick α, β P T and v P Czt0u and let s " v
. Clearly, tn Ñ 0 and v " 2Rnstn for each n P N.
v
Since z and w are independent, tpzn, wnq : n P Z`u is dense in T2 and there is a strictly increasing sequence
tnkukPN of positive integers such that tnk ă 1 for each k, znk Ñ isα´1 and wnk Ñ ´isβ´1 as k Ñ 8. Set
yk " tnk ` ib1 ´ t2
k . Then Rnkpakznk ` bkwnkq " 2stnk Rnk " v
and therefore pak, bk, vq P Λ for each k P N. Since yk Ñ i, znk Ñ isα´1 and wnk Ñ ´isβ´1, we have
ak Ñ α and bk Ñ β. Hence the sequence tpak, bk, vqu of elements of Λ converges to pα, β, vq. Since
α, β P T and v P Czt0u are arbitrary, Λ is dense in T T C.
Lemma 3.2. Let z, w P T be independent, k P N, m P Z, and tRnunPN, trnunPN be two sequences of
positive numbers such that Rn Ñ 8, rn Ñ 8 and Rn
pa, bq P T2 : tRnzmnpazkn ` a´1z´knq ` brnwn : n P Z`u is dense in C( is a dense Gδ-subset of T2.
Proof. Clearly, fnpa, bq " Rnzmnpazkn ` a´1z´knq ` brnwn is a continuous map from T2 to C. By
Theorem U, the proof will be complete if we verify that
rn Ñ 8. Then
Λ " tpa, b, Rnzmnpazkn ` a´1z´knq ` brnwnq : pa, bq P T2, n P Z`u is a dense subset of T T C.
Pick α, β P T and v P C. Since rn Ñ 8, for each sufficiently large n, the circle z´mnv ` rnT intersects
the real line at 2 points with exactly one sn being in p0,8q. Let un P T be the unique number such
that z´mnv ` rnun " sn. Obviously, un Ñ 1 and sn
rn Ñ 1
and rn
Rn ď 2 for all sufficiently large n. For such n, we can consider
qn " sn
Rn Ñ 0. Thus sn
i P T. Clearly, qn Ñ i as n Ñ 8. Let an " qnz´kn. It is easy to verify that
rn Ñ 1. Denote bn " ´zmnw´nun. Since sn
Rn Ñ 0, we have sn
2Rn `b1 ´ s2
n
4R2
n
Rnpanzkn ` a´1
n z´knq ` rnbnwn " v for each sufficiently large n.
Since z and w are independent, tpzkn, wnz´mnq : n P Z`u is dense in T2. Thus there is a strictly increasing
sequence tnjujPN of positive integers such that zknj Ñ iα´1 and wnj z´mnj Ñ ´β´1 as j Ñ 8. Since
un Ñ 1 and qn Ñ i, we have bnj " ´w´nj z´mnj unj Ñ β and anj " qnj z´knj Ñ α. By the above display,
panj , bnj , vq P Λ and therefore pα, β, vq is the limit of a sequence of elements of Λ. Since α, β P T and
v P C are arbitrary, Λ is dense in T T C.
Lemma 3.3. Let u, w, z P T be independent and tRnunPN be a sequence of positive numbers such that
Rn Ñ 8. Then there exist a, b, c P R` such that tRnpaun ` bwn ` cznq : n P Z`u is dense in C.
Proof. Let A " tpa, bq P R2 : a ą 0, b ą 0, a ` b ă 1u. Obviously, A is a non-empty open subset of R2.
Clearly, fn : A Ñ C, fnpa, bq " Rnpaun ` bwn ` cznq is continuous. By Theorem U, the proof will be
complete if we show that
Λ " tpa, b, Rnpaun ` bwn ` p1 ´ a ´ bqznqq : n P Z`, pa, bq P Au is dense in A C.
Clearly, M " tpα, β, γq P T3 : α ‰ β, α ‰ γ, β ‰ γu is a dense open subset of T3. Consider
F : R2 T3 Ñ C,
Fpx, ξq " x1ξ1 ` x2ξ2 ` p1 ´ x1 ´ x2qξ3.
Differentiating F , we easily confirm that
for every px, ξq P A M , the differentials of both Fpx,q at ξ and Fp, ξq at x have rank 2.
(3.1)
Furthermore, for every x P A, Fpx, Mq is an open subset of C containing 0.
12
Let x " pa, bq P A and y P C. Pick ξ " pα, β, γq P M such that Fpx, ξq " 0. According to (3.1), a
standard implicit function argument yields that
for every sequence tξj " pαj , βj , γjqujPN in T3 convergent to ξ and for every
sequence tzjujPN in C convergent to 0, there is a sequence txj " paj , bjqujPN in R2
convergent to x such that Fpxj , ξjq " zj for each sufficiently large j P N.
(3.2)
Rnj
Since u, w, z are independent, tpun, wn, znq : n P Z`u is dense in T3. Thus there is a strictly increasing
sequence tnjujPN of positive integers such that unj Ñ α, wnj Ñ β and znj Ñ γ. By (3.2), we can pick a
sequence txj " paj , bjqujPN in R2 convergent to x such that Fpxj , ξjq " y
for each sufficiently large j,
where ξj " punj , wnj , znjq. Since A is open, xj P A for each sufficiently large j. By the definition of F ,
Rnjpaj unj ` bj wnj ` p1 ´ aj ´ bjqznjq " y and therefore paj , bj , yq P Λ for each sufficiently large j. Since
paj , bj , yq Ñ pa, b, yq and x " pa, bq P A, y P C are arbitrary, Λ is dense in A C.
Lemma 3.4. Let z, w P T be independent, tRnunPN and trnunPN be sequences of positive numbers such
that Rn Ñ 8 and Rn
rn Ñ 8, U be a non-empty open subset of C and ϕ : U Ñ C be continuous. Then
a P U : tpRna ` rnϕpaqqzn ` pRna ` rnϕpaqqwn : n P Nu is dense in C( is a dense Gδ-subset of U .
Proof. Clearly, it is enough to prove the statement with U replaced by an arbitrary non-empty open set
whose closure is contained in U . Thus, without loss of generality, we can assume that ϕ is bounded:
c " suptϕpuq : u P Uu ă 8. Since z, w are independent, we can pick independent s, u P T such that
z " us and w " us´1. Then
for each n P N, where ρn " rn
pRna ` rnϕpaqqzn ` pRna ` rnϕpaqqwn " unRn`pa ` ρnϕpaqqsn ` pa ` ρnϕpaqqs´n
Λ " `a, unRn`pa ` ρnϕpaqqsn ` pa ` ρnϕpaqqs´n : a P U, n P N( is dense in U C.
. By Theorem U, it suffices to verify that
Rn
Fix a P Uzt0u and y P Czt0u. Since s, u are independent, tpsn, unq : n P Nu is dense in T2. Hence,
we can find a strictly increasing sequence tnkukPN of positive integers such that snk Ñ iaa and unk Ñ y
y
as k Ñ 8. Since ias´nk Ñ a P U and U is open, there is ε ą 0 such that for each sufficiently large k,
is´nkaeix P U for every x P r´ε, εs. Thus for each sufficiently large k, we can define the map
ξk : r´ε, εs Ñ R,
For every δ P p0, εq, we have
ξkpxq " ´ yRnk ` 2 Re`iaeix ` ρnk ϕpis´nkaeixqsnk.
min
xPr´ε,´δs
max
xPrδ,εs
ξkpxq ě 2a sin δ ´ yRnk ´ cρnk Ñ 2a sin δ ą 0 as k Ñ 8,
ξkpxq ď ´2a sin δ ´ yRnk ` cρnk Ñ ´2a sin δ ă 0 as k Ñ 8..
(3.3)
By (3.3), ξk takes both positive and negative values for each sufficiently large k. Since each ξk is con-
tinuous, the Intermediate Value Theorem guarantees the existence of xk P r´ε, εs such that ξkpxkq " 0.
Furthermore, the estimates (3.3) imply that xk Ñ 0 as k Ñ 8. Now we set ak " iaeixk s´nk P U and
yk " unk Rnk`pak ` ρnk ϕpakqqsnk ` pak ` ρnk ϕpakqqs´nk P C. Obviously pak, ykq P Λ. Since snk Ñ iaa
and xk Ñ 0, we have ak Ñ a. Next, by definition of ak,
yk " 2Rnk unk Repaksnk ` ρnk ϕpakqsnkq " 2Rnk unk Repiaeixk ` ρnk ϕpiaeixk s´nkqsnkq.
Using the above display together with the equality ξkpxkq " 0 and the definition of ξk, we arrive to
yk " unky. Since unk Ñ y
, we have yk Ñ y. Thus tpak, ykqu is a sequence in Λ convergent to pa, yq.
y
Since a P Uzt0u and y P Czt0u are arbitrary, Λ is dense in U C.
13
Lemma 3.5. Let 1 ă r ă R and k, m be distinct integers. Then
pz, wq P T2 : tRnpzkn ` zmnq ` rnwn : n P Z`u is dense in C( is a dense Gδ subset of T2.
Proof. By Theorem U, it suffices to verify that
n ` tm
Λ " tpz, w, Rnpzkn ` zmnq ` rnwnq : z, w P T, n P Z`u is dense in T T C.
Let a, b P T and y P C. Since k ‰ m, for every n P N, there are sn, tn P T such that sk
n " 0
n " 2. Consider the function Gn : T Ñ C, Gnpzq " Rnr´npzk ` zmq ´ r´ny. Then
and tk
rn ´ yrn ą 1 for each sufficiently large
Gnpsnq " r´ny ă 1 for each sufficiently large n and Gnptnq ě 2 Rn
n. Thus for n large enough the function Gn must attain the value 1. That is, there is cn P T such that
Gnpcnq " dn P T. Hence Rnpck
n q ` rndn " y for each sufficiently large n. Now pick zn P T and
wn P T such that zn
n " y, we
have pzn, wn, yq P Λ. Since zn Ñ a, wn Ñ b and a, b P T, y P C are arbitrary, Λ is dense in T T C.
n " dn, zn ´ a ă π
n and wn ´ b ă π
n . Since Rnpznk
n q ` rnwn
n " cn, wn
n ` znm
n ` sm
n ` cm
3.2 Proof of Proposition 1.16
Let UR " pz, wq P T2 : tRnpzn ` wnq : n P Z`u is dense in C(. We have to show that UR is a dense
Gδ-subset of T2 of Lebesgue measure 0. By Theorem U, in order to show that UR is a dense Gδ-subset
of T2, it suffices to verify that
Λ " tpz, w, Rnpzn ` wnqq : z, w P T, n P Z`u is dense in T T C.
Let y P C and z, w P T. Since T ` T Ą 2D, for every n satisfying 2Rn ą y, we can find an, bn P T such
n " bn, zn ´ z ă π
that Rnpan ` bnq " y. For each such n we can pick zn, wn P T for which zn
and wn ´ w ă π
nq, we have pzn, wn, yq P Λ for every sufficiently
large n. Since pzn, wn, yq Ñ pz, w, yq and y P C, z, w P T are arbitrary, Λ is dense in T T C. By
Theorem U, UR is a dense Gδ-subset of T2.
n . Since y " Rnpan ` bnq " Rnpzn
n " an, wn
n ` wn
n
Now let µ be the normalized Lebesgue measure on T2. A straightforward computation yields
µtpz, wq P T2 : zn ` wn ď cu " 2
π arcsin` c
2 for every c P r0, 2s and n P N.
(3.4)
Pick r P p1, Rq. By (3.4),
T2, zn ` wn ą rn
Rnzn ` wn Ñ 8 is incompatible with the membership in UR, µpURq " 0, as required.
Rn( ă 8. Hence for almost all pz, wq P
Rn for all sufficiently large n and therefore Rnzn ` wn Ñ 8. Since the condition
µ pz, wq P T2 : zn ` wn ď rn
8řn"1
3.3
Infinite sums and integrals
Recall that A Ď Z is called syndetic if there is a finite subset B of Z such that A ` B " Z. Equivalently,
A is syndetic if A " tnk : k P Zu, where nk`1 ą nk for each k P Z and tnk`1 ´ nk : k P Zu is bounded.
Lemma 3.6. Let G be a compact topological group, U be an open subset of G and g P G. Then the set
A " tn P Z : gn P Uu is either empty or syndetic. In particular, if U contains the identity element, then
A is always syndetic.
Proof. Since G is a compact topological group, the subgroup H " tgm : m P Zu is pre-compact with
respect to the induced topology. If V " U X H is empty, then A " ∅. Otherwise, V is a non-empty open
subset of H. Hence we can pick m P Z such that gmV is a neighborhood of the identity in H. Since
H is pre-compact, there is a finite set C " tgn1 , . . . , gnku Ď H such that CgmV " H. It follows that
B ` A " Z, where B " tn1 ` m, . . . , nk ` mu. Hence A is syndetic.
Lemma 3.7. Let M be a subset of T such that 1 is an accumulation point of M , A be a syndetic subset
of Z and k P N. Then tpzn
kq : n P A X N, z1, . . . , zk P Mu is dense in Tk.
1 , . . . , zn
14
Proof. Let u " pu1, . . . , ukq be an arbitrary point in Tk. Then we can pick θj P R such that uj " eθj i for
1 ď j ď k. Fix temporarily δ P p0, 1{2q. Since A is syndetic, there is q P N such that A ` t1, . . . , qu " Z.
Since 1 is an accumulation point of M , we can pick b1pδq, . . . , bkpδq P R such that 0 ă bkpδq ă δ,
0 ă bj´1pδq ă δbjpδq for 2 ď j ď k and ebjpδqi P M for 1 ď j ď k.
First, we choose the smallest positive integer m1 such that distpm1b1pδq ´ θ1, 2πZq ă b1pδq. Then
we pick the smallest integer m2 greater than m1 such that distpm2b2pδq´θ2, 2πZq ă b2pδq. We proceed
the same way. Namely, having chosen m1, . . . , mj´1, we set mj to be the smallest integer greater than
mj´1 such that distpmj bjpδq ´ θj, 2πZq ă bjpδq. By definition of q, we can pick n P A such that
mk ď n ď mk ` q ´ 1. Clearly, mj ´ mj´1 ď 2π
bjpδq
for 2 ď j ď k. Hence
k´j´2řm"1
δm ď q ´ 1 `
0 ď n ´ mj ď q ´ 1 `
křm"j`1
2π
bkpδq ď q ´ 1 ` 2π
bjpδq
2πδ
bjpδqp1´δq
for 1 ď j ď k,
where the third inequality is derived from bm´1pδq ă δbmpδq. Hence
nbjpδq ´ mjbjpδq ď pq ´ 1qbjpδq ` 2πδ
1´δ ď δpq ´ 1 ` 4πq for 1 ď j ď k
j ă ε for 1 ď j ď k.
j ´ uj ă ε and z´mzm
j ´ u2
since bjpδq ă δ. Since distpmj bjpδq´θj , 2πZq ă bjpδq, we have uj´emj bjpδqi ă bjpδq ă δ. According
to the above display, enbjpδqi ´ emj bjpδqi ă δpq ´ 1 ` 4πq. Hence uj ´ zn
j ă δpq ` 4πq, where zj " ebjpδqi.
Since zj P M , n P A X N and δ P p0, 1{2q is arbitrary, the desired result follows.
Corollary 3.8. Let M be an infinite subset of T, A be a syndetic subset of Z, k, n0 P N and u "
pu1, . . . , ukq P Tk. Then for every ε ą 0, there exist n, m P A and z, z1, . . . , zk P M such that m ą n ą n0,
z´nzn
Proof. Since A is syndetic, there is q P N such that A ` t1, . . . , qu " Z. Since T is compact and M is
infinite, there is an accumulation point v P T of M . Then for each δ P p0, 1q, 1 is an accumulation point
of Mδ " tv´1z : z P M, z ´ v ă δu. By Lemma 3.7, there are n P A and w1, . . . , wk P Mδ such that
n ą n0 and wn
v with zj P M and zj ´ v ă δ. Since v
j ´ uj ă δ, wj " zj
is an accumulation point of M , we can find z P M such that z ´ v ă δ
and z ´ v ă δ
j ă 4δ for
1 ď j ď k. Now pick m P A X t2n, 2n ` 1, . . . , 2n ` q ´ 1u. Since z ´ zj ď z ´ v ` v ´ zj ă 2δ, we
have z´2nz2n
j ´ uj ă 2δ for 1 ď j ď k. It follows that z´2nz2n
´ zm´2n ď pm ´ 2nqzj ´ z ă 2qδ. Hence
j ´ u2
Thus for δ satisfying p2q ` 4qδ ă ε, z´nzn
j ´ uj ă δ for 1 ď j ď k. Since wj P Mδ, wj " zj
2n , we have z´nzn
j " zm´2n
j ´ u2
j ă 4δ ` 2qδ " p2q ` 4qδ.
j ´ u2
j ă ε for 1 ď j ď k, as required.
j ` z´2nz2n
j ´ z´mzm
j ´ uj ă ε and z´mzm
j ´ z´mzm
z´mzm
2n . Since wn
j
j ă z´2nz2n
v
j ´ u2
The following lemma looks painfully technical. We need it nonetheless. For each ε ą 0, let
for 1 ď j ď 4u.
5 uj ´ e2πji{5 ă ε and u´1
10 u5`j ´ e4πji{5 ă ε
Pε " tu P T10 : u´1
(3.5)
Lemma 3.9. There exist ε, d ą 0 such that for every u P Pε and every z, w P d D, there are positive
numbers a1, . . . , a5 satisfying a1 ` . . . ` a5 " 1, a1u1 ` . . . ` a5u5 " z and a1u6 ` . . . ` a5u10 " w.
Proof. For every u P C10, consider the R-linear map Su : R4 Ñ C2 given by the formula
Sua "´ 4řj"1puj ´ u5qaj ,
4řj"1pu5`j ´ u10qaj¯.
Let ξ " e2πi{5 and for α, β P T, let upα, βq " pαξ, αξ2, αξ3, αξ4, α, βξ2, βξ4, βξ, βξ3, βq P T10. Then
Supα,βq " STα,β, where Tα,β : C2 Ñ C2, Tαpz, wq " pαz, βwq and
S : R4 Ñ C2, Sa " pξ ´ 1qa1 ` pξ2 ´ 1qa2 ` pξ3 ´ 1qa3 ` pξ4 ´ 1qa4
pξ2 ´ 1qa1 ` pξ4 ´ 1qa2 ` pξ ´ 1qa3 ` pξ3 ´ 1qa4 .
15
First, observe that Tα is an isometry. Identifying C2 with R4 in a natural way, we see that S is the linear
operator on R4 with the matrix
S "
cosp2π{5q ´ 1 cosp4π{5q ´ 1 cosp6π{5q ´ 1 cosp8π{5q ´ 1
cosp4π{5q ´ 1 cosp8π{5q ´ 1 cosp2π{5q ´ 1 cosp6π{5q ´ 1
sinp2π{5q
sinp4π{5q
sinp4π{5q
sinp8π{5q
sinp6π{5q
sinp2π{5q
sinp8π{5q
sinp6π{5q
‹‹‚.
One can verify (with some effort by hand and on the spot using an appropriate software) that the
matrix in the above display is invertible. Thus Supα,βq is invertible for each α, β P T. Since the set of
invertible R-linear operators from R4 to C2 is open, there is ε0 ą 0 such that Su is invertible for every
u P Pε0. Thus for each u P Pε0 and every pz, wq P C2, there exists a unique apu, z, wq P R4 such that
Suapu, z, wq " pz ´ u5, w ´ u10q. Obviously, the map pu, z, wq ÞÑ apu, z, wq is continuous on Pε0 C2. It
5.
is also straightforward to verify that Supα,βq` 1
5 , 1
By continuity, there exist ε P p0, ε0q and d ą 0 such that whenever u P Pε and z, w P d D, the components
of apu, z, wq are positive and sum up to a number less than 1. Now for z, w P dD and u P Pε, we set
4řj"1
aj " ajpu, z, wq for 1 ď j ď 4 and a5 " 1 ´
ajpu, z, wq. It is straightforward to see that the equality
Suapu, z, wq " pz ´ u5, w ´ u10q is equivalent to a1u1 ` . . . ` a5u5 " z and a1u6 ` . . . ` a5u10 " w, which
completes the proof.
5 " p´α,´βq. Hence apupα, βq, 0, 0q "` 1
5 , 1
5 , 1
5 , 1
5 , 1
5 , 1
ÿzPM
Lemma 3.10. Let M be an infinite countable subset of T, ω : R` Ñ R` be a strictly increasing continuous
concave function vanishing at 0 and tRnunPN be a sequence of positive numbers such that Rn Ñ 8. Then
there exists a P ℓ1
`pMq such that
Proof. First, note that QpMq " !a P ℓ1
řzPM
azzk : k P Z`) is dense in C.
ωpazq ă 8 and !Rk ÿzPM
ωpazq ă 8) equipped with the metric ρpa, bq "
`pMq : řzPM
ωpaz ´ bzq is a separable complete metric space. Let
Ω ""a P QpMq : lim inf
kÑ8 ÿzPM
"a P QpMq :ÿzPM
8ďk"m
Then Ω is a non-empty (0 P Ω) Gδ-subset of the Polish space QpMq (it is by no means dense!). Indeed,
azzk " 0*.
n*.
azzk ă
azzk " 0 for some k P N).
Ω " čm,nPN
Ω0 "!a P QpMq : ÿzPM
First, we observe that Ω0 is a subset of Ω. Indeed, let a P Ω0. Then there is k P N such that řzPM
azzk " 0.
Consider the compact metrizable topological group G " TM and g P G defined by gpzq " z for z P M .
By Lemma 3.6, for every neighborhood U of 1G in G there are infinitely many n P N such that gn P U .
Hence there is a strictly increasing sequence tnjujPN of positive integers such that gnj Ñ 1G. Then
Let also
1
lim
jÑ8 ÿzPM
azzk`nj " ÿzPM
azzk " 0 and therefore a P Ω.
Let d, ε ą 0 be the numbers provided by Lemma 3.9. First, we shall verify the following claim:
For each a P Ω, q P N, δ ą 0 and y P δd D, there exist n P N and b P QpMq
such that n ą q, ρpb, 0q ă 5ωpδq, a ` b P Ω0 and řzPMpaz ` bzqzn " y.
(3.6)
16
By Lemma 3.9, there exist positive numbers β1, . . . , β5 such that
azzn ă yδ ` 1
δpδd ´ yq ă d.
δpδd ´ yq " d,
y
azzn ď yδ ` 1
δ řzPM
δ řzPM
δ ´ 1
´ 1
azzm " 1
azzm ă 1
δ řzPM
δ řzPM
δ ÿzPM
5ÿj"1
y
δ ´
j "
βj zn
1
βj " 1,
5ÿj"1
Let a P Ω, q P N, δ ą 0 and y P δdD. Clearly, θ " δd ´ y ą 0. Denote A ""n P Z : řzPM
U " !f P G : řzPM
azzn ă θ*.
azfpzq ă θ). Since U is an open subset of the compact topological group G,
Since a P Ω, the set A is infinite and therefore non-empty. Clearly, A " tn P Z : gn P Uu, where
Lemma 3.6 implies that A is syndetic. Let ξ " e2πi{5. By Corollary 3.8, there exist m, n P A and
z, z1, . . . , z5 P M such that m ą n ą q and zn
j ´ zmξ2j ă ε{2 for 1 ď j ď 5. Then
1 , . . . , zn
pzn
5 q belongs to the set Pε defined in (3.5). Since n, m P A, we have
j ´ znξj ă ε{2, zm
1 , . . . , zm
5 , zm
azzn and
5ÿj"1
βj zm
j " ´
1
δ ÿzPM
azzm.
(3.7)
Define b P QpMq by setting bz " 0 if z P Mztz1, . . . , z5u and bzj " δβj for 1 ď j ď 5. The equation (3.7)
now reads řzPMpaz ` bzqzn " y and řzPMpaz ` bzqzm " 0. The latter implies a ` b P Ω0. Since the non-zero
components of b sum up to δ and there are only 5 of them ρpb, 0q ă 5ωpδq. This concludes the proof of
Claim (3.6).
Since Ω is a non-empty Gδ-subset of a Polish space QpMq, Ω is a Polish space in its own right. Since
for every k P N, the map fk : Ω Ñ C, fkpaq " Rk řzPM
azzk is continuous, Theorem U guarantees that the
proof will be complete if we show that
Λ ""´a, Rk ÿzPM
azzk¯ : a P Ω, k P N* is dense in Ω C.
Since Rn Ñ 8, (3.6) yields that for every δ ą 0, y P C and a P Ω, there are n P N and b P ℓ1
that ρpb, 0q ă δ (and therefore ρpa, a ` bq ă δ), a ` b P Ω and řzPMpaz ` bzqzn " y
Since a P Ω, y P C and δ ą 0 are arbitrary, Λ is dense in Ω C, which completes the proof.
Lemma 3.11. Let pΩ, F , µq be a measure space. Assume that the essential range of g P L8pµq contains a
sequence tλnunPN such that λ1 ą 1, tλnunPN is strictly increasing and λn
λn
there exist a non-negative real valued function a P L1pµq for which
`pMq such
Rn . Hence pa ` b, yq P Λ.
are pairwise distinct. Then
"żΩ
gka dµ : k P Z`* is dense in C.
Applying the above lemma to the situation of a purely atomic measure space, we immediately obtain
the following corollary.
Corollary 3.12. Let tλnunPN be a bounded sequence in C such that λ1 ą 1, the sequence tλnunPN is
strictly increasing and the numbers λn
λn
`pNq such that
are pairwise distinct. Then there exists a P ℓ1
n : k P Z`) is dense in C.
anλk
! 8ÿn"1
Proof of Lemma 3.11. Let rn " λn, zn " λn
and r " lim rn. Let Ω0 " tω P Ω : gpωq ă ru and Z be
λn
the set of all real-valued non-negative f P L1pµq vanishing outside Ω0. Clearly, Z is a closed subset of
17
the Banach space L1pµq. Since for every k P N, the map Fk : Z Ñ C, Fkpfq " şΩ gkf dµ is continuous,
Theorem U guarantees that the proof will be complete if we show that
Λ ""´a,ż gkf dµ¯ : a P Z, k P N* is dense in Z C.
Let U be an arbitrary non-empty open subset of Z C. Since rn Ñ r, we can find pa, yq P U and
m ě 2 such that a vanishes outside tω P Ω : gpxq ă rm´1u. Clearly M " tzj : j ą mu is an infinite
m´1. Since rm´1 ă rm, we can find
subset of T. Since g ď rm´1 on the support of a, ş agk dµ ď }a}1rk
3 and therefore
rq
m´1}a}1
q P N such that
m ` y
m ă 1
rq
rq
y
m ´
rk
m żΩ
1
rk
1
3
D for each k ě q.
An elementary geometric argument ensures that there is ε ą 0 (one can even find the biggest ε) such that
3 D is contained in the convex
hull of u, v and w. Corollary 3.8 applied with A " Z guarantees that for every integer s ě q there are
3 D is contained
s . Since us, vs, ws P M , there are j1,s, j2,s, j3,s P tm ` 1, m ` 2, . . . u
s and
agk dµ P
v ´ e2πi{3 ă ε and w
whenever u, v, w P T satisfy u
us, vs, ws P M and ks P N such that ks ą s, uks
in the convex hull of uks
such that us " zj1,s, vs " zj2,s and ws " zj3,s. Since 1
wks
s , by the above display, there exist α1, α2, α3 P R` such that
j2,s ` α3zks
s ´ e2πi{3 ă ε and uks
s ´ e4πi{3 ă ε. Thus 1
3 D is contained in the convex hull of uks
y
m ´
rks
m żΩ
1
rks
agks dµ.
α1 ` α2 ` α3 " 1 and α1zks
j1,s ` α2zks
v ´ e4πi{3 ă ε, the disk 1
s , vks
s and wks
j3,s "
s , vks
s
vks
s
vks
Using the definitions of rn and zn and multiplying by rks
m , we get
α1rks
m
rks
j1,s
λks
j1,s `
α2rks
m
rks
j2,s
λks
j2,s `
j3,s `żΩ
λks
α3rks
m
rks
j3,s
agks dµ " y.
Note that for every k P N, the set
"żΩ
bgk : b P Z, }b} " 1*
is convex and is a dense subset of the closed convex hull of the set
tzk : z is in the essential range of g and z ă ru.
Thus there is bs P Z such that
and
(3.8)
(3.9)
}bs}1 "
bsgks ´ α1rks
m
rks
j1,s
λks
j1,s `
żΩ
α1rks
m
rks
j1,s
α2rks
m
rks
j2,s
`
α3rks
m
rks
j3,s
`
α2rks
m
rks
j2,s
λks
j2,s `
α3rks
m
rks
j3,s
λks
j3,s ă 2´s.
The above display together with (3.8) gives
ys "żΩpas ` bsqgks dµ Ñ y as s Ñ 8.
Since rl,s ą rm for 1 ď l ď 3, (3.9) implies that }bs}1 Ñ 0 as s Ñ 8. By definition of ys, pa ` bs, ysq P Λ
for each s ě q. Since a ` bs Ñ a, ys Ñ y and U is open, Λ meets U . Since U is an arbitrary open subset
of Z C, Λ is dense in Z C, which completes the proof.
18
For the sake of brevity, we use the following fairly standard notation:
pµpnq "żT
zn dµpzq for n P Z for a Borel measure µ on T.
As usual, for g P L1pµq, gµ stands for the measure absolutely continuous with respect to µ with the
density (=Radon -- Nykodim derivative) g. For a non-negative Borel measure on T, we denote the set of
all real-valued non-negative f P L1pµq by the symbol L1
Lemma 3.13. Let µ be a purely non-atomic Borel probability measure on T and tRnunPN be a sequence
of positive numbers such that Rn Ñ 8. Then there is a P L1
`pµq such that tRnxaµpnq : n P Nu is dense in
C.
`pµq.
Proof. We start by proving the following claim.
For every purely non-atomic Borel probability measure ν on T
(3.10)
By the Fubini theorem,
and each ε ą 0, the set tn P Z : pνpnq ă εu is syndetic.
w´k dνpwq "żTT´ z
w¯k
zk dνpzqżT
dpν νqpz, wq.
pνpkq2 "pνpkqpνpkq "żT
m`n´1ÿk"m
1
n
For each m P Z and n P N, we sum up these equalities for m ď k ď m ` n ´ 1:
pνpkq2 "żTT
n´ z
w¯m n´1ÿk"0
For each δ P p0, 2q, we can split the above integral:
where hm,npz, wq "
1
hm,npz, wq dpν νqpz, wq,
zmpwn ´ znq
zk
wk "
n wm`n´1pw ´ zq
.
hm,npz, wq dpν νqpz, wq `żBδ
hm,npz, wq dpν νqpz, wq,
1
n
m`n´1ÿk"m
pνpkq2 "żAδ
żBδ
.
2
δn
2
δnpν νqpBδq ď
where Aδ " tpz, wq P T2 : z ´ w ă δu and Bδ " tpz, wq P T2 : z ´ w ě δu. Note that w ´ z ě δ and
wn ´ zn ď 2 for pz, wq P Bδ. Hence hm,npz, wq ď 2
δn for pz, wq P Bδ. Thus
hm,npz, wq dpν νqpz, wq ď
On the other hand, for every z, w P T and n P N, hm,npz, wq, being the average of several elements of T,
satisfies hm,npz, wq ď 1. Hence
hm,npz, wq dpν νqpz, wq ď pν νqpAδq.
żAδ
m`n´1ÿk"m
pνpkq2 ď
δÑ0pν νqpAδq " 0. Thus we can find δ P p0, 2q such that
Since ν is purely non-atomic, we have lim
2
δn ` pν νqpAδq.
Combining the last three displays, we get
1
n
pν νqpAδq ă ε2
2 . Having δ fixed, we pick n P N such that 2
δn ă ε2
2 . By the above display,
1
n
m`n´1ÿk"m
pνpkq2 ă
ε2
2 `
ε2
2 " ε2 for every m P Z.
19
We also need the following fact.
According to the above display, among n consecutive numbers pνpkq there is at least one strictly less
than ε. Hence the set tn P Z : pνpnq ă εu is syndetic, which proves Claim (3.10).
Let ν be a Borel probability measure on T with infinite support, r P p0, 1q,
m P N, A Ď Z be syndetic, and twnunPA, nąm be a sequence in D.
Then there exist a P L1
`pνq and n P A satisfying n ą m, }a}1 " 1 and xaνpnq " rwn.
(3.11)
zn ´ e2πij{q ă ε{2 for 1 ď j ď q. It follows that zn
Let K be the support of ν. For each n P Z denote K n " tzn : z P Ku. It is easy to see that the
`pνq, }a}1 " 1u is convex, is contained and is dense in the convex span of K n.
set Bn " txaνpnq : a P L1
In particular, Bn contains the interior of the convex span of K n. Since r ă 1, we can pick q P N large
enough in such a way that rD is contained in the interior of the convex span of the points e2πij{q for
1 ď j ď q. Then there is ε ą 0 such that rD is contained in the convex span of u1, . . . , uq whenever
uj
uq ´ e2πij{q ă ε for 1 ď j ď q ´ 1. By Corollary 3.8 we can find n P A and z, z1, . . . , zq P K such that
q ´ e2πij{q ă ε for 1 ď j ď q ´ 1. Thus
n ą m and zn
rD is in the convex span of zn
a P L1
`pνq, such that }a}1 " 1 and xaνpnq " rwn, which completes the proof of Claim (3.11).
Since each of the maps a ÞÑ Rnxaµpnq from L1
order to complete the proof Lemma 3.13 it suffices to verify that
Λ " tpa, Rnxaµpnqq : n P N, a P L1
`pµq, z P C and ε ą 0. By (3.10), A " n P Z : xaµpnq ă ε
Rnε ´ xaµpnqε
q . Hence rD Ă Bn. In particular, rwn P Bn and therefore there is
`pµq to C is continuous, Theorem U ensures that in
for every n P A satisfying n ą n0, where wn " z
wn P 1
`pνq and n P A satisfying n ą n0, }β}1 " 1 and xβµpnq " wn. Using the definition
By (3.11), there is β P L1
of wn, we can rewrite the last equality as xbµpnq " z, where b " a ` εβ. Then pb, zq P Λ and }b ´ a}1 " ε.
Since ε ą 0 is arbitrary, pa, zq belongs to the closure of Λ. Since a P L1
`pµq and z P C are arbitrary, Λ is
dense in L1
Rn Ñ 8, there is n0 P N such that z
4( is syndetic. Since
`pµqu is dense in L1
`pµq C.
4 for each n ě n0. Thus
Fix arbitrary a P L1
Rnε ă 1
2 D,
j
1 , . . . , zn
j
zn
.
`pµq C, which completes the proof.
4 Proof of Theorems 1.9 and 1.10
We start with the following lemma, which must be known. We include the proof for the reader's conve-
nience and since we were unable to find it in the literature anyway.
Lemma 4.1. Let te1, . . . , enu be a basis in an n-dimensional Banach space X and te1 , . . . , enu be the
dual basis in X : ej pekq " δj,k. Then for every c P Rn
` such that c1 ` . . . ` cn " 1, there is px, fq P ΠpXq
for which ejpxqfpejq " cj for 1 ď j ď n.
Proof. It is easy to check that tpe1pxqfpe1q, . . . , enpxqfpenqq : px, fq P ΠpXqu is a closed subset of Cn.
Hence it is enough to prove the result in the case cj ą 0 for every j. From now on we shall just assume
that cj ą 0 for 1 ď j ď n. Consider the continuous map
F : X Ñ r´8,`8q, Fpwq "
cj log ej pwq.
nřj"1
Since BpXq is compact and the map t ÞÑ Fptxq increases on r0, 1s for each x P SpXq, FBpXq
maximal value at some point x P SpXq. Obviously, ej pxq ‰ 0 for 1 ď j ď n. Consider
attains its
f P X,
fpyq "
cj e
j pyq
e
j pxq
for y P X.
nřj"1
20
Clearly, fpxq " 1 and ej pxqfpejq " cj for 1 ď j ď n. In order to clinch the proof it suffices to show that
f is bounded by 1 on SpXq. Indeed, then px, fq P ΠpXq and px, fq satisfies all desired conditions.
Assume the contrary. Then there is y P SpXq such that fpyq is real and fpyq ą 1. Clearly, x ` tu P
BpXq for every t P r0, 1s, where u " y ´ x. Moreover, fpuq " fpyq ´ fpxq " fpyq ´ 1 ą 0. Since FBpXq
attains its maximum at x, Fpx ` tuq ď Fpxq for t P r0, 1s and therefore
Fpx`tuq´Fpxq
t
ď 0.
lim sup
tÑ0`
e
j pxq
e
j pxq
let C´ " CzzR` and X´ " tw P X : ej pwq P
Choose z P T other than 1 or each of
C´ for 1 ď j ď nu. Set log to be the standard (real on the positive half of the real line) branch of the log-
cj log ejpwq.
Since x` tu P X´ for each sufficiently small t, the function Hptq " Gpx` tuq is defined in a neighborhood
of 0. Differentiating, we obtain
is exactly the real part of the function G : X´ Ñ C, Gpwq "
arithm on C´. Then FX´
nřj"1
,
Fpx`tuq´Fpxq
t
" Re H1p0q " Re
lim
tÑ0
" Re
nřj"1"
cj uj
e
j pxq`te
dt logpej pxq ` tejpuqqıt"0
d
nřj"1"cj
j puqıt"0 " Re
nřj"1
cj e
e
j puq
j pxq " Re fpuq " fpuq ą 0.
The obvious contradiction between the last two displays completes the proof.
1 ` . . . ` cnλk
Lemma 4.2. Let T P LpXq, λ1, . . . , λn P σppTq and c1, . . . , cn P R` satisfy c1 ` . . . ` cn " 1. Then there
exists px, fq P ΠpXq such that fpT kxq " c1λk
Proof. Without loss of generality we can assume that λj are pairwise distinct. Choose ej P SpXq such
that T ej " λjej for 1 ď j ď n. Since λj are pairwise distinct, ej form a basis of the n-dimensional space
E " spante1, . . . , enu. Let e1 , . . . , en be the dual basis in E: ekpejq " δi,j. By Lemma 4.1, there is
px, gq P ΠpEq such that ej pxqgpejq " cj for 1 ď j ď n. By the Hahn -- Banach Theorem, there is f P X
such that fE " g and }f} " 1. Since }x} " 1 and fpxq " gpxq " 1, px, fq P ΠpXq. Finally, using the
equalities ejpxqgpejq " cj, we get
n for every k P Z`.
fpT kxq " gpT nxq " gpe1pxqλk
1 e1 ` . . . ` enpxqλk
nenq " c1λk
1 ` . . . ` cnλk
n for each k P Z`.
4.1 Proof of Theorem 1.9
Let T P LpXq and λ1, λ2 P σppTq be such that λ1 " λ2 " R ą 1 and z " λ1
R are independent
in T. Then we can pick linearly independent e1, e2 P X such that T e1 " λ1e1 and T e2 " λ2e2. By
Lemma 3.1, there are a, b P T for which
R , w " λ2
O " tRnpazn ` bwnq : n P Z`u " taλn
1 ` bλn
2 : n P Z`u is dense in C.
By the Hahn -- Banach Theorem, there is f P X such that fpe1q " a and fpe2q " b. Then OpT, e1`e2, fq "
1 ` bλn
taλn
2 : n P Z`u is dense in C. By Proposition 1.5, T is weakly numerically hypercyclic. That is,
(1.9.1) is sufficient for the weak numeric hypercyclicity of T .
Assume that T P LpXq satisfies (1.9.2). That is, there exist independent λ1, λ2 P T such that
kerpT ´ λ1Iq2 ‰ kerpT ´ λ1Iq and kerpT ´ λ2Iq2 ‰ kerpT ´ λ2Iq. An elementary linear algebra
argument allows us to pick linearly independent e1, e2, e3, e4 P X such that pT ´λ1Iqe1 " pT ´λ2Iqe3 " 0,
pT ´ λ1Iqe2 " λ1e1 and pT ´ λ2Iqe4 " λ2e3. Then T ne2 " λn
2pe4 ` ne3q for
every n P N. By Lemma 3.1, there are a, b P T such that tnpaλn
2q : n P Nu is dense in C. By the
Hahn -- Banach Theorem, there is f P X for which fpe4q " fpe2q " 0, fpe1q " a and fpe3q " b. Then
fpT npe2 ` e4qq " npaλn
2q for every n P N. Hence OpT, e2 ` e4, fq is dense in C and T P WNHpXq
according to Proposition 1.5. Thus (1.9.2) is sufficient for the weak numeric hypercyclicity of T .
1pe2 ` ne1q and T ne4 " λn
1 ` bλn
1 ` bλn
21
λ2
has infinite order in T and s " λ1
Assume that T P LpXq satisfies (1.9.3). Then there exist λ1, λ2, λ3 P σppTq such that λ1 " λ2 "
r are independent. If s and u " λ2
R ą λ3 " r ą 1, λ1
R
are independent, then (1.9.1) is satisfied and therefore T is weakly numerically hypercyclic. It remains
to consider the case when s and u are not independent. In this case there are z, f, g P T and l, j P Z
such that f and g have finite order, s " f zl and u " gzj and l ě j. Since s
u has infinite order, we have
l ‰ j and therefore l ą j. Since s and y are independent, z and y are independent. Since f and g have
finite order, there is an even d P N such that f d " gd " 1. Then sdn " zknzmn, udn " z´knzmn and
ydn " wn for every n P Z`, where k " d
2pl ` jq P Z and w " yd P T. Since z and y are
independent, z and w are independent as well. By Lemma 3.2, there exist a, b P T such that
2pl ´ jq P N, m " d
R , y " λ3
O " tRdnzmnpazkn ` a´1z´knq ` brdnwn : n P Z`u is dense in C.
Since λ1, λ2 and λ3 are distinct eigenvalues of T , we can pick linearly independent e1, e2, e3 P X such
that T e1 " λ1e1, T e2 " λ2e2 and T e3 " λ3e3. By the Hahn -- Banach Theorem, we can find h P X for
which hpe1q " a, hpe2q " a´1 and hpe3q " b. Then for each n P Z`,
hpT dnpe1 ` e2 ` e3qq " aλdn
1 ` a´1λdn
2 ` bλdn
3 " Rdnzmnpazkn ` a´1z´knq ` brdnwn.
Thus O is a subset of OpT, e1 ` e2 ` e3, hq. Hence OpT, e1 ` e2 ` e3, hq is dense in C and T P WNHpXq
according to Proposition 1.5. That is, (1.9.3) is sufficient for the weak numeric hypercyclicity of T .
Finally, assume that T P LpXq satisfies (1.9.4). That is, there exist λ1, λ2 P σppTq and λ3 P T
such that λ1 " λ2 " R ą 1, λ1
R , λ3 are independent and
kerpT ´ λ3Iq2 ‰ kerpT ´ λ3Iq. Exactly as in the proof of the previous part, we can find d, k P N, m P Z
and independent z, w P T such that sdn " zknzmn, udn " z´knzmn and y " wn for every n P Z`, where
u " λ2
R and y " λ3. By Lemma 3.2, there exist a, b P T for which
has infinite order in the group T, s " λ1
λ2
O " tRdnzmnpazkn ` a´1z´knq ` bdnwn : n P Z`u is dense in C.
Pick linearly independent e1, e2, e3, e4 P X such that T e1 " λ1e1, T e2 " λ2e2, pT ´ λ3Iqe3 " 0 and
pT ´λ3Iqe4 " λ3e3. By the Hahn -- Banach Theorem, we can find h P X for which hpe1q " a, hpe2q " a´1,
hpe3q " b and hpe4q " 0. Then for each n P Z`,
1 ` a´1λdn
3 " Rdnzmnpazkn ` a´1z´knq ` bdnwn.
hpT dnpe1 ` e2 ` e4qq " aλdn
2 ` bdnλdn
Thus O is a subset of OpT, e1 ` e2 ` e4, hq. Hence OpT, e1 ` e2 ` e4, hq is dense in C and T P WNHpXq
according to Proposition 1.5. Thus (1.9.4) is sufficient for the weak numeric hypercyclicity of T . The
proof of Theorem 1.9 is complete.
1 ` . . . ` cnλk
4.2 Proof of Theorem 1.10
Assume that T P LpXq satisfies (1.10.1). By Lemma 4.2, there is px, fq P ΠpXq such that fpT kxq "
c1λk
n for each k P Z`. Hence OpT, x, fq is dense in C and T P NHpXq. Since every operator
similar to T satisfies the same conditions, T P SNHpXq. Thus (1.10.1) implies the strong numeric
hypercyclicity of T . For further references we prove the following slight modification of this result.
Lemma 4.3. Let T P LpCnq be the diagonal operator with the numbers λ1, . . . , λn on the diagonal.
Then T P NHpCnq if and only if T P SNHpCnq if and only if there exist c1, . . . , cn P R` such that
tc1λk
Proof. If there exist c1, . . . , cn P R` such that tc1λk
n : k P Z`u is dense in C, (1.10.1) is
satisfied, which implies that T P SNHpCnq and therefore T P NHpCnq. Now assume that T P NHpCnq.
Then there is x P SpCnq for which NOpT, xq is dense in C. A direct calculation shows that NOpT, xq "
tx12λk
n : k P Z`u is dense in C with cj " xj2.
n : k P Z`u. That is, tc1λk
n : k P Z`u is dense in C.
1 ` . . . ` xn2λk
1 ` . . . ` cnλk
1 ` . . . ` cnλk
1 ` . . . ` cnλk
Assume that T P LpXq satisfies (1.10.2). Then there exist λ1, λ2, λ3 P σppTq such that λ1 " λ2 "
R are independent. By Lemma 3.3, there are c1, c2, c3 P R` such that
1 ` c2λk
3 : k P Z`u is dense in C. Thus (1.10.1) is satisfied and T P SNHpXq. That is, (1.10.2)
λ3 " R ą 1 and λ1
tc1λk
implies the strong numeric hypercyclicity of T . The proof of Theorem 1.10 is complete.
2 ` c3λk
R , λ2
R , λ3
22
5 Proof of Propositions 1.11 and 1.12
Throughout this section H is a Hilbert space and T P LpHq.
5.1 Proof of Proposition 1.11
Assume that λ1, λ2 P σppTq are such that λ1 " λ2 " R ą 1, the eigenspaces kerpT ´ λ1Iq and
kerpT ´ λ2Iq are non-orthogonal and λ1
R are independent in T. Pick x, y P SpHq such that T x " λ1x,
T y " λ2y and xx, yy " c ą 0. Note that there is ε " εpcq ą 0 such that the set 1`cu
1`cu´1 : u P T( contains
Jε " teiθ : ´ε ă θ ă εu. By Lemma 3.1, there is a P Jε such that the set O " tλn ` aµn : n P Z`u is
dense in C. Since a P Jε, there is u P T such that a " 1`cu
}x`uy} P SpHq
and denote b " }x ` uy}´2. Then for each n P Z`,
1`cu´1 . Consider the vector q " x`uy
R , λ2
xT nq, qy " bxλnx ` uµny, x ` uyy " bpp1 ` cuqλn ` p1 ` cu´1qµnq " bpp1 ` cu´1qqpλn ` aµnq.
Thus the density of O implies the density of NOpT, qq and therefore T P NHpC2q. This completes the
proof of Proposition 1.11.
, λ2
λ2
5.2 Proof of Proposition 1.12
Assume that λ1, λ2 P C are such that kerpT ´ λj Iq2 ‰ kerpT ´ λjIq for j P t1, 2u, λ1 " λ2 ě 1
and λ1
are independent in T. Since kerpT ´ λj Iq2 ‰ kerpT ´ λj Iq, there are two 2-dimensional
λ1
T -invariant subspaces E1 and E2 of H such that E1 X E2 " t0u, pT ´ λj Iq2Ej " 0 and pT ´ λj IqEj ‰ 0
for j P t1, 2u. In particular, Lj " Ej X kerpT ´ λj Iq are one-dimensional for j P t1, 2u. If L1 and L2 are
non-orthogonal, T P NHpHq according to Proposition 1.11. It remains to consider the case when L1 and
L2 are orthogonal. It is straightforward to see that there are e1, . . . , e4 P SpHq such that e1 P L1, e3 P L2,
te1, e2u is a linear basis in E1, te3, e4u is a linear basis in E2, xe1, e2y " xe1, e3y " xe3, e4y " 0, while
g1,4 " xe1, e4y, g2,3 " xe2, e3y and g2,4 " xe2, e4y are non-negative real numbers. The above properties
of the restrictions of T to E1 and E2 yield the existence of non-zero α, β P C such that T e1 " λ1e1,
T e2 " λ1pe2 ` αe1q, T e3 " λ2e3 and T e4 " λ2pe4 ` βe3q. Then
1`a2 ` b2 ` g1,4ad ` g2,3bc ` g2,4bd ` αnbpa ` g1,4dq
xT nx, xy " λn
2`c2 ` d2 ` g1,4da ` g2,3cb ` g2,4db ` βndpc ` g2,3bq
` λn
for a, b, c, d P C and n P N, where x " ae1 ` be2 ` ce3 ` de4.
(5.1)
Consider the degree 2 R-polynomial map
Φ : C4 Ñ R2 C2, Φ
a
b
c
d
‹‹‚"
Repa2 ` b2 ` g1,4ad ` g2,3bc ` g2,4bdq
Repc2 ` d2 ` g1,4da ` g2,3cb ` g2,4dbq
αbpa ` g1,4dq
βdpc ` g2,3bq
‹‹‚.
First, observe that Φp1, 0, 1, 0q " p1, 1, 0, 0q. Looking at Φ as at a function of 8 real variables Re a, Im a,
Re b, Im b, Re c, Im c, Re d and Im d taking values in R6 with the components ordered as Φ1, Φ2, Re Φ3,
Im Φ3, Re Φ4, Im Φ4, we can calculate the Jacobi matrix d Φ of Φ at the point p1, 0, 1, 0q:
0
0
g2,3
g2,3
2 0
0 0
0 0
2 0
0 0 Re α ´Im α 0 0
0 0
0 0 Im α Re α
0 0 Re β ´Im β
0 0
0 0 Im β Re β
0 0
g1,4
g1,4
0
0
0
0
0
0
0
0
0
0
‹‹‹‹‹‹‚
.
d Φp1, 0, 1, 0q "
Since the complex numbers α and β are non-zero, it is an easy exercise to see that the rank of the above
6 by 8 matrix is 6. Indeed, after removing the two zero columns, we are left with an invertible 6 by 6
23
matrix. The Implicit Function Theorem says that there is ε ą 0 and a smooth (even real-analytic) map
Ψ : W " p1´ ε, 1` εqp1´ ε, 1` εq εD εD Ñ C4 such that Ψp1, 1, 0, 0q " p1, 0, 1, 0q and Φ Ψ " IdW .
Now we define
a, b, c, d : εD Ñ C,
papzq, bpzq, cpzq, dpzqq " Ψp1, 1, z, zq.
The equality Φ Ψ " IdW yields
αbpzqpapzq ` g1,4dpzqq " z,
Repapzq2 ` bpzq2 ` g1,4apzqdpzq ` g2,3bpzqcpzq ` g2,4bpzqdpzqq " 1,
Repcpzq2 ` dpzq2 ` g1,4dpzqapzq ` g2,3cpzqbpzq ` g2,4dpzqbpzqq " 1
βdpzqpcpzq ` g2,3bpzqq " z,
(5.2)
(5.3)
(5.4)
for every z P εD. Denote
ϕpzq " apzq2 ` bpzq2 ` g1,4apzqdpzq ` g2,3bpzqcpzq ` g2,4bpzqdpzq for z P εD.
Using (5.3) and (5.4), we obtain
ϕpzq " cpzq2 ` dpzq2 ` g1,4dpzqapzq ` g2,3cpzqbpzq ` g2,4dpzqbpzq for z P εD.
By the above two displays, (5.1) and (5.2),
xT nxpzq, xpzqy " λn
1pϕpzq ` nzq ` λn
where xpzq " apzqe1 ` bpzqe2` cpzqe3` dpzqe4. Applying Lemma 3.4 with z " λ1
λ1
rn " λ1n and U " εD, we find that there is u P εD such that
2pϕpzq ` nzq for every z P εD and n P N,
, w " λ2
λ2
, Rn " nλ1n,
O " tλn
1pϕpuq ` nuq ` λn
2pϕpuq ` nuq : n P Nu is dense in C.
By the above two displays NOpT, yq, being a positive scalar multiple of O, is dense in C, where y " xpuq
}xpuq}
Thus T P NHpHq. The proof of Proposition 1.12 is complete.
.
6 Proof of Theorem 1.20
Lemma 6.1. Let txnunPN be a weakly convergent sequence in a Banach spaces X, which is not norm-
convergent. Then for every sequence trnunPN of positive numbers such that rn Ñ 8, there is f P X such
that trnfpxnq : n P Nu is dense in C.
Proof. Let sn " suptdistpxm, Lnq : m ą nu, where Ln " spantu, x1, . . . , xnu and u is the weak limit of
txnu. Clearly, tsnu is a decreasing sequence of non-negative numbers and therefore sn Ñ 2s ě 0. It is
an elementary exercise to show that the equality s " 0 implies the norm convergence of txnu to u. Thus
s ą 0 and passing to a subsequence, if necessary, we can assume that distpxn`1, Lnq ě s ą 0 for every
n P N. This estimate combined with the Hahn -- Banach Theorem provides fn P X such that fnpxnq " 1,
fnpuq " 0, fnpxjq " 0 for j ă n and }fn} ď 1
s . Since fnpuq " 0 and xk Ñ u weakly, we have fnpxmq Ñ 0
as m Ñ 8. Thus we can choose a strictly increasing sequence tnkukPN of positive integers such that
fnkpxnmq ă 2´m whenever m ą k and rnk ą 2k for each k.
Now we shall verify that the map R : ℓ1pNq Ñ ℓ1pNq defined by pRaqm "
defined continuous linear operator. Indeed, since tfnku is bounded, for every a P ℓ1pNq, the series 8řk"1
is absolutely convergent in X and therefore defines ga "
is a continuous linear functional on ℓ1pNq. Next,
akfnk P X. Thus a ÞÑ pRaqm " gapxnmq
akfnkpxnmq is a well-
akfnk
8řk"1
8řk"1
}Ra ´ a}1 "
8ÿm"1pRaqm ´ am ď ÿk,mPN, k‰m
akfnkpxnmq.
24
Since fnkpxnmq ă 2´m for m ą k and fnkpxnmq " 0 for m ă k, the above display yields
}Ra ´ a}1 ď
8ÿk"1
ak
8ÿm"k`1
2´m ď
1
2
8ÿk"1
ak " }a}1
2
.
rnk
2 ă 1 and therefore R is invertible.
Thus R is bounded. Furthermore, }R ´ I} ď 1
Pick a dense sequence twkukPN in C such that wk ď k for every k P N. Then the sequence b defined
by bk " wk
belongs to ℓ1pNq since rnk ą 2k. Since R is invertible, there is a P ℓ1pNq such that Ra " b.
That is, gapxnmq " bm for every m P N. According to the definition of bm, gaprnk xnkq " wk for each
k P N. Hence trngapxnq : n P Nu is dense in C.
Lemma 6.2. Let X and Y be Banach spaces and tTnunPN be a sequence of continuous linear operators
from Y to X. Assume also that
Ω ""py, fq P Y X : lim inf
nÑ8
1`fpTnyq
››Tn›› " 0* is not nowhere dense in Y X.
Then there is py, fq P Y X such that tfpTnyq : n P Nu is dense in C.
Proof. Since Ω is not nowhere dense, Ω X U is dense in U for some non-empty open set U Ă Y X.
Let V be an arbitrary non-empty open subset of U . We shall prove that
there exist n P N and px, fq P V such that n ą m and fpTnxq " z.
(6.1)
Since Ω X U is dense in U , we can pick ph, uq P Ω X V . By definition of Ω, there is an infinite subset A
of N such that
nPA }Tn} " 8 and
lim
nÑ8
lim
nÑ8
nPA
hpTnuq
}Tn} " 0.
For every n P A, choose yn P SpY q for which 2}Tnyn} ě }Tn}. For each n P A, the Hahn -- Banach Theorem
provides hn P SpXq such that hnpTnynq " }Tnyn}. In particular, 2hnpTnynq ě }Tn} Ñ 8 as n Ñ 8,
n P A.
Case 1: lim inf
nÑ8
nPA
hpTnynq
hnpTnynq ą 0.
In this case there are an infinite subset B of A and c ą 0 such that hpTnynq ą 2chnpTnynq for
each n P B. For n P B denote δn " z´hpTnuq
. From the above display and the inequality hpTnynq ą
hpTnynq
2chnpTnynq ě c}Tn} it follows that δn Ñ 0 as n Ñ 8, n P B. Then we can choose n P B such that
n ą m and pu ` δnyn, hq P V . Plugging the definition of δn into the expression hpTnpu ` δnynqq, we after
cancellations get hpTnpu ` δnynqq " z. Thus n and px, fq " pu ` δnyn, hq satisfy (6.1).
Case 2: lim inf
nÑ8
nPA
hnpTnuq
hnpTnynq ą 0.
In this case there are an infinite subset B of A and c ą 0 such that hnpTnuq ą 2chnpTnynq for
each n P B. For n P B denote δn " z´hpTnuq
. From the above display and the inequality hpTnynq ą
hnpTnuq
2chnpTnynq ě c}Tn} it follows that δn Ñ 0 as n Ñ 8, n P B. Then we can choose n P B such that
n ą m and pu, h ` δnhnq P V . Plugging the definition of δn into the expression ph ` δnhnqpTnuq, we after
cancellations get ph ` δnhnqpTnuq " z. Thus n and px, fq " pu, h ` δnhnq satisfy (6.1).
hnpTnynq Ñ 0 and hnpTnuq
Case 3: hpTnynq
Fix δ ą 0 such that pu ` δBpY qq ph ` δBpXqq Ď V . In our case hnpTnpu ` δynqq " hnpTnuq `
as n Ñ 8, n P A.
hnpTnynq Ñ 0
δhnpTnynq ‰ 0 for all sufficiently large n P A. This allows us to define
hnpTnpu`δynqq " z´hpTnuq´δhpTnynq
δhnpTnynq`hnpTnuq
rn " z´hpTnpu`δynqq
.
hnpTnynq Ñ 0 and hnpTnuq
Since hpTnynq
as n Ñ 8, n P A, we have rn Ñ 0 as n Ñ 8, n P A.
Thus we can pick n P A such that n ą m and rn ă δ. Set x " u ` δyn and f " h ` rnhn. Since
hnpTnynq Ñ 0
25
pu ` δBpY qq ph ` δBpXqq Ď V , we have px, fq P V . Finally, plugging the definition of rn into the
expression fpTnxq " ph` rnhnqpTnu` δTnynq, we after cancellations get fpTnxq " z. This completes the
proof of Claim (6.1).
According to (6.1),
Λ " tpy, f, fpTnyqq : py, fq P U, n P Z`u is dense in U C.
By Theorem U, W " py, fq P U : tfpTnyq : n P Z`u is dense in C( is a dense Gδ subset of U .
particular, W ‰ ∅ and there is py, fq P Y X such that tfpTnyq : n P Nu is dense in C.
Corollary 6.3. Let T P LpXq, A be an infinite subset of N and txkukPN be a sequence in X such that
T nxk ‰ 0 for every pn, kq P Z` N and
In
}T nx1} Ñ 8 and for each k P N, }T nxk`1}
}T nxk} Ñ 8 as n Ñ 8, n P A.
(6.2)
Then T P WNHpXq.
Proof. Let E " spantxk : k P Nu and Y be the closure of E. By (6.2), }T nx} " o`››T nY›› for each x P E
as n Ñ 8, n P A. By Lemma 6.2, applied to Tn " T nY : Y Ñ X, there are px, fq P Y X such that
tfpT nxq : n P Z`u " OpT, x, fq is dense in C. Hence T P WNHpXq.
}T n} " 0) is dense in X. Then
Corollary 6.4. Let T P LpXq be such that !x P X : lim inf
T P WNHpXq.
Proof. Just apply Lemma 6.2 with Y " X and Tn " T n.
1`}T nx}
nÑ8
n
zk
}T nxk} " zk`1
Assume that (1.20.1) is satisfied. Then there is a sequence tznunPN in σppTq such that 1 ă z1 ă
z2 ă . . . Pick xk P SpXq such that T xk " zkxk for each k P N. Then }T nx1} " z1n Ñ 8 and
}T nxk`1}
Ñ 8 as n Ñ 8. By Corollary 6.3, T P WNHpXq. Thus (1.20.1) implies the weak
numeric hypercyclicity of T .
Assume that (1.20.3) is satisfied. That is, there is a cyclic vector x for T satisfying lim inf
nÑ8
1`}T nx}
}T n} " 0.
1`}T ny}
Let p P Crzs and y " ppTqx. Then }T ny} ď }ppTq}}T nx} and therefore lim inf
}T n} " 0. Since the
nÑ8
set tppTqx : p P Crzsu is dense in X, T P WNHpXq by Corollary 6.4. Thus (1.20.3) implies the weak
numeric hypercyclicity of T .
Assume that (1.20.4) is satisfied. That is, there are x P X and an infinite set A Ď N such that
}T nx}(nPA is weakly convergent but is not norm
}T nx} Ñ 8 as n Ñ 8, n P A and the sequence T nx
convergent. Let rn " }T nx} and xn " T nx
. Then trnunPA converges to 8, while txnunPA converges
}T nx}
weakly but not in the norm. By Lemma 6.1, there is f P X such that trnfpxnq : n P Au is dense
in C. Since trnfpxnq : n P Au is a subset of OpT, x, fq, OpT, x, fq is dense in C. By Proposition 1.5,
T P WNHpXq. Thus (1.20.4) implies the weak numeric hypercyclicity of T .
Finally, assume that (1.20.2) is satisfied. That is, there is r ą 1 such that σppTq X rT is infinite.
Then we can choose a countable infinite set M Ă T and ez P SpXq such that T ez " rzez for each z P M .
Note that ez are linearly independent. Let E " spantez : z P Mu. Then the closure of E is a closed
T -invariant subspace. By Proposition 1.6, it suffices to prove the weak numeric hypercyclicity of the
restriction of T to the closure of E. Thus, without loss of generality, we can assume that E is dense
1`}T nx}
}T n} " 0 for every x P E since }T nx} " Oprnq. In this case
in X. If lim sup
nÑ8
}T n}rn ă 8. Then there is c ě 1
T P WNHpXq by Corollary 6.4. It remains to consider the case lim sup
nÑ8
such that }T n} ď crn for each n P Z`. For every z P M , consider the unique linear functional ez : E Ñ C
such that ezpezq " 1 and ezpewq " 0 if w ‰ z. First, we shall prove that each ez is bounded and satisfies
}ez} ď c. Assume the contrary. Then there is z0 P M and x "
cj zj (zj are assumed pairwise distinct)
}T n}rn " 8, then lim inf
nÑ8
mřj"0
26
such that }x} " 1 and c0 " ez0pxq " a ą c. Since }T n} ď crn, we have }xn} ď c for each n P Z`, where
xn " z´n
0 ezj . Then
cj zn
0 r´nT nx " aez0 `
k´1ÿn"0
j z´n
mřj"1
››››
xn›››› "››››kaez0 `
0 " ck `
mÿj"1cj
k´1ÿn"0
zn
j zn
ka ď kc `
By the triangle inequality,
mÿj"1
cj´k´1ÿn"0
zn
j zn
0¯ezj›››› ď ck for each k P N.
mÿj"1
j z´k
cj1 ´ zk
0
0 ď ck `
1 ´ zj z´1
mÿj"1
2cj
zj ´ z0
for every k P N.
azzk : k P N) is dense in C. Since řzPM
Since the last sum does not depend on k, it follows that a ď c. This contradiction proves that each
ez is bounded and has the norm at most c. Hence we can uniquely extend each ez to X by continuity
?az ă 8
and thus treat them as elements of X. By Lemma 3.10, there is a P ℓ1
and O " !rk řzPM
?az ă 8, }ez} " 1 and }ez} ď c for every
z P M , the series x " řzPM
?azez converge absolutely in the Banach spaces X and
X respectively. Using the equalities T ez " rzez and ezpewq " δz,w for z, w P M , we have
fpT kxq " rkf´ ÿwPM
?azez and f " řzPM
?awwkew¯ " rk ÿz,wPM
?azawwkezpewq " rk ÿzPM
`pMq such that řzPM
azzk for each k P N.
Hence OpT, x, fq " O is dense in C. By Proposition 1.6, T P WNHpXq. That is, (1.20.2) implies the
weak numeric hypercyclicity of T . This completes the proof of Theorem 1.20.
7 Proof of Theorems 1.13 and 1.14
First, we shall establish some obstacles for an operator to be weakly numerically hypercyclic.
Lemma 7.1. Let T P LpXq and X " Y ' Z, where Y and Z are closed T -invariant subspaces of X such
have distinct absolute values. Then T R WNHpXq.
that TY P LpY q is power bounded, Z is finite dimensional and the distinct eigenvalues of TZ P LpZq
Proof. Since the eigenvalues of TZ have distinct absolute values, we can write σpTZq " tz1, . . . , zmu
with z1 ă . . . ă zm. Considering the Jordan normal form of TZ , it is easy to see that for every x P X
and f P X there exist y P Y , g P Y and polynomials p1, . . . , pm P Crzs such that
fpT nxq " gpT nyq `
pjpnqzn
j
for every n P Z`.
mřj"1
If p1 " . . . " pm " 0, tfpT nxqunPZ` " tgpT nyqunPZ` is bounded since TY is power bounded. Thus
OpT, x, fq is bounded in C. Otherwise, we can choose q P N such that 1 ď q ď m, pq ‰ 0 and pj " 0
whenever j ą q. Then
q´1řj"1
pjpnq` zj
zqn¯ for every n P Z`.
fpT nxq " gpT nyq ` zn
q´pqpnq `
Since pq ‰ 0, tgpT nyqunPZ` is bounded and zj
zq ă 1 for each j in the above sum, the last display ensures
that tfpT nxqunPZ` is bounded if zq ă 1 and if zq " 1 and pq is constant and that fpT nxq Ñ 8 as
n Ñ 8 otherwise. In any case OpT, x, fq is non-dense in C and therefore T R WNHpXq.
Lemma 7.2. Let f be a holomorphic function defined in a neighborhood of T and B be a discrete closed
subset of R p"a subset of R with no accumulation pointsq. Then BfpTq is nowhere dense in C.
27
Proof. If f is identically 0, the result is trivial. Otherwise f has only finitely many zeros s1, . . . , sk
(if any) in T. Looking at the Taylor expansion of f about sj, we see that for each j, fpzq " ajpz ´
sjqkjp1 ` Opz ´ sjqq as z Ñ sj, where aj P Czt0u and kj P N. Let v P Czt0u be an accumulation point
of BfpTq, which does not belong to BfpTq. Then it is easy to see that v " lim rnfpwnq, where trnunPN
and twnunPN are sequences in B and T respectively such that rn Ñ 8. Convergence of trnfpwnqu yields
fpwnq Ñ 0. Passing to a subsequence, if necessary, we can assume that wn Ñ sj as n Ñ 8 for some j.
Then rnfpwnq " rnajpwn ´ sjqkjp1 ` Opwn ´ sjqq as n Ñ 8. This equality together with rnfpwnq Ñ v
implies that ajv´1piwjqkj P R. That is v belongs to the union Λ of the lines ajpiwjqkj R for 1 ď j ď k.
Hence BfpTq Y Λ is closed in C. Furthermore, for every line L through the origin, which is not in Λ,
L X pBfpTq Y Λq is countable. It follows that BfpTq Y Λ has empty interior and therefore, being closed,
is nowhere dense in C. Thus BfpTq is nowhere dense in C.
Lemma 7.3. Let R ą r ą 1, a, b, c P C and k, m, j P Z. Then the set Ω " tRnpazk ` bzmq ` crnzj : n P
Z`, z P Tu is nowhere dense in C.
Proof. It is clear that the set of accumulation points of Ω that do not belong to Ω is contained in the set
q, where tzquqPN is a sequence in T and tnquqPN is a strictly increasing
Λ of lim
sequence of positive integers. Since the closure of Ω coincides with ΩYΛ, it suffices to show that ΩYΛ has
empty interior. Since the set Ω does not change if we replace z by z2, we can, without loss of generality,
assume that k and m are even.
qÑ8`Rnqpazk
q q ` crnq zj
q ` bzm
If a ‰ b, then every sequence in the last limit runs off to infinity and therefore Λ " ∅. Same happens
if a " b " 0 and c ‰ 0. In these cases ΩY Λ " Ω has empty interior as a union of countably many smooth
curves. If c " 0, the desired result follows from Lemma 7.2. The case k " m is trivial. Thus without loss
q P Λ.
of generality, we can assume that k ‰ m, a " b " 1 and c ‰ 0. Let v " lim
qÑ8
Since R ą r ą 1, this limit can only exist if azk
q ` bzm
q Ñ 0. Passing to a subsequence, if necessary, we
may assume that zq Ñ w and awk ` bwm " 0. Since k and m are even, we can write k " s ` t and
m " s ´ t, where s " k`m
P Zzt0u. As with every pair of elements of T, we can find
α, β P T such that a " βαt and b " βα´t. Then
P Z and t " k´m
q q`crnq zj
Rnqpazk
q `bzm
2
2
q ` bzm
Rnqpazk
q q ` crnq zj
q, where uq " αzq.
q ` u´t
q qus
q q ` drnq ul
q " βα´sRnqput
q ` u´t
q ` crnq α´j uj
q Ñ gv, where l " j ´ s, d " cαs´j β´1 and
Then uq Ñ x with xt ` x´t " 0 and Rnqput
g " αsβ´1. Since uq Ñ x, we can write uq " xeρq i, where tρqu is a convergent to 0 sequence of real
numbers. The equation xt`x´t " 0 implies that xt " i or xt " ´i. Passing to a subsequence, if necessary,
we can assume that either each ρq is positive or that each ρq is negative. We consider the case when each
ρq is positive and xt " ´i (the other 3 cases are similar). Since 2ixtRnq sinptρqq ` drnq xlelρq i Ñ gv, we
have 2Rnq sinptρqq ` drnq xlelρqi Ñ gv. Since elρqi Ñ 1, it follows that dxl " ´h, where h ą 0 (otherwise
2Rnq sinptρqq ` drnq xlelρq i Ñ 8). Thus p2Rnq sinptρqq ´ hrnq cosplρqqq ´ irnq sinplρqq Ñ gv. Hence
rnq sinplρqq Ñ ´Impgvq and therefore rnq sinptρqq Ñ ´ t
l Impgvq. Then either gv is real or trnq sinptρqqu
converges to a non-zero number. In the latter case the boundedness of t2Rnq sinptρqq ´ hrnq cosplρqqu
(which follows from the convergence) implies that R " r2.
Indeed, if R ‰ r2, one of the sequences
t2Rnq sinptρqqu and thrnq cosplρqqu runs to infinity faster than the other. In the case R " r2, by looking
at the Taylor series expansions of sin and cos, one sees that 2Rnq sinptρqq ´ hrnq cosplρqq " Opr´nqq and
therefore 2Rnq sinptρqq ´ hrnq cosplρqq Ñ 0. Hence Repgvq " 0. Thus in any case gv is either real or
purely imaginary. Hence v belongs to the union of two lines g´1R and ig´1R. Since there are finitely
many possibilities to choose x, Λ is contained in the union of finitely many lines through the origin. Thus
Ω Y Λ is a part of the union of countably many smooth closed curves and finitely many lines. Hence
Ω Y Λ has empty interior, as required.
Lemma 7.4. Let T P LpCnq be a diagonal operator with the diagonal entries d1, . . . , dn such that d1 "
. . . " dn ‰ 0 and for every j, k P t1, . . . , nu, dj
dj
are not independent. Then T R WNHpCnq.
and dk
dk
28
If R ď 1, then every numerical orbit of T is bounded and therefore T is not
Proof. Let R " d1.
weakly numerically hypercyclic. It remains to consider the case R ą 1. Since every pair dj
R is not
independent, there are w P T, u1, . . . , un P T of finite order and k1, . . . , kn P Z such that dj " Rujwkj for
1 ď j ď n. Since uj have finite order, we can pick m P N such that um
n " 1. It is easy to see
that every numerical orbit OpT, f, xq has the shape
1 " . . . " um
R , dk
OpT, f, xq " ta1dk
1 ` . . . ` andk
n : k P Z`u, where aj P C.
Taking into account that dj " Rujwkj and um
1 " . . . " um
n " 1, we see that
OpT, f, xq "
m´1ďq"0
Mq, where Mq " tRkm`qpa1uq
1wqk1wkmk1 ` . . . ` anuq
nwqknwkmknq : k P Z`u.
nřj"1
aj uq
j wqkj zkj . Clearly Mq Ď fqpTqBq,
Now for 0 ď q ď m´1, consider the Laurent polynomials fqpzq "
where Bq " tRkm`q : k P Z`u. By Lemma 7.2, each Mq is nowhere dense in C. By the above display
OpT, f, xq is nowhere dense in C. Hence T R WNHpCnq.
Lemma 7.5. Let T P LpC3q be a diagonal operator with non-zero diagonal entries d1, d2, d3 such that
for every j, k P t1, 2, 3u for which dj dk ‰ 0, dj
dj
Proof. If d1,d2,d3 are pairwise distinct, the result follows from Lemma 7.1. If d1 " d2 " d3, the
result follows from Lemma 7.4. Thus we can assume that d1 " d2 ‰ d3. It is easy to see that every
numerical orbit OpT, f, xq has the shape
OpT, f, xq " ta1dk
are not independent. Then T R WNHpC3q.
3 : k P Z`u, where aj P C.
1 ` a2dk
2 ` a3dk
and dk
dk
1 ` a2dk
If d1 " d2 ă d3 and a3 ‰ 0, then the sequence ta1dk
3u is bounded if d3 ď 1 and
converges to infinity otherwise, ensuring that OpT, f, xq is non-dense in C.
If a3 " 0, then we fall
under the jurisdiction of Lemma 7.4 with n " 2 and again OpT, f, xq is non-dense in C.
In the case
d1 " d2 ą d3 and d3 ď 1, one can proceed similar to the proof of Lemma 7.4 (we leave this easy bit
to the reader). It remains to consider the case d1 " d2 ą d3 ą 1 and a3 ‰ 0. Let R " d1 " d2 and
r " d3. Since every pair dj
is not independent, there is w P T, u1, u2, u3 P T of finite order
dj
and k1, k2, k3 P Z such that d1 " Ru1wk1, d2 " Ru2wk2 and d3 " ru3wk3. Since uj have finite order, we
1 " um
can pick m P N such that um
3 " 1. Then
2 ` a3dk
and dk
dk
2 " um
OpT, f, xq "
Mq, where
m´1ďq"0
Mq " tRkm`qpa1uq
1wqk1 wkmk1 ` a2uq
2wqk2wkmk2q ` rkm`qa3uq
3wqk3 wkmk3 : k P Z`u.
Now each Mq is contained in a set of the shape tRmkpαzk1 ` βzk2q ` rmnγzk3 : k P Z`, z P Tu with
α, β, γ P C being constants. Then Lemma 7.3 guarantees that each Mq is nowhere dense in C. By the
above display OpT, f, xq is nowhere dense in C. Hence T R WNHpC3q.
7.1 Proof of Theorem 1.13
, λ2
Let T P LpC2q and σpTq " tλ1, λ2u with λ1, λ2 P C. By Theorem 1.9, if λ1 " λ2 ą 1 and λ1
λ1
λ2
are independent, then T P WNHpC2q. If λ1 ‰ λ2 or λ1 " λ2, then T R WNHpC2q by Lemma 7.1.
If λ1 " λ2 ď 1 and λ1 ‰ λ2, then T is power bounded and therefore T R WNHpC2q. Finally, if
λ1 " λ2 ą 1, λ1 ‰ λ2 and λ1
are not independent, then T R WNHpC2q according to Lemma 7.4.
λ1
Thus T P WNHpC2q if and only if λ1 " λ2 ą 1 and λ1
λ1
2 : k P Z`u is dense in C, then T P SNHpC2q by Theorem 1.10. Assume now that
T P SNHpC2q. Then T P WNHpC2q and therefore λ1 " λ2 ą 1 and λ1 ‰ λ2. Then T is similar to the
are independent in T.
1 ` λk
If tλk
, λ2
λ2
, λ2
λ2
29
1 ` bλn
2 : n P Z`u is dense in C. Since aλn
diagonal operator with the numbers λ1 and λ2 on the diagonal. By Lemma 4.3, there exist a, b ě 0 such
2 ě a ´ bRn, it follows that a " b. Hence
1 ` bλn
that taλn
2 : n P Z`u is dense in C. Thus T P SNHpC2q if and only if σpTq " tλ1, λ2u with tλk
1 `λn
tλn
2 : k P Z`u
being dense in C.
Now assume that T P WNHpC2q and T is not unitarily equivalent to a diagonal operator. As we
have already shown, λ1 " λ2 ą 1 and λ1
are independent in T. Since T is not unitarily
λ1
equivalent to a diagonal operator, the eigenspaces kerpT ´ λ1Iq and kerpT ´ λ2Iq are non-orthogonal.
By Proposition 1.11, T P NHpC2q. Finally, assume that T P NHpC2q. Then T P WNHpC2q. If T is
unitarily equivalent to a diagonal operator, Lemma 4.3 implies that T P SNHpC2q. These observations
amount to the fact that T P NHpC2q if and only if either T P SNHpC2q or T P WNHpC2q and T is not
unitarily equivalent to a diagonal operator. The proof of Theorem 1.13 is complete.
1`λk
, λ2
λ2
, λ2
λ2
having infinite order in the group T and λ1
λ1
7.2 Proof of Theorem 1.14
Let T P LpC3q. If there are λ1, λ2 P σpTq such that λ1 " λ2 ą 1 and λ1
are independent in
λ1
T, then (1.9.1) is satisfied and Theorem 1.9 implies that T P WNHpC3q. If σpTq " tλ1, λ2, λ3u with
λ1 " λ2 ą λ3 ą 1, λ1
, λ3
being independent in T,
λ3
then (1.9.3) is satisfied and Theorem 1.9 implies that T P WNHpC3q. If distinct members of σpTq have
distinct absolute values, then T R WNHpC3q according to Lemma 7.1. If λj
are pairwise dependent for
λj
non-zero λj, Lemma 7.5 guarantees that T R WNHpC3q. Consider the case λ1 ą λ2 " λ3 " R and
either R ď 1 or λ2
R are not independent. If R ď 1, then each numerical orbit of T is either bounded
or escapes to infinity at the rate λ1n (provided λ1 ą 1). If R ą 1 and λ2
R are not independent,
then T R WNHpC3q. Indeed, each numerical orbit of T has the shape taλn
3 : n P Z`u if
λ2 ‰ λ3 and taλn
2 : n P Z`u if λ2 " λ3. In any case if a ‰ 0, this is a sequence escaping to
infinity. If a " 0 we fall under the jurisdiction of Lemma 7.4 and the numerical orbit is non-dense. If λ3
is 0, the problem is easily reduced to the 2-dimensional situation, already covered by Theorem 1.13. It
remains to notice that we have considered all the possibilities (up to the ordering of the eigenvalues).
R and λ3
1 ` bλn
2 ` cnλn
2 ` cλn
1 ` bλn
λ2
R and λ3
8 Proof of Theorem 1.23
`pNq be such that 8řn"1
Lemma 8.1. Let tenunPN be a Schauder basis in a reflexive Banach space X, tenunPN be the corresponding
sequence of coordinate functionals and a P ℓ1
an " 1. Then there is px, fq P ΠpXq
such that enpxqfpenq " an for each n P N.
Proof. Let Xn " spante1, . . . , enu. By Lemma 4.1, for every n P N, we can find pxn, gnq P ΠpXnq such
that ekpxnqgnpekq " ak for 1 ď k ă n and enpxnqgnpenq " 1 ´ a1 ´ . . . ´ an´1. By the Hahn -- Banach
Theorem, for each n P N, there is fn P SpXq such that fnXn " gn. Then for every n P N,
Since X is reflexive, every bounded sequence in X X has a weakly convergent subsequence. Hence
there is a strictly increasing sequence tnmumPN of positive integers such that txnmumPN converges weakly
to x P X and tfnmumPN converges weakly to f P X. Since BpXq is weakly closed in X and BpXq is
weakly closed in X, we have }x} ď 1 and }f} ď 1. According to the above display, for every k P N,
ekpxnmqfnmpekq " ak for all sufficiently large m. Hence,
pxn, fnq P ΠpXq and ekpxnqfnpekq " ak for 1 ď k ă n.
It follows that
ekpxqfpekq " lim
mÑ8
fpxq " f´ 8řk"1
ekpxqek¯ "
ekpxnmqfnmpekq " ak for each k P N.
8řk"1
ekpxqfpekq "
8řk"1
ak " 1.
Since fpxq " 1, }x} ď 1 and }f} ď 1, px, fq P ΠpXq. Thus px, fq satisfies all desired conditions.
30
Assume that T P LpXq satisfies (1.23.3). That is, there exists λ P C be such that λ ě 1 and
T ´ λI is a semi-Fredholm operator of positive index. Since the index is locally constant, T ´ zI is a
semi-Fredholm operator of positive index for every z P λ ` εD for a sufficiently small ε. Since every
semi-Fredholm operator of positive index is non-injective, z ` εD Ď σppTq. Since λ ě 1, there is r ą 1
such that J " pλ`εDqXprTq ‰ ∅. Then J is a non-trivial open arc of the circle rT. By Proposition 1.16,
we can find z, w P J Ă σppTq such that tzn ` wn : n P Z`u is dense in C. Thus (1.10.1) is satisfied and
therefore T P SNHpXq by Theorem 1.10. Hence (1.23.3) implies the strong numeric hypercyclicity of T .
Assume now that T P LpXq satisfies (1.23.4). That is, X be reflexive and there exists λ P C such
that λ ě 1 and T ´ λI is a semi-Frdholm operator of negative index. It is well-known and easy to see
that if X is reflexive and R P LpXq, then R is semi-Fredholm if and only if R is semi-Fredholm and
ipRq " ´ipRq. Thus T ´ λI is a semi-Fredholm operator of positive index. By the previous part of
the proof, T P SNHpXq. By Proposition 1.6, T P SNHpXq. Thus (1.23.4) implies the strong numeric
hypercyclicity of T .
Assume that T P LpXq satisfies (1.23.1). That is, X is reflexive and there is a Schauder basic sequence
tenunPN in X such that T en " λnen with λn P C for each n P N and for some c P ℓ1
`pNq,
! 8řj"1
cj λk
j : k P Z`) is dense in C.
(8.1)
Without loss of generality, we may assume that 8řn"1
cn " 1. Let Y be the closed linear span of the
sequence tenunPN. By Lemma 8.1, there is px, gq P ΠpY q such that enpxqgpenq " cn for every n P Z`,
where en P Y are the coordinate functionals for the Schauder basis tenunPN. By the Hahn -- Banach
Theorem, there is f P X such that fY " g and px, fq P ΠpXq. Then
8řn"1
nenpxqgpenq "
λk
fpT kxq " gpT kxq " g´ 8řn"1
nenpxqen¯ "
n for each k P Z`.
8řn"1
cnλk
λk
By (8.1), OpT, x, fq is dense in C and T P NHpXq. Since every operator similar to T satisfies the same
conditions, T P SNHpXq. Thus (1.23.1) implies the strong numeric hypercyclicity of T .
Finally, assume that T P LpXq satisfies (1.23.2). That is, X is reflexive and there is a Schauder
basic sequence tenunPN in X such that T en " λnen where λn P C are such that λ1 ą 1, the sequence
tλnunPN is pmaybe non-strictlyq increasing and the numbers λn
are pairwise distinct elements of T.
λn
Passing to a subsequence, if necessary, we can assume that either tλnunPN is strictly increasing or that
tλnunPN is constant. If tλnunPN is strictly increasing, Corollary 3.12 provides c P ℓ1
`pNq satisfying (8.1).
If λn " r ą 1 for every n P N, by Lemma 3.10, there is c P ℓ1
`pNq for which (8.1) holds. In any case
(1.23.1) is satisfied and therefore T P SNHpXq. Thus (1.23.2) implies the strong numeric hypercyclicity
of T . This completes the proof of Theorem 1.23.
9 Proof of Theorem 1.29
Lemma 9.1. Let px, fq P ΠpXq. Then for every y P SpXqXker f , there exists g P X such that gpxq " 0,
gpyq " 1 and }g} ď 2.
Proof. Let E " spantx, yu. Since fpxq " 1 and fpyq " 0, x and y are linearly independent and therefore
form a basis in E. Consider ϕ P E defined by ϕpxq " 0 and ϕpyq " 1. Then u " fpuqx ` ϕpuqy for
every u P E. Hence ϕpuq " }u ´ fpuqx} ď }u} ` }f}}u} " 2}u}. That is, }ϕ} ď 2. By the Hahn -- Banach
Theorem, there is g P X such that }g} ď 2 and gE " ϕ. Obviously, g satisfies all desired conditions.
dimensional subspace E of X such that }TE}LpE,Xq ă ε.
Proof. Since TpXq is non-closed, inft}T x} : x P SpXqu " 0. Hence we can pick x P SpXq for which
4 . By the Hahn -- Banach theorem there is f P X such that px, fq P ΠpXq. Next, observe that
}T x} ă ε
Lemma 9.2. Let T P LpXq be such that TpXq is non-closed in X and ε ą 0. Then there is a 2-
31
inft}T x} : x P SpXqX ker fu " 0. Indeed, otherwise Tpker fq is closed in X and is a subspace of TpXq of
codimension at most 1. Since a subspace of a Banach space with a finite codimensional closed subspace
is closed itself, that would have implied that TpXq is closed.
}T y} ă ε
E " spantx, yu is a 2-dimensional subspace of X. Moreover, for each u P BpEq,
4 ` 2 ε
Thus inft}T x} : x P SpXq X ker fu " 0 and therefore we can choose y P SpXq X ker f such that
4 . By Lemma 9.1, there is g P X such that }g} ď 2, gpxq " 0 and gpyq " 1. Clearly
}T u} " }fpuqT x ` gpuqT y} ď }f}}T x} ` }g}}T y} ă ε
4 " 3ε
4 .
4 ă ε.
Lemma 9.3. Let T P LpXq, E be a 2-dimensional subspace of X and S : E Ñ X be a linear map. Then
Hence }TE}LpE,Xq ď 3ε
there is R P LpXq such that RE " S and }T ´ R} ď 3}S ´ TE}LpE,Xq.
Proof. Take any x P SpEq. By the Hahn -- Banach Theorem, there exists f P X such that px, fq P ΠpXq.
Pick y P SpEq such that fpyq " 0. By Lemma 9.1, there is g P X such that }g} ď 2, gpxq " 0 and
gpyq " 1. Now we define R P LpXq be the formula Ru " T u ` fpuqpSx ´ T xq ` gpuqpSy ´ T yq. It is
straightforward to see that Rx " Sx and Ry " Sy. Since tx, yu is a basis in E, RE " S. Next, for every
u P X, using the relations }x} " }y} " }f} " 1 and }g} ď 2, we have
Hence }T ´ R} ď 3}S ´ TE}LpE,Xq.
}pT ´ Rqu} " }fpuqpT ´ Sqx ` gpuqpT ´ Sqy} ď 3}u} }S ´ TE}LpE,Xq.
Lemma 9.4. Let E be a 2-dimensional Banach space and T P LpEq be such that σpTq " ts, tu and
s " t ě 1. Then there is a sequence tTnunPN in SNHpEq such that }Tn ´ T} Ñ 0.
Proof. Since s " t ě 1, by Proposition 1.16, we can choose two sequence tsnunPN and ttnunPN such
that sn Ñ s, tn Ñ t and tsk
n : k P Z`u is dense in C for every n P N. Since σpTq " ts, tu, we
can choose a linear basis tx, yu in E such that T x " sx and T y " ty ` ax for some a P C. For each
n P N consider Tn P LpXq given by Tnx " snx and Tny " tny ` ax. Since sn Ñ s and tn Ñ t, we have
}Tn ´ T} Ñ 0. On the other hand, σpTnq " tsn, tnu for each n P N. By Theorem 1.13, each Tn belongs
to SNHpEq.
n ` tk
Since WNHpXq Ě NHpXq Ě SNHpXq, we have (1.29.3)ùñ(1.29.2)ùñ(1.29.1). Next, it is easy to
see that the set ΩpXq of T P LpXq satisfying (1.29.4) is operator norm open. Thus in order to verify
that (1.29.4) implies (1.29.3), it suffices to show that ΩpXq X WNHpXq " ∅. Let T P ΩpXq. Then
X " Y ' Z, where Y and Z are closed T -invariant subspaces, TY P LpY q has the spectral radius ă 1,
Z has dimension n P Z` and σ`TZ " tz1, . . . , znu satisfies 1 ď z1 ă . . . ă zn. By Lemma 7.1,
T R WNHpXq, which proves the implication (1.29.4)ùñ(1.29.3).
It remains to verify that (1.29.1) implies (1.29.4). In order to do this, it suffices to show that every
T P LpXqzΩpXq is in the operator norm closure of SNHpXq. Let T P LpXqzΩpXq. The relation T R ΩpXq
can happen for various reasons.
Case 1: There is a 2-dimensional T -invariant subspace E of X such for which σ`TE " ts, tu and
}Sn´ TE} Ñ 0. By Lemma 9.3, there is a sequence tRnunPN in LpXq such that RnE " Sn for each n P N
s " t ě 1. In this case, according to Lemma 9.4, there is a sequence tSnunPN in SNHpEq such that
and }Rn ´ T} Ñ 0. Since the restriction of each Rn to the invariant subspace E is strongly numerically
hypercyclic, Proposition 1.6 implies that Rn P SNHpXq for each n P N. Thus T is in the norm closure of
SNHpXq.
Note that the situation of T having a normal eigenvalue λ of multiplicity ě 2 such that λ ě 1 falls
into Case 1. The same happens if there are two distinct s, t P σppTq such that s " t ě 1. This means
that if Case 1 does not hold, we can pick λ P σpTq such that
λ ě 1, λ is not a normal eigenvalue of T and dim kerpT ´ λIq ď 1.
We keep the notation λ for such a spectral point for the rest of the proof.
32
Case 2:
subspace E of X such that }pT ´ λIqE} ă ε
SE " λIE and }T ´ S} ă 3ε
such that }S ´ R} ă ε
to the norm closure of SNHpXq.
pT ´ λIqpXq is non-closed in X. Let ε ą 0. By Lemma 9.2, we can pick a 2-dimensional
4 . According to Lemma 9.3, there is S P LpXq such that
4 . Now S falls under the jurisdiction of Case 1. Thus there is R P SNHpXq
4 " ε and ε ą 0 is arbitrary, T belongs
4 . Since }T ´ R} ď }T ´ S}`}S ´ R} ă 3ε
Thus it remains to consider the case when pT ´ λIqpXq is closed in X.
Case 3: kerpT ´ λIq is one-dimensional and pT ´ λIqpXq " X. In this case T ´ λI is a Fredholm
operator of index 1. By Theorem 1.23, T P SNHpXq.
Case 4: dimpX{pT ´ λIqpXqq ą dimpkerpT ´ λIqq. In this case T ´ λI is a semi-Fredholm operator
of negative index. Since X is reflexive, Theorem 1.23 implies that T P SNHpXq.
If kerpT ´ λIq " t0u and pT ´ λIqpXq " X, we have λ R σpTq, which is a contradiction. Finally,
if dimpkerpT ´ λIqq " dimpX{pT ´ λIqpXqq " 1, λ is a normal eigenvalue of T , which is again a
contradiction. This completes the proof of the implication (1.29.1)ùñ(1.29.4) and that of Theorem 1.29.
4 ` ε
10 Proof of Theorem 1.26
Lemma 10.1. Let T be a bounded normal operator on a Hilbert space H. Assume also that there exists
a sequence tznunPN in σpTq such that 1 ă z1 ă z2 ă . . . Then T P WNHpXq.
Proof. Let rk " 1
3 mintzk ´zk´1,zk`1 ´zku for k P N, where we assume z0 " 1. For each k P N pick
a continuous function ϕk : C Ñ r0, 1s such that ϕkpzkq " 1 and ϕk vanishes outside the disk zk ` rkD.
The functional calculus for normal operators allows us to consider the operators ϕkpTq. Since zk P σpTq
and ϕkpzkq ‰ 0, ϕkpTq ‰ 0 for every k P N. Thus we can pick xk P SpHqX ϕkpTqpHq. The latter inclusion
and the fact that ϕk vanishes outside the disk zk ` rkD implies that pzk ´ rkqn ď }T nxk} ď pzk ` rkqn
for every k P N and every n P Z`. These estimates together with the inequalities z1 ´ r1 ą 1 and
zk`1 ´ rk`1 ą zk ` rk for k P N imply that }T nx1} Ñ 8 and }T nxk`1}
}T nxk} Ñ 8 as n Ñ 8. By
Corollary 6.3, T P WNHpXq.
Lemma 10.2. Let T be a normal operator on Hilbert space H such that the set t´z : z P σpTqu is
well-ordered and for every r ą 1, the set σpTq X rT has at most one element. Then for each x P H and
f P H, either tfpT nxq : n P Z`u is bounded or fpT nxq Ñ 8. In particular, T R WNHpHq.
Proof. Since every well-ordered subset of R is countable, M " tz P σpTq : z ą 1u is (finite or) countable.
By the Spectral Theorem, T is unitarily equivalent to the orthogonal direct sum of a normal contraction
and a diagonal (with respect to an orthonormal basis) operator with only elements of M on the diagonal.
Taking this into account, we see that for each f P H and x P H,
fpT nxq " bn ` řtPM
attn, where a P ℓ1pMq and b P ℓ8pZ`q.
(10.1)
If a " 0, then tfpT nxq : n P Z`u " tbn : n P Z`u is bounded. Assume now that a ‰ 0. Since
t´z : z P σpTqu is well-ordered, there is c " maxtt : t P M, at ‰ 0u. Since σpTq X cT is a singleton,
there is the unique t0 P M satisfying t0 " c. Then
0 ` řtPM, tăc
0 " at0cnp1 ` op1qq as n Ñ 8.
fpT nxq " at0cn1 ` bn
Since c ą 1, fpT nxq Ñ 8.
Lemma 10.3. Let T be a normal operator on Hilbert space H such that there is a sequence tλnunPN in
σpTq such that 1 ă λ1 ă λ2 ă . . . and the numbers λj
λj
are pairwise distinct. Then T P NHpHq.
at0 tn
attn
at0 tn
33
Proof. By the Spectral Theorem we can assume that H " L2pµq, where pΩ, F , µq is a measure space and
T f " gf for f P H, where g P L8pµq. Then the spectrum of T is exactly the essential range of g. By
Lemma 3.11, we can pick a non-negative real valued function a P L1pµq for which
the set "żΩ
gka dµ : k P Z`* is dense in C.
Without loss of generality, we may assume that a is normalized to satisfy }a}1 " 1. Then x " ?a P SpHq
and xT nx, xy "şΩ gka dµ for every n P Z`. Thus NOpT, xq is dense in C and therefore T P NHpHq.
Now we are ready to prove Theorem 1.26. Let H be a Hilbert space, T P LpHq and k P N be such
that T k is normal. First, assume that there is a sequence tλnunPN in σpTq such that 1 ă λ1 ă λ2 ă . . ..
2 ă . . .. Since T k is normal, Lemma 10.1
Then tλk
yields the weak numeric hypercyclicity of T k. Then T is weakly numerically hypercyclic, which completes
the proof of (1.26.1).
nunPN is a sequence in σpT kq satisfying 1 ă λk
1 ă λk
numbers λj
λj
Next, assume that there is a sequence tλnunPN in σpTq such that 1 ă λ1 ă λ2 ă . . . and the
2 ă . . .
: n P N( is infinite. Then there is a strictly increasing sequence tnju of positive integers
and the set λk
are pairwise distinct. Then tλk
λk
n
nunPN is a sequence in σpT kq satisfying 1 ă λk
1 ă λk
are pairwise distinct. Since T k is normal, Lemma 10.3 implies that T k is
such that the numbers
n
λk
nj
λk
nj
numerically hypercyclic. Hence T is numerically hypercyclic, which completes the proof of (1.26.2).
Now assume that T k is self-adjoint. Then T 2k is self-adjoint and non-negative definite. If t´z : z P
σpTq, z ą 1u is not well-ordered, we can pick a sequence tλnunPN in σpTq such that 1 ă λ1 ă λ2 ă . . .
z ą 1u is well-ordered. Since
By (1.26.1), T P WNHpHq. Assume now that t´z
σpT 2kq Ă r0,8q, t´z : z P σpT 2kq, z ą 1u is well-ordered and σpT 2kqX rT has at most one element for
each r ą 1. By Lemma 10.2, for every x P H and f P H, OpT 2k, x, fq is either bounded or nowhere dense
in C. Now let x P H and f P H. Then OpT, x, fq "
OpT 2k, T jx, fq. Hence OpT, x, fq is the union
of finitely many sets each of which is either bounded or nowhere dense. Thus OpT, x, fq is not dense in
C (the set of accumulation points of OpT, x, fq is bounded). Then T R WNHpHq, which completes the
proof of (1.26.3).
: z P σpTq,
2k´1Ťj"0
Finally, assume that there exist r ą 1 and a T -invariant subspace K such that U " 1
unitary operator with infinite spectrum. If σppUq is infinite, then there are a sequence tznunPN of pairwise
distinct elements of T and an orthonormal sequence tenunPN in K Ď H such that T ken " U en " rznen
for every n P N. Thus T k satisfies (1.10.4) and therefore T k P SNHpHq Ă NHpHq by Theorem 1.10.
Hence T P NHpHq.
It remains to consider the case when σppUq is finite. Since σpUq is infinite, the
Spectral Theorem tells us that the restriction of U to an appropriate invariant subspace K0 is unitarily
equivalent to the multiplication by the argument operator M P LpL2pµqq, M fpzq " zfpzq, where µ is
a Borel probability purely non-atomic measure on T. By Proposition 1.2, in order to prove that T is
numerically hypercyclic, it suffices to verify that rM is numerically hypercyclic. By Lemma 3.13, there
is a P L1
Thus rM is numerically hypercyclic and therefore T is. This completes the proof of (1.26.4) and that of
Theorem 1.26.
`pµq such that }a}1 " 1 and the set O " trnxaµpnq : n P Nu is dense in C. Now it is easy to see
that x " ?a P SpL2pµqq and xM nx, xy " xaµpnq for each n P Z. Hence NOprM, xq " O is dense in C.
r T kK P LpKq is a
11 Proof of Propositions 1.18 and 1.19
Let T P LpC2q. Pick w P T such that at least one of the eigenvalues of wT is real. By Theorem 1.13,
wT R WNHpC2q. This proves Proposition 1.18. The rest of this section is devoted to the proof of
Proposition 1.19, which is more sophisticated. Denote
M ""pr, z, wq P p1,8q T T :rz 0
0 rw P NHpC2q*.
34
By Theorem 1.13, we can rewrite the definition of M in the following equivalent way.
Then the sets Mpz, wq featuring in Proposition 1.19 can be written as
M " pr, z, wq P p1,8q T T : trnpzn ` wnq : n P Z`u is dense in C(.
Mpz, wq ""r ą 1 :rz 0
0 rw P NHpC2q* " tr ą 1 : pr, z, wq P Mu
" r ą 1 : trnpzn ` wnq : n P Z`u is dense in C( for z, w P T.
If tVkukPN is a base of the topology of C, then we can write
M " čk,mPN
8ďn"m pr, z, wq P p1,8q T T : rnpzn ` wnq P Vk(.
(11.1)
This immediately implies that M is a Gδ subset of p1,8q T T and therefore each Mpz, wq is a
Gδ-subset of p1,8q.
Lemma 11.1. The set M is a dense Gδ-subset of p1,8q T T. Furthermore,
W " pz, wq P T2 : Mpz, wq is dense in p1,8q(
is a dense Gδ subset of T2.
Proof. By Proposition 1.16, for every r ą 1, the set rT X M is dense in rT. Hence M is dense in
p1,8q T T. Let trn : n P Nu be a dense subset of p1,8q. By Proposition 1.16, the set Wn " tpz, wq P
T2 : rn P Mpz, wqu is a dense Gδ subset of T2 for every n P N. Hence ŞnPN
Wn is a dense Gδ subset of T2,
Let tVkukPN is a base of the topology of C. It is easy to verify that
being also a subset of W . Hence W is dense in T2. It remains to show that W is a Gδ set.
Aa,b,m,k "
8ďn"m pz, wq P T2 : pzn ` wnqra, bs X Vk ‰ ∅(
is an open subset of T2 whenever m, k P N and 1 ă a ă b ă 8. It remains to notice that
W " čm,kPN ča,bPQ, 1ăaăb
Aa,b,m,k
to conclude that W is a Gδ subset of T2.
a rather typical occurrence.
By Lemma 1.13, Mpz, wq " ∅ if z and w are not independent. Actually, the equality Mpz, wq " ∅ is
Lemma 11.2. The set pz, wq P T2 : Mpz, wq ‰ ∅( has zero Lebesgue measure in T2.
n2( ă 8.
Proof. Let µ be the normalized Lebesgue measure on T2. By (3.4),
Hence for almost all pz, wq P T2, zn` wn ą n´2 for all sufficiently large n and therefore rnzn` wn Ñ 8
for every r ą 1. Hence Mpz, wq " ∅ for almost all z, w P T2.
µ pz, wqPT2 : zn`wn ď 1
8řn"1
On the other hand, it turns out that if Mpz, wq is non-empty, then it is infinite.
Lemma 11.3. Let z, w P T, r P Mpz, wq and k be an odd positive integer. Then r1{k P Mpz, wq.
35
Proof. In this proof, we shall use the symbol or for any sequence tαnu of complex numbers satisfying
lim supαn1{n ď 1
r . Since r P Mpz, wq, the set tyn : n P Z`u is dense in C, where yn " rnpzn ` wnq. Let
A " tn P N : n´1 ď yn ď nu. Then tyn : n P Au is dense in C. The inequality zn ` wn ď nr´n for
n P A yields
(11.2)
wn " ´zn ` or and y2
n " ´yn2z2n ` or as n Ñ 8, n P A.
Since k is odd and rnpzkn ` wknq " rnpzn ` wnq
and (11.2) imply rnpzkn ` wknq " ynzpk´1qn ` or " yk
k´1řj"0p´1qj zjnwpk´1´jqn, the equality yn " rnpzn ` wnq
ynk´1 ` or as n Ñ 8, n P A. Hence,
n
rnpzkn ` wknq " Φpynq ` op1q as n Ñ 8, n P A, where Φ : Czt0u Ñ C, Φpxq " xk
xk´1 .
Since Φ is a homeomorphism of Czt0u onto itself and tyn : n P Au is dense in C, tΦpynq : n P Au is
dense in C. Then by the above display, trnpzkn ` wknq : n P Au is dense in C. Hence the bigger set
trn{kpzn ` wnq : n P Z`u is dense in C and therefore r1{k P Mpz, wq.
Finally, we shall verify that Mpz, wq can not coincide with p1,8q by showing that each Mpz, wq
has zero Lebesgue measure. We need some preparation. Consider the function δ : R Ñ r0, 1s, δpxq "
distpx, 2Z ´ 1q. That is, δpxq is the distance from x to the nearest odd integer.
Lemma 11.4. Let θ P RzQ, 0 ă a ă b ă 1 and A " tn P N : an ď δpnθq ď bnu. Then řnPA
Proof. For every q P N, let ppqq be the unique odd integer p such thatθ´ p
q ă 1
q for each q P N. Consider the set
q and g.c.d.pq, ppqqq " 1).
B "!q P N : αpqq ď bq
řqPB
q . Denote αpqq "θ´ ppqqq .
Then 0 ă αpqq ă 1
First, we shall show that
n ă 8.
q ă 8.
(11.3)
1
1
qj
rj
and γj ď brj
. In particular, ppqjqqj ´ pprjqrj " γj ´ βj. It follows that ppqjqrj´pprjqqj
Assume that (11.3) fails. Then there are two strictly increasing sequences tqjujPN and trjujPN in B such
that qj ă rj for each j P N and qj
rj Ñ 1. Indeed, otherwise members of B written in the increasing order
grow exponentially and (11.3) follows. By definition of B, we have θ " ppqjqqj ` βj " pprjqrj ` γj, where
qj ` brj
βj ď bqj
.
Hence ppqjqrj ´ pprjqqj ď rjbqj ` qjbrj ă 2rj bqj since qj ă rj. Since qj
rj Ñ 1 and b ă 1, we have rj bqj Ñ 0
and therefore ppqjqrj ´ pprjqqj Ñ 0. Since the latter is a sequence of integers, ppqjqrj ´ pprjqqj " 0 for
all sufficiently large j. In particular, there is j P N for which ppqjqrj " pprjqqj. Since pprjq and rj are
relatively prime, rj must divide qj, which is not possible since rj ą qj. This contradiction proves (11.3).
Now n P A if and only if there is an odd integer k such that an ď δpnθq " nθ ´ k ď bn. Let
r " g.c.d. pn, kq. Then we can write n " rq and k " rp, where q P N, p is an odd integer and
g.c.d. pq, pq " 1. Using this notation we can rewrite the last inequality as arq
q , which
happens if and only if q P B, p " ppqq and arq
rq . Thus, we have obtained the following
q ď brq
ď bqj
rq ď bq
qj rj
rj
description of the set A:
rq ď αpqq ď brq
A "!qr : q P B, r P 2N ´ 1, arq
rq ďθ ´ p
rq).
rq ď αpqq ď brq
m for every m P N. Then
Fix d ą 0 small enough in such a way that dm ď am
A Ď A1 " tqr : q P B, r P N, drq ď αpqq ď brqu " tqr : q P B, r P N, βpqq ď r ď cβpqqu,
36
where βpqq " log αpqq
corresponding integral, one easily sees that
q log d and c " log d
log b . Obviously, βpqq ą 0 and c ą 1. Estimating a monotonic sum by the
řmPN
m ď 1 ` log c for every β ą 0 and c ą 1.
βďmďcβ
1
Using the above two displays and (11.3), we obtain
1
řrPN
βpqqďrďcβpqq
1
1
1
q
1
q ă 8.
1
řnPA
1
n ă 8.
n " řqPB
r ď p1 ` log cq řqPB
n ă 8 and therefore řnPA
n ď řnPA1
Corollary 11.5. Let z P T be of infinite order and A " tn P N : can ď 1 ` zn ď dbnu, where
0 ă a ă b ă 1 and c, d ą 0. Then řnPA
Proof. Since z has infinite order in T, z " eiπθ with θ P RzQ. Pick real numbers s, t such that 0 ă s ă
a ă b ă t ă 1 and let A1 " tn P N : sn ď δpnθq ď tnu. It is an elementary exercise to see that AzA1 is
finite. By Lemma 11.4, řnPA1
Lemma 11.6. For every z, w P T, Mpz, wq has Lebesgue measure 0.
Proof. If w´1z has finite order in T, Theorem 1.13 implies that Mpz, wq " ∅. Thus we can assume that
s " w´1z is of infinite order. If r P Mpz, wq, then the number rnzn ` wn " rn1 ` sn belongs to r1, 2s
for infinitely many n P N. Let 0 ă a ă b ă 1. The last observation yields
Mpz, wq X rb´1, a´1s Ď Na,b "
Let A " tn P N : an ď 1 ` sn ď 2 bnu. By Corollary 11.5, řnPA
n ă 8. Note that Kn,a,b " ∅ if n R A.
On the other hand, if n P A, then Kn,a,b is contained in the interval r1 ` sn´1{n, 21{n1 ` sn´1{ns,
whose length is p21{n ´ 1q1 ` sn´1{n ď 21{n´1
an . Thus, the Lebesgue measure µ of the sets Kn,a,b is
Kn,a,b, where Kn,a,b " tr P rb´1, a´1s : 1 ď rn1 ` sn ď 2u.
a ď 1
8ďn"m
8čm"1
1
n ă 8.
1
estimated as
µpKn,a,bq " 0 if n R A and µpKn,a,bq ď
1
an
if n R A.
According to the above two displays, for each m P N,
µpMpz, wq X rb´1, a´1sq ď
µpKn,a,bq ď
8ÿn"m
1
a ÿnPA
něm
1
n Ñ 0 as m Ñ 8.
Hence µpMpz, wq X rb´1, a´1sq " 0 whenever 0 ă a ă b ă 1. Thus µpMpz, wqq " 0.
Proposition 1.19 follows from Lemmas 11.1, 11.3 and 11.6.
12 Further comments and open questions
The next proposition provides a peculiar collection of diagonal numerically hypercyclic operators on C3.
Proposition 12.1. There is a diagonal T P SN HpC3q such that the restriction of T to each invariant
2-dimensional subspace is not weakly numerically hypercyclic.
Proof. Let 1 ă r ă R and k, m be distinct integers. By Lemma 3.5, the set
W " pz, wq P T2 : tRnpzkn ` zmnq ` rnwn : n P Z`u is dense in C(
is a dense Gδ subset of T2. Pick pz, wq in W and consider the diagonal operator T on C3 with the numbers
Rzk, Rzm and rw on the diagonal. Clearly, (1.10.1) is satisfied for T . By Theorem 1.10, T P SNHpC3q.
On the other hand, the possible spectra of the restrictions of T to 2-dimensional invariant subspaces
are tRzk, Rzmu, tRzk, rwu and tRzm, rwu.
In any case, such a restriction is not weakly numerically
hypercyclic according to Theorem 1.13.
37
As we have already observed, a self-adjoint operator is never numerically hypercyclic. Funnily enough
a scalar multiple of a self-adjoint operator can turn out to be numerically hypercyclic.
Proposition 12.2. Let T be a bounded self-adjoint operator on a Hilbert space H such that there exists
a sequence tλnunPN in σpTq satisfying 1 ă λ1 ă ´λ2 ă λ3 ă ´λ4 ă . . . Then for any w P T of infinite
order, wT P NHpHq.
Proof. Let sk " 1
3 mintλk ´ λk´1,λk`1 ´ λku for k P N, where we assume λ0 " 1. Using the
spectral theorem, we can find an orthonormal sequence txnunPN in H such that xT kxn, xny is always
between pλn ´ snqk and pλn ` snqk and xT mxk, T nxjy " 0 provided k ‰ j. Let E be the linear span of
txn : n P Nu and K be the closure of E in H. Fix y P E and t P Czt0u. Let k " maxtj P N : xy, xjy ‰
0u. Since λj ` sj ă λj`1 ´ sj`1 for each j, xT ny, yy Ñ 8 and xT ny, yy " opxT nxk`1, xk`1yq as
n Ñ 8. Furthermore, since λj alternate signs, xT ny, yy and xT nxk`1, xk`1y have opposite signs for every
are positive for sufficiently
sufficiently large odd positive integer n. Hence the numbers
large odd n and form a convergent to 0 sequence. That is, for each sufficiently large odd n, we can find
εn ą 0 such that ε2
. Furthermore, εn Ñ 0. The last equality can be easily rewritten as
xT npy ` εnxk`1q, y ` εnxk`1y " t. Since w has infinite order, we can find a strictly increasing sequence
tnjujPN of odd positive integers such that wnj Ñ t
. Then yj Ñ y and tj " xpwTqnj yj, yjy Ñ t, where
t
yj " y ` εnj xk`1. Thus py, tq is in the closure of the set
n " t´xT ny,yy
xT nxk`1,xk`1y
xT nxk`1,xk`1y
t´xT ny,yy
Λ " tpy,xpwTqny, yyq : y P K, n P Nu.
Since E is dense in K and y P E and t P Czt0u are arbitrary, Λ is dense in K C. By Theorem U, there
is y P K such that txpwTqny, yy : n P Nu is dense in C. Hence wT P NHpHq.
The normal operator T in the following example is strongly numerically hypercyclic, while even its
weak numeric hypercyclicity does not follow from Theorem 1.26. The example illustrates the limitations
of Theorem 1.26.
Example 12.3. Let trnunPN be a strictly decreasing sequence in p1,8q and A " t1, 2, 3, 4, 5u. Then
there exists a sequence taj,nupj,nqPAN of finite order elements of T such that the diagonal operator T on
ℓ2pA Nq with the numbers rnaj,n on the diagonal is strongly numerically hypercyclic.
Proof. Let ε, d ą 0 be the numbers furnished by Lemma 3.9. For each n P N, let Un be the only cyclic
subgroup of T of order n: Un " tz P T : zn " 1u. Clearly, Un is a π
n -net in T. Hence there is n0 P N
such that for each n ě n0, we can find a1, . . . , a5 P Un for which pa1, . . . , a5, a2
5q P Pε, where Pε is
the set defined in (3.5). Fix a sequence tznunPN such that tzn : n P Nu is dense in C and zn ă dn for
every n P N. Using the induction with respect to n, we shall construct the sequences taj,nupj,nqPAN in T,
tcj,nupj,nqPAN of positive numbers, tknunPN of positive integers and tpnunPN of prime numbers such that
(A1) aj,n P Upn for 1 ď j ď 5;
5řj"1
(A2)
(A3) pq divides kn ´ 2kq for 1 ď q ă n;
(A4) pn ą pn´1 and rkn´kn´1
5řj"1
ą n2n if n ě 2.
j,q " 0 and
cj,n " n
j,n " zn
1, . . . , a2
5řj"1
cj,qa2kn
cj,nakn
rkn
n
;
rkn
n
,
n
can choose b1,1, . . . , b5,1 P Up1 such that pb1,1, . . . , b5,1, b2
such that p1 does not divide k1 and r´k1
The basis of the induction. Pick an arbitrary prime number p1 such that p1 ě n0. Then we
5,1q P Pε. Since r1 ą 1, there is k1 P N
1 ă 1. Since z1 P d D, Lemma 3.9 provides positive numbers
5řj"1
5řj"1
tj,1 " 1. Since bj,1 P Up1, p1 does not divide
t1,1, . . . , t5,1 such that
j,1 " bj,1 for 1 ď j ď 5. Set cj,1 " tj,1
k1 and p1 is prime, there are a1,1, . . . , a5,1 P Up1 such that ak1
for
1 ď j ď 5. Now it is straightforward to see that (A1) and (A2) with n " 1 are satisfied. Conditions (A3)
and (A4) with n " 1 are satisfied in a trivial manner. Thus we have our basis of induction.
tj,1bj,1 " z1,
j,1 " 0 and
1,1, . . . , b2
5řj"1
tj,1b2
k1
1
r
38
The induction step. Let n ě 2 and assume that aj,q, cj,q, pq and kq with q ă n satisfying (A1 -- A4)
are already constructed. First, we choose an arbitrary prime number pn ą pn´1. Since pn ą p1 ě n0,
we can pick b1,n, . . . , b5,n P Upn such that pb1,n, . . . , b5,n, b2
5,nq P Pε. Since rn ą 1 and p1, . . . , pn
are pairwise distinct primes, the Chinese Remainder Theorem allows us to choose kn P N such that
rkn´kn´1
n P d D, Lemma 3.9
n
5řj"1
tj,n " 1. Since
provides positive numbers t1,n, . . . , t5,n such that
bj,n P Upn, pn does not divide kn and pn is prime, there are a1,n, . . . , a5,n P Upn such that akn
j,n " bj,n for
1 ď j ď 5. Set cj,n " ntj,n
for 1 ď j ď 5. Now it is straightforward to see that (A1 -- A4) are satisfied.
This completes the construction of the sequences taj,nupj,nqPAN, tcj,nupj,nqPAN, tknunPN and tpnunPN
satisfying (A1 -- A4).
ą n2n, pn does not divide kn and pq divides kn´2kq for 1 ď q ď n´1. Since zn
j,n " 0 and
tj,nbj,n " zn
n ,
1,n, . . . , b2
5řj"1
5řj"1
tj,nb2
rkn
n
For pj, nq P A N, set λj,n " rnaj,n. Since
each m P N, we can write
5řj"1
cj,n " n
rkn
n
and rkn
n ě n2n, we have c P ℓ1
`pA Nq. For
ÿpj,nqPAN
cj,nλkm
j,n " ÿnPN
αn,m, where αn,m " rkm
n
cj,nakm
j,n.
5ÿj"1
First, note that by (A2),
5ÿj"1
Since for n ă m, pn divides km ´ 2kn and apn
αm,m " rkm
m
cj,makm
j,m " rkm
m
zm
m " zm.
rkm
j,n " 1, (A2) yields
αn,m " rkm
n
5ÿj"1
cj,nakm
j,n " rkm
n
5ÿj"1
cj,na2kn
j,n " 0 if n ă m.
Finally, if n ą m, using (A4) and (A2), we obtain
(12.1)
(12.2)
(12.3)
(12.4)
Combining (12.1 -- 12.4), we get
αn,m " rkm
cj,nakm
n
5ÿj"1
zm ´ ÿpj,nqPAN
n
j,n ď rkm
j,n "
cj,nλkm
5ÿj"1
cj,n "
nrkm
n
n ď
rkn
n
rkn´kn´1
n
1
2n if n ą m.
ď
8ÿn"m`1
αn,m ď
8ÿn"m`1
1
2n Ñ 0 as m Ñ 8.
Since tzm : m P Nu is dense in C, the above display implies that! řpj,nqPAN
j,n : k P N) is dense in C.
Thus the diagonal operator T P Lpℓ2pA Nqq with the numbers λj,n on the diagonal satisfies (1.23.1).
By Theorem 1.23, T is strongly numerically hypercyclic.
cj,nλk
Let V be the Volterra operator on L2r0, 1s. Exactly as in Example 1.21, we can show that re´V P
WNHpL2r0, 1sq for every r ą 1. Unexpectedly enough it turns out that these operators are not numeri-
cally hypercyclic.
Example 12.4. Let r ą 1 and T " re´V , where V P LpL2r0, 1sq is the Volterra operator. Then
T R NHpL2r0, 1sq.
Proof. The key element of the proof is the following claim:
}I ´ e´cV } ď 1 for every c P R`.
(12.5)
39
We shall derive this estimate from the following observation:
for every f P L2r0, 1s, the function Ff : R` Ñ R`, Ffpcq " }pI ´ e´cV qf} is increasing.
(12.6)
First, we shall verify (12.6). Obviously, it is enough to show that Gf " F 2
f is increasing. Differentiating
by c the expression Gfpcq " xpI ´ e´cV qf,pI ´ e´cV qfy, we get G1fpcq " xV gc, gcy ` xgc, V gcy, where
gc " pI ´ e´cV qf . Hence G1fpcq " xP gc, gcy, where P " V ` V ‹ and V ‹ is the Hilbert space adjoint of V .
It is easy to see that P is the orthogonal projection onto the one-dimensional space of constant functions.
Then the self-adjoint operator P is non-negative definite and G1fpcq " xP gc, gcy ě 0 for every c P R`.
Hence Gf is increasing and (12.6) follows.
Now we shall prove (12.5). Assume the contrary. Then there is a ą 0 such that }I ´ e´aV } ą 1.
Then there is f P SpL2r0, 1sq such that }pI ´ e´aV qf} " Ffpaq ą 1. By (12.6),
Ffpcq ą 1. On the
other hand, in [13] it is shown that }pI ´ V qng} Ñ 0 for each g P L2r0, 1s. According to the main result
of [5] the operators I ´ V and e´V are similar. Hence }e´nV g} Ñ 0 for each g P L2r0, 1s. In particular,
}e´nV f} Ñ 0. Hence Ffpnq Ñ 1, which contradicts the inequality lim
Ffpcq ą 1. This contradiction
cÑ8
completes the proof of (12.5).
lim
cÑ8
Now let r ą 1 and T " re´V . Then for each f P SpL2r0, 1sq, using (12.5), we have
2 xpe´nV `e´nV ‹
RexT nf, fy " rn
qqf, fyq ě rnp1´}I´e´nV }q ě 0.
Hence NOpT, fq is contained in the right half-plane and therefore can not be dense in C. Thus T R
NHpL2r0, 1sq.
2 p2´xppI´e´nV q`pI´e´nV ‹
qf, fy " rn
At the expense of a number of technical details entering the proof, one can similarly show that
rpI ´ cV q is weakly numerically hypercyclic and not numerically hypercyclic whenever r ą 1 and c ą 0.
We shall raise a number of natural questions.
Question 12.5. Let T P WNHpCnq. Is it true that at least one of the conditions p1.9.1 -- 1.9.4q is satisfied?
Question 12.6. Assume that X is reflexive, T P LpXq is not power bounded and σppTq " ∅. Is it true
that T is weakly numerically hypercyclic?
Question 12.7. Let T P WNHpXq. Is it true that T n P WNHpXq for every n P N?
Question 12.8. Let T P WNHpXq. Is it true that rT P WNHpXq for every r ě 1?
Question 12.9. Let T be a unitary operator such that σpTq is infinite and σppTq " ∅. Is rT strongly
numerically hypercyclic for each r ą 1?
Question 12.10. Let T P LpXq and λ1, λ2 P C are such that kerpT ´ λjIq2 ‰ kerpT ´ λjIq for
j P t1, 2u, λ1 " λ2 ě 1 and λ1
Is it true that T must be strongly
λ1
numerically hypercyclic?
are independent in T.
, λ2
λ2
Question 12.11. Characterize normal numerically hypercyclic operators in terms of their spectrum.
Question 12.12. Characterize numerically hypercyclic operators on L2r0, 1s commuting with the Volterra
operator V . In particular, is I ` V numerically hypercyclic?
The last question is in the spirit of the Invariant Subspace Problem and probably is tough.
Question 12.13. Is there T P Lpℓ2pNqq such that NOpT, xq is dense in C for each x P Spℓ2pNqq?
40
References
[1] C. Badea and V. Muller, On weak orbits of operators, Topology and its Applications 156 (2009), 1381-1385
[2] F. Bayart and E. Matheron, Dynamics of linear operators, Cambridge University Press, 2009
[3] S. Bermudo, A. Montes-Rodr´ıguez and S. Shkarin, Orbits of operators commuting with the Volterra operator,
J. Math. Pures Appl. 89 (2008), 145 -- 173
[4] N. Feldman, N -weakly hypercyclic and N -weakly supercyclic operators, J. Functional Analysis [to appear]
[5] R. Frankfurt and J. Rovnyak, Finite convolution operators, J. Math. Anal. Appl. 49 (1975), 347-374.
[6] K. Grosse-Erdmann, Universal families and hypercyclic operators, Bull. Amer. Math. Soc. 36 (1999), 345 -- 381
[7] K. Grosse-Erdmann and A. Peris, Linear Chaos, Springer, 2011
[8] D. Herrero, Limits of hypercyclic and supercyclic operators, J. Funct. Anal. 99 (1991), 179 -- 190
[9] D. Herrero and Z. Wang, Compact perturbations of hypercyclic and supercyclic operators, Indiana Univ. Math.
J. 39 (1990), 819 -- 829
[10] S. Kim, A. Peris and H. Song, Numerically hypercyclic operators, Integral Equations and Operator Theory,
72 (2012), 393 -- 402
[11] S. Kim, A. Peris and H. Song, Numerically hypercyclic polynomials, Archiv der Mathematik, 99 (2012),
443 -- 452
[12] F. Le´on-Saavedra and V. Muller, Rotations of hypercyclic and supercyclic operators, Integral Equations and
Operator Theory 50 (2004), 385 -- 391
[13] A. Montes-Rodr´ıguez, J. S´anchez- ´Alvarez and J. Zem´anek, Uniform Abel-Kreiss boundedness and the extremal
behaviour of the Volterra operator, Proc. London Math. Soc. 91 (2005), 761-788
[14] G. Prajitura, Limits of weakly hypercyclic and supercyclic operators, Glasg. Math. J. 47 (2005), 255-260
[15] H. Salas, Hypercyclic weighted shifts, Trans. Amer. Math. Soc., 347 (1995), 993 -- 1004
Stanislav Shkarin
Queens's University Belfast
Department of Pure Mathematics
University road, Belfast, BT7 1NN, UK
E-mail address:
[email protected]
41
|
1302.6295 | 2 | 1302 | 2013-11-26T23:28:14 | Spectral analysis of the truncated Hilbert transform with overlap | [
"math.FA",
"math.CA",
"math.SP"
] | We study a restriction of the Hilbert transform as an operator $H_T$ from $L^2(a_2,a_4)$ to $L^2(a_1,a_3)$ for real numbers $a_1 < a_2 < a_3 < a_4$. The operator $H_T$ arises in tomographic reconstruction from limited data, more precisely in the method of differentiated back-projection (DBP). There, the reconstruction requires recovering a family of one-dimensional functions $f$ supported on compact intervals $[a_2,a_4]$ from its Hilbert transform measured on intervals $[a_1,a_3]$ that might only overlap, but not cover $[a_2,a_4]$. We show that the inversion of $H_T$ is ill-posed, which is why we investigate the spectral properties of $H_T$.
We relate the operator $H_T$ to a self-adjoint two-interval Sturm-Liouville problem, for which we prove that the spectrum is discrete. The Sturm-Liouville operator is found to commute with $H_T$, which then implies that the spectrum of $H_T^* H_T$ is discrete. Furthermore, we express the singular value decomposition of $H_T$ in terms of the solutions to the Sturm-Liouville problem. The singular values of $H_T$ accumulate at both $0$ and $1$, implying that $H_T$ is not a compact operator. We conclude by illustrating the properties obtained for $H_T$ numerically. | math.FA | math |
SPECTRAL ANALYSIS OF THE TRUNCATED HILBERT
TRANSFORM WITH OVERLAP
REEMA AL-AIFARI∗ AND ALEXANDER KATSEVICH†
Abstract. We study a restriction of the Hilbert transform as an operator HT from L2(a2, a4) to
L2(a1, a3) for real numbers a1 < a2 < a3 < a4. The operator HT arises in tomographic reconstruc-
tion from limited data, more precisely in the method of differentiated back-projection (DBP). There,
the reconstruction requires recovering a family of one-dimensional functions f supported on compact
intervals [a2, a4] from its Hilbert transform measured on intervals [a1, a3] that might only overlap,
but not cover [a2, a4]. We show that the inversion of HT is ill-posed, which is why we investigate
the spectral properties of HT .
We relate the operator HT to a self-adjoint two-interval Sturm-Liouville problem, for which we
prove that the spectrum is discrete. The Sturm-Liouville operator is found to commute with HT ,
which then implies that the spectrum of H ∗
T HT is discrete. Furthermore, we express the singular
value decomposition of HT in terms of the solutions to the Sturm-Liouville problem. The singular
values of HT accumulate at both 0 and 1, implying that HT is not a compact operator. We conclude
by illustrating the properties obtained for HT numerically.
1. Introduction. In tomographic imaging, which is widely used for medical
applications, a 2D or 3D object is illuminated by a penetrating beam (usually X-
rays) from multiple directions, and the projections of the object are recorded by a
detector. Then one seeks to reconstruct the full 2D or 3D structure from this collection
of projections. When the beams are sufficiently wide to fully embrace the object and
when the beams from a sufficiently dense set of directions around the object can be
used, this problem and its solution are well understood [16]. When the data are more
limited, e.g. when only a reduced range of directions can be used or only a part of
the object can be illuminated, the image reconstruction problem becomes much more
challenging.
This method goes back to a result by Gelfand and Graev [6].
Reconstruction from limited data requires the identification of specific subsets of
line integrals that allow for an exact and stable reconstruction. One class of such
configurations that have already been identified, relies on the reduction of the 2D
and 3D reconstruction problem to a family of 1D problems. The Radon transform
can be related to the 1D Hilbert transform along certain lines by differentiation and
back-projection of the Radon transform data (differentiated back-projection or DBP).
Inversion of the Hilbert transform along a family of lines covering a sub-region of the
object (region of interest or ROI) then allows for the reconstruction within the ROI.
Its application
to tomography was formulated by Finch [4] and was later made explicit for 2D in
[17, 23, 28] and for 3D in [18, 24, 27, 29]. To reconstruct from data obtained by
the DBP method, it is necessary to solve a family of 1D problems which consist of
inverting the Hilbert transform data on a finite segment of the line. If the Hilbert
transform Hf of a 1D function f was given on all of R, then the inversion would be
trivial, since H −1 = −H. In case f is compactly supported, it can be reconstructed
even if Hf is not known on all of R. Due to an explicit reconstruction formula by
Tricomi [22], f can be found from measuring Hf only on an interval that covers the
support of f . However, a limited field of view might result in configurations in which
the Hilbert transform is known only on a segment that does not completely cover
the object support. One example of such a configuration is known as the interior
∗Department of Mathematics, Vrije Universiteit Brussel, Brussels B-1050, Belgium
†Department of Mathematics, University of Central Florida, FL 32816, USA
1
problem [1, 10, 12, 25]. Given real numbers a1 < a2 < a3 < a4, the interior problem
corresponds to the case in which the Hilbert transform of a function supported on
[a1, a4] is measured on the smaller interval [a2, a3].
In this paper, we study a different configuration, namely supp f = [a2, a4] and
the Hilbert transform is measured on [a1, a3]. We will refer to this configuration
as the truncated problem with overlap: the operator HT we consider is given by
P[a1,a3]HP[a2,a4], where H is the usual Hilbert transform acting on L2(R), and PΩ
stands for the projection operator (PΩf )(x) = f (x) if x ∈ Ω, (PΩf )(x) = 0 otherwise.
For finite intervals Ω1, Ω2 on R, the interior problem corresponds to PΩ1 HPΩ2 for
Ω1 ⊂ Ω2. The truncated Hilbert transform with a gap occurs when the intervals
Ω1 and Ω2 are separated by a gap, as in [8]. Figure 1.1 shows the different setups.
Examples of configurations in which the truncated Hilbert transform with overlap and
the interior problem occur are given in Figures 1.2 and 1.3. The truncated problem
with overlap arises for example in the "missing arm" problem. This is the case where
the field of view is large enough to measure the torso but not the arms.
(a) Finite Hilbert transform.
(b) Interior problem.
(c) Truncated Hilbert transform with overlap.
(d) Truncated Hilbert transform with a gap.
Fig. 1.1: Different setups for PΩ1 HPΩ2. The upper interval shows the support Ω2
of the function f to be reconstructed. The lower interval is the interval Ω1 where
measurements of the Hilbert transform Hf are taken. This paper investigates case
(c).
2
(a)
(b)
Fig. 1.2: Two examples of the truncated problem with overlap, Fig. 1.2(a) shows
the missing arm problem. In both cases, the field of view (FOV) does not cover the
object support. On the line intersecting the object, measurements can only be taken
within the FOV, i.e. from a1 to a3. The Hilbert transform is not measured on [a3, a4].
Consequently, a reconstruction can only be aimed at in the grey-shaded intersection
of the FOV with the object support, called the region of interest (ROI).
Fig. 1.3: The interior problem. Here, the FOV also does not cover the object support.
The line intersecting the object is such that the Hilbert transform is only measured in
a subinterval [a2, a3] of the intersection [a1, a4] of the line with the object support. The
ROI is the grey-shaded intersection of the FOV with the object support. In this case
stable reconstruction of the shaded ROI is impossible unless additional information
is available.
3
Fix any four real numbers a1 < a2 < a3 < a4. We define the truncated Hilbert
transform with overlap as the operator
Definition 1.1.
f (y)
y − x
where p.v. stands for the principal value. In short,
p.v. Z a4
(HT f )(x) :=
1
π
a2
dy,
x ∈ (a1, a3)
(1.1)
HT := P[a1,a3]HP[a2,a4],
where H is the ordinary Hilbert transform on L2(R).
As we will prove in what follows, the inversion of HT is an ill-posed problem
in the sense of Hadamard [3].
In order to find suitable regularization methods for
its inversion, it is crucial to study the nature of the underlying ill-posedness, and
therefore the spectrum σ(H ∗
T HT ). An important question that arises here is whether
the spectrum is purely discrete. This question has been answered for similar operators
before, but with two very different answers. In [11], it was shown that the finite Hilbert
transform defined as HF = P[a,b]HP[a,b] has a continuous spectrum σ(HF ) = [−i, i].
On the other hand, in [9], we find the result that for the interior problem HI =
P[a2,a3]HP[a1,a4], the spectrum σ(H ∗
In
addition, we obtain that 0 and 1 are accumulation points of the spectrum. Further-
more, we find that the singular value decomposition (SVD) of the operator HT can
be related to the solutions of a Sturm-Liouville (S-L) problem. For the actual recon-
struction, one would aim at finding f in (1.1) only within a region of interest (ROI),
i.e. on [a2, a3]. A stability estimate as well as a uniqueness result for this setup were
obtained by Defrise et al in [2]. A possible method for ROI reconstruction is the trun-
cated SVD. Thus, it is of interest to study the SVD of HT also for the development
of reconstruction algorithms.
The main result of this paper is that H ∗
T HT has only discrete spectrum.
I HI ) is purely discrete.
In [8] and [9], singular value decompositions are obtained for the truncated Hilbert
transform with a gap P[a3,a4]HP[a1,a2] and for HI . This is done by relating the Hilbert
transforms to differential operators that have discrete spectra. We follow this proce-
dure, but obtain a differential operator that is different in nature. In [8] and [9] the
discreteness of the spectra follows from standard results of singular S-L theory (see
e.g. [26]). In the case of truncated Hilbert transform (1.1) we have to investigate the
discreteness of the spectrum of the related differential operator explicitly.
The idea is to find a differential operator for which the eigenfunctions are the
singular functions of HT on (a2, a4). We define the differential operator similarly
to the one in [8], [9], but then the question is which boundary conditions to choose
in order to relate the differential operator to HT . To answer this question we first
develop an intuition about the singular functions of HT .
Let {σn; fn, gn} denote the singular system of HT that we want to find. The
problem can be formulated as finding a complete orthonormal system {fn}n∈N in
L2(a2, a4) and an orthonormal system {gn}n∈N in L2(a1, a3) such that there exist
real numbers σn for which
HT fn = σngn,
H ∗
T gn = σnfn.
4
At the moment, the gn's only have to be complete in Ran(HT ), but as we will see in
Section 5, Ran(HT ) is dense in L2(a1, a3).
As will be shown in Section 4, the functions fn and gn
(a) can only be bounded or of logarithmic singularity at the points ai,
1 , a−
(b) do not vanish at the edges of their supports (a+
We will now make use of the following results from [5], Sections 8.2 and 8.5:
4 for fn, and a+
2 , a−
3 for gn).
Lemma 1.2 (Local properties of the Hilbert transform). Let f be a function with
support [b, d] ⊂ R. And let c be in the interior of [b, d].
1. If f is Holder continuous (for some Holder index α) on [b, d], then close to b
the Hilbert transform of f is given by
(Hf )(x) = −
1
π
f (b+) lnx − b + H0(x)
(1.2)
where H0 is bounded and continuous in a neighborhood of b.
2. If in a neighborhood of c, the function f is of the form f (x) = f (x) ln x − c
for Holder continuous f , then close to the point c its Hilbert transform is of
the form
(Hf )(x) = H0(x),
where H0 is bounded with a possible finite jump discontinuity at c.
3. If f is of the form f (x) = f (x) ln x − b on [b, c], where f is Holder contin-
uous, then its Hilbert transform at b has a singularity of the order ln2 x − b
if f (b) 6= 0.
Suppose fn has a logarithmic singularity at a+
2 . Since HT integrates over [a2, a4],
the function HT fn would have a singularity at a2 of order ln2 x − a2. Hence, this
would violate the property of gn at a2. Therefore, fn has to be bounded at a+
2 . If
fn does not vanish at a+
2 , this leads to logarithmic singularities of HT fn and gn at
a2. Using the same argument we conclude that gn is bounded at a−
3 and fn has a
logarithmic singularity at a3.
On the other hand, since gn is bounded at a−
3 , HT fn is also bounded there. This
requires that close to a3, fn = fn,1 + fn,2 lnx− a3 for functions fn,i continuous at a3.
A similar argument holds for gn at a2. Close to that point, gn = gn,1 + gn,2 lnx− a2
for functions gn,i continuous at a2.
Clearly, HT fn is bounded at a+
4 . Therefore, fn has to be
bounded at a−
1 and HT gn is bounded at a−
4 and gn must be bounded at a+
1 .
Thus, if we want to show the commutation of HT with a differential operator that
acts on fn(x), x ∈ (a2, a4), we need to impose boundary conditions at a+
4 that
require boundedness and some transmission conditions at a3 that make the bounded
term and the term in front of the logarithm in fn continuous at a3.
2 and a−
Having found these properties of the singular functions of HT (in case the SVD
for HT exists), in Section 2 we introduce a differential operator and find a self-adjoint
extension for this operator. We then show in Section 3 that this self-adjoint differential
operator LS has a discrete spectrum. In Section 4 we establish that LS commutes
with the operator HT . This allows us to find the SVD of HT . In Section 5 we then
5
(a) Sketch of fn's.
(b) Sketch of gn's.
Fig. 1.4: Intuition about the singular functions of HT .
study the accumulation points of the singular values of HT . In particular, we find that
HT is not a compact operator. Finally, we conclude by showing numerical examples
in Section 6.
2. Introducing a differential operator. In this section, we find two differen-
tial operators LS and LS that will turn out to have a commutation property of the
form
HT LS = LSHT .
(2.1)
In order to find the SVD of HT , we will be interested in finding LS and LS with simple
discrete spectra. Initially, it is not apparent whether differential operators with such
properties exist and if so, how to find them. We do not know of a coherent theory that
relates certain integral operators to differential operators via a commutation property
as the above. However, there have been examples of integral operators for which -- by
what seems to be a lucky accident -- such differential operators exist.
One instance is the well-known Landau-Pollak-Slepian (LPS) operator that arises
in signal processing in the study of time- and bandlimited representations of signals
[20, 13, 14]. There, it is of interest to find the largest eigenvalue of the LPS operator
P[−T,T ]F −1P[−W,W ]FP[−T,T ]. Here, F is the Fourier transform, and T and W are
some positive numbers. This operator happens to commute with a second order
differential operator, of which the eigenfunctions and eigenvalues had been studied
long before its connection to the LPS operator was known. The eigenfunctions of this
differential operator are the so-called prolate spheroidal wave functions and they turn
out to be the eigenfunctions of the LPS operator as well. The work of Landau, Pollak
and Slepian has been generalized and extended by Grunbaum et al. [7].
More recent examples of integral operators with commuting differential operators
are the interior Radon transform [15] and two instances of the truncated Hilbert
transform mentioned earlier [8, 9].
To start our search for LS and LS, we follow the procedure in [8, 9] and define a
differential operator
Definition 2.1.
L(x, dx)ψ(x) := (P (x)ψ′(x))′ + 2(x − σ)2ψ(x)
6
(2.2)
where
P (x) =
4
Yi=1
(x − ai),
σ =
1
4
4
Xi=1
ai.
(2.3)
The four points ai are all regular singular, and in a complex neighborhood of each ai
the functions (x− ai)· P ′(x)/P (x) and (x− ai)2· 2(x− σ)2/P (x) are complex analytic.
The term regular singular point is standard in the general theory of differential equa-
tions and, as such, is also used in the theory of S-L equations, see e.g.
[21] for this
and other terminology and basic properties of S-L equations. Consequently, by the
method of Fuchs-Frobenius it follows that for λ ∈ C any solution of Lψ = λψ is either
bounded or of logarithmic singularity close to any of the points ai, see [21]. Away
from the singular points ai the analyticity of the solutions follows from the analyticity
of the coefficients of the differential operator L. More precisely, in a left and a right
neighborhood of each regular singular point ai, there exist two linearly independent
solutions of the form
ψ1(x) = x − aiα1
ψ2(x) = x − aiα2
∞
Xn=0
Xn=0
∞
bn(x − ai)n,
dn(x − ai)n + k lnx − aiψ1(x),
(2.4)
(2.5)
where without loss of generality we can assume b0 = d0 = 1. The exponents α1 and
α2 are the solutions of the indicial equation
where
α2 + (p0 − 1)α + q0 = 0,
p0 = lim
x→ai
q0 = lim
x→ai
(x − ai)P ′(x)/P (x),
(x − ai)2[2(x − σ)2 − λ]/P (x).
(2.6)
(2.7)
With our choice of P , this gives α1 = α2 = 0 which implies k 6= 0. For the bounded
solution in (2.4), α1 = 0 results in ψ1(ai) 6= 0. The radius of convergence of the series
in (2.4) and (2.5) is the distance to the closest singular point different from ai. In a
left and in a right neighborhood of ai, the general form of the solutions of (L−λ)ψ = 0
is
ψ1(x) = ℓ0
ψ2(x) = ℓ1
∞
Xn=0
Xn=0
∞
bn(x − ai)n
dn(x − ai)n + ℓ2 lnx − ai
∞
Xn=0
bn(x − ai)n
(2.8)
(2.9)
for some constants ℓj. Hence we have one degree of freedom for the bounded solution,
and two -- for the unbounded solution. Clearly, for the bounded solutions (2.8),
the coefficients bn are the same on both sides of ai, since we have assumed b0 =
1. However, the bounded part of the unbounded solutions (2.9) may have different
coefficients d−
n to the left and to the right of ai respectively.
n and d+
7
2.1. The Maximal and Minimal Domains and Self-Adjoint Realiza-
tions. Since we are interested in a differential operator that commutes (on some
set to be defined) with HT , we want to consider L on the interval (a2, a4). Due to
the regular singular point a3 in the interior of the interval, standard techniques for
singular S-L problems are not applicable.
It is crucial for our application that we
identify a commuting self-adjoint operator, for which the spectral theorem can be ap-
plied. We therefore wish to study all self-adjoint realizations; we follow the treatment
in Chapter 13 in [26] which gives a characterization of all self-adjoint realizations for
two-interval problems, of which problems with an interior singular point are a special
case.
First of all, one needs to define the maximal and minimal domains on Ij =
(aj, aj+1) (see Chapter 9 in [26]). Let ACloc(I) be the set of all functions that are
absolutely continuous on all compact subintervals of the open interval I. Then,
Dj,max := {ψ : Ij → C : ψ, P ψ′ ∈ ACloc(Ij ); ψ, Lψ ∈ L2(Ij )},
Dj,min := {ψ ∈ Dj,max : supp ψ ⊂ (aj , aj+1)},
and the related maximal and minimal operators are defined as follows:
Lj,max := L(Dj,max) : Dj,max → L2(Ij ),
Lj,min := L(Dj,min) : Dj,min → L2(Ij ).
(2.10)
(2.11)
(2.12)
(2.13)
We shall follow essentially the procedure in Chapter 13 in [26], to which we refer for
more details. On (a2, a4), the maximal and minimal domains and the corresponding
operators are defined as the direct sums:
Definition 2.2. The maximal and minimal domains Dmax, Dmin ⊂ L2(a2, a4)
and the operators Lmax, Lmin are defined as
Dmax := D2,max + D3,max,
Lmax := L2,max + L3,max,
Dmin := D2,min + D3,min
Lmin := L2,min + L3,min
and therefore
Lmax : Dmax → L2(a2, a4),
Lmin : Dmin → L2(a2, a4).
(2.14)
(2.15)
The operator Lmin is a closed, symmetric, densely defined operator in L2(a2, a4) and
Lmax, Lmin form an adjoint pair, i.e. L∗
min = Lmax. In order to
define a self-adjoint extension of Lmin, we need to introduce the notion of the Lagrange
sesquilinear form:
max = Lmin and L∗
where, at the singular points,
[u, v] := uP v′ − vP u′,
[u, v](a+
i ) := lim
α→a+
i
[u, v](a−
i ) := lim
α→a−
i
[u, v](α),
[u, v](α).
8
(2.16)
(2.17)
(2.18)
These limits exist and are finite for all u, v ∈ Dmax. If we choose u, v ∈ Dmax such
that [u, v](ai) = 1 for all the singular points (a+
4 ), then the extension of
Lmin defined by the following conditions
3 , a+
2 , a−
3 , a−
[ψ, u](a+
[ψ, u](a−
[ψ, v](a−
2 ) = 0 = [ψ, u](a−
4 )
3 ) = [ψ, u](a+
3 )
3 ) = [ψ, v](a+
3 )
(2.19)
(2.20)
(2.21)
is self-adjoint. We refer to (2.19) as boundary conditions, and to (2.20) and (2.21)
-- as transmission conditions. The latter connect the two subintervals (a2, a3) and
(a3, a4). Motivated by the conditions mentioned in Section 1, we define a self-adjoint
extension of Lmin:
Lemma 2.3. The extension LS : D(LS) → L2(a2, a4) of Lmin to the domain
D(LS) := {ψ ∈ Dmax : [ψ, u](a+
[ψ, u](a−
2 ) = [ψ, u](a−
3 ) = [ψ, u](a+
4 ) = 0,
3 ), [ψ, v](a−
3 ) = [ψ, v](a+
3 )}
(2.22)
with the following choice of maximal domain functions u, v ∈ Dmax
u(y) := 1,
v(y) :=
4
Xi=1 Yj6=i
j∈{1,...,4}
1
ai − aj
lny − ai,
(2.23)
(2.24)
is self-adjoint.
This choice of maximal domain functions gives [u, v](ai) = 1 for i = 1, . . . , 4. The
boundary conditions simplify to
lim
y→a+
2
P (y)ψ′(y) = lim
y→a−
4
P (y)ψ′(y) = 0.
(2.25)
For an eigenfunction ψ of LS this is equivalent to ψ being bounded at a+
2 and a−
4
(because the only possible singularity is of logarithmic type). Let φ1 and φ2 be
the restrictions of ψ to the intervals (a2, a3) and (a3, a4), respectively. Since ψ is
an eigenfunction, on the corresponding intervals φ1 and φ2 are of the form φi(y) =
φi1(y) + φi2(y) lny − a3. Here, the functions φij are analytic on (a2, a3) for i = 1
and on (a3, a4) -- for i = 2. Having this, the transmission conditions can be simplified
as follows:
3 ) = [ψ, u](a−
3 )
[ψ, u](a+
P (y)ψ′(y) = lim
y→a−
3
P (y)ψ′(y)
lim
y→a+
3
The condition involving v yields
lim
y→a−
3
φ12(y) = lim
y→a+
3
φ22(y).
lim
y→a−
3
[ψ, v](a−
[ψ(y) − v(y)(P ψ′)(y)] = lim
φ11(y) = lim
y→a+
3
lim
y→a−
3
3 ) = [ψ, v](a+
3 )
[ψ(y) − v(y)(P ψ′)(y)]
φ21(y)
y→a+
3
9
(2.26)
(2.27)
(2.28)
Note that on each side of (2.27) the logarithmic terms in φi2 cancel because of the
choice of the constants in v. The properties (2.25), (2.26) and (2.28) are the same
as the ones found for fn in Section 1. Thus, we have constructed an operator LS for
which close to the points a2, a3 and a4, the eigenfunctions behave in the same way
that is expected for the fn's.
Close to a3, an eigenfunction ψ is given by
ψ(y) =
ℓ11
ℓ12
∞
∞
Pm=0
Pm=0
d−
m(y − a3)m + ℓ21 lny − a3
d+
m(y − a3)m + ℓ22 lny − a3
∞
∞
Pm=0
Pm=0
bm(y − a3)m,
bm(y − a3)m,
y < a3
y > a3
(2.29)
where similarly to (2.5), we assume d−
conditions require that
0 = d+
0 = 1 and b0 = 1. The transmission
(2.30)
(2.31)
(2.32)
ℓ11 = ℓ12,
ℓ21 = ℓ22.
We can thus express ψ in a sufficiently small neighborhood of a3 as
∞
∞
ψ(y) = ℓ11 + ℓ21 lny − a3
Xm=0
bm(y − a3)m +
Xm=1
m = ℓ11d−
m, when y > a3 and for ℓ−
ℓ±
m(y − a3)m,
m, when y < a3.
where ℓ±
m stands for ℓ+
m = ℓ11d+
3. The spectrum of LS. In order to prove that the spectrum of the differential
self-adjoint operator LS introduced in Lemma 2.3 is discrete, we need to show that
for some z in the resolvent set, (LS − zI)−1 is a compact operator. To do so, it is
sufficient to prove that the Green's function G of LS − zI, which for z in the resolvent
set exists and is unique, is a function in L2((a2, a4)2). This would allow us to conclude
that the integral operator TG with G as its integral kernel is a compact operator from
L2(a2, a4) to L2(a2, a4), where TG is equivalent to the inversion of LS − zI.
Lemma 3.1. The Green's function G(x, ξ) associated with LS−i is in L2((a2, a4)2)
and consequently, (LS − i)−1 : L2(a2, a4) → D(LS) ⊂ L2(a2, a4) is a compact opera-
tor.
Proof. The self-adjointness of LS is equivalent to LS − i being one-to-one and
onto (Theorem VIII.3 in [19]). Moreover, the ai's are limit-circle points and thus,
the deficiency index d equals 4 (Theorem 13.3.1 in [26]). This means that if we do
not impose boundary and transmission conditions, there are two linearly independent
solutions p1 and p2 of (L− i)p = 0 on (a2, a3) as well as two linearly independent solu-
tions q1 and q2 of (L − i)q = 0 on (a3, a4). Note that none of these four solutions can
be bounded at both of its endpoints because i is not an eigenvalue of the self-adjoint
operator Lj,S : D(Lj,S) → L2(Ij ) with D(Lj,S) = {ψ ∈ Dj,max : limy→a+
P (y)ψ′(y) =
P (y)ψ′(y) = 0}. By taking appropriate combinations, if necessary, we can
limy→a−
eliminate the logarithmic singularity at a+
4 -- of
another solution. We can thus assume that
2 of one of the solutions, and at a−
j+1
j
- on (a2, a3): p1 is bounded at a+
2 and logarithmic at a−
3 , p2 is logarithmic at
both endpoints;
10
- on (a3, a4): q1 is logarithmic at a+
3 and bounded at a−
4 , q2 is logarithmic at
both endpoints.
We next check the restrictions imposed by the transmission conditions at a3. Close
to a3, both functions p1 and q2 are of the form (2.9). Let ℓ11, ℓ21 denote the free
parameters in the expression for p1 and ℓ12, ℓ22 the ones in q2. These can be chosen
such that they satisfy (2.30) and (2.31). Thus, there exists a solution h1(x) on (a2, a4)
given by
q2(x)
h1(x) = (cid:26) p1(x)
2 and logarithmic at a−
for x ∈ (a2, a3)
for x ∈ (a3, a4)
that is bounded at a+
4 . In addition, it is of the form (2.32) close
to a3, i.e. it is logarithmic at a3 and satisfies the transmission conditions (2.26), (2.28)
there. Similarly, with p2 and q1 we can obtain a solution h2 on (a2, a4) that satisfies the
transmission conditions at a3 and is of ln-ln-bounded-type. Thus, imposing only the
transmission conditions, we obtain two linearly independent solutions of (L − i)h = 0
on (a2, a3) ∪ (a3, a4). One of them, h1, is of a bounded-ln-ln-type, and the other one,
h2, is of a ln-ln-bounded-type, at the points a+
4 , respectively. We are now in a
position to consider the Green's function G(x, ξ) of LS − i. Close to a3, we can write
the two functions as hj(x) = hj1(x) + lnx − a3hj2(x) with continuous functions hj1
and hj2. By rescaling if necessary, we can assume h12(a3) = h22(a3). We construct G
from h1 and h2 as follows:
2 , a3, a−
G(x, ξ) = (cid:26) c1(ξ)h1(x)
c2(ξ)h2(x)
for x < ξ
for x > ξ
(3.1)
where ξ ∈ (a2, a3) ∪ (a3, a4) and the functions c1(ξ) and c2(ξ) are chosen such that G
is continuous at x = ξ and ∂G/∂x has a jump discontinuity of 1/P (ξ) at x = ξ:
c1(ξ)h1(ξ) − c2(ξ)h2(ξ) = 0,
c1(ξ)h′
2(ξ) = −
1(ξ) − c2(ξ)h′
1
P (ξ)
.
(3.2)
(3.3)
In other words, G is the solution of (L − i)G = δ, where δ is the Dirac delta function.
For ξ away from a3, G(x, ξ) is continuous in ξ but with logarithmic singularities at
a+
2 and a−
4 . This can be seen as follows. Consider ξ close to a2. There, we can write
h2(ξ) = h21(ξ) + h22(ξ) lnξ − a2
and, since h1 is bounded close to a+
Wh1,h2 denote the Wronskian of h1 and h2, i.e. Wh1,h2 = h1h′
we obtain
2 , it is of the form (2.8), i.e. h1(a+
2 − h′
2 ) 6= 0. Let
1h2. For c1 and c2
c1(ξ) =
c2(ξ) =
h2(ξ)
P (ξ)Wh1,h2(ξ)
h1(ξ)
P (ξ)Wh1,h2(ξ)
,
.
(3.4)
(3.5)
The denominator in the above expressions is bounded by
P (ξ)(h1(ξ)h′
2(ξ) − h2(ξ)h′
1(ξ)) = O((ξ − a2) lnξ − a2) + h1(ξ)h22(ξ)p(ξ),
11
where p(ξ) = P (ξ)/(ξ − a2) and h1(a+
a+
2 ,
2 )h22(a2)p(a2) 6= 0. Thus, in a neighborhood of
c1(ξ) = O(ln ξ − a2),
c2(ξ) = O(1).
Similarly, since h2(a−
4 ) 6= 0, close to a−
4
c1(ξ) = O(1),
c2(ξ) = O(ln ξ − a4).
(3.6)
(3.7)
(3.8)
(3.9)
For each fixed ξ ∈ (a2, a3) ∪ (a3, a4), G(x, ξ) as a function in x is continuous on
[a2, a3) ∪ (a3, a4] and has a logarithmic singularity at a3, due to the singularities in
h1(x) and h2(x). It remains to check what happens as ξ → a3. We need to make
sure that the functions c1(ξ) and c2(ξ) behave in such a way that G ∈ L2((a2, a4)2).
Therefore, we derive the asymptotics of c1(ξ) and c2(ξ) as ξ → a−
3 . For ξ = a3 − ǫ
and small ǫ > 0, equation (3.2) becomes
c1(a3 − ǫ)h1(a3 − ǫ) − c2(a3 − ǫ)h2(a3 − ǫ) = 0.
Since close to a3, hi = hi1 + hi2 ln(ǫ) and the hij are continuous, the ratio c1/c2 is of
the form
a + b ln(ǫ)
c + d ln(ǫ)
,
where b and d are non-zero (because the logarithmic singularity is present). Thus,
the ratio tends to the finite limit b/d as ǫ → 0. Conditions (3.2) and (3.3) together
imply:
c2(a3 − ǫ) =
=
where
h1(a3 − ǫ)
P (a3 − ǫ)Wh1,h2(a3 − ǫ)
h11(a3 − ǫ) + h12(a3 − ǫ) ln(ǫ)
r1(ǫ) + ǫ · r2(ǫ)
1
ǫ
P (a3 − ǫ)(cid:0)h21h12 − h22h11(cid:1)(a3 − ǫ),
r1(ǫ) = −
r2(ǫ) = O(1),
If r1(0) 6= 0, then c2 is of order O(ln(ǫ)), removing a possible
and h12(a3) 6= 0.
obstruction to square integrability of G.
Suppose r1(0) = 0, i.e.
h21(a3)h12(a3) − h22(a3)h11(a3) = 0.
This would imply
h11(a3) = C · h21(a3),
h12(a3) = C · h22(a3)
12
(3.10)
(3.11)
for some constant C. By assumption, h12(a3) = h22(a3), so that C = 1. Now if both
(3.10) and (3.11) hold for C = 1, the function defined by
h(x) = (cid:26) h1(x)
h2(x)
for x ∈ (a2, a3)
for x ∈ (a3, a4)
would be a non-trivial solution of (LS − i)h = 0 (fulfilling both boundary and trans-
mission conditions), i.e. i would be an eigenvalue of LS. But this contradicts the self-
adjointness of LS. We can thus conclude that r1(0) 6= 0. This shows that c2(a3 − ǫ)
is of order O(ln(ǫ)) and therefore also c2 · c1
Analogously, we can find the same asymptotics of c1(ξ) and c2(ξ) as ξ → a+
3 .
Therefore, the properties of the Green's function G(x, ξ) can be summarized as fol-
lows:
= c1 = O(ln(ǫ)).
c2
- G(·, ξ) has logarithmic singularities at a+
- G(x,·) is of logarithmic singularity at a3,
- away from these singularities G(x, ξ) is continuous in x and ξ.
2 , a3 and a−
4 ,
Thus, G is in L2((a2, a4)2). Hence, TG : L2(a2, a4) → L2(a2, a4) is a compact Fred-
holm integral operator.
From this we conclude:
Proposition 3.2. The operator LS has only a discrete spectrum, and the asso-
ciated eigenfunctions are complete in L2(a2, a4).
Proof. By Theorem VIII.3 in [19], the self-adjointness of LS implies that for the
operator (LS − i) : D(LS) → L2(a2, a4) we have
Ker(LS − i) = {0},
Ran(LS − i) = L2(a2, a4).
(3.12)
(3.13)
Consequently, (LS − i)−1 : L2(a2, a4) → D(LS) is one-to-one and onto. Moreover, it
is a normal compact operator and thus we get the spectral representation
(LS − i)−1f =
∞
Xn=0
λnhf, fnifn,
(3.14)
where {fn}n∈N is a complete orthonormal system in L2(a2, a4). This can be trans-
formed into the spectral representation for LS:
LSf =
∞
(
Xn=0
1
λn
+ i)hf, fnifn.
(3.15)
Clearly, the eigenfunctions fn of LS can be chosen to be real-valued. The com-
pleteness of {fn}n∈N is essential for finding the SVD of HT . Another property that
will be needed for the SVD is that the spectrum of LS is simple, i.e. that each eigen-
value has multiplicity 1.
13
Proposition 3.3. The spectrum of LS is simple.
Proof. From the compactness of (LS − i)−1, we know that each eigenvalue has
finite multiplicity. Suppose f1 and f2 are linearly independent eigenfunctions of LS
corresponding to the same eigenvalue λ ∈ R. Then, on all of (a2, a3) ∪ (a3, a4) the
following holds
f1Lf2 − f2Lf1 = 0.
(3.16)
Consequently,
0 = f1Lf2 − f2Lf1 = f1(P f ′
= [f1, f2]′.
2)′ − f2(P f ′
1)′
Thus, [f1, f2] is constant on both (a2, a3) and (a3, a4). From the boundary conditions
that f1 and f2 satisfy, we find that [f1, f2](a+
4 ), which implies
[f1, f2] = 0 on (a2, a3) ∪ (a3, a4). Since [f1, f2] = P (f ′
2 ) = 0 = [f1, f2](a−
2), we get that
1f2 − f1f ′
2 = 0 on (a2, a3) ∪ (a3, a4).
f ′
1f2 − f1f ′
(3.17)
The functions f1 and f2 satisfy the transmission conditions at a3. Consequently, they
can be written as
f1(x) = f11(x) + f12(x) ln x − a3,
f2(x) = f21(x) + f22(x) ln x − a3,
in a neighborhood of a3, where fij are continuous. Since the one-sided derivatives f ′
ij
are bounded at a3, equation (3.17) implies
(cid:0)f12f21 − f11f22(cid:1)(x)
x − a3
+ O(ln2 x − a3) = 0.
(3.18)
Note that the terms containing lnx − a3/(x − a3) cancel. Taking the limit x → a3
in (3.18), we obtain
f12(a3)f21(a3) − f11(a3)f22(a3) = 0.
Thus, for some constant C:
(cid:18) f11(a3)
f12(a3) (cid:19) = C(cid:18) f21(a3)
f22(a3) (cid:19) .
If we take f1 on (a2, a3), then f11(a3) and f12(a3) define a singular initial value
problem on (a3, a4) that is uniquely solvable (Theorem 8.4.1 in [26]). Thus, f1 = C·f2
on (a3, a4). Now, on the other hand, by considering f1 on (a3, a4), the values f11(a3)
and f12(a3) define a singular initial value problem on (a2, a3) which has a unique
solution. Hence, f1 = C · f2 on (a2, a3) ∪ (a3, a4) in contradiction to our assumption.
4. Singular value decomposition of HT . Having introduced the differential
operator LS, we now want to relate it to the truncated Hilbert transform HT . The
main result of this section is that the eigenfunctions of LS fully determine the two
14
families of singular functions of HT . We start by stating the following
Proposition 4.1. On the set of eigenfunctions {fn}n∈N of LS, the following
commutation relation holds:
(HT L(y, dy)fn)(x) = L(x, dx)(HT fn)(x)
for x ∈ (a1, a2) ∪ (a2, a3).
(4.1)
Sketch of proof. This proof follows the same general idea as the proof of Proposition
2.1 in [9]. We therefore provide full details only for those steps where additional care
needs to be taken because of the singularity at a3. The steps that are completely
analogous to those in the proof of Proposition 2.1 in [9] are only sketched here.
Let ψ ∈ {fn}n∈N. The boundedness of ψ at a+
4 implies that P ψ′ → 0
and P ψ → 0 there. Moreover, the transmission conditions at a3 guarantee that P ψ′
is continuous at a3. With these properties, the commutation relation for x ∈ (a1, a2),
i.e. where the Hilbert kernel is not singular, can be shown similarly to the proof of
Proposition 2.1 in [8].
2 and a−
Next, let x ∈ (a2, a3). The main difference from the proof of Proposition 2.1 in [9]
is that now the eigenfunctions are not in C∞([a2, a4]), but are singular at a3. However,
the fact that we exclude the point x = a3 allows us to always have a neighborhood of x
away from a3 on which ψ is bounded. We further note that ψ ∈ C∞([a2, a3)∪(a3, a4]).
Since the Hilbert kernel is singular, we need to use principal value integration and
introduce the following notation: Iǫ(x) := [a2, x − ǫ] ∪ [x + ǫ, a4]. Here ǫ > 0 is so
small that (x − ǫ, x + 2ǫ) ⊂ (a2, a3), i.e. the ǫ-neighborhood of x is well separated
from a3. Then,
π(HT L(y, dy)ψ)(x) = lim
ǫ→0+ZIǫ(x)h (P (y)ψ′(y))′
y − x
+
2(y − σ)2ψ(y)
y − x
idy.
For the first term under the integral, we integrate by parts twice and plug in the
boundary conditions. Again, we use that P ψ′ → 0 and P ψ → 0 at a+
2 and a−
4 :
Z
Iǫ(x)
(P (y)ψ′(y))′
y − x
dy = −
(P ψ′)(x − ǫ) + (P ψ′)(x + ǫ)
ǫ
+
(P ψ)(x − ǫ) − (P ψ)(x + ǫ)
ǫ2
+ Z
Iǫ(x)
ψ(y)
2P (y) − P ′(y)(y − x)
(y − x)3
dy.
(4.2)
The integral on the right-hand side of (4.2) can be related to the derivatives of
In [9] similar relations (cf. eq.
R ψ(y)/(y − x)dy.
(2.7)) were obtained from the
Leibniz integral rule, using explicitly that the integrand was continuous. In our case,
the function ψ is no longer continuous because of the singularity at a3. We can
generalize the argument of [9] by invoking the dominated convergence theorem and
15
rewrite the last term in (4.2) as follows:
2P (y) − P ′(y)(y − x)
dy =
dy +
ψ′(x − ǫ) + ψ′(x + ǫ)
ǫ
dy +
ψ(x − ǫ) + ψ(x + ǫ)
ǫ
ψ(x − ǫ) − ψ(x + ǫ)
ǫ2
i
−
i
(y − x)3
ψ(y)
y − x
ψ(y)
y − x
Iǫ(x)
Iǫ(x)
ψ(y)
Z
= P (x)h d2
dx2 Z
+ P ′(x)h d
− Z
2ψ(y)
Iǫ(x)
dx Z
Iǫ(x)
(y − σ)2 − (x − σ)2
dy
y − x
Putting all pieces together, we obtain:
π(HT L(y, dy)ψ)(x) =
= lim
ǫ→0+n −
ǫ
(P ψ′)(x − ǫ) + (P ψ′)(x + ǫ)
ψ′(x − ǫ) + ψ′(x + ǫ)
ψ(x − ǫ) + ψ(x + ǫ)
+ P (x)(cid:2)
+ P ′(x)
ǫ
ǫ
+
(P ψ)(x − ǫ) − (P ψ)(x + ǫ)
ǫ2
ǫ2
−
+ L(x, dx) Z
ψ(x − ǫ) − ψ(x + ǫ)
(cid:3)
dyo
ψ(y)
y − x
Iǫ(x)
The eigenfunction ψ is in C∞[a2, x + 2ǫ). Following [9], we can thus express the
boundary terms in the above equation by Taylor expansions around x and make use
of the fact that the boundary terms consist only of odd functions in ǫ. The boundary
terms are then of the order O(ǫ). We thus have
π(HT L(y, dy)ψ)(x) = lim
ǫ→0+
L(x, dx)ZIǫ(x)
ψ(y)
y − x
dy.
(4.3)
Since for ǫ > 0 sufficiently small, ψ ∈ C∞([x− ǫ, x + ǫ]), one can interchange the limit
with L(x, dx) as in [9].
Because the spectrum of LS is purely discrete, we have thus found an orthonormal
basis (the eigenfunctions of LS) {fn}n∈N of L2(a2, a4) for which (4.1) holds. Let
us define gn := HT fn/kHT fnkL2(a1,a3). Then, in order to obtain the SVD for HT
(with singular functions fn and gn), it is sufficient to prove that the gn's form an
orthonormal system of L2(a1, a3) (they will then consequently form an orthonormal
basis of L2(a1, a3), see Proposition 5.2).
The orthogonality of the gn's will follow from the commutation relation. Since
fn is an eigenfunction of LS for some eigenvalue λn, we obtain
L(x, dx)gn(x) = λngn(x),
x ∈ (a1, a2) ∪ (a2, a3).
Similarly to LS, we define a new self-adjoint operator that acts on functions supported
on [a1, a3]:
16
Definition 4.2. Let Dmax := D1,max + D2,max and Lmin := L1,min + L2,min. The
operator LS : D( LS) → L2(a1, a3) is defined as the self-adjoint extension of Lmin,
where
D( LS) := {ψ ∈ Dmax : [ψ, u](a+
[ψ, u](a−
1 ) = [ψ, u](a−
2 ) = [ψ, u](a+
3 ) = 0,
2 ), [ψ, v](a−
2 ) = [ψ, v](a+
2 )}
(4.4)
with the maximal domain functions u, v ∈ Dmax as in (2.23), (2.24).
The intuition then is the following. The function fn is bounded at a+
2 and log-
arithmic at a3, where it satisfies the transmission conditions. Consequently, as will
be shown below, gn is bounded at a−
3 , logarithmic at a2 and satisfies the correspond-
ing transmission conditions at a2. Clearly, it is also bounded at a+
1 . Thus, gn is an
eigenfunction of the self-adjoint operator LS. As a consequence, the gn's form an
orthonormal system.
Proposition 4.3.
If LSfn = λnfn, then gn := HT fn/kHT fnkL2(a1,a3) is an
eigenfunction of LS corresponding to the same eigenvalue
LSgn = λngn.
(4.5)
Proof. First of all, the commutation relation for fn yields
L(x, dx)(HT fn)(x) = (HT L(y, dy)fn)(x),
L(x, dx)gn(x) = λngn(x),
x ∈ (a1, a2) ∪ (a2, a3).
ditions. Therefore, we consider p.v.R a4
What remains to be shown is that gn satisfies the boundary and transmission con-
fn(y)/(y − x)dy for x close to a1, a2 and a3.
In a neighborhood of a1 away from [a2, a4] this function is clearly analytic. Next, let
x be confined to a small neighborhood of a2. Since the discontinuity of fn is away
from a2, we can split the above integral into two -- one that integrates over a right
neighborhood of a2 and another one that is an analytic function. The first item in
Lemma 1.2 then implies that
a2
p.v.
a4
Z
a2
fn(y)
y − x
dy = gn,1(x) − fn(a+
2 ) lnx − a2
(4.6)
where gn,1(x) is continuous in a neighborhood of x = a2. Thus, gn satisfies the
transmission conditions (2.26), (2.28).
It remains to check the behavior of gn close to a−
3 . We first express fn as
fn(y) = fn,1(y) + fn,2(y) lny − a3,
where both fn,1 and fn,2 are Lipschitz continuous. Then, in view of Lemma 1.2, both
summands on the right-hand side of the equation
p.v.
a4
Z
a2
fn(y)
y − x
dy =p.v.
a4
Z
a2
fn,1(y)
y − x
dy + p.v.
17
a4
Z
a2
fn,2(y) lny − a3
y − x
dy
remain bounded as x tends to a3.
Since the spectrum of LS is simple, we can conclude that the gn's form an or-
thonormal system and thus the following holds:
Theorem 4.4. The eigenfunctions fn of LS, together with
gn := HT fn/kHT fnkL2(a1,a3) and σn := kHT fnkL2(a1,a3) form the singular value
decomposition for HT :
HT fn = σngn,
H ∗
T gn = σnfn.
(4.7)
(4.8)
5. Accumulation points of the singular values of HT . The main result of
this section is that 0 and 1 are accumulation points of the singular values of HT . To
find this, we first analyze the nullspace and range of HT , which will also prove the
ill-posedness of the inversion of HT . First, we need to state the following
Lemma 5.1.
If the Hilbert transform of a compactly supported f ∈ L2(a, b)
vanishes on an open interval (c, d) disjoint from the object support, then f = 0 on all
of R.
Sketch of proof. A similar statement (and proof) can be found in [1]. The main
difference is that here we consider a more general class of functions f . By dominated
convergence, f ∈ L1(a, b) implies that for any z ∈ Ω = C\((−∞, a) ∪ (b,∞)), the
function g(z) = R b
a f (x)/(x − z)dx is differentiable in a neighborhood of z. Thus, g
is analytic on Ω. The statement then follows in the same way as Lemma 2.1 in [1].
With this property of the Hilbert transform, we can obtain results on the nullspace
and the range of HT :
Proposition 5.2. The operator HT : L2(a2, a4) → L2(a1, a3) has a trivial
nullspace and dense range that is not all of L2(a1, a3), i.e.
Ker(HT ) = {0},
Ran(HT ) 6= L2(a1, a3),
Ran(HT ) = L2(a1, a3).
(5.1)
(5.2)
(5.3)
Proof of (5.1). Suppose HT f = 0. Then
Hχ[a2,a4]f = 0 on (a1, a2),
and by Lemma 5.1, f = 0 on all of [a2, a4]. Thus, f ∈ L2(a2, a4) can always be
uniquely determined from HT f .
Proof of (5.2). Take any g ∈ L2(a1, a3) that vanishes on (a1, a2) and such that
kgkL2(a1,a3) 6= 0. Suppose g ∈ Ran(HT ). By Lemma 5.1, if f ∈ L2(a2, a4) and
HT f = g, then f is zero on [a2, a4]. This implies that g = 0 on (a1, a3), which
contradicts the assumption kgk 6= 0.
Proof of (5.3). The operator H ∗
same general properties. By the above argument, Ker(H ∗
Thus, Ran(HT )⊥ = {0}.
T is also a truncated Hilbert transform with the
T ) = {0}.
18
Equation (5.2) shows the ill-posedness of the problem. It is not true that for every
g ∈ L2(a1, a3) there is a solution f to the equation HT f = g. Since Ran(HT ) is dense,
the solution need not depend continuously on the data. Thus, our problem violates
two properties of Hadamard's well-posedness criteria [3]. These are the existence of
solutions for all data and the continuous dependence of the solution on the data. We
now turn to the spectrum of H ∗
T HT . In what follows, k·k denotes the norm associated
with L2(R), and h·,·i denotes the L2(R) inner product. We begin with proving the
following
Lemma 5.3. The operator H ∗
T HT has norm equal to 1.
Proof. From kHk = 1, we know that kH ∗
T HTk = sup
kH ∗
kψk=1kH ∗
T HT ψk,
T HTk ≤ 1. Since
finding a sequence ψn with kψnk = 1 and kH ∗
T HT ψnk → 1 would prove the assertion.
Take a compactly supported function ψ ∈ L2([−1, 1]) with kψk = 1 and two
vanishing moments, R 1
−1 x· ψ(x)dx. From this, we define a family of
functions, such that the norm is preserved but the supports decrease. More precisely,
for a > 2/(a3 − a2), we set
−1 ψ(x)dx = 0 = R 1
ψa(x) = √aψ(a(x −
a2 + a3
2
)).
(5.4)
These functions satisfy kψak = 1 and supp ψa = [ a2+a3
For their Hilbert transforms we obtain
2 − 1
a , a2+a3
2 + 1
a ] ⊂ [a2, a3].
(Hψa)(x) = √a(Hψ)(a(x −
a2 + a3
2
)).
We can write
H ∗
T HT ψa = −χ[a2,a4]Hχ[a1,a3]Hχ[a2,a4]ψa
= −χ[a2,a4]H(I − (I − χ[a1,a3]))Hχ[a2,a4]ψa
= ψa + χ[a2,a4]H(I − χ[a1,a3])Hχ[a2,a4]ψa;
(I − H ∗
T HT )ψa = −χ[a2,a4]H(I − χ[a1,a3])Hχ[a2,a4]ψa.
Consider the L2-norm of the last expression
(5.5)
(5.6)
k(I − H ∗
T HT )ψak2 = kχ[a2,a4]H(I − χ[a1,a3])Hχ[a2,a4]ψak2
≤ k(I − χ[a1,a3])Hχ[a2,a4]ψak2
= Z(−∞,a1)∪(a3,∞) (Hψa)(x)2dx
= aZ(−∞,a1)∪(a3,∞) (Hψ)(a(x −
= Z a·(a1− a2 +a3
Hψ2dy +Z ∞
)
2
−∞
a2 + a3
2
)2dx
a·(a3− a2 +a3
2
) Hψ2dy.
(5.7)
Because of the ordering of the ai's, we have that a1 − a2+a3
2 > 0.
Since ψ has two vanishing moments, Hψ asymptotically behaves like 1/y3 and hence,
2 < 0 and a3 − a2+a3
19
both integrals in (5.7) are of the order O(a−5). Thus, given any ǫ > 0, one can find
a > 2/(a3 − a2) such that
k(I − H ∗
T HT )ψak < ǫ.
Consequently,
Therefore
kH ∗
T HT ψak ≥ kψak − k(I − H ∗
T HT )ψak > 1 − ǫ.
which implies that kH ∗
kH ∗
T HTk = 1.
T HT ψak → 1 as a → ∞,
We are now in a position to prove
Theorem 5.4. The values 0 and 1 are accumulation points of the singular values
of HT .
T HT ).
Proof. First of all, 0 and 1 are both elements of the spectrum σ(H ∗
T HT ). For
T ) 6= L2(a2, a4). Moreover, since
T HT is self-adjoint, the spectral radius is equal to 1. Thus,
T HT ) ⊂ Ran(H ∗
T HTk = 1 and H ∗
The second step is to show that 0 and 1 are not eigenvalues of H ∗
the value 0, this follows from Ran(H ∗
kH ∗
1 ∈ σ(H ∗
0 is not an eigenvalue:
T HT fi = 0. Since
Ker(HT ) = 0, this implies f = 0. Thus, Ker(H ∗
1 is not an eigenvalue: Suppose there exists a non-vanishing function f ∈ L2(a2, a4),
such that
T HT f = 0, then kHT fk2 = hf, H ∗
T HT ) = {0}.
If H ∗
T HT .
Then,
−χ[a2,a4]Hχ[a1,a3]Hχ[a2,a4]f = f.
kHχ[a2,a4]fk2 = kfk2 = −hχ[a2,a4]Hχ[a1,a3]Hχ[a2,a4]f, fi
= hχ[a1,a3]Hχ[a2,a4]f, Hχ[a2,a4]fi
= kχ[a1,a3]Hχ[a2,a4]fk2.
This implies that Hχ[a2,a4]f is identically zero outside [a1, a3]. By Lemma 5.1, this
implies f = χ[a2,a4]f = 0, contradicting the assumption f 6≡ 0. Therefore, 0 and 1
are accumulation points of the eigenvalues of H ∗
T HT and consequently, of the singular
values of HT .
Since the singular values of HT also accumulate at a point other than zero, the
operator HT is not compact.
6. Numerical illustration. We want to illustrate the properties of the trun-
cated Hilbert transform with overlap obtained above for a specific configuration. We
choose a1 = 0, a2 = 1.5, a3 = 6 and a4 = 7.5. First, we consider two different
discretizations of HT and calculate the corresponding singular values. We choose the
first discretization to be a uniform sampling with 601 partition points in each of the
two intervals [0, 6] and [1.5, 7.5]. Let vectors X and Y denote the partition points of
[0, 6] and [1.5, 7.5] respectively. To overcome the singularity of the Hilbert kernel the
vector X is shifted by half of the sample size. The i-th components of the two vectors
20
100 (i + 1
2
2 ) and Yi = 1.5 + 1
X and Y are given by Xi = 1
100 i; HT is then discretized as
(HT )i,j = (1/π)(Xi − Yj), i, j = 0, . . . , 600. Figure 6.1(a) shows the singular values
for the uniform discretization. We see a very sharp transition from 1 to 0.
The second discretization uses orthonormal wavelets with two vanishing moments.
Let φ denote the scaling function. For the discretization we define a finest scale
J = −7. The scaling functions on [1.5, 7.5] are taken to be φ−7,k for integers k =
192, . . . , 957, i.e. such that supp φ−7,k ⊂ [1.5, 7.5]. On the interval [0, 6] the scaling
for integers ℓ =
functions are shifted in the sense that we take them to be φ−7,ℓ+ 1
0, . . . , 765, i.e. such that supp φ−7,ℓ+ 1
2 ⊂ [0, 6]. Figure 6.1(b) shows a plot of the
singular values of this wavelet discretization of HT . Although the transition is not as
sharp as in 6.1(a), the singular values in both cases very clearly accumulate at 0 and
1.
1.0
0.8
0.6
0.4
0.2
1.0
0.8
0.6
0.4
0.2
100
200
300
400
500
600
100
200
300
400
500
600
700
(a) Uniform discretization of size 601 × 601.
(b) Wavelet discretization of size 766 × 766.
Fig. 6.1: a1 = 0, a2 = 1.5, a3 = 6, a4 = 7.5. Singular values of two discretizations of
HT .
Next, we consider the singular functions. Figure 6.2 shows the singular functions
of the uniform discretization for singular values in the transmission region between
0 and 1. Figure 6.3 illustrates the behavior of singular functions for small singular
values. As anticipated, they are bounded at the two endpoints and singular at the
point of truncation. Figure 6.4 gives two examples of the close to linear behavior in
a log-linear plot of the singular functions. In agreement with the theory in Section 4,
these plots confirm that the singularities are of logarithmic kind.
Based on the numerical experiments conducted, we make the following observa-
tions on the behavior of the singular functions and singular values. First, the singular
functions in Figures 6.2 and 6.3 have the property that two functions with consecu-
tive indices have the number of zeros differing by 1. Moreover, the zeros are located
only inside one subinterval Ij . Furthermore, the plots show that singular functions
with zeros within the overlap region correspond to significant singular values, whereas
those which have zeros outside the overlap region correspond to small singular values.
Finally, we remark that singular functions for small singular values are concentrated
outside the ROI I2 = [a2, a3].
Acknowledgments. RA was supported by an FWO Ph.D. fellowship and would
like to thank Prof. Ingrid Daubechies and Prof. Michel Defrise for their supervision
and contribution. Also, RA would like to thank the Department of Mathematics at
Duke University for hosting several research stays. AK was supported in part by NSF
grants DMS-0806304 and DMS-1211164.
21
0.15
0.10
0.05
-0.05
-0.10
-0.15
0.05
1
2
3
4
5
6
7
1
2
3
4
5
6
7
-0.05
-0.10
-0.15
(a) Singular functions f448 and f449.
(b) Singular functions f450 and f451.
0.15
0.10
0.05
-0.05
-0.10
-0.15
1
2
3
4
5
6
(c) Singular functions g448 and g449.
(d) Singular functions g450 and g451.
Fig. 6.2: Consecutive singular functions for the uniform discretization with 3, 2,
1 and no zeros within the overlap region. The corresponding singular values are
σ448 = 0.999963, σ449 = 0.998782, σ450 = 0.966192, σ451 = 0.542071.
22
0.2
0.1
-0.1
-0.2
0.2
0.1
-0.1
-0.2
1
2
3
4
5
6
7
0.2
0.1
-0.1
-0.2
1
2
3
4
5
6
7
(a) Singular functions f452 and f453.
(b) Singular functions f454 and f455.
1
2
3
4
5
6
0.2
0.1
-0.1
-0.2
1
2
3
4
5
6
(c) Singular functions g452 and g453.
(d) Singular functions g454 and g455.
Fig. 6.3: Consecutive singular functions for the uniform discretization with 1, 2, 3
and 4 zeros outside the overlap region. The corresponding singular values are σ452 =
6.29189 · 10−3, σ453 = 2.83533 · 10−5, σ454 = 1.18274 · 10−7, σ455 = 4.83357 · 10−10.
0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
-0.20
0.005
0.010
0.050
0.100
0.500
0.005
0.010
0.050
0.100
0.500
(a)
(b)
Fig. 6.4: Log-linear plot, i.e. with a logarithmic x scale of the singular functions g450
(left) and g453 (right) on the interval [1.5, 2.3].
23
REFERENCES
[1] M Courdurier, F Noo, M Defrise, and H Kudo. Solving the interior problem of computed
tomography using a priori knowledge. Inverse problems, 24, 2008. 065001 (27pp).
[2] M Defrise, F Noo, R Clackdoyle, and H Kudo. Truncated Hilbert transform and image recon-
struction from limited tomographic data. Inverse Problems, 22(3):1037 -- 1053, 2006.
[3] H W Engl, M Hanke, and A Neubauer. Regularization of Inverse Problems, volume 375.
Springer, 1996.
[4] D V Finch, 2002. Mathematisches Forschungsinstitut Oberwolfach (private conversation).
[5] F D Gakhov. Boundary Value Problems. Dover Publications, 1990.
[6] I M Gelfand and M I Graev. Crofton function and inversion formulas in real integral geometry.
Functional Analysis and its Applications, 25:1 -- 5, 1991.
[7] FA Grunbaum, L Longhi, and M Perlstadt. Differential operators commuting with finite con-
volution integral operators: some nonabelian examples. SIAM Journal on Applied Math-
ematics, 42(5):941 -- 955, 1982.
[8] A Katsevich. Singular value decomposition for the truncated Hilbert transform. Inverse Prob-
lems, 26, 2010. 115011 (12pp).
[9] A Katsevich. Singular value decomposition for the truncated Hilbert transform: part II. Inverse
Problems, 27, 2011. 075006 (7pp).
[10] E Katsevich, A Katsevich, and G Wang. Stability of the interior problem for polynomial region
of interest. Inverse Problems, 28, 2012. 065022.
[11] W Koppelman and J D Pincus. Spectral representations for finite Hilbert transformations.
Mathematische Zeitschrift, 71(1):399 -- 407, 1959.
[12] H Kudo, M Courdurier, F Noo, and M Defrise. Tiny a priori knowledge solves the interior
problem in computed tomography. Phys. Med. Biol., 53:2207 -- 2231, 2008.
[13] H J Landau and H O Pollak. Prolate spheroidal wave functions, Fourier analysis and Uncer-
tainty - II. Bell Syst. Tech. J, 40(1):65 -- 84, 1961.
[14] 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
Syst. Tech. J, 41(4):1295 -- 1336, 1962.
[15] P Maass. The interior Radon transform. SIAM J. Appl. Math., 52(3):710 -- 724, June 1992.
[16] F Natterer. The Mathematics of Computerized Tomography, volume 32. Society for Industrial
Mathematics, 2001.
[17] F Noo, R Clackdoyle, and J D Pack. A two-step Hilbert transform method for 2D image
reconstruction. Physics in Medicine and Biology, 49(17):3903 -- 3923, 2004.
[18] J D Pack, F Noo, and R Clackdoyle. Cone-beam reconstruction using the backprojection of
locally filtered projections. IEEE Transactions on Medical Imaging, 24:1 -- 16, 2005.
[19] M Reed and B Simon. Methods of Modern Mathematical Physics: Vol. 1: Functional Analysis.
Academic Press, 1972.
[20] D Slepian and H O Pollak. Prolate spheroidal wave functions, Fourier analysis and Uncertainty
- I. Bell Syst. Tech. J, 40(1):43 -- 63, 1961.
[21] G Teschl. Ordinary Differential Equations and Dynamical Systems, volume 140. American
Mathematical Society, 2012.
[22] F G Tricomi. Integral Equations, volume 5. Dover publications, 1985.
[23] Y Ye, H Yu, Y Wei, and G Wang. A general local reconstruction approach based on a truncated
Hilbert transform. International Journal of Biomedical Imaging, 2007. 63634.
[24] Y Ye, S Zhao, H Yu, and G Wang. A general exact reconstruction for cone-beam CT via
backprojection-filtration. IEEE Transactions on Medical Imaging, 24:1190 -- 1198, 2005.
[25] Y B Ye, H Y Yu, and G Wang. Exact interior reconstruction with cone-beam CT. International
Journal of Biomedical Imaging, 2007. 10693.
[26] A Zettl. Sturm-Liouville Theory, volume 121 of Mathematical Surveys and Monographs. Amer-
ican Mathematical Society, Providence, Rhode Island, 2005.
[27] T Zhuang, S Leng, B E Nett, and G-H Chen. Fan-beam and cone-beam image reconstruction
via filtering the backprojection image of differentiated projection data. Physics in Medicine
and Biology, 49(24):5489 -- 5503, 2004.
[28] Y Zou, X Pan, and E Y Sidky.
Image reconstruction in regions-of-interest from truncated
projections in a reduced fan-beam scan. Physics in Medicine and Biology, 50(1):13 -- 28,
2005.
[29] Y Zou and X C Pan.
Image reconstruction on PI-lines by use of filtered backprojection in
helical cone-beam CT. Physics in Medicine and Biology, 49:2717 -- 2731, 2004.
24
|
1804.01372 | 1 | 1804 | 2018-04-04T12:43:31 | Subsymmetric weak$^*$ Schauder bases and factorization of the identity | [
"math.FA"
] | Let $X^*$ denote a Banach space with a subsymmetric weak$^*$ Schauder basis satisfying condition~\eqref{eq:condition-c}. We show that for any operator $T : X^*\to X^*$, either $T(X^*)$ or $(I-T)(X^*)$ contains a subspace that is isomorphic to $X^*$ and complemented in $X^*$. Moreover, we prove that $\ell^p(X^*)$, $1\leq p \leq \infty$ is primary. | math.FA | math |
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND
FACTORIZATION OF THE IDENTITY
RICHARD LECHNER
Abstract. Let X∗ denote a Banach space with a subsymmetric weak∗ Schauder
basis satisfying condition (C). We show that for any operator T : X∗ → X∗,
either T (X∗) or (IdX∗ −T )(X∗) contains a subspace that is isomorphic to X∗
and complemented in X∗. Moreover, we prove that 'p(X∗), 1 ≤ p ≤ ∞ is
primary.
1. Introduction and results
In [3], Casazza and Lin showed that for any bounded linear projection Q on a
Banach space S with a subsymmetric basis, either Q(S) or (Id−Q)(S) contains
a subspace which is isomorphic to S and complemented in S. Our first main
result Theorem 1.1 extends their result to Banach spaces having a subsymmetric
weak∗ Schauder basis which satisfies the subsequent condition (C).
j=1, whose normalized associated coordinate functionals (e∗
Condition (C). Let U denote a Banach space with normalized unconditional basis
j)∞
(ej)∞
j=1 form a uncondi-
tional weak∗ Schauder basis of U∗. Given A ⊂ N, we define the bounded projection
PA : U∗ → U∗ by
(cid:16) ∞(cid:88)
j=1
(cid:17) =
∞(cid:88)
j∈A
PA
aje∗
j
aje∗
j ,
(1.1)
We say that U∗ together with its unconditional weak∗ Schauder basis (e∗
where the above series converge in the weak∗ topology of U∗.
j)∞
j=1
satisfies condition (C), if for every infinite set Λ ⊂ N and every θ > 0, we can find
a sequence (Aj)∞
j=1 of pairwise disjoint and infinite subsets of Λ, such that for all
j=1 ⊂ U∗ with kx∗
j=1 ∈ '1 with
j)∞
(x∗
k(aj)∞
j=1k'1 = 1 such that
jkU∗ ≤ 1 there exists a sequence of scalars (aj)∞
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=1
(cid:13)(cid:13)(cid:13)U∗ ≤ θ.
ajPAj x∗
j
(C)
The above series converges in the weak∗ topology of U∗.
Theorem 1.1. Let S∗ denote a Banach space with a subsymmetric weak∗ Schauder
basis satisfying condition (C). Then for any given bounded operator T : S∗ → S∗,
there exist operators M, N : S∗ → S∗ such that for H = T or H = IdS∗ −T, the
Date: July 31, 2018.
2010 Mathematics Subject Classification. 46B25,46B26.
Key words and phrases. Primary, factorization, subsymmetric, weak∗ Schauder basis.
Supported by the Austrian Science Foundation (FWF) Pr.Nr. P28352.
1
2
diagram
R. LECHNER
S∗ IdS∗ /
S∗
kMkkNk ≤ 48K7
uK4
s
(1.2)
N
/ S∗
M
S∗
H
is commutative. Consequently, H(S∗) contains a subspace which is isomorphic to
S∗ and complemented in S∗.
n Sn
We prove Theorem 1.1 in Section 4.
A related concept is the notion of a primary Banach space: a Banach space X
is primary, if for every bounded projection Q : X → X either Q(X) or (Id−Q)(X)
is isomorphic to X (see e.g. [6]). In [2], Casazza, Kottman and Lin showed that for
any Banach space S with a symmetric basis (ej)∞
j=1, the following Banach spaces
are primary:
c0-sum, and S is not isomorphic to '1;
c0-sum, and Sn = span{e1, . . . , en};
. (cid:0)(cid:80) S(cid:1), where the direct sum is either the 'p-sum for 1 < p < ∞ or the
. (cid:0)(cid:80)
(cid:1), where the direct sum is either the 'p-sum for 1 < p < ∞ or the
. (cid:0)(cid:80) '∞(cid:1)
p, 1 ≤ p < ∞.
1 and(cid:0)(cid:80) S(cid:1)
. (cid:0)(cid:80) S(cid:1)
(cid:1)
. (cid:0)(cid:80)
1 and(cid:0)(cid:80)
These results were later complemented by Capon in [1], who showed that the Banach
spaces
are primary. Theorem 1.2 below adds 'p(S∗) to that list (S∗ may be non-separable).
Theorem 1.2. Let S∗ denote a Banach space with a subsymmetric weak∗ Schauder
basis satisfying condition (C). Then for all 1 ≤ p ≤ ∞ and any given bounded
operator T : 'p(S∗) → 'p(S∗), there exist operators M, N : 'p(S∗) → 'p(S∗) such
that for H = T or H = Id'p(S∗) −T, the diagram
∞,
n Sn
n Sn
(cid:1)
∞,
'p(S∗)
M
'p(S∗)
Id'p(S∗)
H
'p(S∗)
N
/ 'p(S∗)
kMkkNk ≤ 48K7
uK4
s ,
(1.3)
is commutative. Consequently, 'p(S∗) is primary.
The proof of Theorem 1.2 is given in Section 5.
The method of proof for both our main theorems is based on the recent result [4].
2. Notation
Let X be a Banach space and let X∗ denote its dual. We say a sequence (x∗
in X∗ converges weak∗ to x∗ ∈ X∗, if
n)∞
n=1
n
x ∈ X.
hx∗
n, xi = hx∗, xi,
lim
In this case we write w*lim x∗
n = x∗.
j=1 ⊂ X∗ is a weak∗ Schauder basis for X∗, if there exists a
j)∞
We say that (e∗
j=1 for X such that the associated coordinate functionals of (ej)j = 1∞
basis (ej)∞
j , eji = 1 and he∗
j=1, i.e. he∗
j , eii = 1, i, j ∈ N, i 6= j. Thus,
n(cid:88)
are (e∗
j)∞
hx∗, ejie∗
j ,
x∗ ∈ X∗.
x∗ = w*lim
n→∞
j=1
/
/
O
O
/
/
/
O
O
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY 3
For more details we refer to [8, 6]. From now on, whenever we encounter a series
in the Banach space
. X, U or S, the mode of convergence is in the norm topology of X, U or S;
. X∗, U∗ or S∗, the mode of convergence is in the weak∗ topology of X∗, U∗
We say the weak∗ Schauder basis (e∗
constant C ≥ 1 such that for all sequences of scalars (aj)∞
j=1 of X∗ is unconditional, if there exists a
j)∞
j ∈ X∗
j=1 aje∗
j=1 with(cid:80)∞
or S∗.
holds that(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)X∗ ≤ C sup
γj(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)X∗ ,
γjaje∗
j=1 ∈ '∞.
We denote the infimum over all such constants C ≥ 1 by Ku. Note that
(γj)∞
aje∗
j∈N
j=1
j=1
j
j
(2.1)
(ej)∞
j=1 is unconditional
if and only if
j)∞
j=1 of X∗ is subsymmetric, if the fol-
j)∞
(e∗
j=1 is unconditional.
We say that a weak∗ Schauder basis (e∗
lowing two conditions are satisfied:
(a) (e∗
(b) Let (nj)∞
j)∞
j=1 is unconditional;
sure of span{e∗
j=1 ⊂ N be an increasing sequence, and let Y(nj)∞
}∞
j=1. Then the operator S(nj)∞
nj
(cid:16) ∞(cid:88)
j=1
(cid:17) =
j=1 : X∗ → Y(nj)∞
∞(cid:88)
aje∗
,
nj
aje∗
j
j=1
j=1 denote the weak∗ clo-
j=1 given by
(2.2)
S(nj)∞
j=1
is isomorphic.
Note that
(ej)∞
We define Ks by
j=1 is subsymmetric
Ks = sup(cid:8)kS(nj)∞
(cid:110)(yn)∞
j=1k,kS−1
(nj)∞
j=1
k : (nj)∞
if and only if
(e∗
j)∞
j=1 is subsymmetric.
j=1 ⊂ N is increasing(cid:9).
< ∞(cid:111)
(cid:17)1/p
kynkX
.
n=1k'p(X) =(cid:16) ∞(cid:88)
n=1
The constant Ks is finite (see [8, Chapter II, Theorem 21.2]).
For all 1 ≤ p ≤ ∞ and for any Banach space X, we define as usual 'p(X) by
n=1 : yn ∈ X and k(yn)∞
We conclude this section with two remarks on condition (C).
Remark 2.1. Any Banach space U∗ with an unconditional weak∗ Schauder basis
(e∗
j) in which '1 does not embed, satisfies condition (C).
Remark 2.2. '∞ with the standard unit vector weak∗ Schauder basis (ej)∞
condition (C).
j=1 satisfies
First, we will show that Remark 2.1 is indeed correct. To this end, assume that
condition (C) does not hold. Then there exists an infinite set Λ ⊂ N and a θ > 0,
such that for any sequence (Aj)∞
j=1 of pairwise disjoint and infinite subsets of Λ,
j=1 ∈ '1
j=1 ⊂ U∗ with kx∗
we can find (x∗
j)∞
with k(aj)∞
jkU∗ ≤ 1 such that for all scalars (aj)∞
j=1k'1 = 1 we have(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=1
(cid:13)(cid:13)(cid:13)U∗ > θ.
ajPAj x∗
j
4
R. LECHNER
If we put y∗
ality
j = PAj x∗
j and (cid:101)y∗
j = y∗
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j /ky∗
jkU∗, j ∈ N, then we obtain by uncondition-
aj(cid:101)y∗
(cid:13)(cid:13)(cid:13)U∗ > θ/Ku,
j
j=1
j )∞
j=1 is equivalent to the standard unit vector basis of '1.
To see that Remark 2.2 is valid, let Λ ⊂ N denote an infinite set, and let θ > 0.
Pick any sequence (Aj)∞
j=1 denote
any sequence of vectors in '∞ with kxjk'∞ ≤ 1, j ∈ N. Choose N ∈ N such that
N ≤ θ, and define aj = 1
1
j=1 of pairwise disjoint subsets of Λ, and let (xj)∞
N if j ≤ N, and aj = 0 whenever j > N. Then clearly
thus, ((cid:101)y∗
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=1
(cid:13)(cid:13)(cid:13)'∞ = 1
N
≤ θ.
ajPAj xj
3. Two subspace annihilation lemmata
We will present two lemmata, which are used in the proof of our first main result
Theorem 1.1 to almost diagonalize an operator T : X∗ → X∗. Although we use
Lemma 3.1 only for m = 1 here, we give the proof for general m ∈ N here, for later
reference.
Lemma 3.1. Let X denote a Banach space with a normalized basis (ej)∞
that the associated coordinate functionals (e∗
X∗. Let Λ0 ⊂ N denote an infinite set, m, n ∈ N, x∗
Then we can find a finite set F ⊂ Λ0 with cardinality F = 2m such that
j=1, such
j)∞
j=1 form a weak∗ Schauder basis of
n ∈ X∗ and η > 0.
1, . . . , x∗
where the set of signs E = E(F) is given by
max
1≤j≤n
(cid:12)(cid:12)(cid:12)(cid:68)(cid:88)
(cid:69)(cid:12)(cid:12)(cid:12) ≤ η,
E =(cid:110)(εk)k∈F : εk ∈ {±1},
(cid:88)
εkek, x∗
k∈F
j
ε ∈ E,
εk = 0(cid:111)
.
k∈F
(cid:111)
Proof. Firstly, we will prove that given x∗ ∈ X∗, there exists an infinite set Λ ⊂ Λ0
such that
hek, x∗i − hek0, x∗i ≤ η
m
,
k, k0 ∈ Λ.
(3.1)
M' =(cid:110)
To this end, we define the sets
j ∈ Λ0 : η
2m
Since hej, x∗i ≤ kx∗k < ∞, there are only finitely many ' ∈ Z for which M' is
'∈Z Ml = Λ0, at least one of the sets M', ' ∈ Z has to
be infinite. Clearly, this infinite set satisfies (3.1).
Now we will just repeatedly use (3.1). We begin by selecting Λ1 ⊂ Λ0 so that
non-empty. Hence, since(cid:83)
(' − 1) ≤ hej, x∗i <
' ∈ Z.
η
2m
'
,
hek, x∗
1i − hek0, x∗
1i ≤ η
m
,
k, k0 ∈ Λ1.
Let 1 ≤ i ≤ n− 1 and assume we have already chosen infinite sets Λ0 ⊃ Λ1 ⊃ Λ2 ⊃
··· ⊃ Λi−1 such that
max
k, k0 ∈ Λi−1.
ji − hek0, x∗
hek, x∗
,
ji ≤ η
m
1≤j≤i−1
By replacing Λ0 with the infinite set Λi−1 in (3.1), we can find an infinite set
Λi ⊂ Λi−1 such that
max
1≤j≤i
hek, x∗
ji − hek0, x∗
ji ≤ η
m
,
k, k0 ∈ Λi.
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY 5
Stopping the induction after n steps, we obtain
ji ≤ η
m
ji − hek0, x∗
(3.2)
We choose any F ⊂ Λn with F = 2m and let ε = (εk)k∈F ∈ E = E(F),
k∈F εk = 0. Then since F ⊂ Λn, we obtain from (3.2)
k, k0 ∈ Λn.
hek, x∗
max
1≤j≤n
,
i.e. εk ∈ {±1} is such that(cid:80)
(cid:12)(cid:12)(cid:12)(cid:88)
(cid:69)(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:68)(cid:88)
εkek, x∗
that
j
k∈F
ji(cid:12)(cid:12)(cid:12) ≤ η,
hek, x∗
ji − (cid:88)
k∈F
εk=−1
hek, x∗
k∈F
εk=1
1 ≤ j ≤ n.
(cid:3)
The following Lemma is an abstract version of an argument which Lindenstrauss
used in [5] (see also [6]) to show that '∞ is prime (which means that every infinite
dimensional complemented subspace of '∞ is isomorphic to '∞).
Lemma 3.2. Let U denote a Banach space with a normalized unconditional basis
j)∞
(ej)∞
j=1 form an uncondi-
tional weak∗ Schauder basis of U∗. Let Λ ⊂ N denote an infinite set, η > 0 and
ϕ ∈ U∗∗. If U∗ together with (e∗
j)∞
j=1 satisfies condition (C), then there exists an
infinite set A ⊂ Λ such that
j=1, such that the associated coordinate functionals (e∗
sup
kx∗kU∗≤1
hϕ, PAx∗i ≤ η.
Proof. Let η > 0, ϕ ∈ U∗∗ \ {0}, and assume the conclusion of the Lemma is false.
Hence, for all infinite sets A ⊂ Λ there exists x∗ with kx∗kU∗ = 1 such that
hϕ, PAx∗i > η.
(3.3)
j=1 with Aj ⊂ Λ, j ∈ N according to
jkU∗ = 1,
j ∈ U∗ with kx∗
We define θ =
condition (C). By our assumption (3.3), we can find x∗
j ∈ N, such that
and choose (Aj)∞
2kϕkU∗∗ Ku
η
ji > η.
By condition (C) we can find a sequence (aj)∞
that
hϕ, PAj x∗
j=1 ∈ '1 with k(aj)∞
(3.4)
j=1k'1 = 1 such
η
2kϕkU∗∗ Ku
.
(3.5)
But on the other hand, we obtain from (3.4) that
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=1
ajPAj x∗
j
(cid:13)(cid:13)(cid:13)U∗ ≤ θ =
(cid:69)
n(cid:88)
ajPAj x∗
j
(cid:68)
ϕ,
j=1
> η,
for a large enough n ∈ N. Combining the latter estimate with (3.5), we arrive at
the contradiction
(cid:68)
n(cid:88)
j=1
η <
ϕ,
ajPAj x∗
j
(cid:69) ≤ kϕkU∗∗
(cid:13)(cid:13)(cid:13) n(cid:88)
j=1
(cid:13)(cid:13)(cid:13)U∗ ≤ η/2.
ajPAj x∗
j
(cid:3)
4. Proof of Theorem 1.1
For convenience of the reader, we restate Theorem 1.1 here below.
Theorem 4.1 (Main result Theorem 1.1). Let S∗ denote a Banach space with a
subsymmetric weak∗ Schauder basis satisfying condition (C). Then for any given
6
bounded operator T : S∗ → S∗, there exist operators M, N : S∗ → S∗ such that for
H = T or H = IdS∗ −T, the diagram
R. LECHNER
S∗ IdS∗ /
S∗
M
S∗
N
/ S∗
H
kMkkNk ≤ 48K7
uK4
s
(4.1)
is commutative. Consequently, H(S∗) contains a subspace which is isomorphic to
S∗ and complemented in S∗.
Proof. In this proof, we use the following constants:
ηi = K−1
u 4−i−1,
i ∈ N.
(4.2)
Step 1: Inductive construction of the block basis. Let (sj)∞
j=1 denote
j)∞
the subsymmetric basis of S, and let (s∗
j=1 denote the associated coordinate
functionals, which form a weak∗ Schauder basis for S∗, such that condition (C) is
satisfied. We put A1 = N, choose B1 = {1, 2}, and we define
1 − s∗
2.
b1 = s1 − s2
b∗
1 = s∗
(4.3)
and
By Lemma 3.2, there exists an infinite collection A2 ⊂ A1 \ {1, 2}, such that
sup
kx∗kS∗≤1
hT ∗b1, PA2 x∗i ≤ η1.
(4.4)
Now assume we have already chosen infinite sets A1 ⊃ A2 ⊃ ··· ⊃ Ai, pairwise
disjoint finite sets Bj ⊂ Aj with Bj = 2, 1 ≤ j ≤ i − 1 and that we have defined
(4.5)
for all 1 ≤ j ≤ i − 1. By Lemma 3.1, we can find a set Bi ⊂ Ai with Bi = 2 so
that if we put
k0, k1 ∈ Bj, k0 < k1,
bj = sk0 − sk1
k0 − s∗
k1 ,
and b∗
j = s∗
k0, k1 ∈ Bi, k0 < k1,
bi = sk0 − sk1
we have the estimate
such that
kx∗kS∗≤1
This completes the inductive step.
The estimates (4.4), (4.7) and (4.9) imply
i−1(cid:88)
∞(cid:88)
j=1
bi, T
j=i+1
(cid:12)(cid:12)(cid:12)(cid:68)
hbi, T b∗
ji ≤ ηi,
(cid:69)(cid:12)(cid:12)(cid:12) ≤ ηi
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=i+1
ajb∗
j
(cid:13)(cid:13)(cid:13)S∗ ,
ajb∗
j
i ∈ N,
i ∈ N,
(4.10a)
(4.10b)
By Lemma 3.2, there exists an infinite set Ai+1 with
j=1
and b∗
hbi, T b∗
k0 − s∗
k1 ,
ji ≤ ηi.
i = s∗
i−1(cid:88)
Ai+1 ⊂ Ai \(cid:110)
k ∈ N : k ≤ max(cid:0) i(cid:91)
(cid:12)(cid:12)(cid:12)(cid:68)
T ∗bi, PAi+1 x∗(cid:69)(cid:12)(cid:12)(cid:12) ≤ ηi.
sup
j=1
(cid:1)(cid:111)
Bi
(4.6)
(4.7)
(4.8)
(4.9)
/
/
O
O
j=1 ⊂ R is such that(cid:80)∞
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY 7
whenever (aj)∞
Step 2: Basic operators. We define the operators B, Q : S∗ → S∗ by
j=i+1 ajb∗
j ∈ S∗.
∞(cid:88)
∞(cid:88)
j=1
j=1
Bx∗ =
Qx∗ =
hx∗, sjib∗
j ,
1
jkS∗
kb∗
jkS∗
kb∗
Bj hbj, x∗is∗
j ,
x∗ ∈ S∗,
x∗ ∈ S∗.
(4.11a)
(4.11b)
We will now show that B and Q are bounded linear operators. To this end, let
(cid:13)(cid:13)(cid:13)S∗ ,
k1(j)
jkS∗ ≥ K−1
(cid:17)
u
,
(cid:13)(cid:13)(cid:13)S∗
1
jkS∗
kb∗
ajs∗
j=1
ajs∗
k1(j)
where {k0(j), k1(j)} = Bj, j ∈ N. By unconditionality, we obtain kb∗
and
x∗ =(cid:80)∞
(cid:13)(cid:13)(cid:13)B
(cid:16) ∞(cid:88)
(cid:13)(cid:13)(cid:13)B
j=1
j=1
j=1
j=1
j=1
j
u
k0(j)
k0(j)
ajs∗
j
ajs∗
ajs∗
ajs∗
1
jkS∗
kb∗
j. Firstly, note that
j=1 ajs∗
(cid:16) ∞(cid:88)
(cid:17)(cid:13)(cid:13)(cid:13)S∗ ≤(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:17)(cid:13)(cid:13)(cid:13)S∗ ≤ K2
(cid:13)(cid:13)(cid:13)B
(cid:16) ∞(cid:88)
kQx∗kS∗ ≤(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
kb∗
jkS∗
Bj hsk0(j), x∗is∗
(cid:13)(cid:13)(cid:13)S∗ +
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:16)(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)S∗ +
(cid:17)(cid:13)(cid:13)(cid:13)S∗ ≤ 2K2
(cid:13)(cid:13)(cid:13)S∗ .
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)S∗ +
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)S∗ + Ks
j)∞
j=1 is subsymmetric, hence
ajs∗
ajs∗
kb∗
jkS∗
Bj hsk0(j), x∗is∗
uKs
k0(j)
j
j=1
j
j=1
j=1
j
j=1
obtain
Since (s∗
nated by
Ks
Since kb∗
j=1
j kS∗
Bj ≤ 1, unconditionality yields
kQx∗kS∗ ≤ 2KsKukx∗kS∗ .
One can easily verify that QB = IdS∗, i.e. the diagram
S∗
B
IdS∗
S∗
S∗
Q
kb∗
jkS∗
Bj hsk1(j), x∗is∗
j
kb∗
jkS∗
Bj hsk1(j), x∗is∗
k1(j)
j=1
(4.12)
(cid:13)(cid:13)(cid:13)S∗ .
(cid:13)(cid:13)(cid:13)S∗ .
(4.13)
(4.14)
Furthermore, by (4.8) we have that sk0(j) < sk0(j+1) and sk1(j) < sk1(j+1), j ∈ N.
The weak∗ Schauder basis (s∗
Secondly, if we write bj = sk0(j) − sk1(j) for k0, k1 ∈ Bj, k0 < k1, j ∈ N, we
j)∞
j=1 is subsymmetric, the right hand side of the latter inequality is domi-
is commutative. Consequently, B is an isomorphism onto its range and its range is
complemented by BQ.
Step 3: Conclusion of the proof. Observe that at least one of the two
following sets is infinite:
{j ∈ N : hbj, T b∗
{j ∈ N : hbj, (IdS∗ −T)b∗
ji ≥ 1}.
ji ≥ 1}
or
/
/
!
!
=
=
8
R. LECHNER
If the left set is infinite we denote it by J and put H = T, and if the left set is
finite, we denote the right set by J and we define H = IdS∗ −T. In either case, we
obtain
J is infinite
and
hbj, Hb∗
ji ≥ 1,
j ∈ J.
(4.15)
Let Y = B(S∗) and note that (4.12), (4.13) and (4.14) yields
S∗
IdS∗
B
Y
IdS∗
Now, define P : S∗ → Y by
P x∗ =(cid:88)
j∈J
S∗
QY
/ Y
kBkkQY k ≤ 4K3
uK2
s .
(4.16)
hbj, x∗i
hbj, Hb∗
ji b∗
j ,
x∗ ∈ S∗,
(4.17)
and observe that by (4.15), the unconditionality of (s∗
ditionality of (b∗
j)∞
j=1) and the definition of B and Q (see (4.11)), we obtain
j)∞
j=1 (and thus, the uncon-
Since hbj, yi ≤ kbjkSkykS∗, we obtain aj ≤ kykS∗. Hence, by (4.15) and the
crucial off diagonal estimate (4.10a), we obtain
aj
i∈J
identity is true:
kP x∗kS∗ ≤ 2KukBQx∗kS∗ ,
x∗ ∈ S∗.
Combining the latter estimate with (4.12) and (4.13) yields
x∗ ∈ S∗.
skx∗kS∗ ,
uK2
j∈J:j<i
(cid:88)
kP x∗kS∗ ≤ 8K4
A straightforward calculation shows that for all y =(cid:80)
(cid:10)bi, H(cid:80)
P Hy − y =(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
(cid:10)bi, H(cid:80)
i +(cid:88)
ji
hbi, Hb∗
i i b∗
hbi, Hb∗
(cid:13)(cid:13)(cid:13)(cid:13)S∗
hbi, Hb∗
ji
i i b∗
hbi, Hb∗
(cid:13)(cid:13)(cid:13) (cid:88)
≤ 2(cid:88)
≤ 2KukykS∗(cid:88)
j∈J:j>i ajb∗
(cid:13)(cid:13)(cid:13)(cid:13)S∗
hbi, Hb∗
i i
(cid:88)
tionality(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
≤ 2kykS∗(cid:88)
i∈J
j∈J:j>i
i∈J
j∈J:j<i
j
b∗
i
(cid:11)
i∈N
i∈J
aj
i
i∈J
ηi
ηi.
i∈N
(4.18)
j ∈ Y , the following
j∈J ajb∗
j∈J:j>i ajb∗
j
hbi, Hb∗
i i
b∗
i .
(4.19)
(cid:11)
ηi.
(4.20)
Using (4.15) and the second off diagonal estimate (4.10b), we obtain by uncondi-
(cid:13)(cid:13)(cid:13)S∗
ajb∗
j
(4.21)
u 4−i−1, i ∈ N. Thus, inserting the esti-
Recall that in (4.2) we put ηi = K−1
mates (4.20) and (4.21) into (4.19) yields
kP Hy − ykS∗ ≤ 1
(4.22)
Let J : Y → S∗ denote the operator given by Jy = y, then, by (4.22), P HJ :
2. Thus, if we define V = (P HJ)−1P, the
Y → Y is invertible and k(P HJ)−1k ≤ 3
3kykS∗ ,
y ∈ Y.
/
/
/
O
O
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY 9
diagram
Y
IdY
(P HJ)−1
Y
P HJ
J
S∗
Y
H
V
P
/ S∗
kJkkV k ≤ 12K4
uK2
s
(4.23)
is commutative. Merging the diagrams (4.16) and (4.23) concludes the proof:
S∗
B
Y
S∗
M
J
P HJ
IdS∗
IdY
Y
H
S∗
QY
Y
N
(P HJ)−1
P
V
/ S∗
kMkkNk ≤ 48K7
uK4
s . (4.24)
(cid:3)
5. Direct sums of Banach spaces with a weak∗ subsymmetric basis
We begin this section by introducing notation specific to the two parameter case.
Then we extend the subspace annihilation lemmata in Section 3 to the two param-
eter case and use them to prove our second main result Theorem 1.2.
5.1. Notation. Let X denote a Banach space with normalized basis (xj)∞
j=1, such
that their normalized associated coordinate functionals (x∗
basis of X∗. For each i ∈ N, let qi : 'p(X∗) → X∗ denote the canonical norm 1
coordinate projection given by
j)∞
j=1 form a weak∗ Schauder
(cid:0)(yn)∞
n=1
(cid:1) = yi,
yi
(yn)∞
n=1 ∈ 'p(X∗).
Let I ⊂ N and define QI : 'p(X∗) → 'p(X∗) as the natural norm 1 projection onto
the coordinates indexed by I, i.e.
n=1 = yi, i ∈ I
qiQI(yn)∞
n=1 ∈ 'p(X∗). For i ∈ N, we will sometimes write Qi instead of Q{i}.
for all (yn)∞
For each i ∈ N, we define Ii : X∗ → 'p(X∗) as the canonical isometric embedding
of X∗ into the i-th coordinate of 'p(X∗), which is given by
k ∈ N, k 6= i,
n=1 = 0, i /∈ I,
qiQI(yn)∞
qiIix∗ = x∗
qkIix∗ = 0,
i ∈ N.
and
and
We define the sequence (eij)∞
i,j=1 ⊂ 'p(X∗) by
eij = Iix∗
j ,
i, j ∈ N;
hence, the associated coordinate functionals fij : 'p(X∗) → R are given by
hfij, yi = hqiy, xji,
Given any sequence of scalars (aij)∞
'p(X∗), we define
∞(cid:88)
aijeij =(cid:16) ∞(cid:88)
i,j=1
j=1
aijx∗
j
y ∈ 'p(X∗), i, j ∈ N.
i,j=1 for which (w*limn→∞(cid:80)n
(cid:17)∞
(cid:17)∞
=(cid:16)w*lim
n(cid:88)
aijx∗
.
j
n→∞
i=1
i=1
j=1
j=1 aijx∗
i=1 ∈
j)∞
(5.1)
/
/
>
>
/
`
`
O
O
/
/
/
/
'
'
O
O
7
7
/
g
g
O
O
Y
Y
10
Thus, we have the identity
∞(cid:88)
hfij, yieij =(cid:16) ∞(cid:88)
i,j=1
j=1
R. LECHNER
(cid:17)∞
i=1
hqiy, xjix∗
j
= y,
y ∈ 'p(X∗).
For K ⊂ N2, we define the projection RK : 'p(X∗) → 'p(X∗) by
∞(cid:88)
RKy =
hfij, yi1K(i, j)eij,
where 1K(i, j) = 1 if (i, j) ∈ K, and 1K(i, j) = 0 if (i, j) /∈ K. If (x∗
tional, then RK is bounded.
j) is uncondi-
i,j=1
5.2. Two subspace annihilation lemmata for direct sums. The following
lemmata are two parameter versions of the subspace annihilation lemmata presented
in Section 3.
Lemma 5.1. Let X denote a Banach space with a normalized basis (xj)∞
j=1, such
j)∞
j=1 form a weak∗ Schauder
that their normalized associated coordinate functionals (x∗
basis of X∗. Let 1 ≤ p ≤ ∞ and let Λ0 ⊂ N denote an infinite set. Then, given
m ∈ N, y1, . . . , ym ∈ 'p(X∗), n0 ∈ N, η > 0 and M ∈ N, there exists a finite set
F ⊂ Λ0 with F = 2M such that
where the set of signs E = E(n0, F) is given by
max
1≤'≤m
(cid:12)(cid:12)(cid:10)(cid:88)
(cid:11)(cid:12)(cid:12) ≤ η,
E =(cid:110)(εn0k)k∈F : εn0k ∈ {±1},
(cid:88)
εn0kfn0k, y'
k∈F
k∈F
(5.2)
ε ∈ E,
εn0k = 0(cid:111)
.
Proof. Since the proof of Lemma 5.1 is completely analogous to that of Lemma 3.1,
(cid:3)
we will omit it.
Lemma 5.2. Let U denote a Banach space with a normalized unconditional basis
j)∞
(uj)∞
j=1 form
an unconditional weak∗ Schauder basis of U∗. Let U∗ satisfy condition (C), 1 ≤
p ≤ ∞, and let Λi ⊂ N, i ∈ N denote infinite sets. Let ϕ ∈ ('p(U∗))∗ and η > 0.
Then there exist infinite sets and Ji ⊂ Λi, i ∈ N such that
j=1, such that their normalized associated coordinate functionals (u∗
sup
kyk'p(U∗)≤1
hϕ, R{i}×Jiyi ≤ η,
i ∈ N.
Proof. If we assume Lemma 5.2 is false, then there exists an i0 ∈ N such that
sup
kyk'p(U∗)≤1
hϕ, R{i0}×Ayi > η,
(5.3)
(5.4)
and choose a sequence
j=1 of pairwise disjoint, infinite subsets of J according to (C). By (5.4), we can
for every infinite set A ⊂ Λi0. We define θ =
(Aj)∞
find a sequence (yj)∞
j=1 in 'p(U∗) with kyjk ≤ 1, j ∈ N such that
2kϕk('p(U∗))∗ Ku
η
hϕ, R{i0}×Aj yji > η,
j ∈ N.
j=1 ∈ '1 with k(aj)∞
(5.5)
j=1k'1 = 1
Next, define u∗
according to (C) such that
j = qi0yj, j ∈ N and choose (aj)∞
(cid:13)(cid:13)(cid:13) ∞(cid:88)
j=1
(cid:13)(cid:13)(cid:13)U∗ ≤ θ =
ajPAj u∗
j
η
2kϕk('p(U∗))∗ Ku
.
(5.6)
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY11
By (5.5), we obtain an integer n ∈ N such that
n(cid:88)
Observe that(cid:80)n
η < hϕ,
j=1
leads to the contradiction
j=1 ajR{i0}×Aj yj = Ii0
ajR{i0}×Aj yji ≤ kϕk('p(U∗))∗
(cid:80)n
(cid:13)(cid:13)(cid:13) n(cid:88)
η ≤ kϕk('p(U∗))∗
ajPAj uj
j=1
ajR{i0}×Aj yj
j=1
(cid:13)(cid:13)(cid:13) n(cid:88)
(cid:13)(cid:13)(cid:13)U∗ ≤ η/2.
j=1 ajPAj uj, hence, combining (5.6) and (5.7)
(cid:13)(cid:13)(cid:13)'p(U∗)
.
(5.7)
(5.8)
(cid:3)
5.3. Proof of main result Theorem 1.2. In this section we prove our second
main result Theorem 1.2, which we repeat here below (see Theorem 5.3). The
proof involves inductively constructing a block basis in each coordinate of 'p(S∗),
1 ≤ p ≤ ∞. The order by which we proceed is determined by the linear order ≺ on
N2, which is defined as follows: Let <' denote the lexicographic order on N2 and
define
(i0, j0) ≺ (i1, j1)
if and only if
(i0 + j0, i0) <' (i1 + j1, i1),
for all (i0, j0), (i1, j1) ∈ N2 (see Figure 1). Let O≺ : N2 → N denote the unique
bijective function that preserves the order ≺, i.e.
O≺((i0, j0)) < O≺((i1, j1))
if and only if
(i0, j0) ≺ (i1, j1),
for all (i0, j0), (i1, j1) ∈ N2.
Figure 1. The first 10 Elements of N2 with respect to the order ≺.
Theorem 5.3 (Main result Theorem 1.2). Let S∗ denote a Banach space with
a subsymmetric weak∗ Schauder basis satisfying condition (C). Then for all 1 ≤
p ≤ ∞ and any given bounded operator T : 'p(S∗) → 'p(S∗), there exist operators
M, N : 'p(S∗) → 'p(S∗) such that for H = T or H = Id'p(S∗) −T, the diagram
'p(S∗)Id'p(S∗)/
'p(S∗)
M
'p(S∗)
N
/ 'p(S∗)
H
kMkkNk ≤ 48K7
uK4
s ,
(5.9)
is commutative. Consequently, 'p(S∗) is primary.
(1,1)(1,2)(2,1)(1,3)(2,2)(3,1)(1,4)(2,3)(3,2)(4,1)/
/
O
O
12
R. LECHNER
Proof. For the definition of eij and fij we refer to Section 5.1. Within this proof,
we identify
ηk ↔ ηk0k1 ,
Bk ↔ Bk0k1,
bk ↔ bk0k1,
k ↔ b∗
b∗
k0k1,
whenever O≺((k0, k1)) = k. We conclude this preliminary step by defining
ηk = K−1
u 4−k−1,
k ∈ N.
(5.10)
Step 1: Inductive construction. We will now inductively (with respect to
the order ≺ on N2) construct sequences (b
)∞
i,j=1. We begin by
putting J1,i = N, i ∈ N, B1 = {(1, 1), (1, 2)} and define
i,j=1 and (b
ij )∞
(ε)
∗(ε)
ij
11 = e11 − e12
(ε)
b
and
∗(ε)
11 = f11 − f12.
b
By Lemma 5.2, there exist infinite sets J2,i ⊂ N, i ∈ N with
J2,i ⊂ J1,i \ {1, 2},
such that
sup
kyk'p(S∗)≤1
hT ∗b
∗(ε)
1
, R{i}×J2,i yi ≤ η1,
i ∈ N.
Let us assume that we have already
. selected infinite sets Jk,i ⊂ N, i ∈ N, 1 ≤ k ≤ K and finite sets Bk ⊂ N2
with Bk = 2, 1 ≤ k ≤ K − 1 such that
Bk = Bk0k1 ⊂ ({k0} × Jk,k0) \ ({k0} × Jk+1,k0),
(cid:47)(cid:110)
whenever 1 ≤ k ≤ K − 1 and O≺((k0, k1)) = k, and that
j ∈ N : j ≤ max(cid:8)j0 ∈ N : ∃i0 ∈ N, (i0, j0) ∈ k(cid:91)
Jk+1,i ⊂ Jk,i
(5.14a)
(5.14b)
(cid:9)(cid:111)
,
B'
'=1
for all i ∈ N and 1 ≤ k ≤ K − 1;
. defined b
∗(ε)
k
by
(ε)
k , b
k = ek0j − ek0j0
(ε)
b
(5.14c)
whenever 1 ≤ k ≤ K−1, O≺((k0, k1)) = k and (k0, j), (k0, j0) ∈ Bk = Bk0k1
with j < j0;
∗(ε)
k = fk0j − fk0j0,
and
b
. obtained the estimates
(5.11)
(5.12)
(5.13)
We now come to the inductive step. To this end, choose (K0, K1) ∈ N2 with
O≺((K0, K1)) = K. The set JK,K0 is infinite (see (5.14a)), hence, by Lemma 5.1
there exists a set F ⊂ JK,K0 with F = 2 such that
(cid:12)(cid:12)(cid:10)fK0j − fK0j0, T b
(cid:11)(cid:12)(cid:12) ≤ ηK,
(ε)
'
K−1(cid:88)
'=1
j, j0 ∈ F.
(5.15)
We define the collection
(5.16)
and note that (with the possible exception of BK ∩ ({K0} × JK+1,K0) = ∅) condi-
tion (5.14a) holds true for K + 1. We will take care of the possible exception by
BK = BK0K1 = {K0} × F,
k−1(cid:88)
, R{i}×Jk+1,iy(cid:11)(cid:12)(cid:12) ≤ ηk,
' i ≤ ηk,
(ε)
∗(ε)
k
hb
, T b
'=1
∗(ε)
k
(cid:12)(cid:12)(cid:10)T ∗b
sup
kyk'p(S∗)≤1
for all 1 ≤ k ≤ K − 1.
(5.14d)
(5.14e)
i ∈ N,
'=1
(ε)
K , b
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY13
appropriately choosing JK+1,i, i ∈ N at a later stage of the proof. Next, we define
the functions b
j, j0 ∈ F, j < j0,
which is in accordance with (5.14c). Combining (5.15) with (5.18) yields
∗(ε)
K and the signs εK0,j, j ∈ F by
∗(ε)
K = fK0j − fK0j0,
K = eK0j − eK0j0,
(ε)
b
K−1(cid:88)
b
(5.17)
(5.18)
(cid:11)(cid:12)(cid:12) ≤ ηK,
(cid:12)(cid:12)(cid:10)b
∗(ε)
(ε)
K , T b
'
JK+1,i ⊂ JK,i
showing that (5.14d) holds for K + 1, as required. Since all the sets Bk, 1 ≤ k ≤ K
are finite, we can apply Lemma 5.2 to find infinite sets JK+1,i, i ∈ N with
(cid:47)(cid:110)
j ∈ N : j ≤ max(cid:8)j0 ∈ N : ∃i0 ∈ N, (i0, j0) ∈ K(cid:91)
(cid:12)(cid:12)(cid:10)T ∗b
K , R{i}×JK+1,i y(cid:11)(cid:12)(cid:12) ≤ ηK,
i ∈ N,
∗(ε)
'=1
sup
kyk'p(S∗)≤1
(cid:9)(cid:111)
,
B'
(5.19)
(5.20)
such that the following estimate is satisfied:
Note that (5.19) and (5.20) imply (5.14b) and (5.14e) for K + 1. Moreover, (5.19)
also implies BK∩({K0}×JK+1,K0) = ∅, which completes the missing part of (5.14a)
for K + 1 (see (5.16)).
Altogether, we completed the inductive step, and we know that (5.14) is true for
K + 1.
Step 2: Basic operators. We define the operators B, Q : 'p(S∗) → 'p(S∗) by
∞(cid:88)
∞(cid:88)
i,j=1
i,j=1
By =
Qy =
1
hfij, yi
kb
ij k'p(S∗)
(ε)
kb
ij k'p(S∗)
(ε)
∗(ε)
Bij
ij
hb
(ε)
b
ij ,
y ∈ 'p(S∗),
, yieij,
y ∈ 'p(S∗).
(5.21a)
(5.21b)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
i,j=1
(cid:13)(cid:13)(cid:13)'p(S∗)
For the mode of convergence of the above series, we refer to (5.1).
y ∈ 'p(S∗) and observe that by (5.14c), kByk'p(S∗) is dominated by
We will now show that B and Q are bounded linear operators. To this end, let
hfij, yi
kb
1
ij k'p(S∗)
(ε)
ekij 'ij
hfij, yi
kb
1
ij k'p(S∗)
(ε)
ekij mij
where {(kij, 'ij), (kij, mij)} = Bij, i, j ∈ N. By unconditionality we obtain that
ij k'p(S∗) ≥ K−1
kb
(ε)
kByk'p(S∗) ≤ K2
u , i, j ∈ N, and subsequently
hfij, yiekij mij
hfij, yiekij 'ij
+ K2
(cid:13)(cid:13)(cid:13) ∞(cid:88)
u
u
i,j=1
Furthermore, by (5.14b) we have that 'ij < 'i(j+1) and mij < mi(j+1), i, j ∈ N.
The weak∗ Schauder basis (s∗
j)∞
j=1 of S∗ is subsymmetric, hence
hfij, yiekij j
(cid:13)(cid:13)(cid:13) ∞(cid:88)
= 2K2
uKs
hfij, yieij
i,j=1
i,j=1
+
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:13)(cid:13)(cid:13)'p(S∗)
i,j=1
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
,
(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:13)(cid:13)(cid:13)'p(S∗)
.
(5.22)
kByk'p(S∗) ≤ 2K2
uKs
uKskyk'p(S∗).
i,j=1
= 2K2
14
R. LECHNER
i,j=1
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
Similarly, we obtain that kQyk'p(S∗) is dominated by
+
(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:13)(cid:13)(cid:13)'p(S∗)
i,j=1
(cid:13)(cid:13)(cid:13) ∞(cid:88)
(cid:13)(cid:13)(cid:13) ∞(cid:88)
kb
hfkij 'ij , yieij
ij k'p(S∗)
(ε)
Bij
kb
ij k'p(S∗)
(ε)
Bij
j=1 is subsymmetric, gives us the following upper bound for kQyk'p(S∗):
j)∞
which, since (s∗
ij k'p(S∗)
(ε)
Bij
ij k'p(S∗)
(ε)
Bij
i,j=1
≤ 1, we obtain by unconditionality
ij k'p(S∗)
(ε)
Bij
i,j=1
Using kb
hfkij mij , yieimij
hfkij 'ij , yiei'ij
hfkij mij , yieij
(cid:13)(cid:13)(cid:13)'p(S∗)
+Ks
kb
kb
Ks
(cid:13)(cid:13)(cid:13)'p(S∗)
,
.
kQyk'p(S∗) ≤ 2KsKukyk'p(S∗).
One can easily verify that QB = Id'p(S∗), i.e. the diagram
'p(S∗)
Id'p(S∗)
'p(S∗)
B
Q
'p(S∗)
(5.23)
(5.24)
is commutative. Consequently, B is an isomorphism onto its range, and its range
is complemented by BQ.
Step 3: Factorization of the identity. With i ∈ N fixed, observe that at
least one of the two following sets is infinite:
Ki = {j ∈ N : hb
Thus, clearly one of the following two sets is infinite:
Li = {j ∈ N : hb
, (Id'p(S∗) −T)b
ij i ≥ 1},
(ε)
ij i ≥ 1}.
(ε)
∗(ε)
ij
∗(ε)
ij
, T b
{i ∈ N : Ki is infinite}
or
{i ∈ N : Li is infinite}
If the left set is infinite, we denote it by I and we put Ji = Ki, i ∈ I as well as
H = T; if the left set is finite, we denote the right set by I, and we define Ji = Li,
i ∈ I as well as H = Id'p(S∗) −T. In either of these two cases, we obtain that
(5.25)
I is infinite,
Let Y = B('p(S∗)) and define A : 'p(S∗) → Y , y 7→ By. Thus, by (5.22), (5.23)
and (5.24), the following diagram is commutative:
Ji is infinite, i ∈ I and hb
ij i ≥ 1, i ∈ I, j ∈ Ji.
(ε)
∗(ε)
ij
, Hb
'p(S∗)
Id'p(S∗)
A
Y
Id'p(S∗)
'p(S∗)
QY
/ Y
kAkkQY k ≤ 4K3
uK2
s .
(5.26)
Now, put K = {O≺(i, j) : i ∈ I, j ∈ Ji} and define P : 'p(S∗) → Y by
P y = (cid:88)
k∈K
∗(ε)
hb
k
∗(ε)
hb
k
, yi
k i b
(ε)
, Hb
(ε)
k ,
y ∈ 'p(S∗).
(5.27)
∗(ε)
Recall that at the beginning of the proof we identified b
k with
∗(ε)
k0k1, whenever O≺((k0, k1)) = k. Observe that by (5.25), unconditionality and the
b
definition of B and Q (see (5.21)), we obtain that
(ε)
k0k1 and b
(ε)
k with b
kP yk'p(S∗) ≤ 2KukBQyk'p(S∗)
/
/
$
$
:
:
/
/
/
O
O
SUBSYMMETRIC WEAK∗ SCHAUDER BASES AND FACTORIZATION OF THE IDENTITY15
Combining the latter estimate with (5.22) and (5.23) yields
(5.28)
uK2
skyk'p(S∗).
kP yk'p(S∗) ≤ 8K4
A straightforward calculation shows that for all y = (cid:80)
,(cid:80)
P Hy − y = (cid:88)
following identity is true:
k +(cid:88)
(cid:10)H∗b
(cid:88)
, Hb
(ε)
hb
hb
∗(ε)
k
∗(ε)
k
' i
(ε)
k i b
(ε)
a'
(ε)
b
k .
(5.29)
j kSkyk'p(S∗), we obtain aj ≤ kyk'p(S∗). Hence, by (5.25)
(ε)
'∈K:'>k a'b
, Hb
∗(ε)
k
k i
(ε)
k∈K
, Hb
'∈K:'<k
k∈K
, yi ≤ kb
Since hb
and the off-diagonal estimate (5.14d), we obtain
(ε)
j
k∈K akb
k ∈ Y , the
(ε)
(cid:11)
(ε)
'
∗(ε)
k
hb
≤ 2kyk'p(S∗)
(cid:88)
k∈N
ηk.
(5.30)
(cid:13)(cid:13)(cid:13) (cid:88)
'∈K:'>k
ηk
(cid:13)(cid:13)(cid:13)'p(S∗)
.
a'b
(ε)
'
(5.31)
k∈K
(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
(cid:10)H∗b
(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
k∈K
(cid:88)
'∈K:'<k
hb
hb
∗(ε)
k
∗(ε)
k
a'
, Hb
, Hb
∗(ε)
k
∗(ε)
hb
k
'∈K:'>k a'b
, Hb
k i
(ε)
(ε)
'
,(cid:80)
(cid:13)(cid:13)(cid:13) (cid:88)
,(cid:80)
b
(ε)
k
(ε)
k
' i
(ε)
k i b
(ε)
(cid:11)
(cid:13)(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:13)(cid:13)(cid:13)(cid:13)'p(S∗)
≤ 2(cid:88)
(cid:13)(cid:13)(cid:13)'p(S∗)
(cid:11)
(cid:13)(cid:13)(cid:13)(cid:13)'p(S∗)
≤ 2K3
≤ 4K3
(ε)
k
b
k∈K
Using (5.25), (5.14b), (5.14a) and the other off-diagonal estimate (5.14e) yields
By (5.21a), estimate (5.22) and by unconditionality, we obtain
a'b
(ε)
'
'∈K:'>k
uKskyk'p(S∗).
The latter estimate together with (5.31) yields
(cid:10)H∗b
(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
k∈K
∗(ε)
k
∗(ε)
hb
k
(ε)
'∈K:'>k a'b
'
, Hb
k i
(ε)
(cid:88)
k∈K
ηkkyk'p(S∗).
(5.32)
uKs
Recall that we put ηk = K−1
u 4−k−1, k ∈ N, (see (5.10)). Combining the esti-
mates (5.30) and (5.32) with (5.29) yields
kP Hy − yk'p(S∗) ≤ 1
3kyk'p(S∗),
y ∈ Y.
(5.33)
Let J : Y → 'p(S∗) denote the operator given by Jy = y, then P HJ : Y → Y is
invertible by (5.33). Thus, if we define V = (P HJ)−1P, the following diagram is
commutative:
Y
J
'p(S∗)
IdY
(P HJ)−1
Y
H
Y
V
P
/ 'p(S∗)
kJkkV k ≤ 12K4
uK2
s .
(5.34)
,
,
;
;
/
b
b
O
O
16
R. LECHNER
The operators J and V are both bounded, thus merging the diagrams (5.26)
and (5.34) yields the commutative diagram
'p(S∗)
Id'p(S∗)
'p(S∗)
A
Y
J
M
'p(S∗)
QY
Y
IdY
(P HJ)−1
Y
H
V
P
/ 'p(S∗)
N
kMkkNk ≤ 48K7
uK4
s .
(5.35)
Step 4: 'p(S∗) is primary. Let Q : 'p(S∗) → 'p(S∗) denote a bounded projection.
Then by (5.9), we know that 'p(S∗) is either isomorphic to a complemented sub-
space of Q('p(S∗)), or isomorphic to a complemented subspace of (Id'p(S∗) −Q)('p(S∗)).
Moreover, since 'p(S∗) is the 'p-sum of a Banach space, 'p(S∗) is isomorphic
to 'p('p(S∗)). Thus, by Pełczyński's decomposition method (see [7]; see also [9,
II.B.24]), we obtain that either Q('p(S∗)) is isomorphic to 'p(S∗), or (Id'p(S∗) −Q)('p(S∗))
(cid:3)
is isomorphic to 'p(S∗).
Acknowledgments.
It is my pleasure to thank P.F.X. Müller for many helpful discussions. Supported
by the Austrian Science Foundation (FWF) Pr.Nr. P28352.
References
[1] M. Capon. Primarité de certains espaces de Banach. Proc. London Math. Soc. (3), 45(1):113 --
130, 1982.
[2] P. G. Casazza, C. A. Kottman, and B. L. Lin. On some classes of primary Banach spaces.
Canad. J. Math., 29(4):856 -- 873, 1977.
[3] P. G. Casazza and B. L. Lin. Projections on Banach spaces with symmetric bases. Studia
[4] R. Lechner. Factorization in SL∞. Israel J. Math., to appear. Preprint available on ArXiv
Math., 52:189 -- 193, 1974.
https://arxiv.org/abs/1611.00622.
[5] J. Lindenstrauss. On complemented subspaces of m. Israel J. Math., 5:153 -- 156, 1967.
[6] J. Lindenstrauss and L. Tzafriri. Classical Banach spaces. I. Springer-Verlag, Berlin-New York,
1977. Sequence spaces, Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 92.
[7] A. Pełczyński. Projections in certain Banach spaces. Studia Math., 19:209 -- 228, 1960.
[8] I. Singer. Bases in Banach spaces. I. Springer-Verlag, New York-Berlin, 1970. Die Grundlehren
der mathematischen Wissenschaften, Band 154.
[9] P. Wojtaszczyk. Banach spaces for analysts, volume 25 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 1991.
Richard Lechner, Institute of Analysis, Johannes Kepler University Linz, Altenberger
Strasse 69, A-4040 Linz, Austria
E-mail address: [email protected]
/
/
,
,
O
O
;
;
/
b
b
O
O
Y
Y
|
1310.7626 | 1 | 1310 | 2013-10-28T21:12:03 | A new resolvent equation for the S-functional calculus | [
"math.FA"
] | The S-functional calculus is a functional calculus for $(n+1)$-tuples of non necessarily commuting operators that can be considered a higher dimensional version of the classical Riesz-Dunford functional calculus for a single operator. In this last calculus, the resolvent equation plays an important role in the proof of several results. Associated with the S-functional calculus there are two resolvent operators: the left $S_L^{-1}(s,T)$ and the right one $S_R^{-1}(s,T)$, where $s=(s_0,s_1,\ldots,s_n)\in \mathbb{R}^{n+1}$ and $T=(T_0,T_1,\ldots,T_n)$ is an $(n+1)$-tuple of non commuting operators. These two S-resolvent operators satisfy the S-resolvent equations $S_L^{-1}(s,T)s-TS_L^{-1}(s,T)=\mathcal{I}$, and $sS_R^{-1}(s,T)-S_R^{-1}(s,T)T=\mathcal{I}$, respectively, where $\mathcal{I}$ denotes the identity operator. These equations allows to prove some properties of the S-functional calculus. In this paper we prove a new resolvent equation for the S-functional calculus which is the analogue of the classical resolvent equation. It is interesting to note that the equation involves both the left and the right S-resolvent operators simultaneously. | math.FA | math |
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL
CALCULUS
DANIEL ALPAY, FABRIZIO COLOMBO, JONATHAN GANTNER, AND IRENE SABADINI
Contents
Introduction
1.
2. Preliminary results
3. The case of several non commuting operators
3.1. Some applications
4. The case of commuting operators and the quaternionic case
4.1. The case of several commuting operators
4.2. The quaternionic setting
References
1
4
6
12
20
20
22
22
L (s, T ) and the right one S−1
Abstract. The S-functional calculus is a functional calculus for (n + 1)-tuples of
non necessarily commuting operators that can be considered a higher dimensional
version of the classical Riesz-Dunford functional calculus for a single operator. In this
last calculus, the resolvent equation plays an important role in the proof of several
results. Associated with the S-functional calculus there are two resolvent operators:
the left S−1
R (s, T ), where s = (s0, s1, . . . , sn) ∈ Rn+1
and T = (T0, T1, . . . , Tn) is an (n + 1)-tuple of non commuting operators. These two
S-resolvent operators satisfy the S-resolvent equations S−1
L (s, T ) = I,
and sS−1
R (s, T )T = I, respectively, where I denotes the identity operator.
These equations allows to prove some properties of the S-functional calculus. In this
paper we prove a new resolvent equation for the S-functional calculus which is the
analogue of the classical resolvent equation. It is interesting to note that the equation
involves both the left and the right S-resolvent operators simultaneously.
L (s, T )s − T S−1
R (s, T )−S−1
1. Introduction
The S-resolvent operators are a key tool in the definition of the higher dimensional
version of the Riesz-Dunford functional calculus called S-functional calculus. This
calculus works for (n+1)-tuples (T0, T1, . . . , Tn) of non necessarily commuting operators
and is based on the so-called S-spectrum, see [14, 17]. In the case of a single operator
the S-functional calculus reduces to the Riesz-Dunford functional calculus (see [21, 32]).
When the operators (T0, T1, . . . , Tn) commute among themselves, this calculus admits a
commutative version called SC-functional calculus. In this case the S-resolvent operator
and the S-spectrum have a simpler expression, see [15].
1991 Mathematics Subject Classification. MSC: 47S10, 30G35.
Key words and phrases. n-tuples of non commuting operators, quaternionic operators, S-spectrum,
right S-resolvent operator, left S-resolvent operator, resolvent equation, projectors.
1
2
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
The class of functions on which this calculus is based is the so called set of slice
hyperholomorphic (or slice monogenic) functions which are defined on subsets of the
Euclidean space Rn+1 and have values in the Clifford algebra Rn.
For more details on the S-functional calculus and the function theory on which it is
based see the the monograph [19].
As it happens for the classical theory of monogenic functions (see [10, 18, 20, 25]),
also in the class of slice hyperholomorphic functions there is the notion of left as well
as of right hyperholomorphicity. But despite what happens in the monogenic case,
for slice hyperholomorphic functions the Cauchy formulas for left and for right slice
hyperholomorphic functions have two different kernels; moreover each of these kernels
can be written in two different ways.
The calculus admits a quaternionic version, which works for quaternionic linear op-
erators and is based on slice hyperholomorphic (or slice regular) functions defined on
subsets of the real algebra of quaternions H with values in the quaternions, see [11, 13].
To explain our new result and its consequences, let us focus, at the moment, on the
quaternionic setting which is simpler to illustrate.
Let us denote by V a two sided quaternionic Banach space and let T : V → V be a
bounded right (or left) linear operator. We recall that the S-spectrum is defined as
σS(T ) = {s ∈ H : T 2 − 2Re(s)T + s2I is not invertible},
where s = s0 + s1i + s2j + s3k is a quaternion, Re(s) = s0, s2 = s2
left and the right S-resolvent operators are defined as
0 + s2
1 + s2
2 + s2
3. The
S−1
L (s, T ) := −(T 2 − 2Re(s)T + s2I)−1(T − sI),
s ∈ H \ σS(T )
(1.1)
and
S−1
R (s, T ) := −(T − sI)(T 2 − 2Re(s)T + s2I)−1,
s ∈ H \ σS(T ),
(1.2)
respectively. The left S-resolvent operator satisfies the equation
S−1
L (s, T )s − T S−1
L (s, T ) = I,
s ∈ H \ σS(T ),
and the right S-resolvent operator satisfies
sS−1
R (s, T ) − S−1
R (s, T )T = I,
s ∈ H \ σS(T ).
(1.3)
(1.4)
Consider the complex plane CI := R + I R, for I ∈ S, where S is the unit sphere of
purely imaginary quaternions. Observe that CI can be identified with a complex plane
since I 2 = −1 for every I ∈ S. Let U ⊂ H be a suitable domain that contains the
S-spectrum of T . We define for left slice hyperholomorphic functions f : U → H (see
the precise definition in the sequel) the quaternionic functional calculus as
f (T ) =
1
2π Z∂(U ∩CI )
S−1
L (s, T ) dsI f (s),
(1.5)
where dsI = −dsI, and for right slice hyperholomorphic functions f : U → H, we
define
f (T ) =
1
2π Z∂(U ∩CI )
f (s) dsI S−1
R (s, T ).
(1.6)
These definitions are well posed since the integrals do not depend neither on the open
set U and nor on the complex plane CI and can be extended to the case of (n+1)-tuples
of operators, using slice hyperholomorphic functions with values in a Clifford algebra.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
3
Using a similar notion of hyperholomorphicity and the S-spectrum in [23] the authors
introduce the continuous functional calculus in a quaternionic Hilbert space.
The S-resolvent equations (1.3), (1.4) are useful to prove several properties of the S-
functional calculus. However it is natural to ask if it is possible to obtain an analog of
the classical resolvent equation
(λI − G)−1(µI − G)−1 =
(λI − G)−1 − (µI − G)−1
µ − λ
, λ, µ ∈ C \ σ(G),
(1.7)
where G is a complex operator on a Banach space, which might be useful to prove
other properties of the calculus. The main goal of this paper it to show that (1.7) can
be generalized in this non commutative setting, but it involves both the left and the
right S-resolvent operators. Precisely, we will show that
S−1
R (s, T )S−1
L (p, T )]p
L (p, T )]](p2 − 2s0p + s2)−1,
L (p, T ) = [[S−1
− s[S−1
R (s, T ) − S−1
R (s, T ) − S−1
for s, p ∈ H \ σS(T ).
It is also worthwhile to mention that the S-resolvent operator plays an important role in
the definition of the quaternionic version of the counterpart of the operator (I − zA)−1
in the realization s(z) = D + zC(I − zA)−1B for Schur multipliers, see [2]. The reader
is referred to [2, 3, 4] for Schur analysis in the slice hyperholomorphic setting and to
[1] and [7] for an overview of Schur analysis in the complex setting.
It is interesting to note that in literature there are other cases in which the authors
consider two resolvent operators. We mention in particular the case of Schur analysis
in the setting of upper triangular operators and in the setting of compact Riemann sur-
faces. In the first case, the role of complex numbers is played by diagonal operators and
there are two "point evaluations" of an operator at a diagonal, one left and one right,
each corresponding to an associated resolvent operator; see [8, (2.4)-(2.6), p. 256], but
the resolvent equation is related just with one resolvent at a time; see [8, Corollary
2.9, p 266]. In the setting of compact Riemann surfaces (see [35, 29] for the general
setting) there is a resolvent operator associated to every meromorphic function on the
given Riemann surface X (see [9, (4.1), p. 307], and one needs two such operators,
associated to a pair of functions which generate the field of meromorphic functions on
X, to study underlying spaces; see [9, §5]. The same resolvent equation is satisfied by
all the resolvent operators; see [9, Theorem 4.2, p. 309].
In this setting both S-resolvent operators enter the resolvent equation.
The plan of the paper is as follows.
In Section 2 we recall some preliminary results on slice hyperholomorphic functions.
In Section 3 we state and prove the new resolvent equation and we show that there are
two possible versions which are equivalent. We prove our results for the S-functional
calculus for (n + 1)-tuples of non commuting operators and we show some applications
of the resolvent equation.
In Section 4 we consider the commutative version of the S-functional calculus, the so
called SC-functional calculus and we reformulate our main results for the quaternionic
functional calculus. Since the proofs follow the lines of the ones for the case of (n + 1)-
tuples of non commuting operators we will omit them in both cases.
4
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
2. Preliminary results
In this section we recall the notion of slice hyperholomorphic functions and their Cauchy
formulas, see [19].
Let Rn be the real Clifford algebra over n imaginary units e1, . . . , en satisfying the
relations eiej + ejei = 0, i 6= j, e2
i = −1. An element in the Clifford algebra will
i1 < . . . < ir is a
multi-index and eA = ei1ei2 . . . eir , e∅ = 1. An element (x0, x1, . . . , xn) ∈ Rn+1 will be
j=1 xjej ∈ Rn called paravector and
the real part x0 of x will also be denoted by Re(x). The norm of x ∈ Rn+1 is defined as
x2 = x2
j=1 xjej.
Let
be denoted by PA eAxA where A = {i1 . . . ir} ∈ P{1, 2, . . . , n},
identified with the element x = x0 + x = x0 + Pn
n. The conjugate of x is defined by ¯x = x0 − x = x0 −Pn
1 + . . . + x2
0 + x2
S = {x = e1x1 + . . . + enxn x2
1 + . . . + x2
n = 1};
for I ∈ S we obviously have I 2 = −1. Given an element x = x0 + x ∈ Rn+1 let us set
Ix = x/x if x 6= 0, and given an element x ∈ Rn+1, the set
[x] := {y ∈ Rn+1 : y = x0 + Ix, I ∈ S}
is an (n − 1)-dimensional sphere in Rn+1. The vector space R + I R passing through
1 and I ∈ S will be denoted by CI and an element belonging to CI will be indicated
by u + Iv, for u, v ∈ R. With an abuse of notation we will write x ∈ Rn+1. Thus,
if U ⊆ Rn+1 is an open set, a function f : U ⊆ Rn+1 → Rn can be interpreted as a
function of the paravector x.
Definition 2.1 (Slice hyperholomorphic functions). Let U ⊆ Rn+1 be an open set and
let f : U → Rn be a real differentiable function. Let I ∈ S and let fI be the restriction
of f to the complex plane CI.
The function f is said to be left slice hyperholomorphic (or slice monogenic) if, for
every I ∈ S, on U ∩ CI it satisfies
1
2 (cid:18) ∂
∂u
fI(u + Iv) + I
∂
∂v
fI(u + Iv)(cid:19) = 0.
We will denote by SM(U) the set of left slice hyperholomorhic functions on the open
set U or by SML(U) when confusion may arise.
The function f is said to be right slice hyperholomorphic (or right slice monogenic) if,
for every I ∈ S, on U ∩ CI, it satisfies
1
2 (cid:18) ∂
∂u
fI(u + Iv) +
∂
∂v
fI(u + Iv)I(cid:19) = 0.
We will denote by SMR(U) the set of right slice hyperholomorphic functions on the
open set U.
Slice hyperholomorphic functions possess good properties when they are defined on
suitable domains which are introduced in the following definition.
Definition 2.2 (Axially symmetric slice domain). Let U be a domain in Rn+1. We
say that U is a slice domain (s-domain for short) if U ∩ R is non empty and if U ∩ CI
is a domain in CI for all I ∈ S. We say that U is axially symmetric if, for all x ∈ U,
the (n − 1)-sphere [x] is contained in U.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
5
Definition 2.3 (Cauchy kernel for left slice hyperholomorphic functions). Let x, s ∈
Rn+1 be such that x 6∈ [s]. Let S−1
L (s, x) be the function defined by
S−1
L (s, x) := −(x2 − 2xRe[s] + s2)−1(x − s).
(2.8)
We say that S−1
written in form I.
L (s, x) is the Cauchy kernel (for left slice hyperholomorphic functions)
Proposition 2.4. Suppose that x and s ∈ Rn+1 are such that x 6∈ [s]. The following
identity holds:
− (x − ¯s)(x2 − 2Re(s)x + s2)−1 = (s2 − 2Re(x)s + x2)−1(s − ¯x).
(2.9)
Remark 2.5. By Proposition 2.4 S−1
L (s, x) can also be written as
S−1
L (s, x) := (s − ¯x)(s2 − 2Re(x)s + x2)−1.
(2.10)
In this case, we will say S−1
L (s, x) is written in form II.
Proposition 2.6. The function S−1
x and right slice hyperholomorphic in the variable s for x 6∈ [s].
L (s, x) is left slice hyperholomorphic in the variable
The case of the Cauchy kernel for right slice hyperholomorphic functions is similar.
Definition 2.7 (Cauchy kernel for right slice hyperholomorphic functions). Let x, s ∈
Rn+1 be such that x 6∈ [s]. The Cauchy kernel S−1
R (s, x) for right slice hyperholomorphic
functions is defined by
S−1
R (s, x) := −(x − ¯s)(x2 − 2Re(s)x + s2)−1.
(2.11)
We say that S−1
R (s, x) is written in form I.
Remark 2.8. An analog of Proposition 2.4 holds in fact:
− (x − ¯s)(x2 − 2Re(s)x + s2)−1 = (s2 − 2Re(x)s + x2)−1(s − ¯x),
(2.12)
for x, s ∈ Rn+1 such that x 6∈ [s].
Thus S−1
R (s, x) can be written as
S−1
R (s, x) = (s2 − 2Re(x)s + x2)−1(s − ¯x),
and in this case we say that S−1
R (s, x) is written in form II.
Theorem 2.9 (The Cauchy formula with slice hyperholomorphic kernel). Let U ⊂
Rn+1 be an axially symmetric s-domain. Suppose that ∂(U ∩ CI) is a finite union of
continuously differentiable Jordan curves for every I ∈ S. Set dsI = −dsI for I ∈ S.
• If f is a (left) slice hyperholomorphic function on a set that contains U then
f (x) =
1
2π Z∂(U ∩CI )
S−1
L (s, x)dsIf (s)
(2.13)
and the integral does not depend on U and on the imaginary unit I ∈ S.
• If f is a right slice hyperholomorphic function on a set that contains U , then
f (x) =
1
2π Z∂(U ∩CI )
f (s)dsIS−1
R (s, x)
(2.14)
and the integral does not depend on U and on the imaginary unit I ∈ S.
6
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
The above Cauchy formulas are the starting point to define the S-functional calculus.
A crucial fact of slice hyperholomorphic functions in the representation formula (also
called structure formula). This formula will be used in the sequel to give applications
of the new resolvent equation.
Theorem 2.10 (Representation Formula). Let U be an axially symmetric s-domain
U ⊆ H.
• Let f be a (left) slice hyperholomorphic function on U. Choose any J ∈ S.
Then the following equality holds for all x = u + Iv ∈ U:
f (u + Iv) =
1
2hf (u + Jv) + f (u − Jv)i + I
2hJ[f (u − Jv) − f (u + Jv)]i.
1
(2.15)
Moreover, for all u, v ∈ R such that u + vS ⊆ U, there exist Rn-valued functions
α, β depending on u, v only such that for all K ∈ S
1
2hf (u + Kv) + f (u − Kv)i = α(u, v) and
2hK[f (u − Kv) − f (u + Kv)]i = β(u, v).
(2.16)
• Let f be a right slice hyperholomorphic function on U. Choose any J ∈ S. Then
1
1
the following equality holds for all x = u + Iv ∈ U:
f (u + Iv) =
1
2hf (u + Jv) + f (u − Jv)i +
2h[f (u − Jv) − f (u + Jv)]JiI.
(2.17)
Moreover, for all u, v ∈ R such that u + vS ⊆ U, there exist Rn-valued functions
α, β depending on u, v only such that for all K ∈ S
1
2hf (u + Kv) + f (u − Kv)i = α(u, v) and
1
2h[f (u − Kv) − f (u + Kv)]Ki = β(u, v).
(2.18)
3. The case of several non commuting operators
In the sequel, we will consider a Banach space V over R with norm k·k. It is possible to
endow V with an operation of multiplication by elements of Rn which gives a two-sided
module over Rn. A two-sided module V over Rn is called a Banach module over Rn,
if there exists a constant C ≥ 1 such that kvak ≤ Ckvka and kavk ≤ Cakvk for all
v ∈ V and a ∈ Rn. By Vn we denote V ⊗ Rn over Rn; Vn turns out to be a two-sided
Banach module .
i1 < . . . < ir is a multi-index). The multiplications of an element v ∈ Vn with a scalar
V .
We denote by B(V ) the space of bounded R-homomorphisms of the Banach space V
to itself endowed with the natural norm denoted by k · kB(V ). Given TA ∈ B(V ), we
An element in Vn is of the type PA vA ⊗ eA (where A = i1 . . . ir, iℓ ∈ {1, 2, . . . , n},
a ∈ Rn are defined by va = PA vA ⊗ (eAa) and av = PA vA ⊗ (aeA). For simplicity,
we will write PA vAeA instead of PA vA ⊗ eA. Finally, we define kvk2
Vn = PA kvAk2
can introduce the operator T = PA TAeA and its action on v = P vBeB ∈ Vn as
T (v) = PA,B TA(vB)eAeB. The operator T is a right-module homomorphism which is
In the sequel, we will consider operators of the form T = T0 + Pn
j=1 ejTj where Tj ∈
B(V ) for j = 0, 1, . . . , n. The subset of such operators in B(Vn) will be denoted by
B0,1(Vn). We define kT kB0,1(Vn) = Pj kTjkB(V ). Note that, in the sequel, we will omit
the subscript B0,1(Vn) in the norm of an operator. Note also that kT Sk ≤ kT kkSk.
a bounded linear map on Vn.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
7
Definition 3.1. Let T ∈ B0,1(Vn). We define the left Cauchy kernel operator series or
S-resolvent operator series as
S−1
L (s, T ) = Xn≥0
T ns−1−n,
and the right Cauchy kernel operator series as
S−1
R (s, T ) = Xn≥0
s−1−nT n,
for kT k < s.
(3.19)
(3.20)
The Cauchy kernel operator series are the power series expansion of the S-resolvent
operators. Their sum is computed in the following result:
Theorem 3.2. Let T ∈ B0,1(Vn) and let s ∈ H. Then, for kT k < s, we have
T ms−1−m = −(T 2 − 2Re(s)T + s2I)−1(T − sI),
s−1−mT m = −(T − sI)(T 2 − 2Re(s)T + s2I)−1.
(3.21)
(3.22)
Xm≥0
Xm≥0
We observe that the sum of the above series are independent of the fact that the
components of the paravector operator T commute. Moreover the operators on right
hand sides of (3.21) and (3.22) are defined on a subset of Rn+1 that is larger then
{s ∈ Rn+1 : kT k < s}. This fact suggests the definition of S-spectrum, of S-resolvent
set and of S-resolvent operators.
Definition 3.3 (The S-spectrum and the S-resolvent set). Let T ∈ B0,1(Vn). We
define the S-spectrum σS(T ) of T as:
σS(T ) = {s ∈ Rn+1
: T 2 − 2 Re(s)T + s2I
is not invertible}.
The S-resolvent set ρS(T ) is defined by
ρS(T ) = Rn+1 \ σS(T ).
Definition 3.4 (The S-resolvent operators). Let T ∈ B0,1(Vn) and s ∈ ρS(T ). We
define the left S-resolvent operator as
S−1
L (s, T ) := −(T 2 − 2Re(s)T + s2I)−1(T − sI),
and the right S-resolvent operator as
S−1
R (s, T ) := −(T − sI)(T 2 − 2Re(s)T + s2I)−1.
(3.23)
(3.24)
The operators S−1
L (s, T ) and S−1
R (s, T ) satisfy the equations below, see [19]:
Theorem 3.5. Let T ∈ B0,1(Vn) and let s ∈ ρS(T ). Then, the left S-resolvent operator
satisfies the equation
and the right S-resolvent operator satisfies the equation
S−1
L (s, T )s − T S−1
L (s, T ) = I,
sS−1
R (s, T ) − S−1
R (s, T )T = I.
(3.25)
(3.26)
8
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
Our goal is to establish the analogue of the classical resolvent equation. To this end,
we need some preliminary results. A crucial fact is the following Theorem 3.6 that
will give us the hint to discover what is the structure of the resolvent equation in this
non commutative setting at least in the case the S-resolvent operators are expressed in
power series.
Theorem 3.6. Let A, B ∈ B(Vn) and let s, p ∈ Rn+1. Then, for p < s, we have
and
Xm≥0
Xm≥0
pmAs−1−m = −(p2 − 2Re(s)p + s2)−1(pA − As),
(3.27)
s−1−mBpm = −(Bp − sB)(p2 − 2Re(s)p + s2)−1.
(3.28)
Moreover, (3.28) can be written as
s−1−mBpm = (s2 − 2Re(p)s + p2)−1(sB − Bp).
(3.29)
Xm≥0
Proof. To verify (3.27) define
X := (p2 − 2Re(s)p + s2)Xm≥0
pmAs−1−m
and observe that
X = Xm≥0
(p2 − 2Re(s)p + s2)pmAs−1−m
= p2As−1 − 2Re(s)pAs−1 + s2As−1
+ p3As−2 − 2Re(s)p2As−2 + s2pAs−2
+ p4As−3 − 2Re(s)p3As−3 + s2p2As−3 + . . .
= −(pA − As) + Xm≥2
pmA(s2 − 2Re(s)s + s2)s−1−m.
(3.30)
Since any paravector s satisfies
s2 − 2Re(s)s + s2 = 0
we deduce that
X = (p2 − 2Re(s)p + s2)Xm≥0
pmAs−1−m = −(pA − As)
and the statement follows. The equality in (3.28) can be verified by setting
Y := Xm≥0
s−1−mBpm(p2 − 2Re(s)p + s2)
and observing that
Y = −(Bp − sB) + Xm≥0
s−1−mBpm(p2 − 2Re(p)p + p2) = −(Bp − sB).
With similar computations one can verify equality (3.29).
(cid:3)
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
9
Corollary 3.7. Let A, B ∈ B(Vn) and let s, p be paravectors. Then, for p < s, the
following equations hold
m
Xj=0
m
Xj=0
and
pjAs−1−j = −(p2 − 2Re(s)p + s2)−1(pA − As)
+ pm+1(p2 − 2Re(s)p + s2)−1(pA − As)s−1−m,
s−1−jBpj = −(Bp − sB)(p2 − 2Re(s)p + s2)−1
+ s−1−m(Bp − sB)(p2 − 2Re(s)p + s2)−1pm+1.
Moreover, (3.32) can also written as
m
Xj=0
s−1−jBpj = (s2 − 2Re(p)s + p2)−1(sB − Bp)
− s−1−m(s2 − 2Re(p)s + p2)−1(sB − Bp)pm+1.
(3.31)
(3.32)
(3.33)
Proof. Identity (3.31) follows from
m
Xj=0
pjAs−1−j =
∞
Xj=0
pjAs−1−j −
∞
Xj=m+1
pjAs−1−j,
that can be written as
m
Xj=0
pjAs−1−j =
∞
Xj=0
pjAs−1−j − pm+1(cid:16)
∞
Xj=0
pjAs−1−j(cid:17)s−1−m,
but now we use (3.27) to get the result. Identity (3.32) and (3.33) follow with similar
computations.
(cid:3)
We now prove the new S-resolvent equation. In the proof we first consider the case in
which the S-resolvent operators admit the power series expansion
S−1
L (s, T ) = Xm≥0
T ms−1−m,
S−1
R (s, T ) = Xm≥0
s−1−mT m,
that is for kT k < s. Then we verify that such equation holds in general.
Theorem 3.8. Let T ∈ B0,1(Vn) and let s and p ∈ ρS(T ). Then we have
R (s, T )S−1
S−1
L (p, T ))p−s(S−1
R (s, T )−S−1
L (p, T ) = ((S−1
R (s, T )−S−1
L (p, T )))(p2 −2s0p+s2)−1.
(3.34)
Moreover, the resolvent equation can also be written as
R (s, T )−S−1
R (s, T )S−1
S−1
L (p, T ) = (s2−2p0s+p2)−1(s(S−1
L (p, T ))−(S−1
R (s, T )−S−1
L (p, T ))p).
(3.35)
Proof. We prove the theorem in two steps.
STEP I. First we assume that the S-resolvent operators are expressed in power series.
If kT k < p < s then the S-resolvent operators have power series expansion and so
S−1
R (s, T )S−1
L (p, T ) = (Xj≥0
s−1−jT j)(Xj≥0
T jp−1−j).
(3.36)
10
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
By setting
(3.36) can be written as
Λm(s, p; T ) :=
m
Xj=0
s−1−j(T mp−1−m)pj
S−1
R (s, T )S−1
L (p, T ) = Xm≥0
Λm(s, p; T ).
Formula (3.32) with B = T mp−1−m and some computations give
Λm(s, p; T ) = −((T mp−1−m)p − s(T mp−1−m))(p2 − 2Re(s)p + s2)−1
+ s−1−m((T mp−1−m)p − s(T mp−1−m))(p2 − 2Re(s)p + s2)−1pm+1
= −[(T mp−1−m)p − s(T mp−1−m)
+ (s−1−mT m)p − s(s−1−mT m)](p2 − 2Re(s)p + s2)−1.
(3.37)
From the chain of equalities
S−1
R (s, T )S−1
L (p, T ) = Xm≥0
Λm(s, p; T )
= −[(Xm≥0
+ (Xm≥0
(T mp−1−m)p − sXm≥0
s−1−mT m)p − sXm≥0
(T mp−1−m))]
s−1−mT m)](p2 − 2Re(s)p + s2)−1
(3.34) follows.
To prove that the resolvent equation can be written in the second form (3.35) observe
that Λm(s, p; T ) can also be written using (3.29) as
Λm(s, p; T ) = (s2 − 2Re(p)s + p2)−1(s(T mp−1−m) − (T mp−1−m)p)
− s−1−m(s2 − 2Re(p)s + p2)−1(s(T mp−1−m) − (T mp−1−m)p)pm+1.
(3.39)
(3.38)
L (p, T ) defined in (3.23) and (3.24), respectively.
so taking the sum Pm≥0 Λm(s, p; T ) we get the second version of the resolvent equation.
STEP II. We prove that, for s and p ∈ ρS(T ), (3.34) and (3.35) hold with S−1
and S−1
Let us verify (3.34). Since s and p ∈ ρS(T ) the left and right S-resolvent operators
defined by (3.23) and (3.24) satisfy the left and the right resolvent equations (3.25)
and (3.26), respectively. To verify (3.34) we have to show that S−1
L (p, T )(p2 −
2s0p + s2) equals
R (s, T )S−1
R (s, T )
(S−1
R (s, T ) − S−1
L (p, T ))p − s(S−1
R (s, T ) − S−1
L (p, T )).
To do this we use the left and the right S-resolvent equations (3.25), (3.26). Indeed,
using the left S-resolvent equation, written as
S−1
L (p, T )p = T S−1
L (p, T ) + I,
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
11
we have
L (p, T )(p2 − 2s0p + s2) = S−1
R (s, T )[S−1
L (p, T )p]p
S−1
R (s, T )S−1
− 2s0S−1
= S−1
+ s2S−1
R (s, T )S−1
R (s, T )[T S−1
R (s, T )S−1
L (p, T )p + s2S−1
L (p, T ) + I]p − 2s0S−1
L (p, T )
R (s, T )S−1
L (p, T )
R (s, T )[T S−1
L (p, T ) + I]
(3.40)
and using again the left S-resolvent equation
L (p, T )(p2 − 2s0p + s2) = S−1
L (p, T ) + I] + s2S−1
S−1
R (s, T )S−1
− 2s0S−1
R (s, T )[T S−1
R (s, T )T [T S−1
R (s, T )S−1
L (p, T )
L (p, T ) + I] + S−1
R (s, T )p
we obtain
S−1
R (s, T )S−1
R (s, T )T ]T S−1
L (p, T )(p2 − 2s0p + s2)
= [S−1
− 2s0[[S−1
+ s2S−1
R (s, T )T ]S−1
R (s, T )S−1
L (p, T ).
Now we use the right S-resolvent equation
L (p, T ) + S−1
R (s, T )T ] + S−1
R (s, T )p
L (p, T ) + S−1
R (s, T )]
(3.41)
(3.42)
(3.43)
(3.44)
(3.45)
we obtain
S−1
R (s, T )T = sS−1
R (s, T ) − I
S−1
R (s, T )S−1
L (p, T )(p2 − 2s0p + s2)
= [[sS−1
R (s, T ) − I]T ]S−1
− 2s0[[sS−1
R (s, T ) − I]S−1
+ s2S−1
L (p, T ).
R (s, T )S−1
L (p, T ) + sS−1
L (p, T ) + S−1
R (s, T ) − I] + S−1
R (s, T )]
R (s, T )p
Iterating the use of the above right S-resolvent equation we get
S−1
R (s, T )S−1
R (s, T ) − I] − T ]S−1
L (p, T )(p2 − 2s0p + s2)
= [s[sS−1
− 2s0[[sS−1
+ s2S−1
R (s, T )S−1
R (s, T )S−1
L (p, T ),
L (p, T ) + sS−1
R (s, T ) − I] + S−1
R (s, T )p
L (p, T ) − S−1
L (p, T )] + S−1
R (s, T )]
which leads to
S−1
R (s, T )S−1
L (p, T )(p2 − 2s0p + s2)
= (s2 − 2s0s + s2)S−1
+ [S−1
R (s, T ) − S−1
R (s, T )S−1
L (p, T )
L (p, T )]p − s[S−1
R (s, T ) − S−1
L (p, T )],
and since s2 − 2s0s + s2 = 0 we obtain (3.34). With similar computations we can
show that also (3.35) holds.
(cid:3)
We now observe that in the commutative case besides the resolvent equation, also the
following relation between the resolvent operators
(λI − G)−1(µI − G)−1 = (µI − G)−1(λI − G)−1,
for λ, µ ∈ ρ(G)
12
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
holds. In the non commutative case we cannot aspect the validity of such a relation,
however we will show that an analogous equation holds for the so-called pseudo S-
resolvent operators defined below.
Definition 3.9. Let T ∈ B0,1(Vn). We define, for s ∈ ρS(T ), the pseudo S-resolvent
operator of T is defined as
Qs(T ) := (T 2 − 2Re(s)T + s2I)−1.
With the above definition the resolvents S−1
L (s, T ) and S−1
R (s, T ) become
and
S−1
L (s, T ) := −Qs(T )(T − sI),
s ∈ ρS(T ),
S−1
R (s, T ) := −(T − sI)Qs(T ),
s ∈ ρS(T ).
We now prove the following:
Theorem 3.10. Let T ∈ B0,1(Vn) and let s, p ∈ ρS(T ). Then we have
(T − sI)Qs(T )Qp(T )(T − pI) = (T − sI)Qp(T )Qs(T )(T − pI).
Proof. It follows from the fact that
(T 2 − 2Re(s)T + s2I)(T 2 − 2Re(p)T + p2I)
= (T 2 − 2Re(p)T + p2I)(T 2 − 2Re(s)T + s2I).
Since s, p ∈ ρS(T ) we can take the inverse and the statement follows.
Remark 3.11. Observe that the function FT (s, p) defined by
(3.46)
(3.47)
(3.48)
(cid:3)
FT (s, p) := S−1
R (s, T )S−1
L (p, T )
is left slice hyperholomorphic in s and it is right slice hyperholomorphic in p with values
in B(Vn). The function
GT (s, p) := S−1
L (p, T )S−1
R (s, T )
is not slice hyperholomorphic neither in p nor in s.
Remark 3.12. Using the star products left and right in the variables s, p, which will
be denoted by ⋆s,lef t, ⋆p,right respectively, see [6], the resolvent equation (3.34) can be
written as
S−1
R (s, T )S−1
L (p, T ) = [S−1
R (s, T ) − S−1
L (p, T )] ⋆s,lef t (p − s)(p2 − 2Re(s)p + s2)−1I,
or
S−1
R (s, T )S−1
L (p, T ) = (s − p)(s2 − 2Re(p)s + p2)−1I ⋆p,right [S−1
R (s, T ) − S−1
L (p, T )].
3.1. Some applications. Here we recall the formulations of the S-functional calculus
and the we use the resolvent equation to deduce some results.
We first recall two important properties of the S-spectrum.
Theorem 3.13 (Structure of the S-spectrum). Let T ∈ B0,1(Vn) and suppose that
p = p0 + p belongs σS(T ) with p 6= 0. Then all the elements of the (n − 1)-sphere [p]
belong to σS(T ).
This result implies that if p ∈ σS(T ) then either p is a real point or the whole (n − 1)-
sphere [p] belongs to σS(T ).
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
13
Theorem 3.14 (Compactness of S-spectrum). Let T ∈ B0,1(Vn). Then the S-spectrum
σS(T ) is a compact nonempty set. Moreover, σS(T ) is contained in {s ∈ Rn+1 : s ≤
kT k }.
Definition 3.15. Let Vn be a two sided Banach module, T ∈ B0,1(Vn)and let U ⊂
Rn+1 be an axially symmetric s-domain that contains the S-spectrum σS(T ) such that
∂(U ∩ CI) is the union of a finite number of continuously differentiable Jordan curves
for every I ∈ S. In this case we say that U is a T -admissible open set.
We can now introduce the class of functions for which we can define the two versions
of the S-functional calculus.
Definition 3.16. Let Vn be a two sided Banach module, T ∈ B0,1(Vn) and let W be an
open set in Rn+1.
(i) A function f ∈ SML(W ) is said to be locally left hyperholomorphic on σS(T )
if there exists a T -admissible domain U ⊂ Rn+1 such that U ⊂ W , on which f
is left slice hyperholomorphic. We will denote by SML
σS (T ) the set of locally left
hyperholomorphic functions on σS(T ).
(ii) A function f ∈ SMR(W ) is said to be locally right regular on σS(T ) if there
exists a T -admissible domain U ⊂ Rn+1 such that U ⊂ W , on which f is right
slice hyperholomorphic. We will denote by SMR
σS(T ) the set of locally right slice
hyperholomorphic functions on σS(T ).
Definition 3.17 (The S-functional calculus). Let Vn be a two sided Banach module
and T ∈ B0,1(Vn). Let U ⊂ Rn+1 be a T -admissible domain and set dsI = −dsI. We
define
and
f (T ) =
1
2π Z∂(U ∩CI )
f (T ) =
1
2π Z∂(U ∩CI )
S−1
L (s, T ) dsI f (s),
for f ∈ SML
σS (T ),
(3.49)
f (s) dsI S−1
R (s, T ),
for f ∈ SMR
σS (T ).
(3.50)
We now define the Riesz projectors for the S-functional calculus. We begin with a
preliminary lemma.
Lemma 3.18. Let B ∈ B(Vn) and let G be an axially symmetric s-domain such that
p ∈ G. Then
(sB − Bp)(p2 − 2s0p + s2)−1 = (s2 − 2p0s + p2)−1(sB − Bp),
p 6∈ [s],
(3.51)
and
1
2π Z∂(G∩CI )
dsI(sB − Bp)(p2 − 2s0p + s2)−1 = B.
(3.52)
14
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
Proof. Formula (3.51) is obtained by direct computation. Let us prove (3.52), so we
write
1
2π Z∂(G∩CI )
dsI(sB − Bp)(p2 − 2s0p + s2)−1
=
=
+
1
1
1
2π Z∂(G∩CI )
2π Z∂(G∩CI )
2π Z∂(G∩CI )
dsI(s2 − 2p0s + p2)−1(sB − Bp)
dsI(s2 − 2p0s + p2)−1(s − p)B
dsI(s2 − 2p0s + p2)−1(pB − Bp)
but observe that
1
2π Z∂(G∩CI )
dsI(s2 − 2p0s + p2)−1(s − p)B =
1
2π Z∂(G∩CI )
dsIS−1
R (s, p)B = B
and moreover by the residue theorem it is
1
2π Z∂(G∩CI )
dsI(s2 − 2p0s + p2)−1 = 0
so we get the statement.
(cid:3)
Theorem 3.19. Let T ∈ B0,1(Vn) and let σS(T ) = σ1S(T ) ∪ σ2S(T ), with
dist (σ1S(T ), σ2S(T )) > 0.
Let U1 and U2 be two axially symmetric s-domains such that σ1S(T ) ⊂ U1 and σ2S(T ) ⊂
U2, with U 1 ∩ U 2 = ∅. Set
Pj :=
Tj :=
1
2π Z∂(Uj ∩CI )
2π Z∂(Uj ∩CI )
1
S−1
L (s, T ) dsI,
j = 1, 2,
S−1
L (s, T ) dsI s,
j = 1, 2.
(3.53)
(3.54)
Then Pj are projectors and T Pj = PjT for j = 1, 2.
Proof. Let σjS(T ) ⊂ G1 and G2 be two T -admissible open sets such that G1∪∂G1 ⊂ G2
and G2 ∪ ∂G2 ⊂ Uj, for j = 1 or 2. Thanks to the structure of the S-spectrum we will
assume that G1 and G2 are axially symmetric and s-domains.
Take p ∈ ∂(G1 ∩ CI) and s ∈ ∂(G2 ∩ CI) and observe that, for I ∈ S, we have
Pj :=
1
2π Z∂(G2∩CI )
dsIS−1
R (s, T )
but we can also write Pj as
Pj =
1
2π Z∂(G1∩CI )
S−1
L (p, T )dpI.
Now consider P 2
j written as
P 2
j =
1
(2π)2 Z∂(G2∩CI )
dsI Z∂(G1∩CI )
S−1
R (s, T )S−1
L (p, T )dpI.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
15
Using the resolvent equation we write:
P 2 =
−
1
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
1
dsI Z∂(G1∩CI )
dsI Z∂(G1∩CI )
[S−1
R (s, T ) − S−1
L (p, T )]p(p2 − 2s0p + s2)−1dpI
s[S−1
R (s, T ) − S−1
L (p, T )](p2 − 2s0p + s2)−1dpI.
Now observe that
1
(2π)2 Z∂(G2∩CI )
dsIS−1
R (s, T )Z∂(G1∩CI )
p(p2 − 2s0p + s2)−1dpI = 0
and
−
1
(2π)2 Z∂(G2∩CI )
dsIsS−1
R (s, T )Z∂(G1∩CI )
(p2 − 2s0p + s2)−1dpI = 0
since the functions
p 7→ p(p2 − 2s0p + s2)−1,
p 7→ (p2 − 2s0p + s2)−1
are slice hyperholomorphic and do not have singularities inside ∂(G1 ∩ CI). So P 2
be written as
j can
P 2
j =
−
=
1
1
dsI Z∂(G1∩CI )
dsI Z∂(G1∩CI )
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )Z∂(G1∩CI )
1
−S−1
L (p, T )p(p2 − 2s0p + s2)−1dpI
−sS−1
L (p, T )(p2 − 2s0p + s2)−1dpI,
dsI(sS−1
L (p, T ) − S−1
L (p, T )p)(p2 − 2s0p + s2)−1dpI.
Applying now Lemma 3.18 with B := S−1
have
L (p, T ) and observing that p ∈ G2, we finally
P 2
j =
1
2π Z∂(G1∩CI )
S−1
L (p, T )dpI = Pj.
Let us now prove that T Pj = PjT . Observe that the functions f (s) = sm, for m ∈ N0
are both right and left slice hyperholomorphic. So the operator T can be written as
T =
1
2π Z∂(U ∩CI )
S−1
L (s, T ) dsI s =
1
2π Z∂(U ∩CI )
s dsI S−1
R (s, T );
analogously, as already observed, for the projectors Pj we have
Pj =
1
2π Z∂(Uj ∩CI )
S−1
L (s, T ) dsI =
1
2π Z∂(Uj ∩CI )
dsI S−1
R (s, T ).
From the identity
Tj =
1
2π Z∂(Uj ∩CI )
S−1
L (s, T ) dsI s =
1
2π Z∂(Uj ∩CI )
s dsI S−1
R (s, T )
we can compute T Pj as:
T Pj =
1
2π Z∂(Uj ∩CI )
T S−1
L (s, T ) dsI
16
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
and using the resolvent equation (3.25) it follows
T Pj =
=
=
1
1
1
2π Z∂(Uj ∩CI )
2π Z∂(Uj ∩CI )
2π Z∂(Uj ∩CI )
[S−1
L (s, T ) s − I] dsI
S−1
L (s, T ) s dsI
S−1
L (s, T ) dsI s
Now consider
= Tj.
PjT =
1
2π Z∂(Uj ∩CI )
dsI S−1
R (s, T )T
and using the resolvent equation (3.26) we obtain
PjT =
=
1
2π Z∂(Uj ∩CI )
2π Z∂(Uj ∩CI )
1
dsI [s S−1
R (s, T ) − I]
dsI s S−1
R (s, T )
so the equality PjT = T Pj holds.
= Tj,
(3.55)
(3.56)
(cid:3)
Remark 3.20. The properties that the Riesz projectors commute with the operator
T has been proved for the quaternionic version of the S-functional calculus in [4], while
the property that P 2 = P given in [19] is obtained heuristically. This fact shows the
importance of this new resolvent equation.
As it is well known for hyperholomorphic functions the product of two hyperholomor-
phic functions is not in general hyperholomorphic. Here we recall a class of functions
for which the pointwise multiplication remains slice hyperholomorphic.
Definition 3.21. Let f : U → Rn be a slice hyperholomorphic function, where U is an
open set in Rn+1. We define
N (U) = {f ∈ SM(U) : f (U ∩ CI) ⊆ CI, ∀I ∈ S}.
Proposition 3.22. Let U be an open set in Rn+1. Let f ∈ N (U), g ∈ SM(U), then
f g ∈ SM(U).
First of all let us observe that functions in the subclass N (U) are both left and right
slice hyperholomorphic. When we take the power series expansion of this class of
functions at a point on the real line the coefficients of the expansion are real numbers.
Now observe that for functions in f ∈ N (U) we can define f (T ) using the left but also
the right S-functional calculus. It is
f (T ) =
=
1
2π Z∂(U ∩CI )
2π Z∂(U ∩CI )
1
S−1
L (s, T ) dsI f (s)
f (s) dsI S−1
R (s, T ).
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
17
Lemma 3.23. Let B ∈ B(Vn). Let G be an axially symmetric s-domain and assume
that f ∈ N (G). Then, for p ∈ G, we have
1
2π Z∂(G∩CI )
f (s)dsI(sB − Bp)(p2 − 2s0p + s2)−1 = Bf (p).
Proof. Recalling formula (3.51) we write
1
2π Z∂(G∩CI )
f (s)dsI(sB − Bp)(p2 − 2s0p + s2)−1
=
=
+
1
1
2π Z∂(G∩CI )
2π Z∂(G∩CI )
2π Z∂(G∩CI )
1
f (s)dsI(s2 − 2p0s + p2)−1(sB − Bp)
f (s)dsI(s2 − 2p0s + p2)−1(s − p)B
f (s)dsI(s2 − 2p0s + p2)−1(pB − Bp)
:= J1 + J2
but observe that
J1 =
=
1
2π Z∂(G∩CI )
2π Z∂(G∩CI )
1
f (s)dsI(s2 − 2p0s + p2)−1(s − p)B
f (s)dsIS−1
R (s, p)B = f (p)B.
Consider now the second integral. Taking s = u + Iv then the solutions of the equation
s2 − 2p0s + p2 = 0 are s1 = α and s2 = α where α = p0 + Ip, so
J2 =
=
1
2π Z∂(G∩CI )
2π Z∂(G∩CI )
1
f (s)dsI(s2 − 2p0s + p2)−1(pB − Bp)
f (s)
(s − α)(s − α)
dsI(pB − Bp),
by the residues theorem we get
J2 =
1
2π Z∂(G∩CI )
f (s)dsI(s2 − 2p0s + p2)−1(pB − Bp)
=
I
2p
[f (p0 − Ip) − f (p0 + Ip)](pB − Bp).
Now we recall the structure formula that shows that a slice hyperholomorphic function
can be written as
f (p) = α(p0, p) + Ipβ(p0, p)
where
α(p0, p) =
β(p0, p) =
1
2
I
2
[f (p0 − Ip) + f (p0 + Ip)],
[f (p0 − Ip) − f (p0 + Ip)]
18
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
and in the case of functions f ∈ N the functions α and β are real valued. Observe that
J1 + J2 = f (p)B +
I
2p
[f (p0 − Ip) − f (p0 + Ip)](pB − Bp)
= α(p0, p)B + Ipβ(p0, p)B +
= α(p0, p)B + Ipβ(p0, p)B +
= B(α(p0, p) + Ipβ(p0, p))
= Bf (p),
β(p0, p)
p
β(p0, p)
p
(pB − Bp)
((p0 − Ipp)B − B(p0 − Ipp))
so we get the statement.
(cid:3)
Remark 3.24. If we assume that f ∈ N (B(0, r)) where B(0, r) is the open ball in
Rn+1 centered at 0 and of radius r > 0 and s ∈ B(0, r), then the proof of the above
theorem follows in a shorter way. Indeed we have
(sB − Bp)(p2 − 2s0p + s2)−1 = Xm≥0
s−1−mBpm,
p < s.
So
but
1
2π Z∂(G∩CI )
f (s)dsI Xm≥0
s−1−mBpm,
p < s,
1
2π Z∂(G∩CI )
Xm≥0
f (s)dsIs−1−mBpm = Xm≥0
1
m!
f (m)(0)Bpm
and for functions in N (B(0, r)) the derivatives f (m)(0) are real numbers and so they
commute with B. We get
1
m!
Xm≥0
f (m)(0)Bpm = B Xm≥0
1
m!
f (m)(0)pm = Bf (p).
We now offer a different proof of the theorem that shows that (f g)(T ) = f (T )g(T ),
under suitable assumptions of f, g. Originally, see [19], the proof was based on the
Cauchy formula and the resolvent equations (3.25), (3.26).
Theorem 3.25. Let T ∈ B0,1(Vn) and assume f ∈ NσS (T ) and g ∈ SMσS (T ). Then we
have
(f g)(T ) = f (T )g(T ).
Proof. Let σS(T ) ⊂ G1 and G2 be two T -admissible open sets such that G1 ∪ ∂G1 ⊂ G2
and G2 ∪ ∂G2 ⊂ U. Take p ∈ ∂(G1 ∩ CI) and s ∈ ∂(G2 ∩ CI) and observe that, for
I ∈ S, we have
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
19
f (T )g(T ) =
=
−
−
+
1
1
1
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
1
1
f (s) dsI S−1
R (s, T )Z∂(G1∩CI )
S−1
L (p, T ) dpI g(p)
f (s) dsI Z∂(G1∩CI )
f (s) dsI Z∂(G1∩CI )
f (s) dsI Z∂(G1∩CI )
f (s) dsI Z∂(G1∩CI )
S−1
R (s, T )p(p2 − 2s0p + s2)−1dpI g(p)
S−1
L (p, T )p(p2 − 2s0p + s2)−1dpI g(p)
sS−1
R (s, T )(p2 − 2s0p + s2)−1dpI g(p)
sS−1
L (p, T )(p2 − 2s0p + s2)−1dpI g(p)
where we have used the resolvent equation. But now observe that
1
(2π)2 Z∂(G2∩CI )
f (s) dsI Z∂(G1∩CI )
S−1
R (s, T )p(p2 − 2s0p + s2)−1dpI g(p) = 0
and
−
1
(2π)2 Z∂(G2∩CI )
f (s) dsI Z∂(G1∩CI )
sS−1
R (s, T )(p2 − 2s0p + s2)−1dpI g(p) = 0.
so it follows that
f (T )g(T ) = −
S−1
L (p, T )p(p2 − 2s0p + s2)−1dpI g(p)
1
(2π)2 Z∂(G2∩CI )
(2π)2 Z∂(G2∩CI )
f (s) dsI Z∂(G1∩CI )
f (s) dsI Z∂(G1∩CI )
1
+
sS−1
L (p, T )(p2 − 2s0p + s2)−1dpI g(p)
which can be written as
1
f (T )g(T ) =
(2π)2 Z∂(G2∩CI )
f (s) dsI Z∂(G1∩CI )
[sS−1
L (p, T ) − S−1
L (p, T )p]×
× (p2 − 2s0p + s2)−1dpI g(p).
Using Lemma 3.23 we get
f (T )g(T ) =
1
2π Z∂(G1∩CI )
S−1
L (p, T )dpI f (p) g(p)
which gives the statement.
(cid:3)
In the original proof of the above theorem we have used the fact that for functions
f ∈ NσS (T ) the left S-resolvent equation gives
f (T )T m =
1
2π Z∂(U ∩CI )
S−1
L (p, T )dpI f (p)pm
from which one obtains
f (T )T mt−1−m =
1
2π Z∂(U ∩CI )
S−1
L (p, T )dpI f (p)pmt−1−m.
20
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
By taking the sum and considering t ∈ ρS(T ), we have
f (T )S−1
L (t, T ) =
1
2π Z∂(U ∩CI )
S−1
L (p, T )dpI f (p)S−1
L (t, p).
Using this equality and the Cauchy formula we obtain the statement.
4. The case of commuting operators and the quaternionic case
In this last section we state the resolvent equation in the case of commuting operators
and for the quaternionic functional calculus. We also take the occasion to make some
comments that show how the S-functional calculus turns out to be a natural extension
of the Riesz-Dunford functional calculus.
4.1. The case of several commuting operators. We denote by BC0,1(Vn) the subset
of B0,1(Vn) consisting of paravector operators with commuting components. Given an
operator in paravector form T = T0 + e1T1 + . . . + enTn, its so-called conjugate T is
defined by T = T0 − e1T1 − . . . − enTn. When T ∈ BC0,1(Vn) the operator T T is well
defined and T T = T T = T 2
Theorem 4.1. Let T ∈ BC0,1(Vn) and s ∈ Rn+1 be such that s < kT k. Then
n and T + T = 2T0.
0 + T 2
1 + . . . + T 2
T ms−1−m = (sI − ¯T )(s2I − s(T + T ) + T T )−1,
s−1−mT m = (s2I − s(T + T ) + T T )−1(sI − ¯T ).
Xm≥0
Xm≥0
(4.57)
(4.58)
The above theorem follows from the fact that the Cauchy kernels for slice hyperholo-
morphic functions can be written in two possible ways, see Section 2 and [15]. In the
case of commuting operators the two expressions are equivalent. The advantage of
this approach is that one can work with the so called F-spectrum which is easier to
compute than the S-spectrum. In fact it can be computed over a complex plane CI,
taking s = u + Iv, and then extended to Rn. This is a consequence of the fact that the
F-spectrum takes into account the commutativity of the operators Tj, j = 0, 1, ..., n.
The F-spectrum is suggested by Theorem 4.1 and it is described below.
Definition 4.2 (The F -spectrum and the F -resolvent sets). Let T ∈ BC0,1(Vn). We
define the F -spectrum of T as:
σF (T ) = {s ∈ Rn+1
:
s2I − s(T + T ) + T T
is not invertible }.
The F -resolvent set of T is defined by
ρF (T ) = Rn+1 \ σF (T ).
The main properties of the F -spectrum are similar to those of the S-spectrum as it is
proved in the next results:
Theorem 4.3 (Structure of the F -spectrum). Let T ∈ BC0,1(Vn) and let p = p0 +p1I ∈
[p0 + p1I] ⊂ Rn+1 \ R, such that p ∈ σF (T ). Then all the elements of the (n − 1)-sphere
[p0 + p1I] belong to σF (T ).
Theorem 4.4 (Compactness of F -spectrum). Let T ∈ BC0,1(Vn). Then the F -spectrum
σF (T ) is a compact nonempty set. Moreover σF (T ) is contained in {s ∈ Rn+1 : s ≤
kT k }.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
21
The relation between the S-spectrum and the F-spectrum is contained in the following
result:
Proposition 4.5. Let T ∈ BC0,1(Vn) . Then σF (T ) = σS(T ).
Definition 4.6. (The SC-resolvent operator) Let T ∈ BC0,1(V ) and s ∈ ρF (T ). We
define the SC-resolvent operator as
S−1
C,L(s, T ) := (sI − T )(s2I − s(T + T ) + T T )−1.
C,R(s, T ) := (s2I − s(T + T ) + T T )−1(sI − T ).
S−1
(4.59)
(4.60)
Theorem 4.7. Let T ∈ BC0,1(Vn) and s, p ∈ ρF (T ). Then S−1
S-resolvent equation
C,L(s, T ) satisfies the left
S−1
C,L(s, T )s − T S−1
C,L(s, T ) = I,
(4.61)
and S−1
C,R(s, T ) satisfies the right S-resolvent equation
Moreover, for p 6∈ [s], it satisfies the resolvent equation
sS−1
C,R(s, T ) − S−1
C,R(s, T )T = I.
S−1
C,R(s, T )S−1
C,L(p, T ) = ((S−1
− s(S−1
C,R(s, T ) − S−1
C,R(s, T ) − S−1
C,L(p, T ))p
C,L(p, T )))(p2 − 2s0p + s2)−1.
which can be written as
S−1
C,R(s, T )S−1
C,L(p, T ) = (s2 − 2p0s + p2)−1(s(S−1
C,L(p, T ))p).
C,R(s, T ) − S−1
− (S−1
C,R(s, T ) − S−1
C,L(p, T ))
(4.62)
We conclude this subsection with a couple of considerations on the case of unbounded
operators.
(I) Suppose that T is a closed operator with domain D(T ). As one can clearly see, the
non commutative version of S−1
L (s, T ), that is
S−1
L (s, T ) := −(T 2 − 2Re(s)T + s2I)−1(T − sI),
is defined on the domain of T , and not on Vn as it is in the classical case. So we have
to consider the extension to Vn writing S−1
L (s, T ) as follows
S−1
L (s, T ) := −(T (T 2 − 2Re(s)T + s2I)−1 − (T 2 − 2Re(s)T + s2I)−1sI).
In this case S−1
L (s, T ) turns out to be defined on Vn. Observe now that if T is a closed
operator with domain D(T ) and with commuting components the left S-resolvent op-
erator S−1
C,L(s, T ) turns out to be already defined on Vn. For a more detailed discussion
see the original papers [13, 15]. In the case of the right S-resolvent we have the opposite
situation. With the above consideration the new resolvent equation remains the same
also for unbounded operators.
(II) The F-spectrum is also useful to defined the so called F-functional calculus, see
[5, 16]. This calculus is defined using the Fueter-Sce-Qian mapping theorem in integral
form, see [22, 31, 34]. It is a hyperholomorphic functional calculus in the spirit of A.
McIntosh, B. Jefferies and their coauthors (see [27, 28, 30, 33], the monograph [26]
and the references therein) who first used the theory of hyperholomorphic functions,
see [10, 18, 20, 25], to define a hyperholomorphic functional calculus for n-tuples of
operators.
22
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
4.2. The quaternionic setting. What we have previously proved in the paper can
be rephrased also for the quaternionic functional calculus. We simply point out that
in this case slice hyperholomorphic functions are defined on an open set U ⊆ H and
have values in the quaternions H. The resolvent operators are as in the introduction
of this paper. Here it is important to consider right linear operators as well as left
linear operators T . The possible formulations of the quaternionic functional calculus
has been carried out in [13]. The resolvent equations in Theorem 3.8 hold in this
setting where instead of the paravector operator T = T0 + T1e1 + . . . + Tnen we replace
quaternionic operators. We finally mention for sake of completeness one more analogy
with the classical case. As it is well known the Laplace transform of a semigroup etG
where for simplicity we take a bounded operator G defined on a Banach space X is the
resolvent operator (λI − G)−1. In the quaternionic case we have the analogue result
for the two S-resolvent operators: Let T ∈ B(V ) and let s0 > kT k. Then the left
S-resolvent operator S−1
L (s, T ) is given by
S−1
L (s, T ) = Z +∞
0
and S−1
R (s, T ) is given by
S−1
R (s, T ) = Z +∞
0
etT e−ts dt,
e−tset T dt.
We point out that the theory of the quaternionic evolution operators is developed in
[12] where it is also studied the case in which the generator is unbounded. Recently
the case of sectorial operators has been treated in [24].
References
[1] D. Alpay, The Schur algorithm, reproducing kernel spaces and system theory, American Mathe-
matical Society, Providence, RI, 2001. Translated from the 1998 French original by Stephen S.
Wilson, Panoramas et Synth`eses.
[2] D. Alpay, F. Colombo and I. Sabadini, Schur functions and their realizations in the slice hyper-
holomorphic setting, Integral Equations and Operator Theory 72 (2012), 253-289.
[3] D. Alpay, F. Colombo and I. Sabadini, Pontryagin de Branges-Rovnyak spaces of slice hyperholo-
morphic functions, J. d'Analyse Math´ematique, 2013.
[4] D. Alpay, F. Colombo and I. Sabadini, Krein-Langer factorization and related topics in the slice
hyperholomorphic setting, J. of Geom. Anal., to appear.
[5] D. Alpay, F. Colombo and I. Sabadini, On some notions of convergence for n-tuples of operators,
Math. Meth. Appl. Sci. (2013/14), to appear.
[6] D. Alpay, F. Colombo, I. Lewkowicz, and I. Sabadini. Realizations of slice hyperholomorphic
generalized contractive and positive functions. arXiv:1310.1035.
[7] D. Alpay, A. Dijksma, J. Rovnyak, and H. de Snoo. Schur functions, operator colligations, and
reproducing kernel Pontryagin spaces, volume 96 of Operator theory: Advances and Applications.
Birkhauser Verlag, Basel, 1997.
[8] D. Alpay, Y. Peretz, Realizations for Schur upper triangular operators centered at an arbitrary
point, Integral Equations Operator Theory 37 (2000), 251–323.
[9] D. Alpay and V. Vinnikov, Finite dimensional de Branges spaces on Riemann surfaces. J. Funct.
Anal., 189(2):283–324, 2002.
[10] F. Brackx, R. Delanghe, F. Sommen, Clifford analysis. Research Notes in Mathematics, 76,
Pitman, Advanced Publishing Program, Boston, MA, 1982.
[11] F. Colombo, I. Sabadini, On some properties of the quaternionic functional calculus, J. Geom.
Anal., 19 (2009), 601–627.
A NEW RESOLVENT EQUATION FOR THE S-FUNCTIONAL CALCULUS
23
[12] F. Colombo, I. Sabadini, The quaternionic evolution operator, Adv. Math., 227 (2011), 1772–
1805.
[13] F. Colombo, I. Sabadini, On the formulations of the quaternionic functional calculus, J. Geom.
Phys., 60 (2010), 1490–1508.
[14] F. Colombo, I. Sabadini, The Cauchy formula with s-monogenic kernel and a functional calculus
for noncommuting operators, J. Math. Anal. Appl., 373 (2011), 655–679.
[15] F. Colombo, I. Sabadini, The F-spectrum and the SC-functional calculus, Proc. Roy. Soc. Edin-
burgh Sect. A, 142 (2012), 479–500.
[16] F. Colombo, I. Sabadini, F. Sommen, The Fueter mapping theorem in integral form and the
F -functional calculus, Math. Meth. Appl. Sci., 33 (2010), 2050–2066.
[17] F. Colombo, I. Sabadini, D.C. Struppa, A new functional calculus for noncommuting operators,
J. Funct. Anal., 254 (2008), 2255–2274.
[18] F. Colombo, I. Sabadini, F. Sommen, D.C. Struppa, Analysis of Dirac Systems and Computational
Algebra, Progress in Mathematical Physics, Vol. 39, Birkhauser, Boston, 2004.
[19] F. Colombo, I. Sabadini, D. C. Struppa, Noncommutative Functional Calculus. Theory and Appli-
cations of Slice Hyperholomorphic Functions, Progress in Mathematics, Birkhauser, Basel, 2011.
[20] R. Delanghe, F. Sommen, V. Soucek, Clifford algebra and spinor-valued functions. A function
theory for the Dirac operator. Mathematics and its Applications, 53. Kluwer Academic Publishers
Group, Dordrecht, 1992. xviii+485 pp.
[21] N. Dunford, J. Schwartz, Linear Operators, part I: General Theory , J. Wiley and Sons (1988).
[22] R. Fueter, Analytische Funktionen einer Quaternionenvariablen, Comm. Math. Helv., 4 (1932),
9–20.
[23] R. Ghiloni, V. Moretti, A. Perotti, Continuous slice functional calculus in quaternionic Hilbert
spaces, Reviews in Mathematical Physics, Vol. 25, No. 4 (2013), pp.1350006–1–1350006–83.
[24] R. Ghiloni, V. Recupero, Semigroups over real alternative *-algebras: generation theorems and
spherical sectorial operators, Preprint 2013.
[25] K. Gurlebeck, K. Habetha, W. Sprossig, Holomorphic Functions in the Plane and n-dimensional
space, Birkhauser, Basel, 2008.
[26] B. Jefferies, Spectral properties of noncommuting operators, Lecture Notes in Mathematics, 1843,
Springer-Verlag, Berlin, 2004.
[27] B. Jefferies, A. McIntosh, The Weyl calculus and Clifford analysis, Bull. Austral. Math. Soc., 57
(1998), 329–341.
[28] B. Jefferies, A. McIntosh, J. Picton-Warlow, The monogenic functional calculus, Studia Math.,
136 (1999), 99–119.
[29] M.S. Livsic, N. Kravitski, A. Markus, and V. Vinnikov. Commuting nonselfadjoint operators and
their applications to system theory. Kluwer, 1995.
[30] A. McIntosh, A. Pryde, A functional calculus for several commuting operators, Indiana Univ.
Math. J., 36 (1987), 421–439.
[31] T. Qian, Generalization of Fueter's result to Rn+1, Rend. Mat. Acc. Lincei, 8 (1997), 111–117.
[32] M. Reed, B. Simon, Functional Analysis, Methods of Modern Mathematical Physics, Academic
Press; REV edition, 1980.
[33] C. Li, A. McIntosh, T. Qian, Clifford algebras, Fourier transforms and singular convolution
operators on Lipschitz surfaces, Rev. Mat. Iberoamericana 10 (1994), 665–721.
[34] M. Sce, Osservazioni sulle serie di potenze nei moduli quadratici, Atti Acc. Lincei Rend. Fisica ,
23 (1957), 220–225.
[35] V. Vinnikov, Commuting nonselfadjoint operators and algebraic curves, volume 59 of Operator
Theory: Advances and Applications, pages 348–371. Birkhauser Verlag, Basel, 1992.
24
D. ALPAY, F. COLOMBO, J. GANTNER, AND I. SABADINI
(DA) Department of Mathematics, Ben-Gurion University of the Negev, Beer-Sheva
84105 Israel
E-mail address: [email protected]
(FC) Politecnico di Milano, Dipartimento di Matematica, Via E. Bonardi, 9, 20133
Milano, Italy
E-mail address: [email protected]
(JG) Vienna University of Technology, Institute for Analysis and Scientific Comput-
ing, Wiedner Hauptstrasse 8 - 10, 1040 Wien, Austria
E-mail address: [email protected]
(IS) Politecnico di Milano, Dipartimento di Matematica, Via E. Bonardi, 9, 20133
Milano, Italy
E-mail address: [email protected]
|
1908.07786 | 1 | 1908 | 2019-08-21T10:47:53 | On the Banach lattice c_0 | [
"math.FA"
] | We show that $c_0$ is not a projective Banach lattice, answering a question of B. de Pagter and A. Wickstead. On the other hand, we show that $c_0$ is complemented in the free Banach lattice generated by itself (seen as a Banach space). As a consequence, the free Banach lattice generated by $c_0$ is not projective. | math.FA | math | ON THE BANACH LATTICE c0
ANTONIO AVIL´ES, GONZALO MART´INEZ-CERVANTES,
AND JOS´E DAVID RODR´IGUEZ ABELL ´AN
Abstract. We show that c0 is not a projective Banach lattice, answering a question of
B. de Pagter and A. Wickstead. On the other hand, we show that c0 is complemented in
the free Banach lattice generated by itself (seen as a Banach space). As a consequence,
the free Banach lattice generated by c0 is not projective.
9
1
0
2
g
u
A
1
2
]
.
A
F
h
t
a
m
[
1
v
6
8
7
7
0
.
8
0
9
1
:
v
i
X
r
a
1. Introduction
The purpose of this paper is to answer negatively Question 12.11 proposed by B. de
Pagter and A. Wickstead in [3] (notice that this also answers negatively [3, Question 12.10]).
We prove that c0, seen as a Banach lattice, is not projective. Moreover, we show that it
can be (isometrically) embedded as a Banach lattice into the free Banach lattice generated
by itself seen as a Banach space, which is denoted by F BL[c0]. This embedding composed
with the natural quotient from F BL[c0] onto c0 gives the identity map on c0. Thus, c0 is
complemented in F BL[c0]. As a consequence, we will obtain that F BL[c0] is not projective.
The concepts of free and projective Banach lattices were introduced in [3]. If A is a set
with no extra structure, the free Banach lattice generated by A, denoted by F BL(A), is
a Banach lattice together with a bounded map u : A −→ F BL(A) having the following
universal property: for every Banach lattice Y and every bounded map v : A −→ Y there
is a unique Banach lattice homomorphism S : F BL(A) −→ Y such that S ◦ u = v and
kSk = sup {kv(a)k : a ∈ A}. The same idea is applied by A. Avil´es, J. Rodr´ıguez and P.
Tradacete to define the concept of the free Banach lattice generated by a Banach space E,
F BL[E]. This is a Banach lattice together with a bounded operator u : E −→ F BL[E]
such that for every Banach lattice Y and every bounded operator T : E −→ Y there
is a unique Banach lattice homomorphism S : F BL[E] −→ Y such that S ◦ u = v and
kSk = kT k.
2010 Mathematics Subject Classification. 46B43, 06BXX.
Key words and phrases. c0; F BL[c0]; Banach lattice; Free Banach lattice; Projectivity.
Authors supported by project MTM2017-86182-P (Government of Spain, AEI/FEDER, EU) and project
20797/PI/18 by Fundaci´on S´eneca, ACyT Regi´on de Murcia. Third author supported by FPI contract of
Fundaci´on S´eneca, ACyT Regi´on de Murcia.
1
2
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
In [2] and [3], the corresponding authors show that both objects exist and are unique up
to Banach lattices isometries. Moreover, A. Avil´es, J. Rodr´ıguez and P. Tradacete give an
explicit description of them in [2].
Let A be a non-empty set. For x ∈ A, let δx : [−1, 1]A −→ [−1, 1] be the evaluation
function given by δx(x∗) = x∗(x) for every x∗ ∈ [−1, 1]A, and for every f : [−1, 1]A −→ R
define
kf k = sup( n
Xi=1
f (x∗
i ) : n ∈ N, x∗
1, . . . , x∗
n ∈ [−1, 1]A, sup
x∈A
x∗
i (x) ≤ 1) .
n
Xi=1
The Banach lattice F BL(A) is the Banach lattice generated by the evaluation functions
δx inside the Banach lattice of all functions f : [−1, 1]A −→ R with finite norm. The
natural identification of A inside F BL(A) is given by the map u : A −→ F BL(A) where
u(x) = δx. Since every function in F BL(A) is a uniform limit of such functions, they are all
continuous (with respect to the product topology) and positively homogeneous, i.e. they
commute with multiplication by positive scalars.
Now, let E be a Banach space. For a function f : E∗ −→ R consider the norm
kf kF BL[E] = sup( n
Xi=1
f (x∗
i ) : n ∈ N, x∗
1, . . . , x∗
n ∈ E∗,
sup
x∈BE
n
Xi=1
x∗
i (x) ≤ 1) .
The Banach lattice F BL[E] is the closure of the vector lattice in RE ∗ generated by the
evaluations δx : x∗ 7→ x∗(x) with x ∈ E. These evaluations form the natural copy of E inside
F BL[E]. All the functions in F BL[E] are positively homogeneous and weak∗-continuous
when restricted to the closed unit ball BE ∗.
The notions of free and projective objects are closely related in the general theory of
In the setting of Banach lattices, de Pagter and Wickstead [3] introduced
categories.
projectivity in the following form:
Definition 1.1. A Banach lattice P is projective if whenever X is a Banach lattice, J a
closed ideal in X and Q : X −→ X/J the quotient map, then for every Banach lattice
homomorphism T : P −→ X/J and ε > 0, there is a Banach lattice homomorphism
T : P −→ X such that T = Q ◦ T and k T k ≤ (1 + ε) kT k.
Some examples of projective Banach lattices given in [3] include F BL(A), ℓ1, all finite
dimensional Banach lattices and Banach lattices of the form C(K), where K is a compact
neighborhood retract of Rn. They also prove that ℓ∞ and c are not projective.
In this
paper we will focus on c0 and F BL[c0].
ON THE BANACH LATTICE c0
3
2. Non-projectivity of c0 as a Banach lattice
In this section we are going to prove that c0, seen as a Banach lattice, is not projective.
We will use the following fact (see [1, Proposition 2.1]):
Proposition 2.1. Let P be a projective Banach lattice, I an ideal of P and π : P −→ P/I
the quotient map. The quotient P/I is projective if and only if for every ε > 0 there exists
a Banach lattice homomorphism uε : P/I −→ P such that π ◦ uε = idP/I and kuεk ≤ 1 + ε.
Let L = P +
f in(ω) = Pf in(ω) \ {∅} be the set of the finite parts of ω without the empty
set.
For A ∈ L let us define the map χA : L −→ [−1, 1] given by χA(B) = 1 if B ⊂ A and
χA(B) = 0 if B 6⊂ A.
Let Φ : F BL(L) −→ c0 be the map given by
Φ(f ) =(cid:0)f(cid:0)(χA({1}))A∈L(cid:1) , f(cid:0)(χA({2}))A∈L(cid:1) , . . .(cid:1) =(cid:0)f(cid:0)(χA({n}))A∈L(cid:1)(cid:1)n∈N
for every f : [−1, 1]L −→ R ∈ F BL(L).
Lemma 2.2. The map Φ : F BL(L) −→ c0 has the following properties:
(1) Φ(δA) =Pi∈A ei ∈ c0 for every A ∈ L.
f ∈ F BL(L).
(2) Φ is a Banach lattice homomorphism that is well-defined, i.e., Φ(f ) ∈ c0 for every
(3) Φ is surjective.
Proof. The first assertion follows from the definition of Φ. For every n ∈ N let gn :
L −→ [−1, 1] be the function gn = (χA({n}))A∈L . Then the sequence (gn)n∈N is pointwise
convergent to zero, so f (gn) converges to zero and Φ(f ) ∈ c0. Since Φ preserves linear
combinations, suprema, infima and kΦ(f )k ≤ kf k for every f ∈ F BL(L), we have that Φ
is a Banach lattice homomorphism whose image is in c0.
Let us prove property (3). Let x = (x1, x2, . . .) ∈ c0 and suppose, without loss of
generality, that x ≥ 0. Fix a sequence of natural numbers (ni)i∈N such that xn1 ≥ xn for
every n ∈ N and xni+1 ≥ xn for every n ∈ N \ {n1 . . . , ni}.
Now, let Ai = {n1, . . . , ni} and λi = xni − xni+1 for every i = 1, 2, . . .
For A ∈ L, if we put eA :=Pi∈A ei, we have that x =P∞
Xj=1
Φ(λjδAj ) = Φ(cid:0)
λjΦ(δAj ) =
Xj=1
∞
Xj=1
x =
∞
j=1 λjeAj , and then,
∞
λjδAj(cid:1),
4
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
where the last element P∞
j=1 λjδAj is well-defined since P∞
norm one.
j=1 λj < ∞ and each δAj has
(cid:3)
Thus, Φ is a quotient map. We are going to prove that there is no bounded Banach
lattice homomorphism ϕ : c0 −→ F BL(L) such that Φ ◦ ϕ = idc0. This fact will be a
consequence of the following Lemma:
Lemma 2.3. Let A be an infinite set, (x∗
in F BL(A) with the following properties:
n)n∈N a sequence in [−1, 1]A and (fn)n∈N a sequence
(1) fn ≥ 0 for every n ∈ N;
(2) fn(x∗
(3) For every finite set F ⊆ A there is a natural number n such that x∗
n) = 1 for every n ∈ N;
nF = 0, i.e. the
restriction of x∗
n to F is null.
Then for every ε > 0 there is a subsequence (fnk)k∈N such that
≥ m − ε for every m ∈ N.
m
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1
fnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Proof. Fix ε > 0 and fn1 := f1. Since the elements of F BL(A) are continuous with respect
1 such that f1(x∗) ≥ 1− ε
to the product topology, there is a neighborhood U1 of x∗
2 whenever
x∗ ∈ U1. In particular, there is a finite set F1 ⊆ A such that f1(x∗) ≥ 1 − ε
2 whenever
x∗F1 = x∗
1F1.
We recursively construct the subsequence (fnk)k∈N and the sequence of sets (Fk)k∈N. Sup-
niF1∪F2∪...∪Fi−1 =
pose that we have fn1, . . . , fnk and F1, . . . , Fk finite subsets of A such that x∗
0 and fni(x∗) ≥ 1 − ε
2i whenever x∗Fi = x∗
niFi.
Property (3) guarantees the existence of a number nk+1 ∈ N such that x∗
nk+1F1∪F2∪...∪Fk =
0. It follows from property (2) that there is a finite set Fk+1 ⊆ A such that fnk+1(x∗) ≥
1 − ε
2k+1 whenever x∗Fk+1 = x∗
nk+1Fk+1.
For each k ∈ N define y∗
k : A −→ [−1, 1] such that y∗
kFk = x∗
k(x) = 0 whenever
2k for every k ∈ N. On the other hand, if k < k′ and
k′(x) = 0. It
k) and therefore x∗
nk′ (x) = 0, so y∗
nkFk and y∗
x ∈ A \ Fk. Notice that fnk(y∗
y∗
k(x) 6= 0 then x ∈ Fk (by the definition of y∗
follows that y∗
k) ≥ 1 − ε
k and y∗
k′ have disjoint supports. In particular,
sup
x∈A
m
Xk=1
y∗
k(x) ≤ 1.
Thus,
ON THE BANACH LATTICE c0
5
fnk(z∗
: n ∈ N, z∗
1, . . . , z∗
n ∈ [−1, 1]A, sup
x∈A
n
z∗
i (x) ≤ 1) ≥
m
Xk=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
fnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
m
Xi=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= sup( n
Xk=1
Xi=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xk=1
≥
m
m
i )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for every m ∈ N.
fnk(y∗
(1)
≥
fnk (y∗
k) ≥
m
Xk=1
m
Xk=1(cid:16)1 −
Xi=1
2k(cid:17) ≥ m − ε
ε
(cid:3)
Theorem 2.4. The Banach lattice c0 is not projective.
Proof. We argue by contradiction. Suppose c0 is projective. Since F BL(L) is projective,
it follows from Proposition 2.1 and Lemma 2.2 the existence of a bounded Banach lattice
homomorphism ϕ : c0 −→ F BL(L) such that Φ ◦ ϕ = idc0. Set fn := ϕ(en). Since ϕ is
a Banach lattice homomorphism and each en is positive, we have that fn = ϕ(en) ≥ 0 for
every n ∈ N. It follows from the equality Φ(fn) = (Φ ◦ ϕ)(en) = en and the definition of Φ
that
for every n ∈ N. Set x∗
F ⊆ L we have x∗
fn(cid:0)(χA({n}))A∈L(cid:1) = en(n) = 1
n = (χA({n}))A∈L for every n ∈ N. Notice that for every finite set
n(S) = 0 for every S ∈ F whenever n /∈SS∈F S.
Thus, Lemma 2.3 asserts that for every ε > 0 there is a subsequence (fnk)k∈N such that
On the other hand, since ϕ is bounded, there is a constant C > 0 such that
m
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
fnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
enk!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ϕ m
Xk=1
m
Xk=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
fnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≥ m − ε for every m ∈ N.
m
≤ C(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1
enk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞
= C for every m ∈ N.
Thus, m − ε ≤ C for every ε > 0 and every m ∈ N, which yields to a contradiction.
(cid:3)
3. Complementability of c0 in F BL[c0]
This section is devoted to the proof that c0 is lattice-embeddable in F BL[c0] as a Banach
lattice, that is to say, there exist a Banach lattice homomorphism u : c0 −→ F BL[c0] and
two constants K, M ≥ 0 such that
K kxk∞ ≤ ku(x)kF BL[c0] ≤ M kxk∞
6
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
for every x ∈ c0. Moreover, we will prove that c0 is complemented in F BL[c0].
By [4, Theorem 4.50] we know that the Banach lattice c0 is lattice-embeddable in a
Banach lattice E if and only if there exists a disjoint sequence (xn)n∈N ∈ E+ (the positive
cone of E) such that
a) (xn)n∈N does not converge in norm to zero, and
b) the sequence of partial sums of (xn)n∈N is norm bounded, i.e., there exists some
M > 0 satisfying kPn
i=1 xikE ≤ M for every n ∈ N.
Thus, what we are going to do is to construct a sequence (fn)n∈N ∈ F BL[c0] with the
desired properties. The following lemma will be very useful:
Lemma 3.1. Let A be a set and f : [−1, 1]A −→ R a continuous and positively homogeneous
function that depends on a finite amount of coordinates, i.e., there exists a finite subset
A0 ⊂ A and f : [−1, 1]A0 −→ R such that f (x∗) = f (x∗A0). Then, f is in F BL(A).
Proof. The function f : [−1, 1]A0 −→ R is continuous and positively homogeneous. By [3,
Proposition 5.3], f is in F BL(A0).
Let T : F BL(A0) −→ F BL(A) be the map given by T (g)(x∗) = g(x∗A0) for every
g : [−1, 1]A0 −→ R and x∗ ∈ [−1, 1]A.
Clearly, f = T ( f ), and then f is in F BL(A).
(cid:3)
Let (Nn)n∈N be a strictly increasing sequence of natural numbers.
For r ∈ R let r+ = max {r, 0} be the positive part of r, and for every n ∈ N let
fn : c∗
0 −→ R be the map given by
fn(x∗) = (x∗
n − Nn max {x∗
m : m < n})+ · Πm>ngnm(x∗)
1, x∗
for every x∗ = (x∗
2, . . .) ∈ c∗
such that gnm(x∗) = 0 if Nmx∗
gnm( x∗
kx∗k) whenever x∗ 6= 0.
0 = ℓ1, where gnm : c∗
n ≤ x∗
m, gnm(x∗) = 1 if x∗
0 −→ [0, 1] is any continuous function
n and gnm(x∗) =
m ≤ (Nm − 1)x∗
Let us see that (fn)n∈N is a disjoint sequence of positive elements which satisfies both
properties a) and b):
Lemma 3.2. fn ≥ 0 for every n ∈ N and fn ∧ fl = 0 for every n 6= l.
ON THE BANACH LATTICE c0
7
Proof. The first assertion is clear. For the second one, suppose, for example, that n < l,
and let x∗ ∈ c∗
In
particular, x∗
n. Now, if fn(x∗) 6= 0, we have that gnm(x∗) 6= 0 for every m > n,
and then, that x∗
0 such that fl(x∗) 6= 0. We have that x∗
n for every m > n. Taking m = l we have a contradiction.
l > Nl max {x∗
m < Nmx∗
m : m < l}.
l > Nlx∗
Lemma 3.3. fn is in F BL[c0] for every n ∈ N.
Proof. Fix n ∈ N. We are going to find a sequence of functions (hk)k∈N ∈ F BL[c0] such
that limk→+∞ khk − fnkF BL[c0] = 0. Then, we will have that fn ∈ F BL[c0].
Let hk : c∗
0 −→ R be the map given by
(cid:3)
hk(x∗) = (x∗
n − Nn max {x∗
m : m < n})+ · Πn<m≤n+k gnm(x∗)
for every x∗ = (x∗
1, x∗
2, . . .) ∈ c∗
0 = ℓ1. Let us see that hk ∈ F BL[c0] for every k ∈ N.
1 = y∗
1, . . . , x∗
n+k = y∗
n+k. Let P : F BL(Bc0) −→ Rc∗
Notice that hk is continuous, positively homogeneous and satisfies that hk(x∗) = hk(y∗)
whenever x∗
0 be the map given
by P (f )(x∗) = f (( x∗(y)
kx∗k )y∈Bc0 ) · kx∗k if x∗ 6= 0 and P (f )(0) = f (0), where 0 denotes
the identically zero function in the corresponding space. We have that P (F BL(Bc0)) ⊂
F BL[c0] because P maps the evaluation functions in F BL(Bc0) to the evaluations func-
tions in F BL[c0], preserves linear combinations, the lattice structure and kP (f )kF BL[c0] ≤
kf kF BL(Bc0 ).
Now, let hk : [−1, 1]Bc0 −→ R be the map given by
hk(z∗) = hk((z∗(e1), . . . , z∗(en+k), 0, 0, . . .))
for every z∗ ∈ [−1, 1]Bc0 . Since hk is continuous, positively homogeneous and depends only
on finitely many coordinates, by Lemma 3.1 we have that hk ∈ F BL(Bc0). It follows that
P ( hk) = hk ∈ F BL[c0].
Now, by definition, we have that
khk − fnkF BL[c0] = sup( l
Xi=1
(hk − fn)(x∗
i ) : l ∈ N, x∗
1, . . . , x∗
l ∈ Bℓ1, sup
x∈Bc0
x∗
i (x) ≤ 1) .
l
Xi=1
Take x∗
1, . . . , x∗
l ∈ Bℓ1, with x∗
i = (x∗
i1, x∗
i2, . . .) for every i = 1, . . . , l and such that
(hk − fn)(x∗
i ) 6= 0 for every i = 1, . . . , l and supx∈Bc0Pl
i=1 x∗
i (x) ≤ 1.
Note that hk ≥ fn, so we can remove absolute values in the previous expression. Then,
we have that
8
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
in − Nn max {x∗
im : m < n})+ ·
i ) − Πn<m gnm(x∗
i )(cid:3)
in − Nn max {x∗
im : m < n})+ ·
l
Xi=1
(hk − fn)(x∗
i ) =
·
=
·
=
l
l
(x∗
(x∗
Xi=1
(cid:2)Πn<m≤n+k gnm(x∗
Xi=1
(cid:2)Πn<m≤n+k gnm(x∗
Xi=1
(x∗
l
in − Nn max {x∗
im : m < n})+ ·
· Πn<m≤n+k gnm(x∗
≤
l
Xi=1
x∗
in(cid:2)1 − Πn+k<m gnm(x∗
i )(cid:3).
i )(cid:2)1 − Πn+k<m gnm(x∗
i )(cid:3)
i ) − Πn<m≤n+k gnm(x∗
i )Πn+k<m gnm(x∗
i )(cid:3)
Since (hk − fn)(x∗
i ) 6= 0 for every i = 1, . . . , l, we have that x∗
i ) 6= 0. Thus, for every i there exists mi > n + k such that gnmi(x∗
1 − Πn+k<m gnm(x∗
in > Nmi − 1. Since Nmi > Nn+k, this implies that x∗
Nn+k−1x∗
imi
that is to say,
in <
x∗
x∗
in 6= 0 and also that
i ) 6= 1,
imi.
1
Thus,
(hk − fn)(x∗
i ) ≤
l
Xi=1
x∗
in <
1
Nn+k − 1
l
Xi=1
x∗
imi.
l
Xi=1
Therefore, since limk→+∞
1
Nn+k−1 = 0, the proof will follow from the following Claim.
Claim. For every x∗
1, x∗
2, . . . , x∗
l ∈ Bℓ1 and every natural numbers m1, m2, . . . , ml ∈ N
we have
x∗
imi ≤ 1
l
Xi=1
whenever supx∈Bc0Pl
i=1 x∗
i (x) ≤ 1.
Proof of the Claim. Fix m = max{mi : 1 ≤ i ≤ l}. We show first that
l
Xi=1
x∗
imi ≤ max( l
Xi=1 (cid:12)(cid:12)
m
Xj=1
ε(j)x∗
ij(cid:12)(cid:12) : ε ∈ {−1, +1}m) .
In fact,
ON THE BANACH LATTICE c0
9
max( l
Xi=1 (cid:12)(cid:12)
m
Xj=1
ε(j)x∗
ij(cid:12)(cid:12) : ε ∈ {−1, +1}m) ≥
=
≥
=
l
m
1
1
l
m
ε(mi)=ε
ε(j)x∗
Xj=1
Xi=1 (cid:12)(cid:12)
2m Xε∈{−1,+1}m
ij(cid:12)(cid:12)
Xj=1
2m Xε∈{−1,+1}
Xi=1 Xε∈{−1,+1}m
(cid:12)(cid:12)
Xi=1
2 Xε∈{−1,+1}
2m−1(cid:12)(cid:12) Xε∈{−1,+1}m
Xi=1
imi.
ε(mi)=ε
x∗
1
1
l
l
ε(j)x∗
ij(cid:12)(cid:12)
m
Xj=1
ε(j)x∗
ij(cid:12)(cid:12)
Let ε ∈ {−1, +1}m be the function which gives the maximum above. We have that
x∗
imi ≤
l
Xi=1
m
Xj=1
l
Xi=1 (cid:12)(cid:12)
ε(j)x∗
ij(cid:12)(cid:12) =
x∗
i (x) ≤ 1,
l
Xi=1
taking, in the equality, x = (ε(1), ε(2), . . . , ε(m), 0, . . .) ∈ Bc0.
i=1 fikF BL[c0] ≤ 1 for every n ∈ N.
= sup( k
Xj=1
n
(
Xi=1
fi)(x∗
j ) : k ∈ N, x∗
1, . . . , x∗
k ∈ Bℓ1, sup
x∈Bc0
(cid:3)
x∗
j (x) ≤ 1) .
k
Xj=1
Proof. By definition, we have that
Lemma 3.4. kPn
fi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
n
Fix x∗
1, x∗
2, . . . , x∗
disjoint, for each j = 1, 2, . . . , k there is at most one ij ∈ {1, 2, . . . , n} such that fij (x∗
Thus,
k ∈ Bℓ1 with supx∈Bc0Pk
j=1 x∗
j (x) ≤ 1. Since the functions fi are
j ) 6= 0.
k
n
(
Xj=1
Xi=1
fi)(x∗
j ) =
fij (x∗
j ).
k
Xj=1
Without loss of generality, we suppose that fij (x∗
j ) 6= 0 for every j = 1, 2, . . . , k.
10
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
Notice that each fij (x∗
j ) ≤ x∗
jij for every j = 1, 2, . . . , k, so
k
n
(
Xj=1
Xi=1
fi)(x∗
j ) =
k
Xj=1
k
fij (x∗
j ) ≤
x∗
jij .
Xj=1
k asserts that Pk
Now the Claim in Lemma 3.3 applied to x∗
1, . . . , x∗
j=1 x∗
jij ≤ 1 and
therefore
i=1 fikF BL[c0] ≤ 1.
Thus, kPn
k
n
(
Xj=1
Xi=1
fi)(x∗
j ) ≤
k
Xj=1
x∗
jij ≤ 1.
(cid:3)
Lemma 3.5. The sequence (fn)n∈N does not converge in norm to zero. Furthermore,
kfnkF BL[c0] = 1 for every n ∈ N.
Proof. We know that
kfnkF BL[c0] = sup( k
Xj=1
fn(x∗
j ) : k ∈ N, x∗
1, . . . , x∗
k ∈ Bℓ1, sup
x∈Bc0
x∗
j (x) ≤ 1) .
k
Xj=1
Taking e∗
n ∈ Bℓ1 we have that fn(e∗
n) = 1, so that kfnkF BL[c0] ≥ 1. In general, since
fn(x∗) ≤ x∗(en) for every x∗ ∈ Bℓ1, we have that
fn(x∗
j ) ≤
k
Xj=1
k
Xj=1
x∗
j (en) ≤ sup
x∈Bc0
k
Xj=1
x∗
j (x) ≤ 1,
(cid:3)
so kfnkF BL[c0] = 1.
Thus, the Banach lattice c0 is lattice-embeddable into the free Banach lattice F BL[c0]
[4, Theorem 4.50]. For the sake of completeness we include a proof that, indeed, this
embedding has norm one:
Theorem 3.6. The operator u : c0 −→ F BL[c0] given by u(x) = P+∞
x = (x1, x2, . . .) ∈ c0 is a Banach lattice embedding with kuk = 1.
i=1 xifi for every
Proof. It is enough to prove that
n
Xi=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
aifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
= max
1≤i≤n
ai
ON THE BANACH LATTICE c0
11
for every n ∈ N and every a1, a2, . . . , an ∈ R. Fix any a1, a2, . . . , an ∈ R. Without loss
of generality, suppose that a1 = max1≤i≤n ai. Since each function fi is positive and has
norm one, we have that
max
1≤i≤n
n
ai = a1 = ka1f1kF BL[c0] ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
aifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
= (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
= (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n
n
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
aifi(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
Xi=1
aifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
Xi=1
.
On the other hand,
n
Xi=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
aifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
n
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
aifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
n
a1fi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
≤ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
fi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F BL[c0]
= a1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
≤ a1 = max
1≤i≤n
ai.
n
(cid:3)
1), f (e∗
The Banach lattice homomorphism T : F BL[c0] −→ c0 given by the formula T (f ) =
(f (e∗
0 −→ R ∈ F BL[c0] is surjective (notice that F LB[c0]
contains a natural copy of c0 and T c0 is the identity map), so it is a quotient map. Since
T ◦ u = idc0, we have proved the following theorem:
2), . . .) for every f : c∗
Theorem 3.7. The Banach lattice c0 is complemented in F BL[c0].
As a consequence of this fact, that kuk = 1 and that c0 is not projective, applying again
Proposition 2.1, we conclude that:
Theorem 3.8. F BL[c0] is not projective.
References
[1] A. Avil´es, J. D. Rodr´ıguez Abell´an, Projectivity of the free Banach lattice generated by a lattice,
Archiv der Mathematik. To appear.
[2] A. Avil´es, J. Rodr´ıguez, P. Tradacete, The free Banach lattice generated by a Banach space, J. Funct.
Anal. 274 (2018), 2955 -- 2977.
12
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J.D. RODR´IGUEZ ABELL ´AN
[3] B. de Pagter, A. W. Wickstead, Free and projective Banach lattices, Proc. Royal Soc. Edinburgh
Sect. A, 145 (2015), 105 -- 143.
[4] C. D. Aliprantis, O. Burkinshaw, Positive Operators, Handbook of the Geometry of Banach Spaces,
Springer 2006.
Universidad de Murcia, Departamento de Matem´aticas, Campus de Espinardo 30100 Mur-
cia, Spain.
E-mail address: [email protected]
Universidad de Murcia, Departamento de Matem´aticas, Campus de Espinardo 30100 Mur-
cia, Spain.
E-mail address: [email protected]
Universidad de Murcia, Departamento de Matem´aticas, Campus de Espinardo 30100 Mur-
cia, Spain.
E-mail address: [email protected]
|
1008.3326 | 3 | 1008 | 2011-10-28T06:17:43 | On typical properties of Hilbert space operators | [
"math.FA",
"math.GN"
] | We study the typical behavior of bounded linear operators on infinite dimensional complex separable Hilbert spaces in the norm, strong-star, strong, weak polynomial and weak topologies. In particular, we investigate typical spectral properties, the problem of unitary equivalence of typical operators, and their embeddability into C_0-semigroups. Our results provide information on the applicability of Baire category methods in the theory of Hilbert space operators. | math.FA | math |
On typical properties of Hilbert space operators
KdV Institute for Mathematics, University of Amsterdam, P.O. Box 94248, 1090 GE Amsterdam,
Tanja Eisner1
The Netherlands.
Tam´as M´atrai∗
University of Toronto, Mathematics Department BA 6290, 40 St. George St., M5S 2E4 Toronto,
Ontario, Canada.
Abstract
We study the typical behavior of bounded linear operators on infinite dimensional
complex separable Hilbert spaces in the norm, strong-star, strong, weak polynomial
and weak topologies. In particular, we investigate typical spectral properties, the
problem of unitary equivalence of typical operators, and their embeddability into C0-
semigroups. Our results provide information on the applicability of Baire category
methods in the theory of Hilbert space operators.
Keywords: Baire category, typical behavior, Hilbert space, norm topology,
strong-star topology, strong topology, weak polynomial topology, weak topology,
contraction, unitary operator, unitary equivalence, C0-semigroup
2000 MSC: 47A05, 47A10, 47A65, 47D03, 54A10, 54H05
∗The author was partially supported by the OTKA grants K 61600, K 49786 and K 72655,
by the NSERC grants 129977, A-7354, RGPIN 3185-10 and 402762, and by the NSF Grant DMS
0600940.
Email addresses: [email protected] (Tanja Eisner), [email protected] (Tam´as
M´atrai)
URL: http://staff.science.uva.nl/~eisner/ (Tanja Eisner),
http://www.renyi.hu/~matrait (Tam´as M´atrai)
1The author was supported by the European Social Fund and by the Ministry of Science, Re-
search and the Arts Baden-Wurttemberg.
Preprint submitted to Elsevier
October 22, 2018
1. Introduction
Given a property Φ on the points of a Baire space X, we say that a typical point
of X satisfies Φ, or simply that Φ is typical, if the set {x ∈ X : x satisfies Φ} is
co-meager in X, i.e. if {x ∈ X : x does not satisfy Φ} is of first category in X. Many
important and classical results in analysis are concerned with typical properties in
particular topological spaces. Examples include the Banach -- Mazurkiewicz theorem
(see e.g. [2] and [23]) stating that the set of continuous nowhere differentiable func-
tions are residual in (C([0, 1]),k · k∞) (see also [4] for a primer on typical properties
of continuous functions), or the famous result by P. R. Halmos [16] and V. A. Rohlin
[28] in ergodic theory on the existence of weakly mixing but not strongly mixing
transformations.
In this paper we continue the investigations of the first author and study the
typical properties of contractive linear operators on infinite dimensional complex
separable Hilbert spaces in the norm, strong-star, strong, weak polynomial and weak
topologies (for the definitions, see Definition 2.1). Typical properties of various
classes of operators have been studied previously: see e.g. [5] for typical properties
of measure preserving transformations; [3] and [21] for typical mixing properties of
Markov semigroups; [9], [12] and [13] for typical stability properties; [19], [27] and
[30] for typical spectral properties in various very special families of operators. Our
research is different from these works in several respects. We study typical proper-
ties of contractions as a whole, and we carry out our analysis in several topologies.
Surprisingly, in contrast to classical results, we mostly obtain "good" properties as
being typical, and it turns out that the typical properties may change drastically if
the reference topology is changed.
We obtain the following results. In Section 3, we recall some results obtained by
the first author (see [8], [10]) about typical properties in the weak topology. In this
topology, a typical contraction is unitary, it has maximal spectrum and empty point
spectrum, it can be embedded into a C0-semigroup, and typical contractions are not
unitarily equivalent. Our results make use of the theory of typical properties of mea-
sures developed by M. G. Nadkarni [24, Chapter 8] (see also [5]). The importance of
the weak topology in operator theory is an obvious motivation for our investigations.
In Section 4 we consider the weak polynomial topology. Our main observations
are that the contractions endowed with this topology form a Polish space, where the
set of unitary operators is a co-meager subset. Since on the set of unitary operators
the weak and weak polynomial topologies coincide, we conclude that the typical
properties of contractions in the weak and weak polynomial topologies coincide. This
part of our work is motivated by the increasing interest in this unusual topology.
Section 5 treats the strong topology. We show that a typical contraction is
2
unitarily equivalent to the infinite dimensional backward unilateral shift operator. An
analogous result for strongly continuous semigroups is obtained as well. In particular,
the point spectrum of a typical contraction is the open unit disk, typical contractions
are unitarily equivalent and can be embedded into a C0-semigroup. Contrast these
results with the behavior in the weak topology. Just as for the weak topology, our
interest in the strong topology necessitates no clarification.
We study the strong-star topology in Section 6. Our main result gives that the
theory of typical properties of contractions in the strong-star topology can be reduced
to the theories of typical properties of unitary and positive self-adjoint operators
in the strong topology. As a corollary, we obtain that a typical contraction has
maximal spectrum and empty point spectrum, and two typical contractions are not
unitarily equivalent. We included the strong-star topology in our research because
it may play the most important role while extending our investigations into general
Banach⋆-algebras.
Section 7 contains our results on the norm topology, the only non-separable topol-
ogy we consider. We obtain that in this topology there are no such non-trivial typical
structural properties as for the separable topologies. Intuitively, the reason for this
phenomenon is that the norm topology is fine enough to allow for the coexistence of
many different properties on non-meager sets. We close the paper with an outlook
to typical properties of operators on general Banach spaces, and with a list of open
problems.
Before turning our attention to the proofs, let us justify our settings. We choose
contractions as the underlying set of operators because it becomes a Baire space in
all the five topologies we consider. By writing the set of bounded linear operators
as a countable union of scaled copies of the set of contractions, suitable extensions
of our result can be easily obtained. In our investigations of contractions, several
other important classes of operators (e.g. isometries, positive self-adjoint operators,
unitary operators) come into play, and we obtain information about typical properties
in these subclasses, as well.
We work only in infinite dimensional separable Hilbert spaces because removing
any of these assumptions invalidates most of our results.
Infinite dimensionality
guarantees that the families of operators under consideration are sufficiently rich.
Separability is essential for our descriptive set theoretic arguments. The Hilbert
space structure not only allows us to use a well-developed spectral theory, but also
facilitates the construction of operators using orthogonal decomposition.
It is of
limited importance that our Hilbert spaces are over the complex field; analogous
results hold in real Hilbert spaces, as well. We defer the further discussion of possible
extensions of our work until Section 8.
3
To conclude this introduction we note that in research works studying contrac-
tions, it is customary to point out that contractions in general are hard to study.
The theory of contractions as a whole is often contrasted to the theories of normal,
self-adjoint or unitary operators where a satisfactory classification can be obtained,
e.g. via spectral measures (see e.g. [24] and [33]). Our results provide an explana-
tion of this intuitive observation. As we pointed out above, as far as Baire category
methods are concerned, in the four separable topologies we consider, the theory of
contractions is reduced to the theory of unitary operators (weak and weak poly-
nomial topologies), to the theory of one shift operator (strong topology) or to the
theories of unitary and positive self-adjoint operators (strong-star topology). Thus if
the oversimplified pictures captured by these separable topologies are dissatisfactory
for an analyst, then necessarily the very fine norm topology has to be used, in which
case non-separability can be made responsible for being complicated. Since only such
properties can be studied using Baire category arguments which are non-trivial in
a suitable Baire topology, i.e. which hold at least on a non-meager set, our results
outline a limitation of Baire category methods in operator theory.
Acknowledgment. We thank Mariusz Lema´nczyk and Nicolas Monod for help-
ful discussions and for calling our attention to the problem of unitary equivalence of
typical operators, as well as the referee for careful reading and helpful suggestions.
2. Preliminaries
As general references, see [20] for descriptive set theory, [33] for functional analy-
sis, and [15] for semigroup theory. Recall that a topological space X is a Baire space
if every non-empty open set in X is non-meager, or equivalently if the intersection
of countably many dense open sets in X is dense (see e.g. [20, (8.1) Proposition
and (8.2) Definition p. 41]). Polish spaces, i.e. separable complete metric spaces, are
well-known examples of Baire spaces.
A set in a topological space is Gδ if it can be obtained as an intersection of
countably many open sets. We will often use the observation that in every topological
space, every dense Gδ set is co-meager.
Let X be a Baire space and let Φ be a property on the points of X. We say
that Φ is a typical property on X, or that a typical element of X satisfies Φ if
{x ∈ X : x satisfies Φ} is a co-meager subset of X. Note that if Φn (n ∈ N) are typical
properties on X then a typical element of X satisfies all Φn (n ∈ N) simultaneously.
In the sequel (H,k · k) always denotes an infinite dimensional complex sepa-
rable Hilbert space. The scalar product on H is denoted by h·,·i. For every
U ⊆ H, span{U} denotes the linear subspace of H generated by U, and U ⊥ =
4
{x ∈ H : hx, ui = 0 (u ∈ U)}. For every U, V ⊆ H, we write U⊥V if hu, vi = 0
(u ∈ U, v ∈ V ).
If V ≤ H is a subspace, we define BV = {v ∈ V : kvk ≤ 1},
SV = {v ∈ V : kvk = 1}.
Let B(H), C(H), U(H) and P (H) denote the sets of bounded, contractive, uni-
tary and contractive positive self-adjoint linear H → H operators. The identity
operator is denoted by Id. For every V ≤ H, the orthogonal projection onto V is
denoted by PrV . For every A ∈ B(H) the adjoint of A is denoted by A⋆. For every
A ∈ B(H) and U ⊆ H we set A[U] = {Ax : x ∈ U}, ker A = {x ∈ H : Ax = 0} and
Ran A = A[H]. For every A ∈ B(H), let O(A) = {UAU −1 : U ∈ U(H)}.
For every A ∈ B(H) the spectrum, the point spectrum, the continuous spectrum,
and the residual spectrum of A is denoted by σ(A), Pσ(A), Cσ(A), and Rσ(A) (see
e.g. [33, Definition p. 209]).
We recall the definitions and some elementary properties of the weak, weak poly-
nomial, strong and strong-star topologies (see e.g. [33, Definition 3 p. 112], [25,
Definition p. 1142], [33, Definition p. 69], and [34, Definition p. 220]).
Definition 2.1. Let A, An ∈ B(H) (n ∈ N) be arbitrary.
1. We say that {An : n ∈ N} converges to A weakly, A = w-limn∈N An in notation,
if for every x, y ∈ H, limn∈NhAnx, yi = hAx, yi. The topology corresponding
to this notion of convergence is called the weak topology. Topological notions
referring to the weak topology are preceded by w-.
2. We say that {An : n ∈ N} converges to A weakly polynomially, A = pw-
n = Ak. The topol-
limn∈N An in notation, if for every k ∈ N, w-limn∈N Ak
ogy corresponding to this notion of convergence is called the weak polynomial
topology. Topological notions referring to the weak polynomial topology are
preceded by pw-.
3. We say that {An : n ∈ N} converges to A strongly, A = s-limn∈N An in notation,
if for every x ∈ H, limn∈N Anx = Ax. The topology corresponding to this
notion of convergence is called the strong topology. Topological notions referring
to the strong topology are preceded by s-.
4. We say that {An : n ∈ N} converges to A in the strong-star sense, A = s⋆-
n = A⋆. The topology
limn∈N An in notation, if s-limn∈N An = A and s-limn∈N A⋆
corresponding to this notion of convergence is called the strong-star topology.
Topological notions referring to the strong-star topology are preceded by s⋆-.
Proposition 2.2. ([32, Section 2 p. 67], [34, Section 7.f p. 121]) Let {ei : i ∈ N} ⊆
H be an orthonormal basis.
5
1. For every A, B ∈ B(H), set
dw(A, B) = Xi,j∈N
2−i−jhAei, eji − hBei, eji.
Then dw is a complete separable metric on C(H) which generates the weak
topology.
2. For every A, B ∈ B(H), set
ds(A, B) =Xi∈N
2−ikAei − Beik.
Then ds is a complete separable metric on C(H) which generates the strong
topology.
The strong-star topology on uniformly bounded sets is generated by the metric
ds⋆(A, B) = ds(A, B) + ds(A⋆, B⋆).
It is easy to see that ds⋆ is a complete met-
ric on C(H). We will see in Section 4 that C(H) endowed with the weak polynomial
topology is a Polish space as well.
Note that the weak, weak polynomial, strong, strong-star, and norm topologies
all refine the preceding topologies in this list. We also need the following.
Proposition 2.3. ([32, Remark 4.10 p. 84]) On U(H) the weak, weak polynomial,
strong, and strong-star topologies coincide. With this topology, U(H) is a Polish
space.
3. The weak topology
In the weak topology, the theories of typical properties of contractions and of
unitary operators coincide.
Theorem 3.1. ([8, Theorem 2.2 p. 2]) A w-typical contraction is unitary.
Equivalently, Theorem 3.1 says that U(H) is a w-co-meager subset of C(H).
By Proposition 2.3, U(H) endowed with the weak topology is also a Polish space.
Hence the notions related to Baire category make sense relative to U(H), and a set
A ⊆ C(H) is w-co-meager in C(H) if and only if A∩ U(H) is w-co-meager in U(H).
Unitary operators are well-understood. E.g. the theory of spectral measures al-
lows a detailed description of the spectral properties and conjugacy classes of unitary
6
operators. We refer to [33, Chapter XI.4 p. 306] and [24, Chapter 2 p. 17] for an
introduction to spectral measures. The following proposition briefly summarizes how
the spectral properties of w-typical contractions can be obtained from the theory of
spectral measures. We set S1 = {λ ∈ C : λ = 1}. For a measure µ on S1, supp µ
denotes the closed support of µ. For measures µ, ν on S1, we write µ⊥ν if µ and ν
are mutually singular.
Proposition 3.2. A w-typical contraction U satisfies Cσ(U) = S1.
Proof. Let U be a w-typical contraction. By Theorem 3.1, U is unitary. Let µU
denote the maximal spectral type of U (see [24, Section 8.22 p. 55]), which is a Borel
measure on S1. E.g. by the uniqueness of the measure class of µU , λ ∈ σ(U) if and
only if λ ∈ supp µU , and λ ∈ Pσ(U) if and only if λ is an atom of µU .
By [24, 8.25 Theorem (a) p. 56],
{U ∈ U(H) : Pσ(U) = ∅} = {U ∈ U(H) : µU is atomless}
is a w-co-meager set in U(H). By a similar argument, using [24, 7.7 Corollary p. 46],
we obtain that
{U ∈ U(H) : σ(U) = S1} = {U ∈ U(H) : supp µU = S1}
is also a w-co-meager set in U(H). Since σ(U) = Pσ(U) ∪ Cσ(U) ∪ Rσ(U) and
Rσ(U) = ∅ for every unitary operator, we conclude that a w-typical contraction U
satisfies Cσ(U) = S1, as required.(cid:4)
By analogous applications of spectral measures, one can isolate numerous addi-
tional w-typical properties of contractions (see e.g. [24, 8.25 Theorem p. 56]). We
refer to e.g. [11, Section IV.3] for asymptotic properties of w-typical contractions,
and mention that P. Zorin [35] showed recently that a w-typical contraction admits
a fixed cyclic vector. Here we restrict ourselves to pointing out that despite the
abundance of w-typical properties, typical contractions are not unitarily equivalent.
Proposition 3.3. For every U ∈ C(H), O(U) is w-meager in C(H). In particular,
for a w-typical pair of contractions (U1, U2) ∈ C(H)× C(H) we have that U1 and U2
are not unitarily equivalent.
Proof. Let U ∈ C(H) be arbitrary. By Theorem 3.1, the statement follows if
U /∈ U(H). So for the rest of the argument, we can assume U ∈ U(H). By the
conjugacy invariance of the measure class of the maximal spectral type, for every
7
V ∈ U(H) we have that µU and µV U V −1 are mutually absolutely continuous. By [24,
8.25 Theorem (b) p. 56], for every measure ν on S1, the set {V ∈ U(H) : µV ⊥ν} is
w-co-meager in U(H). Thus O(U) is w-meager in U(H) and so in C(H), as well.
Finally consider the set
E = {(U1, U2) ∈ C(H) × C(H) : U1, U2 are unitarily equivalent}.
The set E is clearly analytic hence has the Baire property (see e.g. [20, (29.14)
Corollary p. 229]). As we have seen above, for every U1 ∈ C(H) we have {U2 ∈
C(H) : (U1, U2) ∈ E} is w-meager in C(H). So by the Kuratowski-Ulam theorem
(see e.g. [20, (29.14) Corollary p. 229]) we get that E is w-meager in C(H) × C(H).
This completes the proof.(cid:4)
4. The weak polynomial topology
The main result of this section is the following.
Theorem 4.1. The set C(H) endowed with the weak polynomial topology is a Polish
space. Moreover, U(H) is a pw-co-meager pw-Gδ subset of C(H).
This result immediately implies that the theories of pw-typical and w-typical
properties of contractions coincide.
Corollary 4.2. A set C ⊆ C(H) is pw-co-meager in C(H) if and only if C ∩ U(H)
is w-co-meager in U(H). In particular, a property Φ of contractions is pw-typical if
and only if Φ is w-typical.
Proof. By Theorem 4.1, U(H) is a pw-co-meager subset of the Polish space C(H).
By Proposition 2.3, the weak and weak polynomial topologies coincide on U(H)
and in these topologies U(H) is a Polish space. So the notions related to Baire
category make sense relative to U(H). By U(H) ⊆ C(H) being pw-co-meager, a set
M ⊆ U(H) is pw-meager in U(H) if and only if M ⊆ C(H) is pw-meager in C(H).
We obtained that C ⊆ C(H) is pw-co-meager in C(H) if and only if C ∩ U(H) ⊆
C(H) is pw-co-meager in C(H). This is equivalent to C ∩ U(H) ⊆ U(H) being pw-
co-meager in U(H), which is the same as C ∩ U(H) ⊆ U(H) being w-co-meager in
U(H). Finally by Theorem 3.1, this is equivalent to C ⊆ C(H) being w-co-meager
in C(H). This completes the proof.(cid:4)
By Corollary 4.2, for the pw-typical properties of contractions one can refer to
Section 3.
To prove the first part of Theorem 4.1 we need the following lemmas on the weak
topology.
8
Lemma 4.3. Let A, An ∈ B(H) (n ∈ N) satisfy A = w-limn∈N An. If for every
x ∈ SH we have kAxk ≥ lim supn∈N kAnxk then A = s-limn∈N An.
Proof. For x ∈ SH we have
kAx − Anxk2 = kAxk2 + kAnxk2 − 2RehAx, Anxi,
so
0 ≤ lim supn∈N kAx − Anxk2 = kAxk2 + lim supn∈N kAnxk2−
2 limn∈N RehAx, Anxi ≤ 2kAxk2 − 2RehAx, Axi = 0.
Since x ∈ SH was arbitrary, A = s-limn∈N An follows. (cid:4)
Lemma 4.4. Let n > 0, {xi : i < n} ⊆ SH and B ∈ C(H) be arbitrary. Then for
every ε > 0 there exists a w-open set W ⊆ C(H) such that B ∈ W and for every
A ∈ W we have kAxik ≥ kBxik − ε (i < n).
Proof. We prove the statement for n = 1 only; for n > 1 the required set W
can be obtained by intersecting the sets Wi satisfying the conditions of the lemma
for each xi (i < n) separately.
If Bx0 = 0 the statement is trivial; so we can assume Bx0 6= 0. Consider the set
W = {A ∈ C(H) : hAx0, Bx0i > kBx0k2 − kBx0kε};
then W is w-open and B ∈ W . For every A ∈ W we have
≥ kBx0k − ε
kAx0k ≥ hAx0, Bx0i
kBx0k
as required. (cid:4)
Proof of Theorem 4.1. By [8, Theorem 2.2], U(H) is a w-co-meager w-Gδ
subset of C(H). Since the weak polynomial topology is finer than the weak topology,
U(H) ⊆ C(H) is pw-Gδ, as well. By [25, Theorem 1 p. 1142], U(H) is pw-dense in
C(H), thus U(H) is pw-co-meager in C(H).
To prove that C(H) is Polish in the weak polynomial topology we use [20, (8.18)
Theorem p. 45] stating that a non-empty, second countable topological space is Polish
if and only if it is T1, regular and strong Choquet (for the definition, see [20, (8.14)
Definition p. 44.]). Since (C(H), pw) is a metric space, it is second countable, T1 and
regular. So in order to show it is Polish, we have to prove it is strong Choquet, i.e.
that player II has a winning strategy in the strong Choquet game.
9
We define the strategy for player II as follows. As in Proposition 2.2, let dw
denote the complete metric on C(H) which induces the weak topology on C(H). Let
{xn : n ∈ N} be a dense subset of SH. Suppose the nth move of player I is (An, Un),
where An ∈ Un ⊆ C(H) and Un is a pw-open set. Let Wn be a w-open set with the
following properties:
1. An ∈ Wn ⊆ clw(Wn) ⊆ {A : dw(A, An) < 1/(n + 1)};
2. clw(Wn) ⊆ Wn−1;
3. for every A ∈ Wn we have kAxik ≥ kAnxik − 1/(n + 1) (i ≤ n);
such a set exists by Lemma 4.4. Then let player II respond by playing a pw-open set
Vn ⊆ Un such that
4. An ∈ Vn ⊆ clpw(Vn) ⊆ Un ∩ Wn.
We show that this strategy is winning for player II. Let {(An, Un), Vn : n ∈ N}
be a run in the game in which player II follows the above strategy. By An+1 ∈
Vn ⊆ Wn (n ∈ N) and conditions 1 and 2, An is weakly convergent, say A = w-
limn∈N An. Again by conditions 1 and 2, A ∈ Wn (n ∈ N); thus by condition 3,
kAxik ≥ lim supn∈N kAnxik (i ∈ N). Since {xn : n ∈ N} ⊆ SH is dense and A, An
(n ∈ N) are contractions, we get kAxk ≥ lim supn∈N kAnxk (x ∈ SH). Thus by
in particular, A = pw-limn∈N An. Since
Lemma 4.3, we have A = s-limn∈N An;
A ∈ clpw(Vn) ⊆ Vn−1 for each n by condition 4, we get A ∈ Tn∈N Vn. This shows
that the strategy is winning for player II, and finishes the proof. (cid:4)
Note that, by the same argument as in the above proof, C(H) is Polish for any
metrizable topology on C(H) which is finer than the weak topology and coarser
than the strong one. In addition, for such topology U(H) is a co-meager Gδ subset
of C(H) if and only if U(H) is dense in C(H) in this topology.
5. The strong topology
Probably the most surprising observation in the present paper is that some typical
properties of contractions in the strong and weak topologies are completely different.
While typical contractions in the weak topology are not unitarily equivalent, typical
contractions in the strong topology are unitarily equivalent to an infinite dimensional
backward unilateral shift operator, hence the investigation of s-typical properties of
contractions is reduced to the study of one particular operator. We introduce this
operator in the following definition. Since infinite dimensional complex separable
10
Hilbert spaces are isometrically isomorphic, we can restrict ourselves to the study of
a particular one.
Definition 5.1. Set H = ℓ2(N × N) and denote the canonical orthonormal basis of
H by {ei(n) : i, n ∈ N}. We define the infinite dimensional backward unilateral shift
operator S ∈ C(H) by Se0(n) = 0 and Sei+1(n) = ei(n) (i, n ∈ N).
The main result of this section is the following. Recall O(A) = {UAU −1 : U ∈
U(H)}.
Theorem 5.2. The set O(S) is an s-co-meager subset of C(H).
By Theorem 5.2, a property of contractions is s-typical if and only if S has this
property. In the following corollary, we recall only the properties of S we usually
concern in this paper.
Corollary 5.3. An s-typical contraction A satisfies that
1. Pσ(A) = {λ ∈ C : λ < 1}, and for every λ ∈ Pσ(A) we have dim ker(λ · Id −
A) = ∞;
2. Cσ(A) = S1;
3. A can be embedded into a strongly continuous semigroup.
Moreover, typical contractions are unitarily equivalent.
Proof. Statements 1 and 2 follow from Theorem 5.2 and the results of [29], while 3
is a corollary of [8, Proposition 4.3]. (cid:4)
Our strategy to prove Theorem 5.2 is the following. The main observation is
that for an s-typical contraction A, its adjoint A⋆ is an isometry. Then by the Wold
decomposition theorem (see e.g. [31, Theorem 1.1 p. 3]), A is unitarily equivalent
to a direct sum of unitary and backward unilateral shift operators, and the number
of shifts in the direct sum depends on the dimension of ker A. Since an s-typical
contraction A is strongly stable, the unitary part is trivial. So we complete the proof
by showing dim ker A = ∞.
11
5.1. Elementary observations
We collect here some elementary results we need later in our analysis.
Lemma 5.4. Let x, y ∈ SH satisfy x 6= −y. Set α = (2 + 2 · Rehx, yi)−1/2. Then
kα(x + y)k = 1.
Proof. We have kα(x+y)k2 = α2(kxk2+kyk2+2·Rehx, yi) = α2(2+2·Rehx, yi) = 1,
as required.(cid:4)
Lemma 5.5. Let A ∈ C(H) and let (bn)n∈N ⊆ SH , z ∈ SH satisfy limn∈N Abn = z.
Then (bn)n∈N is convergent.
Proof. It is enough to prove that for every ε > 0 there is an N ∈ N such that for
every n, m ≥ N we have kbn − bmk2 ≤ ε. So let ε > 0 be arbitrary. For every n, m
sufficiently large we have bn 6= −bm, so αn,m = (2 + 2 · Rehbn, bmi)−1/2 is defined. By
Lemma 5.4 we have kαn,m(bn + bm)k = 1. So since A is a contraction,
αn,m · (2 − kAbn − zk − kAbm − zk) ≤
αn,m · k2 · z + (Abn − z) + (Abm − z)k =
(1)
Let N ∈ N be such that for every n ≥ N we have kAbn − zk ≤ ε/8. Then by (1),
for every n, m ≥ N we get αn,m ≤ 1/(2− 2ε/8), i.e. 2 + 2· Rehbn, bmi ≥ 4· (1− ε/8)2.
Thus
αn,m · kAbn + Abmk = kαn,m · A(bn + bm)k ≤ 1.
kbn − bmk2 = 2 − 2 · Rehbn, bmi ≤ 4 − 4 · (1 − ε/8)2 ≤ ε,
as required.(cid:4)
The following lemma will be helpful to show that the kernel of a typical contrac-
tion is infinite dimensional.
Lemma 5.6. Let n ∈ N \ {0} and let {ei : i < n} ⊆ H be an orthonormal family.
Let {fi : i < n} ⊆ H satisfy kfi − eik < 1/n (i < n). Then {fi : i < n} are linearly
independent.
Proof. Let αi ∈ C (i < n) be arbitrary satisfying Pi<n αi > 0. We have
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi<n
αiei −Xi<n
≤ Xi<n
αifi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2
2
<
αikei − fik!2
αi!2
1
nXi<n
≤
1
nXi<n
αi2 =
.
1
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi<n
αiei(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
12
So by (cid:13)(cid:13)Pi<n αiei(cid:13)(cid:13) > 0 we have (cid:13)(cid:13)Pi<n αifi(cid:13)(cid:13) > 0, as required.(cid:4)
Next we point out a trivial sufficient condition for the direct sum of contractions
to be a contraction.
Lemma 5.7. Let V0, V1 ≤ H be subspaces satisfying V0⊥V1. Let Ai : Vi → H (i < 2)
be contractive linear operators such that A0[V0]⊥A1[V1]. Then A : span{V0, V1} → H,
A(αv0 + βv1) = αA0v0 + βA1v1 (α, β ∈ C)
is also contractive.
Proof. Let v ∈ span{V0, V1}, kvk = 1 be arbitrary. Then there are α, β ∈ C with
α2 + β2 = 1 and vi ∈ SVi (i < 2) such that v = αv0 + βv1. We have
kAvk2 = kA(αv0 + βv1)k2 = kαA0v0 + βA1v1k2 =
α2kA0v0k2 + β2kA1v1k2 ≤ α2 + β2 = 1,
so the proof is complete.(cid:4)
Finally we point out that strongly stable contractions form a s-co-meager s-Gδ
subset of C(H).
Definition 5.8. A contraction A is strongly stable if s-limn∈N An = 0, i.e., for every
x ∈ SH and ε > 0 there is an n ∈ N such that kAnxk < ε. The set of strongly stable
contractions is denoted by S.
Lemma 5.9. The set of strongly stable contractions is an s-co-meager s-Gδ subset
of C(H).
Proof. By A = limn∈N(1 − 2−n)A (A ∈ C(H)), the set of contractions A satisfying
kAk < 1 is a norm dense and hence an s-dense subset of C(H). Since every such
operator is strongly stable, it remains to show that S is s-Gδ. To this end, let
{xi : i ∈ N} be a dense subset of SH . Note that for each x ∈ H and A ∈ S,
the sequence kAnxk (n ∈ N) is monotonically decreasing, so limn∈N kAnxk = 0 is
equivalent to inf n∈N kAnxk = 0. Thus
which completes the proof.(cid:4)
S =Ti,j∈NSn∈N{A ∈ C(H) : kAnxik < 2−j},
(2)
We remark that Lemma 5.9 holds in every separable Banach space. It is interest-
ing to note the difference to the weak operator topology in which the set of weakly
stable contractions is w-meager in C(H) (see [12, Theorem 4.3]).
13
5.2. Mapping properties of s-typical contractions
The purpose of this section is to prove the following.
Theorem 5.10. Let G denote the set of contractive operators A satisfying the fol-
lowing properties:
1. for every y ∈ SH there exists an x ∈ SH such that Ax = y, i.e., SH ⊆ A[SH ];
2. dim Ker A = ∞.
Then G is an s-co-meager subset of C(H).
To prove Theorem 5.10, we need a geometric lemma saying that every contraction
defined on a finite dimensional subspace of H can be extended to a contraction which
is surjective in a very strong sense.
Lemma 5.11. Let V ≤ H be a finite dimensional subspace and let A : V → H
be a contractive linear operator. Let W = A[V ] and let Y ≤ H be an arbitrary
subspace satisfying Y ⊥W . Then for every X ≤ H satisfying X⊥V and dim X =
dim W + dim Y there exists a contractive linear operator A : span{V, X} → H such
that AV = A and A[Bspan{V,X}] = Bspan{W,Y }.
Proof. We handle first the special Y = {0} case. Let dim V = n and dim W = m.
The Gram-Schmidt orthogonalization theorem states that there exists an orthonor-
mal base {vi : i < n} ⊆ V such that {Avi : i < n} ⊆ W are pairwise orthogonal,
and
(i) kAv0k = max{kAvk : v ∈ SV };
(ii) for every j < n − 1, kAvj+1k = max{kAvk : v ∈ SV ∩ span{vi : i ≤ j}⊥}.
By (i) and (ii) we have kAvik > 0 if and only if i < m. Set wi = Avi/kAvik (i < m).
Let X ≤ H satisfy X⊥V and dim X = m. Fix an orthonormal base {xi : i < m}
in X and define A : span{V, X} → H such that AV = A and Axi =p1 − kAvik2wi
First we show A is a contraction. Let u ∈ span{V, X} satisfy kuk = 1. Then
u =Pi<n αivi +Pi<m βixi where αi ∈ C (i < n), βi ∈ C (i < m) satisfyPi<n αi2 +
(i < m).
14
x =Xi<m
kxk =Xi<m(cid:16)β2
βi(kAvikvi +p1 − kAvik2xi).
ip1 − kAvik2
i kAvik2 + β2
2(cid:17) ≤ 1,
Then
and
Ax =Xi<m
2
=
αivi +Xi<m
Pi<m βi2 = 1. Then
k Auk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
βixi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
A Xi<n
(αikAvikwi + βip1 − kAvik2wi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi<m
(αikAvik + βip1 − kAvik2)2
By the Cauchy-Schwarz inequality, for every 0 ≤ p, q, r ≤ 1 we have (pr+q√1 − r2)2 ≤
p2 + q2. So k Auk ≤Pi<m αi2 + βi2 ≤ 1, as required.
Next we show that for every y ∈ BW there is an x ∈ Bspan{V,X} such that Ax = y.
Let y =Pi<m βiwi where βi ∈ C (i < m) satisfy Pi<m βi2 ≤ 1. Let
≤Xi<m
2
βi(kAvikAvi +p1 − kAvik2 Axi) =
Xi<m
βi(kAvikkAvikwi +p1 − kAvik2p1 − kAvik2wi) = y,
as required. This completes the proof of the special Y = {0} case.
In the general case write X = X0 ⊕ X1 where X0⊥X1 and dim X0 = dim W ,
dim X1 = dim Y . By the special case above, there is a contraction A : span{V, X0} →
H such that AV = A and A[Bspan{V,X0}] = BW . Extend further A by setting
AX1 : X1 → Y be any isometric isomorphism. By Lemma 5.7, A is a contraction
which clearly satisfies A[Bspan{V,X}] = Bspan{W,Y }. This completes the proof.(cid:4)
From Lemma 5.11, we immediately get the following.
Proposition 5.12. The set of contractive operators A such that for every y ∈ SH
there exists an x ∈ SH such that Ax = y is an s-co-meager subset of C(H).
15
Proof. Let
M = {A ∈ C(H) : ∀ε > 0, y ∈ SH ∃x ∈ SH (ky − Axk < ε)}.
First we show that M is an s-dense s-Gδ subset of C(H).
Fix y ∈ SH and ε > 0. The set
C(y, ε) = {A ∈ C(H) : ∃x ∈ SH (ky − Axk < ε)}
is s-open. We show that it is s-dense.
Let U ⊆ C(H) be any non-empty s-open set. By passing to a subset, we can
assume U = {A ∈ C(H) : kyi − Axik < εi (i ∈ I)} where xi, yi ∈ H, εi > 0 (i ∈ I)
and I is finite. Let V = span{xi : i ∈ I} and take an arbitrary A ∈ U. By restricting
A to V we can assume AV ⊥ = 0. Set W = A[V ].
Let Y ≤ H be an at most one dimensional subspace such that Y ⊥W and y ∈
span{W, Y }. Let X ≤ H be a dim W + dim Y dimensional subspace such that
X⊥V . By Lemma 5.11, there exists a contraction A : span{V, X} → H such that
AV = AV and A[Bspan{V,X}] = Bspan{W,Y }. In particular, there is an x ∈ Bspan{V,X}
such that Ax = y. Since A is a contraction and kyk = 1, we get x ∈ SH. Extend
further A by setting Aspan{V,X}⊥ = 0. Then A ∈ U ∩ C(y, ε); i.e. we concluded that
C(y, ε) is s-dense.
Let D ⊆ SH be a countable dense set. We have
M =T{C(y, 2−n) : y ∈ D, n ∈ N},
so by the Baire category theorem, M is an s-dense s-Gδ subset of C(H).
It now suffices to show that every A ∈ M satisfies SH ⊆ A[SH ]. To this end,
let A ∈ M and z ∈ SH be arbitrary. By the definition of M, there is a sequence
(bn)n∈N ⊆ SH such that limn∈N Abn = z. By Proposition 5.5, (bn)n∈N is convergent,
say limn∈N bn = x. Then Ax = z, which completes the proof. (cid:4)
To prove that a typical contraction has infinite dimensional kernel, we need a
lemma showing that a typical contraction approximates the zero operator on arbi-
trarily large finite dimensional subspaces.
Lemma 5.13. The set of contractive operators A such that for every n ∈ N and
ε > 0 there exists Z ≤ H with dim Z ≥ n and kAZk < ε is an s-dense s-Gδ subset
of C(H).
Proof. For every n ∈ N and ε > 0, the set
C(n, ε) = {A ∈ C(H) : ∃Z ≤ H (dim Z ≥ n, kAZk < ε)}
16
is s-open. We show that C(n, ε) (n ∈ N, ε > 0) are s-dense.
Fix arbitrary n ∈ N and ε > 0. Let U ⊆ C(H) be any non-empty s-open set.
By passing to a subset, we can assume U = {A ∈ C(H) : kyi − Axik < εi (i ∈ I)}
where xi, yi ∈ H, εi > 0 (i ∈ I) and I is finite. Let A ∈ U be arbitrary. Let
V = span{xi : i ∈ I}, and define B ∈ C(H) by BV = AV , BV ⊥ = 0. Since
dim V ⊥ = ∞, we obtained B ∈ C(n, ε) ∩ U; i.e. we concluded C(n, ε) is s-dense.
The set of contractive operators A such that for every n ∈ N and ε > 0 there
exists Z ≤ H with dim Z ≥ n and kAZk < ε is
So by the Baire category theorem, this is an s-dense s-Gδ subset of C(H), which
completes the proof.(cid:4)
T{C(n, 2−m) : n, m ∈ N}.
We are ready to prove the second part of Theorem 5.10.
Proposition 5.14. The set of contractions A which satisfy dim ker A = ∞ is an
s-co-meager subset of C(H).
Proof. By Lemma 5.13 and Proposition 5.12, the set of contractions A which satisfy
that
1. for every n ∈ N and ε > 0 there exists Z ≤ H with dim Z ≥ n and kAZk <
ε/n;
2. for every y ∈ SH there exists an x ∈ SH such that Ax = y;
is an s-co-meager subset of C(H). We show that every member A of this set satisfies
dim ker A = ∞; this will complete the proof.
Fix an arbitrary n ∈ N \ {0}. Set ε = 1 and let Z ≤ H satisfy 1 for this n and
ε. Let {ei : i < n} ⊆ Z be an orthonormal family. We find fi ∈ H (i < n) such
that kfi − eik < 1/n (i < n) and Afi = 0. Then by Lemma 5.6, {fi : i < n} are
linearly independent. Then dim ker A ≥ n, and since n was arbitrary, we concluded
dim ker A = ∞.
Fix i < n; we define fi as follows. If Aei = 0, set fi = ei. Else observe that by 2,
there exists xi ∈ SH with Axi = Aei/kAeik. Define now
fi = ei − kAeikxi.
Then kfi − eik = kAeik < 1/n and we have Afi = Aei − kAeikAxi = 0, so the
construction is complete.(cid:4)
Proof of Theorem 5.10. The statement follows from Proposition 5.12 and
Proposition 5.14.(cid:4)
17
5.3. Unitary equivalence of s-typical contractions to a shift
Operators in the set G introduced in Theorem 5.10 have the following properties.
Proposition 5.15. Let A ∈ G. Then AA⋆ = Id and A⋆A is the projection onto
Ran A⋆, which is an infinite dimensional and infinite co-dimensional subspace of H.
In particular, A⋆ is an isometry, hence A is a co-isometry and in addition, A is an
isometry on (Ker A)⊥.
Proof. Let {ei : i ∈ N} be an orthonormal basis of H. By the definition of G, for
every i ∈ N there is an ai ∈ SH such that Aai = ei. Note that for every i ∈ N,
1 = hei, eii = hAai, Aaii = hA⋆Aai, aii. By A, A⋆ being contractions, this is possible
only if A⋆ei = A⋆Aai = ai (i ∈ N), thus AA⋆ei = Aai = ei (i ∈ N). This proves
that AA⋆ = Id, Ran A⋆ = span{A⋆ei : i ∈ N} = span{ai : i ∈ N}, and that A⋆A is the
projection onto Ran A⋆. Again by A, A⋆ ∈ C(H), this implies that A⋆ is an isometry,
and A is an isometry on Ran A⋆. By the definition of G, Ran A⋆ = (Ker A)⊥ is infinite
dimensional and infinite co-dimensional. This completes the proof. (cid:4)
Proof of Theorem 5.2. By Lemma 5.9 and Theorem 5.10, it is enough to show
that every A ∈ S ∩ G is unitarily equivalent to the operator S of Definition 5.1. By
Proposition 5.15, A is a co-isometry. So by the Wold decomposition theorem (see
e.g. [31, Theorem 1.1 p. 3]), we have H = Hu ⊕ Hs such that AHu is unitary and
AHs is unitarily equivalent to the backward unilateral shift operator on l2(N, ker A),
i.e. the Hilbert space of square summable N → ker A functions. By A ∈ S we have
Hu = {0}. By Theorem 5.10.2, dim ker A = ∞ hence l2(N, ker A) is isometrically
isomorphic to ℓ2(N × N). This completes the proof.(cid:4)
5.4. The continuous case
In this section we show that a typical strongly continuous contraction semigroup
on H is unitarily equivalent to an infinite dimensional backward unilateral shift
semigroup.
Let C c(H) denote the set of contractive C0-semigroups on H. Here we endow
C c(H) with the topology induced by the uniform strong convergence on compact
time intervals, i.e., by the metric
dc
s(T (·), S(·)) =Pj,n∈N 2−(j+n) supt∈[0,n] kT (t)ej − S(t)ejk,
where {ej : j ∈ N} is an orthonormal basis of H. With respect to this topology,
C c(H) is a Polish space. With an abuse of notation, topological notions referring to
this topology are also preceded by s-.
18
We again restrict ourselves without loss of generality to a particular infinite di-
mensional complex separable Hilbert space and introduce on it the infinite dimen-
sional backward unilateral shift semigroup.
Definition 5.16. Set H = L2(R2
unilateral shift semigroup S(·) ∈ C c(H) by
[S(t)f ](s, w) = f (s + t, w)
(t, s, w ∈ R+, f ∈ H).
+). We define the infinite dimensional backward
We also set O(S(·)) = {(U −1S(t)U)t∈R+ : U ∈ U(H)} ⊆ C c(H).
The following shows that an s-typical contraction semigroup is unitarily equiva-
lent to S(·).
Theorem 5.17. The set O(S(·)) is an s-co-meager subset of C c(H).
Note that for every t > 0, S(t) is unitarily equivalent to the backward unilateral
shift operator S of Definition 5.1. So by Theorem 5.17, s-typical contractive C0-
semigroups T (·) satisfy that for every t > 0, the operator T (t) has the same properties
as S. Moreover, s-typical contraction semigroups are unitarily equivalent.
To prove Theorem 5.17, we need to introduce the following not so well-known
concept from semigroup theory.
Definition 5.18. Let T (·) be a C0-semigroup on H. Let the generator A of T (·)
satisfy 1 ∈ ρ(A). Then the operator V ∈ B(H) defined by
V = (A + Id)(A − Id)−1 = Id + 2(A − Id)−1
is called the cogenerator of T (·).
The cogenerator is a bounded operator which determines the semigroup uniquely.
Moreover, it shares many properties of T (·) such as being contractive, unitary, self-
adjoint, normal, isometric, strongly stable etc. (see [31, Section III.8-9] for details).
We will use the following (see e.g. [31, Theorem III.8.1]).
Lemma 5.19. With the notation of Definition 5.18, V is the cogenerator of a con-
traction semigroup if and only if V is contractive and satisfies 1 /∈ Pσ(V ).
Our key observation is the following.
19
Lemma 5.20. Set
V = {V ∈ C(H) : 1 /∈ Pσ(V )},
endowed with the strong topology. Define J : C c(H) → V by J(T (·)) = V where V is
the cogenerator of T (·) (T (·) ∈ C c(H)). Then J is a homeomorphism.
Proof. The statement follows from Lemma 5.19 and the First Trotter -- Kato Theo-
rem (see e.g. [15, Theorem III.4.8]).(cid:4)
Proof of Theorem 5.17. By Corollary 5.3, V is s-co-meager in C(H). So with
the notation of (2) and Theorem 5.10, by Lemma 5.9 and Theorem 5.10 we have
that S ∩ G ∩ V is s-co-meager in V. Hence the set J −1(S ∩ G ∩ V) is s-co-meager in
C c(H). So it is enough to show that J −1(S ∩ G ∩ V) ⊆ O(S(·)).
Let T (·) ∈ J −1(S∩G∩V) be arbitrary. Let V denote the cogenerator of T (·); then
V ∈ S ∩ G ∩ V. By Proposition 5.15, V is a strongly stable co-isometry. Since the
cogenerator of T ⋆(·) is V ⋆, and since the semigroup and the cogenerator share strong
stability and the isometric property (see [31, Theorem III.9.1]), T (·) is a strongly
stable co-isometric contraction semigroup. By Wold's decomposition for semigroups
(see [31, Theorem III.9.3]), T (·) is unitarily equivalent to the backward unilateral
shift semigroup on L2(R+, Y ), where Y = (Ran V ⋆)⊥. Since (Ran V ⋆)⊥ = Ker V is
infinite dimensional, the statement follows. (cid:4)
6. The strong-star topology
As we have seen in the previous section, the theory of s-typical properties of
contractions is reduced to the study of one particular operator. The reason behind
this phenomenon is that s-convergence does not control the adjoint, i.e. the function
A 7→ A⋆ is not s-continuous. A straightforward remedy to this problem is to refine
the strong topology such that taking adjoint becomes a continuous operation. This
naturally leads to the investigation of the strong-star topology.
As one may expect, the structure of an s⋆-typical contraction is more complicated
than the structure of an s-typical contraction. We show that the theory of s⋆-typical
properties of contractions can be reduced to the theories of typical properties of
unitary and positive self-adjoint operators in the better understood strong topology.
Theorem 6.1. There exist an s⋆-co-meager s⋆-Gδ set H ⊆ C(H) and an s-co-
meager s-Gδ set P ⊆ P (H) such that the function Ψ : U(H) × P → H, Ψ(U, P ) =
U · P is a homeomorphism, where U(H) and P are endowed with the strong topology
and H is endowed with the strong-star topology.
Moreover, if (ψ0, ψ1) : H → U(H) × P denotes the inverse of Ψ, then for every
A ∈ H and U ∈ U(H) we have UAU −1 ∈ H and ψi(UAU −1) = Uψi(A)U −1 (i < 2).
20
Note that Theorem 6.1 immediately implies the following. Recall O(A) = {UAU −1 : U ∈
U(H)}.
Corollary 6.2. For every A ∈ C(H), O(A) is an s⋆-meager subset of C(H). In
particular, s⋆-typical contractions are not unitarily equivalent.
Proof. We follow the notation of Theorem 6.1. Let A ∈ C(H) be arbitrary.
If
A /∈ H then by the unitary invariance of H we have O(A) ⊆ C(H) \ H, i.e. O(A) is
s⋆-meager. So we can assume A ∈ H. We have
Ψ−1(O(A)) = {(Uψ0(A)U −1, Uψ1(A)U −1) : U ∈ U(H)} ⊆
{Uψ0(A)U −1 : U ∈ U(H)} × {Uψ1(A)U −1 : U ∈ U(H)} = O(ψ0(A)) × O(ψ1(A)).
By Proposition 3.3, O(ψ0(A)) ⊆ U(H) is s-meager. Hence Ψ−1(O(A)) is s×s-meager
in U(H) × P. By Ψ being a homeomorphism, we obtain that O(A) is s⋆-meager in
H and so in C(H), as well. The further corollary that unitarily equivalent pairs are
s⋆ × s⋆-meager in C(H) × C(H) follows as in the proof of Proposition 3.3.(cid:4)
Similarly to our approach in the previous section, for the proof of Theorem 6.1 we
need to describe the mapping properties of s⋆-typical contractions. These investiga-
tions will also help us to determine the s⋆-typical spectral properties of contractions.
6.1. Mapping properties of s⋆-typical contractions
We prove the following.
Proposition 6.3. The set
T = {A ∈ C(H) : ∀λ ∈ C (Ran (A − λ · I) is dense in H)}
is s⋆-co-meager and s⋆-Gδ in C(H).
We need the following elementary property of the backward unilateral shift op-
erator. It follows from the fact that its adjoint, the unilateral shift operator, on ℓ2
has empty point spectrum (see e.g. [26]).
Lemma 6.4. Let {ei : i ∈ N} be an orthonormal base in H. Define D ∈ C(H) by
De0 = 0, Dei+1 = ei (i ∈ N). Then for every λ ∈ C, Ran (D − λ · Id) is dense in H.
21
Proof of Proposition 6.3. First we show that T is s⋆-dense in C(H). Let
U ⊆ C(H) be a non-empty s⋆-open set. Then there exist {xi : i < n} ⊆ SH, ε > 0
and A ∈ C(H) such that for every B ∈ C(H), kBxi − Axik ≤ ε (i < n) and
kB⋆xi − A⋆xik ≤ 2ε (i < n) imply B ∈ U. It is enough to find a B ∈ U such that
for every λ ∈ C, Ran (B − λ · Id) is dense in H.
Set V = span{xi, Axi, A⋆xi : i < n}. Let Q : V → V be defined by Q = (1 − ε) ·
PrV AV . Let dim V = m and let {xi(0) : i < m} be an orthonormal base in V . Let
{xi(j) : i < m, j ∈ N \ {0}} be an orthonormal base in V ⊥. Define T : V ⊥ → H by
T xi(j) = xi(j − 1) (i < m, j ∈ N \ {0}). Set B = Q ⊕ ε · T ; we show that B fulfills
the requirements.
We have kBk ≤ kBV k + kBV ⊥k = kQk + ε · kTk ≤ 1 − ε + ε = 1, so B ∈ C(H).
For every λ ∈ C,
Ran (B − λ · Id) ⊇ Ran (BV ⊥ − λ · IdV ⊥) ⊇ Ran (cid:18)T −
ε · IdV ⊥(cid:19) .
Since T is an m-fold sum of the operator D of Lemma 6.4, Ran (cid:0)T − λ
λ
dense in H, as required.
For every i < n,
ε · IdV ⊥(cid:1) is
kBxi − Axik = kQxi − Axik = kQxi − PrV AV xik ≤ ε,
kB⋆xi − A⋆xik = kQ⋆xi + ε · T ⋆xi − A⋆xik ≤
ε · kT ⋆xik + kQ⋆xi − A⋆xik ≤ ε + kQ⋆xi − PrV A⋆V xik ≤ 2ε,
i.e. B ∈ U, as required.
and L ≥ 0 set
It remains to show that T is s⋆-Gδ in C(H). To this end, for every y ∈ SH , δ > 0
R(y, δ, L) = {A ∈ C(H) : ∃λ ∈ C (λ ≤ L, dist(Ran (A − λ · Id), y) ≥ δ}.
(3)
We show that R(y, δ, 1) (y ∈ SH, δ > 0) are s-closed. Let {An : n ∈ N} ⊆ R(y, δ, 1)
and A ∈ C(H) be such that s-limn∈N An = A; we show that A ∈ R(y, δ, 1). By
An ∈ R(y, δ, 1) (n ∈ N), we have λn ∈ C, λn ≤ 1 (n ∈ N) such that dist(Ran (An −
λn · Id), y) ≥ δ.
By passing to a subsequence, we can assume {λn : n ∈ N} is convergent, say
limn∈N λn = λ; then λ ≤ 1.
It's enough to prove dist(Ran (A − λ · Id), y) ≥ δ.
Suppose this is not the case, i.e. there is an x ∈ H such that kAx − λx − yk <
22
δ. Since limn∈N Anx = Ax and limn∈N λn = λ, for an n sufficiently large we have
kAnx − λnx − yk < δ, which contradicts dist(Ran (An − λn · Id), y) ≥ δ.
Thus R(y, δ, 1) (y ∈ SH, δ > 0) are s-closed, and so s⋆-closed, as well. Let
Y ⊆ SH be a dense countable set. Since T = C(H) \S{R(y, 2−n, 1) : y ∈ Y, n ∈ N},
the statement follows. (cid:4)
For technical reasons, we state the following corollary of Proposition 6.3.
Corollary 6.5. The set
E = {A ∈ C(H) : Ran A is dense in H}
is s⋆-co-meager and s⋆-Gδ in C(H).
Proof. By Proposition 6.3, E is s⋆-co-meager in C(H). With the notation of (3),
E = C(H) \S{R(y, 2−n, 0) : y ∈ Y, n ∈ N}, so the statement follows.(cid:4)
We also need a similar result for positive self-adjoint operators.
Proposition 6.6. The set
P = {P ∈ P (H) : Ran P is dense in H}
is s-co-meager and s-Gδ in P (H).
Proof. First we show that P is s-dense in P (H). To this end, let U ⊆ P (H) be
a non-empty s-open set. Then there exist {xi : i < n} ⊆ SH , ε > 0 and A ∈ P (H)
such that for every B ∈ P (H), kBxi − Axik ≤ ε (i < n) implies B ∈ U. We find a
B ∈ U such that Ran B is dense in H.
Set V = span{xi, Axi : i < n}. Let Q : V → V be an invertible contractive pos-
itive self-adjoint operator such that kQ − PrV AV k ≤ ε; such a Q exists, e.g. any
Q = (1−δ)·PrV AV +δ′·IdV fulfills the requirements for suitable 0 < δ, δ′ ≤ ε/2. Set
B = Q ⊕ IdV ⊥. Then B is a positive self-adjoint operator, kBV k ≤ 1, kBV ⊥k = 1
and B is invertible, hence B ∈ P (H) and Ran B is dense in H. For every i < n,
kBxi − Axik = kQxi − Axik = kQxi − PrV AV xik ≤ ε,
i.e. B ∈ U, as required.
It remains to show that P is s-Gδ. Observe that for every y ∈ SH and δ > 0, the
set R(y, δ) = {A ∈ P (H) : dist(Ran A, y) ≥ δ} is s-closed. Let Y ⊆ SH be a dense
countable set, then
which completes the proof.(cid:4)
P = P (H) \S{R(y, 2−n) : y ∈ Y, n ∈ N},
23
6.2. Proof of Theorem 6.1
With the notation of Corollary 6.5, set
H = E ∩ {A⋆ : A ∈ E} = {A ∈ C(H) : Ran A and Ran A⋆ are dense in H}.
By Corollary 6.5, H is an s⋆-co-meager s⋆-Gδ set in C(H). Define ψ1 : H → P (H)
by ψ1(A) = (A⋆A)1/2 (A ∈ H). For every A ∈ C(H), (A⋆A)1/2 is a positive self-
adjoint contraction so the definition makes sense. Moreover, for A ∈ H we have
Ran (A⋆A)1/2 ⊇ Ran (A⋆A) is dense in H, so with the notation of Proposition 6.6, ψ1
maps into P. Note that the mapping A 7→ A⋆A is s⋆-continuous. We need that ψ1 is
s⋆-continuous. Since we couldn't find a reference, we outline a proof of the following.
Lemma 6.7. The function ·1/2 : P (H) → P (H), A 7→ A1/2 (A ∈ P (H)) is s-
continuous.
Proof. Let A ∈ P (H) be arbitrary; note that σ(A) ⊆ [0, 1]. By [33, Theorem XI.6.1
p. 313], [33, Theorem XI.4.1 p. 307], [33, Proposition XI.5.2 p. 310] and [33, Theorem
XI.8.1 p. 319], there is a s-left-continuous function F : [0, 1] → P (H) such that F is
projection-valued, F (0) = 0, F (1) = Id, F (ϑ)F (ϑ′) = F (min{ϑ, ϑ′}) (ϑ, ϑ′ ∈ [0, 1]),
and for every continuous function f : [0, 1] → R, R 1
0 f (ϑ)dF (ϑ) exists as the s-limit
of Riemann-Stieltjes sums and defines a bounded linear operator, denoted by f (A),
satisfying kf (A)k ≤ kfk∞. Moreover, by [33, Example XI.12.1 p. 338] and [33,
Theorem XI.12.3 p. 343], this defines a functional calculus compatible with power
series expansion, in particular A1/2A1/2 = A.
Let {An : n ∈ N} ⊆ P (H) satisfy s-limn∈N An = A. Let x ∈ SH and ε > 0 be
arbitrary. Let p : [0, 1] → R be a polynomial satisfying maxx∈[0,1] p(x) − x1/2 < ε.
n = Ak (k ∈ N), we have limn∈N p(An)x = p(A)x. As we observed
Since s-limn∈N Ak
above, kA1/2 − p(A)k,kA1/2
n − p(An)k < ε (n ∈ N). This implies lim supn∈N kA1/2x −
A1/2
n xk < 2ε; since ε was arbitrary, the statement follows.(cid:4)
Corollary 6.8. The function ψ1 : H → P is s⋆-continuous.
Proof. Since the strong and strong-star topologies coincide on P (H), the statement
follows.(cid:4)
Let A ∈ H be arbitrary. As we observed above, ψ1(A) is a positive self-adjoint
operator with dense range. Hence ψ1(A)−1 is a closed densely defined positive self-
adjoint operator. Consider the densely defined operator A · ψ1(A)−1. It has a dense
range, and for every x ∈ Ran ψ1(A) we have
hA · ψ1(A)−1x, A · ψ1(A)−1xi = hA⋆A · ψ1(A)−1x, ψ1(A)−1xi =
hψ1(A)−1 · A⋆A · ψ1(A)−1x, xi = hx, xi.
24
Hence A · ψ1(A)−1 is a densely defined isometry with dense range, i.e. it is closable
and its closure is an unitary operator. Let ψ0(A) ∈ U(H) denote the closure of
A · ψ1(A)−1. Note that by definition, A = ψ0(A) · ψ1(A).
Corollary 6.9. Every A ∈ H can be written as A = ψ0(A) · ψ1(A) where ψ0(A) ∈
U(H) and ψ1(A) ∈ P. This decomposition is unique and unitary invariant, i.e. for
every A ∈ H and U ∈ U(H) we have ψi(UAU −1) = Uψi(A)U −1 (i < 2).
Proof. To see uniqueness, let U, V ∈ U(H) and P, Q ∈ P be arbitrary. Then
U · P = V · Q implies P 2 = P ⋆U ⋆UP = Q⋆V ⋆V Q = Q2, so P = Q by the uniqueness
of positive square root. By P, Q having dense range, we get U = V .
If A ∈ H and U ∈ U(H) then UAU −1 ∈ H and UAU −1 = Uψ0(A)U −1 ·
Uψ1(A)U −1. Hence by the uniqueness of this decomposition, we have ψ0(UAU −1) =
Uψ0(A)U −1 and ψ1(UAU −1) = Uψ1(A)U −1. (cid:4)
Recall Ψ : U(H) × P → H, Ψ(U, P ) = U · P , which is an s⋆-continuous function.
By Corollary 6.9, Ψ is injective. It is obvious that (ψ0, ψ1) is the inverse of Ψ, so the
following completes the proof of Theorem 6.1.
Lemma 6.10. The function ψ0 is s⋆-continuous.
Proof. Let {An : n ∈ N} ⊆ H and A ∈ H be such that s⋆-limn∈N An = A. For every
y ∈ Ran ψ1(A), say y = ψ1(A)x, we have
kψ0(An)y − ψ0(A)yk = kψ0(An)ψ1(A)x − ψ0(A)ψ1(A)xk ≤
kψ0(An)ψ1(A)x − ψ0(An)ψ1(An)xk + kψ0(An)ψ1(An)x − ψ0(A)ψ1(A)xk ≤
kψ0(An)k · kψ1(A)x − ψ1(An)xk + kAnx − Axk.
Since ψ1 is s⋆-continuous, the statement follows from Ran ψ1(A) ⊆ H being dense.(cid:4)
6.3. Spectral properties of s⋆-typical contractions
Our main result in this section is the following.
Proposition 6.11. An s⋆-typical contraction A satisfies Cσ(A) = {λ ∈ C : λ ≤ 1}.
We start with a lemma.
Lemma 6.12. The set
S = {A ∈ C(H) : ∀λ ∈ C,λ ≤ 1 (λ ∈ σ(A))}
is s⋆-co-meager and s⋆-Gδ in C(H).
25
Proof. First we show that S is s⋆-dense in C(H). Let U ⊆ C(H) be a non-empty
s⋆-open set. Then there exist {xi : i < n} ⊆ SH , ε > 0 and A ∈ C(H) such that
for every B ∈ C(H), kBxi − Axik < ε (i < n) and kB⋆xi − A⋆xik < ε (i < n)
imply B ∈ U. It is enough to find a B ∈ U such that for every λ ∈ C with λ ≤ 1,
λ ∈ σ(B).
Set V = span{xi, Axi, A⋆xi : i < n}, and let Q : V → V be defined by Q =
PrV AV . Let T ∈ C(V ⊥) be arbitrary satisfying λ ∈ σ(T ) (λ ∈ C,λ ≤ 1), say the
backward unilateral shift operator, and set B = Q ⊕ T . We show that B fulfills the
requirements.
Since σ(T ) ⊆ σ(B), we have λ ∈ σ(B) (λ ∈ C,λ ≤ 1). For every i < n,
kBxi − Axik = kQxi − Axik = k PrV AV xi − Axik = 0 < ε,
kB⋆xi − A⋆xik = kQ⋆xi − A⋆xik = k PrV A⋆V xi − A⋆xik = 0 < ε,
i.e. B ∈ U, as required.
It remains to show that S is s⋆-Gδ in C(H). For every δ > 0, set
S(δ) = {A ∈ C(H) : ∃λ ∈ C (λ ≤ 1 and k(A−λ·I)xk,k(A⋆−λ·I)xk ≥ δ (x ∈ SH))}.
It is clear that S(δ) (δ > 0) are s⋆-closed. So the proof will be complete if we show
S = C(H) \Sn∈N S(2−n).
Let A ∈ C(H) be arbitrary. If λ /∈ σ(A) for a λ ∈ C, λ ≤ 1 then (A − λ · I)−1
and (A⋆ − λ · I)−1 exist so A ∈ S(δ) for some sufficiently small δ > 0. This proves
C(H) \ S ⊆Sn∈N S(2−n).
Similarly, if A ∈ S(δ) for some δ > 0, say λ ∈ C, λ ≤ 1 witnesses this, then
λ /∈ Pσ(A) ∪ Cσ(A) and λ /∈ Pσ(A⋆). By Ker (A⋆ − λ · I) = Ran (A − λ · I)⊥,
λ /∈ Pσ(A⋆) implies λ /∈ Rσ(A). So we obtained λ /∈ σ(A), i.e. A /∈ S. This proves
Sn∈N S(2−n) ⊆ C(H) \ S and finishes the proof.(cid:4)
Proof of Proposition 6.11. The map A 7→ A⋆ is a s⋆-homeomorphism of
C(H). So by Proposition 6.3, a typical contraction A satisfies that for every λ ∈ C,
Ran (A − λ · I) and Ran (A⋆ − λ · I) are dense in H. Since Ker (A − λ · I) =
Ran (A⋆ − λ · I)⊥, we get Ker (A − λ · I) = ∅ (λ ∈ C). Hence Pσ(A) = Pσ(A⋆) = ∅
and Rσ(A) = Rσ(A⋆) = ∅. So Cσ(A) = {λ ∈ C : λ ≤ 1} is an immediate corollary
of Lemma 6.12.(cid:4)
7. The norm topology
In the norm topology it is not possible to give a simple description of the spectral
properties of typical operators. An intuitive explanation may be that the norm
26
topology is non-separable, hence several different properties can coexist on non-
meager sets.
In this section, every topological notion refers to the norm topology. We prove
the following.
Theorem 7.1. Let λ ∈ C be arbitrary. Then the following sets of operators have
non-empty interior.
1. {A ∈ B(H) : λ /∈ σ(A)};
2. {A ∈ B(H) : λ ∈ Rσ(A)};
3. {A ∈ B(H) : λ ∈ Pσ(A)}.
In particular, the following sets have non-empty interior.
4. {A ∈ B(H) : Ran (A − λ · I) = H};
5. {A ∈ B(H) : Ran (A − λ · I) is not dense in H}.
On the other hand, the following sets of operators are nowhere dense.
6. {A ∈ B(H) : λ ∈ Cσ(A)};
7. {A ∈ B(H) : Ran (A − λ · I) is dense in H but not equal to H}.
We need some terminology in advance.
Definition 7.2. Let A ∈ B(H) and λ ∈ C be arbitrary. We say λ ∈ σ(A) is stable if
there is an ε > 0 such that for every D ∈ B(H) with kDk < ε we have λ ∈ σ(A + D).
Similarly, we say λ ∈ Pσ(A) is stable if there is an ε > 0 such that for every D ∈ B(H)
with kDk < ε we have λ ∈ Pσ(A + D).
With this terminology, [17, Theorem 2 p. 912] can be reformulated as follows.
Proposition 7.3. Let A ∈ B(H) and λ ∈ C be arbitrary. Then the following are
equivalent.
1. λ ∈ σ(A) is stable;
2. dim Ker (A − λ · I) 6= dim Ran (A − λ · I)⊥ and there is an ε > 0 such that for
every x ∈ Ker (A − λ · I)⊥ we have k(A − λ · I)xk ≥ ε · kxk.
27
We prove a similar result for the point spectrum.
Proposition 7.4. Let A ∈ B(H) and λ ∈ C be arbitrary. Then the following are
equivalent.
1. λ ∈ Pσ(A) is stable;
2. dim Ran (A − λ · I)⊥ < dim Ker (A − λ · I) and there is a ε > 0 such that for
every x ∈ Ker (A − λ · I)⊥ we have k(A − λ · I)xk ≥ ε · kxk.
Proof. By linearity, we can assume λ = 0. Suppose first 0 ∈ Pσ(A) is stable.
Then 0 ∈ σ(A) is stable so by Proposition 7.3, dim Ker (A) 6= dim Ran (A)⊥ and
there is a ε > 0 such that for every x ∈ Ker (A)⊥ we have kAxk ≥ ε · kxk.
If
dim Ker (A) < dim Ran (A)⊥ then for every ε > 0 there is a D ∈ B(H) such that
kDk < ε, DKer (A)⊥ = 0 and for every x ∈ Ker (A)\{0}, Dx ∈ Ran (A)⊥ \{0}. Then
for every x ∈ H \{0} we have (A + D)x 6= 0, hence 0 /∈ Pσ(A + D). This contradicts
the assumption that 0 ∈ Pσ(A) is stable. Thus dim Ran (A)⊥ < dim Ker (A), as
required.
To see the converse, suppose the conditions of statement 2 hold. Then we have
dim Ker (A) > 0 hence 0 ∈ Pσ(A). We also have that A : Ker (A)⊥ → Ran (A) is a
bounded invertible operator, in particular Ran (A) is a closed co-finite dimensional
subspace of H. Let k = 1 + dim Ran (A)⊥.
Let D ∈ B(H) be arbitrary with kDk < ε/2k2. For every x ∈ Ker (A) consider
the following inductive definition of a sequence {xn : n ∈ N} ⊆ H. Set x0 = x. Let
n ∈ N be arbitrary and suppose that xn ∈ H is defined. Write Dxn = un + vn where
un ∈ Ran (A) and vn ∈ Ran (A)⊥. Set xn+1 = A−1un. This completes the inductive
step of the definition of {xn : n ∈ N}.
We set ξ(x) = Pn∈N(−1)nxn, ρ(x) = Pn∈N(−1)nvn. Note that kxn+1k ≤
kunk/ε ≤ kDxnk/ε < kxnk/2k2, hence kxnk < kxk/(2k2)n (n ∈ N) so the defi-
nitions make sense. Moreover, the functions ξ and ρ are linear, ρ(x) ∈ Ran (A)⊥ and
kξ(x)− xk < 1/k (x ∈ BKer (A)). So by Lemma 5.6, if {x(i) : i < k} is an orthonormal
system in Ker (A) then {ξ(x(i)) : i < k} are linearly independent.
Observe that
(A + D)ξ(x) = (A + D)Xn∈N
Ax0 + Xn∈N\{0}
(−1)nxn =Xn∈N
(−1)nAxn +Xn∈N
(−1)n(Axn + Dxn) =
(−1)n(un + vn) =Xn∈N
(−1)nvn = ρ(x).
28
By dim Ker (A) ≥ k, there is an orthonormal system {x(i) : i < k} in Ker (A). Since
dim Ran (A)⊥ < k, there is an x ∈ span{x(i) : i < k}\{0} such that ρ(x) = 0. Then
(A + D)ξ(x) = 0 shows 0 ∈ Pσ(A + D), so the proof is complete.(cid:4)
Proof of Theorem 7.1. By linearity, it is enough to consider the λ = 0 case.
The set of invertible operators shows statements 1 and 4. By Proposition 7.4, a
neighborhood of the backward unilateral shift operator shows statement 3.
The set of statement 7 contains the set of statement 6 so it is sufficient to prove
statement 7. Let A ∈ B(H) be arbitrary satisfying Ran (A) is dense in H but not
equal to H. Then A cannot be invertible. Moreover, note that condition kAxk ≥
ε · kxk (x ∈ Ker (A)⊥) in Proposition 7.3.2 would imply Ran (A) is closed. Since
Ran (A) cannot be closed, A cannot satisfy Proposition 7.3.2. So the set of statement
7 is contained in the boundary of the set of invertible operators hence it is nowhere
dense.
Finally notice that a neighborhood of the unilateral shift operator D⋆ is contained
in {A ∈ B(H) : 0 ∈ σ(A) \ Pσ(A)}, since the stability of 0 ∈ σ(D⋆) is implied by
Proposition 7.3 while the stability of 0 /∈ Pσ(D⋆) follows from D⋆ being an isometry.
So by statement 6, statements 2 and 5 follow.(cid:4)
Observe that the backward unilateral shift operator D of Lemma 6.4 satisfies the
conditions of statement 2 of Proposition 7.4 for every λ ∈ C with λ < 1, hence
every such λ ∈ Pσ(D) is stable. So by taking direct sums of scaled and translated
copies of D we can obtain arbitrarily complicated stable spectra. We also obtain
the following. Here an operator T is called embeddable, if it can be embedded into a
strongly continuous semigroup, i.e. if T = T (1) holds for some strongly continuous
semigroup (T (t))t≥0.
Corollary 7.5. The set of embeddable operators and the set of non-embeddable op-
erators have non-empty interior.
Proof. By [10, Theorem 2.1 p. 452], each operator in a small neighborhood of Id is
embeddable. To show a non-empty open set of non-embeddable operators, we aim
to use [10, Theorem 3.1 p. 454]. Recall the backward unilateral shift operator D
of Lemma 6.4. By Proposition 7.4, each operator A in a small neighborhood of D
satisfies 0 ∈ Pσ(A). So we get A is non-embeddable if we show dim Ker (A) ≤ 1.
Note that if kA − Dk < 1/√2 then for every x, y ∈ Ker (A) ∩ SH we have
kDxk,kDyk < 1/√2. This implies kx − D⋆Dxk,ky − D⋆Dyk > 1/√2, and since
D⋆D is an orthogonal projection, we have
hx, yi = hx − D⋆Dx, y − D⋆Dyi + hD⋆Dx, D⋆Dyi ≥
hx − D⋆Dx, y − D⋆Dyi − hD⋆Dx, D⋆Dyi > 0.
29
Thus dim Ker (A) ≤ 1, which completes the proof. (cid:4)
The results of this section provide no information about the size of the spectrum
of a typical operator. We state two questions related to this in Problem 8.2. Here
we only point out the following.
Proposition 7.6. The following sets are dense.
1. {A ∈ B(H) : Pσ(A) 6= ∅};
2. {A ∈ B(H) : Rσ(A) 6= ∅}.
Proof. Since A 7→ A⋆ (A ∈ B(H)) is a homeomorphism of B(H) and Pσ(A⋆) =
Rσ(A) (A ∈ B(H)), it suffices to show 1.
Let B ∈ B(H) be arbitrary. Since the approximative point spectrum of B is
[6, Proposition VII.6.7 p. 210]), there exists a λ ∈ σ(B) and a
nonempty (see e.g.
sequence {xn : n ∈ N} ⊆ SH such that limn∈N kBxn − λxnk = 0. For every n ∈ N, let
Pn denote the orthogonal projection onto span{xn}, and set An = B(Id−Pn)+λ·Pn.
We have Anxn = λxn and kAn − Bk = kλ · Pn − BPnk = kλxn − Bxnk (n ∈ N).
Hence Pσ(An) 6= ∅ (n ∈ N) and limn∈N kAn − Bk = 0. This completes the proof.(cid:4)
8. An outlook to general Banach spaces
The proofs in the previous section could convince the reader that an attempt
to extend our investigations to tackle the typical properties of contractions on arbi-
trary Banach spaces could encounter considerable technical difficulties. In addition,
such an endeavor has to face some problems of more fundamental nature. Recent
developments in the theory of Banach spaces resulted in numerous spaces exhibiting
surprising functional analytic properties. On the famous Banach space of S. A. Argy-
ros and R. G. Haydon, every bounded linear operator is of the form λ· Id + K where
λ ∈ C and K is a compact operator (see [1]). Spaces were constructed with only
trivial isometries (see e.g. [7] and the references therein), moreover even a renorming
of a Banach space may result in an arbitrary isometry group (see [18] and the refer-
ences therein). The relevance of isometries comes from the important role played by
unitary operators in the theory developed in the previous section.
Observe that renorming does not change the topology of the underlying Banach
space, so the five topologies we consider on operators remain unchanged, as well.
Instead, renorming affects the size of various classes of operators. So in a sense, the
study of typical properties of operators in various topologies is more related to the
geometry of the underlying Banach space than to the topology it carries. Therefore,
30
reasonable extensions of our investigations should be pursued in Banach spaces where
the geometry of the space is of special significance.
The most obvious such spaces are the Lp spaces (1 ≤ p ≤ ∞). The functional
analysis of these spaces is well-developed, isometric operators are characterized (see
e.g. [22]). Nevertheless, we expect that none of our main results extends to arbitrary
Lp spaces.
Another promising extension of the theory of typical behavior of operators could
tackle Banach⋆-algebras. Developing our results to that generality could separate
typical properties which are of operator theoretic nature from typical properties
exploiting the geometry of the underlying space.
Finally let us propose some concrete problems which stem from the results of the
previous section.
Problem 8.1. We say A, B ∈ B(H) are similar if there is an invertible operator
T ∈ B(H) such that A = T BT −1.
1. Is it true that w-typical contractions are similar?
2. Is it true that s⋆-typical contractions are similar?
Problem 8.2. Determine the s-typical properties of self-adjoint and positive self-
adjoint operators.
Problem 8.3. Is a s⋆-typical contraction embeddable into a strongly continuous
semigroup?
Problem 8.4.
1. Is the set {A ∈ B(H) : Pσ(A) 6= ∅} co-meager in the norm topology?
2. Is the set {A ∈ B(H) : Cσ(A) = ∅} co-meager in the norm topology?
It would be instructive to examine whether suitable analogues of our results hold
for strongly continuous semigroups instead of single operators. Let C c(H) denote the
set of contractive C0-semigroups. Given any metric d on C(H), one can endow C c(H)
with the topology generated by uniform d-convergence on compact time intervals.
This topology is induced by the metric
dc(T (·), S(·)) =Pn∈N 2−n supt∈[0,n] d(T (t), S(t)).
Problem 8.5. Under which of these topologies is Cc(H) a Baire space? What are
the typical properties of contractive C0-semigroups in these topologies?
Recall that the strong topology case was treated in Section 5.4. Some results
related to Problem 8.5 for the weak topology can be found in [8, Section 4], [12] and
[14].
31
References
[1] S. A. Argyros, R. G. Haydon, A hereditarily indecomposable L∞-space that
solves the scalar-plus-compact problem, Acta Math. 206 (2011), 1-54.
[2] S. Banach, Uber die Baire'sche Kategorie gewisser Funktionenmengen, Studia
Math. 3 (1931), 174 -- 179.
[3] W. Bartoszek and B. Kuna, On residualities in the set of Markov operators on
C1, Proc. Amer. Math. Soc. 133 (2005), 2119 -- 2129.
[4] A. Bruckner, Differentiation of real functions, Second edition, CRM Monograph
Series, 5, American Mathematical Society, Providence, RI, 1994.
[5] J. R. Choksi, M. G. Nadkarni, Baire category in spaces of measures, unitary
operators, and transformations, Proc. Int. Conference on Invariant Subspaces
and Allied Topics (1986), Narosa Publishers, New Delhi.
[6] J. B. Conway, A Course in Functional Analysis. Second edition, Graduate Texts
in Mathematics 96, Springer-Verlag, New York, 1990.
[7] G. Diestel, A. Koldobsky, Sobolev spaces with only trivial isometries, Positivity
10 (2006), no. 1, 135 -- 144.
[8] T. Eisner, A "typical" contraction is unitary, Enseign. Math. (2) 56 (2010),
403-410.
[9] T. Eisner, Rigidity of contractions on Hilbert spaces, arXiv:0909.4695v1.
[10] T. Eisner, Embedding operators into strongly continuous semigroups, Arch.
Math. (Basel) 92 (2009), 451 -- 460.
[11] T. Eisner, Stability of Operators and Operator Semigroups, Operator Theory:
Advances and Applications, 209, Birkhauser Verlag, Basel, 2010.
[12] T. Eisner, A. Ser´eny, Category theorems for stable operators on Hilbert spaces,
Acta Sci. Math. (Szeged) 74 (2008), 259 -- 270.
[13] T. Eisner, A. Ser´eny, Category theorems for stable semigroups, Ergodic Theory
Dynamical Systems 29 (2009), 487 -- 494.
[14] T. Eisner, A. Ser´eny, On the weak analogue of the Trotter -- Kato theorem, Tai-
wanese J. Math. 14 (2010), 1411 -- 1416.
32
[15] K.-J. Engel, R. Nagel, One-parameter semigroups for linear evolution equations,
Graduate Texts in Mathematics, 194, Springer-Verlag, New York, 2000.
[16] P. R. Halmos, In general a measure preserving transformation is mixing, Ann.
Math. 45 (1944), 786 -- 792.
[17] J. Feldman, R. V. Kadison, The closure of the regular operators in a ring of
operators, Proc. Amer. Math. Soc. 5 (1954), 909 -- 916.
[18] V. Ferenczi, E. M. Galego, Countable groups of isometries on Banach spaces,
Trans. Amer. Math. Soc. 362 (2010), 4385-4431.
[19] S. Jitomirskaya, B. Simon, Operators with singular continuous spectrum. III.
Almost periodic Schrdinger operators, Comm. Math. Phys. 165 (1994), no. 1,
201 -- 205.
[20] A. S. Kechris, Classical Descriptive Set Theory, Graduate Texts in Mathematics,
156, Springer-Verlag, 1994.
[21] B. Kuna, On residualities in the set of Markov continuous semigroups on C1,
Demonstratio Math. 39 (2006), no. 2, 439 -- 453.
[22] J. Lamperti, On the isometries of certain function-spaces, Pacific J. Math. 8
(1958) 459 -- 466.
[23] S. Mazurkiewicz, Sur les fonctions non d´erivables, Studia Math. 3 (1931), 92 -- 94.
[24] M. G. Nadkarni, Spectral theory of dynamical systems, Birkhauser Advanced
Texts: Basler Lehrbucher, Birkhauser Verlag, Basel, 1998.
[25] V. V. Peller, Estimates of operator polynomials in the space Lp with respect to
the multiplicative norm, J. Math. Sciences 16 (1981), 1139 -- 1149.
[26] W. C. Ridge, Approximate point spectrum of a weighted shift, Trans. Amer.
Math. Soc. 147 (1970), 349 -- 356.
[27] R. Del Rio, N. Makarov, B. Simon, Operators with singular continuous spec-
trum. II. Rank one operators, Comm. Math. Phys. 165 (1994), no. 1, 59 -- 67.
[28] V. A. Rohlin, A "general" measure-preserving transformation is not mixing,
Doklady Akad. Nauk SSSR 60 (1948), 349 -- 351.
33
[29] A. L. Shields, Weighted shift operators and analytic function theory, Topics in
Operator Theory, Math. Surveys No. 13, Amer. Math. Soc., Providence, R.I.,
1974, 49 -- 128.
[30] B. Simon, Operators with singular continuous spectrum: I. General operators,
Annals Math. 141 (1995), 131 -- 145.
[31] B. Szokefalvi-Nagy, C. Foia¸s, Harmonic Analysis of Operators on Hilbert Space,
North-Holland, Akad´emiai Kiad´o, Amsterdam, Budapest, 1970.
[32] M. Takesaki, Theory of operator algebras I, Springer-Verlag, New York-
Heidelberg, 1979.
[33] K. Yosida, Functional analysis, Sixth edition, Grundlehren der Mathematischen
Wissenschaften, 123, Springer-Verlag, Berlin-New York, 1980.
[34] Y.-C. Wong, Introductory theory of topological vector spaces, Monographs and
Textbooks in Pure and Applied Mathematics, 167, Marcel Dekker, Inc., New
York, 1992.
[35] P. Zorin-Kranich, Typical bounded operators admit common cyclic vectors, sub-
mitted, arXiv:1008.2878.
34
|
1006.5803 | 1 | 1006 | 2010-06-30T08:36:36 | On the two-dimensional moment problem | [
"math.FA"
] | In this paper we obtain an algorithm towards solving the two-dimensional moment problem. This algorithm gives the necessary and sufficient conditions for the solvability of the moment problem. It is shown that all solutions of the moment problem can be constructed using this algorithm. | math.FA | math |
On the two-dimensional moment problem.
S.M. Zagorodnyuk
1
Introduction.
In this paper we analyze the two-dimensional moment problem. Recall that
this problem consists of finding a non-negative Borel measure µ in R2 such
that
xm
1 xn
2 dµ = sm,n,
m, n ∈ Z+,
(1)
ZR2
where {sm,n}m,n∈Z+ is a prescribed sequence of complex numbers.
The two-dimensional moment problem and the (closely related to this
subject) complex moment problem have an extensive literature, see books [1],
[2], [3], surveys [4],[5] and [6]. Some conditions of solvability for this moment
problem were obtained by Kilpi and by Stochel and Szafraniec, see e.g. [2]
and [6]. However, these conditions are hard to check. To the best of our
knowledge, the description of all solutions was not obtained.
Firstly, we obtain a solvability criterion for an auxiliary extended two-
dimensional moment problem. It is shown that this problem is always de-
terminate and its solution can be constructed explicitly. An idea of our
algorithm for solving the two-dimensional moment problem is to extend
the prescribed sequence of moments (1) to be a moment sequence of the
extended moment problem.
It is shown that all solutions of the moment
problem (1) can be constructed on this way. The method of proof uses an
abstract operator approach, see [7]. Roughly speaking, the final algorithm
reduces to the solving of finite and infinite linear systems of equations with
parameters.
Notations.
As usual, we denote by R, C, N, Z, Z+ the sets of real
numbers, complex numbers, positive integers, integers and non-negative in-
tegers, respectively. The real plane will be denoted by R2. For a subset S
of R2 we denote by B(S) the set of all Borel subsets of S. Everywhere in
this paper, all Hilbert spaces are assumed to be separable. By (·,·)H and
k · kH we denote the scalar product and the norm in a Hilbert space H,
respectively. The indices may be omitted in obvious cases. For a set M in
H, by M we mean the closure of M in the norm k · kH . For {xm,n}m,n∈Z+ ,
xm,n ∈ H, we write Lin{xm,n}m,n∈Z+ for the set of linear combinations of
elements {xm,n}m,n∈Z+ and span{xm,n}m,n∈Z+ = Lin{xm,n}m,n∈Z+. The
identity operator in H is denoted by EH . For an arbitrary linear operator
1
A in H, the operators A∗,A,A−1 mean its adjoint operator, its closure and
its inverse (if they exist). By D(A) and R(A) we mean the domain and the
range of the operator A. The norm of a bounded operator A is denoted by
kAk. By P H
= PH1 we mean the operator of orthogonal projection in H on
a subspace H1 in H. By L2
µ we denote the usual space of square-integrable
complex functions f (x1, x2), x1, x2 ∈ R2, with respect to the Borel measure
µ in R2.
H1
2 The solution of an extended two-dimensional
moment problem.
Consider the following moment problem: to find a non-negative Borel mea-
sure µ in R2 such that
ZR2
1 (x1 + i)k(x1 − i)lxn
xm
2 (x2 + i)r(x2 − i)tdµ = um,k,l;n,r,t,
m, n ∈ Z+, k, l, r, t ∈ Z,
(2)
where {um,k,l;n,r,t}m,n∈Z+,k,l,r,t∈Z is a prescribed sequence of complex num-
bers. This problem is said to be the extended two-dimensional moment
problem.
We set
Ω = {(m, k, l; n, r, t) : m, n ∈ Z+, k, l, r, t ∈ Z},
Ω0 = {(m, k, l; n, r, t) : m, n ∈ Z+, k, l, r, t ∈ Z, k = l = r = t = 0},
Ω′ = Ω\Ω0.
Let the moment problem (2) have a solution µ. Choose an arbitrary function
αm,k,l;n,r,txm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)t,
where all but finite number of complex coefficients αm,k,l;n,r,t are zeros. Then
P (x1, x2) = X(m,k,l;n,r,t)∈Ω
0 ≤ZR2 P (x1, x2)2dµ =
xm+m′
1
(x1 + i)k+l′
∗ZR2
=
X
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
αm,k,l;n,r,tαm′,k′,l′;n′,r′,t′
X
(x1 − i)l+k′
αm,k,l;n,r,tαm′,k′,l′;n′,r′,t′ um+m′,k+l′,l+k′;n+n′,r+t′,t+r′.
(x2 − i)t+r′
dµ
xn+n′
2
(x2 + i)r+t′
2
Therefore
X
αm,k,l;n,r,tαm′,k′,l′;n′,r′,t′ um+m′,k+l′,l+k′;n+n′,r+t′,t+r′ ≥ 0,
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
(3)
for arbitrary complex coefficients αm,k,l;n,r,t, where all but finite number of
αm,k,l;n,r,t are zeros. The latter condition on the coefficients αm,k,l;n,r,t in
infinite sums will be assumed in similar situations.
We shall use the following important fact (e.g. [8, pp.361-363]).
Theorem 1 Let a sequence of complex numbers {um,k,l;n,r,t}(m,k,l;n,r,t)∈Ω
satisfy condition (3). Then there exist a separable Hilbert space H with
a scalar product (·,·)H and a sequence {xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω in H, such
that
(xm,k,l;n,r,t, xm′,k′,l′;n′,r′,t′)H = um+m′,k+l′,l+k′;n+n′,r+t′,t+r′ ,
(m, k, l; n, r, t), (m′, k′, l′; n′, r′, t′) ∈ Ω,
(4)
and span{xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω = H.
Proof (We do not claim the originality of the idea of this proof). Choose
an arbitrary infinite-dimensional linear vector space V (for instance, one
may choose the space of all complex sequences (un)n∈N, un ∈ C). Let
X = {xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω be an arbitrary infinite sequence of linear in-
dependent elements in V which is indexed by elements of Ω. Set LX =
Lin{xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω. Introduce the following functional:
[x, y] =
αm,k,l;n,r,tβm′,k′,l′;n′,r′,t′
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
X
∗ um+m′,k+l′,l+k′;n+n′,r+t′,t+r′,
(5)
for x, y ∈ LX,
y =
x = X(m,k,l;n,r,t)∈Ω
X(m′,k′,l′;n′,r′,t′)∈Ω
αm,k,l;n,r,txm,k,l;n,r,t,
βm′,k′,l′;n′,r′,t′ xm′,k′,l′;n′,r′,t′,
where αm,k,l;n,r,t, βm′,k′,l′;n′,r′,t′ ∈ C. Here all but finite number of indices
αm,k,l;n,r,t, βm′,k′,l′;n′,r′,t′ are zeros.
The set LX with [·,·] will be a quasi-Hilbert space. Factorizing and making
the completion we obtain the desired space H (e.g. [3]). ✷
3
Let the moment problem (2) be given and the condition (3) hold. By
Theorem 1 there exist a Hilbert space H and a sequence {xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω,
in H, such that relation (4) holds. Set L = Lin{xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω. In-
troduce the following operators
A0 X(m,k,l;n,r,t)∈Ω
B0 X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,txm,k,l;n,r,t = X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,txm,k,l;n,r,t = X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,txm+1,k,l;n,r,t,
αm,k,l;n,r,txm,k,l;n+1,r,t,
(6)
(7)
where all but finite number of complex coefficients αm,k,l;n,r,t are zeros. Let
us check that these definitions are correct. Indeed, suppose that
x = X(m,k,l;n,r,t)∈Ω
We may write
αm,k,l;n,r,txm,k,l;n,r,t =
X(m′,k′,l′;n′,r′,t′)∈Ω
βm′,k′,l′;n′,r′,t′ xm′,k′,l′;n′,r′,t′.
(8)
X(m,k,l;n,r,t)∈Ω
= X(m,k,l;n,r,t)∈Ω
= X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,txm+1,k,l;n,r,t, xa,b,c;d,e,fH
αm,k,l;n,r,tum+1+a,k+c,l+b;n+d,r+f,t+e
αm,k,l;n,r,t(xm,k,l;n,r,t, xa+1,b,c;d,e,f )H
= (x, xa+1,b,c;d,e,f )H ,
(a, b, c; d, e, f ) ∈ Ω.
In the same manner we obtain:
X(m′,k′,l′;n′,r′,t′)∈Ω
αm′,k′,l′;n′,r′,t′xm′+1,k′,l′;n′,r′,t′ , xa,b,c;d,e,fH
= (x, xa+1,b,c;d,e,f )H ,
(a, b, c; d, e, f ) ∈ Ω.
Therefore the definition of A0 is correct. The correctness of the definition
of B0 can be checked in a similar manner. Notice that
(A0xm,k,l;n,r,t, xa,b,c;d,e,f )H = (xm+1,k,l;n,r,t, xa,b,c;d,e,f )H
4
= um+1+a,k+c,l+b;n+d,r+f,t+e = (xm,k,l;n,r,t, xa+1,b,c;d,e,f )H
= (xm,k,l;n,r,t, A0xa,b,c;d,e,f )H ,
(m, k, l; n, r, t), (a, b, c; d, e, f ) ∈ Ω.
Therefore A0 is symmetric. The same argument implies that B0 is symmet-
ric, as well.
Suppose that the following conditions hold:
um+1+a,k+c,l+b;n+d,r+f,t+e + ium+a,k+c,l+b;n+d,r+f,t+e
= um+a,k+1+c,l+b;n+d,r+f,t+e,
um+1+a,k+c,l+b;n+d,r+f,t+e − ium+a,k+c,l+b;n+d,r+f,t+e
= um+a,k+c,l+1+b;n+d,r+f,t+e,
um+a,k+c,l+b;n+1+d,r+f,t+e + ium+a,k+c,l+b;n+d,r+f,t+e
= um+a,k+c,l+b;n+d,r+1+f,t+e,
(9)
(10)
(11)
um+a,k+c,l+b;n+1+d,r+f,t+e − ium+a,k+c,l+b;n+d,r+f,t+e
(12)
for all (m, k, l; n, r, t), (a, b, c; d, e, f ) ∈ Ω. These conditions are equivalent to
conditions
= um+a,k+c,l+b;n+d,r+f,t+1+e,
(xm+1,k,l;n,r,t + ixm,k,l;n,r,t, xa,b,c;d,e,f )H
= (xm,k+1,l;n,r,t, xa,b,c;d,e,f )H ,
(xm+1,k,l;n,r,t − ixm,k,l;n,r,t, xa,b,c;d,e,f )H
= (xm,k,l+1;n,r,t, xa,b,c;d,e,f )H ,
(xm,k,l;n+1,r,t + ixm,k,l;n,r,t, xa,b,c;d,e,f )H
= (xm,k,l;n,r+1,t, xa,b,c;d,e,f )H ,
(xm,k,l;n+1,r,t − ixm,k,l;n,r,t, xa,b,c;d,e,f )H
(13)
(14)
(15)
(16)
for all (m, k, l; n, r, t), (a, b, c; d, e, f ) ∈ Ω. The latter conditions are equiva-
lent to the following conditions:
= (xm,k,l;n,r,t+1, xa,b,c;d,e,f )H ,
xm+1,k,l;n,r,t + ixm,k,l;n,r,t = xm,k+1,l;n,r,t,
xm+1,k,l;n,r,t − ixm,k,l;n,r,t = xm,k,l+1;n,r,t,
xm,k,l;n+1,r,t + ixm,k,l;n,r,t = xm,k,l;n,r+1,t,
(17)
(18)
(19)
5
xm,k,l;n+1,r,t − ixm,k,l;n,r,t = xm,k,l;n,r,t+1,
for all (m, k, l; n, r, t) ∈ Ω. The last conditions mean that
(A0 + iEH )xm,k,l;n,r,t = xm,k+1,l;n,r,t,
(A0 − iEH )xm,k,l;n,r,t = xm,k,l+1;n,r,t,
(B0 + iEH)xm,k,l;n,r,t = xm,k,l;n,r+1,t,
(B0 − iEH)xm,k,l;n,r,t = xm,k,l;n,r,t+1,
(20)
(21)
(22)
(23)
(24)
for all (m, k, l; n, r, t) ∈ Ω. The latter conditions imply that
(A0 ± iEH )L = L,
(B0 ± iEH )L = L.
Therefore operators A0 and B0 are essentially self-adjoint. The condi-
tions (21)-(24) also imply that
(A0 + iEH )−1xm,k,l;n,r,t = xm,k−1,l;n,r,t,
(A0 − iEH )−1xm,k,l;n,r,t = xm,k,l−1;n,r,t,
(B0 + iEH)−1xm,k,l;n,r,t = xm,k,l;n,r−1,t,
(B0 − iEH )−1xm,k,l;n,r,t = xm,k,l;n,r,t−1,
for all (m, k, l; n, r, t) ∈ Ω.
Consider the Cayley transformations of A0 and B0:
VA0 = (A0 − iEH )(A0 + iEH )−1,
(25)
(26)
(27)
(28)
(29)
VB0 = (B0 − iEH )(B0 + iEH )−1,
D(A0) = D(B0) = L.
(30)
By virtue of relations (22),(24),(25),(27) we obtain:
VA0 VB0 xm,k,l;n,r,t = xm,k−1,l+1;n,r−1,t+1 = VB0 VA0 xm,k,l;n,r,t,
for all (m, k, l; n, r, t) ∈ Ω. Therefore
VA0 VB0 x = VB0 VA0 x,
x ∈ L.
(31)
By continuity we extend the isometric operators VA0 and VB0 to unitary
operators UA0 and VB0 in H, respectively. By continuity we conclude that
UA0 UB0 x = UB0 UA0 x,
x ∈ H.
(32)
6
Set A = A0, B = B0. The Cayley transformations of the self-adjoint oper-
rators A and B coincide on L with UA0 and UB0, respectively. Thus, the
Cayley transformations of A and B are UA0 and UB0, respectively. There-
fore, operators A and B commute.
Notice that
xm,k,l;n,r,t = Am(A + i)k(A − i)lBn(B + i)r(B − i)tx0,0,0;0,0,0,
for all (m, k, l; n, r, t) ∈ Ω. In fact, by induction we can check that
(33)
xm,k,l;n,r,t = (B − iEH )txm,k,l;n,r,0,
t ∈ Z,
for any fixed m, n ∈ Z+, k, l, r ∈ Z;
xm,k,l;n,r,0 = (B + iEH)rxm,k,l;n,0,0,
r ∈ Z,
for any fixed m, n ∈ Z+, k, l ∈ Z;
xm,k,l;n,0,0 = Bnxm,k,l;0,0,0,
n ∈ Z+,
for any fixed m ∈ Z+, k, l ∈ Z;
xm,k,l;0,0,0 = (A − iEH )lxm,k,0;0,0,0,
l ∈ Z,
for any fixed m ∈ Z+, k ∈ Z;
xm,k,0;0,0,0 = (A + iEH )lxm,0,0;0,0,0,
k ∈ Z,
for any fixed m ∈ Z+;
xm,0,0;0,0,0 = Amx0,0,0;0,0,0,
m ∈ Z+,
and then by substitution of each relation into previous one we obtain rela-
tion (33).
For the commuting self-adjoint operators A and B there exists an or-
thogonal operator spectral measure E(x) on B(R2) such that
A =ZR2
x1dE(x), B =ZR2
x2dE(x).
(34)
Then
=(cid:18)ZR2
um,k,l;n,r,t = (xm,k,l;n,r,t, x0,0,0;0,0,0)H
xm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)tdE(x)x0,0,0;0,0,0, x0,0,0;0,0,0(cid:19)H
7
=ZR2
Hence, the Borel measure
xm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)td(E(x)x0,0,0;0,0,0, x0,0,0;0,0,0)H .
µ = (E(x)x0,0,0;0,0,0, x0,0,0;0,0,0)H ,
(35)
is a solution of the moment problem (2).
Theorem 2 Let the extended two-dimensional moment problem (2) be given.
The moment problem has a solution if and only if conditions (3) and (9)-
(12) are satisfied. If these conditions are satisfied then the solution of the
moment problem is unique and can be constructed by (35).
Proof. The sufficiency of conditions (3) and (9)-(12) for the existence of
a solution of the moment problem (2) was shown before the statement of
the Theorem. The necessity of condition (3) was proved, as well. Let us
check that conditions (9)-(12) are necessary for the solvability of the moment
problem (2).
Let µ be a solution of the moment problem (2). Consider the space L2
the following subsets in L2
µ:
µ and
Lµ = Lin{xm
We denote
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)t}(m,k,l;n,r,t)∈Ω, Hµ = Lµ.
(36)
ym,k,l;n,r,t := xm
1 (x1 + i)k(x1− i)lxn
2 (x2 + i)r(x2− i)t,
(m, k, l; n, r, t) ∈ Ω.
(37)
Notice that
(ym,k,l;n,r,t, ym′,k′,l′;n′,r′,t′)L2
µ
= um+m′,k+l′,l+k′;n+n′,r+t′,t+r′,
(38)
for all (m, k, l; n, r, t), (m′, k′, l′; n′, r′, t′) ∈ Ω. Consider the operators of
multiplication by the independent variable in L2
µ:
Aµf (x1, x2) = x1f (x1, x2), Bµf (x1, x2) = x2f (x1, x2),
f ∈ L2
µ.
Notice that
µ
(Aµ + iEL2
(Aµ − iEL2
(Bµ + iEL2
µ
µ
)ym,k,l;n,r,t = ym,k+1,l;n,r,t,
)ym,k,l;n,r,t = ym,k,l+1;n,r,t,
)ym,k,l;n,r,t = ym,k,l;n,r+1,t,
8
(39)
(40)
(41)
(42)
(Bµ − iEL2
µ
)ym,k,l;n,r,t = ym,k,l;n,r,t+1,
(43)
for all (m, k, l; n, r, t), (m′, k′, l′; n′, r′, t′) ∈ Ω.
Since conditions (3) are satisfied, by Theorem 1 there exist a Hilbert space
H and a sequence of elements {xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω, in H, such that
relation (4) holds. Repeating arguments after the Proof of Theorem 1 we
construct operators A0 and B0 in H. Consider the following operator:
W0 X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,tym,k,l;n,r,t = X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,txm,k,l;n,r,t,
(44)
where all but finite number of complex coefficients αm,k,l;n,r,t are zeros. Let
us check that this operator is defined correctly. In fact, suppose that
αm,k,l;n,r,tym,k,l;n,r,t =
X(m′,k′,l′;n′,r′,t′)∈Ω
βm′,k′,l′;n′,r′,t′ ym′,k′,l′;n′,r′,t′,
(45)
where βm′,k′,l′;n′,r′,t′ ∈ C. We may write
x = X(m,k,l;n,r,t)∈Ω
0 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=
=
∗(αm′,k′,l′;n′,r′,t′ − βm′,k′,l′;n′,r′,t′)(ym,k,l;n,r,t, ym′,k′,l′;n′,r′,t′)L2
µ
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
(αm,k,l;n,r,t − βm,k,l;n,r,t)
(αm,k,l;n,r,t − βm,k,l;n,r,t)ym,k,l;n,r,t(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X(m,k,l;n,r,t)∈Ω
X
X
(αm,k,l;n,r,t − βm,k,l;n,r,t)xm,k,l;n,r,t(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)H
X(m,k,l;n,r,t)∈Ω
Thus, the operator W0 is defined correctly. Ifex ∈ H and
ex = X(m,k,l;n,r,t)∈Ω
(αm,k,l;n,r,t − βm,k,l;n,r,t)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
γm,k,l;n,r,tym,k,l;n,r,t,
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
2
L2
µ
.
∗(αm′,k′,l′;n′,r′,t′ − βm′,k′,l′;n′,r′,t′)(xm,k,l;n,r,t, xm′,k′,l′;n′,r′,t′)H
9
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
αm,k,l;n,r,tγm′,k′,l′;n′,r′,t′
X
where γm,k,l;n,r,t ∈ C, then
(W0x, W0ex)H =
=
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
X
∗(xm,k,l;n,r,t, xm′,k′,l′;n′,r′,t′)H
αm,k,l;n,r,tγm′,k′,l′;n′,r′,t′(ym,k,l;n,r,t, ym′,k′,l′;n′,r′,t′)H
By continuity we extend W0 to a unitary operator W which maps Hµ onto
H. Observe that
.
µ
= (x,ex)L2
W −1A0W ym,k,l;n,r,t = ym+1,k,l;n,r,t = Aµym,k,l;n,r,t,
(46)
W −1B0W ym,k,l;n,r,t = ym,k,l;n+1,r,t = Bµym,k,l;n,r,t,
(47)
for all (m, k, l; n, r, t) ∈ Ω. By using the last relations in relations (40)-
(43) we obtain relations (21)-(24). The latter relations are equivalent to
conditions (9)-(12).
Let us check that the solution of the moment problem is unique. Consider
the following transformation
T : (x1, x2) ∈ R2 7→ (ϕ, ψ) ∈ [0, 2π) × [0, 2π),
eiϕ =
x1 + i
x1 − i
,
eiψ =
x2 + i
x2 − i
;
ν(∆) = µ(T −1(∆)), ∆ ∈ B([0, 2π) × [0, 2π)).
(48)
(49)
and set
Since T is a bijective continuous transformation, then ν is a non-negative
Borel measure on [0, 2π) × [0, 2π). Moreover, we have
x1 − i(cid:19)k(cid:18) x2 + i
x2 − i(cid:19)l
u0,k,−k;0,l,−l =ZR2(cid:18) x1 + i
for all k, l ∈ Z. Let eµ be another solution of the moment problem (2) andeν
dµ =Z[0,2π)×[0,2π)
be defined by
eikϕeilψdν,
(50)
(51)
eν(∆) =eµ(T −1(∆)), ∆ ∈ B([0, 2π) × [0, 2π)).
10
By relation (50) we obtain that
Z[0,2π)×[0,2π)
eikϕeilψdν =Z[0,2π)×[0,2π)
eikϕeilψdeν,
k, l ∈ Z.
(52)
By the Weierstrass theorem we can approximate ϕm and ψn, for some fixed
m, n ∈ Z+, by trigonometric polynomials Pk(ϕ) and Rk(ψ), respectively:
1
k
,
ψ∈[0,2π)ψm − Rk(ψ) ≤
max
1
k
,
k ∈ N.
(53)
Then
ϕ∈[0,2π)ϕm − Pk(ϕ) ≤
max
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z[0,2π)×[0,2π)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z[0,2π)×[0,2π)
ϕmψndν −Z[0,2π)×[0,2π)
(ϕm − Pk(ϕ))ψndν +Z[0,2π)×[0,2π)
ψ∈[0,2π)ψm
ν([0, 2π)) + max
≤ max
1
k
ϕ∈[0,2π)Pk(ϕ)
1
k
Pk(ϕ)Rk(ψ)dν(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Pk(ϕ)(ψn − Rk(ψ))dν(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ν([0, 2π))
as k → ∞. Hence, we conclude that
(54)
as k → ∞. In the same manner we get
≤ max
ψ∈[0,2π)ψm
1
k
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z[0,2π)×[0,2π)
Z[0,2π)×[0,2π)
ν([0, 2π)) → 0,
k
k
+ max
ν([0, 2π)) +(cid:18) 1
ϕmψndeν −Z[0,2π)×[0,2π)
ϕmψndν =Z[0,2π)×[0,2π)
ϕ∈[0,2π)ϕm(cid:19) 1
Pk(ϕ)Rk(ψ)deν(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) → 0,
ϕmψndeν, m, n ∈ Z+.
Since the two-dimensional moment problem on a rectangular has a unique
solution, we get ν =eν and µ =eµ. ✷
3 An algorithm towards solving the two-dimensional
moment problem.
As a first application of our results on the extended two-dimensional moment
problem we get the following theorem.
11
Theorem 3 Let the two-dimensional moment problem (1) be given. The
moment problem has a solution if and only if there exists a sequence of com-
plex numbers {um,k,l;n,r,t}(m,k,l;n,r,t)∈Ω, which satisfies conditions (3), (9)-
(12) and
(55)
um,0,0;n,0,0 = sm,n,
m, n ∈ Z+.
The proof is obvious and left to the reader.
Let the two-dimensional moment problem (1) be given. As it is well
known (and can be checked in the same manner as for the relation (3)) the
necessary condition for its solvability is the following:
Xm,n,m′,n′∈Z+
αm,nαm′,n′sm+m′,n+n′ ≥ 0,
(56)
for arbitrary complex coefficients αm,n, where all but finite number of αm,n
are zeros.
We assume that the condition (56) holds. Repeating arguments of the proof
of Theorem 1 we can state that there exist a Hilbert space H0 and a sequence
{hm,n}m,n∈Z+ such that
(hm,n, hm′,n′)H0 = sm+m′,n+n′, m, n, m′, n′ ∈ Z+.
Consider the following Hilbert space:
(57)
(58)
H = H0 ⊕
∞Mj=1
Hj ,
where Hj are arbitrary one-dimensional Hilbert spaces, j ∈ N. We shall call
it the model space for the two-dimensional moment problem.
Introduce an arbitrary indexation in the set Ω′ by the unique positive integer
index j:
j ∈ N 7→ w = w(j) = (m, k, l; n, r, t)(j) ∈ Ω′.
(59)
Suppose that the two-dimensional moment problem (1) has a solution µ.
Consider the space L2
µ and the following subsets in L2
µ:
Lµ,0 = Lin{xm
1 xn
2}m,n∈Z+ , Hµ,0 = Lµ,0.
We denote
Notice that
ym,n := xm
1 xn
2 ,
m, n ∈ Z+.
(ym,n, ym′,n′)L2
µ
= sm+m′,n+n′,
12
(60)
(61)
(62)
for all m, n, m′, n′ ∈ Z+. We shall also use the notations from (36),(37).
Define the following numbers
um,k,l;n,r,t :=ZR2
xm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)tdµ,
(m, k, l; n, r, t) ∈ Ω.
(63)
For these numbers conditions (3) hold and repeating arguments after the
relation (43) we construct a Hilbert space H and a sequence of elements
{xm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω, in H, such that relation (4) holds. We introduce
the operator W as after (44). The operator W maps Hµ onto H. Set
H0 = span{xm,0,0;n,0,0}m,n∈Z+ ⊆ H.
(64)
Let us construct a sequence of Hilbert spaces Hj, j ∈ N, in the following
way.
Step 1. We set
f1 = xw(1) − P H
H0 xw(1), H1 = span{f1},
where w(·) is the indexation in the set Ω′.
Step r, with r ≥ 2. We set
fr = xw(r) − P H
H0⊕(⊕1≤t≤r−1Ht)xw(1), Hr = span{fr}.
Then we get a representation
Observe that Hj is either a one-dimensional Hilbert space or Hj = {0}. We
denote
(68)
Λµ = {j ∈ N : Hj 6= {0}},
Λ′
µ = N\Λµ.
Then
We shall construct a unitary operator U which maps H onto the following
subspace of the model space H:
(65)
(66)
(67)
(69)
(70)
H = H0 ⊕
∞Mj=1
Hj .
H = H0 ⊕Mj∈Λµ
bH = H0 ⊕Mj∈Λµ
Hj .
Hj ⊂ H.
13
Choose an arbitrary element
x = Xm,n∈Z+
αm,nxm,0,0;n,0,0 + Xj∈Λµ
βj
fj
kfjkH
,
with αm,n, βj ∈ C. Set
U x = Xm,n∈Z+
αm,nhm,n + Xj∈Λµ
βj ej ,
(71)
(72)
where ej ∈ Hj, kejkH = 1, are chosen arbitrarily.
Let us check that this definition is correct. Suppose that x has another
representation:
(73)
fj
kfjkH
,
Xm,n∈Z+
= Xm,n,m′,n′∈Z+
with eαm,n,eβj ∈ C. By orthogonality we have βj = eβj, j ∈ N. Then
x = Xm,n∈Z+eαm,nxm,0,0;n,0,0 + Xj∈Λµeβj
0 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(αm,n −eαm,n)xm,0,0;n,0,0(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(αm,n −eαm,n)(αm′,n′ −eαm′,n′)(xm,0,0;n,0,0, xm′,0,0;n′,0,0)H
= Xm,n,m′,n′∈Z+
(αm,n −eαm,n)(αm′,n′ −eαm′,n′)(hm,n, hm′,n′)H
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(αm,n −eαm,n)hm,n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)H
Xm,n∈Z+
Thus, the operator U is defined correctly. If bx ∈ H and
bx = Xm,n∈Z+bαm,nxm,0,0;n,0,0 + Xj∈Λµbβj
fj
kfjkH
,
2
H
.
where bαm,n,bβj ∈ C, then
(x,bx)H = Xm,n,m′,n′∈Z+
αm,nbαm′,n′(xm,0,0;n,0,0, xm′,0,0;n′,0,0)H + Xj∈Λµ
βjbβj
14
= Xm,n,m′,n′∈Z+
αm,nbαm′,n′(hm,n, hm′,n′)H + Xj∈Λµ
= (U x, Ubx)H.
βjbβj
By continuity we extend U to a unitary operator which maps H onto bH.
Then the operator U W is a unitary operator which maps Hµ onto bH. We
could define this operator directly, but we prefer to underline an abstract
structure of the corresponding spaces and this maybe explains where the
model space comes from.
We set
hm,k,l;n,r,t := U W ym,k,l;n,r,t,
(m, k, l; n, r, t) ∈ Ω.
Observe that
hm,0,0;n,0,0 = hm,n, m, n ∈ Z+; U W Hµ,0 = H0.
(74)
(75)
(76)
(77)
Since
then
xw(r) ∈ H0 ⊕ Mj∈Λµ: j≤r
Hj ,
Hj ,
hw(r) ∈ H0 ⊕ Mj∈Λµ: j≤r
r ∈ N.
Observe that {ym,k,l;n,r,t}(m,k,l;n,r,t)∈Ω satisfy relations (17)-(20) (with y in-
stead of x). Therefore {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω satisfy relations, as well.
Notice that
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (ym,k,l;n,r,t, ym′,k′,l′;n′,r′,t′)L2
µ
= um+m′,k+l′,l+k′;n+n′,r+t′,t+r′,
(78)
for all (m, k, l; n, r, t), (m′, k′, l′; n′, r′, t′) ∈ Ω.
Theorem 4 Let the two-dimensional moment problem (1) be given. Choose
an arbitrary model space H with a sequence {hm,n}m,n∈Z+, satisfying (57)
and fix it. The moment problem has a solution if and only if there exists
a sequence {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω, in H such that the following conditions
hold:
1) hm,0,0;n,0,0 = hm,n, m, n ∈ Z0;
15
2) hw(r) ∈ H0 ⊕(cid:16)Lj∈Λ: j≤r Hj(cid:17) ,
Here w(r) is an arbitrary indexation in Ω′ by indices from N.
r ∈ N, for some subset Λ ⊆ N.
3) The sequence {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω satisfies conditions (17)-(20) (with
h instead of x).
4) There exists a complex function ϕ(m, k, l; n, r, t), (m, k, l; n, r, t) ∈ Ω,
such that
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = ϕ(m+m′, k+l′, l+k′; n+n′, r+t′, t+r′),
for all (m, k, l; n, r, t), (m′, k′, l′; n′, r′, t′) ∈ Ω.
The necessity of conditions 1)-4) for the solvability of the two-
Proof.
dimensional moment problem was established before the statement of the
Theorem.
Let conditions 1),3),4) be satisfied. Consider the extended two-dimensional
moment problem (2) with
um,k,l;n,r,t := ϕ(m, k, l; n, r, t),
(m, k, l; n, r, t) ∈ Ω,
(79)
where ϕ is from the condition 4). Then
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
=
(m,k,l;n,r,t),(m′,k′,l′;n′,r′,t′)∈Ω
αm,k,l;n,r,tαm′,k′,l′;n′,r′,t′ um+m′,k+l′,l+k′;n+n′,r+t′,t+r′
αm,k,l;n,r,tαm′,k′,l′;n′,r′,t′(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H
X
X
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X(m,k,l;n,r,t)∈Ω
αm,k,l;n,r,thm,k,l;n,r,t(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≥ 0,
2
H
for arbitrary complex coefficients αm,k,l;n,r,t, where all but finite number of
αm,k,l;n,r,t are zeros.
By conditions 3) and 4) we conclude that conditions (9)-(12) hold. By
Theorem 2 we obtain that there exists a non-negative Borel measure µ in
R2 such that (2) holds. In particular, using conditions 4),1) we get
ZR2
1 xn
xm
2 dµ = um,0,0;n,0,0 = ϕ(m, 0, 0; n, 0, 0)
= (hm,0,0;n,0,0, h0,0,0;0,0,0)H = (hm,n, h0,0)H = sm,n, m, n ∈ Z+.
16
✷
Observe that condition 2) can be removed from the statement of Theo-
rem 4. However, it will be used later.
Denote a set of sequences {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω, in H satisfying con-
ditions 1)-4) by X = X(H). As we have seen in the proof of Theorem 4,
for an arbitrary {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω ∈ X(H), the unique solution of the
extended two-dimensional moment problem with moments (79) gives a solu-
tion of the two-dimensional moment problem. Observe that all solutions of
the two-dimensional moment problem can be constructed in this manner. In-
deed, let µ be an arbitrary solution of the two-dimensional moment problem.
Repeating arguments from relation (60) till the statement of Theorem 4 we
may write
ZR2
xm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)tdµ = (ym,k,l;n,r,t, y0,0,0;0,0,0)L2
µ
= (U W ym,k,l;n,r,t, U W y0,0,0;0,0,0)H = (hm,k,l;n,r,t, h0,0,0;0,0,0)H
= ϕ(m, k, l; n, r, t) = um,k,l;n,r,t,
for all (m, k, l; n, r, t) ∈ Ω. Here the operators U ,W , the sequence {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω,
the function ϕ(m, k, l; n, r, t) and the moments um,k,l;n,r,t, of course, depend
on the choice of µ. Notice that {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω ∈ X(H).
For such constructed parameters, the measure µ is a solution of the extended
two-dimensional moment problem considered in the proof of Theorem 4.
Since the solution of this moment problem is unique, µ will be reconstructed
in the above described manner.
Notice that condition 4) of Theorem 4 is equivalent to the following
conditions:
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (h em,k,l;n,r,t, h em′,k′,l′;n′,r′,t′)H,
(80)
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (hm,ek,l;n,r,t, hm′,k′,el′;n′,r′,t′)H,
if m, m′,em,em′, n, n′ ∈ Z+, k, l, r, t, k′, l′, r′, t′ ∈ Z: m + m′ = em + em′;
if m, m′, n, n′ ∈ Z+, k,ek, l, r, t, k′, l′,el′, r′, t′ ∈ Z: k + l′ =ek +el′;
if m, m′, n, n′ ∈ Z+, k, l,el, r, t, k′,ek′, l′, r′, t′ ∈ Z: l + k′ =el +ek′;
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (hm,k,el;n,r,t, hm′,ek′,l′;n′,r′,t′)H,
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (hm,k,l;en,r,t, hm′,k′,l′;en′,r′,t′)H,
(81)
(82)
(83)
17
(84)
(85)
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (hm,k,l;n,er,t, hm′,k′,l′;n′,r′,et′)H,
if m, m′, n, n′,en,en′ ∈ Z+, k, l, r, t, k′, l′, r′, t′ ∈ Z: n + n′ =en +en′;
if m, m′, n, n′ ∈ Z+, k, l, r,er, t, k′, l′, r′, t′,et′ ∈ Z: r + t′ =er +et′;
if m, m′, n, n′ ∈ Z+, k, l, r, t,et, k′, l′, r′,er′, t′ ∈ Z: t + r′ =et +er′.
(hm,k,l;n,r,t, hm′,k′,l′;n′,r′,t′)H = (hm,k,l;n,r,et, hm′,k′,l′;n′,er′,t′)H,
Let {gn}∞
As we can see, the solving of the two-dimensional moment problem re-
duces to a construction of the set X(H). Let us describe an algorithm for a
construction of sequences from X(H).
n=1 be an arbitrary orthonormal basis in H0 obtained by the
Gram-Schmidt orthogonalization procedure from the sequence {hm,n}m,n∈Z+
indexed by a unique index.
Choose an arbitrary j ∈ N. Let w(j) = (m, k, l; n, r, t)(j) ∈ Ω′. If we
had constructed {hm,n}m,n∈Z+ ∈ X(H), then the two-dimensional moment
problem has a solution µ and
H = kxm
1 (x1 + i)k(x1 − i)lxn
2 (x2 + i)r(x2 − i)tk2
L2
µ
dj := khw(j)k2
=ZR2
x2m
1 (x2
1 + 1)k(x2
1 + 1)lx2n
2 (x2
2 + 1)r(x2
2 + 1)tdµ.
Therefore dj are bounded by some constants Mj = Mj(S) depending on the
prescribed moments S := {sm,n}m,n∈Z+ . (Notice that e.g. (x2
1 + 1)l ≤ 1,
for l < 0, and for non-negative m, k, l; n, r, t the values of dj are determined
uniquely).
Step 0. We set
hm,0,0;n,0,0 = hm,n,
m, n ∈ Z0.
(86)
We check that conditions (13)-(16) (with h instead of x) and (80)-(85)
are satisfied for hm,0,0;n,0,0, m, n ∈ Z0. If they are not satisfied, the two-
dimensional moment problem has no solution and we stop the algorithm.
Step 1. We seek for hw(1) in the following form:
hw(1) =
∞Xn=1
α1;ngn + β1e1,
(87)
with some complex coefficients α1;n, β1.
18
Conditions (13)-(16) (with h instead of x) and (80)-(85) which include hw(1)
and the already constructed hm,k,l;n,r,t are equivalent to a set L1 of linear
equations with respect to α1;n, n ∈ N , and d1 = khw(1)k2
H. Notice that they
depend on β1 only by d1. Denote the set of solutions of these equations by
S1 = {(α1;n, n ∈ N ; d1) :
equations from L1 are satisfied} .
(88)
Set
Finally, we set
bS1 =((α1;n, n ∈ N ; d1) ∈ S1 :
G1 =
α1;ngn + eiθ d1 −
∞Xn=1
∞Xn=1
∞Xn=1
α1;n2! 1
2
hw(r) =
∞Xn=1
αr;ngn +
(89)
α1;n2 ≤ d1, d1 ≤ M1) .
e1 : (α1;n, n ∈ N ; d1) ∈ bS1, θ ∈ [0, 2π)
rXj=1
βj ej ,
(91)
(90)
.
Step r, with r ≥ 2. We seek for hw(r) in the following form:
with some complex coefficients αr;n, βj.
Conditions (13)-(16) (with h instead of x) and (80)-(85) which include hw(r)
and the already constructed hm,k,l;n,r,t are equivalent to a set Lr of linear
equations with respect to αr;n, n ∈ N , βj, 1 ≤ j ≤ r− 1, and dr = khw(r)k2
H,
and depending on parameters (hw(1), hw(2), ..., hw(r−1)) ∈ Gr−1 .
Notice that these linear equations depend on βr only by dr. Denote the set
of solutions of these equations by
(hw(1), hw(2), ..., hw(r−1)) ∈ Gr−1,
Sr =(cid:8)(αr;n, n ∈ N ; βj , 1 ≤ j ≤ r − 1; dr; hw(1), hw(2), ..., hw(r−1)) :
and equations from Lr with parameters (hw(1), hw(2), ..., hw(r−1)), are satisfied(cid:9) .
bSr =(cid:8)(αr;n, n ∈ N ; βj , 1 ≤ j ≤ r − 1; dr; hw(1), hw(2), ..., hw(r−1)) ∈ Sr :
(92)
Set
(93)
∞Xn=1
αr;n2 +
r−1Xj=1
βr;j2 ≤ dr, dr ≤ Mr
.
19
αr;ngn +
r−1Xj=1
∞Xn=1
1
2
Finally, we set
Gr =(cid:8)(cid:0)hw(1), hw(2), ..., hw(r−1),
βr;jej + eiθdr −
r−1Xj=1
αr;n2 −
er :
∞Xn=1
(αr;n, n ∈ N ; βj , 1 ≤ j ≤ r − 1; dr; hw(1), hw(2), ..., hw(r−1)) ∈ bSr, θ ∈ [0, 2π)o .
βr;j2
Final step. Consider a space H of sequences
(94)
h = (h1, h2, h3, ...),
hr ∈ H, r ∈ N,
with the norm given by
khkH = sup
r∈N
1
√Mr khrkH < ∞.
(95)
(96)
For arbitrary (h1, ..., hr) ∈ Gr, we put into correspondence elements h ∈ H
of the following form
h = (h1, ..., hr, gr+1, gr+2, ...) : gj ∈ H, kgjkH ≤pMj , j > r.
Thus, the set Gr is mapped onto a set Gr ⊂ H. Observe that all elements
of Gr has the norm less or equal to 1. Set
(97)
G =
Gr.
∞\r=1
(98)
If G 6= ∅, then to each (g1, g2, ...) ∈ G, we put into correspondence a se-
quence H = {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω such that (86) holds and
hw(r) := gr,
r ∈ N.
(99)
We state that H ∈ X(H). In fact, conditions (13)-(16) (with h instead of x)
and (80)-(85) are satisfied for hm,0,0;n,0,0, m, n ∈ Z+, by Step 0. If one of
these equations include hw(r) with r ≥ 1, then we choose the maximal ap-
pearing index r. Since (hw(1), ..., hw(r)) ∈ Gr, then this equation is satisfied.
Condition 2) is satisfied by the construction.
Thus, if G 6= ∅, then using H we can construct a solution of the two-
dimensional moment problem in the described above manner.
20
Theorem 5 Let the two-dimensional moment problem (1) be given. Choose
an arbitrary model space H with a sequence {hm,n}m,n∈Z+, satisfying (57)
and fix it. The moment problem has a solution if and only if conditions (13)-
(16) (with h instead of x) and (80)-(85) are satisfied for hm,0,0;n,0,0 := hm,n,
m, n ∈ Z+, and
(100)
G 6= ∅,
where G is constructed by (98) according to the algorithm.
If the latter conditions are satisfied then to each (g1, g2, ...) ∈ G, we put
into correspondence a sequence H = {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω such that (86)
holds and
(101)
This sequence belongs to X(H) and the unique solution of the extended
two-dimensional moment problem with moments (79) gives a solution µ of
the two-dimensional moment problem. Moreover, all solutions of the two-
dimensional moment problem can be obtained in this way.
hw(r) := gr,
r ∈ N.
Proof. The sufficiency of the conditions in the statement of the Theorem
for the solvability of the two-dimensional moment problem was shown be-
fore the statement of the Theorem. Let us show that these conditions are
necessary.
Let µ be a solution of the two-dimensional moment problem. By Theorem 4
stead of x) and (80)-(85) are satisfied. In particular, they are satisfied for
the set X(H) is not empty. Choose an arbitrarybH = {bhm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω ∈
X(H). By conditions 3),4) we see that conditions (13)-(16) (with bh in-
bhm,0,0;n,0,0 = hm,n, m, n ∈ Z+.
Comparing condition 2) with Steps 1 and r for r ≥ 2, we see that
Therefore elements
(bhw(1), ...,bhw(r)) ∈ Gr,
r ∈ N.
(102)
(103)
(104)
where gj ∈ H : kgjkH ≤pMj, for j > r are arbitrary. Thus, the element
(bhw(1), ...,bhw(r), gr+1, gr+2, ...) ∈ Gr,
bh := (bhw(1),bhw(2),bhw(3), ...) ∈ Gr,
r ∈ N,
r ∈ N.
and
Gr = G.
(105)
bh ∈ \r∈N
21
Therefore G 6= ∅.
If the conditions of the Theorem are satisfied then to each g = (g1, g2, ...) ∈
G, we put into correspondence a sequence H = H(g) = {hm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω
such that (86) and (101) hold. Then H ∈ X(H), as it was shown before the
statement of the Theorem. The sequence H generates a solution of the
extended two-dimensional moment problem and of the two-dimensional mo-
ment problem, see considerations after the proof of Theorem 4.
It remains to show that all solutions of the two-dimensional moment problem
can be obtained in this way. Since elements of X(H) generate all solutions
of the two-dimensional moment problem (see considerations after the proof
of Theorem 4), it remains to prove that
{H(g) : g ∈ G} = X(H).
(106)
On the other hand, choose an arbitrary eH = {ehm,k,l;n,r,t}(m,k,l;n,r,t)∈Ω ∈
Denote the set on the left-hand side by X1. It was shown that X1 ⊆ X(H).
X(H). Repeating the construction at the beginning of this proof we obtain
that
(107)
Observe that
eh := (ehw(1),ehw(2),ehw(3), ...) ∈ G.
H(eh) =eH.
Therefore X(H) ⊆ X1 and relation (106) holds. ✷
Remark. The truncated two-dimensional moment problem can be consid-
ered in a similar manner. Moreover, the set of indices of the known elements
hm,n will be finite in this case and therefore equations in the r-th step of the
algorithm will form finite systems of linear equations. Thus, the r-th step
could be easily performed using computer.
(108)
References
[1] J. A. Shohat, J. D. Tamarkin, The Problem of Moments, Amer. Math.
Soc., New York City, 1943.
[2] N. I. Akhiezer, Classical Moment Problem, Fizmatlit., Moscow, 1961.
(in Russian).
[3] Ju. M. Berezanskii, Expansions in Eigenfunctions of Selfadjoint Op-
erators, Amer. Math. Soc., Providence, RI, 1968. (Russian edition:
Naukova Dumka, Kiev, 1965).
22
[4] B. Fuglede, The multidimensional moment problem, Expo. Math. 1
(1983), 47 -- 65.
[5] Yu. M. Berezansky, Spectral theory of the infinite block Jacobi type nor-
mal matrices, orthogonal polynomials on the complex domain, and the
complex moment problem, Operator Theory: Advances and Applica-
tions 191 (2009), 37 -- 50.
[6] J. Stochel, F. H. Szafraniec, The complex moment problem and subnor-
mality: a polar decomposition approach, J. of Functional Analysis 159
(1998), 432 -- 491.
[7] S. M. Zagorodnyuk, Positive definite kernels satisfying difference equa-
tions, Methods of Functional Analysis and Topology 16 (2010), 83 -- 100.
[8] N. I. Akhiezer, I.M. Glazman, Theory of Linear Operators in a Hilbert
Space, Gos. izdat. teh.-teor. lit., Moscow, Leningrad, 1950. (in Russian).
On the two-dimensional moment problem.
S.M. Zagorodnyuk
In this paper we obtain an algorithm towards solving the two-dimensional
moment problem. This algorithm gives the necessary and sufficient condi-
tions for the solvability of the moment problem. It is shown that all solutions
of the moment problem can be constructed using this algorithm.
23
|
1809.04829 | 1 | 1809 | 2018-09-13T08:25:56 | Weighted composition operators on the Fock space | [
"math.FA"
] | In this paper, we study weighted composition operators on the Fock space. We show that a weighted composition operator is cohyponorma if and only if it is normal. Moreover, we give a complete characterization of closed range weighted composition operators. Finally, we find norms of some weighted composition operators. | math.FA | math |
Weighted composition operators on the Fock space
Mahsa Fatehi
September 14, 2018
Abstract
In this paper, we study weighted composition operators on the Fock
space. We show that a weighted composition operator is cohyponorma
if and only if it is normal. Moreover, we give a complete characteriza-
tion of closed range weighted composition operators. Finally, we find
norms of some weighted composition operators.
1
1
Introduction
Let D denote the open unit disk in the complex plane C. For an analytic map
ϕ, let ϕ0 be the identity function, ϕ1 = ϕ and ϕn+1 = ϕ◦ ϕn for n = 1, 2, ....
We call them the iterates of ϕ. It is well-known that if ϕ, neither the identity
nor an elliptic automorphism of D (i.e., ϕ is an automorphism of D with a
fixed point in D), is an analytic map on the unit disk into itself, then there
exists a point w in D such that ϕn converges to w uniformly on compact
subsets of D. The point w is called the Denjoy-Wolff point of ϕ. The Denjoy-
Wolff point w is the unique fixed point of ϕ in D so that ϕ′(w) ≤ 1 (see
[7]).
Recall that the Fock space F 2 is a Hilbert space of all entire functions on
C that are square integrable with respect to the Gaussian measures dµ(z) =
π−1e−z2
dA(z), where dA is the usual Lebesgue measure on C. The Fock
space F 2 is a reproducing kernel Hilbert space with inner product
hf, gi =ZC
f (z)g(z)dµ(z)
1AMS Subject Classifications. Primary 47B33.
Key words and phrases: Fock space, Weighted composition operator, Normaloid, Cohy-
ponormal, Norm, Closed range.
1
(see [16]).
and reproducing kernel function Kw(z) = ewz for any w ∈ C. Note that
for any w ∈ C, kKwk = ew2/2. For each w ∈ C, we define the normalized
reproducing kernel as kw(z) = Kw(z)
= ewz−w2/2. For each nonnegative
kKwk
integer n, let en(z) = zn/√n!. The set {en} is an orthonormal basis for F 2
For entire function ϕ on C, the composition operator Cϕ on F 2 is defined
as Cϕ(f ) = f ◦ ϕ for any f ∈ F 2; moreover, for ψ ∈ F 2, the weighted
composition operator Cψ,ϕ is defined by Cψ,ϕf = ψ · (f ◦ ϕ). There exists
some literature on composition operators acting on the Hardy and Bergman
spaces. The books [7] and [11] are the important references.
Carswell et al. [4] characterized the bounded and compact composition
operators on the Fock space over Cn. Specified to the one-dimensional case,
they stated that Cϕ is bounded if and only if ϕ(z) = az, where a = 1
or ϕ(z) = az + b with a < 1. In [12], Ueki found the criteria to charac-
terize the boundedness and compactness of weighted composition operators
on the Fock space. After that in [10], Le obtained much easier criteria for
the boundedness and compactness of weighted composition operators on
the Fock space. On the Hardy space, normal weighted composition opera-
tors were studied; moreover, unitary weighted composition operators were
characterized (see [3]). Also in [6], cohyponormal weighted composition op-
erators were obtained. After that in [8], hyponormal weighted composition
operators were investigated on the Hardy and weighted Bergman spaces.
Unitary weighted composition operators Cψ,ϕ on the Fock space were ob-
tained in [13]. Also invertible weighted composition operators on the Fock
space were characterized in [14]. In the second section, we find normaloid,
hyponormal and cohyponormal composition operators. After that we obtain
all hyponormal weighted composition operators Cψ,ϕ, where ψ = Kc for each
c ∈ C. Moreover, we study a class of normaloid weighted composition oper-
ators. Next, we show that for ϕ(z) = az + b, Cψ,ϕ is cohyponormal if and
. Closed range composition operators were studied
only if ψ = ψ(0)Kb a−1
on the Hardy and weighted Bergman spaces in [1], [9] and [17]. In the third
section, we characterize closed range weighted composition operators on the
Fock space. In the fourth section, we find norm of Cψ,ϕ on F 2, when ψ = Kc
for any c ∈ C.
a−1
2
2 Normaloid weighted composition operators
Suppose that T is a bounded operator on a Hilbert space. Throughout
this paper, the spectrum of T , the essential spectrum of T , and the point
spectrum of T are denoted by σ(T ), σe(T ), σp(T ) respectively. Also the
spectral radius of T is denoted by r(T ).
Le [10] studied the boundedness of weighted composition operator on the
Fock space. His result shows that if Cψ,ϕ is bounded on F 2, then ϕ(z) =
az + b, where a ≤ 1; furthermore, he proved that if a = 1 and Cψ,az+b is
bounded on F 2, then ψ = ψ(0)K−ab. We use these facts frequently in this
paper and so throughout this paper, we assume that ϕ(z) = az + b, where
a ≤ 1.
[4], [13] and [15] were written on another Fock space (see [16, p.
33]), but their results hold for F 2 which is considered in this paper, with
identical arguments.
In the following proposition, we investigate σp(Cψ,ϕ), when ϕ(z) = az +b
and a < 1. Note that in the case that a = 1, as we saw in the preceding
paragraph, ψ = ψ(0)K−ab. Then Cψ,ϕ is a constant multiple of a unitary
operator from [13, Corollary 1.2]. Moreover, in [13, Corollary 1.4], the spec-
trum of unitary weighted composition operators were characterized.
Proposition 2.1. Let ψ and ϕ be entire functions on C and Cψ,ϕ be a
bounded weighted composition operator on F 2. Suppose that ϕ(z) = az + b,
where a < 1 and b ∈ C. If λ ∈ σp(Cψ,ϕ), then λ ≤ ψ( b
1−a ). Moreover, if
ψ( b
1−a ) = 0 and ϕ and ψ are nonconstant, then Cψ,ϕ has no eigenvalues.
Proof. Suppose that a < 1. First we find a weighted composition
operator C eψ, eϕ which is unitary equivalent to Cψ,ϕ such that the fixed point
of eϕ lies in D and eϕ is a self-map of D. There exists a positive integer N
such that a + 1
N 1 − a < 1. It is not hard to see that there is a complex
number u ∈ C such that u + b
N . By [13, Corollary 1.2], we know
that Cku,z−u is unitary (the operator Cku,z−u is known as the Weyl unitary).
By [10, Proposition 3.1], C∗ku,z−u = Ck−u,z+u, so
1−a < 1
Cku,z−uCψ,ϕCk−u,z+u =
=
=
1
1
kKuk2 Ceuz,z−uCψ,ϕCe−uz,z+u
kKuk2 euz · ψ(z − u) · e(−u(az+b))◦(z−u)C(z+u)◦(az+b)◦(z−u)
kKuk2 euz · ψ(z − u) · e−u(az−au+b)Caz+u(1−a)+b.
(1)
1
3
Let eϕ(z) = az + u(1 − a) + b and
1
eψ(z) =
kKuk2 euz · ψ(z − u) · e−u(az−au+b).
(2)
It is easy to see that the fixed point of eϕ is u + b
1−a which belongs to
D. Because a + u(1 − a) + b < 1, eϕ is a self-map of D. Since Cψ,ϕ
is unitary equivalent to C eψ, eϕ, σp(Cψ,ϕ) = σp(C eψ, eϕ). Assume that λ is a
nonzero eigenvalue for C eψ, eϕ with corresponding eigenvector h. We obtain
for each z ∈ C and positive integer n. For any fixed point z ∈ C, we obtain
λnh(z) = Πn−1
j=0 eψ(eϕj(z))h(eϕn(z))
h(eϕn(z)) = hh ◦ eϕn, Kzi
≤ kh ◦ eϕnkkKzk
= kh ◦ eϕnke
= kC eϕn(h)ke
≤ kC eϕnkkhke
z2
2 .
z2
2
z2
2
(3)
(4)
1−a is the Denjoy-Wolff point of eϕ, eϕj(z) → u + b
Since h is not the zero function, we can choose z ∈ D such that h(z) 6=
0. Since u + b
1−a and
eψ(eϕj(z)) → eψ(u + b
1−a ) as j → ∞. Take n-th roots of the absolute value
each side of Equation (3), use Equation (4) and let n → ∞, we get λ ≤
eψ(u + b
1−a )r(Cϕ). From Equation (2) and [15, Theorem 1.1], we see that
λ ≤ eψ(u + b
1−a )r(Cϕ) = ψ( b
1−a ). Then
b
).
(5)
λ ≤ ψ(
1−a ) = ψ( b
1 − a
Now assume that eψ(u + b
1−a ) = 0 and ϕ and ψ are nonconstant.
Thus, by Equation (5), λ = 0 is the only possible eigenvalue for Cψ,ϕ. Since
ψ is not the zero function and ϕ is not constant, the Open Mapping Theo-
rem implies that 0 cannot be an eigenvalue for Cψ,ϕ (see idea of the proof
of [2, Lemma 4.1]).
(cid:3)
Suppose that T is a bounded operator. The operator T is hyponor-
mal (cohyponormal) if T ∗T ≥ T T ∗ (T ∗T ≤ T T ∗). Also T is normaloid if
kTk = r(T ). It is well known that hyponormal (cohyponormal) operators
4
are normaloid. In the following proposition, we characterize hyponormal,
cohyponormal and normaloid composition operators on F 2.
Proposition 2.2. Let ϕ(z) = az + b, where a ≤ 1 and b ∈ C. Then
Cϕ is a bounded normaloid (hyponormal or cohyponormal) operator if and
only if b = 0.
1
2
b2
Proof. Let Cϕ be normaloid (hyponormal or cohyponormal). Suppose
that b 6= 0. Then [4, Theorem 1] states that a < 1. By [4, Theorem 4],
1−a2 (note that the inner product for F 2 in this paper is different
kCϕk = e
from [4], so the norm of a composition operator is not exactly the same as [4,
Theorem 4]; furthermore, in Remark 4.1, kCϕk will be described). Also [15,
Theorem 1.1] implies that r(Cϕ) = 1. Since Cϕ is normaloid (hyponormal
b2
1−a2 = 1. Hence b = 0 which is a contradiction.
or cohyponormal), e
Conversely, suppose that ϕ(z) = az, where a ≤ 1. Invoking [4, Lemma 2],
C∗ϕ = Caz. Then Caz is normal and the result follows.
(cid:3)
1
2
Suppose that Cψ,ϕ is a bounded weighted composition operator and
ϕ(z) = az + b. Note that if a = 1, then from [13, Corollary 1.2] and as
we saw in the second paragraph of this section, Cψ,ϕ is a constant multiplie
of a unitary operator. Thus, Cψ,ϕ is normal, normaloid, hyponormal and
cohyponormal. Hence we state the following proposition for a 6= 1.
Proposition 2.3. Let ψ = Kc and ϕ(z) = az + b, where a ≤ 1, a 6= 1
and b, c ∈ C. Suppose that Cψ,ϕ is bounded on F 2. Then the following are
equivalent.
(a) Cψ,ϕ is hyponormal.
(b) Cψ,ϕ is cohyponormal.
(c) Cψ,ϕ is normaloid.
(d) c = b a−1
a−1 .
e−u2
We can see that Cψ,ϕ is unitarily equivalent to ψ( b
Proof. There is u ∈ C such that c = u(a − 1). By Equation (1)
and some calculation, Cψ,ϕ is unitarily equivalent to C eψ, eϕ, where eψ(z) =
euz · ψ(z − u) · e−u(az−au+b) = ψ( b
1−a ) and eϕ(z) = az + u(1 − a) + b.
(a) ⇒ (d). Suppose that Cψ,ϕ is hyponormal. Then Caz+u(1−a)+b is
hyponormal. Proposition 2.2 implies that u(1 − a) + b = 0. Since u = c
a−1 ,
we conclude that c = b a−1
a−1 .
1−a )Caz+u(1−a)+b.
5
(d) ⇒ (a). Assume that c = b a−1
a−1 . Let u = b
and some calculation, Cψ,ϕ is unitarily equivalent to e
az. We infer from Proposition 2.2 that C eϕ is hyponormal and so Cψ,ϕ is
hyponormal.
a−1 . By Equation (1)
1−a C eϕ, where eϕ(z) =
b2
(b) ⇔ (d). The idea of the proof is similar to (a) ⇔ (d).
(c) ⇔ (d). The idea of the proof is similar to (a) ⇔ (d).
Suppose that ϕ(z) = az + b, where a = 1 and ψ is an entire function.
If Cψ,ϕ is bounded on F 2, then as we saw in the second paragraph of this
section, ψ(z) = ψ(0)K−β , where β = ab. Hence Cψ,ϕ is a constant multiple
of the unitary operator (see [13, Corollary 1.2]). It shows that in this case
Cψ,ϕ is normaloid and so in the next theorem, we assume that a < 1.
(cid:3)
Theorem 2.4. Suppose that ψ is an entire function and is not identi-
cally zero. Assume that ϕ(z) = az + b, where a < 1 and b ∈ C. Let for
each λ ∈ σe(Cψ,ϕ), λ ≤ ψ( b
1−a ). Then Cψ,ϕ is normaloid if and only if
ψ = ψ(0)Kb a−1
.
a−1
Proof. Let p = b
1−a (note that it is obvious that p is the fixed point of
ϕ). By [13, Corollary 1.2], Ckp,z−p is unitary. Furthermore, [10, Proposition
3.1] states that C∗kp,z−p = Ck−p,z+p. We obtain
H : = C∗kp,z−pCψ,ϕCkp,z−p
= Ck−p,z+pCψ,ϕCkp,z−p
= Cq, eϕ,
where
and
eϕ(z) = ϕ(z + p) − p = a(z + p) + b − p = az
q(z) = k−p(z)kp(ϕ(z + p))ψ(z + p)
= epb+p2(a−1)ep(a−1)zψ(z + p)
= ep(a−1)zψ(z + p).
(6)
(7)
(8)
It is not
Since Cψ,ϕ is unitary equivalent to Cq, eϕ, σe(Cψ,ϕ) = σe(Cq, eϕ).
hard to see that q(0) = ψ(p). Now suppose that Cψ,ϕ is normaloid. Then
Cq, eϕ is normaloid. Since Cψ,ϕ and Cq, eϕ are unitary equivalent, for each
6
λ ∈ σp(Cq, eϕ), λ ≤ q(0). We infer from [5, Proposition 6.7, p. 210] and
[5, Proposition 4.4, p. 359] that r(Cq, eϕ) ≤ q(0). Since Cq, eϕ is normaloid,
q(0) ≥ kCq, eϕk ≥ kCq, eϕ(1)k = kqk. We know that { zm
: m ≥ 0} is an
√m!
orthonormal basis for F 2. Then kqk ≥ q(0). It shows that q must be con-
stant. Thus, Equation (8) shows that ψ(z) · ep(a−1)(z−p) is constant. Then
ψ(z) = ψ(0)e−p(a−1)z = ψ(0)eb a−1
Conversely, suppose that ψ = ψ(0)Kb a−1
. By Equation (8), q is constant.
Equations (6) and (7) state that Cψ,ϕ is unitarily equivalent to a constant
multiple of Caz. Since by Proposition 2.2, Caz is normaloid, Cψ,ϕ is also
normaloid.
(cid:3)
a−1 z = ψ(0)Kb a−1
(z).
a−1
a−1
We know that for each c ∈ C, CKc,ϕ is compact, where ϕ(z) = az + b
and a < 1 (see [15, Corollary 2.4 ]). Hence σe(CKc,ϕ) = {0}. Thus, CKc,ϕ
satisfies the conditions of Theorem 2.4 and so if CKc,ϕ is normaloid, then c
must be b a−1
a−1 (see also Proposition 2.3).
Corollary 2.5. Suppose that ψ is an entire function and is not identi-
cally zero. Assume that ϕ(z) = az + b, where a < 1 and b ∈ C. Then Cψ,ϕ
is compact and normaloid if and only if ψ = ψ(0)Kb a−1
.
a−1
Suppose that ψ is an entire function and ϕ(z) = az + b, where a ≤ 1.
In the next theorem, we show that Cψ,ϕ is cohyponormal if and only if Cψ,ϕ
is normal (see [10, Theorem 3.3] and note that if a = 1 and a 6= 1, then
a−1
a−1 = −a).
Theorem 2.6. Suppose that ψ is an entire function and ϕ(z) = az + b,
a−1
a−1
where a ≤ 1. Then Cψ,ϕ is cohyponormal if and only if ψ = ψ(0)Kb
for a 6= 1 and ψ = ψ(0)K−b for a = 1.
Proof. We break the proof into two cases. First assume that a = 1. If
ψ = ψ(0)K−b, then by [13, Corollary 1.2], Cψ,ϕ is a constant multiple of a
unitary operator. Thus, Cψ,ϕ is cohyponormal. Now let Cψ,ϕ be cohyponor-
mal. As we saw in the second paragraph of this section ψ = ψ(0)K−b.
Now assume that a 6= 1. Suppose that Cψ,ϕ is cohyponormal. Let Cq, eϕ
be as in Equation (6), where q and eϕ were obtained in Equations (7) and (8).
It is obvious that Cq, eϕ is also cohyponormal. Then kC∗q, eϕK0k ≥ kCq, eϕK0k.
We obtain q(0) ≥ kqk. As we saw in the proof of Theorem 2.4, q must be
constant, so Equation (8) stats that ψ(z) = ψ(0)eb a−1
a−1 z.
7
The other direction follows easily from Proposition 2.3.
(cid:3)
3 Closed range weighted composition operator
In this section, we prove that Cψ,ϕ has closed range if and only if Cψ,ϕ is a
constant multiple of a unitary operator (see [13, Corollary 1.2 ]).
Theorem 3.1. Let ϕ and ψ be entire functions on C such that ψ is
not identically zero. Suppose that Cψ,ϕ is bounded on F 2. Then Cψ,ϕ has
closed range if and only if and ϕ(z) = az + b, with a = 1, b ∈ C and
ψ = ψ(0)K−ab.
Proof. First suppose that a = 1 and ψ = (0)K−ab. By [13, Corollary
1.2 ], we have Cψ,ϕ is a constant multiple of a unitary operator. Therefore,
Cψ,ϕ has closed range.
Conversely, let Cψ,ϕ have closed range on F 2. As we stated in the second
paragraph of Section 2, ϕ(z) = az+b, with a ≤ 1. Suppose that a < 1. By
Equations (6) and (7), Cψ,ϕ is unitarily equivalent to Cq, eϕ, where eϕ(z) = az,
so without loss of generality, we assume that ϕ(z) = az (note that Cψ,ϕ has
closed range if and only if Cq, eϕ has closed range). Since Cψ,ϕ is bounded
on F 2, Cψ,az(ez) = ψ(z)eaz belongs to F 2. Now we define a bounded linear
functional Fψ(z)eaz by Fψ(z)eaz (f ) = hf (z), ψ(z)eazi for each f ∈ F 2. We
know that Kw
kKwk
converges to zero weakly as w → ∞. Then
It shows that
lim
w→∞h
Kw
kKwk
, ψ(z)eazi = 0.
ψ(w)eaw
ew2/2
lim
w→∞
= 0.
(9)
Now if a = aeiθ, we take w = re−iθ, where r is a positive real number.
Then eaw = eaw. Equation (9) shows that
r→∞ψ(re−iθ)2ear2−r2
= 0.
(10)
lim
From Equation (10), we obtain
lim
r→∞kC∗ψ,ϕ
Kre−iθ
kKre−iθkk2 = lim
r→∞ψ(re−iθ)2kereiθazk2
er2
8
= lim
r→∞ψ(re−iθ)2er2(a2−1)
= 0.
(11)
Since ψ is not identically zero, it is easy to see that there exits a sequence
{rn} such that for any n, rn is a positive real number, rn → ∞ as n → ∞ and
Karne−iθ
ψ(rne−iθ) 6= 0 for each n. We have C∗ψ,ϕ
kKrne−iθk 6= 0
Krne−iθ
kKrne−iθk 6∈ Ker(C∗ψ,ϕ). Equation (11) and [5, Proposition 6.1, p.
and so
363 ] show that C∗ψ,ϕ does not have closed range. Thus, Cψ,ϕ does not have
closed range (see [5, Proposition 6.2, p. 364 ]). Hence a = 1 and the result
follows from the second paragraph of Section 2.
Krne−iθ
kKrne−iθk
= ψ(rne−iθ)
(cid:3)
4 Norm of weighted composition operator
Suppose that ϕ and ψ are entire functions on C and Cψ,ϕ is bounded on F 2.
It is well-known that for any w ∈ C,
C∗ψ,ϕKw = ψ(w)Kϕ(w).
We use this formula in the following remark. Furthermore, in this section
for each c ∈ C, MKc is multiplication by the kernel function Kc.
Remark 4.1. For ϕ(z) = az + b, where a ≤ 1 and b, c ∈ C, kCϕk
was found in [4, Theorem 4]. In [4], the inner product of the Fock space is
different from ours. An analogue of [4, Theorem 4] holds for F 2 with our
definition. One can follow the outline of the proof of [4, Theorem 4] to find
kCϕk on F 2. Moreover, we state another proof for finding kCϕk. We break
it into two cases.
First assume that ϕ(z) = az+b, where a < 1. By [4, Theorem 2], Cϕ is com-
pact. Then [4, Lemma 2] implies that C∗ϕCϕ = MKbCazCaz+b = MKbCa2z+b
is compact ( note that by the similar proof which was stated in [4, Lemma
2], we can see that C∗az+b = CKb,az). We know that kCϕk2 = kC∗ϕCϕk =
r(MKbCa2z+b). Since MKbCa2z+b is compact, r(MKbCa2z+b) = sup{λ :
λ ∈ σp(MKbCa2z+b)} (see [5, Theorem 7.1, p. 214 ]). By Proposition
b2
1−a2 . We have
2.1, for each λ ∈ σp(MKbCa2z+b), λ ≤ Kb(
(MKbCa2z+b)∗K b
. Then e
1−a2 ) = e
1−a2 ∈ σp((MKbCa2z+b)∗)
b2
1−a2 K b
= e
1−a2
1−a2
b
b2
9
b2
1−a2 ∈ σp(MKbCa2z+b). Thus,
and so by [5, Theorem 7.1, p. 214 ], e
1
2
b2
1−a2 .
kCaz+bk = e
Now assume that ϕ(z) = az + b with a = 1. By [4, Theorem 1], b = 0.
From [4, Lemma 2], one can easily see that kCϕk2 = kC∗ϕCϕk = kCazCazk =
kCa2zk = kCzk = 1. Then in this case kCϕk = 1.
In the preceding sections, we saw that among weighted composition op-
erators, CKc,ϕ is much important, when c ∈ C. Hence in the following
theorem, we try to find kCKc,ϕk.
Theorem 4.2. Let ψ = Kc and ϕ(z) = az + b, where a ≤ 1 and
b, c ∈ C. Suppose that Cψ,ϕ is bounded on F 2.
(a) If a < 1, then kCψ,ϕk = ec b
(b) If a = 1 and a 6= 1, then kCψ,ϕk = e
(c) If a = 1, then kCψ,ϕk = e
1−ae
a−1
1−a2
b2
2 .
1−a.
c(1−a)
b2
1
2
+b2
.
1
2
c(1−a)
+b2
a−1
1−a2
Proof.
1−a )e
(see also Remark 4.1).
1−a )Caz+u(1−a)+b, where u = c
(a) Assume that a < 1. By the proof of Proposition 2.3,
a−1 . Since
1−a )kCaz+u(1−a)+bk =
Cψ,ϕ is unitarily equivalent to ψ( b
a < 1, [4, Theorem 4] implies that kCψ,ϕk = ψ( b
ψ( b
(b) Assume that a = 1 and a 6= 1. Again by the proof of Proposition 2.3,
kCψ,ϕk = ψ( b
1−a )kCaz+ca+bk. Since a = 1 and
CKc,ϕ is bounded, from the second paragraph of Section 2, c = −ab. Then
ca + b = 0. Therefore, kCψ,ϕk = ψ( b
1−a ) (see [4, Theorem
4] and Remark 4.1). Thus, kCψ,ϕk = e
(c) Assume that a = 1. As we saw in the second paragraph of Section 2,
c = −ab = −b. Then ψ(z) = e−bz. Now we must find kCe−bz,z+bk. By [13,
Corollary 1.2], e
1−a )kCazk = ψ( b
1−a.
1−a = e
a−1 +bk = ψ( b
2 Ce−bz,z+b is unitary. Hence kCψ,ϕk = e
1−a )kC
az+ c(1−a)
b2
2 .
−ab2
−b2
b2
(cid:3)
References
[1] J. R. Akeroyd and S. R. Fulmer, Closed-range composition operators on weighted
Bergman spaces, Integr. Equ. Oper. Theory 72 (2012), 103-114.
10
[2] P. S. Bourdon, Spectra of some composition operators and associated weighted com-
position operators, J. Oper. Theory 67 (2) (2012), 537-560.
[3] P. S. Bourdon and S. K. Narayan, Normal weighted composition operators on the
Hardy space H 2(D), J. Math. Anal. Appl. 367 (2010), 278-286.
[4] B. J. Carswell, B. D. MacCluer and A. Schuster, Composition operators on the Fock
space, Acta Sci. Math. (Szeged) 69 (2003), 871-887.
[5] J. B. Conway, A Course in Functional Analysis, Second Edition, Springer-Verlag,
New York, 1990.
[6] C. C. Cowen, S. Jung, and E. Ko, Normal and cohyponormal weighted composition
operators on H 2, Operator Theory: Advances and Applications 240 (2014), 69-85.
[7] C. C. Cowen and B. D. MacCluer, Composition Operators on Spaces of Analytic
Functions. Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1995.
[8] M. Fatehi and M. Haji Shaabani, Norms of hyponormal weighted composition op-
erators on the Hardy and weighted Bergman spaces, Operators and Matrices 12(4)
(2018), 997-1007.
[9] P. Ghatage and M. Tjani, Closed range composition operators on Hilbert function
spaces, J. Math. Anal. Appl. 431 (2) (2015), 841-866.
[10] T. Le, Normal and isometric weighted composition operators on the Fock space, Bull.
London Math. Soc. 46 (2014), 847-856.
[11] J. H. Shapiro, Composition Operators and Classical Function Theory, Springer-
Verlag, New York, 1993.
[12] S. Ueki, Weighted composition operator on the Fock space, Proc. Amer. Math. Soc.
135 (2007), 1405-1410.
[13] L. Zhao, Unitary weighted composition operators on the Fock space of Cn, Complex
Anal. Oper. Theory 8 (2014), 581-590.
[14] L. Zhao, Invertible weighted composition operators on the Fock space of Cn, J. Funct
Spaces 2015. Art. ID 250358.
[15] L. Zhao and C. Pang, A class of weighted composition operators on the Fock space,
Journal of Mathematical Research with Applications, 35 (3) (2015), 303-310.
[16] K. Zhu, Analysis on Fock Spaces, Graduate Texts in Mathematics 263, Springer,
New York, 2012.
[17] N. Zorboska, Compositon operators with closed range, Trans. Amer. Math. Soc. 344
(1994), 791-801.
M. Fatehi, Department of Mathematics, Shiraz Branch, Islamic Azad University, Shi-
raz, Iran.
E-mail: [email protected]
11
|
1111.3990 | 2 | 1111 | 2012-09-25T17:51:32 | Uncertainty Principles in Finitely generated Shift-Invariant Spaces with additional invariance | [
"math.FA"
] | We consider finitely generated shift-invariant spaces (SIS) with additional invariance in $L^2(\R^d)$. We prove that if the generators and their translates form a frame, then they must satisfy some stringent restrictions on their behavior at infinity. Part of this work (non-trivially) generalizes recent results obtained in the special case of a principal shift-invariant spaces in $L^2(\R)$ whose generator and its translates form a Riesz basis. | math.FA | math |
UNCERTAINTY PRINCIPLES IN FINITELY GENERATED
SHIFT-INVARIANT SPACES WITH ADDITIONAL INVARIANCE
ROMAIN TESSERA AND HAICHAO WANG
Abstract. We consider finitely generated shift-invariant spaces (SIS) with
additional invariance in L2(Rd). We prove that if the generators and their
translates form a frame, then they must satisfy some stringent restrictions on
their behavior at infinity. Part of this work (non-trivially) generalizes recent
results obtained in the special case of a principal shift-invariant spaces in L2(R)
whose generator and its translates form a Riesz basis.
1. Introduction
Finitely generated shift-invariant spaces have been widely used in approximation
theory, numerical analysis, sampling theory and wavelet theory ( see e.g., [1, 3, 4,
9, 11, 12, 15, 16] and the references therein). Shift-invariant spaces with additional
invariance have been studied in the context of wavelet analysis and sampling theory
[6, 10, 13, 17], and have been given a complete algebraic description in [2] for L2(R)
and in [7] for L2(Rd). As a tool for showing our main results, we will prove a slightly
different but useful characterization.
It is well-known that the Paley-Wiener space P W is translation-invariant. More-
over, Shannon's sampling theorem easily implies that P W is principal, i.e. gener-
ated by the single function sinc. It turns out that the fact that sinc is non-integrable
is not a coincidence. Actually, for a principal shift-invariant space in L2(R) which
is translation-invariant, any frame generator is non-integrable (see for instance [5]).
This observation holds in any dimension.
frame generator has to satisfy for a.e. ω ∈ Rd
Indeed, the Fourier transform bφ of a
C−11A(ω) ≤ bφ(ω) ≤ C1A(ω)
1
for some C ≥ 1 and some finite measure subset A. In particular, bφ is not continuous.
Such a condition also prevents bφ from being in the Sobolev space1 H
2 (Rd) (see
2 −ǫ(Rd) for every ǫ > 0 when A is a Euclidean ball
[14]), whereas it belongs to H
of positive radius.
1
Our first result is a straightforward generalisation of this fact to shift-invariant
spaces generated by several functions.
Theorem 1.1. Let Λ be a lattice in Rd. If a finitely generated Λ-invariant space
of L2(Rd) is translation-invariant, then at least one of its frame generators has a
Key words and phrases. Finitely generated shift-invariant spaces; Frame; Additional invari-
ance; Uncertainty principle.
1i.e. R φ(x)2(1 + x)dx = ∞.
1
2
ROMAIN TESSERA AND HAICHAO WANG
non continuous Fourier transform, and in particular is not in L1. Moreover this
generator satisfies R φ(x)2(1 + x)dx = ∞.
The slow spatial-decay of the generators of shift-invariant spaces that are also
translation-invariant is a disadvantage for the numerical implementation of some
analysis and processing algorithms. This is a motivation for considering instead
shift-invariant spaces that are only 1
n Z-invariant and hoping the generators will
have better time-frequency localization. Indeed, it was shown in [5] that for every
n, one can construct a principal shift-invariant space with an orthonormal genera-
tor which is in L1(R) although the space is 1
n Z-invariant. However, there are still
obstructions if we require more regularity on the Fourier transform of the gener-
ators. Asking for fractional differentiability yields the following Balian-Low type
obstructions (see [8] and the reference therein). Compare [5, Theorem 1.2].
Theorem 1.2. Let Λ < Γ be two lattices2 in Rd. Suppose φi ∈ L2(Rd) are such that
{φi(· + λ)λ ∈ Λ, i = 1, . . . , r} forms a frame for the closed subspace V Λ(Φ) spanned
by these functions. Let ρ (≤ r) be the minimal number of generators of V Λ(Φ).
Assume that [Γ : Λ] is not a divisor of ρ, and suppose that V Λ(Φ) is Γ-invariant.
Then there exists i0 ∈ {1, . . . , r} such that
ZRd φi0 (x)2xd+ǫdx = +∞
for all ǫ > 0. I.e. bφi0 is not in H
d
2 +ǫ(Rd).
One can also ask for a combinaison of regularity, namely continuity and of control
of the decay at infinity. In this spirit, we obtain the following result (compare [5,
Theorem 1.3]).
Theorem 1.3. Keep the same assumptions as in the last theorem, and suppose
moreover that bφi is continuous for every i = 1, . . . , r. Then there exits i0 ∈
{1, . . . , r} such that ω
d
2 +ǫbφi0 (ω) 6∈ L∞(Rd) for any ǫ > 0, i.e.,
ess sup ω∈Rbφi0 (ω)ω
2 +ǫ = +∞.
d
For d = 1, the exponent is sharp up to the ǫ in both theorems but it does not seem
to be the case for larger d. We actually conjecture that the right exponent should
be the same for all dimensions.
Notice that the condition [Γ : Λ] ∤ ρ in the above theorems is essential. Indeed all
regularity constraints trivially disappear when ρ = k[Γ : Λ] for any integer k > 0.
To see why, start with some Γ-invariant space generated by exactly k orthogonal
generators φ1, . . . , φk, with -- say -- smooth and compactly generated Fourier trans-
forms. Then note that as a Λ-invariant space, V Γ(φ1, . . . , φk) is generated by the
orthogonal generators φi(· − f ) for f ∈ F and i = 1, . . . , k, where F is a section of
Γ/Λ in Γ.
The previous results state that under additional invariance there always exists at
least one (frame) generator whose Fourier transform has poor regularity. One may
wonder if at least some generators can be chosen with good properties. We do not
know what the optimal proportion of good generators should be. The following
2The notation Λ < Γ is standard to denote subgroups.
UNCERTAINTY PRINCIPLES
3
proposition gives a lower bound on the number of good generators. Observe that
this bound gets worse with the dimension.
Proposition 1.4. For every d ≥ 1, and every k ∈ N there exists an SIS V (Φ) in
L2(Rd) generated by an orthonormal basis Φ consisting of r = (2k)d functions, kd
of which have smooth and compactly supported Fourier transforms, and such that
V (Φ) is translation-invariant. Moreover all the generators can be chosen so that
their Fourier transforms are in H
2 −ǫ(Rd) for all ǫ > 0.
1
We also have the following result,
Proposition 1.5. For d ≥ 1, let Γ(> Zd) be a lattice of Rd and r ≥ 1. Then
there exists an SIS V (Φ) in L2(Rd) generated by r orthonormal generators φi's,
all of which are in L1 (hence they have continuous Fourier transforms) and satisfy
ω
2bφi(ω) ∈ L∞(Rd). Moreover, V (Φ) is Γ-invariant.
1
To summarize, while Proposition 1.4 states that it is possible to construct translation-
invariant SIS with a portion of the generators having smooth and compactly sup-
ported Fourier transforms, Proposition 1.5 shows that we can construct Γ-invariant
SIS with all its generators having certain pointwise decay in Fourier domain.
Organization. In the following section, we state and prove a convenient charac-
terization (similar to the one given in [2]) of SIS with additional invariance.
In Section 3, we refine this characterization under the assumption that the genera-
tors and their translates form a frame.
In Section 4, we prove a useful property of the Gramian under the sole assumption
that the generators have continuous Fourier transform.
Sections 5 and 6 are dedicated to the proof of the results stated in the introduction.
2. Finitely generated shift-invariant spaces with additional
invariance
Given a closed subgroup Λ of Rd, we define its Fourier transform Λ∗ as the closed
subgroup of Rd defined by
Λ∗ = {x ∈ Rd, e2πihλ,xi = 1 ∀λ ∈ Λ}.
Note that the map Λ → Λ∗ is an involution. The Fourier transform reverses the
inclusions, namely if Λ < Γ then Γ∗ < Λ∗. Observe for instance that for d = 1, the
Fourier transform of nZ is 1
n Z, or that for d = 2, the Fourier transform of R × {0}
is {0} × R.
In this paper, we consider the following general setting: Λ < Γ are two closed
subgroups of Rn. Let φ1, . . . , φr ∈ L2(Rd), and denote by Φ the column vector
whose components are the φi's. We will denote by V Λ(Φ) the smallest closed Λ-
invariant subspace containing the φi's. It is common to call V Λ(Φ) shift-invariant
when Λ = Zd, and translation-invariant when Λ = Rd. To allege notation, we will
omit the subscript Zd, when Λ = Zd.
We will be interested in the situation where V Λ(Φ) is in addition Γ-invariant. As we
will see later, this prevents frame generators from having nice decay at infinity. We
will start by providing a short and self-contained proof of a result essentially due
4
ROMAIN TESSERA AND HAICHAO WANG
to [2, 7]. Observe that this problem is only non-trivial when the quotient Rd/Λ is
compact since otherwise, the only finitely generated Λ-invariant space is {0}. This
is equivalent to the fact that Λ∗ (hence Γ∗) is discrete.
In this paper we will mainly focus on the cases when Λ is a lattice and when Γ is
either a (larger) lattice or all of Rd.
Before stating this result, let us introduce some notation. The Gramian associated
to Φ and Λ is a measurable field of r×r matrices whose general coefficient is defined
for a.e. ω ∈ R by
For short,
GΛ
i,j(ω) = Xl∈Λ∗ bφi(ω + l)cφj(ω + l).
GΛ(ω) = Xl∈Λ∗ bΦ(ω + l)bΦ∗(ω + l).
A(ω) = Xg∈Γ∗ bΦ(ω + g)bΦ∗(ω + g).
For a.e. ω ∈ R, let A(ω) to be the r × r matrix defined by
Now, let F be a subset of Λ∗ consisting of representatives of the quotient Λ∗/Γ∗.
For instance, for d = 1, Λ = Z and Γ = 1
n Z, one has Λ∗ = Z and Γ∗ = nZ, and for
F , one can take {0, . . . n − 1}.
We have
A(ω + f ).
(2.1)
GΛ(ω) = Xf ∈F
Now let us state the main result of this section. Although it could easily be deduced
from the main results in [2, 7], we chose to write down a short self-contained proof.
Theorem 2.1. Let Λ < Γ be closed cocompact subgroups of Rd. The space V Λ(Φ)
is Γ-invariant if and only if the following equality holds for a.e. ω ∈ Rd.
rank A(ω + f ).
(2.2)
rank GΛ(ω) = Xf ∈F
Proof. Recall that a function ϕ belongs to V Λ(Φ) if and only if there exist bounded
measurable Λ∗-periodic functions P1, . . . , Pr such that bϕ = P1cφ1 + . . . + Prcφr.
Fiberwise, this is equivalent to saying that for a.e. ω, the vector (bϕ(ω + l))l∈Λ∗ ∈
ℓ2(Λ∗) lies in the subspace bV (ω) spanned by the r vectors (bφi(ω + l))l∈Λ∗ for i =
1, . . . , r.
Now the space V Λ(Φ) is Γ-invariant if for every bounded Γ∗-periodic function Q,
and for every i = 1, . . . , r, there are bounded measurable Λ∗-periodic functions
Pi,1, . . . , Pi,r such that
Qbφi = Pi,1cφ1 + . . . + Pi,rcφr.
Again, this is equivalent to saying that for a.e. ω, and for every bounded Γ∗-periodic
θ : Λ∗ → C the vector (θ(ω + l)bφi(ω + l))l∈Λ∗ ∈ ℓ2(Λ∗) lies in the subspace bV (ω).
Let us denote cW (ω) the subspace of ℓ2(Λ∗) spanned by (θ(ω + l)bφi(ω + l))l∈Λ∗ for
all bounded Γ∗-periodic θ, and every i = 1, . . . , r. Clearly bV (ω) is a subspace of
cW (ω), and the two coincide for a.e. ω exactly when V Λ(Φ) is Γ-invariant.
UNCERTAINTY PRINCIPLES
5
The rest of the proof amounts to showing that the left-hand term in (2.2) corre-
sponds to the dimension of bV (ω) (which is obvious) and that the right-hand term
of the equality corresponds to the dimension of cW (ω). A basis for the space of
bounded Γ∗-periodic functions on Λ∗ consists of the functions (θf )f ∈F , where each
θf (l) equals 1 if l ∈ f + Γ∗ and 0 elsewhere. Observe that for all f 6= f ′, and all
ϕ, ϕ′ ∈ W bΦ(ω), θf ϕ and θf ′ϕ′ are orthogonal. Hence cW (ω) decomposes as a direct
sum
cW (ω) = Mf ∈F cW f (ω),
where cW f (ω) is the subspace of functions of the form θf ϕ, for ϕ ∈ cW (ω). But it
comes out that cW f (ω) is precisely the subspace of ℓ2(Λ∗) spanned by (bφi(ω + f +
g))g∈Γ∗ , whose dimension equals the rank of A(ω + f ). This finishes the proof of
the theorem.
(cid:3)
Although we will be mostly interested in the case where both Λ and Γ are lattices,
we will also use the following special case of Theorem 2.1, where Λ is a lattice and
Γ = Rd (so Γ∗ = {0}).
Corollary 2.2. Let Λ be a lattice, and let V Λ(Φ) be a Λ-invariant space generated
by φ1, . . . , φr. Then it is translation-invariant if and only if the following equality
holds for a.e. ω ∈ Rd.
(2.3)
(2.4)
rank GΛ(ω) = Xf ∈Λ∗
rank A(ω + f )
= {f ∈ Λ∗, Φ(ω + f ) 6= 0} .
3. Frame generators and additional invariance
The main goal of this section is to reformulate Theorem 2.1 under the additional
assumption that the generators form a frame for the space V Λ(Φ).
Recall that the family of all Λ-translates of the generators φ1, . . . , φr form a Riesz
basis if and only if there exists s ≥ 1 such that
s−1I ≤ GΛ(ω) ≤ sI, a.e. ω ∈ R.
(3.1)
In this case the functions φi's are called Riesz generators for V Λ(Φ).
Similarly, {φi(· + λ) λ ∈ Λ, i = 1, . . . , r} is a frame if and only there exists s ≥ 1
such that
s−1GΛ(ω) ≤ (GΛ(ω))2 ≤ sGΛ(ω)
(3.2)
for almost every ω ∈ Rd (see [9]). Here the φi's are called frame generators for
V Λ(Φ).
Let us start by an easy lemma. Given a non-negative self-adjoint matrix A, denote
qA its associated quadratic form, µ−(A) its smallest non-zero eigenvalue and kA
the dimension of its kernel. Denote the unit sphere of a Euclidean space V by SV .
Lemma 3.1. Let C = A + B be three d× d non-negative self-adjoint matrices such
that rank C = rank A + rank B. Then µ−(C) ≤ min{µ−(A), µ−(B)}.
6
ROMAIN TESSERA AND HAICHAO WANG
Proof. Note that we can assume without loss of generality that C has full rank.
Observe that KerA and KerB are in direct sum. By the min-max theorem,
µ−(A) =
min
dim V =kA+1
max
x∈SV
qA(x).
Now let V0 be a subspace minimizing the above expression. Let x ∈ V0 ∩ KerB of
norm 1. We have
µ−(A) ≥ qA(x) = qA(x) + qB(x) = qC (x) ≥ µ−(C).
We conclude since the roles of A and B are symmetric.
(cid:3)
The following theorem will play a central role in the sequel.
Theorem 3.2. Let Λ < Γ be closed cocompact subgroups of Rd. Let Φ be finite set
of generators of V Λ(Φ). Suppose the space V Λ(Φ) is Γ-invariant and Φ are frame
generators. Then, there exists s ≥ 1 such that the formula (2.2) and the following
inequality hold for a.e. ω ∈ Rd
s−1A(ω) ≤ A(ω)2 ≤ sA(ω).
(3.3)
Proof. By Theorem 2.1, it is enough to prove that (3.2) implies (3.3) (up to chang-
ing the constant s). First observe that the previous lemma can be extended (by
induction) to a sum of more than two matrices. We deduce that the lower bound
in (3.2) implies that of (3.3). The equivalence between the upper bounds follows
from the fact that the number of non-zero matrices in the right-hand side of (2.1)
is bounded by r.
(cid:3)
We immediately get the following corollary.
Corollary 3.3. Let Λ be a closed cocompact subgroups of Rd. Assume that the
space V Λ(Φ) is translation-invariant and Φ are frame generators. Then T r(A(ω))
is not continuous, nor in H
2 .
1
Proof. By Theorem 3.2, the map ω → T r(A(ω)) is larger than s−1 on a set of
positive (but finite) measure, and equals zero elsewhere.
In particular it is not
continuous and not in H
(cid:3)
2 (by [14]).
1
Lemma 3.4. Let (M (x))
x∈Rd be a continuous family of r × r non-negative self-
adjoints matrices with complex coefficients. Assume that there exists s ≥ 1 such
that s−1M (x) ≤ M 2(x) for all x ∈ Rd. Then the rank of M is constant.
Proof. Since x → M (x) is continuous, its rank is lower semi-continuous. Let F
be the (closed) subset where rank M reaches its minimal value. Assume that F is
not open, which means that there exists a sequence xn ∈ F c converging to some
x0 ∈ F . It follows that the rank of M (xn) is strictly larger than the rank of M (x0),
which therefore implies that the range of M (xn) intersects non-trivially the kernel
of M (x0). Let un be a sequence of unit vectors lying in this intersection. On one
hand, because un is in the range of M (xn), we must have kM (xn)unk ≥ s−1. On
the other hand, the continuity of M together with the fact that the limit lies in the
kernel implies M (xn)un → 0, contradiction. So F is open which implies F = Rd
hence the lemma.
(cid:3)
UNCERTAINTY PRINCIPLES
7
Corollary 3.5. Let Λ < Γ be closed cocompact subgroups of Rd. Assume that the
space V Λ(Φ) is Γ-invariant and Φ are frame generators. Then if A(ω) is continuous,
its rank is constant.
Proof. The corollary follows immediately from Lemma 3.4 and Theorem 3.2.
(cid:3)
4. Properties of the Gramian when the bφi's are continuous
Although this section mainly serves as preparation for Theorem 1.2, we believe that
it is of independent interest and could be useful elsewhere.
From now on, Λ will always denote a lattice in Rd. The following statement is
essentially trivial (and was observed for instance in [9, 12]).
Proposition 4.1. Let Λ be a lattice in Rd, and let Φ = (φ1, . . . , φr) be a generating
set for the Λ-invariant space V Λ(Φ). Then the minimal number ρ of generators of
V Λ(Φ) equals the essential supremum of rank GΛ(ω). Equivalently, if GΛ(ω) is non
invertible a.e., then one can find a generating set Φ′ with r − 1 generators.
Proof. This is an essentially trivial statement. Let K be a fundamental domain for
the action of Λ∗ (for example take K = [0, 1)d if Λ∗ = Zd). For every ω ∈ K, let
V Λ(ω) be the subspace of ℓ2(Λ∗) spanned by the r vectors vi(ω) = bφi(ω + l)l∈Λ∗
for i = 1, . . . , r. Saying that GΛ is non-invertible a.e. amounts to the fact that
dim V Λ(ω) ≤ r − 1 for a.e. ω ∈ K. The idea is to remove for a.e. ω one vector
vi(ω) and to recombine the other ones in order to get a new set of r − 1 generators
for V Λ(Φ). Now, in order to do this in a measurable way, we pick the vi(ω) of
minimal index with the property that it lies in the vector space spanned by the
other ones. Then we relabel the remaining ones respecting their order: for instance
if v2 is removed, then v1 becomes v′
2 and so on. Now since every
point of Rd can be written as ω + l, for some unique (ω, l) ∈ K × Λ∗, we can define
our new r − 1 generators by the formula φ′
We deduce from this proposition that the essential supremum of rank G(ω) equals
ρ.
i(ω))l, for all l ∈ Λ∗.
1, v3 becomes v′
i(ω + l) = (v′
(cid:3)
Lemma 4.2. Suppose that the functions bφi are continuous for i = 1, . . . , r. Then
the maps ω → rank A(ω) and ω → rank GΛ(ω) from Rd to N are lower semi-
continuous.
Proof. Since GΛ(ω) = PF A(ω + f ) it is enough to show that rank A(ω) is lower
semi-continuous. Observe that A(ω) is the sum over g ∈ Γ∗ of the continuous
positive semi-definite matrices bΦ(ω + g)bΦ(ω + g)∗. Therefore the rank of A(ω) is
the supremum of the ranks of all partial finite sums. Since lower semi-continuity is
stable under taking supremums, we deduce that rank A(ω) is lower semi-continuous.
(cid:3)
Proposition 4.3. Suppose that the functions bφi are continuous, and that V Λ(Φ) is
Γ-invariant for some lattice Γ such that [Γ : Λ] does not divide ρ. Then the subset
{ω, rank GΛ(ω) < ρ} is a non-empty closed subset of Rd.
8
ROMAIN TESSERA AND HAICHAO WANG
Proof. The fact that {ω, rank GΛ(ω) ≤ ρ − 1} is closed results from Lemma 4.2.
We therefore only have to prove that it is non-empty. Let us assume on the contrary
that GΛ(ω) has rank ρ for all ω. We know by Theorem 2.1 that for a.e. ω ∈ Rd,
(4.1)
rank A(ω + f ).
ρ = Xf ∈F
Moreover the equality is an inequality ≤ for all ω. But again Lemma 4.2 implies
that the ω's for which the inequality is strict form an open set, which therefore has
to be empty.
Now observe that lower semi-continuous functions with integer values can only
increase locally. In other words for every ω0 the set
{ω, rank A(ω) ≥ rank A(ω0)} = {ω, rank A(ω) > rank A(ω0) − 1/2}
is a non-empty open set. Together with (4.1), this immediatly implies that rank A(ω)
is locally constant, hence constant on Rd (since Rd is connected!). Let us call m
the corresponding integer. We therefore have ρ = m[Γ : Λ], contradiction.
(cid:3)
5. Proofs of The main results
5.1. Proof of Theorem 1.1. Immediately follows from Corollary 3.3.
d
5.2. Proof of Theorem 1.3. Suppose by contradiction that there exists ǫ > 0 such
that ω
we conclude by Corollary 3.5: namely since rank A(ω) is constant it must divide
the rank of G(ω), which by (2.2) would imply that the index of Λ in Γ divides r.
2 +ǫbφi(ω) ∈ L∞(Rd) for all i. This easily implies that A(ω) is continuous, so
5.3. Proof of Theorem 1.2. The proof of Theorem 1.2 relies on the following
result of harmonic analysis.
Its proof for d = 1 is essentially the proof of [5,
Theorem1.2]. Since it extends without change to any d, we do not reproduce it
here.
Lemma 5.1. Let Λ be a lattice in Rd. Suppose that a function f ∈ L2(Rd) satisfies
the following properties
(1) there exists ǫ > 0 such that
ZRd f (x)2xd+ǫdx < ∞.
Xk∈Λ∗ bf (ω + k)2 ≤ C.
(2) There exists a constant C such that for a.e. ω ∈ Rd,
(5.1)
(3) There exists ω0 is such that
Then for all η > 0, there exists δ0 > 0 only depending on η, ǫ and C such that for
all 0 < δ ≤ δ0
for all k ∈ Λ∗,
bf (ω0 + k) = 0
δd ZB(ω0,δ) Xk∈Λ∗ bf (ω + k)2dω < η.
1
UNCERTAINTY PRINCIPLES
9
Proof of Theorem 1.2. We shall use the notation introduced in Section 3 (before
stating Lemma 3.1). Let us suppose by contraction that there exits ǫ > 0 such that
ZRd φi(x)2xd+ǫdx < ∞,
(5.2)
for every i. Note that this condition implies that the φi are in L1, so that their
Fourier transforms are uniformly continuous.
Define E = {ω, rank GΛ(ω) < ρ}, which by Proposition 4.3, is a non-empty closed
subset of Rd. Take an open ball B ⊂ Ec whose boundary interests E in a point ω0.
Observe that at least one fourth of the volume of a sufficiently small ball centered
in ω0 is contained in Ec. We will use this remark at the end of the proof.
Since ω0 ∈ E, there exists an (r − ρ + 1)-dimensional subspace V0 of Rr on which
GΛ(ω0) vanishes. Since GΛ(ω) = F (ω)F ∗(ω) where F (ω) is the r × Λ∗ matrix
whose (i, l)-coefficient is F (ω)il = bφi(ω + l), we have F ∗(ω0)a = 0 for all a ∈ V0.
This implies that
rXi=1
1
for all l ∈ Λ∗.
aibφi(ω0 + l) = 0
In other words, the function fa(x) = Pr
i=1 aiφi(x) satisfies bfa(ω0 + l) = 0 for all
l ∈ Λ∗ and for all a ∈ V0. Now clearly the functions fa for a ∈ SV0 satisfy the
assumptions of the lemma with C independent of a. We therefore get for δ > 0
small,
δd ZB(ω0,δ)Xk∈ bΛ
bfa(ω + k)2dω < η.
(5.3)
Suppose a1, . . . , ar−ρ+1 is an orthonormal basis of V0. Then for a.e. ω ∈ B, the
frame condition implies
s−1 ≤ µ−(GΛ(ω)) ≤
r−ρ+1Xi=1
hGΛ(ω)ai, aii.
By the above remark, we see that this contradicts (5.3) provided η is small enough.
(cid:3)
6. Proofs of Proposition 1.4 and Proposition 1.5
6.1. Proof of Proposition 1.4. It is actually enough to provide a construction for
d = 1. Indeed given such φ1, . . . , φ2k in L2(R), we define generators in dimension d
by considering tensor products of these functions: namely φi1,...,id = φi1 ⊗ . . .⊗ φid .
Then we get that V Φ is simply the tensor product V ⊗d
Φ , whose generators are clearly
orthonormal. Translation invariance of V Φ therefore follows from that of VΦ.
Hence, let us focus on the one-dimensional case. Let g be an infinitely-differentiable
function that satisfies g(x) = 0 when x ≤ 0, g(x) = 1 when x ≥ 1, and g2(x) +
g2(1 − x) = 1 when 0 ≤ x ≤ 1. Define bφ1(ω) = g(ω)g(2 − ω) and bφ2(ω) =
g(1 − ω)χ[0,1) − g(ω − 1)χ[1,2).
Then for i ≤ r, define bφi(ω) = bφ1(ω − i + 1) when i is odd and bφi(ω) = bφ2(ω −
i + 2) when i is even (bφ1 and bφ2 are plotted in Figure 1). It is easy to check that
Pl∈Z bφi(ω + l)2 = 1 a.e. ω for all i, and that Pl∈Z bφ1(ω + l)bφ2(ω + l) = 0. We
10
ROMAIN TESSERA AND HAICHAO WANG
deduce thatPl∈Z bφi(ω + l)bφj(ω + l) = 0 for i 6= j. This shows that G(ω) = I which
amounts to saying that the functions φi are orthonormal generators for V (Φ).
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Figure 1. The function bφ1 being red and the function bφ2 being blue.
Since all the functions are supported on [0, r], we can show that, for a.e. ω ∈ [0, 1),
rank A(ω + f ) = 1 when 0 ≤ f ≤ r − 1 and rank A(ω + f ) = 0 elsewhere. Hence
Pf ∈Z rank A(ω + f ) = r = rank GΛ(ω). From Corollary 2.2, V (Φ) is translation-
invariant.
Observe that when i is odd, bφi is compactly supported and infinitely-differentiable.
When i is even, bφi can be written as a product of a compactly supported and
infinitely-differentiable function and the characteristic function χ[i−2,i). Since χ[i−2,i)
belongs to H
2 −ǫ for any ǫ > 0.
1
1
2 −ǫ, we have bφi ∈ H
Z×···× 1
6.2. Proof of Proposition 1.5. We assume that Γ = 1
nd
n1
first need the following lemma which treats the case when d = 1 and r = 1.
Lemma 6.1 ([5]). For integer n1, there exists a function φ1 ∈ L1 ∩ L2 (and hence
bφ1 is continuous), such that φ1 is an orthonormal generator for its generating space
V (φ1), V (φ1) is 1
n1
Z-invariant and ω
Z. We
Z× 1
n2
1
2bφ1(ω) ∈ L∞(R).
Since our method relies on the construction of φ1 in Lemma 6.1, we will describe
φ1 explicitely here. Let g be the function as in the proof of Proposition 1.4. Define
g0(x) = g(x + 1)g(−x + 1) and g1(x) = g(x + 1)g(−2x + 1). Then the Fourier
transform of φ11 is defined to be
cφ1(ω) = h0(ω) +
∞Xj=1
+
∞Xj=1
4j −1Xl=0
2−jhj(ω − n1(γj + l))
4j −1Xl=0
2−jhj(−ω − n1(γj + l)),
(6.1)
UNCERTAINTY PRINCIPLES
11
1
4 , 1
k=0 4k, h0(ω) = g0(4ω) and hj(ω) = g1(2j+1ω − 2j + 1).
where γj =Pj−1
It is not hard to check that cφ1 is an even function such that bφ1 ≤ 1 and has
support on [− 1
4 ] ∪ (∪∞
2 −
2j+2 + n1(γj + l)] and E′
From one to several generators (in dimension 1). Now we want to construct a
1
Z-invariant SIS with r orthonormal generators in L1∩L2 satisfying our pointwise
n1
decay property in Fourier domain.
Define φi for 2 ≤ i ≤ r to be such that bφi(ω) = 0 when ω ∈ [0, n1γ2i−2 − 1] and
1j ), where E1j = ∪4j −1
l=0 [ 1
1
2j − 1
2 − n1(γj + l),
2j + n1(γj + l), 1
2 − n1(γj + l)].
j=1E1j ∪ E′
1j = ∪4j −1
l=0 [ 1
2 − 1
2j+2 − 1
2−(2i−2)h0(ω − n1(γ2i−2 + l))(cid:17)
bφi(ω) =
∞Xj=1
+
1
√2(cid:16) 42i−2−1Xl=0
4j+2i−2−1Xl=0
2−(j+2i−2)hj(ω − n1(γj+2i−2 + l))
(6.2)
when ω ≥ n1γ2i−2 − 1, and bφi(ω) = bφi(−ω) when ω ≤ 0.
It is easy to see that each bφi obeys the pointwise decay property and that it is an
orthonormal generator for the principal SIS V (φi). In fact,
Xl∈Z
bφi(ω + l)2 =Xl∈Z
bφ(ω + l)2 = 1
for a.e. ω. One can also check that for 2 ≤ i ≤ r, bφi has support
(cid:16) ∪∞
j=1 Eij ∪ E′
l=0
ij(cid:17)[(cid:16) ∪42i−2−1
1
[−
4 − n1(γ2i−2 + l),
[(cid:16) ∪42i−2−1
1
+ n1(γ2i−2 + l),
[−
4
2 − 1
[ 1
2 − n1(γj+2i−2 + l),
2j + n1(γj+2i−2 + l), 1
2j+2 − 1
l=0
1
1
4 − n1(γ2i−2 + l)](cid:17)
+ n1(γ2i−2 + l)](cid:17),
1
4
(6.3)
l=0
where Eij = ∪4j+2i−2−1
l=0
ij = ∪4j+2i−2−1
2j − 1
[ 1
E′
In particular, each V (φi) is 1
n1
Fourier domain. Hence V (Φ1) is also 1
n1
whose components are the φi's. In fact
2j+2 + n1(γj+2i−2 + l)] and
2 − 1
2 − n1(γj+2i−2 + l)].
-invariant and all the φi's have disjoint support in
-invariant, where Φ1 is the column vector
V (Φ1) =Mi≤r
V (φi)
(6.4)
From Lemma 6.1, for each j ≥ 2, we can construct V (ψj) with orthonormal gener-
ator ψj having the desired properties and that V (ψj ) is 1
nj
Higher dimension. Like in the previous section, we let φi be the d-fold tensor
product φi ⊗ ψ2 . . .⊗ ψd, for i = 1, . . . , r. By construction, the shift-invariant space
generated by these r orthonormal function is Γ-invariant. Since all φi's and ψj's
are in L1(R), it follows that the φi's are in L1(Rd). Pointwise decay results from
the fact that all these functions have bounded Fourier transforms.
Z-invariant.
12
ROMAIN TESSERA AND HAICHAO WANG
Acknowledgement The authors would like to thank Akram Aldroubi and Qiyu
Sun for valuable discussions. The authors would also like to thank Carolina Mos-
quera and Victoria Paternostro for their comments on the proof of Theorem 1.2.
References
[1] A. Aldroubi, C. A. Cabrelli, D. Hardin, and U. M. Molter, Optimal shift invariant spaces and
their parseval frame generators, Appl. Comput. Harmon. Anal., 23 (2007), 273 -- 283.
[2] A. Aldroubi, C. A. Cabrelli, C. Heil, K. Kornelson and U. M. Molter, Invariance of a shift
invariant space, J. Fourier Anal. and Appl., 16 (2010), 60-75.
[3] A. Aldroubi and K. Grochenig, Non-uniform sampling in shift-invariant space, SIAM Rev.,
43 (2001), no. 4, 585 -- 620.
[4] A. Aldroubi, Q. Sun and W.-S. Tang, Convolution, average sampling, and Calderon resolution
of the identity of shift-invariant spaces, J. Fourier Anal. Appl., 11 (2005), 215 -- 244.
[5] A. Aldroubi, Q. Sun, and H. Wang, Uncertainty principles and Balian-Low type theorems in
principal shift-invariant spaces, Appl. Comput. Harmon. Anal., 30(2011), 337 -- 347.
[6] A. Aldroubi, I. Krishtal, R. Tessera, and H. Wang, Principal Shift-Invariant Spaces with Extra
Invariance Nearest to Observed Data, Collectanea Mathematica, 63(2012), 393-401.
[7] M. Anastasio, C. Cabrelli, and V. Paternostro, Invariance of a shift-invariant space in several
variables, Complex Anal. Oper. Theor. 2010.
[8] J. Benedetto, W. Czaja, P. Gadzinski and A. Powell, The Balian-Low Theorem and regularity
of Gabor systems, J. Geom. Anal., 13 (2003), 239 -- 254.
[9] M. Bownik, The structure of shift-invariant subspaces of L2(Rn), J. Funct. Anal., 177(2000),
282 -- 309.
[10] C. K. Chui and Q. Sun, Tight frame oversampling and its equivalence to shift-invariance of
affine frame operators, Proc. Amer. Math. Soc., 131 (2003), 1527 -- 1538.
[11] C. K. Chui and Q. Sun, Affine frame decompositions and shift-invariant spaces, Appl. Com-
put. Harmon. Anal., 20 (2006), 74 -- 107.
[12] C. de Boor, R.A. DeVore, and A. Ron, The structure of finitely generated shift-invariant
spaces in L2(Rd), J. Funct. Anal. 119 (1994), no. 1, 37 -- 78.
[13] J. Hogan and J. Lakey, Non-translation-invariance and the synchronization problem in
wavelet sampling Acta Appl. Math., 107 (2009), 373 -- 398.
[14] M. Kolountzakis and T. Wolff, On the Steinhaus tiling problem. Mathematika 46 (1999),
253 -- 280.
[15] A. Ron, Z. Shen, Frames and stable bases for shift invariant subspaces of L2(Rd), Canadian
Journal of Mathematics, 47(1995), 1051-1094.
[16] M. Unser and T. Blu, Fractional splines and wavelets, SIAM Rev., 42 (2000), 43 -- 67.
[17] E. Weber, On the translation invariance of wavelet subspaces, J. Fourier Anal. Appl., 6
(2000), 551 -- 558.
(Romain Tessera) Department of Mathematics, Ecole Normale Superieure de Lyon, France
E-mail address: [email protected]
(Haichao Wang) Department of Mathematics, University of California, Davis, CA 95616
E-mail address: [email protected]
|
1502.05909 | 1 | 1502 | 2015-02-20T15:43:20 | Extensions of the Inequalities of Hardy and Hilbert | [
"math.FA",
"math.OA"
] | In this note we produce generalized versions of the classical inequalities of Hardy and of Hilbert and we establish their equivalence. Our methods rely on the H^1-BMOA duality. We produce a class of examples to establish that the generalizations are non-trivial. | math.FA | math |
EXTENSIONS OF THE INEQUALITIES OF HARDY AND
HILBERT
VERN I. PAULSEN AND DINESH SINGH
Abstract. In this note we produce generalized versions of the classical
inequalities of Hardy and of Hilbert and we establish their equivalence.
Our methods rely on the H 1
− BM OA duality. We produce a class of
examples to establish that the generalizations are non-trivial.
1. Introduction
The classical inequality of Hardy states that there exists a constant A so
n=0 anzn is a function in the Hardy space H 1(D), then
that if f (z) =P∞
(1.1)
an
n + 1 ≤ Akfk1.
∞
Xn=0
The well known inequality of Hilbert states that there exists a constant B
so that if (αn), (βn) are two sequences in the HIlbert space ℓ2
+ = ℓ2(Z+),
then
(1.2)
∞
∞
Xm=0
Xn=0
αnβn
n + m + 1 ≤ Bk(αn)k2k(βn)k2.
In this paper we establish generalizations of both of these inequalities
by replacing the sequences (cid:8) 1
n(cid:9) and n 1
n+m+1o with a larger class of se-
quences. We then prove that the two generalized inequalities are equivalent.
In particular, it follows that the two classical inequalities (1.1) and (1.2)
are equivalent. The literature frequently mentions the fact that Hilbert’s
inequalitiy (1.2) implies Hardy’s inequality (1.1), but we found no mention
of the converse.
The inequalities of Hardy (1.1) and Hilbert (1.2) are extremely important
for a variety of reasons and they are still of current interest. For some classic
expositions, we refer the reader to [1], [2], [3] and [4]. For a sampling of the
current research interest, we refer the reader to [5], [6], [7], [8], [9], and [10].
2000 Mathematics Subject Classification. Primary 26D15; 30H10; Secondary 30H35.
Key words and phrases. generalized Hardy inequality; generalized Hilbert inequality;
equivalence; H 1
− BM OA duality.
Research supported in part by NSF .
1
2
V. I. PAULSEN AND D. SINGH
2. Preliminary Results
The Lebesque spaces on the unit circle T in the complex plane shall be de-
noted by Lp, 1 ≤ p ≤ ∞. The Hardy spaces H p are the subspaces consisting
of all f in Lp whose negative Fourier coefficients vanish. In particular, the
+, via the map en → zn, n ≥ 0.
Hilbert space H 2 will often be identified with ℓ2
The H p spaces can be identified as spaces of holomorphic functions on the
open unit disk D that satisfy the condition:
(2.1)
sup
0<r<1(cid:18) 1
2π Z 2π
0
f (reiθ)pdθ(cid:19)1/p
< ∞,
The dual of H 1 is the Banach space BM OA defined as all functions
and for f ∈ H p this supremum is equal to the norm of f in Lp.
f ∈ H 1 such that
kfk∗ = sup
I
I(f − I(f )) < ∞,
where I varies over all subarcs of T, I denotes the normalized arclength of
I and
I(f ) =
f (θ)dθ.
1
IZI
The norm of a function f ∈ BM OA is given by
(2.2)
kfk = f (0) + kfk∗.
We recall that by the famous duality theorem of Fefferman, BM OA can be
identified isometrically and conjugate linearly with the dual of H 1 via the
dual pairing given by
hf, pi =
1
2π Z 2π
0
p(eiθ)f (eiθ)dθ,
Finally, we shall also need Fefferman’s second characterization of BMOA.
for f ∈ BM OA and p ∈ H 1 a polynomial.
Namely, that a function f ∈ H 1 is in BM OA if and only if (1−r2)f ′(reiθ)2drdθ
is a Carleson measure on D.
For the details of these facts, we refer the reader to [1], [2], and [3].
3. Statement of the Main Results
We state our three main results below, together with their immediate
corollaries, but we defer their proofs until later sections. We let X denote
the space of sequences {cn} such that
(3.1)
k=0(k + 1)2ck2
k{cn}k2 = sup(cid:26)Pn
n + 1
: n ∈ N(cid:27) < ∞.
It is not hard to see that X is a Banach space in this norm and that ck = 1
k+1
is an element of unit norm in this space.
HARDY AND HILBERT
3
Theorem 3.1 (Generalized Hardy Inequality). There is a constant A so
that if (cn) is a sequence in X, then for any f (z) =P∞
∞
k=0 anzn in H 1,
ancn ≤ Ak(cn)kkfk1.
Xn=0
Corollary 3.2 (Hardy’s Inequality). There is a constant δ such that for
any f (z) =P∞
n=0 anzn in H 1,
∞
an
n + 1 ≤ δkfk1.
Xn=0
Proof. Set cn = 1
n+1 and apply Theorem 3.1.
(cid:3)
Theorem 3.3 (Generalized Hilbert Inequality). There is a constant B so
that if (cn) ∈ X and (an), (bn) in ℓ2
+, then
∞
Xn,m=0
anbmcn+m ≤ Bk(cn)kk(an)k2k(bn)k.
Corollary 3.4 (Hilbert’s Inequality). There is a constant B so that for any
(an), (bn) in ℓ2
+,
∞
Xm,n=0
anbm
n + m + 1 ≤ Bk(an)k2k(bn)k2.
Proof. Apply Theorem 3.3 with cn = 1
n+1 .
(cid:3)
Theorem 3.5. For each (cn) in X, the corresponding generalized Hardy
inequality is equivalent to the corresponding generalized Hilbert inequality.
Moreover, the least constant A in the generalized Hardy inequality is equal
to the least constant B in the generalized Hilbert inequality.
Corollary 3.6. The Hardy inequality and Hilbert inequality are equivalent.
4. Proof of Theorem 3.5
We present the proof of Theorem 3.5 first because it is the most direct and
demonstrates the equivalence of the classic Hardy and Hilbert inequalities.
We first prove that the generalized Hilbert inequality implies the general-
+ so that by the generalized
ized Hardy inequality. Let (an) and (bn) be in ℓ2
Hilbert inequality(Theorem 3.3),
∞
anbncn+m ≤ Bk(cn)kk(an)k2k(bn)k2,
Xn,m=0
where k(cn)k is defined by (3.1).
Let f (z) =P∞
as f (z) = g(z)h(z) where g(z) =P∞
n=0 αnzn be in H 1. It is well known that f can be factorized
n=0 γnzn are both
n=0 βnzn and h(z) =P∞
4
V. I. PAULSEN AND D. SINGH
in H 2 and kfk1 = kgk2khk2. Applying the generlized Hilbert inequality to
the sequences (βn) and (γn) we have that
∞
Xn,m=0
βnγncn+m ≤ Bk(cn)kk(βn)k2k(γn)k2.
This inequality may be rewritten as
∞
m
Xm=0(cid:0)
Xk=0
cmβkγm−k(cid:1) ≤ Bk(cn)kkgk2khk2 = Bk(cn)kkf1k.
By the Cauchy product formula,
αm =
m
Xk=0
∞
βkγm−k ≤
m
Xk=0
βkγm−k,
and hence,
αmcm ≤ Bk(cm)kkfk1,
Xm=0
so that the generalized Hardy inequality follows. This also shows that A ≤
B.
We now prove that the generalized Hardy inequality implies the gener-
alized Hilbert inequality. Note that to prove the generalized Hilbert in-
equality it is sufficient to consider sequences in ℓ2
+ of non-negative terms.
Choosing any two sequences, (an), (bn) in ℓ2
+ with an ≥ 0 and bn ≥ 0.
n=0 bnzn we obtain two functions
n=0 dnzn is in H 1, where
n=0 anzn and h(z) = P∞
Setting g(z) = P∞
in H 2. Hence the function f (z) = g(z)h(z) = P∞
dn =Pn
Xn=0
Thus, by the generalized Hardy inequality(Theorem 3.1),
∞
dncn ≤ Ak(cn)kkfk1 ≤ Ak(cn)kkgk2khk2 = Ak(cn)kk(an)k2k(bn)k2.
k=0 akbn−k.
Substituting for dn we have
∞
∞
Xn,m=0
cn+manbm =
n
Xn=0(cid:0)
Xk=0
cnakbn−k(cid:1) ≤ Ak(cn)kk(an)k2k(bn)k2,
and so the generalized Hilbert inequality follows. Moreover, we see that
B ≤ A and so the best constants must be equal.
5. Proofs of Theorems 3.1 and 3.3
We begin with the proof of the generalized Hardy inequality. Note that
K = sup
0≤r<1
r
1 − r2[ 1
1−r
]!2
is finite, where [t] denotes the greatest integer less than or equal to t.
HARDY AND HILBERT
5
Given I an arc on the unit circle, we let
z
R(I) = {z ∈ D :
z ∈ I, 1 − z ≤ I},
2π
where I = length(I)
denotes the normalized arc length. Recall that a pos-
itive Borel measure µ on D is called a Carleson measure if there is a
constant k so that µ(R(I)) ≤ kI for all subarcs.
to BM OA if and only if
The theorems of Fefferman and Garsia say that a function f in H 1 belongs
(1 − r2)f ′(reiθ)2rdrdθ
is a Carleson measure. Moreover, in this case if we let η(f )2 be the least
constant k in the Carleson condition, then η(f ) is essentially, the Garsia
norm and hence is a norm on BM OA that is equivalent to kfk∗.
Proposition 5.1. Let (cn) ∈ X, and set g(z) = P∞
n=0 cnzn, then g ∈
BM OA and
Z ZR(I)
(1 − r2)g′(reiθ)2rdrdθ ≤ 2Kk(cn)k2I.
k=m k2ck2 ≤ An. Note that for 0 < r < 1,
Consequently, the linear map L : X → BM OA defined by L((cn)) = g(z) is
bounded.
Proof. Set A = k(cn)k2 so thatPm+n
Z 2π
k2ck2r2k =
Xp=0
g′(reiθ)2dθ =
m2cm2r2m ≤
Xm=p[ 1
1 − r2[ 1
)2 ≤ AK
1
1 − r
]+2A(p + 1)[
1
1 − r
1
1 − r
∞
Xp=0
∞
Xk=1
] ≤ A[
(p+1)[ 1
1−r
]
r2p[ 1
1−r
1
]
1−r
0
∞
]+1
1−r
]r2(
.
Hence,
r + r2dr ≤ 2AKI
(1 − r2)g′(reiθ)2rdrdθ ≤ AKZ 1
Z ZR(I)
and we have shown the Carleson estimate. Hence g ∈ BM OA with η(g) ≤
√2Kk(cn)k.
Proof of the generalized Hardy inequality. Let (cn) ∈ X, and let f (z) =
P∞
n=0 anzn ∈ H 1. Write an = αnan where αn = 1. Set g(z) =P∞
so that by Proposition 4.1, g is in BM OA with norm independent of the
particular choice of α’s.
1−I
(cid:3)
n=0 αncnzn
We have that
hf, gi = lim
r→1−Z 2π
0
f (reiθ)g(reiθ)dθ =
∞
Xn=0
ancn.
6
V. I. PAULSEN AND D. SINGH
Hence,
∞
ancn ≤ kfk1kgk∗ ≤ kLkkfk1k(cn)k
Xn=0
and the proof is complete.
Proof of the generalized Hilbert Inequality. Given (an), (bn) ∈ ℓ2
(cn) ∈ X, set f (z) =P∞
H 2 and hence, f (z)h(z) =P∞
Defining g(z) =P∞
Xn,m=0
anbmcn+m = hf h, gi ≤
n=0 anzn and h(z) =P∞
n=0 cnzn and arguing as above, we have that
n,m=0 anbmzn+m is in H 1.
+ and
n=0 bnzn. Then f, h are in
∞
kLkk(cn)kkf hk1 ≤ kLkk(cn)kk(an)k2k(bn)k2,
and the result follows.
6. Slow Decay in X
The importance of the classic Hardy inequality, which corresponds to the
sequence cn = 1
n+1 , is that it gives us information about the rate of decay of
coefficients of H 1 functions. For this reason it is interesting to seek elements
(cn) ∈ X which tend to 0 at a slower rate. In particular, we shall show that
there exist sequences in X so that sup{nscn : n ∈ N} = ∞ for any s > 1/2.
To this end fix an r, 1/2 ≤ r ≤ 1 and fix β > 1. We define cn inductively
as follows: Set c1 = 1 so that 12c12 ≤ 1β and 12c12 + (1+1)2
(1+1)2r ≤ (1 + 1)β
Assume that c1, ..., cn have been defined so that
m
(6.1)
k2ck2 ≤ mβ for m = 1, 2, ..., n.
Xk=1
k=1 k2ck2 + (n+1)2
(n+1)2r ≤ β(n + 1), then we set cn+1 = 1
(n+1)r and we
have that (6.1) holds for m = 1, ..., (n + 1). If this inequality is false, then we
set cn+1 = 1
k=1 k2ck2 + (n + 1)2cn+12 ≤ βn + 1 ≤
β(n + 1), so again (6.1) holds for m = 1, ..., (n + 1).
In this way we define a sequence cn that satisfies (6.1) for all n with
n+1 and observe that Pn
If Pn
cn = 1
nr or cn = 1
n for each n.
Now suppose that ck = 1
k for n ≤ k ≤ n + m. Then
k2ck2 +
(n + m + 1)2
(n + m + 1)2r =
n
Xk=1
k2ck2 + (m − 1) + (n + m + 1)2−2r
≤ βn + (m − 1) + (n + m + 1)2−2r.
n+m
Xk=1
Since,
βn + (m − 1) + (n + m − 1)2−2r ≤ β(n + m + 1),
HARDY AND HILBERT
7
if and only if (n + m− 1)2−2r ≤ (β − 1)m + β + 1 and 2− 2r < 1 we see that
eventually, there will exist an m for which this inequality holds and then
cn+m+1 =
1
(n+m+1)r .
Thus, cn = 1
nr infinitely often. In particular, sup{nscn : n ∈ N} will be
infinite for any s > r ≥ 1/2.
References
[1] P.L. Duren, The Theory of H p spaces, Academic Press, London-New York, 1970.
[2] J.B. Garnet, Bounded Analytic Functions, Academic Press, New York, 1970.
[3] G.H. Hardy, J.E. Littlewood, and G. Polya, Inequalities, Cambridge University Press,
1952.
[4] H. Helson, Harmonic Analysis, Hindustan Book Agency, 1995.
[5] H.L. Montgomery and J.D. Vaaler, A further generalization of Hilbert’s inequality,
Mathematika 46(1999), No.1, 35-39.
[6] K. Oleszkiewicz, An elementary proof of Hilbert’s inequality, Amer. Math. Monthly
100(1993), No. 3, 276-280.
[7] M. Sabahheh, On an argument of Korner and Hardy’s inequality, Analysis Mathe-
matica 34(2008), 51-57.
[8] S. Saeki and H. Yamaguchi, On Hardy’s inequality and Paley’s gap theorem, Hokkaida
Math. Journal 19(1990), No. 2, 289-296.
[9] Bicheng Yang, On a relation between Hilbert’s inequality and a Hilbert type inequality,
Appl. Math. Letters 21(2008), No. 5, 483-488.
[10] Kehe Zhu, Translating inequalities between Hardy and Bergman spaces, Amer. Math.
Monthly 111(2004), 520-525.
Department of Mathematics, University of Houston, Houston, Texas 77204-
3476, U.S.A.
E-mail address: [email protected]
Department of Mathematics, University of Delhi, Delhi, India
|
1312.3042 | 1 | 1312 | 2013-12-11T06:21:10 | Perturbation of Browder Spectrum of Upper-triangular Operator Matrices | [
"math.FA"
] | Let $M_{C}=\left(\begin{array}{cc}A&C\\0&B\\\end{array} \right)$ is a 2-by-2 upper triangular operator matrix acting on the Banach space $X\oplus Y$ or Hilbert space $H\oplus K$. For the most import spectra such as spectrum, essential spectrum and Weyl spectrum, the characterizations of the perturbations of $M_{C}$ on Banach space have been presented. However, the characterization for the Browder spectrum on the Banach space is still unknown. The goal of this paper is to present some necessary and sufficient conditions for $M_C$ to be Browder for some $C\in B(Y,X)$ by using an alternative approach based on matrix representation of operators and the ghost index theorem. Moreover, in the Hilbert space case the characterizations of the Fredholm and invertible perturbations of Browder spectrum are also given. | math.FA | math | Perturbation of Browder Spectrum of
Upper-triangular Operator Matrices∗
Shifang Zhang1†, Huaijie Zhong1, Lin Zhang 2
1School of Mathematics and Computer Science, Fujian Normal University, Fuzhou 350007, P. R. China
2Institute of Mathematics, Hangzhou Dianzi University, Hangzhou 310018, P. R. China
3
1
0
2
c
e
D
1
1
]
.
A
F
h
t
a
m
[
1
v
2
4
0
3
.
2
1
3
1
:
v
i
X
r
a
Abstract Let MC = (cid:18) A C
0 B (cid:19) is a 2-by-2 upper triangular operator matrix act-
ing on the Banach space X ⊕ Y or Hilbert space H ⊕ K. For the most import spectra
such as spectrum, essential spectrum and Weyl spectrum, the characterizations of
the perturbations of MC on Banach space have been presented. However, the char-
acterization for the Browder spectrum on the Banach space is still unknown. The
goal of this paper is to present some necessary and sufficient conditions for MC to
be Browder for some C ∈ B(Y, X) by using an alternative approach based on matrix
representation of operators and the ghost index theorem. Moreover, in the Hilbert
space case the characterizations of the Fredholm and invertible perturbations of
Browder spectrum are also given.
Keywords: operator matrices, Browder spectrum, left semi-Browder spectrum.
1 Introduction
It is well-known that if H is a Hilbert space, T is a bounded linear operator on H, and H1
is an invariant closed subspace of T , then T is of the form
T = (cid:18) ∗ ∗
0 ∗ (cid:19) : H1 ⊕ H ⊥
1 → H1 ⊕ H ⊥
1 ,
which motivated the interest in 2 × 2 upper-triangular operator matrices (see [1-9, 11-20]).
The problems related to the perturbation of spectra of 2-by-2 upper triangular operator
matrices were first studied by H. K. Du and J. Pan in [5], and they given the characterization
of MC to be invertible for some C ∈ B(K, H) in the Hilbert spece case. Some years later, in
∗This work is supported by the NSF of China (Grant Nos. 11301077,11301124, 11226113, 11171066 and 10771191
) , the FED of Fujian Province(Grant No. JA12074), the NSF of Fujian Province (Grant No.2012J05003).
†Corresponding author: E-mail: [email protected]
1
[13] J. K. Han et al. generalized the above result to the Banach space case. In 2002, D. S.
Djordjevi´c in [5] future studied this field for some other spectra and presented the necessary
and sufficient conditions for MC to be a Fredholm and Weyl operator for some C ∈ B(Y, X),
respectively. Also in [5], some sufficient conditions for MC to be a Browder operator were
given. Recently, Cao [1] gave the necessary and sufficient condition for MC to be a Browder
operator in the Hilbert space case. Till now, although for the most import and widely used
spectra such as spectrum, essential spectrum and Weyl spectrum, the characterizations of
the perturbations of upper triangular operator matrices on Banach space have been given,
the characterization for the Browder spectrum is still unknown. The goal of this paper
is to present a necessary and sufficient condition for MC to be Browder for some C ∈
B(Y, X) by using an alternative approach based on matrix representation of operators and
the ghost index theorem, and our result is an extension of the main result in [1] and the
corresponding results from [5, 22]. Moreover, the characterizations of the Fredholm and
invertible perturbations of Browder spectrum are also given.
Throughout this paper, let H and K be the complex infinite dimensional separable Hilbert
spaces, X and Y be the complex infinite dimensional Banach spaces, and B(X, Y ) be the set
of all bounded linear operators from X into Y. For simplicity, we write B(X, X) as B(X).
For T ∈ B(X), T is called Drazin invertible if there exists an operator T D ∈ B(X) such that
T T D = T DT,
T DT T D = T D,
T k+1T D = T k,
for some nonnegative integer k([5]). T ∈ B(X, Y ) is called regular( also called relatively
regular in [5]) if there is an operator B ∈ B(Y, X) such that T = T BT ([4]). It is well-known
that T is regular if and only if R(T ) and N(T ), respectively, are closed and complemented
subspaces of Y and X.
For T ∈ B(X, Y ), let R(T ) and N(T ) be the range and kernel of T , respectively, and
denote α(T ) = dim N(T ) and β(T ) = dim Y /R(T ). If T ∈ B(X), the ascent, denoted by
asc(T ), and the descent, denoted by des(T ), of T are the non-negative integers k defined
respectively by
asc(T ) = inf {k : N(T k) = N(T k+1)},
des(T ) = inf {k : R(T k) = R(T k+1)}.
If such k does not exist, then asc(T ) = ∞, respectively des(T ) = ∞. If the ascent and
the descent of T are finite, then they are equal (see [5]). For T ∈ B(X), if R(T ) is closed
and α(T ) < ∞, then T is said to be upper semi-Fredholm, if β(T ) < ∞, then T is said
to be semi-Fredholm.
If T ∈ B(X) is either upper or lower semi-Fredholm , then T is
said to be semi-Fredholm. For a semi-Fredholm operator T , its index ind (T ) is defined as
ind (T ) = α(T ) − β(T ). The sets of invertible operators, left invertible operators and right
invertible operators on X, respectively, are denoted by G(X), Gl(X) and Gr(X).
The sets of all Fredholm operators, Weyl operators, left semi-Fredholm operators and
right semi-Fredholm operators on X are defined, respectively, by
Φ(X) := {T ∈ B(X) : α(T ) < ∞ and β(T ) < ∞};
Φ0(X) := {T ∈ Φ(X) : ind(T ) = 0};
Φl(X) := {T ∈ B(X) : R(T ) is a closed and complemented subspace of X and α(T ) < ∞};
Φr(X) := {T ∈ B(X) : N(T ) is a closed and complemented subspace of X and β(T ) < ∞}.
2
The sets of all Browder operators, left semi-Browder operators and right semi-Browder
operators on X are defined, respectively, by
Φb(X) := {T ∈ Φ(X) : asc(T ) = des(T ) < ∞};
Φlb(X) := {T ∈ Φl(X) : asc(T ) < ∞};
Φrb(X) := {T ∈ Φr(X) : des(T ) < ∞}.
The spectrum, left spectrum, right spectrum, essential spectrum, Weyl spectrum, left semi-
Fredholm spectrum, right semi-Fredholm spectrum, Browder spectrum, left semi-Browder
spectrum, right semi-Browder spectrum of an operator T ∈ B(H) are respectively denoted
by σ(T ), σl(T ), σr(T ), σe(T ), σw(T ),σle(T ), σre(T ), σb(T ), σlb(T ) and σrb(T ).
One of the main results is : MC is a Browder operator for some C ∈ B(Y, X) if and only
if the following statements are satisfied
(a) A ∈ B(X) is a left semi-Browder operator;
(b) B ∈ B(Y ) is a right semi-Browder operator;
(c) N(A) × N(B) ∼= X/R(A) × Y /R(B).
extension of the corresponding results from [1, 5, 22], that is,
Using the above result, we characterize completely the set TC∈B(Y, X) σb(MC), which is the
σb(MC) = σlb(A)∪σrb(B)∪{λ ∈ C : N(A−λ)×N(B−λ) 6∼= X/R(A−λ)×Y /R(B−λ)}.
\C∈B(Y, X)
Moreover, for any given A ∈ B(H) and B ∈ B(K), it is proved that
\C∈G(K, H)
σb(MC) = \C∈Φ(K, H)
σb(MC)
= σlb(A) ∪ σrb(B) ∪ {λ ∈ C : α(A − λ) + α(B − λ) 6= β(A − λ) + β(B − λ)}.
2 Main Results and Proofs
We begin with some lemmas, which are useful for the proofs of the main results.
Lemma 2.1. ([23] or [2]) For any given A ∈ B(X) and B ∈ B(Y ), if MC is Drazin invertible
for some C ∈ B(Y, X), then
(1) des(B) < ∞ and asc(A) < ∞;
(2) des(A∗) < ∞ and asc(B∗) < ∞.
Lemma 2.2. ([5] Lemma 2.3) Let M and N be finite dimensional spaces. If dim M = dim N
and X × M ∼= Y × N, then X ∼= Y.
Lemma 2.3. ([13]) If T ∈ B(X, Y ), S ∈ B(Y, Z) and ST ∈ B(X, Z) are regular, then
3
N(T ) × N(S) × Z/R(ST ) ∼= N(ST ) × Y /R(T ) × Z/R(S).
Some well-known facts are collected in the following lemma.
Lemma 2.4. ([24]) For any given A ∈ B(X), B ∈ B(Y ) and C ∈ B(Y, X), we have:
(1) if any two of operators A, B and MC are invertible (resp., Fredholm, Weyl, Brow-
der,Drazin invertible ), then so is the third;
(2) if A is Browder, then B is left semi-Browder if and only if so is MC ;
(3) if B is Browder, then A is right semi-Browder if and only if so is MC.
Lemma 2.5([5]). Let A ∈ B(X) and B ∈ B(Y ) be given and consider the statements:
(i) MC is a Weyl operator for some C ∈ B(Y, X)
(ii) A ∈ B(X) and B ∈ B(Y ) satisfy the following conditions:
(a) A is a left semi-Fredholm operator;
(b) B is a right semi-Fredholm operator;
(c) N(A) × N(B) ∼= X/R(A) × Y /R(B).
Then (i) ⇐⇒ (ii)
Lemma 2.6 ([13]). Let A ∈ B(X) and B ∈ B(Y ) be given and consider the statements:
(i) MC is invertible for some C ∈ B(Y, X)
(ii) A ∈ B(X) and B ∈ B(Y ) satisfy the following conditions:
(a) A is left invertible;
(b) B is right invertible;
(c) N(B) ∼= X/R(A).
Then (i) ⇔ (ii)
Lemma 2.7. For T ∈ B(X), T is left semi-Browder if and only if T can be decomposed
into the following form with respect to space decomposition X = X1 ⊕ X2
T = (cid:18) T1 T12
T2 (cid:19) ,
0
where dim(X1) < ∞, T1 is nilpotent, and T2 is left invertible.
Proof. Necessity. Suppose that T is left semi-Browder. Then we can assume p = asc(T ) <
∞. Let X1 = N(T p). Then dim X1 < ∞ since T is left semi-Fredholm. Therefore, there
exists a closed subspace X2 of X such that X1 ⊕ X2 = X. With respect to the space
decomposition X = X1 ⊕ X2, we have
T = (cid:18) T1 T12
T2 (cid:19) : X1 ⊕ X2 → X1 ⊕ X2.
0
Obviously, T1 is nilpotent and hence T1 ∈ Φb(X1). Moreover, it follows from Lemma 2.4 that
T2 ∈ Φlb(X2). Next, we shall show T2 is injective. If not, there exists some 0 6= x ∈ X2 such
4
that T2x = 0. Put z = (cid:18) 0
x (cid:19) ∈
x (cid:19). Then T p+1(cid:18) 0
x (cid:19) = T p(cid:18) T12x
0 (cid:19) = 0, and so 0 6= (cid:18) 0
N(T p+1). This is in contradiction with the assumption that asc(T ) = p < ∞, thus, T2 is left
invertible.
Sufficiency. It is evident.
Lemma 2.8. For T ∈ B(X), T is right semi-Browder if and only if T can be decomposed
into the following form with respect to some space decomposition X = X1 ⊕ X2
T = (cid:18) T1 T12
T2 (cid:19) ,
0
where dim(X2) < ∞, T1 is right invertible, and T2 is nilpotent.
Proof. Necessity. Suppose T is right semi-Browder. Then we can assume p = des(T ) < ∞.
Now put X1 = R(T p). Then dim X/X1 < ∞ since T p is a right semi-Browder operator.
Therefore, there exists a closed finite dimensional subspace X2 of X such that X1 ⊕ X2 = X.
Corresponding the space decomposition X = X1 ⊕ X2, we have
T = (cid:18) T1 T12
T2 (cid:19) : X1 ⊕ X2 → X1 ⊕ X2.
0
Obviously, T1 is surjective and T P
Φb(X2), it follows from Lemma 2.4 that T1 ∈ Φrb(X2), and so T1 is right invertible.
2 = 0. Moreover, noting that dim X2 < ∞ implies T2 ∈
Sufficiency. It is evident.
One of our main results is as follows
Theorem 2.9. Let A ∈ B(X) and B ∈ B(Y ) be given and consider the statements:
(i) MC is a Browder operator for some C ∈ B(Y, X)
(ii) A ∈ B(X) and B ∈ B(Y ) satisfy the following conditions:
(a) A is a left semi-Browder operator;
(b) B is a right semi-Browder operator;
(c) N(A) × N(B) ∼= X/R(A) × Y /R(B).
Then (i) ⇔ (ii).
Proof. Necessity. Noting the fact that an operator T is a Browder operator if and only
if T is a Drazin invertible Weyl operator, the implication follows from Lemmas 2.1 and 2.5
immediately.
Sufficiency. Since A is a left semi-Browder operator, from Lemma 2.7, we have
A = (cid:18) A1 A12
0 A2 (cid:19) : X1 ⊕ X2 → X1 ⊕ X2,
5
where A1 is nilpotent, A2 is left invertible and dim(X1) < ∞. Note that
A = (cid:18) I
0 A2 (cid:19)(cid:18) A1 A12
I (cid:19) .
0
0
Applying Lemma 2.3, we have
0
N((cid:18) I
Since (cid:18) I
0 A2 (cid:19))×N((cid:18) A1 A12
0 A2 (cid:19) is injective,
0
0
I (cid:19))×X/R(A) ∼= X/R(cid:18) I
0 A2 (cid:19)×X/R(cid:18) A1 A12
I (cid:19)×N(A).
0
0
0
N((cid:18) A1 A12
I (cid:19)) × X/R(A) ∼= X/R(cid:18) I
I (cid:19) × N(A).
It follows from dim(X1) < ∞ that A1 ∈ Φb(X1). By Lemma 2.4, we obtain (cid:18) A1 A12
0 A2 (cid:19) × X/R(cid:18) A1 A12
0
0
0
I (cid:19) ∈
Φb(X1 ⊕ X2), which implies
α((cid:18) A1 A12
I (cid:19)) = β((cid:18) A1 A12
I (cid:19)) < ∞.
0
0
Using Lemma 2.2, we have
X/R(A) ∼= X/R(cid:18) I
0 A2 (cid:19) × N(A),
0
and hence
X/R(A) ∼= X2/R(A2) × N(A).
(1)
Meanwhile, since B ∈ B(Y ) is a right semi-Browder operator, from Lemma 2.8 we have
B = (cid:18) B1 B12
0 B2 (cid:19) : Y1 ⊕ Y2 → Y1 ⊕ Y2,
where B2 is nilpotent, B1 is right invertible and dim(Y2) < ∞. Similar to the preceding
arguments, we have
B = (cid:18) I B12
0 B2 (cid:19)(cid:18) B1 0
I (cid:19) ,
0
and
N((cid:18) I B12
0 B2 (cid:19)) × N((cid:18) B1 0
I (cid:19)) × Y /R(B) ∼= Y /R(cid:18) I B12
0 B2 (cid:19) × Y /R(cid:18) B1 0
I (cid:19) × N(B).
0
0
6
Since (cid:18) B1 0
0
I (cid:19) is surjective,
0 B2 (cid:19)) × N((cid:18) B1 0
N((cid:18) I B12
0
I (cid:19)) × Y /R(B) ∼= Y /R(cid:18) I B12
0 B2 (cid:19) × N(B).
It follows from dim(Y2) < ∞ that B2 ∈ Φb(Y2). By Lemma 2.4, we have
which implies
(cid:18) I B12
0 B2 (cid:19) ∈ Φb(Y ),
α((cid:18) I B12
0 B2 (cid:19)) = β((cid:18) I B12
0 B2 (cid:19)) < ∞.
Then it follows from Lemma 2.2 that
N(B1) × Y /R(B) ∼= N(B).
(2)
Combining (1) and (2) with the assumption N(A) × N(B) ∼= X/R(A) × Y /R(B), it is easy
to show that
N(A) × N(B1) × Y /R(B) ∼= X2/R(A2) × N(A) × Y /R(B).
Moreover, since α(A) and β(B) are finite, it results from Lemma 2.2 that
N(B1) ∼= X2/R(A2).
(3)
Observe that B1 is right invertible and A2 is left invertible. It follows from Lemma 2.6 that
there exists some C0 ∈ B(Y1, X2) such that
(cid:18) A2 C0
0 B1 (cid:19) ∈ G(X2 ⊕ Y1).
Define C ∈ B(Y, X) as
C = (cid:18) 0
C0 0 (cid:19) : Y1 ⊕ Y2 → X1 ⊕ X2.
0
Then MC can be written as the following form:
MC =
: X1 ⊕ X2 ⊕ Y1 ⊕ Y2 → X1 ⊕ X2 ⊕ Y1 ⊕ Y2.
A1 B12
0
0 A2 C0
0
0
0
0
0 B1 B12
0
0 B2
We conclude from Lemma 2.4 that MC = (cid:18) A C
0 B (cid:19) ∈ Φb(X ⊕ Y ). This completes the
proof.
7
The next corollary is immediate from Theorem 2.9, which characterizes the setTC∈B(Y, X) σb(MC).
Corollary 2.10. Let A ∈ B(X) and B ∈ B(Y ). Then we have
\C∈B(Y, X)
σb(MC) = σlb(A)∪σrb(B)∪{λ ∈ C : N(A−λ)×N(B−λ) 6∼= X/R(A−λ)×Y /R(B−λ)}.
As a particular case of Corollary 2.10, we can rewrite the main result in [1]
Corollary 2.11. Let H and K be the complex infinite dimensional separable Hilbert space,
A ∈ B(H) and B ∈ B(K). Then
\C∈B(K, H)
σb(MC) = σlb(A) ∪ σrb(B) ∪ {λ ∈ C : α(A − λ) + α(B − λ) 6= β(A − λ) + β(B − λ)}.
can be equal to spectrum, essential spectrum, Weyl spectrum and so on. Similarly, the sets
In [25], we have studied the sets TC∈G(K, H) σ∗(MC) and TC∈Φ(K, H) σ∗(MC), where σ∗(·)
TC∈G(K, H) σb(MC) and TC∈Φ(K, H) σb(MC) are studied in the next theorem. Firstly, we give
a well-known result which is needed in the proof of the next theorem.
Lemma 2.12. [18 Corollary 2.2] Let A ∈ B(H) and B ∈ B(K). If R(B) is closed, then MC
is left invertible if and only if both A and C1 are left invertible, where C = (cid:18) C1 C2
C3 C4 (cid:19) :
N(B) ⊕ N(B)⊥ → R(A)⊥ ⊕ R(A).
Duality, we have
Corollary 2.13. Let A ∈ B(H) and B ∈ B(K).
If R(A) is closed, then MC is right
invertible if and only if both B and C1 are right invertible, where C = (cid:18) C1 C2
C3 C4 (cid:19) :
N(B) ⊕ N(B)⊥ → R(A)⊥ ⊕ R(A).
Combing Lemma 2.12 and Corollary 2.13, we have
Corollary 2.14. Let A ∈ B(H) and B ∈ B(K). Then MC is invertible if and only if
A is left invertible, B is right invertible and C1 is invertible, where C = (cid:18) C1 C2
C3 C4 (cid:19) :
N(B) ⊕ N(B)⊥ → R(A)⊥ ⊕ R(A).
We are now ready to give the last main result.
Theorem 2.15. For any given A ∈ B(H) and B ∈ B(K) we have
8
\C∈G(K, H)
σb(MC) = \C∈Φ(K, H)
σb(MC) = SP R,
(4)
where SP R = σlb(A) ∪ σrb(B) ∪ {λ ∈ C : α(A − λ) + α(B − λ) 6= β(A − λ) + β(B − λ)}.
Proof. Noting the fact
\C∈G(K, H)
σb(MC) ⊇ \C∈Φ(K, H)
σb(MC) ⊇ \C∈B(K, H)
σb(MC),
it follows from Corollary 2.11 that in order to prove (4), we only need to prove
\C∈G(K, H)
σb(MC) ⊆ σlb(A) ∪ σrb(B) ∪ {λ ∈ C : α(A − λ) + α(B − λ) 6= β(A − λ) + β(B − λ)}.
For this, it suffices to show that, if
0 6∈ σlb(A) ∪ σrb(B) ∪ {λ ∈ C : α(A − λ) + α(B − λ) 6= β(A − λ) + β(B − λ)},
then
0 6∈ \C∈G(K, H)
σb(MC).
Now suppose that A ∈ Φlb(H) and B ∈ Φrb(K) with α(A) + α(B) = β(A) + β(B). We shall
prove that there exists some C ∈ G(K, H) such that MC ∈ Φb(H ⊕ K).
From Lemmas 2.7 and 2.8, we have
A = (cid:18) A1 A12
0 A2 (cid:19) : H1 ⊕ H2 → H1 ⊕ H2, B = (cid:18) B1 B12
0 B2 (cid:19) : K1 ⊕ K2 → K1 ⊕ K2,
where dim(H1) < ∞, dim(K2) < ∞, A2 is left invertible, B1 is right invertible, and both A1
and B2 are nilpotent. Similar to the proof of Theorem 2.9, we have
Moreover, considering the operator C ′ ∈ B(K1, H2) with the following expression
α(B1) = β(A2).
(5)
C ′ = (cid:18) C ′
C ′
1 C ′
2
3 C ′
4 (cid:19) : N(B1) ⊕ N(B1)⊥ → R(A2)⊥ ⊕ R(A2),
it follows from Corollary 2.11 (replacing operators A, B, C by A1, B1, C ′, respectively) that
(cid:18) A2 C ′
0 B1 (cid:19) is invertible if and only if so is C ′
1,
since A2 is left invertible and B1 is right invertible.
Furthermore, since dim N(B1)⊥ = dim R(A2) = ∞, we have
dim K2 + dim N(B1)⊥ = dim R(A2) + dim H1 = ∞.
(6)
9
It follows from (5) and (6) that there exist two invertible operators Q1 ∈ B(N(B1), R(A2)⊥)
and
Q2 = (cid:18) T1 T2
T3 T4 (cid:19) : N(B1)⊥ ⊕ K2 → R(A2) ⊕ H1.
Define an operator
C : K → H by C =
Q1
0
0
0
0
T1 T2
T3 T4
: N(B1) ⊕ N(B1)⊥ ⊕ K2 → R(A2)⊥ ⊕ R(A2) ⊕ H1.
Obviously C ∈ B(K, H) is invertible. Next we will prove that MC is a Browder operator.
Observing that
A1 A12
0
0 A′
2
0
0
0
0
T3
0
0
0 Q1
0
T1
0 B′
0
T4
0
T2
1 B12
0 B2
: H1 ⊕ H2 ⊕ N(B1) ⊕ N(B1)⊥ ⊕ K2 → H1 ⊕ R(A2)⊥ ⊕
MC =
R(A2) ⊕ K1 ⊕ K2.
=
A1 A12 C1
T4
0 A2 C3 C2
0 B1 B12
0
0
0
0 B2
: H1 ⊕ H2 ⊕ K1 ⊕ K2 → H1 ⊕ H2 ⊕ K1 ⊕ K2,
where A2 = (cid:18) 0
From Corollary 2.14 we have that (cid:18) A2 C3
2 (cid:19) , C1 = (cid:0) 0 T3 (cid:1) , C3 = (cid:18) Q1
T1 (cid:19) , C4 = (cid:18) 0
T2 (cid:19) , B1 = (cid:0) 0 B′
1 (cid:1) .
0 B1 (cid:19) is invertibility, and then it follows from
A′
0
0
Lemma 2.4 that MC is a Browder operator. This ends the proof.
References
[1] X. H. Cao. Browder spectra for upper triangular operator matrices, J. Math. Anal.
Appl., 342 (2008), 477-484.
[2] D. S. Djordjevi´c. Perturbations of spectra of operator matrices, J. Operator Theory, 48
(2002), 467-486.
[3] D. S. Djordjevi´c, M. Z. Kolundzija. Generalized invertibility of operator matrices, Ark.
Mat., 50 (2012), 259-267.
[4] S. V. Djordjevi´c, Y. M. Han. spectral continuity for operator matrices, Glasg. Math. J.,
43 (2001), 487-490.
[5] S. V. Djordjevi´c, B. P. Duggal, Drazin invertibility of the diagonal of an operator, Linear
Multilinear Algebra 60 (2012), 65-71.
10
[6] H. K. Du, J. Pan. Perturbation of spectrum of 2 × 2 operator matrices, Proc. Amer.
Math. Soc., 121 (1994), 761-766.
[7] G. J. Hai, A. Chen. Perturbations of right and left spectra for operator matrices, J.
Operator Theory, 67 (2012), 207-214.
[8] G. J. Hai, A. Chen. On the right (left) invertible completions for operator matrices.
Integral Equations Operator Theory, 67 (2010), 79-93.
[9] J. K. Han, H. Y. Lee, W. Y. Lee. Invertible completions of 2 × 2 upper triangular
operator matrices. Proc. Amer. Math. Soc., 128 (1999), 119-123.
[10] R. E. Harte. The ghost of an index theorem. Proc. Amer. Math. Soc., 106 (1989),
1031-1034.
[11] I. S. Hwang, W. Y. Lee. The boundedness below of 2 × 2 upper triangular operator
matrices, Integr. Equ. Oper. Theory, 39 (2001), 267-276.
[12] W. Y. Lee. Weyl's theorem for operator matrices, Integr. equ. oper. theory, 32 (1998),
319-331.
[13] W. Y. Lee. Weyl spectra of operator matrices, Proc. Amer. Math. Soc., 129 (2000),
131-138.
[14] Y. Li, X. H. Sun, H. K. Du. The intersection of left(right) spectra of 2 × 2 upper
triangular operator matrices, Linear Algebra Appl., 418(2006), 112-121.
[15] E. H. Zerouali, H. Zguitti. Perturbation of spectra of operator matrices and local spectral
theory, J. Math. Anal. Appl., 324(2006), 992-1005.
[16] S. F. Zhang, H. J. Zhong, Q. F. Jiang. Drazin spectrum of operator matrices on the
Banach space, Linear Algebra Appl., 429 (2008), 2067-2075.
[17] S. F. Zhang , Z. Y. Wu, H. J. Zhong, Continuous spectrum, point spectrum and residual
spectrum of operator matrices, Linear Algebra Appl., 433 (2010), 653-661.
[18] S. F. Zhang, H. J. Zhong, J. D. Wu. Spectra of 2 × 2 Upper Triangular Matrices, Acta
Math. Sinica (in Chinese), 54(1) (2011) 41-60.
[19] S. F. Zhang, J. D. Wu. Samuel Multiplicites and Browder spectrum of operator matrices,
Operators and Matrices, 6 (2012), 169-179.
[20] S. F. Zhang, Z. Q. Wu. Characterizations of perturbations of spectra of 2 × 2 upper
triangular operator matrices, J. Math. Anal. Appl. 392 (2012), 103-110.
11
|
1606.06845 | 1 | 1606 | 2016-06-22T08:18:29 | Means refinements via convexity | [
"math.FA"
] | The main goal of this article is to find the exact difference between a convex function and its secant, as a limit of positive quantities. This idea will be expressed as a convex inequality that leads to refinements and reversals of well established inequalities treating different means. The significance of these inequalities is to write one inequality that brings together and refine almost all known inequalities treating the arithmetic, geometric, harmonic and Heinz means, for numbers and operators. | math.FA | math |
MEANS REFINEMENTS VIA CONVEXITY
M. SABABHEH
Abstract. The main goal of this article is to find the exact difference between
a convex function and its secant, as a limit of positive quantities. This idea will
be expressed as a convex inequality that leads to refinements and reversals of
well established inequalities treating different means. The significance of these
inequalities is to write one inequality that brings together and refine almost
all known inequalities treating the arithmetic, geometric, harmonic and Heinz
means, for numbers and operators.
1. introduction
Convex functions and their inequalities have played a major role in the study
of various topics in Mathematics; including applied Mathematics, Mathematical
Analysis and Mathematical Physics. Means and their comparison is indeed an
important application of convexity.
Recall that a function f : I → R, defined on a real interval I, is said to be convex if
f (αx1+βx2) ≤ αf (x1)+βf (x2), when x1, x2 ∈ I and α, β ≥ 0 satisfying α+β = 1.
On the other hand, f : I → R+ is said to be log-convex if g(x) = log f (x) is
convex, or equivalently if f (αx1 + βx2) ≤ f α(x1)f β(x2) for the above parameters.
Speaking of means, the comparison between the weighted arithmetic, geometric
and harmonic means is an immediate consequence of convexity or log-convexity
of the functions x∇ty = (1 − t)x + ty, x#ty = x1−tyt and x!ty = ((1 − t)x−1 +
ty−1)−1, x, y > 0, defined for 0 ≤ t ≤ 1. Adopting these notations, we drop t
when t = 1
2 .
Convexity of the function f (t) = x#ty implies the well known Young's inequality
x#t ≤ x∇ty. On the other hand, convexity of the function g(t) = x!ty implies
the arithmetic-harmonic mean inequality x!ty ≤ x∇ty, while log-convexity of g
implies the geometric-harmonic mean inequality x!ty ≤ x#ty.
These inequalities, though very simple, have some significant applications. For
example, the above Young's inequality implies the celebrated Holder's inequality
kf gk1 ≤ kfkpkgkq for f ∈ Lp(X) and g ∈ Lq(X), for the conjugate exponents
p, q, where X is some measure space.
Among the most interesting applications of the above mean inequalities is the
possible comparison between operators acting on a finite dimensional Hilbert
In the sequel, Mn will denote the space of operators acting on an
space H.
n−deimentional Hilbert space H, M+
n will denotes the cone of semi positive
operators in Mn while M++
n will denotes the cone of strictly positive opera-
tors in Mn. Then the above numerical inequalities have their operator versions
2010 Mathematics Subject Classification. 15A39, 15B48, 26D15, 26B25, 47A30, 47A63.
Key words and phrases. convex functions, means inequalities, norm inequalities.
1
2
M. SABABHEH
2 BA− 1
1
A
1
2 (cid:16)A− 1
n , A∇tB = (1 − t)A + tB and
2 . In this context, we say that A ≤ B for two self-
such as A#tB ≤ A#tB, where A, B ∈ M++
A#tB = A
adjoint operators A and B if B − A ∈ M+
n .
Obtaining the operator versions from the corresponding numerical versions can
be done in different approaches, among which is the application of the following
lemma [2].
2(cid:17)t
Lemma 1.1. Let X ∈ Mn be self-adjoint and let f and g be continuous real
valued functions such that f (t) ≥ g(t) for all t ∈ Sp(X), the spectrum of X.
Then f (X) ≥ g(X).
Recent studies of the topic have investigated possible refinements of the above
inequalities, where adding a positive term to the left side becomes possible. This
idea has been treated in [3, 5, 6, 7, 9, ?, 10, 11, 12], where not only refinements
have been investigated, but reversed versions and much more have been discussed.
Keeping our paper concise, we will not go through the exact results done in
the above references now, however we will comment later how the results in this
paper generalize almost all results in these references, regarding the refinements
and the reverses of the above mean inequalities.
The main goal of this article is to avoid dealing with the specific means, and to
treat a general convexity argument that leads to these refinements. In particular,
we prove that for certain positive quantities Aj(ν)∆jf (ν; a, b), we have
f ((1 − ν)a + νb) +
N
Xj=1
Aj(ν)∆jf (ν; a, b) ≤ (1 − ν)f (a) + νf (b), N ∈ N,
for the convex function f : [a, b] → R. This provides N refining terms of the
inequality f ((1 − ν)a + νb) ≤ (1 − ν)f (a) + νf (b), which follows from convexity
of f . Furthermore, we prove a reversed version and we prove that as N → ∞ the
above inequality becomes an equality. As a natural consequence, we obtain some
refinements and reverses for log-convex functions.
As we will see, the above inequality and its consequences happen to be general-
izations that imply almost all inequalities in the references [3, 5, 6, 9, ?, 10, 11, 12].
This is our main motivation behind this work; to find a formula that implies and
generalizes all other formulae and hence, to enhance our understanding of these
inequalities.
We remark that the proof of the first main result in this work is inspired by
our recent work in [?].
2. main results
For the rest of the paper, the following notations will be adopted. For 0 ≤ ν ≤ 1
and j ∈ N, let
(
kj(ν) = [2j−1ν], rj(ν) = [2jν] and
Aj(ν) = (−1)rj(ν)2j−1ν + (−1)rj (ν)+1h rj(ν)+1
2
i .
(2.1)
MEANS REFINEMENTS VIA CONVEXITY
3
Moreover, if f : [a, b] → R is any function, define
∆jf (ν; a, b) = f(cid:18)(cid:18)1 −
− 2f(cid:18)(cid:18)1 −
kj(ν)
2j−1 (cid:19) a +
2kj(ν) + 1
2j
kj(ν)
2j−1
(cid:19) a +
b(cid:19) + f(cid:18)(cid:18)1 −
2kj(ν) + 1
kj(ν) + 1
2j−1 (cid:19) a +
b(cid:19) , 0 ≤ ν ≤ 1.
2j
kj(ν) + 1
2j−1
b(cid:19)
(2.2)
2.1. Convex functions. We discuss first the inequalities that govern convex
functions, then we apply these inequalities to log-convex functions.
Lemma 2.1. If f : [a, b] → R is convex, then ∆jf (ν; a, b) ≥ 0 for j ∈ N and
0 ≤ ν ≤ 1.
2j−1 b, yj(ν) = (cid:16)1 − kj(ν)+1
Proof. Letting xj(ν) = (cid:16)1 − kj (ν)
and zj(ν) = (cid:16)1 − 2kj(ν)+1
∆jf (ν; a, b) = f (xj(ν)) + f (yj(ν)) − 2f (zj(ν)) ≥ 0, by convexity of f .
2j−1 (cid:17) a + kj (ν)+1
b, it is easy that zj(ν) = xj(ν)+yj (ν)
2j−1(cid:17) a + kj(ν)
(cid:17) a + 2kj(ν)+1
2j−1 b
. The
(cid:3)
2j
2j
2
Remark 2.2. When f : [a, b] → R, we adopt the convention that f (x) = 0 for
x 6∈ [a, b]. This convention will be needed, for example, in the next lemma, when
N = 1 and ν = 1.
Lemma 2.3. Let f : [0, 1] → R be a function and let N ∈ N. Then
N
(1 − ν)f (0) + νf (1) −
Aj(ν)∆jf (ν; 0, 1)
Xj=1
= (cid:0)[2N ν] + 1 − 2N ν(cid:1) f(cid:18) [2N ν]
2N (cid:19) +(cid:0)2N ν − [2N ν](cid:1) f(cid:18)[2N ν] + 1
2N
(cid:19) .
(2.3)
2 , r1(ν) = 0 and k1(ν) = 0. Hence A1(ν) = ν and
Proof. We proceed by induction on N.
When N = 1 and 0 ≤ ν < 1
∆1f (ν; a, b) = f (a) + f (b)− 2f(cid:0) a+b
2 (cid:1) . Then direct computations show the result.
Now if 1
2 ≤ ν < 1, then r1(ν) = 1 and k1(ν) = 0, hence A1(ν) = 1 − ν and
∆1f (ν; a, b) = f (a)+f (b)−2f(cid:0) a+b
2 (cid:1) . Again, direct computations show the result.
When ν = 1, the result follows immediately.
Now assume that (2.3) is true for some N ∈ N. We assert its truth for N + 1.
4
M. SABABHEH
Notice that, using the inductive step,
(1 − ν)f (0) + νf (1) −
N +1
Xj=1
Aj(ν)∆jf (ν; 0, 1)
N
Aj(ν)∆jf (ν; 0, 1) − AN +1(ν)∆N +1f (ν; 0, 1)
= (1 − ν)f (0) + νf (1) −
Xj=1
= (cid:0)[2N ν] + 1 − 2N ν(cid:1) f(cid:18)[2N ν]
2N (cid:19) +(cid:0)2N ν − [2N ν](cid:1) f(cid:18)[2N ν] + 1
− (cid:18)(−1)[2N +1ν]2N ν + (−1)[2N +1ν]+1(cid:20)[2N +1ν] + 1
(cid:21)(cid:19) ×
× (cid:18)f(cid:18) [2N ν]
(cid:19) − 2f(cid:18) 2[2N ν] + 1
2N +1 (cid:19)(cid:19) .
2N (cid:19) + f(cid:18) [2N ν] + 1
2N
2N
2
(cid:19)
(2.4)
Now we treat two cases.
Case I If [2N +1ν] is odd, then we easily see that [2N ν] = [2N +1ν]−1
2
. Therefore,
f(cid:18)[2N ν] + 1
2N
(cid:19) = f(cid:18)[2N +1ν] + 1
2N +1
(cid:19) and f(cid:18)2[2N ν] + 1
2N +1 (cid:19) = f(cid:18) [2N +1ν]
2N +1 (cid:19) .
Substituting these values in (2.4) and simplifying imply
N +1
(1 − ν)f (0) + νf (1) −
Aj(ν)∆j f (ν; 0, 1)
Xj=1
Xj=1
= (cid:0)2N +1ν − [2N +1ν](cid:1) f(cid:18) [2N +1ν] + 1
2N +1
(cid:19) +(cid:0)[2N +1ν] + 1 − 2N +1ν(cid:1) f(cid:18) [2N +1ν]
2N +1 (cid:19) ,
which completes the proof, when [2N +1ν] is odd.
Case II If [2N +1ν] is even, then 2[2N ν] = [2N +1ν] and
f(cid:18)[2N ν]
2N (cid:19) = f(cid:18) [2N +1ν]
2N +1 (cid:19) and f(cid:18) 2[2N ν] + 1
2N +1 (cid:19) = f(cid:18)[2N +1ν] + 1
2N +1
(cid:19) .
Substituting these values in (2.4) and simplifying imply
N +1
(1 − ν)f (0) + νf (1) −
Aj(ν)∆j f (ν; 0, 1)
= (cid:0)2N +1ν − [2N +1ν](cid:1) f(cid:18) [2N +1ν] + 1
2N +1
This completes the proof.
(cid:19) +(cid:0)[2N +1ν] + 1 − 2N +1ν(cid:1) f(cid:18) [2N +1ν]
2N +1 (cid:19) .
(cid:3)
Corollary 2.4. Let f : [0, 1] → R be convex and let N ∈ N. Then
f (ν) +
N
Xj=1
Aj(ν)∆jf (ν; 0, 1) ≤ (1 − ν)f (0) + νf (1).
(2.5)
MEANS REFINEMENTS VIA CONVEXITY
5
Proof. From Lemma 2.3 and convexity of f , we have
N
(1 − ν)f (0) + νf (1) −
Aj(ν)∆j f (ν; 0, 1)
Xj=1
= (cid:0)[2N ν] + 1 − 2N ν(cid:1) f(cid:18)[2N ν]
≥ f(cid:18)(cid:0)[2N ν] + 1 − 2N ν(cid:1)
2N (cid:19) +(cid:0)2N ν − [2N ν](cid:1) f(cid:18) [2N ν] + 1
(cid:19)
2N +(cid:0)2N ν − [2N ν](cid:1)
[2N ν] + 1
[2N ν]
2N
2N
= f (ν).
This completes the proof.
(cid:19)
(cid:3)
Now our first main result in its general form can be stated as follows.
Theorem 2.5. Let f : [a, b] → R be convex. Then for each N ∈ N and 0 ≤ ν ≤ 1,
we have
f ((1 − ν)a + νb) +
N
Xj=1
Aj(ν)∆jf (ν; a, b) ≤ (1 − ν)f (a) + νf (b).
(2.6)
Proof. For the given f , define g : [0, 1] → R by g(x) = f ((1− x)a + xb). Then g is
convex on [0, 1]. Applying Corollary 2.4 on the function g implies the result. (cid:3)
Remark 2.6. We remark that a negative version of the above theorem has been
recently shown in [8]. Namely, it was proved
N
(1 + ν)f (a) − νf (b) +
(cid:19)
(2.7)
for the convex function f : R → R. However, the method of proof is considerably
easier than the above proofs and the applications are different.
(cid:17)
f (a) + f(cid:16) (2j−1−1)a+b
Xj=1
≤ f ((1 + ν)a − νb) , ν ≥ 0, a < b,
− f(cid:18)(2j − 1)a + b
2jν
2j−1
2j
2
Our next step is to prove a reversed version of (2.6).
Theorem 2.7. Let f : [a, b] → R be convex and let N ∈ N. Then for 0 ≤ ν ≤ 1
2 ,
N
we have
f ((1 − ν)a + νb) + (1 − A1(ν))∆1f (ν; a, b)
Xj=1
≥ (1 − ν)f (a) + νf (b) +
2 ≤ ν ≤ 1, we have
On the other hand, if 1
f ((1 − ν)a + νb) + (1 − A1(ν))∆1f (ν; a, b)
Xj=1
≥ (1 − ν)f (a) + νf (b) +
N
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
a + b
2
, b(cid:19) .
Aj(2 − 2ν)∆jf(cid:18)2 − 2ν; a,
a + b
2 (cid:19) .
6
M. SABABHEH
2 , we have
Proof. For 0 ≤ ν ≤ 1
= 2νf(cid:18) a + b
≥
N
f ((1 − ν)a + νb) + (1 − A1(ν))∆1f (ν; a, b) − ((1 − ν)f (a) + νf (b))
2 (cid:19) + (1 − 2ν)f (b) + f ((1 − ν)a + νb) − 2f(cid:18) a + b
2 (cid:19)
+ (1 − 2ν)b(cid:19)
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
, b(cid:19) + f(cid:18)2ν
a + b
a + b
2
2
N
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
Xj=1
2 (cid:19)
+f ((1 − ν)a + νb) − 2f(cid:18) a + b
Xj=1
2 (cid:19)
+f ((1 − ν)a + νb) − 2f(cid:18) a + b
Xj=1
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
a + b
a + b
2
2
N
=
≥
, b(cid:19) + f (νa + (1 − ν)b)
, b(cid:19) ,
where the last line follows from convexity of f , where one has
f (νa + (1 − ν)b) + f ((1 − ν)a + νb)
≥ 2f(cid:18) νa + (1 − ν)b + (1 − ν)a + νb
2 ≤ ν ≤ 1.
2
This completes the proof for 0 ≤ ν ≤ 1
inequality for 1
(cid:19) = 2f(cid:18) a + b
2 (cid:19) .
2 . Similar computations imply the desired
(cid:3)
In fact, the above reversed version turns out to be equivalent to convexity.
Proposition 2.8. Let f : I → R be a function defined on the interval I. Assume
that for all a < b in I and all 0 ≤ ν ≤ 1, we have
f ((1 − ν)a + νb) + (1 − A1(ν))∆1f (ν; a, b) ≥ (1 − ν)f (a) + νf (b),
(2.8)
then f is convex on I.
Proof. Observe that when 0 ≤ ν ≤ 1
f ((1 − ν)a + νb) + (1 − ν)(cid:18)f (a) + f (b) − 2f(cid:18) a + b
2 , (2.8) is equivalent to
or
On the other hand, if 1
f(cid:18) a + b
1
2 − 2ν
2 (cid:19) ≤
f ((1 − ν)a + νb) +
2 ≤ ν ≤ 1, (2.8) is equivalent to
2 (cid:19) ≤
2ν − 1
f (a) +
1
2ν
2ν
f(cid:18) a + b
2 (cid:19)(cid:19) ≥ (1 − ν)f (a) + νf (b),
1 − 2ν
2 − 2ν
f (b).
(2.9)
f ((1 − ν)a + νb) .
(2.10)
MEANS REFINEMENTS VIA CONVEXITY
7
Let x1 < x2 ∈ I and let 0 < λ < 1. We assert that f ((1 − λ)x1 + λx2) ≤
(1 − λ)f (x1) + λf (x2).
If 0 < λ ≤ 1
2 , let
ν =
1 − 2λ
2(1 − λ)
Then one can easily check that when 0 < λ ≤ 1
With these choices, we have
, a = (2 − 2λ)x1 + (2λ − 1)x2 and b = x2.
2 , we have 0 < ν ≤ 1
2 and a < b.
a + b
2
= (1 − λ)x1 + λx2 and (1 − ν)a + νb = x1.
Substituting these quantities in (2.9) implies f ((1− λ)x1 + λx2) ≤ (1− λ)f (x1) +
λf (x2). This proves the desired inequality for 0 < λ ≤ 1
2 .
Now if 1
2 ≤ λ < 1, let
ν =
1
2λ
, a = x1 and b = (1 − 2λ)x1 + 2λx2.
With these choices, we have 1
quantities in (2.10) implies the desired inequality for 1
the proof.
2 < ν ≤ 1 and a < b. Now substituting these
2 ≤ λ < 1. This completes
(cid:3)
As for the geometric meaning of these refinements, it turns out we are dealing
with the interpolation of the function f over the dyadic partition.
Proposition 2.9. Let f : [0, 1] → R be any function, and let N ∈ N. Then, if
νi = i
2N for some i = 0, 1,· · · , 2N , we have
f (νi) +
N
Xj=1
Aj(νi)∆jf (νi; 0, 1) = (1 − νi)f (0) + νif (1).
(2.11)
Proof. Observe that when νi = i
we have
2N , we have [2N νi] = 2N νi = i. From Lemma 2.3,
N
(1 − νi)f (0) + νif (1) −
Aj(νi)∆jf (νi; 0, 1)
Xj=1
= (cid:0)[2N νi] + 1 − 2N νi(cid:1) f(cid:18)[2N νi]
= f(cid:18) i
2N(cid:19) = f (νi).
2N (cid:19) +(cid:0)2N νi − [2N νi](cid:1) f(cid:18) [2N νi] + 1
2N
(cid:19)
This completes the proof.
Proposition 2.10. Let f : [0, 1] → R be a given function. If f is continuous,
then
(cid:3)
f (ν) + lim
N→∞
uniformly in ν ∈ [0, 1].
N
Xj=1
Aj(ν)∆jf (ν; 0, 1) = (1 − ν)f (0) + νf (1),
(2.12)
8
M. SABABHEH
Proof. Let N ∈ N and define the function
gN (ν) =
N
Xj=1
Aj(ν)∆jf (ν; 0, 1).
From Proposition 2.9, we have g(νi) = (1− νi)f (0) + νif (1)− f (νi), when νi = i
for some i = 1,· · · , 2N . Noting the definitions of Aj and ∆jf , one can easily see
that gN is linear on each dyadic interval Ii := (cid:2) i
2N(cid:3) , i = 0,· · · , 2N − 1. Now
since gN is linear on Ii and gN coincides with the continuous function h(ν) :=
(1− ν)f (0) + νf (1)− f (ν), it follows that gN is the linear interpolation of h at the
dyadic partition of [0, 1]. Since f is continuous, it follows that gN → h uniformly,
completing the proof.
2N , i+1
2N
(cid:3)
2.2. Log-Convex function. The proof of the following result follows from The-
orem 2.5 on replacing f by log f.
Corollary 2.11. Let f : [a, b] → (0,∞) be log-convex. Then for 0 ≤ ν ≤ 1 and
N ∈ N, we have
f ((1 − ν)a + νb)
N
Yj=1(cid:18) f (xj(ν))f (yj(ν))
f 2(zj(ν))
(cid:19)Aj (ν)
≤ f 1−ν(a)f ν(b),
(2.13)
where xj(ν), yj(ν) and zj(ν) are as in the proof of Lemma 2.1.
On the other hand, applying Theorem 2.7 implies the following.
Corollary 2.12. Let f : [a, b] → (0,∞) be log-convex. Then for 0 ≤ ν ≤ 1
N ∈ N, we have
f ((1 − ν)a + νb) f (a)f (b)
2 ) !1−A1(ν)
≥ f 1−ν(a)f ν(b)
Yj=1(cid:18) f (tj(ν))f (uj(ν))
f 2(wj(ν))
(2.14)
where tj(ν), uj(ν) and wj(ν) are obtained from the above xj(ν), yj(ν) and zj(ν)
f 2( a+b
2 and
(cid:19)Aj (1−2ν)
N
On the other hand, if 1
on replacing (ν, a, b) by (cid:0)1 − 2ν, a+b
2 , b(cid:1) .
2 ≤ ν ≤ 1, we have
2 ) !1−A1(ν)
f ((1 − ν)a + νb) f (a)f (b)
≥ f 1−ν(a)f ν(b)
f 2( a+b
(2.15)
where tj(ν), uj(ν) and wj(ν) are obtained from the above xj(ν), yj(ν) and zj(ν)
on replacing (ν, a, b) by (cid:0)2 − 2ν, a, a+b
2 (cid:1) .
The following is a squared additive version for log-convex functions. This
inequality will help prove some squared versions of certain means.
N
Yj=1(cid:18) f (tj(ν))f (uj(ν))
f 2(wj(ν))
(cid:19)Aj (2−2ν)
,
,
MEANS REFINEMENTS VIA CONVEXITY
9
Theorem 2.13. Let f : [a, b] → [0,∞) be log-convex. Then for 0 ≤ ν ≤ 1 and
N ≥ 2, we have
N
f 2((1 − ν)a + νb) + A2
1(ν)∆1f 2(ν; a, b) +
Xj=2
≤ ((1 − ν)f (a) + νf (b))2 .
Aj(ν)∆j f 2(ν; a, b)
Proof. We prove the result for 0 ≤ ν ≤ 1
g = f 2 is log-convex too, and hence is convex. Therefore, Theorem 2.5 implies
2 . Since f is log-convex, it follows that
N
g((1 − ν)a + νb) +
Xj=1
which implies, for 0 ≤ ν ≤ 1
2 ,
Aj(ν)∆jg(ν; a, b) ≤ (1 − ν)g(a) + νg(b),
N
f 2((1 − ν)a + νb) + ν2∆1f 2(ν; a, b) +
Aj(ν)∆jf 2(ν; a, b)
Xj=2
≤ ((1 − ν)f (a) + νf (b))2 + H(ν; a, b),
(2.16)
where
H(ν; a, b) = (1 − ν)f 2(a) + νf 2(b) + ν2∆1f 2(ν; a, b) − ν∆1f 2(ν; a, b)
− ((1 − ν)f (a) + νf (b))2
= 2ν(1 − ν)(cid:18)f 2(cid:18) a + b
≤ 0,
2 (cid:19) − f (a)f (b)(cid:19)
where the last inequality follows from log-convexity of f . Since H(ν; a, b) ≤ 0, it
follows from (2.16) that
f 2((1 − ν)a + νb) + ν2∆1f 2(ν; a, b) +
Xj=2
≤ ((1 − ν)f (a) + νf (b))2 .
Similar computations imply the result for 1
2 ≤ ν ≤ 1.
N
Aj(ν)∆jf 2(ν; a, b)
(cid:3)
Then reversed squared versions maybe obtained in a similar way from Theorem
we have
2.7 as follows.
Theorem 2.14. Let f : [a, b] → [0,∞) be log-convex and N ∈ N. If 0 ≤ ν ≤ 1
2 ,
f 2((1 − ν)a + νb) + (1 − ν)2∆1f 2(ν; a, b) + 2ν(1 − ν)(cid:18)f (a)f (b) − f 2(cid:18) a + b
2 (cid:19)(cid:19)
≥ ((1 − ν)f (a) + νf (b))2
N
+
Xj=1
Aj(1 − 2ν)∆jf 2(cid:18)1 − 2ν;
a + b
2
, b(cid:19) .
10
M. SABABHEH
If 1
2 ≤ ν ≤ 1, we have
f 2((1 − ν)a + νb) + ν2∆1f 2(ν; a, b) + 2ν(1 − ν)(cid:18)f (a)f (b) − f 2(cid:18) a + b
2 (cid:19)(cid:19)
≥ ((1 − ν)f (a) + νf (b))2
N
+
Xj=1
Aj(2 − 2ν)∆jf 2(cid:18)2 − 2ν; a,
a + b
2 (cid:19) .
3. Application
3.1. Refinements of means inequalities. In this section we present some in-
teresting applications of the above inequalities. The first result is the following
refinement of Young's inequality.
Corollary 3.1. Let x, y > 0, N ∈ N and 0 ≤ ν ≤ 1. Then
x#ν y +
N
Xj=1
Aj(ν)(cid:18) 2jqykj(ν)x2j−1−kj (ν) − 2jqykj(ν)+1x2j−1−kj(ν)−1(cid:19)2
≤ x∇νy.
(3.1)
Proof. This follows from Theorem 2.5, on letting f (t) = x1−tyt, a = 0, b = 1.
Then f is convex. Moreover, direct computations show that
∆jf (ν; 0, 1) =(cid:18) 2jqykj(ν)x2j−1−kj(ν) − 2jqykj(ν)+1x2j−1−kj (ν)−1(cid:19)2
.
(cid:3)
The above theorem has been recently proved in [?] as a refinement of Young's
inequality. This inequality refines the corresponding refinements appearing in [5]
and [11], where the inequality was proved only for N = 1, 2.
On the other hand, letting f (t) = x!ty, the weighted harmonic mean, we obtain
the following refinement of the arithmetic-harmonic mean inequality.
Corollary 3.2. Let x, y > 0, N ∈ N and 0 ≤ ν ≤ 1. Then
N
x!ν y +
Aj(ν)(cid:0)x!αj (ν)y + x!βj(ν)y − 2x!γj (ν)y(cid:1) ≤ x∇ν y,
Xj=1
2j−1 , βj(ν) = [2j−1ν]+1
2j−1
2
and γj(ν) = αj (ν)+βj (ν)
.
where αj(ν) = [2j−1ν]
(3.2)
This inequality is a significant refinement of the corresponding inequality in
[12], where the inequality was proved only for N = 1.
Now noting log-convexity of the function t 7→ x!ty on [0, 1], and applying
Corollary 2.11, we get the following multiplicative refinement of the geometric-
harmonic mean inequality.
Corollary 3.3. Let x, y > 0, N ∈ N and 0 ≤ ν ≤ 1, we have
≤ x#ν y,
x!ν y
N
Yj=1(cid:18)(x!αj (ν)y)(x!βj(ν)y)
(x!γj (ν)y)2
(cid:19)Aj(ν)
MEANS REFINEMENTS VIA CONVEXITY
11
where αj(ν) = [2j−1ν]
2j−1 , βj(ν) = [2j−1ν]+1
2j−1
and γj(ν) = αj (ν)+βj (ν)
2
.
When N = 1, Corollary 3.3, reduces to
and
x#y(cid:19)2ν
(x!ν y)(cid:18)x∇y
x#y(cid:19)2(1−ν)
(x!ν y)(cid:18)x∇y
≤ x#ν y, 0 ≤ ν ≤
1
2
≤ x#ν y,
1
2 ≤ ν ≤ 1.
x#y(cid:17)2
The constant (cid:16) x∇y
appearing in these inequalities is called the Kantorovich
constant, and has appeared in recent refinements of these mean inequalities. One
can see [6] as a recent reference treating some inequalities using this constant.
As for the squared version, applying Theorem 2.13 to the log-convex functions
t 7→ x#ty and t 7→ x!ty implies the following. The first inequality refines the
corresponding results in [3] and [11], while the other inequality is new.
Corollary 3.4. Let x, y > 0, 0 ≤ ν ≤ 1, N ≥ 2 and
αj(ν) =
Then
kj(ν)
2j−1
, βj(ν) =
and γj(ν) =
kj(ν) + 1
2j−1
αj(ν) + βj(ν)
.
2
(x#ν y)2 + A2
1(ν)(x − y)2 +
≤ (x∇ν y)2,
and
N
Xj=2
Aj(ν)(cid:0)x1−αj (ν)yαj(ν) − x1−βj (ν)yβj(ν)(cid:1)2
(x!ν y)2 + 2A2
1(ν)(x2∇y2 − (x!y)2) +
≤ (x∇νy)2.
N
Xj=2
Aj(ν)(cid:0)(x!αj (ν)y)2 + (x!βj (ν)y)2 − 2(x!γj (ν)y)2(cid:1)
have
3.2. Reversed Version. Applying Theorem 2.7 to the function f (t) = x#ty
implies the following reversed version of Young's inequality.
Corollary 3.5. For x, y > 0, let f (t) = x#ty and let N ∈ N. If 0 ≤ ν ≤ 1
x#ν y + (1 − ν)(√x − √y)2 ≥ x∇ν y +
, 1(cid:19) .
Xj=1
2 ≤ ν ≤ 1, we have
Xj=1
x#ν y + ν(√x − √y)2 ≥ x∇ν y +
Aj(2 − 2ν)∆jf(cid:18)2 − 2ν; 0,
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
On the other hand, if 1
1
2(cid:19) .
2 , we
1
2
N
N
These inequalities refine those in [5] and [11]. Then an arithmetic-harmonic
reversed version maybe obtained by applying Theorem 2.7 to the function f (t) =
x!ty as follows.
12
M. SABABHEH
Xj=1
2 ≤ ν ≤ 1, we have
Xj=1
N
have
Corollary 3.6. For x, y > 0, let f (t) = x#ty and let N ∈ N. If 0 ≤ ν ≤ 1
, 1(cid:19) .
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
x!ν y + (1 − ν) (x + y − 2x!y) ≥ x∇ν y +
1
2
N
2 , we
On the other hand, if 1
x!νy + ν(x + y − 2x!y) ≥ x∇νy +
Aj(2 − 2ν)∆jf(cid:18)2 − 2ν; 0,
1
2(cid:19) .
These inequalities refine those in [6].
Similarly, noting log-convexity of the function f (t) = x!ty, we may apply Corol-
lary 2.12 to obtain reversed multiplicative version of the harmonic-geometric mean
inequality. We leave the application to the reader.
Following the same guideline, we may obtain reversed squared versions by
applying Theorem 2.14 to the functions t 7→ x#ty and x 7→ x!ty. Observe that
when f (t) = x#ty we have f (a)f (b)−f 2(cid:0) a+b
2 (cid:1) = 0. Therefore, applying Theorem
2.14 implies the following inequalities, which refine the corresponding inequalities
in [3] and [11].
Corollary 3.7. Let x, y > 0 and N ≥ 1. If 0 ≤ ν ≤ 1, we have
(x#ν y)2 + (1 − ν)2(x − y)2 ≥ (x∇ν y)2 +
If 1
Aj(1 − 2ν)∆jf 2(cid:18)1 − 2ν;
, 1(cid:19) .
Xj=1
1
2
N
2 ≤ ν ≤ 1, we have
(x#ν y)2 + ν2(x − y)2 ≥ (x∇ν y)2 +
N
Xj=1
Aj(2 − 2ν)∆jf 2(cid:18)1 − 2ν; 0,
1
2(cid:19) .
Now letting g(t) = x!ty we obtain the following new inequalities for the arithmetic-
harmonic means.
Corollary 3.8. Let x, y > 0 and N ≥ 1. If 0 ≤ ν ≤ 1, we have
(x!ν y)2 + 2(1 − ν)2(x2∇y2 − (x!y)2) + 2ν(1 − ν) xy −(cid:18) 2xy
x + y(cid:19)2!
≥ (x∇ν y)2 +
N
Xj=1
Aj(1 − 2ν)∆jg2(cid:18)1 − 2ν;
1
2
, 1(cid:19) .
If 1
2 ≤ ν ≤ 1, we have
(x!νy)2 + 2ν2(x2∇y2 − (x!y)2) + 2ν(1 − ν) xy −(cid:18) 2xy
x + y(cid:19)2!
2(cid:19) .
Aj(2 − 2ν)∆jg2(cid:18)1 − 2ν; 0,
≥ (x∇ν y)2 +
1
N
Xj=1
MEANS REFINEMENTS VIA CONVEXITY
13
3.3. Some Lp inequalities. Let (X,M, µ) be a measure space, and let 0 < p <
q < r. Then Lp ∩ Lr ⊂ Lq and
pkfk1−ν
, where f ∈ Lp ∩ Lr and ν =
kfkq ≤ kfkν
.
r
q−1 − r−1
p−1 − r−1
This inequality can be modified using Corollary 2.11 and a reversed version can
be obtained using Corollary 2.12.
Proposition 3.9. Let (X,M, µ) be a measure space, 0 < p < q < r and ν be as
above. If f ∈ Lp ∩ Lr and N ∈ N, then we have
Proof. It is easy to check that the function h(t) = kfk1/t is log-convex on [r−1, p−1].
Then direct application of Corollary 2.11 implies the result.
(ν)kfky−1
pkfk1−ν
≤ kfkν
kfkx−1
Yj=1
kfk2
kfkq
Aj (ν)
z−1
j
(ν)
(ν)
(cid:3)
N
.
r
j
j
In particular, when N = 1, the above proposition implies
kfkq ≤( kfk1−2ν
kfk2ν−1
2pr
p+r
kfk2ν
kfk2−2ν
2pr
p+r
p
r
,
,
2
0 ≤ ν ≤ 1
2 ≤ ν ≤ 1 ) ≤ kfkν
1
pkfk1−ν
r
.
2 can be interpreted as 2pr
p+r ≤ q ≤ r, while 1
2 ≤ ν ≤ 1
The condition 0 ≤ ν ≤ 1
means p ≤ q ≤ 2pr
p+r .
Moreover, a reversed version maybe obtained using Corollary 2.12.
Proposition 3.10. Let (X,M, µ) be a measure space, 0 < p < q < r and
ν = q−1−r−1
p−1−r−1 . If 2pr
p+r ≤ q ≤ r, f ∈ Lp ∩ Lr and N ∈ N, then
Aj (1−2ν)
where tj, uj and zj are obtained from xj, yj and zj by replacing (ν, a, b) with
kfkq ≥ kfk2−2ν
2pr
j
j
j
p
N
(ν)
(ν)
w−1
kfk2
Yj=1
kfkt−1
(ν)kfku−1
p+r kfk2ν−1
2pr , p−1(cid:17) . On the other hand, if p ≤ q ≤ 2pr
p+rkfk1−2ν
(ν)kfku−1
kfkt−1
Yj=1
kfk2
w−1
(ν)
(ν)
2pr
N
r
j
j
j
(cid:16)1 − 2ν, p+r
kfkq ≥ kfk2ν
p+r , then
Aj (1−2ν)
≥ kfk2−2ν
p+r kfk2ν−1
2pr
p
,
≥ kfk2ν
p+rkfk1−2ν
2pr
r
,
where tj, uj and zj are obtained from xj, yj and zj by replacing (ν, a, b) with
2pr(cid:17) .
(cid:16)2 − 2ν, r−1, p+r
Propositions 3.9 and 3.10 have been obtained using log-convexity of the func-
tion h(t) = kfkt−1. In fact, noting log-convexity of the function h(t) = kfkt
t, we
obtain the same results! This is due to the equivalence of log-convexity of the
functions t 7→ kfkt−1 and that of kfkt
t. We refer the reader to [9] where these
relations between the different log-convex function criteria have been discussed.
The celebrated three lines lemma of Hadamard states the following.
14
M. SABABHEH
Lemma 3.11. Let D = {z ∈ C : 0 ≤ ℜz ≤ 1} and let ϕ : D → C be continuous
on D and analytic in the interior of D. Then the function f : [0, 1] → R defined
by f (x) = supy ϕ(x + iy) is log-convex.
This lemma is an extremely useful tool in the theory of complex functions. In
particular, this lemma becomes handy in proving different interpolation versions
of bounded linear operators between Lp spaces.
Log-convexity implied by Lemma 3.11 allows us to apply our refined and re-
versed versions for log-convex functions. In the following proposition, we present
one term refinement and reverse.
1
and
Proposition 3.12. Let D = {z ∈ C : 0 ≤ ℜz ≤ 1} and let ϕ : D → C be
continuous on D and analytic in the interior of D. Then the function f : [0, 1] →
R defined by f (x) = supy ϕ(x + iy) satisfies the following
f (x) ≤(cid:26) f 1−2x(0)f 2x(cid:0) 1
0 ≤ x ≤ 1
2(cid:1) ,
f 2x−1(1)f 2−2x(cid:0) 1
2(cid:1) ,
2 ≤ x ≤ 1
2(cid:1) , 0 ≤ x ≤ 1
f (x) ≥(cid:26) f 2x−1(1)f 2−2x(cid:0) 1
f 1−2x(0)f 2x(cid:0) 1
2(cid:1) ,
2 ≤ x ≤ 1
3.4. Operator versions. The following theorem provides a refinement of the
well known Heinz inequality and its reverse. The proof follows immediately noting
convexity of the Heinz means, see [1].
Theorem 3.13. For A, B ∈ M+
norm k k, let
n , X ∈ Mn, 0 ≤ ν ≤ 1 and any unitarily invariant
.
2
2
1
Then we have the following refinement of Heinz inequality
f (ν) = kAνXB1−ν + A1−ν XBνk.
kAνXB1−ν + A1−νXBνk +
N
Xj=1
Aj(ν)∆j f (ν; 0, 1) ≤ kAX + XBk.
2 , we have
Moreover, if 0 ≤ ν ≤ 1
kAνXB1−ν + A1−ν XBνk + 2(1 − ν)(cid:16)kAX + XBk − k
N
≥ kAX + XBk +
On the other hand, if 1
2 ≤ ν ≤ 1, we have
Xj=1
kAνXB1−ν + A1−ν XBνk + 2ν(cid:16)kAX + XBk − k
Xj=1
≥ kAX + XBk +
N
√AX√Bk(cid:17)
Aj(1 − 2ν)∆jf(cid:18)1 − 2ν;
1
2
, 1(cid:19) .
√AX√Bk(cid:17)
Aj(2 − 2ν)∆jf(cid:18)2 − 2ν; 0,
1
2(cid:19) .
MEANS REFINEMENTS VIA CONVEXITY
15
In [9], it is shown that for A, B ∈ M+
n and X ∈ Mn, the functions
t → kAtXB1−tk, t → kAtXB1−tk kA1−tX Btk, t → tr(AtXB1−tX∗)
are log-convex on [0, 1]. Therefore, we may apply Corollaries 2.11 and 2.12 to
obtain refinements and reversed versions for such functions.
For the k k2 norm, we can prove log convexity of the Heinz means, which allows
us to obtain further refinements of the Heinz inequality by applying Corollaries
2.11 and 2.12.
Proposition 3.14. Let A, B ∈ M+
A1−νXBνk2. Then f is log-convex on [0, 1].
Proof. Since A, B ∈ M+
n , there are diagonal matrices D1 := diag(λi), D2 :=
diag(µi) and unitarily matrices U, V such that λi, µi ≥ 0, A = U D1U∗ and
B = V D2V ∗. Letting Y = U∗XV, we have
Aν XB1−ν + A1−νXBν = U(λν
n and X ∈ Mn, and define f (ν) = kAνXB1−ν +
yijµν
j + λ1−ν
i
j )V ∗.
i yijµ1−ν
Since k k2 is a unitarily invariant norm, we have
j + λ1−ν
j + λ1−ν
i µν
i yijµ1−ν
i µ1−ν
i
f 2(ν) = kU(λν
= Xi,j (cid:0)λν
2
yijµν
j )V ∗k2
j(cid:1)2 yij2.
Notice that each summand is log-convex, being the square of a log-convex func-
tion. This implies that f 2 is log-convex. Consequently, f is log-convex.
(cid:3)
Letting f (ν) = kAνXB1−ν + A1−νXBνk2 and applying Theorem 2.13 imply
the following squared version of Heinz inequality.
Corollary 3.15. Let A, B ∈ M+
kAνXB1−ν + A1−νXBνk2
n , X ∈ Mn, 0 ≤ ν ≤ 1 and N ≥ 2. Then
2 + 2A2
2 XB
2 − 2kA
1
1(ν)(cid:16)kAX + XBk2
1
2(cid:17)
2k2
+
N
Xj=2
Aj(ν)∆jf 2(ν; 0, 1) ≤ kAX + XBk2
2.
We leave the application of Corollaries 2.11 and 2.12 to the reader.
Further operator versions maybe obtained using Lemma 1.1. The following
operator versions refine the corresponding results in [5] and [11].
Proposition 3.16. Let A, B ∈ M++
n and 0 ≤ ν ≤ 1. Then for αj(ν) = kj(ν)
2j−1 , βj(ν) =
kj(ν)+1
2j−1
, γj(ν) = αj(ν)+βj (ν)
2
and N ∈ N, we have
A#νB +
N
Xj=1
Aj(ν)(cid:0)A#αj (ν) + A#βj(ν) − 2A#γj (ν)B(cid:1) ≤ A∇νB.
Proof. In Corollary 3.1, let x = 1, expand the summand and apply Lemma 1.1
with y replaced by X = A− 1
2 . Then the result follows upon conjugating
both sides with A
(cid:3)
2 BA− 1
2 .
1
16
M. SABABHEH
In a similar way one may obtain reversed versions by applying Corollary 3.5.
This provides refinements of the reversed versions of [11]. The following is an
operator arithmetic-harmonic version, refining the corresponding results in [12].
Proposition 3.17. Let A, B ∈ M++
n and 0 ≤ ν ≤ 1. Then for αj(ν) = kj(ν)
2j−1 , βj(ν) =
kj(ν)+1
2j−1
, γj(ν) = αj(ν)+βj (ν)
2
and N ∈ N, we have
A!νB +
N
Xj=1
Aj(ν)(cid:0)A!αj (ν) + A!βj(ν) − 2A!γj (ν)B(cid:1) ≤ A∇νB.
The proof follows immediately on applying Lemma 1.1 together with Corollary
3.2. On the other hand, applying Corollary 3.6 implies the following refinement
of the corresponding inequalities in [6].
Proposition 3.18. Let A, B ∈ M++
and N ∈ N. If 0 ≤ ν ≤ 1, we have
n
A!νB + (1 − ν)(A + B − 2A!B)
N
Xj=1
≥ A∇νB +
2(cid:16)1 − kj(1−2ν)
2j−1 (cid:17) + kj(1−2ν)
where αj(ν) = 1
and γj(ν) = αj (ν)+βj (ν)
.
On the other hand, if 1
2 ≤ ν ≤ 1, we have
2j−1
2
Aj(1 − 2ν)(cid:0)A!αj (ν)B + A!βj(ν)B − 2A!γj (ν)B(cid:1) ,
, βj(ν) = 1
2(cid:16)1 − kj(1−2ν)+1
2j−1
(cid:17) + kj (1−2ν)+1
2j−1
A!νB + ν(A + B − 2A!B)
N
≥ A∇νB +
Xj=1
Aj(2 − 2ν)(cid:0)A!αj (ν)B + A!βj(ν)B − 2A!γj (ν)B(cid:1) ,
where αj(ν) = kj (2−2ν)
2j−1
, βj(ν) = kj(2−2ν)+1
2j−1
and γj(ν) = αj(ν)+βj (ν)
2
.
The following is an interesting one-term multiplicative refinement of the oper-
ator geometric-harmonic mean inequality.
Theorem 3.19. Let A, B ∈ M++
n
and 0 ≤ ν ≤ 1. Then
(A!νB)(cid:18) A−1B + 2I + B−1A
4
(cid:19)r
≤ A#ν B,
where r = min{ν, 1 − ν}.
Proof. We prove the desired inequality for 0 ≤ ν ≤ 1
2√y(cid:17)2ν
and x = 1, to get (1!νy)(cid:16) 1+y
≤ 1#νy, or
2 . In Corollary 3.3, let N = 1
1
4ν (cid:0)(1 − ν) + νy−1(cid:1)−1(cid:0)y + 2 + y−1(cid:1)ν ≤ yν.
(3.3)
Let X = A− 1
2 BA− 1
1
1
=
= A− 1
1
1
2 B−1A
2 BA− 1
2 (A!νB)A− 1
2(cid:17)−1(cid:16)A− 1
4ν (cid:16)(1 − ν)I + νA
4ν hA− 1
2 (cid:0)A−1B + 2I + B−1A(cid:1)ν
(cid:19)ν
2 (A!ν B)(cid:18) A−1B + 2I + B−1A
2ihA
A− 1
2 .
4
1
A− 1
2i
2 + 2I + A
1
2 B−1A
1
2(cid:17)ν
(3.4)
MEANS REFINEMENTS VIA CONVEXITY
17
2 and apply Lemma 1.1. The left hand side of (3.3) becomes
On the other hand, the right hand side of (3.3) is simply (cid:16)A− 1
. This
together with (3.4) imply the desired inequality, upon conjugating both sides with
A
(cid:3)
2 . This completes the proof.
2 BA− 1
1
2(cid:17)ν
References
[1] R. Bhatia, Matrix analysis, Springer-Verlag, New York, 1997.
[2] T. Furuta, J. Micic Hot and J. Pecaric, Mond-Pecaric Method in Operator Inequalities,
Element, Zagreb, 2005.
[3] O. Hirzallah and F. Kittaneh, Matrix Young inequalities for the Hilbert-Schmidt norm,
Linear Algebra Appl. 308 (2000), 77 -- 84.
[4] F. Kittaneh, On the convexity of the Heinz mean, Integr. Equ. Oper. Theory 68 (2010),
519 -- 527.
[5] F. Kittaneh and Y. Manasrah, Improved Young and Heinz inequalities for matrices, J. Math.
Anal. Appl. 36 (2010), 262 -- 269.
[6] W. Liao and J. Wu, Reverse arithmetic-harmonic mean and mixed mean operator inequali-
ties, J. Inequal. Appl., 2015:215.
[7] Y. Manasrah and F. Kittaneh, A generalization of two refined Young inequalities, Positivity
19 (2015), 757 -- 768.
[8] M. Sababheh, Convex functions and means of matrices, Math. Ineq. Appl., accepted.
[9] M. Sababheh, Log and harmonically log-convex functions related to matrix norms, Operators
and Matrices, in press.
[10] M. Sababheh, Integral inequalities of the Heinz means as convex functions, J. Math. Ineq.,
10 (2) (2016), 313 -- 325.
[11] J. Zhao and J. Wu, Operator inequalities involving improved Young and its reverse inequal-
ities, J. Math. Anal. Appl. 421 (2015) 1779 -- 1789.
[12] H. Zuo, G. Shi and M. Fujii, Refined Young inequality with Kantorovich constant, J. Math.
Inequal. 5(4)(2011), 551-556 .
Department of Basic Sciences, Princess Sumaya University For Technology,
Al Jubaiha, Amman 11941, Jordan.
E-mail address: [email protected], [email protected]
|
1708.06885 | 1 | 1708 | 2017-08-23T05:05:24 | Completeness of Unbounded Convergences | [
"math.FA"
] | As a generalization of almost everywhere convergence to vector lattices, unbounded order convergence has garnered much attention. The concept of boundedly uo-complete Banach lattices was introduced by N. Gao and F. Xanthos, and has been studied in recent papers by D. Leung, V.G. Troitsky, and the aforementioned authors. We will prove that a Banach lattice is boundedly uo-complete iff it is monotonically complete. Afterwards, we study completeness-type properties of minimal topologies; minimal topologies are exactly the Hausdorff locally solid topologies in which uo-convergence implies topological convergence. | math.FA | math |
COMPLETENESS OF UNBOUNDED CONVERGENCES
M.A. TAYLOR
Abstract. As a generalization of almost everywhere convergence
to vector lattices, unbounded order convergence has garnered much
attention. The concept of boundedly uo-complete Banach lattices
was introduced by N. Gao and F. Xanthos, and has been studied
in recent papers by D. Leung, V.G. Troitsky, and the aforemen-
tioned authors. We will prove that a Banach lattice is bound-
edly uo-complete iff it is monotonically complete. Afterwards, we
study completeness-type properties of minimal topologies; mini-
mal topologies are exactly the Hausdorff locally solid topologies in
which uo-convergence implies topological convergence.
1. Introduction
In the first half of the paper, we study when norm bounded uo-
Cauchy nets in a Banach lattice are uo-convergent. The section starts
with a counterexample to a question posed in [LC], and culminates in a
proof that a Banach lattice is (sequentially) boundedly uo-complete iff
it is (sequentially) monotonically complete. This gives the final solution
to a problem that has been investigated in [Gao14], [GX14], [GTX17],
and [GLX].
The latter half of this paper focuses on the "extremal" topologies
of a vector lattice X. For motivation, recall that corresponding to a
dual pair hE, E∗i is a family of topologies on E "compatible" with
duality. The two most important elements of this family are the weak
and Mackey topologies, which are defined by their extremal nature.
Analogously, given a vector lattice X, it is often possible to equip X
Date: October 3, 2018.
2010 Mathematics Subject Classification. 46A40, 46A16, 46B42.
Key words and phrases. uo-convergence, unbounded topology, minimal topology,
completeness, boundedly uo-complete, monotonically complete, Levi.
The author acknowledges support from NSERC and the University of Alberta.
1
2
M.A. TAYLOR
with many topologies compatible (in the sense of being locally solid and
Hausdorff) with the lattice structure. It is easy to see that whenever X
admits some Hausdorff locally solid topology, the collection of all Riesz
pseudonorms on X generates a finest Hausdorff locally solid topology
on X. This "greatest" topology appears in many applications. Indeed,
analogous to the theory of compatible locally convex topologies on a
Banach space - where the norm topology is the Mackey topology -
the norm topology on a Banach lattice X is the finest topology on X
compatible with the lattice structure. This is [AB03, Theorem 5.20].
On the opposite end of the spectrum, a Hausdorff locally solid topol-
ogy on a vector lattice X is said to be minimal if there is no coarser
Hausdorff locally solid topology on X; it is least if it is coarser than
every Hausdorff locally solid topology on X. Least topologies were
introduced in [AB80] and studied in [AB03]; minimal topologies were
studied in [Lab87], [Con05], [Tay], and [KT]. An important example
of a least topology is the unbounded norm topology on an order con-
tinuous Banach lattice. The unbounded absolute weak∗-topology on
L∞[0, 1] is a noteworthy example of a minimal topology that is not
least. In the next subsection, we briefly recall some facts about mini-
mal and unbounded topologies; for a detailed exposition the reader is
referred to [Tay] and [KT].
1.1. Notation. Throughout this paper, all vector lattices are assumed
o−→ x if
Archimedean. For a net (xα) in a vector lattice X, we write xα
(xα) converges to x in order ; that is, there is a net (yβ), possibly
over a different index set, such that yβ ↓ 0 and for every β there exists
uo−→ x
α0 such that xα − x ≤ yβ whenever α ≥ α0. We write xα
o−→ 0 for
and say that (xα) uo-converges to x ∈ X if xα − x ∧ u
every u ∈ X+. For facts on uo-convergence, the reader is referred to
[GTX17].
In particular, [GTX17, Theorem 3.2] will be used freely.
Recall that a Banach lattice X is (sequentially) boundedly uo-
complete if norm bounded uo-Cauchy nets (respectively, sequences)
in X are uo-convergent in X.
Given a locally solid topology τ on a vector lattice X, one can asso-
ciate a topology, uτ , in the following way. If {Ui}i∈I is a base at zero
UNBOUNDED CONVERGENCES
3
for τ consisting of solid sets, for each i ∈ I and u ∈ X+ define
Ui,u := {x ∈ X :
x ∧ u ∈ Ui}.
As was proven in [Tay, Theorem 2.3], the collection N0 = {Ui,u :
i ∈ I, u ∈ X+} is a base of neighbourhoods at zero for a new lo-
cally solid topology, denoted by uτ , and referred to as the unbounded
τ -topology . Noting that the map τ 7→ uτ from the set of locally solid
topologies on X to itself is idempotent, a locally solid topology τ is
called unbounded if there is a locally solid topology σ with τ = uσ
or, equivalently, if τ = uτ. The following connection between minimal
topologies, unbounded topologies, and uo-convergence was proven in
[Tay, Theorem 6.4]. Recall that a locally solid topology τ is Lebesgue
if order null nets are τ -null.
Theorem 1.1. Let τ be a Hausdorff locally solid topology on a vector
lattice X. TFAE:
(i) uo-null nets are τ -null;
(ii) τ is Lebesgue and unbounded;
(iii) τ is minimal.
In particular, a vector lattice can admit at most one minimal topology.
Interestingly, the process of unbounding a topology can convert the
greatest topology into the least topology; this happens with the norm
topology on an order continuous Banach lattice.
All other undefined terminology is consistent with [AB03]. In par-
ticular, we say that a locally solid topology τ on a vector lattice X is
Levi if τ -bounded increasing nets in X+ have supremum. Levi and
monotonically complete are synonymous; the latter terminology is
that of [MN91], and is used in [GLX].
2. Boundedly uo-complete Banach lattices
Results equating the class of boundedly uo-complete Banach lattices
to the class of monotonically complete Banach lattices have been ac-
quired, under technical assumptions, by N. Gao, D. Leung, V.G. Troit-
sky, and F. Xanthos. The sharpest result is [GLX, Proposition 3.1];
it states that a Banach lattice whose order continuous dual separates
4
M.A. TAYLOR
points is boundedly uo-complete iff it is monotonically complete. In
this section, we remove the restriction on the order continuous dual.
The following question was posed as Problem 2.4 in [LC]:
Question 2.1. Let (xα) be a norm bounded positive increasing net in
a Banach lattice X. Is (xα) uo-Cauchy in X?
If Question 2.1 is true, it is easily deduced that a Banach lattice is
boundedly uo-complete iff it is monotonically complete. However, the
next example answer this question in the negative, even for sequences.
Example 2.2. Let S be the set of all non-empty finite sequences of
natural numbers. For s ∈ S define λ(s) = length(s). If s, t ∈ S, define
s ≤ t if λ(s) ≤ λ(t) and s(i) = t(i) for i = 1, . . . , λ(s). For s ∈ S with
λ(s) = n and i ∈ N, define s ∗ i = (s(1), . . . , s(n), i). Put
(2.1)
X = {x ∈ ℓ∞(S) : lim
i→∞
x(s ∗ i) =
1
2
x(s) for all s ∈ S}.
It can be verified that X is a closed sublattice of (ℓ∞(S), k · k∞) and
for t ∈ S the element et : S → R defined by
et(s) =
( 1
2)λ(s)−λ(t)
0
if t ≤ s
otherwise
is an element of X with norm 1. Define f1 = e(1), f2 = e(1) ∨ e(2) ∨
e(1,1) ∨ e(1,2) ∨ e(2,1) ∨ e(2,2), and, generally,
fn = sup{et : λ(t) ≤ n and t(k) ≤ n ∀k ≤ λ(t)}.
The sequence (fn) is increasing and norm bounded by 1; it was shown
in [BL88, Example 1.8] that (fn) is not order bounded in X u. There-
fore, (fn) cannot be uo-Cauchy in X for if it were then it would be
uo-Cauchy in X u and hence order convergent in X u by [GTX17, The-
orem 3.10]. Since it is increasing, it would have supremum in X u; this
is a contradiction as (fn) is not order bounded in X u.
Under some mild assumptions, however, Question 2.1 has a positive
solution. Recall that a Banach lattice is weakly Fatou if there exists
K ≥ 1 such that whenever 0 ≤ xα ↑ x, we have kxk ≤ K sup kxαk.
UNBOUNDED CONVERGENCES
5
Proposition 2.3. Let X be a weakly Fatou Banach lattice. Then every
positive increasing norm bounded net in X is uo-Cauchy.
If 0 ≤ ( 1
Proof. Let K be such that 0 ≤ xα ↑ x implies kxk ≤ K sup kxαk. Now
assume that 0 ≤ uα ↑ and kuαk ≤ 1. Let u > 0 and pick n such that
kuk > K
n . Therefore, there
n .
exists 0 < w ∈ X such that ( 1
n uα) ∧ u ≤ u − w for all α. But then
n uα) ∧ u ↑α u, then kuk ≤ K
n uα) ∧ u(cid:3) ≥ nw > 0 for all α, so that (uα) is
(nu − uα)+ = n(cid:2)u − ( 1
and hence uα ↑ bu for some bu ∈ X u. This proves that (uα) is uo-Cauchy
dominable. By [AB03, Theorem 7.37], (uα) is order bounded in X u,
in X u, hence in X.
(cid:3)
Proposition 2.4. Let X be a weakly σ-Fatou Banach lattice. Then
every positive increasing norm bounded sequence in X is uo-Cauchy.
Proof. The proof is similar and, therefore, omitted.
(cid:3)
Even though Question 2.1 is false, the equivalence between bound-
edly uo-complete and Levi still stands. We show that now:
Theorem 2.5. Let X be a Banach lattice. TFAE:
(i) X is σ-Levi;
(ii) X is sequentially boundedly uo-complete;
(iii) Every increasing norm bounded uo-Cauchy sequence in X+ has
a supremum.
xn
n=1
1
2n
parts. Define e = P∞
Proof. (i)⇒(ii): Let (xn) be a norm bounded uo-Cauchy sequence in
X. WLOG, (xn) is positive; otherwise consider positive and negative
1+kxnk and consider Be, the band generated
by e. Then (xn) is still norm bounded and uo-Cauchy in Be. Also, Be
has the σ-Levi property for if 0 ≤ yn ↑ is a norm bounded sequence in
Be, then yn ↑ y for some y ∈ X as X is σ-Levi. Since Be is a band,
y ∈ Be and yn ↑ y in Be. We next show that there exists u ∈ Be such
that xn
uo−→ u in Be, and hence in X.
For each m, n, n′ ∈ N, since xn ∧ me − xn′ ∧ me ≤ xn − xn′ ∧ me,
the sequence (xn ∧me)n is order Cauchy, hence order converges to some
um in Be since the σ-Levi property implies σ-order completeness. The
6
M.A. TAYLOR
sequence (um) is increasing and
kumk ≤ K sup
n
kxn ∧ mek ≤ K sup
n
kxnk < ∞
where we use that σ-Levi implies weakly σ-Fatou. Since Be is σ-Levi,
(um) increases to an element u ∈ Be. Fix m. For any N, N ′ define
xN,N ′ = supn≥N,n′≥N ′ xn − xn′ ∧ e. Since (xn) is uo-Cauchy, xN,N ′ ↓ 0.
Now, for each m,
xn ∧ me − xn′ ∧ me ∧ e ≤ xn − xn′ ∧ e ≤ xN,N ′ ∀n ≥ N, n′ ≥ N ′.
Taking order limit in n′ yields
xn ∧ me − um ∧ e ≤ xN,N ′
Taking order limit in m now yields:
xn − u ∧ e ≤ xN,N ′, ∀n ≥ N,
from which it follows that xn − u ∧ e
in Be since e is a weak unit of Be.
o−→ 0 in Be. This yields xn
uo−→ u
The implication (ii)⇒(iii) is clear. For the last implication it suf-
fices, by [AW97, Theorem 2.4], to verify that every norm bounded
laterally increasing sequence in X+ has a supremum. Let (xn) be a
norm bounded laterally increasing sequence in X+. By [AW97, Propo-
sition 2.2], (xn) has supremum in X u, hence is uo-Cauchy in X u. It
follows that (xn) is uo-Cauchy in X and, therefore, by assumption,
uo-converges to some x ∈ X. It is then clear that xn ↑ x in X.
(cid:3)
Theorem 2.6. Let X be a Banach lattice. TFAE:
(i) X is Levi;
(ii) X is boundedly uo-complete;
(iii) Every increasing norm bounded uo-Cauchy net in X+ has a
supremum.
Proof. If X is Levi, then X is boundedly uo-complete by [GLX, Propo-
sition 3.1]. It is clear that (ii)⇒(iii) and the proof of (iii)⇒(i) is the
same as in the last theorem but with [AW97, Theorem 2.4] replaced
with [AW97, Theorem 2.3].
(cid:3)
UNBOUNDED CONVERGENCES
7
3. Completeness of minimal topologies
Throughout this section, X is a vector lattice and τ denotes a locally
solid topology on X. We begin with a brief discussion on relations
between minimal topologies and the B-property. Proposition 3.3 will
be of importance as many properties of locally solid topologies are
stated in terms of positive increasing nets. For minimal topologies,
these properties permit a uniform and efficient treatment.
The B-property was introduced as property (B,iii) by W.A.J. Lux-
emburg and A.C. Zaanen in [LZ64]. It is briefly studied in [AB03] and,
in particular, it is shown that the Lebesgue property does not imply
the B-property. We prove, however, that if τ is unbounded then this
implication does indeed hold true:
Definition 3.1. A locally solid vector lattice (X, τ ) satisfies the B-
property if it follows from 0 ≤ xn ↑ in X and (xn) τ -bounded that
(xn) is τ -Cauchy. An equivalent definition is obtained if sequences are
replaced with nets.
Proposition 3.2. If X is a vector lattice admitting a minimal topology
τ , then τ satisfies the B-property.
Proof. Suppose τ is minimal and (xn) is a τ -bounded sequence satisfy-
ing 0 ≤ xn ↑. By [AB03, Theorem 7.50], (xn) is dominable. By [AB03,
uo−→ u for some
Theorem 7.37], (xn) is order bounded in X u so that xn
u ∈ X u. In particular, (xn) is uo-Cauchy in X u. It follows that (xn) is
uo-Cauchy in X. Since τ is Lebesgue, (xn) is uτ -Cauchy in X. Finally,
since τ is unbounded, (xn) is τ -Cauchy in X.
(cid:3)
Proposition 3.3. Let X be a vector lattice admitting a minimal topol-
ogy τ , and (xα) an increasing net in X+. TFAE:
(i) (xα) is τ -bounded;
(ii) (xα) is τ -Cauchy.
Proof. It remains to prove (ii)⇒(i): Let (xα) be an increasing τ -Cauchy
net in X+. By [Tay, Corollary 5.7], τ extends to a complete Hausdorff
τ u
−→ x for some x ∈ X u.
Lebesgue topology τ u on X u. It follows that xα
Since (xα) is increasing, xα ↑ x in X u.
In particular, (xα) is order
8
M.A. TAYLOR
bounded in X u, hence dominable in X by [AB03, Theorem 7.37]. By
[Tay, Theorem 5.2], (xα) is τ -bounded in X.
(cid:3)
Recall the following definition, taken from [AB03, Definition 2.43].
Definition 3.4. A locally solid vector lattice (X, τ ) is said to satisfy
the monotone completeness property (MCP) if every increasing
τ -Cauchy net of X+ is τ -convergent in X. The σ-MCP is defined
analogously with nets replaced with sequences.
Remark 3.5. By Proposition 3.3, a minimal topology has MCP iff it
is Levi.
Proposition 3.6. Let τ be a Hausdorff locally solid topology on X. If
uτ satisfies MCP then so does τ . If uτ satisfies σ-MCP then so does
τ .
Proof. Suppose 0 ≤ xα ↑ is a τ -Cauchy net. It is then uτ -Cauchy and
τ−→ x.
hence uτ -converges to some x ∈ X. Therefore, xα ↑ x and xα
Replacing nets with sequences yields the σ-analogue.
(cid:3)
Recall by [AB03, Theorem 2.46 and Exercise 2.11] that a Hausdorff
locally solid vector lattice (X, τ ) is (sequentially) complete iff order
intervals are (sequentially) complete and τ has (σ)-MCP. Therefore,
since τ -convergence agrees with uτ -convergence on order intervals, uτ
being (sequentially) complete implies τ is (sequentially) complete.
Lemma 3.7. Let (X, τ ) be a Hausdorff locally solid vector lattice. If
τ is unbounded then TFAE:
(i) τ has MCP and is pre-Lebesgue;
(ii) τ is Lebesgue and Levi.
Proof. It is sufficient, by [DL98, Theorem 2.5], to prove that (X, τ )
contains no lattice copy of c0. Suppose, towards contradiction, that X
does contain a lattice copy of c0, i.e., there is a homeomorphic Riesz
isomorphism from c0 onto a sublattice of X. This leads to a contradic-
tion as the standard unit vector basis is not null in c0, but the copy in
X is by [Tay, Theorem 4.2].
(cid:3)
UNBOUNDED CONVERGENCES
9
Lemma 3.7 is another way to prove that a minimal topology has
MCP iff it is Levi. We next present the sequential analogue:
Lemma 3.8. Let (X, τ ) be a Hausdorff locally solid vector lattice. If
τ is unbounded then TFAE:
(i) τ has σ-MCP and is pre-Lebesgue;
(ii) τ is σ-Lebesgue and σ-Levi.
Proof. (i)⇒(ii) is similar to the last lemma; apply instead [DL98, Propo-
sition 2.1 and Theorem 2.4].
(ii)⇒(i): It suffices to show that τ is pre-Lebesgue. For this, suppose
that 0 ≤ xn ↑≤ u; we must show that (xn) is τ -Cauchy. Since τ is σ-
Levi and order bounded sets are τ -bounded, xn ↑ x for some x ∈ X.
Since τ is σ-Lebesgue, xn
(cid:3)
τ−→ x.
Putting pieces together from other papers, we next characterize se-
quential completeness of uo-convergence.
Theorem 3.9. Let X be a vector lattice. TFAE:
(i) X is sequentially uo-complete;
(ii) Every positive increasing uo-Cauchy sequence in X uo-converges
in X;
(iii) X is universally σ-complete.
In this case, uo-Cauchy sequences are order convergent.
Proof. (i)⇒(ii) is clear. (ii)⇒(iii) by careful inspection of [LC, Propo-
sition 2.8], (iii)⇒(i) and the moreover clause follow from [GTX17, The-
orem 3.10].
(cid:3)
Remark 3.10. Recall that by [AB03, Theorem 7.49], every locally
solid topology on a universally σ-complete vector lattice satisfies the
pre-Lebesgue property. Using uo-convergence, we give a quick proof of
this. Suppose τ is a locally solid topology on a universally σ-complete
vector lattice X; we claim that uo-null sequences are τ -null. This fol-
lows since τ is σ-Lebesgue and uo and o-convergence agree for sequences
by [GTX17, Theorem 3.9]. In particular, since disjoint sequences are
uo-null, disjoint sequences are τ -null.
10
M.A. TAYLOR
We next give the topological analogue of Theorem 3.9:
Lemma 3.11. Let X be a vector lattice admitting a minimal topology
τ . TFAE:
(i) τ is σ-Levi;
(ii) τ has σ-MCP;
(iii) X is universally σ-complete.
(iv) (X, τ ) is sequentially boundedly uo-complete in the sense that
τ -bounded uo-Cauchy sequences in X are uo-convergent in X.
in X, and define xn = Pn
Proof. (i)⇔(ii) follows from Lemma 3.8. We next deduce (iii). Since τ
is σ-Levi, X is σ-order complete; we prove X is laterally σ-complete.
Let {an} be a countable collection of mutually disjoint positive vectors
k=1 ak. Then (xn) is a positive increasing
sequence in X, and it is uo-Cauchy, as an argument similar to [LC,
Proposition 2.8] easily shows. By Theorem 1.1, (xn) is τ -Cauchy, hence
τ−→ x for some x ∈ X since τ has σ-MCP. Since (xn) is increasing
xn
and τ is Hausdorff, xn ↑ x. Clearly, x = sup{an}.
(iii)⇒(iv) follows from Theorem 3.9; (iv)⇒(i) is easy.
(cid:3)
The following question(s) remain open:
Question 3.12. Let X be a vector lattice admitting a minimal topol-
ogy τ . Are the following equivalent?
(i) (X, τ ) is sequentially complete;
(ii) X is universally σ-complete.
Question 3.13. Let (X, τ ) be Hausdorff and Lebesgue. Are the fol-
lowing equivalent?
(i) Order intervals of X are sequentially τ -complete;
(ii) X is σ-order complete.
Remark 3.14. Question 3.12 and Question 3.13 are equivalent. In-
deed, in both cases it is known that (i)⇒(ii). If Question 3.13 is true
then Question 3.12 is true since we have already established that min-
imal topologies have σ-MCP when X is universally σ-complete. Sup-
pose Question 3.12 is true. If X is σ-order complete, then X is an ideal
in its universal σ-completion, X s. Indeed, it is easy to establish that
UNBOUNDED CONVERGENCES
11
if Y is a σ-order complete vector lattice sitting as a super order dense
sublattice of a vector lattice Z, then Y is an ideal of Z; simply modify
the arguments in [AB03, Theorem 1.40]. By [AB03, Theorem 4.22] we
may assume, WLOG, that τ is minimal. τ then lifts to the universal
completion and can be restricted to X s.
Question 3.13 is a special case of Aliprantis and Burkinshaw's [AB78,
Open Problem 4.2], which we state as well:
Question 3.15. Suppose τ is a Hausdorff σ-Fatou topology on a σ-
order complete vector lattice X. Are the order intervals of X sequen-
tially τ -complete?
The case of complete order intervals is much easier than the sequen-
tially complete case. The next result is undoubtedly known, but fits in
nicely; we provide a simple proof that utilizes minimal topologies.
Proposition 3.16. Suppose τ is a Hausdorff Lebesgue topology on X.
Order intervals of X are complete iff X is order complete.
Proof. If X is order complete then order intervals are complete by
[AB03, Theorem 4.28].
By [AB03, Theorem 4.22] we may assume, WLOG, that τ is minimal.
If order intervals are complete then X is an ideal of bX = X u by [AB03,
Theorem 2.42] and [Tay, Theorem 5.2]. Since X u is order complete, so
is X.
(cid:3)
Remark 3.17. If X is an order complete and laterally σ-complete
vector lattice admitting a minimal topology τ , then τ is sequentially
complete. Although these conditions are strong, they do not force X
to be universally complete. This can be seen by equipping the vector
lattice of [AB03, Example 7.41] with the minimal topology given by
restriction of pointwise convergence from the universal completion.
The key step in the proof of Theorem 2.5 is [AW97, Theorem 2.4]
which states that a Banach lattice is σ-Levi if and only if it is laterally
σ-Levi. A sequence (xn) in a vector lattice is said to be laterally
increasing if it is increasing and (xm − xn) ∧ xn = 0 for all m ≥ n.
12
M.A. TAYLOR
We say that a locally solid vector lattice (X, τ ) has the lateral σ-
Levi property if sup xn exists whenever (xn) is laterally increasing
and τ -bounded. For minimal topologies, the σ-Levi and lateral σ-Levi
properties do not agree, as we now show:
Proposition 3.18. Let X be a vector lattice admitting a minimal topol-
ogy τ . TFAE:
(i) X is laterally σ-complete;
(ii) τ has the lateral σ-Levi property;
(iii) Every disjoint positive sequence, for which the set of all possible
finite sums is τ -bounded, must have a supremum.
Proof. (i)⇒(iii) is clear, as is (ii)⇔(iii); we prove (ii)⇒(i). Assume (ii)
and let (xn) be a disjoint sequence in X+. Since (xn) is disjoint, (xn)
has a supremum in X u. Define yn = x1 ∨ · · · ∨ xn. The sequence (yn)
is laterally increasing and order bounded in X u. By [AB03, Theorem
7.37], (yn) forms a dominable set in X+. By [Tay, Theorem 5.2(iv)],
(yn) is τ -bounded, and hence has supremum in X by assumption. This
implies that (xn) has a supremum in X and, therefore, X is laterally
σ-complete.
(cid:3)
In [Lab84] and [Lab85], many completeness-type properties of locally
solid topologies were introduced. For entirety, we classify the remaining
properties, which he refers to as "BOB" and "POB".
Definition 3.19. A Hausdorff locally solid vector lattice (X, τ ) is said
to be boundedly order-bounded (BOB) if increasing τ -bounded
nets in X+ are order bounded in X. (X, τ ) satisfies the pseudo-order
boundedness property (POB) if increasing τ -Cauchy nets in X+
are order bounded in X.
Remark 3.20. It is clear that a Hausdorff locally solid vector lattice
is Levi iff it is order complete and boundedly order-bounded. It is also
clear that BOB and POB coincide for minimal topologies.
Proposition 3.21. Let X be a vector lattice admitting a minimal topol-
ogy τ . TFAE:
(i) (X, τ ) satisfies BOB;
UNBOUNDED CONVERGENCES
13
(ii) X is majorizing in X u.
Proof. (i)⇒(ii): Let 0 ≤ u ∈ X u. Since X is order dense in X u, there
exists a net (xα) in X such that 0 ≤ xα ↑ u. In particular, (xα) is order
bounded in X u, hence dominable in X by [AB03, Theorem 7.37]. By
[Tay, Theorem 5.2], (xα) is τ -bounded. By assumption, (xα) is order
bounded in X, hence, (xα) ⊆ [0, x] for some x ∈ X+. It follows that
u ≤ x, so that X majorizes X u.
(ii)⇒(i): Suppose (xα) is an increasing τ -bounded net in X+.
It
follows from [AB03, Theorem 7.50] that (xα) is dominable, hence order
bounded in X u. Since X majorizes X u, (xα) is order bounded in X. (cid:3)
Remark 3.22. By [AB03, Theorem 7.15], laterally complete vector
lattices majorize their universal completions.
Remark 3.23. If τ is a Hausdorff Fatou topology on X, it is easy to
see that (X, τ ) satisfies BOB iff every increasing τ -bounded net in X+
is order Cauchy in X. Compare with Question 2.1.
We next state the σ-analogue of Proposition 3.21.
Proposition 3.24. Let X be an almost σ-order complete vector lattice
admitting a minimal topology τ . TFAE:
(i) (X, τ ) satisfies σ-BOB;
(ii) X is majorizing in the universal σ-completion X s of X.
Proof. (i)⇒(ii) is similar to Proposition 3.21.
(ii)⇒(i): Suppose (xn) is an increasing τ -bounded sequence in X+.
It is then dominable in X, hence in X s by [AB03, Lemma 7.11]. It
follows by [AB03, Theorem 7.38] that (xn) is order bounded in X s.
Since X is majorizing in X s, (xn) is order bounded in X.
(cid:3)
The next definition is standard in the theory of topological vector
spaces:
Definition 3.25. Let (E, σ) be a Hausdorff topological vector space. E
is quasi-complete if every σ-bounded σ-Cauchy net is σ-convergent.
Remark 3.26. Since Cauchy sequences are bounded, there is no se-
quential analogue of quasi-completeness.
14
M.A. TAYLOR
We finish with the full characterization of completeness of minimal
topologies:
Theorem 3.27. Let X be a vector lattice admitting a minimal topology
τ . TFAE:
(i) X is universally complete;
(ii) τ is complete;
(iii) τ satisfies MCP;
(iv) τ is Levi;
(v) τ is quasi-complete;
(vi) (X, τ ) is boundedly uo-complete in the sense that τ -bounded
uo-Cauchy nets in X are uo-convergent in X.
Proof. (i)⇔(ii) by [Tay, Corollary 5.3] combined with [Tay, Theorem
6.4]. Clearly, (ii)⇒(iii)⇔(iv). (iii)⇒(ii) since if τ satisfies MCP then
τ is topologically complete by [AB03, Corollary 4.39]. We have thus
established that (i)⇔(ii)⇔(iii)⇔(iv).
It is clear that (ii)⇒(v), and
(v)⇒(iii) by Proposition 3.3.
(ii)⇒(vi): Let (xα) be a uo-Cauchy net in X; (xα) is then τ -Cauchy
and hence τ -convergent. The claim then follows from [Tay, Remark
2.26].
(vi)⇒(iv): Suppose 0 ≤ xα ↑ is τ -bounded. (xα) is then uo-Cauchy,
(cid:3)
hence uo-convergent to some x ∈ X. Clearly, x = sup xα.
Remark 3.28. This is in good agreement with Proposition 3.6.
If
the minimal topology satisfies MCP then Proposition 3.6 states that
every Hausdorff Lebesgue topology satisfies MCP. Universally complete
spaces, however, admit at most one Hausdorff Lebesgue topology by
[AB03, Theorem 7.53].
References
[AW97]
[AB80]
[AB78]
Y.D. Abramovich and A.W. Wickstead, When each continuous operator
is regular, II, Indag. Math., 1997.
C.D. Aliprantis and O. Burkinshaw, Minimal topologies and L
Illinois J. Math. 24 (1980), 164 -- 172.
C.D. Aliprantis and O. Burkinshaw, Locally Solid Riesz Spaces, Aca-
demic Press, New York, 1978.
p-spaces,
UNBOUNDED CONVERGENCES
15
[AB03]
[BL88]
[Con05]
[DL98]
C.D. Aliprantis and O. Burkinshaw, Locally Solid Riesz Spaces with
Applications to Economics, 2nd ed., AMS, Providence, RI, 2003.
G. Buskes and I. Labuda, On Levi-like properties and some of their
applications in Riesz space theory, Canad. Math. Bull. Vol. 31 (4), 1988.
J.J. Conradie, The Coarsest Hausdorff Lebesgue Topology, Quaestiones
Math. 28, 287-304, 2005.
L. Drewnowski and I. Labuda, Copies of c0 and ℓ∞ in topological Riesz
spaces, Trans. Amer. Math. Soc. 350 , p. 3555-3570, 1998.
[Frem74] D. Fremlin, Topological Riesz spaces and measure theory, Cambridge
[Gao14]
[GLX]
University Press, 1974.
N. Gao, Unbounded order convergence in dual spaces, J. Math. Anal.
Appl., 419(1) (2014), 347-354.
N. Gao, D. Leung and F. Xanthos, Duality for unbounded order con-
vergence and applications, preprint. arXiv:1705.06143 [math.FA].
[KT]
[Lab85]
[Lab87]
[LC]
[GX14]
[Lab84]
[GTX17] N. Gao, V.G. Troitsky and F. Xanthos, Uo-convergence and its applica-
tions to Ces`aro Means in Banach Lattices, Israel J. Math., 220 (2017),
649 -- 689.
N. Gao and F. Xanthos, Unbounded order convergence and application
to martingales without probability, J. Math. Anal. Appl., 415 (2014),
931-947.
M. Kandi´c and M.A. Taylor, Metrizability of minimal and unbounded
topologies.
I. Labuda, Completeness type properties of locally solid Riesz spaces,
Studia Math. 77 (1984), 349-372.
I. Labuda, On boundedly order-complete locally solid Riesz spaces, Stu-
dia Math. 81 (1985), 245-258.
I. Labuda, Submeasures and locally solid topologies on Riesz spaces,
Math. Z. 195 (1987), 179-196.
H. Li and Z. Chen, Some Loose Ends on Unbounded Order Conver-
gence, Positivity, to appear. arXiv:1609.09707v2 [math.FA]
P. Meyer-Nieberg, Banach lattices, Springer-Verlag, Berlin, 1991.
M.A. Taylor, Unbounded topologies and uo-convergence in locally solid
vector lattices, preprint. arXiv:1706.01575 [math.FA].
W.A.J. Luxemburg and A.C. Zaanen, Notes on Banach function space,
Nederl. Akad. Wetensch, Proc. Ser. A, Note XI 67, 1964.
[MN91]
[Tay]
[LZ64]
Department of Mathematical and Statistical Sciences, University
of Alberta, Edmonton, AB, T6G 2G1, Canada.
E-mail address: [email protected]
|
1303.6944 | 1 | 1303 | 2013-03-27T19:56:27 | Sharp extensions for convoluted solutions of abstract Cauchy problems | [
"math.FA"
] | In this paper we give sharp extension results for convoluted solutions of abstract Cauchy problems in Banach spaces. The main technique is the use of algebraic structure (for usual convolution product $\ast$) of these solutions which are defined by a version of the Duhamel formula. We define algebra homomorphisms from a new class of test-functions and apply our results to concrete operators. Finally, we introduce the notion of $k$-distribution semigroups to extend previous concepts of distribution semigroups. | math.FA | math |
(1.1)
(1.2)
SHARP EXTENSIONS FOR CONVOLUTED SOLUTIONS OF
ABSTRACT CAUCHY PROBLEMS
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Abstract. In this paper we give sharp extension results for convoluted solutions of
abstract Cauchy problems in Banach spaces. The main technique is the use of alge-
braic structure (for usual convolution product ∗) of these solutions which are defined
by a version of the Duhamel formula. We define algebra homomorphisms from a new
class of test-functions and apply our results to concrete operators. Finally, we intro-
duce the notion of k-distribution semigroups to extend previous concepts of distribution
semigroups.
Let A be a closed linear operator on a Banach space X and τ > 0. It is well known (see
1. introduction
[6, Theorem 3.1], [3] or [2, Theorem 2.1]) that if for every x ∈ X, the Cauchy problem
(u′(t) = Au(t) + x, 0 ≤ t < τ
u(0) = 0
has a unique solution u ∈ C 1([0, τ ), X) ∩ C([0, τ ), D(A)) (where D(A) is endowed with
the graph norm), then A is the generator of a strongly continuous semigroup.
This means that the solutions, initially obtained on [0, τ ), admit extensions to [0, ∞)
without loss of regularity. Moreover, these solutions are (uniformly) exponentially bounded
in the sense that there exist M > 0, ω ∈ R independent of u(.) such that all solutions u(.)
satisfy ku(t)k ≤ M eωt, t ≥ 0. Also note that in this case, A is necessarily densely defined.
In 1833, J.M.C. Duhamel considered the following evolution problem corresponding to
the initial-boundary value problem for the heat equation in a domain Ω (Ω is an open
subset of Rn):
= ∆u,
∂u
∂t
u(0, x) = u0(x),
u(t,·)∂Ω = g(t,·),
(t, x) ∈ R+ × Ω
x ∈ Ω;
t > 0.
2010 Mathematics Subject Classification. Primary 47D62, 47D06; Secondary 44A10, 44A35 .
Key words and phrases. Abstract Cauchy problem; convoluted semigroups; Laplace transform, distri-
bution semigroups.
P.J. Miana and L. S´anchez-Lajusticia have been partly supported by supported by Project MTM2010-
16679, DGI-FEDER, of the MCYTS; Project E-64, D.G. Arag´on, and JIUZ-2012-CIE-12, Universidad
de Zaragoza, Spain. This paper was written during several visits of P. J. Miana to the Department of
Mathematics, University of Puerto Rico, R´ıo Piedras Camppus. He is grateful to the Department for its
hospitality.
1
2
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
and proposed the following formula to express the solution of (1.2).
u(t, x) =Z t
0
∂
∂t
u(λ, t − λ, x)dλ,
t > 0,
where u(λ, t, x) is a solution of (1.2) for a particular function g(·, λ0) fixed λ0 ([11]).
This formula allows one to reduce the Cauchy problem for an in-homogeneous partial
differential equation to the Cauchy problem for the corresponding homogeneous equation.
This formula (known also as Duhamel's principle) is of widespread use in partial differential
equations and has been studied in a large number of papers, see for example [12, 34]. In
the [34], the author extends the Duhamel principle to fractional order equations.
Let k : (0, T ) → C be a locally integrable function, X a Banach space and x ∈ X. The
Cauchy problem
(1.3)
(v′(t) = Av(t) + K(t)x
v(0) = 0,
0 < t < T,
0 k(s)xds for 0 < t < T .
is called K-convoluted Cauchy problem where K(t) = R t
If
there exists a solution of the abstract initial value problem u′(t) = Au(t) for 0 < t < T,
u(0) = x then, as usual for a nonhomogeneous equation, we have v = u ∗ K (∗ is the
usual convolution in R+). Local k-convoluted semigroups are defined using a version of
Duhamel's formula and were introduced in [8, 9]. This class of semigroups includes C0-
semigroups and integrated semigroups as particular examples, see a complete treatment in
[27, Section 1.3.1], [19, Chapter 2] and other details in [18, 20]. The concept of regularized
semigroups is covered by taking k(t) ≡ C, 0 ≤ t < τ, where C is a bounded and injective
operator on X.
Contrary to what happens in the case of equation (4.12), if k : (0, ∞) −→ C is lo-
cally integrable and for every x ∈ X there exists a unique solution u ∈ C 1([0, τ ), X) ∩
C([0, τ ), D(A)) for (1.3), it is generally not the case that these solutions can be extended to
[0, ∞), nor that exponential boundedness is achieved in case one can extend the solutions.
In this case, we say that A is the generator of a local k−convoluted semigroup. However,
there is an underlying algebraic structure of k-convoluted semigroups which leads to the
following the extension property: the solution of k-convoluted problem on [0, T ) is used
to express the solution of the k ∗ k-convoluted problem on [0, 2T ), see [9, Section 2] and
[19, Theorem 2.1.1.9]. Stated otherwise, when (1.3) is well posed on [0, τ ), the equation
in which we replace k(.) with (k ∗ k)(.) is well posed on [0, 2τ ). Our result (Theorem 4.4)
provide a sharpening of this extension property.
Γ(α) with α > 0 defines the α-times integrated semigroup.
Originally they were the first example of convoluted semigroups. An extension formula
for n-times integrated semigroups (for n ∈ N) was given in [2, Section IV, (4.2)] and for
α-times integrated semigroup in [29, Formula (5)] with α > 0. Extensions of local α-times
integrated C-semigroups were given in [24, Theorem 6.1] and automatic extension of local
regularized semigroups appears in [36, Section 2].
The special case of k(t) = tα−1
The main objective of this paper is to illustrate the algebraic structure of local k-
convoluted semigroups. In [18, Section 5], authors consider global exponentially bounded
convoluted semigroups and algebra homomorphisms defined via these classes of semi-
groups; in fact both concepts are equivalent, see [18, Theorem 5.7].
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
3
In the context of local convoluted semigroups as well as global non-exponentially bounded
convoluted semigroups, this point of view is not so evident. This is due to the fact that
the Laplace transform is an essential tool in the global exponential case. First we need
some technical identities which involve convolution products in Section 2. In Section 3, we
introduce a new test function space, Dk∗∞(in Definition 3.4) which will play a fundamental
role in this paper, see Theorem 5.1 and Appendix. In the fourth section, we give one of
the main results of this paper. We derive a sharp extension theorem for local convoluted
semigroups (Theorem 4.4). In Section 5, we use the extension formula to define algebra
homomorphism from the test function space Dk∗∞ via local convoluted semigroup (The-
orem 5.1). In Section 6, we apply our results to four concrete operators which generate
(local and global) convoluted semigroups. In the global case and under the exponential
boundedness assumption, the Laplace transform is used as a crucial tool. This is no longer
the case when one consider the local case or the global one without the assumption of ex-
ponential boundedness. In the case we consider in this paper, algebras concerned are no
longer Banach algebras but only locally convex algebras.
Historically distribution semigroups were introduced by J.L. Lions in [25] in the early
sixties in the exponential case with the Laplace transform of distributions as an important
tool. The paper [7] by J. Chazarain presents an extension to the non exponential case
and goes further to introduce the ultradistributional framework (see also the monograph
[26]). This class of vector-valued distribution (with a suitable algebraic structure) gives a
equivalent approach to local integrated semigroups as was proven in [2, Theorem 7.2]. For
local convoluted semigroups, we present a similar approach in Appendix where we intro-
duce k-distributions semigroups and we present their connections with local convoluted
semigroups. The interest in the local case stems from the fact that for the general classes
of generalized semigroups that have been introduced following Lions' paper, by using the
local approach, one is able to obtain a Banach space valued formulation that captures
almost all the situations involved. The monographs [19], [27] and the references cited
therein contain more information on distribution as well as ultradistribution semigroups.
They also explore the ways in which they relate to local convoluted semigroups.
A similar and independent approach may be followed in the abstract Cauchy problem
of second order or wave problem. In this case we need to consider local convoluted co-
sine functions and distribution cosine function (mainly algebra homomorphism for cosine
convolution product) see more details in [30].
2. Some identities for convolution products
Let L1
loc(R+) the space of complex valued locally integrable functions on R+ and we
consider the usual convolution product ∗, given by
(f ∗ g)(t) =Z t
0
f (t − s)g(s)ds, t ≥ 0,
loc(R+). We write k∗2 instead of k ∗ k and then k∗n = k ∗(cid:0)k∗(n−1)(cid:1) for
where f, g ∈ L1
n > 2 is the n−fold convolution power of k. The convolution product is associative and
commutative. We also follow the notation ◦ to denote the dual convolution product of ∗
given by
(f ◦ g)(t) =Z ∞
t
f (s − t)g(s)ds, t ≥ 0,
4
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
where f, g ∈ L1(R+). Note that
max{t t ∈ supp(f ◦ g)} ≤ max{t t ∈ supp(g)},
f, g ∈ L1(R+).
corresponds to the Heaviside function. Moreover, for α > 0, we set jα(t) =
We denote by χ the constant function equal to 1, i.e., χ(t) = 1 for t ∈ R+. This
, t ≥ 0.
It will be convenient to set j0 = δ0, the Dirac measure concentrated at the origin. Observe
that the following semigroup property holds: jα ∗ jβ = jα+β, α, β ≥ 0. The following
lemma will be used for the proof of the main result in Section 4.
Lemma 2.1. Take 0 ≤ τ ≤ t and f, g ∈ L1
Z t−τ
g(t−s) (χ ∗ f ) (s)ds = (g ∗ (χ ∗ f )) (t)−(χ ∗ g) (t−τ ) (χ ∗ f ) (τ ).
loc(R+). Then
tα−1
Γ(α)
0
0
f (t−s) (χ ∗ g) (s)ds+Z τ
dsZ t
f (t − u)du = Z τ
t−τ
d
0
s
Proof. Observe that
have: Z t
Z t−τ
0
f (t − s)Z s
0
parts in the following integral to obtain,
for 0 ≤ τ ≤ t. Note that
Z t−τ
0
g(s)Z t−s
0
f (t − u)du = −f (t − s) and by simple change of variable we
f (u)du. We integrate by
f (s)ds and Z t
s
f (t − u)du =Z t−s
0
0
0
0
t−τ
f (t − s)dsZ t−τ
g(u)duZ t−τ
g(u)du +Z t−τ
g(u)duds = −Z t
f (t − u)du +Z t−τ
= −Z t−τ
= − (χ ∗ g) (t − τ ) (χ ∗ f ) (τ ) +Z t−τ
f (x)dxds −Z t
g(s)Z t
g(s)Z s
g(s)Z t−s
g(s)Z t−s
f (x)dxds = Z t
g(s)Z t−s
0
0
0
0
0
0
s
f (x)dxds
f (t − u)duds
f (t − u)duds
f (u)duds,
0
0
= (g ∗ (χ ∗ f )) (t) −Z τ
= (g ∗ (χ ∗ f )) (t) −Z τ
0
0
t−τ
g(t − u)Z u
0
f (x)dxdu
g(t − u) (χ ∗ f ) (u)du,
(cid:3)
and this concludes the proof.
Taking f = jα and g = jβ with α, β > 0 in Lemma 2.1, we get the equality
sβ
Γ(β + 1)
ds+Z τ
0
(t − s)β−1
Γ(β)
sα
Γ(α + 1)
ds =
tα+β
Γ(α + β + 1)−
(t − τ )β
Γ(β + 1)
τ α
Γ(α + 1)
(t − s)α−1
Z t−τ
0
Γ(α)
for 0 ≤ τ ≤ t.
If we set f = g in Lemma 2.1, we obtain:
Corollary 2.2. Take 0 ≤ τ ≤ t and f ∈ L1
loc(R+). Then
(2.4) (cid:18)Z t−τ
0
0 (cid:19) f (t − s) (χ ∗ f ) (s)ds = (f ∗ (χ ∗ f )) (t) − (χ ∗ f ) (t − τ ) (χ ∗ f ) (τ ).
+Z τ
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
5
Further specializing to f = jα for α > 0, yields the identity:
(cid:18)Z t−τ
0
0 (cid:19) (t − s)α−1
+Z τ
Γ(α)
for 0 ≤ τ ≤ t.
sα
Γ(α + 1)
ds =
t2α
Γ(2α + 1) −
(t − τ )α
Γ(α + 1)
τ α
Γ(α + 1)
As a consequence of the last corollary, we obtain another proof of the following result
given in [19, Lemma 2.1.12] for continuous functions and in [23, Lemma 3.1] for f (t) =
tα−1
Γ(α)
Corollary 2.3. For f ∈ L1
and α > 0.
in particular for α > 0 and s, u ≥ 0, we get that
loc(R+) and s, u ≥ 0 we have
0 −Z u
−Z s
0 −Z u
−Z s
0 (cid:19) f (u + s − r)f (r)dr = 0;
0 (cid:19) (u + s − r)α−1rα−1dr = 0;
0
(cid:18)Z s+u
(cid:18)Z s+u
0
Proof. By change of variable, we write the identity (2.4) as
(f ∗ (χ ∗ f )) (t) − (χ ∗ f ) (t − τ ) (χ ∗ f ) (τ ) =(cid:18)Z t
τ
t−τ(cid:19) f (x)(χ ∗ f )(t − x)dx
+Z t
for 0 ≤ τ ≤ t.
Now we observe that
Similarly, we have
d
dt
(f ∗ (χ ∗ f )) (t) = (f ∗ f )(t).
d
dt
((χ ∗ f )(t − τ )(χ ∗ f )(τ )) = f (t − τ )(χ ∗ f )(τ ),
d
dtZ τ
t
f (u)(χ ∗ f )(t − u)du = −f (t)(χ ∗ f )(0) −Z t
τ
f (u)f (t − u)du
= Z τ
t
f (u)f (t − u)du,
and
d
dtZ t
t−τ
f (u)(χ ∗ f )(t − u)du = f (t)(χ ∗ f )(0) +Z t
0
f (u)f (t − u)du
−f (t − τ )(χ ∗ f )(τ ) −Z t−τ
= Z t
−Z t−τ
f (u)f (t − u)du
0
0
0
f (u)f (t − u)du − f (t − τ )(χ ∗ f )(τ )
f (u)f (t − u)du
6
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Differentiating with respect to the variable t and using the above, we have:
(f ∗ f )(t) = Z t
= Z t−τ
f (x)f (t − x)dx +Z t
f (t − s)f (s)ds +Z τ
t−τ
0
0
τ
f (x)f (t − x)dx
f (t − s)f (s)ds
for 0 ≤ τ ≤ t. Now take t = s + u, and τ = s and we conclude the proof.
(cid:3)
Take k ∈ L1
loc([0, τ )), and we define (kt)t∈[0,τ ) ⊂ L1
(2.5)
kt(s) := k(t − s)χ[0,t](s),
loc([0, τ )) by
s ∈ [0, τ ).
A similar result was considered in [18, Proposition 2.2] for functions belong to L1
Here we present a direct proof for L1
Theorem 2.4. Take k ∈ L1
loc([0, τ )) and (kt)t∈[0,τ ) defined by (2.5). Then
loc([0, τ )).
loc(R+).
kt ∗ ks(x) =Z t+s
t
k(t + s − r)kr(x)dr −Z s
0
k(t + s − r)kr(x)dr,
0 ≤ x < τ,
for 0 ≤ s, t ≤ t + s < τ .
Proof. We consider (without loss of generality) that 0 ≤ s ≤ t. First we consider 0 ≤ x ≤ s.
Then
k(t − (x − y))k(s − y)dy,
0 ≤ x < τ,
and
0
kt ∗ ks(x) =Z x
0 (cid:19) k(t + s − r)kr(x)dr =(cid:18)Z t+s
−Z s
k(u)k(t + s − u − x)du −Z s−x
k(u)k(t + s − u − x)du =Z x
0
0
t
t
(cid:18)Z t+s
=Z s
=Z s
s−x
0
x (cid:19) k(t + s − r)k(r − x)dr
−Z s
k(t + s − x − y)k(y)dy
k(s − y)k(t − x + y)dy
where we have changed variables in several equalities. The other cases s ≤ x ≤ t, t ≤
In particular we remark that
x ≤ t + s and t + s ≤ x < τ are made following ideas.
kt ∗ ks(x) = 0 for t + s ≤ x < τ .
(cid:3)
Remark 2.5. By Proposition 4.3 and Theorem 2.4, we may conclude that (kt)t∈[0,τ ) is a
local k-convoluted semigroup in L1
loc([0, τ )). In fact, note that kt = δt ∗ k where (δt)t≥0
is the Dirac measure concentrated at t = 0. In this sense, (kt)t∈[0,τ ) is the canonical local
k-convoluted semigroup.
3. The Laplace transform and k-Test function spaces
Let k ∈ L1
loc(R+). We write bybk the usual Laplace transform of k, given by
e−λtk(t)dt,
N→∞Z N
0
bk(λ) = lim
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
7
in the case that there exists for λ ∈ C; abs(k) is defined by abs(k) := inf{ℜλ; existbk(λ)},
see [3, Section 1.4]. In the case that k(t) ≤ M eωt for a.e. t ≥ 0 and M, ω > 0, we have
that
bk(λ) =Z ∞
0
e−λtk(t)dt,
ℜλ > ω.
For λ ∈ C, we write eλ(t) = e−λt, t ≥ 0.
Lemma 3.1. Take k ∈ L1
Then
In particular, ifbk(λ) = 0 for some ℜλ > ω, then k ◦ e−λ = 0.
Proof. Take λ ∈ C with ℜλ > ω and
ℜλ > ω.
loc(R+) such that k(t) ≤ M eωt for a.e. t ≥ 0 and M, ω > 0.
k ◦ e−λ =bk(λ)e−λ,
k(s − t)e−λsds =Z ∞
k(u)e−λudue−λt =bk(λ)e−λ(t)
(cid:3)
0
k ◦ e−λ(t) =Z ∞
t
for t > 0.
We write by supp(h) the usual support of a function h defined in R; D is the space
of C(∞) functions with compact support on R and D0 is the set of C(∞) functions with
compact support on [0,∞), D0 ⊂ D. The space D will be equipped with the Schwartz
topology which turns it into a complete topological vector space. We denote the topology
by τ. In particular, sequential convergence in D is described by: let (φn)n≥1 ⊂ D, φ ∈ D,
then φn →τ φ if and only if
(i) there exists a compact subset K ⊂ R such that supp(φn), supp(φ) ⊂ K.
(ii) for any j ≥ 0, φ(j)
n → φ(j) uniformly on compacts sets.
Note that D0 is a closed subspace of D and then (D0, τ ) is a complete topological space
(we keep the same notation for the topology τ and its restriction to the subspace D0).
We denote by D+ the set of functions defined by φ+ : [0,∞) → C, given by φ+(t) := φ(t)
for t ≥ 0 and φ ∈ D and define K : D → D+ by K(φ) = φ+ for φ ∈ D. Due to the extension
theorem of R. T. Seeley [32], there exists a linear continuous operator Λ : D+ → D, such
that KΛ = ID+; in particular if ψ is a C(∞) functions on [0,∞) and compact support then
ψ ∈ D+. The space D+ is also a complete topological vector spaces equipped with the
τ -topology of uniform convergence on compact subsets.
We define the operator T ′k : D → D by f 7→ T ′k(f ) := k ◦ f, that is,
T ′k(f )(t) =Z ∞
t
k(s − t)f (s)ds, t ≥ 0.
We shall also use the same notation for the restriction to D+ that is, T ′k : D+ → D+;
however T ′k : D0 6→ D0. Note that T ′k(fu) = (T ′k(f ))u, where fu(t) = f (u + t) for u, t ≥ 0
and f ∈ D+.
In the case that 0 ∈ supp(k), we have that T ′k : D+ → D+ is an injective, linear and
continuous homomorphism such that
T ′k(f ◦ g) = f ◦ T ′k(g),
f, g ∈ D+,
8
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
see [18, Theorem 2.5]. Then, we define the space Dk by Dk := T ′k(D+) ⊂ D+ and the right
inverse map of T ′k, i.e., Wk : Dk → D+ by
f (t) = T ′k(Wk(f ))(t) =Z ∞
k(s − t)Wkf (s)ds,
see [18, Definition 2.7]. Note that the operator Wk : D+ → D+ is a closed operator
(D(Wk) = Dk), but we cannot apply the open mapping theorem to conclude that it is
continuous.
It is clear that the subspace Dk is also a topological algebra: take f, g ∈ Dk, then
f ∗ g ∈ Dk ([18, Theorem 2.10]) and the map (f, g) → f ∗ g is continuous in Dk. Moreover
Wk(k ◦ f ) = f for f ∈ D+ and fu ∈ Dk, with
Wk(fu) = (Wk(f ))u,
(3.6)
f ∈ Dk,
t ≥ 0,
t
f ∈ Dk,
u ≥ 0.
We have the following property to the effect that Wk does not increase the support.
loc(R+) be such that 0 ∈ supp(k) and let a > 0. Then supp(f ) ⊂
Lemma 3.2. Let k ∈ L1
[0, a] if and only if supp(Wkf ) ⊂ [0, a] for f ∈ Dk.
Proof. Take f ∈ Dk such that supp(f ) ⊂ [0, a]. Then
0 = f (t) = k ◦ Wkf (t) =Z ∞
t
k(s − t)Wkf (s)ds,
t ≥ a,
Wkf = l ◦ Wk∗lf,
f ∈ Dk∗l.
see [18, Lemma 2.8]. A consequence of (3.8) is the following lemma.
Lemma 3.3. Take k ∈ L1
(i) Dk∗n ֒→ Dk∗m for n ≥ m ≥ 1.
(ii) Wk∗mf = kn−m ◦ Wk∗nf = Wk∗n(kn−m ◦ f ) and if supp(Wk∗nf ) ⊂ I with I an interval
in R+ then supp(Wk∗mf ) ⊂ I for f ∈ Dk∗n and n ≥ m ≥ 1.
loc(R+) such that 0 ∈ supp(k). Then
The next definition gives the test function space will be used later to obtain new dis-
tribution spaces and corresponding distribution semigroups.
We write t = a + r for t ≥ a and r ≥ 0,
0 =Z ∞
t
k(s − t)Wkf (s)ds =Z ∞
0
k(x)Wkf (x + a + r)dx =Z ∞
r
k(x − r)(Wkf )a(x)dx,
where (Wkf )a(x) := (Wkf )(x+a) for x > 0. We apply the Titchmarsh-Foia¸s [18, Theorem
2.4] to conclude that (Wkf )a(x) = 0 for x > 0, i.e., supp(Wkf ) ⊂ [0, a]. Conversely,
suppose that supp(Wkf ) ⊂ [0, a]. It then follows from the representation
f (t) =Z ∞
t
k(s − t)Wkf (s)ds,
t ≥ 0,
(cid:3)
that supp(f ) ⊂ [0, a].
Note that in the case that f ∈ Dk then f (n) ∈ Dk and
Wk(f (n)) = (Wkf )(n),
n ≥ 1;
loc(R+) such that 0 ∈ supp(k)∩ supp(l). Then 0 ∈ supp(k ∗ l), Dk∗l ⊂ Dk ∩Dl
(3.7)
take k, l ∈ L1
and
(3.8)
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
9
Definition 3.4. Take k ∈ L1
defined by
loc(R+) such that 0 ∈ supp(k). We denote by Dk∗∞ the space
Dk∗∞ :=
Dk∗n.
∞\n=1
It is clear that Dk∗∞ is also a topological algebra (equipped with the τ -topology) and
Dk∗∞ ֒→ Dk∗n ֒→ D+. In fact, Dk∗∞ is the inverse (or projective) limit of the family
(Dk∗n)n≥1. By Lemma 3.3, Wk∗n : Dk∗∞ → Dk∗∞ and k∗n ◦ Wk∗nf = f for f ∈ Dk∗∞ and
n ∈ N. Note that if f ∈ Dk∗∞ then fu ∈ Dk∗∞ for u ≥ 0, see formula (3.6).
Example 3.5. In the case that Dk = D+, then Dk∗n = D+ for n ∈ N and Dk∗∞ = D+.
Take z ∈ C, ez(t) := ezt for t ≥ 0, Then Dkez = {ezf f ∈ Dk} = Dk and
Wkez f = ezWk(e−zf ),
f ∈ Dk;
in the case that Dk = D+, then Dkez = D+, see [18, Proposition 2.9].
(i) Recall that α > 0 and jα(t) := tα−1
of order α, Wα and Djα = Dj ∗∞
of usual derivation, see more details for example in [31].
α = D+; note that for α ∈ N, Wα = (−1)α dα
(ii) Given α > 0 and z ∈ C, we have that Dezjα = D(ezjα)∗∞ = D+ and
Γ(α) ; the map Wjα is the Weyl fractional derivative
dtα , the α-iterate
Wezjαf = ezWα(e−zf ),
f ∈ D+;
for α = 1, 2 see explicit expressions in [18, Section 2].
(iii) It is straightforward to check that T ′χ(0,1)
Dχ(0,1)∗n = D+ and
(f )(t) =R t+1
t
f (s)ds for f ∈ D+, Dχ(0,1) =
Wχ(0,1)f (t) = −
∞Xn=0
f′(t + n),
f ∈ D+,
t ≥ 0.
Now let f, g ∈ Dk. Then f ∗ g ∈ Dk and
(3.9)
Wk(f ∗ g)(s) =Z s
−Z ∞
0
s
Wkg(r)Z s
Wkg(r)Z ∞
s−r
s
k(t + r − s)Wkf (t)dtdr
k(t + r − s)Wkf (t)dtdr,
see [18, Theorem 2.10].
Under some conditions on the function k, some Banach algebras under the convolution
product may be considered as the next theorem shows.
Theorem 3.6. ([18, Theorems 3.4 and 3.5]) Let k ∈ L1
supp(k) and abs(k) < ∞. Then the formula
kfkk,eβ :=Z ∞
0
Wkf (t)eβtdt,
f ∈ Dk,
loc(R+) be a function with 0 ∈
for β > max{abs(k), 0} defines an algebra norm on Dk for the convolution product ∗. We
denote by Tk(eβ) the Banach space obtained as the completion of Dk in the norm k · kk,eβ ,
and then we have Tk(eβ) ֒→ L1(R+).
10
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Note that in the three examples below, the space Dk = D+. However, as the following
result shows, there are functions k such that Dk D+ and Dk∗∞ = {0}.
Theorem 3.7. Take k ∈ L1
loc(R+) such that 0 ∈ supp(k) and 0 ≤ abs(k) < ∞.
bk(λ0) = 0 for some ℜλ0 > abs(k) then Dk D+ and Dk∗∞ = {0}.
Proof. We suppose that Dk = D+ and bk(λ0) = 0 for some ℜλ0 > abs(k). Take β ∈ R
such that abs(k) < β < ℜλ0, There exists fn ⊂ D+ such that
(3.10)
If
Wkfn(t) − e−λ0(t)eβtdt → 0,
n → ∞,
As a consequence of Theorem 3.6, we obtain that fn 6→ 0 in L1(R+). On the other hand,
by [18, Theorem 2.5 (ii)] and (3.10), we get that
fn(t) − k ◦ e−λ0(t)eβtdt → 0,
n → ∞.
By Lemma 3.1 k ◦ e−λ0 = 0, and then fn → 0 in L1(R+). We conclude that Dk D+.
Now take f ∈ Dk∗∞. Then there exists a sequence (gn) ⊂ D+, such that f = k∗n ∗ gn
for n ≥ 1. Then bf (λ0) =(cid:16)bk(λ0)(cid:17)n
zero on λ0 of order n at least for any n ≥ 1. We conclude that f = 0.
Example 3.8. The following example was presented in [4, Section 5] and appeared later in
other references in connection to convoluted semigroups (see [20, Example 6.1] and [19,
Example 2.8.1]). Let
bgn(λ0) = 0 for any n ≥ 1. We conclude that bf has a
(cid:3)
Z ∞
0
Z ∞
0
K(λ) :=
1
λ2
n2 − λ
n2 + λ
,
ℜλ > 0.
∞Yn=0
Then there exists a continuous and exponentially bounded function κ in [0,∞) such that
We note that for the cases k(t) = jα(t), t > 0 (corresponding to local integrated semi-
bκ = 1/K. Moreover, 0 ∈ supp(κ) and we apply Theorem 3.7 to conclude that Dκ D+.
groups) andbk(λ) ==
in [8]) and will be presented in Section 6), we havebk(λ) 6= 0 for all ℜλ > abs(k).
(where l > 0, 0 < a < 1 and which are considered
∞Yj=1 1 +
4. Local convoluted semigroups
a!−1
lz
1
j
The definition of global k-convoluted semigroups was introduced by the first time by I.
Cioranescu [8] and subsequently developed in [9] (see also [17] and the monographs [19]
and [27]). We will consider the following definition of local k-convoluted semigroup as
appears in [20, Definition 2.1].
Definition 4.1. Let 0 < τ ≤ ∞, k ∈ L1
loc([0, τ )) and A be a closed operator. (Sk(t))t∈[0,τ ) ⊂
B(X) a strongly continuous operator family. The family (Sk(t))t∈[0,τ ) is a local k-convoluted
semigroup (or local k-semigroup in short) generated by A if Sk(t)A ⊂ ASk(t),R t
0 Sk(s)xds ∈
D(A) for t ∈ [0, τ ) and x ∈ X and
(4.11)
AZ t
0
Sk(t)xdt = Sk(t)x −Z t
0
k(s)dsx,
x ∈ X,
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
11
for t ∈ [0, τ ); in this case the operator A is called the generator of (Sk(t))t∈[0,τ ). We say
that (Sk(t))t∈[0,τ ) is non degenerate if S(t)x = 0 for all 0 ≤ t < τ implies x = 0.
Alternatively, in relation to Problem (1.3), we note that when the problem is well posed
in the sense that for every x ∈ X, there exists a unique solution v ∈ C 1([0, τ ), X) ∩
C([0, τ ), D(A)), we set S(t)x = v′(t), 0 ≤ t < τ, x ∈ X. It follows from the Closed Graph
Theorem that S(t) ∈ B(X), 0 ≤ t < τ. Clearly, t 7→ S(t) is strongly continuous from
[0, τ ) to B(X). The local convoluted semigroup defined in this manner is necessarily non
degenerate, due to the uniqueness assumption.
It is easy to prove that if A generates a k-convoluted semigroup (Sk(t))t∈[0,τ ), then
Sk(0) = 0 and Sk(t)x ∈ D(A) for t ∈ [0, τ ) and x ∈ X. See more details, for example in
[17] and [20].
Remark 4.2. (i) For α > 0 and k(t) = tα−1
Γ(α) for t > 0, we get α-times integrated semigroups
which were introduced in [14]. We follow the usual notation (Sα(t))t∈[0,τ ) for α-times local
integrated semigroups.
(ii) If C ∈ B(X) is an injective operator and we set k(t) ≡ C, 0 ≤ t < τ then we recover
the concept of local C−regularized semigroups. Local C−regularized semigroups were
first studied in [33].
(iii) One condition in the definition of local convoluted semigroup, equation (4.11) may be
interpreted as a Duhamel formula for the abstract Cauchy problem. More precisely, if we
are interested in the (non-homogeneous) initial value problem
(u′(t) = Au(t) + F (t), 0 ≤ t < τ
u(0) = x ∈ X,
where F is an X−valued function and τ ∈ (0, ∞], and K(.) takes values in B(X) with the
additional assumptions that K(t)K(s) = K(s)K(t), t, s ∈ [0, τ ); AK(t)x = K(t)Ax, t ∈
[0, τ ), x ∈ D(A), we can consider the regularized problem
(v′(t) = Av(t) + K(t)x + FK (t), 0 ≤ t < τ
v(0) = 0,
in which FK(t) = (K ∗ F )(t) =R t
0 K(t − s)F (s)ds, 0 ≤ t < τ More details can be found in
the reference [9]. We hall be concerned only with the situation where K(t) = φ(t)I where
I is the identity operator on X. Spectral criteria for the generation of local convoluted
semigroups involving the resolvent of the generator can be found in the references [8], [9],
[17] and [19].
(iv) Other equivalent definitions of local convoluted semigroup, using the composition
property (see Proposition 4.3) or the Laplace transform ([17, Theorem 3.2]) show this
algebraic aspect in a straightforward way.
The next characterization of local k-semigroups has the advantage to offer an algebraic
character which is crucial in the development of the theory as we will see in Theorem 5.1.
The proof runs parallel as in the global case presented in [20, Proposition 2.2], see also
[19, Proposition 2.1.5].
Proposition 4.3. Let 0 < τ ≤ ∞, k ∈ L1
loc([0, τ )), A a closed linear operator and
(Sk(t))t∈[0,τ ) a non-degenerate strongly continuous operator family. Then (Sk(t))t∈[0,τ ) is
12
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
a local k-convoluted semigroup generated by A if and only if Sk(0) = 0 and
k(t + s − r)Sk(r)xdr −Z s
0
k(t + s − r)Sk(r)xdr,
x ∈ X,
(4.12) Sk(t)Sk(s)x =Z t+s
for 0 ≤ s, t ≤ t + s < τ .
Note that if k ∈ L1
t
loc([0, τ )), and we define (kt)t∈[0,τ ) ⊂ L1
loc([0, τ )) by
kt(s) := k(t − s)χ[0,t](s),
s ∈ [0, τ ).
By Proposition 4.3 and Corollary 2.3, we may conclude that (kt)t∈[0,τ ) is a local k-
convoluted semigroup in L1
loc([0, τ )). In this paper, we only consider local k-convoluted
semigroups which are non-degenerate.
The next theorem is the main result in this paper and shows how a local k-convoluted
semigroup (Sk(t))t∈[0,τ ) is extended to [0, nτ ); in fact we get a local k∗n-convoluted semi-
group in [0, nτ ) for n ∈ N. Note that we improve previous results ([9, Section 2] and [19,
Theorem 2.1.1.9]): our approach is sharper that n-iterations of these theorems.
Theorem 4.4. Let n ∈ N, 0 < τ ≤ ∞, k ∈ L1
loc([0, (n+1)τ )) and (Sk(t))t∈[0,τ ) be a local k-
convoluted semigroup generated by A. Then the family of operators (Sk∗(n+1)(t))t∈[0,(n+1)κ]
defined by
Sk∗(n+1)(t)x =Z t
0
k(t − s)Sk∗n(s)xds,
x ∈ X,
for t ∈ [0, nκ] and
0
k(t − s)Sk∗n(s)xds +Z t−nκ
Sk∗(n+1)(t)x = Sk∗n(nκ)Sk(t− nκ)x +Z nκ
k∗n(t − s)Sk(s)xds,
for x ∈ X and t ∈ [nκ, (n + 1)κ] is a local k∗(n+1)-semigroup generated by A for any κ < τ .
Then we conclude that A generates a local k∗(n+1)-semigroup (Sk∗(n+1)(t))t∈[0,(n+1)τ ).
Proof. Note that the family of operators (Sk∗(n+1)(t))t∈[0,(n+1)κ] is strongly continuous. It is
known that (Sk∗(n+1)(t))t∈[0,nκ] is a local k∗(n+1)-semigroup generated by A, see for example
[19, Proposition 2.1.3] and [18, Proposition 5.2]. Now take t ∈ [nκ, (n + 1)κ] and x ∈ X.
0 Sk∗(n+1)(r)xdr ∈ D(A).
Since
0
Sk∗(n+1)(r)xdr,
0
Sk∗(n+1)(r)xdr =Z nκ
It is clear that Sk∗(n+1)(t)A ⊂ ASk∗(n+1)(t) and we show that R t
Z t
we check thatR t
nκ(cid:18)Z nκ
+Z t
Sk∗(n+1)(r)xdr +Z t
nκ Sk∗(n+1)(r)xdr ∈ D(A), i.e,
Sk∗n(nκ)Sk(r − nκ)xdr
Z t
nκ
nκ
0
0
k(r − s)Sk∗n(s)xds +Z r−nκ
Sk∗n(nκ)Z t
nκ
0
Sk(r − nκ)xdr ∈ D(A).
As (Sk(t))t∈[0,τ ) is a local k-convoluted semigroup generated by A, we get that
k∗n(r − s)Sk(s)xds(cid:19) dr ∈ D(A).
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
13
0 k(r − s)Sk∗n(s)xdsdr ∈ D(A). We apply the Fubini theorem,
Now we prove thatR t
change the variable u = r − s, to get that
0
nκR nκ
k(r − s)Sk∗n(s)xdsdr =Z nκ
Z t
nκZ nκ
k(u)duds =Z t
Sk∗n(s)xZ t−s
Z nκ
k∗n(r − s)Sk(s)xdsdr =Z t
Z t
nκZ r−nκ
nκ−s
0
0
0
0
In a similar way, it is shown that
nκ
Sk∗n(s)xZ t
k(u)Z min{t−u,nκ}
k∗n(u)Z t−u
max{nκ−u,0}
nκ
0
k(r − s)drds
Sk∗n(s)xdsdu ∈ D(A).
Sk(s)xdsdu ∈ D(A).
To finish the proof we prove the equality (4.11) for t ∈ [nκ, (n + 1)κ] and x ∈ X. Note
that
0
We apply the Fubini theorem to get that
Sk∗(n+1)(s)xds = Sk∗(n+1)(nκ)x −Z nκ
AZ t
Sk∗(n+1)(s)xds = Sk∗n(nκ)Z t−nκ
Z t
+Z t−nκ
k∗(n+1)(s)dsx + AZ t
Sk∗n(r)xZ t
k∗n(s − r)dsdr.
Sk(u)xdu +Z nκ
Sk(r)xZ t
(4.13)
r+nκ
nκ
nκ
0
0
0
0
nκ
We apply the operator A to the first summand to get that
Sk∗(n+1)(s)xds.
k(s − r)dsdr
Sk(u)xdu = Sk∗n(nκ)Sk(t − nκ)x − Sk∗n(nκ)xZ t−nκ
0
k(u)du.
We apply the operator A and the Fubini theorem to obtain that
k(u)du.
k(u)du−R nκ−r
0
Sk∗n(r)xdrdu
0
nκ k(s− r)ds =R t−r
k(u)Z min{nκ,t−u}
k∗n(y)dyx(cid:19) du
k(u)du
0
0
0
0
0
Sk∗n(nκ)AZ t−nκ
In the second summand of (4.13) we writeR t
k(u)dudr = AZ t
k∗n(y)dyx(cid:19)Z t−nκ
k(u)(cid:18)Sk∗n(t − u)x −Z t−u
k(u)du +Z nκ
k∗n(y)dy(cid:19)(cid:18)Z t−nκ
k(u)dudr = Z nκ
Sk∗n(r)xZ t−r
AZ nκ
=(cid:18)Sk∗n(nκ)x −Z nκ
+Z t
= Sk∗n(nκ)xZ t−nκ
−(cid:18)Z nκ
Sk∗n(r)xZ nκ−r
t−nκ
0
0
0
0
0
0
0
0
AZ nκ
0
Using similar ideas we also get that
k(t − r)Sk∗n(r)xdr
0
k(u)du(cid:19) x −Z nκ
k(t − r)Z r
k(u)(cid:18)Sk∗n(nκ − u)x −Z nκ−u
0
0
k∗(n+1)(r)drx.
0
= Sk∗(n+1)(nκ)x −Z nκ
0
k∗n(y)dydrx.
k∗n(r)drx(cid:19) du
We conclude that
To finish the proof we join together all summands to have that
14
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
We apply the operator A and the Fubini theorem to obtain that
In the third summand of (4.13) we write R t
k∗n(s − r)dsdr = AZ t
k∗n(u)du +Z t−nκ
Sk(r)xZ t
AZ t−nκ
= Sk(t − nk)xZ nκ
0
0
0
0
r
r+nκ k∗n(s − r)ds = R t
r
r −R r+nκ
k∗n(u)Z min{t−nκ,t−u}
k∗n(t − r)Z r
0
0
0
k∗n(t − r)Sk(r)xdr
and finally we get
k(y)dy.
k∗n(s − r)ds.
Sk(r)xdrdu
−(cid:18)Z nκ
0
k∗n(u)du(cid:19)(cid:18)Z t−nκ
0
k(u)du(cid:19) −Z t−nκ
AZ t−nκ
0
r
0
0
0
r+nκ
k∗n(s − r)dsdr
AZ t−nκ
Sk(r)xZ r+nκ
= Sk(t − nκ)xZ nκ
Sk(r)xZ t
=Z t−nκ
k∗n(u)du −(cid:18)Z nκ
k∗n(s − r)dsdr
k∗n(t − r)Sk(r)xdr −Z t−nκ
Sk∗(n+1)(s)xds = Sk∗n(nκ)Sk(t − nκ)x +Z nκ
+Z t−nκ
−Z nκ
k∗n(t − r)Sk(r)xdr −(cid:18)Z nκ
k∗n(y)dydrx −Z t−nκ
k(t − r)Z r
= Sk∗(n+1)(t)x −Z t
0
k∗(n+1)(s)xds,
AZ t
0
0
0
0
0
0
0
0
0
k∗n(u)du(cid:19)(cid:18)Z t−nκ
0
k(u)du(cid:19) .
k∗n(t − r)Z r
0
k(y)dy.
k(t − r)Sk∗n(r)xdr
k∗n(y)dy(cid:19)(cid:18)Z t−nκ
k∗n(t − r)Z r
0
0
k(u)du(cid:19) x
k(y)dydr
where we have used the Lemma 2.1. This proves the claim.
(cid:3)
In fact, the expression of the (Sk∗(n+1)(t))t∈[0,(n+1)κ] is not unique as shown next result.
Both proof are similar to the proof of Theorem 4.4 and are left to the reader.
Theorem 4.5. Let n ≥ 2, 0 < τ ≤ ∞, k ∈ L1
loc([0, nτ )) and (Sk(t))t∈[0,τ ) is a local k-
convoluted semigroup generated by A. Then the family of operators (Sk∗n(t))t∈[0,nκ] defined
in Theorem 4.4 verify that
Sk∗n(t)x =Z t
0
k∗(n−j)(t − s)Sk∗j (s)xds,
x ∈ X,
for t ∈ [0, jκ] and
Sk∗n(t)x = Sk∗j (nκ)Sk∗(n−j)(t − jκ)x
+Z jκ
0
k∗(n−j)(t − s)Sk∗j (s)xds +Z t−jκ
0
for x ∈ X, 1 ≤ j ≤ n − 1 and t ∈ [jκ, nκ] and any κ < τ .
k∗j(t − s)Sk∗(n−j)(s)xds,
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
15
Corollary 4.6. Let 0 < τ ≤ ∞, k ∈ L1
loc([0, 2τ )) and (Sk(t))t∈[0,τ ) be a local k-convoluted
semigroup. Then the family of operators (Sk∗k(t))t∈[0,2κ] defined in Theorem 4.4 verifies
that
Sk∗k(t + s)x = Sk(t)Sk(s)x +(cid:18)Z t
0
for t, s ∈ [0, κ) and x ∈ X.
0 (cid:19) k(t + s − u)Sk(u)xdu
+Z s
The next theorem was given in [29, Theorem 2] in the case n = 1.
Corollary 4.7. Let n ∈ N, 0 < τ ≤ ∞, (Sα(t))t∈[0,τ ) is a local α-times integrated semi-
group generated by A. Then the family of operators (S(n+1)α(t))t∈[0,(n+1)κ] defined by
(t − s)α−1
Γ(α)
Snα(s)xds,
x ∈ X,
S(n+1)α(t)x =Z t
S(n+1)α(t)x = Snα(nκ)Sα(t−nκ)x+Z nκ
for t ∈ [0, nκ] and
0
0
(t − s)α−1
Γ(α)
Snα(s)xds+Z t−nκ
0
(t − s)nα−1
Γ(nα)
Sα(s)xds,
for x ∈ X and t ∈ [nκ, (n + 1)κ] is a local α-times integrated semigroup generated by A
for any κ < τ . Then we conclude that A generates a local α-times integrated semigroup
generated by A, (S(n+1)α(t))t∈[0,(n+1)τ ).
Remark 4.8. In the case α ∈ N, and t ∈ [nκ, (n + 1)κ], note that
and
0
Z nκ
Z t−nκ
0
(t − s)α−1
Γ(α)
Snα(s)xds =
(t − s)nα−1
Γ(nα)
Sα(s)xds =
α−1Xj=0
nα−1Xj=0
(t − nκ)j
j!
S(n+1)α−j(nκ)
(nκ)j
j!
S(n+1)α−j(t − nκ)
for x ∈ X and we recover the extension given in [2, Theorem 4.1] for the case n = 1. To
finish this section, we mention that other extension result (for scalar α-times semigroups
in this case) may be found in [21, Lemma 4.4],
t ∗ I n
I n
s = I 2n
s+t −
n−1Xj=0(cid:18) sj
j!
I 2n−j
t
tj
j!
−
I 2n−j
s
(cid:19) ,
t, s > 0,
where I n
t (r) := (t−r)n
n! χ[0,t](r), r ∈ R+ and n ∈ N ∪ {0}.
5. Algebra homomorphisms and local convoluted semigroups
As the following theorem shows, local k-convoluted semigroups induce algebra homo-
morphisms from certain spaces of test functions Dk∗∞. Note that the extension theorem
(Theorem 4.4) is necessary to define the algebra homomorphisms from functions defined
on R+. The space Dk∗∞ is introduced in Definition 3.4.
16
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Theorem 5.1. Let k ∈ L1
loc(R+) with 0 ∈ supp(k), and (Sk(t))t∈[0,τ ] a non-degenerate
local k-convoluted semigroup generated by A. We define the map Gk : Dk∗∞ → B(X) by
Gk(f )x :=Z nτ
x ∈ X, f ∈ Dk∗∞,
Wk∗nf (t)Sk∗n(t)xdt,
0
where supp(f ) ⊂ [0, nτ ] and (Sk∗n(t))t∈[0,nτ ] is defined in Theorem 4.4 for some n ∈ N.
Then the following properties hold.
(i) The map Gk is well defined, linear and bounded.
(ii) The map Gk(f ∗ g) = Gk(f )Gk(g) for f, g ∈ Dk∗∞.
(iii) Gk(f )x ∈ D(A) and AGk(f )x = −Gk(f′)x − f (0)x for any f ∈ Dk∗∞ and x ∈ X.
Proof. Take f ∈ Dk∗∞ and supp(f ) ⊂ [0, nτ ] for some n ∈ N. First, we prove that Gk is
well defined. Let m ≥ n, k∗m = k∗n ∗ k∗(m−n), and k∗(m−n) ◦ Wk∗mf = Wk∗nf by Lemma
3.3 (ii). Now we apply the Lemma 3.2 to conclude supp(Wk∗mf ) ⊂ [0, nτ ]. By Theorem
4.5 and the Fubini theorem, we get that
0
Z mτ
=Z nτ
Wk∗mf (t)Sk∗m(t)xdt =Z nτ
k∗(m−n) ◦ Wk∗mf (t)Sk∗n(t)xdt =Z nτ
0
0
0
Wk∗mf (t)(k∗(m−n) ∗ Sk∗n)(t)xdt
Wk∗nf (t)Sk∗n(t)xdt
for x ∈ X.
It is direct to check that Gk is linear. Now take (fn)n≥1 ⊂ Dk∗∞, and f ∈ Dk∗∞ such
that fn → f . Then there exists n ∈ N such that supp(fn), supp(f ) ⊂ [0, nτ ]. Note that
the map t 7→ Sk∗n(t)x, [0, nτ ] → X, is continuous and
0
Wk∗nfn(t)−Wk∗nf (t)kSk∗n(t)xkdt ≤ CxZ nτ
kGk(fn)x−Gk(f )xk ≤Z nτ
for x ∈ X. Now consider the operator T ′k∗n : L1[0, nτ ] → L1[0, nτ ], f 7→ T ′k∗n(f ) =
k∗n ◦ f given in Section 3. By the open mapping theorem, T ′k∗n is open; we conclude that
Wk∗nfn → Wk∗nf in L1[0, nτ ], the map Gk is bounded and the part (i) is proved.
Take f, g ∈ Dk∗∞, i.e. f, g ∈ Dk∗n and then f ∗ g ∈ Dk∗n (see [18, Theorem 2.10]) for
n ≥ 1. Then f ∗ g ∈ Dk∗∞. Now we show that Gk(f ∗ g) = Gk(f )Gk(g). Take n ∈ N such
that supp(f ), supp(g) ⊂ [0, nτ ] and by Lemma 3.2, supp(Wk∗2nf ), supp(Wk∗2ng) ⊂ [0, nτ ].
Then supp(f ∗ g) ⊂ [0, 2nτ ] and supp(Wk∗2n(f ∗ g)) ⊂ [0, 2nτ ]. By (3.9) we have that
Wk∗n(fn−f )(t)dt
0
0
Gk(f ∗ g)x = Z 2nτ
Wk∗2ng(r)Z t
= Z 2nτ
Z t
−Z 2nτ
Z 2nτ
0
0
0
t
t−r
Wk∗2ng(r)Z 2nτ
t
Wk∗2n(f ∗ g)(t)Sk∗2n (t)xdt
k∗2n(s + r − t)Wk∗2nf (s)dsdrSk∗2n(t)xdt
k∗2n(s + r − t)Wk∗2nf (s)dsdrSk∗2n(t)xdt.
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
17
By Fubini theorem, we obtain the four integral expressions
Gk(f ∗ g)x = Z 2nτ
0
Wk∗2nf (s)Z s+r
r
k∗2n(s + r − t)Sk∗2n(t)xdtdsdr
r
s
0
0
k∗2n(s + r − t)Sk∗2n(t)xdtdsdr
Wk∗2nf (s)Z s+r
Wk∗2nf (s)Z s
k∗2n(s + r − t)Sk∗2n(t)xdtdsdr
Wk∗2nf (s)Z r
k∗2n(s + r − t)Sk∗2n(t)xdtdsdr
k∗2n(s + r − t)Sk∗2n(t)xdt(cid:19) dsdr
Wk∗2nf (s)(cid:18)Z s+r
−Z s
k∗2n(s + r − t)Sk∗2n(t)xdt(cid:19) dsdr
Wk∗2nf (s)(cid:18)Z s+r
−Z r
0
0
0
s
r
r
0
0
0
0
Wk∗2ng(r)Z 2nτ
Wk∗2ng(r)Z r
Wk∗2ng(r)Z 2nτ
Wk∗2ng(r)Z r
+Z 2nτ
−Z 2nτ
−Z 2nτ
Wk∗2ng(r)Z r
Wk∗2ng(r)Z nτ
Wk∗2ng(r)Sk∗2n(r)Z r
Wk∗2ng(r)Sk∗2n(r)Z nτ
Wk∗2ng(r)Sk∗2n(r)Z 2nτ
0
0
0
r
r
0
0
= Z 2nτ
+Z 2nτ
= Z 2nτ
+Z 2nτ
= Z 2nτ
0
0
0
Wk∗2nf (s)Sk∗2n(s)xdsdr
Wk∗2nf (s)Sk∗2n(s)xdsdr
Wk∗2nf (s)Sk∗2n(s)xdsdr = Gk(g)Gk(f )x,
where we have applied the formula (4.12) and the part (ii) is shown.
Now we consider f ∈ Dk∗∞, supp(f ) ⊂ [0, nτ ] and x ∈ X. We apply the formulae (3.7)
and (4.11) to get
0
AGk(f )x = = −AZ nτ
= − Z nτ
= −Gk∗n(f′)x −Z nτ
(Wk∗nf )′(t)Z t
Wk∗nf′(t)(cid:18)Sk∗n(t)x −Z t
0
0
0
0
Sk∗n(s)xdsdt
k∗n(s)dsx(cid:19) dt
Wk∗nf (t)k∗n(t)dtx = −Gk∗n(f′)x − f (0)x,
and the part (iii) is proven.
(cid:3)
The previous theorem allows to show as a consequence one of main results in [18].
Remark 5.2. When the operator A generates a global convoluted semigroup (Sk(t))t≥0,
the homomorphism Gk is defined from Dk to B(X) by
Gk(f )x =Z ∞
0
Wkf (t)Sk(t)xdt,
x ∈ X,
f ∈ Dk,
see [18, Theorem 5.5]. Under some conditions of the boundedness of (Sk(t))t≥0, the ho-
momorphism Gk may be extended to a bounded Banach algebra homomorphism, see [18,
Theorem 5.6].
18
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Let k, l ∈ L1
loc([0, τ )) with 0 ∈ supp(k), and (Sk(t))t∈[0,τ ) a non-degenerate local k-
convoluted semigroup generated by A. Then (l ∗ Sk(t))t∈[0,τ ] is a non-degenerate local
k-convoluted semigroup generated by A, see a similar proof in [18, Proposition 5.2].
Corollary 5.3. Let k, l ∈ L1
non-degenerate local k-convoluted semigroups generated by A. Then
loc(R+) with 0 ∈ supp(k)∩supp(l), and (Sk(t))t∈[0,τ ), (Sk∗l(t))t∈[0,τ )
f ∈ D(k∗l)∗∞,
Gk∗l(f ) = Gk(f ),
where Gk(f ),Gk∗l are defined in Theorem 5.1.
Proof. Take f ∈ D(k∗l)∗∞ and supp(f ) ⊂ [0, nτ ]. By (3.8), f ∈ D(k∗l)∗∞ ⊂ Dk∗∞ and
Wk∗n∗l∗nf (t)Sk∗n∗l∗n(t)xdt
(l∗n ◦ Wk∗n∗l∗nf )(t)Sk∗n(t)xdt
W(k∗l)∗nf (t)S(k∗l)∗n(t)xdt =Z nτ
Wk∗n∗l∗nf (t)(l∗n ∗ Sk∗n)(t)xdt =Z nτ
0
0
0
0
Gk∗l(f )x =Z nτ
=Z nτ
=Z nτ
0
Wk∗nf (t)Sk∗n(t)xdt = Gk(f )x
where we have applied the formula (3.8).
(cid:3)
Corollary 5.4. Let (Sα(t))t∈[0,τ ] a non-degenerate local α-times integrated semigroup gen-
erated by A. We define the map Gα : D+ → B(X) by
Wαnf (t)Sαn(t)xdt,
x ∈ X, f ∈ D+,
Gα(f )x :=Z nτ
0
where supp(f ) ⊂ [0, nτ ] and (Snα(t))t∈[0,nτ ] is defined in Theorem 4.4 for some n ∈ N.
Then the map Gα is well defined, linear, bounded and Gk(f ∗ g) = Gk(f )Gk(g) for f, g ∈
Dk∗∞. Moreover, Gk(f )x ∈ D(A) and
AGk(f )x = −Gk(f′)x − f (0)x,
f ∈ D+,
x ∈ X.
6. Examples and applications
In this section we consider different examples of convoluted semigroups which have been
presented in the literature. Our results are applied in all these examples to illustrate its
importance.
6.1. Differential operators on Lp(RN ). [1, 3, 14] Let E be one of the spaces Lp(Rn) (1 ≤
p ≤ ∞), C0(Rn), BU C(Rn), or Cb(Rn) and AE the associated operator to a differential
operator and defined by Fourier multipliers, see details in [14, Section 4], [3, Chapter
8]. Under some conditions, the operator AE generates an α-times integrated semigroup
(Sα(t))t≥0 on E and kSα(t)k ≤ CE(1 + tα) for some constant CE and certain α > 0, see
In this case the map Gα : D+ → B(Lp(RN ))
[1, Theorem 6.3] and [14, Theorem 4.2].
(Corollary 5.4) extends to a Banach algebra homomorphism Gα : Tα(1 + tα) → B(Lp(RN ))
where Tα(1 + tα) is the completion of D+ in the norm
W αf (t)(1 + tα)dt,
f ∈ D+,
kfkα,0,α :=Z ∞
0
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
19
where Wα is the Weyl derivation of order α, see [28, Proposition 4.7]. Other examples of
global α-times integrated semigroups may be found in [10, Theorem 5.3] and [15, Propo-
sition 8.1]. Earlier results on the Schorodinger equation using distribution semigroups can
be found in [5] where differential operators are treated using Fourier multipliers in Lp
spaces.
6.2. Multiplication local integrated semigroup in ℓ2. [27, Example 1.2.6], [29, Ex-
ample 1]. Let ℓ2 be the Hilbert space of all complex sequences x = (xm)m=1 such that
∞Xm=1
xm2 < ∞,
with the Euclidean norm kxk :=(cid:0)P∞m=1 xm2(cid:1) 1
T(cid:17)2! 1
+ i (cid:18) em
m(cid:19)2
−(cid:16) m
(t − s)α−1eamsxmds(cid:19)m=1
Γ(α)Z t
where i2 = −1. For any α > 0 let (Sα(t))t>0 be defined by
Sα(t)x =(cid:18) 1
, m ∈ N,
am =
m
T
0
,
2 . Take T > 0 and define
2
for x ∈ D(Sα(t)) where D(Sα(t)) = {x ∈ ℓ2 ; Sα(t)x ∈ ℓ2}. Then (Sα(t))t∈[0,αT ) is a
local α-times integrated semigroup on ℓ2, (Sα(t))t∈[0,αT ) ⊂ B(ℓ2), such that (Sα(t))t∈[0,αT )
cannot be extended to t ≥ αT , see [29, Example 1]. We may apply the Corollary 4.7 to
define (Snα(t)) for t < nαT and Corollary 5.4 to define the map Gα : D+ → B(X) by
Gα(f )x :=Z nτ
0
Wαnf (t)Sαn(t)xdt,
x ∈ X, f ∈ D+,
with supp(f ) ⊂ [0, nτ ]. Other examples of local integrated semigroups defined by multi-
plication may be found in [2, Example 4.4 (c), (d)].
6.3. The Laplacian on L2[0, π] with Dirichlet boundary conditions. [4, Section 3,
5], [20, Example 6.1]. The operator −∆ on L2[0, π] with Dirichlet or Neumann boundary
conditions generates a polynomially bounded κ-convoluted semigroups (Sκ(t))t≥0, where
kSκ(t)k ≤ C(1 + t3) where C > 0; note that κ is given in Example 3.8 and κ(t) ≤ Ceβt
for t ≥ 0 and some C, β > 0. By Remarks 5.2, there exists an algebra homomorphism
Gκ : Dκ → B(L2[0, π]) such that it extends to a Banach algebra homomorphism Gκ :
Tκ(eβ) → B(L2[0, π]), see Theorem 3.6 and [18, Theorem 6.5].
Other examples of generators of k-convoluted semigroups (which does not generate
integrated semigroups) may be found in [20, Example 6.2], or [19, Section 2.8], where
k(t) =
e−a2/4t
2√πt3
=
1
2πiZ r+i∞
r−i∞
eλt/a2−
√λdλ,
t > 0,
for some a > 0 (where in the integral, r is any positive real number). In this case, we
have:
1
a
e−a√λ, ℜλ > 0.
bk(λ) =
20
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
6.4. Ultradistributions in the Gevrey classes. [9, Section 5. Applications ], [8, 4.
Example and final comments]. Let Mk, k = 0, 1, 2, ... be a Gevrey type sequences, i.e.,
a sequence of positive numbers such that M0 = 1, which is logarithmically convex and
non-quasianalytic:
M 2
k ≤ Mk−1Mk+1,
and
∞Xk=1
Mk−1/Mk < ∞,
for example (k!s), (kks) and Γ(1 + ks) for s > 1.Let mk = Mk/Mk−1 for k ∈ N,
P (z) =
∞Yj=1(cid:18)1 +
z
mj(cid:19) ,
ℜz > 0,
(i) Let X = L2(R) and A = i d4
dt4 − d2
differential polynomial. The following two operators are generators of local K-convoluted
semigroups and appear in [8, 4. Example and final comments]:
and the function K defined by bK(z) = 1/P (z). The entire function P (z) is called an ultra-
j2(cid:17) for some l > 0.
on [0, τ ), where bK1(z) = 1/P1(z) where P1(z) =Q∞j=1(cid:16)1 + lz
In the context of ultradistribution semigroups, this example was first proposed
by J. Chazarain ([7, Remarque 6.4]). It was observed in [16] that one can take A =
iB where B is the generator of a strongly continuous cosine function. Additional
results are given in [19], [20] including the case of Beurling ultradistributions.
Other developments are discussed in [19], including a generalization of the abstract
Weierstrass formula.
dt2 . Then A generates an K1-convoluted semigroup
(ii) Let ω be a σ-finite measure, the Lebesgue space X = Lp(Ω) (1 ≤ p ≤ ∞) and
m : Ω → C a measurable function. We define Af = mf where D(A) = {f ∈
Lp(Ω); mf ∈ Lp(Ω)} and
{z ∈ C;ℜz ≥ αza + β} ⊂ ρ(A).
for α, β > 0 and 0 < a < 1. For every τ > 0 there is l > 0 such that A generates a
local K2-convoluted semigroup on [0, τ ) where
P2(z) =
∞Yj=1 1 +
a! ,
lz
1
j
where bK2(z) = 1/P2(z).
The following estimates are valid for the Gevrey sequences Mj = (j!s), Mj = (jjs) and
Mj = Γ(1 + js) where s > 1 is given.
e(lz)a
≤ P (z) ≤ e(Lz)a
, ℜz ≥ 0,
where L is a positive constant, see e.g. [8, (1.1)].
We apply the Theorem 4.4 to conclude that A generates a local K∗n-convoluted semi-
group on [0, nτ ).
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
21
7. Appendix: k-Distribution Semigroups
Let X be a Banach space. Vector valued algebraic distributions, i.e., linear and continu-
ous maps from a test function space to the space of bounded linear operators, B(X), which
satisfy an algebraic property (similar to Theorem 5.1 (ii)) have been studied deeply in a
large numbers of papers, see [13, Chapter 8 ]. In this sense, distribution semigroups (in
the sense of Lions, DS-L) were introduced by J.L. Lions in [25], see also [7], [2, Definition
7.1]. P.C. Kunstmann considered pre-distribution semigroups (or quasidistribution semi-
group in the terminology of S. W. Wang) as linear and continuous maps G : D 7→ B(X),
G ∈ D′(B(X)), satisfying
see [21, Definition 2.1] and [35, Definition 3.3]. In fact, a pre-distribution G can be regarded
as a continuous linear map from D+ into B(X), G : D+ 7→ B(X), such that
(i) G(φ ∗ ψ) = G(φ)G(ψ) for φ, ψ ∈ D,
(ii) ∩{ker(G(θ)) θ ∈ D0} = {0},
(i) G(φ ∗ ψ) = G(φ)G(ψ) for φ, ψ ∈ D+.
(ii) ∩{ker(G(θ)) θ ∈ D+} = {0}.
see [35, Remark 3.4]. The differences between quasi-distribution and distribution semi-
groups in the sense of Lions may be found in [21, Remark 3.13].
Classes of Distribution semigroups on (0,∞) (in short DS on (0,∞)) were considered
in [22, Definition 1]. The subspace D′+(B(X)) ⊂ D′(B(X)) is formed of the elements
supported in [0,∞). A distribution semigroup on (0,∞), G ∈ D′+(B(X)), is a continuous
linear map from D into B(X) G : D 7→ B(X), supported in [0,∞), such that
(i) G(φ ∗ ψ) = G(φ)G(ψ) for φ, ψ ∈ D0.
(ii) ∩{ker(G(θ)) θ ∈ D0} = {0}.
Distribution semigroups on (0,∞) and quasi-distribution semigroups are not the same
concept, see [22, Remark 4]. However, a particular class of distribution semigroups on
(0,∞) (simply called distribution semigroups, see [22, Definition 2]) may be identified
with quasi-distribution semigroups, [22, Theorem 1].
Keeping in mind these definitions and Theorem 5.1, we introduce the concept of k-
distribution semigroup.
loc(R+) such that 0 ∈ supp(k). We say that a linear and
Definition 7.1. Let k ∈ L1
continuous map Gk : Dk∗∞ 7→ B(X) is a k-distribution semigroup, in short k-DS, if it
satisfies the following conditions.
(i) Gk(φ ∗ ψ) = Gk(φ)Gk(ψ) for φ, ψ ∈ Dk∗∞.
(ii) ∩{ker(Gk(θ)) θ ∈ Dk∗∞} = {0}.
In the case that Gk is a k-distribution semigroup on (0,∞), then Dk∗∞ 6= {0} by (ii).
Remark 7.2. Let G : D → B(X) be a distribution semigroup (or quasi-distribution semi-
group in the sense of Wang). Then G◦Λ is a k-distribution semigroup for any k ∈ L1
loc(R+)
such that 0 ∈ supp(k) and Dk∗∞ = D+; in particular G ◦ Λ is a jα − DS for any α > 0 (see
definition of Λ and jα in Section 3).
For a given k-DS Gk, define the operator A′ by
(i) D(A′) := ∪{Im(Gk(θ)) θ ∈ Dk∗∞}.
(ii) A′Gk(θ)x := −Gk(θ′)x − θ(0)x, for x ∈ X and θ ∈ Dk∗∞.
22
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
Proposition 7.3. The operator A′ is well defined and closable.
Proof. Assume that Gk(φ)x = Gk(ψ)y for some x, y ∈ X and φ, ψ ∈ Dk∗∞. Now take
θ ∈ Dk∗∞. Since
θ ∗ φ′(t) = θ′ ∗ φ(t) + θ(0)φ(t) − φ′(0)θ(t),
t ≥ 0,
see, for example [35, Proposition 3.1 (iii)], we get that
and hence
Similarly,
Gk(θ)Gk(φ′)x = Gk(θ′)Gk(φ)x + θ(0)Gk(φ)x − φ′(0)Gk(θ)x
(cid:0)−Gk(θ′) − θ(0)(cid:1) Gk(φ)x = Gk(θ)(cid:0)−Gk(φ′)x − φ′(0)x(cid:1) .
(cid:0)−Gk(θ′) − θ(0)(cid:1)Gk(ψ)y = Gk(θ)(cid:0)−Gk(ψ′)y − ψ′(0)y(cid:1) .
−Gk(φ′)x − φ′(0)x = −Gk(ψ′)y − ψ′(0)y
By Definition 7.1 (ii), we conclude that
and A′ is well defined.
To prove that A′ is closable, let (xn)n≥1 ⊂ D(A′) be such that xn → 0, A′xn → y. We
write xn = Gk(φn)zn with (φn)n≥1 ⊂ Dk∗∞ and (zn)n≥1 ⊂ X. Take θ ∈ Dk∗∞ and then
Gk(θ)y = lim
= lim
n→∞Gk(θ)A′xn = lim
n→∞Gk(θ)A′Gk(φn)zn
n→∞Gk(θ)(cid:0)−Gk(φ′n)zn − φn(0)zn(cid:1)
n→∞(cid:0)−Gk(θ′) − θ(0)(cid:1)Gk(φn)zn = lim
= lim
n→∞(cid:0)−Gk(θ′) − θ(0)(cid:1) xn = 0.
This implies that y = 0 by Definition 7.1 (ii).
Definition 7.4. The closure of A′, denoted by A, is called the generator of a Gk.
(cid:3)
Ax = y,
Other definitions of generator of distribution semigroups are given using approximate
units (see [2, Definition 7.1]) or the distribution −δ′0 ([21, Definition 3.3] and [22, Propo-
sition 1]). In our case, given a k-DS Gk and its generator (A, D(A)), then
D(A) ⊂ {x ∈ X exists y ∈ X such that Gk(θ)y = −Gk(θ′)x − θ(0)x for any θ ∈ Dk∗∞};
x ∈ D(A),
loc(R+) with 0 ∈ supp(k), (Sk(t))t∈[0,τ ] a non-degenerate local k-
Theorem 7.5. Let k ∈ L1
convoluted semigroup generated by A and Gk : Dk∗∞ → B(X) the map defined in Theorem
5.1. Then Gk is a k-DS generated by A.
Proof. The first condition in Definition 7.1 appears in Theorem 5.1 (ii). Take x ∈ X such
that Gk(φ)(x) = 0 for any φ ∈ Dk∗∞. Since φ′ ∈ Dk∗∞, we apply Theorem 5.1 (iii) to
conclude 0 = φ(0)x for any φ ∈ Dk∗∞ and then 0 = φu(0)x for any u ≥ 0. We conclude
that 0 = φ(u)x for any u ≥ 0 and φ ∈ Dk∗∞. Then x = 0 and the condition Definition 7.1
(ii) holds.
Note that AGk(θ)x := −Gk(θ′)x− θ(0)x, for x ∈ X and θ ∈ Dk∗∞, see Theorem 5.1 (iii).
As A is a closed operator, we conclude that A is the generator of Gk.
A straightforward consequence of Theorem 7.5 is the next corollary.
(cid:3)
SHARP EXTENSIONS FOR SOLUTIONS OF ABSTRACT CAUCHY PROBLEMS
23
Corollary 7.6. Let (Sα(t))t∈[0,τ ] a non-degenerate local α-times integrated semigroup gen-
erated by A and Gα : D+ → B(X) the map defined in Corollary 5.4. Then Gα is a jα-DS
generated by A with α > 0.
When the generator A is densely defined in Corollary 7.6, it is equivalent that A gener-
ates a n-times integrated semigroup for some n ∈ N and A is the generator of a distribution
semigroup in the sense of Lions, see [2, Theorem 7.2, Corollary 7.3].
To consider the test-function space Dk∗∞ may be given for a wide class of vector-
valued distributions such that different distribution semigroups fall into the scope of this
approach.
References
[1] W. Arendt, H. Kellermann, Integrated solutions of Volterra integrodifferential equations and appli-
cations, in: Volterra Integrodifferential Equations in Banach spaces and Applications, Trento, 1987,
in: Pitman Res. Notes Math. Ser., vol. 190, Longman sci. Tech., Harlow, 1989, pp. 21-51.
[2] W. Arendt, O. El-Mennaoui and V. Keyantuo, Local integrated semigroups: Evolution with jumps
of regularity. J. Math. Anal. Appl. 186 (1994), 572-595.
[3] W. Arendt, C. Batty, M. Hieber, F. Neubrander. "Vector-valued Laplace Transforms and Cauchy
Problems". Monographs in Mathematics. vol.96. Birkhauser, Basel, 2001.
[4] B. Baumer, Approximate solutions to the abstract Cauchy problems, in: Evolution Equations and
their Application in Physical and Life Sciences, Lecture Notes in Pure and Applied Mathematics Vol
215(Marcel Dekker, New York, 2001), 33 -- 41.
[5] M. Balabane, H. A. Emamirad, Lp estimates for Schrodinger evolution equations. Trans. Amer.
Math. Soc. 292(1985), 357-373.
[6] J. A. van Casteren, "Generators of strongly continuous semigroups". Research Notes in Mathematics,
115. Boston-London-Melbourne: Pitman Advanced Publishing Program, 1985.
[7] J. Chazarain, Probl`emes de Cauchy abstraits et applications `a quelques probl`emes mixtes, J. Funct.
Anal. 7 (1971), 386-446.
[8] I. Cioranescu, Local convoluted semigroups, in: Evolution Equations (Baton Rouge, LA, 1992),
107 -- 122, Dekker, New York, 1995.
[9] I. Cioranescu, G. Lumer, Probl`emes d'´evolution r´egularis´es par un noyau g´en´eral K(t). Formule de
Duhamel, prolongements, th´eor`emes de g´en´eration, C. R. Acad. Sci. Paris S´er. I Math. 319 (1995),
1273 -- 1278.
[10] G. Carron, T. Coulhon and E.-L. Ouhabaz, Gaussian estimates and Lp-boundedness of Riesz means,
J. Evol. Equ. 2 (2002) 299-317.
[11] J. M. C. Duhamel, M´emoire sur la m´ethode g´en´erale relative au mouvement de la chaleur dans les
corps solides plong´es dans les milieux dont la temp´erature varie avec le temps. J. Ec. Polyt. Paris
14, Cah. 22, 20 (1833).
[12] R.M. Dubois and G. Lumer, Formule de Duhamel abstraite, Arch. Math. 43 (1984) 49-56.
[13] H.O. Fattorini, "The Cauchy Problem", Addison-Wesley Publishing Co., Reading, Mass., 1983.
[14] M. Hieber: Integrated semigroups and differential operators on Lp(RN ), Math. Ann. 291 (1991),
1-16.
[15] C. Kaiser and L. Weis, Perturbation theorems for α-times integrated semigroups, Arch. Math. 81
(2003) 215-228.
[16] V. Keyantuo, Integrated semigroups and related partial differential equations, J. Math. Anal. Appl.
212(1997) 135-153.
[17] V. Keyantuo, C. Muller and P. Vieten, The Hille-Yosida theorem for local convoluted semigroups.
Proc. Edinb. Math. Soc. 46 (2003), 357-372.
[18] V. Keyantuo, C. Lizama, P. Miana, Algebra homomorphisms defined via semigroups and cosine
functions, J. Funct. Anal. 257 (2009) 3454-3487.
[19] M. Kosti´c, "Generalized semigroups and cosine functions". Mathematical Institute, Belgrade, 2011.
[20] M. Kosti´c and S. Pilipovi´c, Global convoluted semigroups. Math. Nachr. 15 (2007), 1727-1743.
24
VALENTIN KEYANTUO, PEDRO J. MIANA, AND LUIS S ´ANCHEZ-LAJUSTICIA
[21] P. C. Kunstmann, Distribution semigroups and abstract Cauchy Problem. Trans. Amer. Math. Soc
351 (1999), 837 -- 856.
[22] P. C. Kunstmann, M. Mijatovic and S. Pilipovic, Classes of distribution semigroups. Studia Math.
187 (2008), 37 -- 58.
[23] C.C. Kuo and S.Y. Shaw, On α-times integrated C-semigroups and the abstract Cauchy problem.
Studia Math. 142 (2000), 201 -- 217.
[24] Y-C. Li and S-Y Shaw, On local α-times integrated C-semigroup, Abst. Appl. Anal. 207 (2007) 1-18,
doi 10.1155/2007/34890.
[25] J. L. Lions, Les semi-groupes distributions, Portugalia Math. 19 (1960) 141-164.
[26] J. L. Lions and E. Magenes, "Non-homogeneous boundary value problems and applications". Vol.
III., Die Grundlehren der mathematischen Wissenschaften, Band 183. Springer-Verlag, New York-
Heidelberg, 1973.
[27] Melnikova, I., and A. Filinkov:
"Abstract Cauchy problems:
three approaches", Chapman-
Hall/CRC, New York, 2001.
[28] P.J. Miana, α-Times integrated semigroups and fractional derivation, Forum Math. 14 (2002), 23-46.
[29] P.J. Miana, Local and global solutions of well-posed integrated Cauchy problems. Studia Math. 187
(2008), 219-232.
[30] P. J. Miana and V. Poblete, Sharp extensions for convoluted solutions of wave equations, Preprint,
2013.
[31] S. G. Samko, A. A. Kilbas and O. I. Marichev: "Fractional Integrals and Derivatives. Theory and
Applications". Gordon-Breach, New York (1993).
[32] R. T. Seeley, Extension of C∞ functions defined in a half space. Proc. Amer. Math. Soc. 15 (1964),
625-626.
[33] N. Tanaka, N. Okazawa, Local C-semigroups and local integrated semigroups. Proc. London Math.
Soc. 61 (1990), 63-90.
[34] S. Umarov, On fractional Duhamels principle and its applications, J. Differential Equations 252
(2012), 5217-5234.
[35] S. W. Wang, Quasi-distribution semigroups and integrated semigroups. J. Funct. Anal. 146 (1997),
352-381.
[36] S. W. Wang and M.C. Gao, Automatic extensions of local regularized semigroups and local regular-
ized cosine funtions. Proc. Amer. Math. Soc. 127(6) (1999), 1651-1663.
University of Puerto Rico, Department of Mathematics, Faculty of Natural Sciences,
Box 70377, San Juan PR 00936-8377 U.S.A.
E-mail address: [email protected]
Universidad de Zaragoza, Departamento de Matem´aticas & I.U.M.A., 50.009, Zaragoza.
E-mail address: [email protected]
Universidad de Zaragoza, Departamento de Matem´aticas & I.U.M.A., 50.009, Zaragoza.
E-mail address: [email protected]
|
1211.4329 | 1 | 1211 | 2012-11-19T08:18:40 | Dimension free boundedness of Riesz transforms for the Grushin operator | [
"math.FA"
] | Let $G = - \Delta_{\xi} - |\xi|^2 \frac{\partial^2}{\partial \eta^2}$ be the Grushin operator on $\R^n \times \R.$ We prove that the Riesz transforms associated to this operator are bounded on $L^p (\R^{n+1}), 1 < p < \infty$ and their norms are independent of the dimension $n$. | math.FA | math | Dimension free boundedness of
Riesz transforms for the Grushin operator
P. K. Sanjay and S. Thangavelu
Abstract. Let G = −∆ξ − ξ2 ∂ 2
∂η2 be the Grushin operator on Rn × R. We prove
that the Riesz transforms associated to this operator are bounded on Lp(Rn+1), 1 <
p < ∞ and their norms are independent of the dimension n.
1. Introduction
We consider the Grushin operator G = −∆ξ − ξ2 ∂2
∂η2 on Rn × R which can be
written formally as
2
1
0
2
v
o
N
9
1
]
.
A
F
h
t
a
m
[
1
v
9
2
3
4
.
1
1
2
1
:
v
i
X
r
a
where
and
where
is the scaled Hermite operator on Rn and
is the inverse Fourier transform of f in the η variable. In the light of this decomposition,
we can define the Riesz transforms associated to the operator G as
Gf (ξ, η) =
1
2πZR
e−iληH(λ)f λ(ξ)dλ
H(λ) = −∆ + λ2ξ2
f λ(ξ) =ZR
f (ξ, η)eiληdη
Rjf (ξ, η) =Z ∞
R∗j f (ξ, η) =Z ∞
−∞
−∞
e−iληRj(λ)f λ(ξ)dλ
e−iληR∗j (λ)f λ(ξ)dλ
for j = 1, 2, 3, . . . , n. Here,
Rj(λ) = Aj(λ)H(λ)− 1
2 , R∗j (λ) = Aj(λ)∗H(λ)− 1
2
are the Riesz transforms associated to the Hermite operator which has the decompo-
sition
1
2
H(λ) =
Xj=1
Aj(λ) = −
n
(Aj(λ)Aj(λ)∗ + Aj(λ)∗Aj(λ))
∂
∂ξj
+ λξj, Aj(λ)∗ =
∂
∂ξj
+ λξj
1991 Mathematics Subject Classification. 42C, 42C05 , 43A65.
Key words and phrases. Riesz Transforms, Grushin operator, Hermite operator, Transference
method.
1
2
P. K. SANJAY AND S. THANGAVELU
are the creation and annihilation operators on Rn. The boundedness of these Riesz
transforms on Lp(Rn+1) is known [JST, BG]. In [BG], the authors have proved
the boundedness of Riesz transforms associated to a much larger class of smooth
locally subelliptic diffusion operators on smooth connected non-compact manifold.
In [JST] these Riesz transforms were treated as operator valued Fourier multipli-
ers for Lp(Rn) valued functions on R and the boundedness was proved with the aid
of a result of L. Weis [W] on the operator valued Fourier multipliers. The aim
of this paper is to prove the following theorem about the dimension free bound-
edness of the vector of Riesz transforms Rf . That is, we consider the operator
R = (R1, R2,· · · , Rn, R∗1, R∗2,· · · , R∗n) with
Rf (ξ, η) = n
Xj=1
Rjf (ξ, η)2 +
n
Xj=1
and prove:
R∗j f (ξ, η)2!1/2
Theorem 1.1. For each 1 < p < ∞, there exists a constant Cp independent of the
dimension n, such that for all f ∈ Lp(Rn × R),
kRfkp ≤ Cpkfkp.
E. M. Stein introduced the notion of the dimension free boundedness for the vec-
tor of Riesz transforms in [S1]. Stein considered the Euclidean Riesz transforms and
proved the result using the technique of g-functions. Later this technique was adapted
for a non-commutative situation by Harboure, et al.
[HRST] to prove a similar re-
sult for the Riesz transforms associated to the Hermite operator. Alternate proofs
for the Euclidean Riesz transforms were provided using the method of rotations by
J. Duoandikoetxea and J. L. Rubio de Francia [DR] and using a transference argu-
ment by Gilles Pisier [P]. Similar techniques have been used to prove the dimension
free boundedness of Riesz transforms in other contexts as well. Coulhon, Muller and
Zienkiewicz [CMZ] proved the boundedness for the Riesz transforms associated to
the sub-Laplacian on the Heisenberg group.
In two works [LP2] and [LP1] Fran-
coise Lust-Piquard provided alternate proofs for the Riesz transforms associated to
the Hermite operator and the Riesz transforms associated to the sub-Laplacian. Our
proof follows a method very similar to that of [CMZ],[LP1] and [LP2]. The method
of rotations applied to the Euclidean Riesz transforms [DR] involves expressing the
Riesz transforms as the average of certain directional Hilbert transforms and using the
boundedness (independent of the direction) of these directional Hilbert transforms.
An expression similar to this in the Heisenberg group context is obtained in [CMZ].
In this paper we obtain a similar representation for Riesz transforms associated to
Grushin operators .
We also note that Theorem 1.1 in turn implies that the Riesz transforms associated
to the Hermite operator satisfy dimension free bounds. Since the Riesz transform for
the Grushin operator is an operator valued multiplier for the Fourier transform on
R, the corresponding multiplier operator for the Fourier series is also bounded by
the same norm. This can be proved using a generalisation of a transference result
of Karel de Leeuw [L]. The proof is similar to the proof in [ST] of the boundedness
of the Riesz transforms on the reduced Heisenberg group using transference from the
Heisenberg group. We refer to Theorem 2.1 and Theorem 2.2 of [ST] for details. Then
the proof of the Hermite Riesz transforms follows by looking at functions of the form
F (ξ, η) = f (ξ)eikη on Rn × [0, 2π).
RIESZ TRANSFORMS
3
2. Riesz transforms for the Grushin operator
We first consider the individual Riesz transforms Rj and R∗j for j = 1, 2,· · · , n and
show that they satisfy dimension free bounds on Lp(Rn+1) for 1 < p < ∞. In order to
j and R∗ǫ
do this we introduce the operators Rǫ
j which we will call the truncated Riesz
transforms. We only give details of Rǫ
j as the other one is similar. Note that the Riesz
transform Rj(λ) associated to the Hermite operator H(λ) can be written as
Rj(λ) = Aj(λ)H(λ)−1/2 =
e−rH(λ)r−1/2dr.
Aj(λ)
√π Z ∞
0
Here e−rH(λ) is the Hermite semigroup. For ǫ > 0, we define the truncated Riesz
transforms Rǫ
j(λ) by
Rǫ
j(λ) =
e−rH(λ)r−1/2dr.
Then the truncated Riesz transforms Rǫ
j for the Grushin operator are defined as
Aj(λ)
ǫ2
√π Z 1/ǫ2
2πZ ∞
−∞
1
Rǫ
jf (ξ, η) =
We first prove
e−iληRǫ
j(λ)f λ(ξ)dλ.
Proposition 2.1. For every f ∈ L2(Rn+1), Rǫ
Proof. It follows from the definition that
jf (ξ, η) − Rjf (ξ, η)2dξ dη =ZR(cid:18)ZRn Rǫ
ZRZRn Rǫ
The proposition follows once we show that for every λ ∈ R∗
jf → Rjf in L2(Rn+1) as ǫ → 0.
j(λ)f λ(ξ) − Rj(λ)f λ(ξ)2dξ(cid:19) dλ.
and
ZRn Rǫ
ZRn Rǫ
j(λ)f λ(ξ) − Rj(λ)f λ(ξ)2dξ → 0 as ǫ → 0
j(λ)f λ(ξ) − Rj(λ)f λ(ξ)2dξ ≤ 4ZRn f λ(ξ)2dξ.
To see these, we expand f λ in terms of scaled Hermite functions Φλ
definition) and use the fact that
α (see [T] for
Aj(λ)Φλ
α = (2αj + 2)1/2λ1/2Φλ
α+ej .
where ej is the canonical unit vector of Rn with 1 in the jth entry and zero elsewhere.
Then
Rǫ
(2αj + 2)1/2λ1/2
(cid:18)ZAǫ
j(λ)f λ(ξ)−Rj(λ)f λ(ξ) = Xα∈Nn
where Aǫ = (0, ǫ2) ∪ (1/ǫ2,∞). From this, it follows that
2α + n(f λ, Φλ
j(λ)f λ − Rj(λ)f λk2
2αj + 2
kRǫ
√π
2 =
1
π Xα∈Nn
e−(2α+n)λrr−1/2dr(cid:19) (f λ, Φλ
α)Φλ
α+ej .
α)2 ZAǫ,α
e−rr−1/2dr!2
.
where Aǫ,α = (0, (2α + n)λǫ2) ∪ ((2α + n)λǫ−2,∞). From the above equation it
is clear that
kRǫ
j(λ)f λ − Rj(λ)f λk2 ≤ 2kf λk2
4
P. K. SANJAY AND S. THANGAVELU
We also note that when
f λ(ξ) = X(2α+n)λ≤N
is a finite linear combination of Φλ
α,
(f λ, Φλ
α)Φλ
α(ξ)
kRǫ
j(λ)f λ − Rj(λ)f λk2
α)2(cid:18)Z(0,N ǫ2)∪(N ǫ−2,∞)
which goes to 0 as ǫ → 0. As such functions are dense in L2(Rn), we get
π X(2α+n)λ≤N
(f λ, Φλ
2 ≤
1
e−rr−1/2dr(cid:19)2
kRǫ
j(λ)f λ − Rj(λ)f λk2 → 0, as ǫ → 0
for any f ∈ L2(Rn+1).
For the individual Riesz transforms, we have the following result:
Theorem 2.2. For j = 1, 2, 3,· · · , n we have
kRjfkp + kR∗j fkp ≤ Cpkfkp, 1 < p < ∞
for all f ∈ Lp(Rn+1) where Cp is independent of the dimension.
In order to prove this result, we claim that it is enough to prove
(cid:3)
kRǫ
jfkp + kR∗ǫ
j fkp ≤ Cpkfkp
for the truncated Riesz transforms. A proof of this will be given in section 5 where
jf(cid:1) is Cauchy in Lp(Rn+1) and hence there exists an operator
we will show that (cid:0)Rǫ
Sj, bounded on Lp(Rn+1) such that Rǫ
jf → Sjf as ǫ → 0. In view of Proposition 2.1,
Rǫ
jf → Rjf for f ∈ L2(Rn+1) and hence Sjf = Rjf . This will prove the stated
boundedness of Rj ( and R∗j ). Now to prove the boundedness of Rǫ
j and R∗ǫ
j we
express these operators as a superposition of certain truncated Hilbert transforms.
3. A representation for the truncated Riesz transforms
The representation we obtain is very similar to the representation obtained in
[CMZ] for the Riesz transforms on the Heisenberg group Hn. Before stating this
result, we recall some definitions and notation. For further details and proofs we refer
to [T].
Recall that as a manifold Hn = Cn × R and hence we write (z, t), z = x + iy ∈
Cn, t ∈ R to denote the elements of Hn. The sub-Laplacian L on the Heisenberg
group Hn can be written as the sum of the differential operators Xj =(cid:16) ∂
∂t(cid:17)
and Yj =(cid:16) ∂
j ). The sub-Laplacian is homogeneous of
degree 2 with respect to the non-isotropic dilations δr(z, t) = (rz, r2t) of the Heisenberg
group. Hence ps(z, t), the heat kernel associated to L satisfies
∂t(cid:17) as L =Pn
∂yj − 1
j=1(X 2
j + Y 2
+ 1
2xj
2 yj
∂xj
∂
∂
pr2s(z, t) = r−(2n+2)ps(z/r, t/r2).
This also follows from the following explicit expression for qs(z, λ), the inverse Fourier
transform of ps(z, t) in the t variable,
qs(z, λ) =ZR
ps(z, t)eiλtdt = (4π)−n(cid:18) λ
sinh λs(cid:19)n
e− 1
4 λ coth(sλ)z2
.
RIESZ TRANSFORMS
5
By πλ, we denote the Schrodinger representation of the Heisenberg group Hn acting
on L2(Rn) in the following manner
πλ(x + iy, t)φ(ξ) = eiλteiλ(x.ξ+ 1
2 x.y)φ(ξ + y).
When t = 0, we will denote πλ(z, t) by πλ(z). πλ also defines a representation of the
group algebra L1(Hn) as
πλ(f ) =ZHn
f (z, t)πλ(z, t)dz dt = Wλ(f λ)
where Wλ(g) = RCn g(z)πλ(z)dz is the Weyl transform of g. The representation given
in [CMZ] for the Riesz transforms associated to the sub-Laplacian L is
Xip1(w, s)H(w,s)f (z, t)dw ds
1
(XiL− 1
2 f )(z, t) = −
4(2π)n+1 ZHn
where H(w,s) is the Hilbert transform along a curve in the Heisenberg group. For the
Grushin operator, we obtain a similar representation involving certain operators T (z,t)
and certain differential operators
ǫ
with
Zj = i Xj + Yj; Z∗j = i Xj − Yj
xj
2
∂t(cid:19) ; Yj =(cid:18) ∂
∂xj −
yj
2
∂yj
+
∂
Xj =(cid:18) ∂
∂
∂t(cid:19) .
Proposition 3.1.
Rǫ
jf (ξ, η) =
R∗ǫ
j f (ξ, η) =
f (ξ, η) =Zǫ<r<1/ǫ
1
1
f (ξ, η) Zjp1(z, t)dz dt
T (z,t)
ǫ
√π ZHn
√π ZHn
f(cid:16)ξ + ry, η + rx · ξ + r2(cid:16)t +
T (z,t)
ǫ
f (ξ, η) Z∗j p1(z, t)dz dt
x · y
2 (cid:17)(cid:17) dr
r
and
where
T (z,t)
ǫ
We follow the method in [LP2] in proving this proposition. First we prove the
following lemma which will be used in the proof of Proposition 3.1.
Lemma 3.2. For any φ ∈ S(Rn),
(i) r ∂
∂ξj
(πλ(rz)φ(ξ)) =(cid:16) ∂
(ii) irλξjπλ(rz)φ(ξ) =(cid:16) ∂
+ iλr2xj
∂xj − iλr2yj
∂yj
2 (cid:17) (πλ(rz)φ(ξ))
2 (cid:17) πλ(rz)φ(ξ).
Proof. We first look at the derivative of πλ(rz) with respect to the variable yj
and see that
∂
∂yj(cid:0)πλ(rz)φ(ξ)(cid:1) =
∂
∂yj (cid:16)eiλ(rx·ξ+ 1
2 r2x·y)φ(ξ + ry)(cid:17)
πλ(rz)φ(ξ) + eiλ(rx·ξ+ 1
=
iλr2xj
2
2 r2x·y) ∂
∂yj
φ(ξ + ry)
=
iλr2xj
2
πλ(rz)φ(ξ) + rπλ(rz)
∂φ
∂ξj
(ξ).
6
Writing
πλ(rz)
and noting that
P. K. SANJAY AND S. THANGAVELU
∂φ
∂ξj
(ξ) =(cid:18)πλ(rz)
∂
∂ξj −
∂
∂ξj
πλ(rz)(cid:19) φ(ξ) +
∂
∂ξj
πλ(rz)φ(ξ)
we see that
∂
∂φ
∂ξj
(ξ)
∂
∂ξj(cid:0)πλ(rz)φ(ξ)(cid:1) = iλrxjπλ(rz)φ(ξ) + πλ(rz)
∂yj(cid:0)πλ(rz)φ(ξ)(cid:1) =(cid:18)r
∂xj(cid:0)πλ(rz)φ(ξ)(cid:1) = iλrξjπλ(rz)φ(ξ) +
∂
∂ξj −
iλr2xj
iλr2yj
∂
2
2 (cid:19)(cid:0)πλ(rz)φ(ξ)(cid:1)
πλ(rz)φ(ξ)
which proves part (i). Part (ii) of the lemma follows from
(cid:3)
Since Rǫ
Now using the previous lemma,
Proof of Proposition 3.1. The heat kernels for the sub-Laplacian and the Her-
mite operator are related via the group Fourier transform on the Heisenberg group as
follows:
ps(z, t)πλ(z, t)dz dt = e−sH(λ)
We refer to [T] for a proof of this. Using the homogeneity of the heat kernel we get
ps(λ) =ZHn
e−r2H(λ) = r−(2n+2)ZHn
= r−2nZCn
r
∂ξj
Rǫ
R∗ǫ
1
1
t
,
r
∂
∂ξj
j(λ) =
j (λ) =
r2(cid:19) dz dt
,−λr2(cid:17) dz =ZCn
πλ(rz)q1(cid:0)z,−λr2(cid:1) dz.
j(λ) = 1√π Aj(λ)Rǫ<r<1/ǫ e−r2H(λ)dr, we see that
+ λξj(cid:19) πλ(rz)q1(cid:0)z,−λr2(cid:1) dz dr;
+ λξj(cid:19) πλ(rz)q1(cid:0)z,−λr2(cid:1) dz dr.
πλ(z, t)p1(cid:18)z
πλ(z)q1(cid:16)z
√πZǫ<r<1/ǫZCn(cid:18)−
√π Zǫ<r<1/ǫZCn(cid:18) ∂
λξjπλ(rz)q1(cid:16)z
,−λr2(cid:17) dz
2 (cid:19) πλ(rz)q1(z,−λr2)dz
∂xj −
2 (cid:19) q1(z,−λr2)dz
πλ(rz)(cid:18) ∂
2 (cid:19)(cid:18)ZR
πλ(rz)(cid:18) ∂
πλ(rz)(cid:18)ZR(cid:18) ∂
p1(z, t)eiλr2tdt(cid:19) dz
∂t(cid:19) p1(z, t)eiλr2tdt(cid:19) dz
irZCn
= ZCn(cid:18) ∂
= −ZCn
= −ZCn
= −ZCn
= −ZHn
πλ(rz) Xjp1(z, t)eiλr2tdz dt.
∂xj −
iλr2yj
iλr2yj
iλr2yj
∂
yj
2
∂xj
∂xj
+
+
r
RIESZ TRANSFORMS
7
∂ξj
rZCn(cid:18) ∂
= ZCn(cid:18) ∂
= −ZHn
∂yj
πλ(rz)(cid:19) q1(cid:16) z
r
iλr2xj
+
,−λr2(cid:17) dz
2 (cid:19) πλ(rz)q1(z,−λr2)dz
πλ(rz) Yjp1(z, t)eiλr2tdz dt.
πλ(rz) Zjp1(z, t)eiλr2tdz dt
πλ(rz) Z∗j p1(z, t)eiλr2tdz dt
dr
r
dr
r
.
Rǫ
R∗ǫ
j (λ) =
1
1
j(λ) =
√π Zǫ<r<1/ǫZHn
√πZǫ<r<1/ǫZHn
√π ZR(cid:18)Zǫ<r<1/ǫZHn
1
Similarly,
Hence
and
Thus
Rǫ
jf (ξ, η) =
πλ(rz)f λ(ξ) Zjp1(z, t)eiλr2tdz dt
dr
r (cid:19) eiληdλ.
Now the proposition follows from the fact that
ZR
πλ(rz)f λ(ξ)eiλr2teiληdλ = f(cid:16)ξ + ry, η + rx · ξ + r2(cid:16)t +
x · y
2 (cid:17)(cid:17) .
(cid:3)
4. A transference result
In the previous section, we obtained a representation for the Riesz transforms as
the superposition of certain operators T (z,t)
. To prove the boundedness of the vector of
Riesz transforms, using the method of rotations, we need to prove that these operators
are bounded uniformly in (z, t).
ǫ
Proposition 4.1. For 1 < p < ∞, there exists a constant Cp independent of
(z, t) ∈ Hn, ǫ > 0 and the dimension n such that,
ǫ
(cid:13)(cid:13)T (z,t)
f(cid:13)(cid:13)Lp(Rn+1) ≤ Cp kfkLp(Rn+1) .
Proof. This will be proved using Calder´on's method of transferring an operator
on Lp(X), for a measure space X, to Lp(G) when the group G acts on X by measure
preserving transformations. The measure space Rn+1, the group Hn and the group
action
U(x + iy, t)(ξ, η) = (ξ − y, η − t +
x · y
2 − x.ξ)
are all the same as used by Ratnakumar and Thangavelu in [RT] and we follow their
notation. Accordingly, for f ∈ Lp(Rn+1) and (ξ, η) ∈ Rn+1, we define the transferred
function F(ξ,η) on Hn as
F(ξ,η)(z, t) = f (U(z, t)(ξ, η)) .
For T ∈ B(Lp(Hn)), the transferred operator T0 on B(Lp(Rn+1)) is defined as
T0f (ξ, η) = (T F(ξ,η))(0).
8
P. K. SANJAY AND S. THANGAVELU
For a curve γ = {γ(t) ∈ Hn, t ∈ R} and a function f on Hn, the Hilbert transform of
f along γ is defined as
Hγf (w, s) =ZR
f(cid:0)(w, s)γ(r)−1(cid:1)
dr
r
.
When γ is the curve (rz, r2t), we will denote Hγ by H(z,t). By H ǫ
(z,t) we denote the
truncated Hilbert transform where the integration is only over {r ∈ R, ǫ < r < 1/ǫ}.
(z,t) to an operator on Lp(Rn+1) under the U action
When we transfer this operator H ǫ
defined above, we get
(H ǫ
(z,t))0f (ξ, η) = H ǫ
(z,t)F(ξ,η)(0) =Zǫ<r<1/ǫ
dr
F(ξ,η)(−rz,−r2t)
r
x · y
2 (cid:17)(cid:17) dr
f(cid:16)ξ + ry, η + r(x · ξ) + r2(cid:16)t +
= Zǫ<r<1/ǫ
r
.
Hence we see that, (H ǫ
(z,t))0 = T (z,t)
ǫ
. Now our aim is to prove
kT (z,t)
ǫ
fkLp(Rn+1) ≤ CpkfkLp(Rn+1)
This can be obtained using the transference method from the uniform boundedness of
the operator H ǫ
(z,t). To be precise, we use the fact that there exist a finite constant Cp,
independent of (z, t) ∈ Hn, n and ǫ > 0 such that
kH ǫ
(z,t)fkp ≤ Cpkfkp.
A proof of the above fact is indicated in [CMZ]. We will present it at the end of this
section after proving the boundedness of T (z,t)
. Then
ǫ
ZRn+1T (z,t)
ǫ
f (ξ, η)pdξ dη = ZRn+1 H ǫ
= ZRn+1 H ǫ
(z,t)F(ξ,η)(0)pdξ dη
(z,t)FU (w,s)(ξ,η)(0)pdξ dη
because of the invariance of the measure dξ dη under the U action. Hence
ZRn+1T (z,t)
ǫ
f (ξ, η)pdξ dη =
=
1
BR(0)ZBR(0)ZRn+1H ǫ
BR(0)ZBR(0)ZRn+1H ǫ
1
(z,t)FU (w,s)(ξ,η)(0)pdξ dη dw ds
(z,t)F(ξ,η)(w, s)pdξ dη dw ds
where BR(0) is the ball centred at origin in the Heisenberg group and radius R under
the Koranyi norm k(w, s)k = (w4 + s2)1/4 . When (w, s) ∈ BR(0), and ǫ < r <
1/ǫ, (w, s) (δr(z, t))−1 ∈ BR+ h
(0) where h is the Koranyi norm of (z, t). Hence, in
the above equality F(ξ,η) can be replaced by F(ξ,η) = F(ξ,η) · χBR+h/ǫ(0). Now by an
application of Fubini, we get
ǫ
kT (z,t)
ǫ
fkp
p ≤
1
BR(0)ZRn+1ZHnH ǫ
(z,t)
F(ξ,η)(w, s)pdw ds dξ dη
Now, from the uniform boundedness of the truncated Hilbert transforms, we get,
RIESZ TRANSFORMS
9
kT (z,t)
ǫ
fkp
p ≤ Cp
= Cp
1
1
BR(0)ZRn+1ZHn F(ξ,η)(w, s)pdw ds dξ dη
BR(0)ZBR+h/ǫ(0)ZRn+1 f (U(w, s)(ξ, η))pdξdηdwds
BR(0) kfkp
p = Cp(cid:18)R + h/ǫ
R (cid:19)2n+2
kfkp
p
= CpBR+h/ǫ(0)
again by the invariance dξdη under the U action. Letting R → ∞, we see that
kT (z,t)
ǫ
fkLp(Rn) ≤ CpkfkLp(Rn)
Coming back to the boundedness of H ǫ
to reduce this to the boundedness of H ǫ
Recall that
(z,t), we can use a technique in Lemma 3.1 of [S2],
γ on Lp(R2) for the curve γ = {(t, t2), t ∈ R}.
H(z,t)f (w, s) =ZR
f (w − rz, s − r2t − r Im (w.¯z))
dr
r
.
For σ ∈ U(n), define ρ(σ)f (w, s) = f (σw, s). Then
ρ(σ−1)H(z,t)(ρ(σ)f )(w, s) = H(z,t)(ρ(σ)f )(σ−1w, s)
H(z,t)(ρ(σ)f )(w, s) = ZR
= ZR
= ZR
f (σw − rσz, s − r2t − r Im (w.¯z))
dr
r
.
f (w − rσz, s − r2t − r Im (σ−1w.¯z))
dr
r
f (w − rσz, s − t − r Im (w. ¯σz))
.
dr
r
.
Since kρ(σ)fkp = kfkp and since there exists σ ∈ U(n) such that ρ(σ)(x + iy) = x1e1,
it is enough to consider the operator
f →ZR
f (u1 − rx1, w′, s − r2t − rv1x1)
dr
r
which is equivalent to the operator
T(x,t,v)f (u, s) =ZR
f (u − rx, s − r2t − rvx)
dr
r
acting on functions defined on R2. For λ1, λ2 > 0, let
δλ1,λ2f (u, s) = f (λ1u, λ2s).
Then kδλ1,λ2fkp = (λ1λ2)−1/pkfkp and
δ1/λ1,1/λ2T(x,t,v)δλ1,λ2f (u, s) =ZR
f (u − rλ1x, s − r2λ2t − rλ2vx)
dr
r
Since we can choose λ1, λ2 such that λ1x = λ2t = 1 it is enough to get uniform esti-
mates for operators of the form T(1,1,a). That would imply that
10
P. K. SANJAY AND S. THANGAVELU
kT(x,t,v)fkp = kδλ1,λ2T(1,1,vx/t)δ1/λ1,1/λ2fkp
= (λ1λ2)−1/pkT(1,1,vx/t)δ1/λ1,1/λ2fkp
≤ (λ1λ2)−1/pCpkδ1/λ1,1/λ2fkp
= (λ1λ2)−1/pCp(λ1λ2)1/pkfkp.
Now consider τaf (u, s) = f (u, s + au) so that kτafkp = kfkp. Since
a f )(u − r, s − r2 − ar)
(τ−1
f (u − r, s − au − r2)
a f )(u, s) =ZR
=ZR
T(1,1,a)(τ−1
dr
r
,
dr
r
we have
τaT(1,1,a)(τ−1
a f )(u, s) = T(1,1,a)(τ−1
a f )(u, s + au)
=ZR
f (u − r, s − r2)
dr
r
.
This is the Hilbert transform along the parabola γ(r) = (r, r2) in R2. Since
γ(r) =
δr(1, 1)
0
δ−r(−1, 1)
if r > 0
if r = 0
if r < 0
and the linear space spanned by {γ(r)}r>0 and the linear space spanned by {γ(r)}r<0
are the same, namely R2, γ(r) is a two-sided homogeneous curve (see Def 3.1 in
[SW, p. 1261]). Now we can appeal to Theorem 11 of [SW, p. 1271] to get the
(z,t) on Lp(Hn). That is,
uniform boundedness of the truncated Hilbert transforms H ǫ
there exist a finite constant Cp, independent of (z, t) ∈ Hn, n and ǫ > 0 such that
kH ǫ
(z,t)fkp ≤ Cpkfkp.
(cid:3)
5. Proof of the main theorem
We first prove Theorem 2.2 regarding the boundedness of the individual Riesz
transforms Rj and R∗j .
Proof of Theorem 2.2. We first consider the operators Rǫ
j . In the light
of Proposition 3.1 and Proposition 4.1, we just need to prove that there exist a finite
C, independent of n such that
j and R∗ǫ
As mentioned in Lemma 3 of [CMZ], this follows from the fact that
k Zjp1(z, t)kL1(Hn),k Z∗j p1(z, t)kL1(Hn) ≤ C.
where pn
r is the heat kernel on Hn. Now for the operators Rj and R∗j , we need only
r(z1,·) ∗ p1
pn
r (z1, z2,· · · , zn, t) = p1
jf(cid:1) , (cid:0)R∗ǫ
fkp and so it is enough to prove that T (z,t)
j f(cid:1) are Cauchy in Lp(Rn+1). We have already seen that
f is Cauchy in Lp(Rn+1).
r(z2,·),∗ · · · ,∗p1
r(zn,·)(t)
ǫ
to show that (cid:0)Rǫ
jfkp ≤ CkT (z,t)
kRǫ
ǫ
RIESZ TRANSFORMS
11
During the course of the proof of Proposition 4.1, we had observed that the operator
T (z,t)
ǫ
is obtained by applying a transference on H ǫ
fkLp(Rn+1) ≤ kH ǫ
kT (z,t)
(z,t) and that
(z,t)fkLp(Hn).
ǫ
Being truncated Hilbert transforms H ǫ
and consequently, so is T (z,t)
f .
ǫ
(z,t)f is Cauchy in Lp(Hn) [SW, Theorem 11],
(cid:3)
Now we complete the proof of our main theorem. As mentioned during the dis-
cussion of Theorem 2.2, we only need to prove the boundedness of the operator Rǫ
obtained by replacing Rj and R∗j by their truncated versions. We will be closely
following [LP1] in proving this and so will skip some details. Using the property
1
r
∂p1
∂r
=
1
xj
∂p1
∂xj
=
1
yj
∂p1
∂yj
+
2√πZHn
(xj − yj)T (z,t)
ǫ
f (ξ, η)
∂p1
∂t
(z, t)dzdt
1
r
1
1
r
1
where r = z, we can rewrite
Rǫ
jf (ξ, η) =
1
√π ZHn
(ixj + yj)T (z,t)
ǫ
f (ξ, η)
∂p1
∂r
(z, t)dzdt
and
R∗ǫ
j f (ξ, η) =
1
√πZHn
(ixj − yj)T (z,t)
ǫ
f (ξ, η)
∂p1
∂r
(z, t)dzdt
−
2√π ZHn
(xj + yj)T (z,t)
ǫ
f (ξ, η)
∂p1
∂t
(z, t)dzdt
For a fixed (ξ, η) we can choose λ1, λ2, . . . , λ2n such that P2n
n
jf (ξ, η) + λj+nR∗ǫ
Rǫf (ξ, η) =
Xj=1(cid:16)λjRǫ
j=1 λj2 = 1 and
j f (ξ, η)(cid:17)
Now using the triangle inequality, Holder's inequality and also Lemma 2(a) of [LP1],
we get
Rǫf (ξ, η) ≤ CkT (z,t)
f (ξ, η)kLp( 1
∂p1
ǫ
r
∂r dzdt)kx1kLp′ ( 1
+ CkT (z,t)
ǫ
r
∂p1
∂r dzdt)
f (ξ, η)kLp( ∂p1
∂t dzdt)kx1kLp′ ( ∂p1
∂t dzdt)
where C is a universal constant. Now the theorem follows from Proposition 4.1 and
an application of Minkowski inequality along with the fact (proved in [LP1])that
ZHn x1p 1
r
∂p1
∂r
dz dt,ZHn x1p ∂p1
∂t
where Ap is independent of the dimension.
dz dt ≤ Ap, p ≥ 0,
Acknowledgments
The authors are thankful to the referee for pointing out some errors and typos. The
work of the first author is supported by the All India Council for Technical Education
(AICTE). The work of the second author is supported by J. C. Bose Fellowship from
the Department of Science and Technology (DST) and also by a grant from UGC via
DSA-SAP.
12
P. K. SANJAY AND S. THANGAVELU
References
[BG] F. Baudoin and N. Garofalo, A note on the boundedness of Riesz transform for some subel-
liptic operators, International Mathematics Research Notices (to appear); Arxiv preprint
arXiv:1105.0467 (2012).
[CMZ] T. Coulhon, D. Muller, and J. Zienkiewicz, About Riesz transforms on the Heisenberg groups,
Math. Ann. 305 (1996), no. 2, 369 -- 379. MR1391221 (97f:22015)
[L] Karel de Leeuw, On Lp multipliers, Ann. of Math. (2) 81 (1965), 364 -- 379. MR0174937 (30
#5127)
[DR] J. Duoandikoetxea and Jos´e L. Rubio de Francia, Estimations ind´ependantes de la dimension
pour les transform´ees de Riesz, C. R. Acad. Sci. Paris S´er. I Math. 300 (1985), no. 7, 193 --
196. MR780616 (86e:42028)
[HRST] E. Harboure, L. de Rosa, C. Segovia, and J. L. Torrea, Lp-dimension free boundedness for
Riesz transforms associated to Hermite functions, Math. Ann. 328 (2004), no. 4, 653 -- 682.
MR2047645 (2006a:42017)
[JST] K. Jotsaroop, P K Sanjay, and S. Thangavelu, Riesz transforms and multipliers for
the Grushin operator, Journal d'Analyse Mathematique (to appear); Arxiv preprint
arXiv:1105.3227 (2011).
[LP1] F. Lust-Piquard, Riesz transforms on generalized Heisenberg groups and Riesz transforms
associated to the CCR heat flow, Publ. Mat. 48 (2004), no. 2, 309 -- 333. MR2091008
(2005g:43014)
[LP2]
, Dimension free estimates for Riesz transforms associated to the harmonic oscillator
on Rn, Potential Anal. 24 (2006), no. 1, 47 -- 62. MR2218202 (2006k:42012)
[P] G. Pisier, Riesz transforms: a simpler analytic proof of P.-A. Meyer's inequality, S´eminaire
de Probabilit´es, XXII, 1988, pp. 485 -- 501. MR960544 (89m:60178)
[RT] P. K. Ratnakumar and S. Thangavelu, Spherical means, wave equations, and Hermite-
Laguerre expansions, J. Funct. Anal. 154 (1998), no. 2, 253 -- 290. MR1612697 (99h:33047)
[S1] E. M. Stein, Some results in harmonic analysis in Rn, for n → ∞, Bull. Amer. Math. Soc.
[S2] R. S. Strichartz, Lp harmonic analysis and Radon transforms on the Heisenberg group, J.
(N.S.) 9 (1983), no. 1, 71 -- 73. MR699317 (84g:42019)
Funct. Anal. 96 (1991), no. 2, 350 -- 406. MR1101262 (92d:22015)
[ST] P. K Sanjay and S. Thangavelu, Revisiting Riesz transforms on Heisenberg groups, Revista
Mathem´atica Iberoamericana 28 (2012), no. 4, 1091 -- 1108, available at arXiv:1110.3236.
[SW] E. M. Stein and S. Wainger, Problems in harmonic analysis related to curvature, Bull. Amer.
Math. Soc. 84 (1978), no. 6, 1239 -- 1295. MR508453 (80k:42023)
[T] S. Thangavelu, An introduction to the uncertainty principle, Progress in Mathematics,
vol. 217, Birkhauser Boston Inc., Boston, MA, 2004. Hardy's theorem on Lie groups, With
a foreword by Gerald B. Folland. MR2008480 (2004j:43007)
[W] L. Weis, Operator-valued Fourier multiplier theorems and maximal Lp-regularity, Math. Ann.
319 (2001), no. 4, 735 -- 758. MR1825406 (2002c:42016)
Department of Mathematics, Indian Institute of Science, Bangalore-560 012.
E-mail address, Sanjay P. K.: [email protected]
E-mail address, S. Thangavelu: [email protected]
Permanent address(Sanjay P. K.):Department of Mathematics, National Institute
of Technology, Calicut- 673 601.
|
1111.6812 | 4 | 1111 | 2012-09-25T14:46:48 | Atomic representations in function spaces and applications to pointwise multipliers and diffeomorphisms, a new approach | [
"math.FA"
] | In Chapter 4 of [25] Triebel proved two theorems concerning pointwise multipliers and diffeomorphisms in function spaces of Besov and Triebel-Lizorkin type. In each case he presented two approaches, one via atoms and one via local means. While the approach via atoms was very satisfactory concerning the length and simplicity, only the rather technical approach via local means proved the theorems in full generality. In this paper we generalize two extensions of these atomic decompositions, one by Skrzypczak (see [22]) and one by Triebel and Winkelvoss (see [30]) so that we are able to give a short proof using atomic representations getting an even more general result than in the two theorems in [25].
References:
[22] L. Skrzypczak. Atomic decompositions on manifolds with bounded geometry. Forum Math., 10(1):19-38, 1998.
[25] H. Triebel. Theory of Function Spaces II. Birkh\"auser, Basel, 1992.
[30] H. Triebel and H. Winkelvo{\ss}. Intrinsic atomic characterizations of function spaces on domains. Math. Z., 221(1):647-673, 1996. | math.FA | math |
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO
POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH
BENJAMIN SCHARF
ABSTRACT. In Chapter 4 of [28] Triebel proved two theorems concerning pointwise multipliers and diffeomor-
phisms in function spaces Bs
p,q(Rn). In each case he presented two approaches, one via atoms and
one via local means. While the approach via atoms was very satisfactory concerning the length and simplicity,
only the rather technical approach via local means proved the theorems in full generality.
p,q(Rn) and Fs
In this paper we generalize two extensions of these atomic decompositions, one by Skrzypczak (see [25]) and
one by Triebel and Winkelvoss (see [33]) so that we are able to give a short proof using atomic representations
getting an even more general result than in the two theorems in [28].
INTRODUCTION
The aim of this paper is to generalize the atomic decomposition theorem from Triebel [28, 29] for Besov
p,q(Rn) and to present two applications to pointwise multipliers
p,q(Rn). For a detailed (historical)
p,q(Rn) we refer to Triebel [27, 28], for an introduction to atoms we
and Triebel-Lizorkin spaces Bs
and diffeomorphisms as continuous linear operators in Bs
treatment of the spaces Bs
refer to Frazier and Jawerth [5, 6].
p,q(Rn) resp. F s
p,q(Rn) and Fs
p,q(Rn) and Fs
According to Triebel [28]
Pj
: f 7→ j
· f
maps Bs
vector function j
p,q(Rn) into Bs
: Rn → Rn
p,q(Rn) if s > s p and j ∈ Ck(Rn) with k > s. Furthermore, the superposition with a
Dj
: f 7→ f ◦ j
is a k-diffeomorphism and k is large enough in dependence of s and p.
p,q(Rn) if j
maps Bs
There are similar results for F s
p,q(Rn) to Bs
p,q(Rn).
The main idea for an easy proof is the atomic decomposition theorem. Mainly one has to show that
a multiplication of an atom an ,m with a function j
is still an atom with
similar properties. But there was one problem: If s ≤ s p resp. s ≤ s p,q, then atoms need to fulfil moment
conditions, i.e.
the superposition with j
resp.
(1)
for L ∈ N0 and L > s p − s resp. L > s p,q − s. But these properties are not preserved by multiplication
resp. superposition. By Skrzypczak [25] these moment conditions were replaced by the more general
assumptions
x
a(x) dx = 0 if b ≤ L − 1
b
ZRn
(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
y (x)a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · 2−n (cid:16)s+L+n(cid:16)1− 1
p(cid:17)(cid:17)ky CL(Rn)k
for all y ∈ CL(Rn). Now the situation changes: These conditions remain true after multiplication resp.
superposition.
This replacement is typical when thinking of atomic, in particular wavelet representations as repre-
sentations of functions not mapping from Rn, but from more general manifolds, see the remarks on the
cancellation property in [3, Section 3.1].
2010 Mathematics Subject Classification. 46E35.
Key words and phrases. Atomic decompositions, pointwise multipliers, diffeomorphisms, Sobolev spaces, Besov and Lizorkin-
Triebel spaces.
Benjamin Scharf: Mathematisches Institut, Fakultät für Mathematik und Informatik, Friedrich-Schiller-Universität Jena, D-07737
Jena, Germany, [email protected].
1
2
BENJAMIN SCHARF
In this paper we go a step further. We show that one can replace the usual CK(Rn)-conditions on atoms
by Hölder-conditions (C K(Rn)-spaces) in the following way:
A function a : Rn → C is called (s, p)K,L-atom located at Qn ,m if
supp a ⊂ d · Qn ,m
ka(2−n
·)C K(Rn)k ≤ C · 2−n (s− n
p )
and for every y ∈ C L(Rn) it holds
(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
y (x)a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · 2−n (cid:16)s+L+n(cid:16)1− 1
p(cid:17)(cid:17)ky C L(Rn)k.
This generalizes the known definitions of atoms from Triebel, Skrzypczak and Winkelvoss [29, 25, 33].
Furthermore, there is an existing theory generalizing the conditions ka(2−n
p,p(Rn)k
with K > s, mainly in connection with spline representations. For instance, see the books by Kahane and
Lemarie-Rieusset [11, part II, Section 6.5], Triebel [31, Section 2.2] and the recent paper by Schneider and
Vybiral [22]. Of these, only the first book incorporates the usual moment conditions as in (1).
·)C K(Rn)k by ka(·)BK
In Section 3, as corollaries of the atomic representation theorem with these more general atoms from
Section 2 we are able to extend the key theorems on pointwise multipliers and diffeomorphisms from [28].
It is not the aim of our observations to give best conditions or even exact characterizations for pointwise
p,q(Rn). For this we refer to Strichartz [26], Peetre [18] as
multipliers in function spaces Bs
well as to Maz'ya and Shaposhnikova [15, 16] for the classical Sobolev spaces, while for Bs
p,p(Rn) we refer
to Franke [4], Frazier and Jawerth [6], Netrusov [17], Koch, Runst and Sickel [19, 23, 24, 12] as well as to
Triebel [31, Section 2.3.3] for general function spaces Bs
We obtain for pointwise multipliers with respect to Bs
Let 0 < p ≤ ¥ and r > max(s,s p − s). Then there exists a positive number c such that
p,q(Rn) and Fs
p,q(Rn):
p,q(Rn) and F s
p,q(Rn).
for all j ∈ C
r
(Rn) and all f ∈ Bs
f Bs
p,q(Rn).
kj
p,q(Rn)k ≤ ckj C
r
(Rn)k · k f Bs
p,q(Rn)k
For further sufficient results on diffeomorphisms including characterizations for classical Sobolev spaces
W k
p (Rn) we refer to Gol'dshtein, Reshetnyak, Romanov, Ukhlov and Vodop'yanov [8, 9, 10, 35, 34], [7,
p,q(Rn)
Chapter 4], Markina [13] as well as to Maz'ya and Shaposhnikova [14, 15], while for Besov spaces Bs
with 0 < s < 1 we refer to Vodop'yanov, Bordaud and Sickel [35, 1]. A special case of our result (Lipschitz
diffeomorphisms) can be found in Triebel [30, Section 4].
We will prove (in case of Bs
Let 0 < p ≤ ¥
, r ≥ 1 and r > max(s,1 +s p − s). If j
p,q(Rn)):
is a r -diffeomorphism, then there exists a constant
c such that
k f (j (·))Bs
p,q(Rn)k ≤ c · k f Bs
p,q(Rn)k.
for all f ∈ Bs
p,q(Rn). Hence Dj maps Bs
p,q(Rn) onto Bs
p,q(Rn).
Furthermore, at the end of section 2 we are able to give a simple proof of a local mean theorem very
similar to Triebel's result [32, Theorem 1.15], which paved the way for the wavelet characterization of
p,q(Rn) and F s
Bs
p,q(Rn) - where we are also using the more general Hölder-space conditions.
1. PRELIMINARIES
Let Rn be the euclidean n-space, Z be the set of integers,N be the set of natural numbers and N0 =
N ∪ {0}. By x we denote the usual euclidean norm of x ∈ Rn, by kxXk the (quasi)-norm of an element x
of a (quasi)-Banach space X.
By S (Rn) we mean the Schwartz space on Rn, by S ′(Rn) its dual. The Fourier transform of f ∈
f . The convolution of f ∈ S ′(Rn) and j ∈ S (Rn)
S ′(Rn) resp. its inverse will be denoted by f resp.
will be denoted by f ∗ j .
By Lp(Rn) for 0 < p ≤ ¥ we denote the usual quasi-Banach space of p-integrable complex-valued
functions with respect to the Lebesgue measure m with quasi-norm
f (x)p dx(cid:19)
k f Lp(Rn)k :=(cid:18)ZRn
1
p
.
Let X,Y be quasi-Banach spaces. By the notation X ֒→ Y we mean that X ⊂ Y and that the inclusion
map is bounded.
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH3
Throughout the paper all unimportant constants will be called c,c′,C etc. Only if extra clarity is desirable,
the dependency of the parameters will be stated explicitly. The concrete value of these constants may vary
in different formulas but remains the same within one chain of inequalities.
1.1. Hölder spaces of differentiable functions. Let k ∈ N0. Then by Ck(Rn) we denote the space of all
functions f : Rn → C which are k-times continuously differentiable (continuous, if k = 0) such that the
norm
k f Ck(Rn)k := (cid:229)
a ≤k
a
sup D
f (x)
is finite, where the sup is taken over x ∈ Rn.
(Rn) is defined by
Furthermore, the set C¥
C
(Rn) := \k∈N0
Ck(Rn).
Definition 1.1. Let 0 < s ≤ 1 and f : Rn → C be continuous. We define
f (x) − f (y)
s
k f lip
(Rn)k := sup
x,y∈Rn,x6=y
s
x − y
.
If s ∈ R, then there are uniquely determined ⌊s⌋ ∈ Z and {s} ∈ (0,1] with s = ⌊s⌋ + {s}.
Let s > 0. Then the Hölder space with index s is given by
C s(Rn) =n f ∈ C⌊s⌋(Rn) : k f C s(Rn)k < ¥ o with
k f C s(Rn)k := k f C⌊s⌋(Rn)k + (cid:229)
kD
a
f lip{s}(Rn)k.
a =⌊s⌋
If s = 0, then C 0(Rn) := L¥ (Rn), which is sufficient for the later statements, see e.g. Theorem 2.12.
1.2. Besov and Triebel-Lizorkin function spaces on Rn. Let j
j for j ∈ N0 be elements of S (Rn) with
(2)
supp j 0 ⊂ {x ≤ 2},
supp j
j ⊂ {2 j−1 ≤ x ≤ 2 j+1} for j ∈ N,
j
j(x ) = 1 for all x ∈ Rn,
j=0
a j
D
j(x ) ≤ ca 2− ja for all a ∈ Nn
0.
Then we call {j
j=0 a smooth dyadic resolution of unity. For instance one can choose Y ∈ S (Rn) with
j}
(x ) = 1 for x ≤ 1 and supp Y ⊂ {x ≤ 2} and set
(x /2) − Y
, s ∈ R and {j
(x ),
Definition 1.2. Let 0 < p ≤ ¥
Bs
p,q(Rn) is the collection of all f ∈ S ′(Rn) such that the quasi-norm
j 1(x ) := Y
, 0 < q ≤ ¥
j 0(x ) := Y
(x ),
j}
j
j(x ) := j 1(2− j+1x ) for j ∈ N.
j=0 be a smooth dyadic resolution of unity. Then
k f Bs
p,q(Rn)k := ¥
j=0
2 jsqk(j
1
q
j f )Lp(Rn)kq!
(modified if q = ¥
) is finite.
Definition 1.3. Let 0 < p < ¥
F s
p,q(Rn) is the collection of all f ∈ S ′(Rn) such that the quasi-norm
, s ∈ R and {j
, 0 < q ≤ ¥
j}
j=0 be a smooth dyadic resolution of unity. Then
(modified if q = ¥
) is finite.
k f F s
p,q(Rn)k :=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
j=0
2 jsq(cid:12)(cid:12)(j
j f )(·)(cid:12)(cid:12)
1
q
q!
(cid:12)(cid:12)
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
¥
(cid:229)
¥
Y
¥
(cid:229)
¥
(cid:229)
4
BENJAMIN SCHARF
One can show that the introduced quasi-norms1 for two different smooth dyadic resolutions of unity are
equivalent for fixed p, q and s, i. e. that the so defined spaces are equal. This follows from Fourier multiplier
theorems, see [27], Section 2.3.2., p. 46. Furthermore, the so defined spaces are (quasi)-Banach spaces.
The next proposition, the so called Fatou property, is a classical observation for function spaces Bs
p,q(Rn)
and F s
p,q(Rn), see [4].
Definition 1.4. Let A be a quasi-Banach space with S (Rn) ֒→ A ֒→ S ′(Rn). Then we say that A has the
Fatou property if there exists a constant c such that: If a sequence { fn}n∈N ⊂ A converges to f with respect
to the weak topology in S ′(Rn) and if k fnAk ≤ D, then f ∈ A and k f Ak ≤ c · D.
Proposition 1.5. Let s ∈ R,0 < q ≤ ¥ and 0 < p ≤ ¥
. Then Bs
the Fatou property.
Remark 1.6. If r > 0 and r /∈ N, then C
see [28, Sections 1.2.2, 2.6.5] or for the original source [36, Lemma 4].
,¥ (Rn). This is a classical observation, for instance
resp. 0 < p < ¥
p,q(Rn) and F s
p,q(Rn) have
(Rn) = B
r
r
Moreover, we set
where a+ = max(a,0).
s p = n(cid:18) 1
p
− 1(cid:19)+
,
s p,q = n(cid:18)
1
min(p,q)
− 1(cid:19)+
,
2. ATOMIC DECOMPOSITIONS
At first we describe the concept of atoms as one can find it in [29], Definition 13.3, p. 73, now generalized
using ideas from [25] and [33].
In particular, this gives the possibility to omit the distinction between n = 0 and n ∈ N and now the usual
parameters K and L are nonnegative real numbers instead of natural numbers.
2.1. General atoms. Let Qn ,m := {x ∈ Rn : xi − 2−n mi ≤ 2−n −1} be the cube with sides parallel to the
axes, with center at 2−n m and side length 2−n for m ∈ Zn and n ∈ N0.
Definition 2.1. Let s ∈ R, 0 < p ≤ ¥
A function a : Rn → C is called (s, p)K,L-atom located at Qn ,m if
(3)
, K,L ∈ R and K,L ≥ 0. Furthermore let d > 1, C > 0, n ∈ N0,m ∈ Z.
supp a ⊂ d · Qn ,m
ka(2−n
·)C K(Rn)k ≤ C · 2−n (s− n
p )
Remark 2.2. If L = 0, then condition (5) is neglectable since it follows from (3) and (4) with K = 0.
p(cid:17).
The constant in the exponent will be shortened by κL := s + L + n(cid:16)1 − 1
If K = 0, then by Definition 1.1 we only require a to be suitably bounded.
Later on, we will choose one (s, p)K,L-atom for every n ∈ N0 and m ∈ Zn. Then the parameter d > 1
shall be the same for all these atoms - it describes the overlap of these atoms on one fixed level n ∈ N0.
Remark 2.3. The usual formulation of (4) as in [29] was
a
a(x) ≤ 2−n (cid:16)s− n
p(cid:17)+a n
for all a ≤ K
D
(6)
for K ∈ N0. The modification here was suggested in [33]. It is easy to see that (4) follows from (6) if K is a
natural number, since CK(Rn) ֒→ C K(Rn).
Remark 2.4. The usual formulation of (5) as in [29] was
(7)
for n ∈ N, so n
6= 0. The modification here was suggested in Lemma 1 of [25] for natural numbers L + 1
(using CL(Rn) instead of C L(Rn)). Now we extended this definition to general positive L. For natural L − 1
x
a(x) dx = 0 if b ≤ L − 1
ZRn
b
1In the following we will use the term "norm" even if we only have quasi-norms for p < 1 or q < 1.
(4)
and for every y ∈ C L(Rn) it holds
(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
(5)
y (x)a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · 2−n (cid:16)s+L+n(cid:16)1− 1
p(cid:17)(cid:17)ky C L(Rn)k.
¥
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH5
one can derive (5) from (7) using a Taylor expansion, see [25, Lemma 1, (12) and (14)] or the upcoming
Lemma 2.8. Hence formulation (5) is a generalization.
An alternative formulation of (5) is given by
(8)
(x − 2−n
b
m)
≤ C · 2−n κL if b ≤ ⌊L⌋.
Zd·Qn ,m
(cid:12)(cid:12)(cid:12)(cid:12)
a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
Obviously, this condition is covered by condition (5). For the other direction see [25, Lemma 1, (12)
and (14)] or the upcoming Remark 2.9, in particular (10). It is also possible to assume this condition for
all b ∈ Nn since the statements for b ≥ L follow from the support condition (3) and the boundedness
condition included in (4).
This shows that both conditions (4) and (5) are ordered in K resp. L, i.e. the conditions get stricter for
increasing K resp. L.
Now the question will be whether these more general atoms allow analogous results regarding atomic
decompositions.
2.2. Sequence spaces. We introduce the sequence spaces bp,q and f p,q, whose use will become clear in the
following. For this we refer to [29], Definition 13.5, p. 74.
Definition 2.5. Let 0 < p ≤ ¥
, 0 < q ≤ ¥
and
l = {l n ,m ∈ C : n ∈ N0,m ∈ Zn} .
We set
and
l
bp,q :=
: kl bp,qk =
: kl f p,qk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
n =0
), where c (p)
l
f p,q :=
n =0 (cid:229)
m∈Zn
l n ,mp!
l n ,mc (p)
n ,m(·)q!
m∈Zn
1
q
q
< ¥
p
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
< ¥
(cid:12)(cid:12)
1
q
(modified in the case p = ¥ or q = ¥
cube Qn ,m, i.e
n ,m is the Lp(Rn)-normalized characteristic function of the
c (p)
n ,m = 2
n n
p if x ∈ Qn ,m and c (p)
n ,m = 0 if x /∈ Qn ,m.
2.3. Local means. Let N ∈ N0 be given. We choose k0,k ∈ S (Rn) with compact support - e.g. supp k0,supp k ⊂
e · Q0,0 for a suitable e > 0 - such that
(9)
while k0(0) 6= 0. Furthermore, let there be an e > 0 such that k(x) 6= 0 for 0 < x < e .
a k(0) = 0 if a < N,
D
Such a choice is possible, see [29, 11.2]. We set k j(x) := 2 jnk(2 jx) for j ∈ N.
Proposition 2.6. Let N ∈ N0 and N > s.
. Then
(i) Let 0 < p ≤ ¥ and 0 < q ≤ ¥
k f Bs
p,q(Rn)kk0,k := kk0 ∗ f Lp(Rn)k + ¥
j=1
2 jsqkk j ∗ f Lp(Rn)kq!
1
q
(modified for q = ¥
) is an equivalent norm for k · Bs
p,q(Rn)k. It holds
(ii) Let 0 < p < ¥ and 0 < q ≤ ¥
k f F s
Bs
. Then
p,q(Rn) =(cid:8) f ∈ S ′(Rn) : k f Bs
p,q(Rn)kk0,k := kk0 ∗ f Lp(Rn)k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
p,q(Rn) =(cid:8) f ∈ S ′(Rn) : k f F s
) is an equivalent norm for k · Fs
F s
p,q(Rn)kk0,k < ¥ (cid:9) .
q!
2 jsq(cid:12)(cid:12)(k j ∗ f )(·)(cid:12)(cid:12)
p,q(Rn)kk0,k < ¥ (cid:9) .
p,q(Rn)k. It holds
j=1
1
q
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)
(modified for q = ¥
¥
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
6
BENJAMIN SCHARF
Remark 2.7. This proposition is due to [20]. Some minor technicalities of the proof where modified in the
fourth step of [21, Theorem 2.1] (for the more general vector-valued case).
2.4. A general atomic representation theorem. We start with a lemma which helps us to understand
the relation between conditions like (5) and (7) and which will be heavily used in the proof of the atomic
representation theorem. It also shows that local means and atoms are related, see condition (9).
Lemma 2.8. Let j ∈ N0. If k0 and k j = 2 jnk(2 j·) for j ∈ N are local means as in Definition 2.3, then
p(cid:17)(cid:17) · k j is an (s, p)K,L-atom located at Q j,0 for arbitrary K > 0 and for L ≤ N + 1 with N from
2− j(cid:16)s+n(cid:16)1− 1
(9).
Proof. For j = 0 there is nothing to prove since the moment condition (5) follows from (3) and (4). So we
can concentrate on j ∈ N: The support condition (3) follows from the compact support of k. Furthermore,
kk j(2− j)C K(Rn)k = 2 jnkkC K(Rn)k ≤ C 2 jn
since K is arbitrarily often differentiable. Hence, the Hölder-condition (4) is shown.
Now we have to show condition (5). There is nothing to prove for j = 0. Hence, we can use the moment
conditions (9). Let L > 0, L = ⌊L⌋ + {L} as in Section 1.1 and let y ∈ C L(Rn). We expand the ⌊L⌋-times
continuously differentiable function y
into its Taylor series of order ⌊L⌋ − 1. Then there exists a q ∈ (0,1)
with
y (x) = (cid:229)
b ≤⌊L⌋−1
b y (0) · x
b
D
1
b !
+ (cid:229)
b =⌊L⌋
1
b !
b y (q x) · x
b
D
.
Hence
y (x) − (cid:229)
b ≤⌊L⌋
b y (0) · x
D
1
b !
(cid:12)(cid:12)(cid:12)
b (cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)
b =⌊L⌋
1
b !(cid:0)D
≤ c · ky C L(Rn)k · xL.
b y (q x) − D
b y (0)(cid:1)x
b (cid:12)(cid:12)(cid:12)
Using (9) for k j and ⌊L⌋ ≤ N we can insert the polynomial terms into the integral and get
(10)
y (x)k j(x) dx(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)Zd·Q j,0
p(cid:17)(cid:17)k j fulfils condition (5). The constant C does not depend on j ∈ N0.
Hence 2− j(cid:16)s+n(cid:16)1− 1
Remark 2.9. If we take a look at the proof, we see that instead of (9) it suffices to have
≤ c ky C L(Rn)kZd·Q j,0
k j(x) · xL dx ≤ C · 2− jLky C L(Rn)k.
(cid:3)
(11)
≤ C · 2− jL if b ≤ ⌊L⌋.
b
x
(cid:12)(cid:12)(cid:12)Zd·Q j,0
k j(x) dx(cid:12)(cid:12)(cid:12)
· y C L(Rn)k ≤ C if b ≤ ⌊L⌋, where
(Rn) is a cutoff function, i.e. with compact support and y (x) = 1 for x ∈ supp k, hence for x ∈
In fact, this condition is equivalent to condition (5) for k j since kxb
y ∈ C¥
supp k j, too.
Now we will see what happens if an atom is dilated.
Lemma 2.10. Let j ∈ N0 and j ≤ n . If an ,m is an (s, p)K,L-atom located at the cube Qn ,m, then 2 j(s− n
an ,m(2− j·) is an (s, p)K,L-atom located at Qn − j,m.
p ) ·
Proof. The support condition (3) and the Hölder-condition (4) are easy to verify. Considering the moment
condition (5) we have
(cid:12)(cid:12)(cid:12)Zd·Qn − j,m
y (x)an ,m(2− jx) dx(cid:12)(cid:12)(cid:12)
= 2 jn ·(cid:12)(cid:12)(cid:12)Zd·Qn ,m
y (2 jx)an ,m(x) dx(cid:12)(cid:12)(cid:12)
≤ C · 2 jn · 2−n κL · ky (2 j·)C L(Rn)k
≤ C · 2 jn · 2−n κL · 2 jL · ky C L(Rn)k
= C · 2−(n − j)κL · 2− j(s− n
p ) · ky C L(Rn)k.
This is what we wanted to prove.
(cid:3)
(cid:229)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH7
Now we come to the essential part - showing the atomic representation theorem. We will use an approach
as in Theorem 13.8 of [29]. Using the more general form of the atoms we are able to simplify the proof:
One has to estimate
Z k j(x − y)an ,m(y) dy,
where k j are the local means from Section 2.3 and an ,m are atoms located at Qn ,m. One has to distinguish
between j ≥ n and j < n as in the original proof - but now both cases can be proven very similarly with our
more general approach of atoms.
At first we prove the convergence of the atomic series in S ′(Rn).
Lemma 2.11. Let 0 < p ≤ ¥
Then
resp. 0 < p < ¥
, 0 < q ≤ ¥
and s ∈ R. Let K ≥ 0, L ≥ 0 with L > s p − s.
l n ,man ,m
n =0
m∈Zn
converges unconditionally in S ′(Rn), where an ,m are (s, p)K,L-atoms located at Qn ,m and l ∈ bp,q or
l ∈ f p,q.
Proof. Let j ∈ S (Rn). Having in mind (3) and (5) we obtain
m (cid:12)(cid:12)(cid:12)(cid:12)
ZRn
l n ,man ,m(x)j (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · 2−n κL (cid:229)
m
l n ,m · kj
n
· y (2
· −m)C L(Rn)k,
(Rn), y (x) = 1 for x ∈ d · Q0,0 and supp y ∈ (d + 1) · Q0,0.
where y ∈ C¥
Observing κL = s + L + n(cid:16)1 − 1
(12)
Furthermore, since j ∈ S (Rn) we have
kj
· y (2n
0 < p ≤ 1
p(cid:17) and L > s p − s we get
κL >(0,
.
n(cid:16)1 − 1
p(cid:17) , 1 < p ≤ ¥
· −m)C L(Rn)k ≤ CM ·(cid:0)1 + 2−n
where M ∈ N0 is at our disposal and CM does not depend on n and m.
Let at first be 0 < p ≤ 1. Then we choose M = 0 and get
,
m(cid:1)−M
l n ,m ≤ C′ · 2−n κL(cid:18)(cid:229)
≤ C′ · kl bp,¥ k.
1
p
.
l n ,mp(cid:19)
m
(13)
we get
m
Summing up over n ∈ N0 using κL > 0 we finally arrive at
≤ C′ · 2−n κL (cid:229)
l n ,man ,m(x)j (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
m (cid:12)(cid:12)(cid:12)(cid:12)
ZRn
n
l n ,man ,m(x)j (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
m (cid:12)(cid:12)(cid:12)(cid:12)
ZRn
l n ,man ,m(x)j (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
ZRn
m (cid:12)(cid:12)(cid:12)(cid:12)
In the case 1 < p ≤ ¥ we choose M ∈ N0 such that Mp′ > n, where 1 = 1
p + 1
p′ . Using Hölder's inequality
m
≤ CM · 2−n κL (cid:229)
≤ C′ · 2−n κL(cid:18)(cid:229)
≤ C′′ · 2−n κL · 2
l n ,m ·(cid:0)1 + 2−n
m (cid:0)1 + 2−n
p′ ·(cid:18)(cid:229)
m(cid:1)−M
·(cid:18)(cid:229)
m(cid:1)−M p′(cid:19)
l n ,mp(cid:19)
n n
1
p′
m
1
p
.
1
p
l n ,mp(cid:19)
m
By (12) the exponent is smaller than zero. Hence summing over n ∈ N0 gives the same result as in (13).
Since
we have shown the absolut and hence unconditional convergence in S ′(Rn).
bp,q ֒→ bp,¥
resp. f p,q ֒→ f p,¥
(cid:3)
¥
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
8
BENJAMIN SCHARF
Theorem 2.12. (i) Let 0 < p ≤ ¥
Then f ∈ S ′(Rn) belongs to Bs
, 0 < q ≤ ¥
and s ∈ R. Let K,L ∈ R, K,L ≥ 0, K > s and L > s p − s.
p,q(Rn) if and only if it can be represented as
f =
n =0
m∈Zn
l n ,man ,m with convergence in S ′(Rn).
Here an ,m are (s, p)K,L-atoms located at Qn ,m (with the same constants d > 1 and C > 0 in Definition 2.1
for all n ∈ N0,m ∈ Z) and kl bp,qk < ¥
. Furthermore, we have in the sense of equivalence of norms
k f Bs
p,q(Rn)k ∼ inf kl bp,qk,
where the infimum on the right-hand side is taken over all admissible representations of f .
(ii) Let 0 < p < ¥
, 0 < q ≤ ¥ and s ∈ R. Let K,L ∈ R, K,L ≥ 0, K > s and L > s p,q −s. Then f ∈ S ′(Rn)
belongs to F s
p,q(Rn) if and only if it can be represented as
f =
n =0
m∈Zn
l n ,man ,m with convergence in S ′(Rn).
Here an ,m are (s, p)K,L-atoms located at Qn ,m (with the same constants d > 1 and C > 0 in Definition 2.1
for all n ∈ N0,m ∈ Z) and kl f p,qk < ¥
. Furthermore, we have in the sense of equivalence of norms
where the infimum on the right-hand side is taken over all admissible representations of f .
k f F s
p,q(Rn)k ∼ inf kl f p,qk,
Proof. We rely on the proof of Theorem 13.8 of [29], now modified keeping in mind the more general
conditions (4) and (5) instead of (6) and (7). There are two directions we have to prove.
At first, let us assume that f from Bs
p,q(Rn) is given. Then we know from Theorem 13.8
of [29] that f can be written as an atomic decomposition, with atoms now fulfilling conditions (6) and
(7) for given natural numbers K′ > s and L′ + 1 > s p − s resp. L′ + 1 > s p,q − s. Hence, because of
CK′
(Rn) ⊂ C K′
Conditions (5) are generalizations of the classical moment conditions (7) and are ordered in L, see
(Rn), condition (4) is fulfilled for all K ≤ K′.
p,q(Rn) or F s
Remark 2.4.
Thus, every classical (s, p)K′,L′-atom is an (s, p)K,L atom in the sense of definition 2.1 for K ≤ K′ and
p,q(Rn) for
L ≤ L′ + 1 and this immediately shows that we find a decomposition of f from Bs
arbitrary K and L in terms of the general atoms we introduced.
p,q(Rn) or Fs
Now we come to the essential part of the proof. We have to show that, although we weakened the
conditions on the atoms, a linear combination of atoms is still an element of Bs
p,q(Rn). We
modify the proof of Theorem 13.8 of [29] or into [21] where some minor technical details are modified
(for the more general vector-valued case). There one uses the equivalent characterization by local means
k0,k j := 2 jnk(2 j·) with a suitably large N(see Proposition 2.6) and distinguishes between the cases j ≥ n
and j < n . In both cases the crucial part is the estimate of
p,q(Rn) resp. Fs
Z k j(x − y)an ,m(y) dy,
where an ,m is an (s, p)K,L-atom centered at Qn ,m. The idea now is to use that not only an ,m but also k j can
been interpreted as atoms and admit estimates as in (4) and (5), see Lemma 2.8.
Let at first be j ≥ n . The function k has compact support and fulfils moment conditions (7). At first we
transform the integral, having in mind the form of condition (4) of an ,m,
2 jsZ k j(y)an ,m(x − y) dy = 2 jsZ k j−n (y)an ,m(x − 2−n
y) dy.
Surely, this integral vanishes for x /∈ c · Qn ,m for a suitable c > 0 because of j ≥ n . So we concentrate on
x ∈ c · Qn ,m: By Lemmata 2.8 and 2.10 the function
2−( j−n )(cid:16)s+n(cid:16)1− 1
p(cid:17)(cid:17) · k j−n = 2−( j−n )(cid:16)s+n(cid:16)1− 1
p(cid:17)(cid:17) · 2−n n · k j(2−n
·)
is an (s, p)M,N-atom located at Q j−n ,0 for M arbitrarily large and N from (9), so that also N may be arbitrarily
large, but fixed. Now we will use the moment condition (5) for k j−n and the Hölder-condition (4) for an ,m.
¥
(cid:229)
(cid:229)
¥
(cid:229)
(cid:229)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH9
Hence, with y (y) = an ,m(x − 2−n y) and N ≥ K we have
2 js(cid:12)(cid:12)(cid:12)Z k j−n (y)an ,m(x − 2−n
y) dy(cid:12)(cid:12)(cid:12)
≤ C · 2 js · 2−( j−n )K · kan ,m(x − 2−n
= C · 2 js · 2−( j−n )K · kan ,m(2−n
≤ C · 2 js · 2−( j−n )K · 2−n (s− n
p )
= C′ · 2−( j−n )(K−s) · c (p)(c · Qn ,m),
·)C K(Rn)k
·)C K(Rn)k
where c (p)(c ·Qn ,m) is the Lp(Rn)-normalized characteristic function of c ·Qn ,m. This inequality is certainly
true for x /∈ c · Qn ,m. Hence (13.37) in [29] is shown.
Now let j < n . We will interchange the roles of k j and an ,m using condition (4) now for k j and (5) for
an ,m. Hence we start with
2 jsZ k j(x − y)an ,m(y) dy = 2 jsZ k(2 jx − y)an ,m(2− jy) dy.
Surely, this integral vanishes for x /∈ c · 2n − j · Qn ,m. So we concentrate on x ∈ c · 2n − j · Qn ,m: By Lemma
2.10 we know that 2 j(s− n
p ) · an ,m(2− j·) is an (s, p)K,L-atom located at Qn − j,m while k is an (s, p)M,N -atom
located at Q0,0. Thus, using (4) for k with M ≥ L, we get
2 js(cid:12)(cid:12)(cid:12)Z k(2 jx − y)an ,m(2− jy) dy(cid:12)(cid:12)(cid:12)
p ) · kk(2 jx − ·)C L(Rn)k
≤ c · 2 js · 2−(n − j)κL · 2− j(s− n
≤ c · 2−(n − j)(L+s) · 2
= c · 2−(n − j)(L+s) · 2
n n
p · 2−(n − j)n
n n
p · 2−(n − j)n · c (c · 2n − j · Qn ,m).
where c (c · 2n − j · Qn ,m) is the characteristic function of c · 2n − j · Qn ,m. This estimate is the same as (13.41)
combined with (13.42) in [29] or (72) and (73) in [21], observing that we use L instead of L + 1 in the
atomic representation theorem.
Starting with these two estimates we can follow the steps in [29] or [21] and finish the proof, since K > s
and L > s p − s resp. L > s p,q − s. Strictly speaking, we arrive (in the Bs
p,q(Rn)-case) at
n ≤n 0
m≤m0
(cid:13)(cid:13)(cid:13)
l n ,man ,m(cid:12)(cid:12)
Bs
p,q(Rn)(cid:13)(cid:13)(cid:13)
≤ C · kl bp,qk
for all n 0,m0 ∈ N0 with a constant C independent of n 0 and m0. Using Lemma 2.11 and the Fatou property
of the spaces Bs
p,q(Rn) (see Proposition 1.5) we are finally done, i.e.
p,q(Rn) resp. F s
n ∈N0
m∈Zn
(cid:13)(cid:13)(cid:13)
l n ,man ,m(cid:12)(cid:12)
Bs
p,q(Rn)(cid:13)(cid:13)(cid:13)
≤ C · kl bp,qk.
(cid:3)
Remark 2.13. The conditions (4) and (5) for the atomic representation theorem can be slightly modified: If
K > 0, then it is possible to replace k · C K(Rn)k by k · BK¥
,¥ (Rn)k in condition (4). This is clear for K /∈ N,
see Remark 1.6. If K ∈ N, this follows from
C K(Rn) ֒→ BK¥
,¥ (Rn) ֒→ C K−e
(Rn)
for e > 0.
A similar result holds true for L > 0, L /∈ N and condition (5) by trivial means. If L ∈ N, then k ·C L(Rn)k
can be replaced by k · CL(Rn)k, where the condition needs to be true for all y ∈ CL(Rn). This follows from
the fact, that both conditions imply (8). Hence they are equivalent.
It is not clear to the author whether k · C L(Rn)k can be replaced by k · BL¥
,¥ (Rn)k for L ∈ N.
Remark 2.14. In the proof of Theorem 2.12 we assumed that the local means k j are arbitrarily often
differentiable and fulfil as many moment conditions as we wanted. But if we take a look into the proof, we
see that we did not use the specific structure k j = 2 jnk(2 j·). It is sufficient to know that there are constants
c and C such that for all j ∈ N0 it holds supp k j ⊂ c · Q j,0, that
(14)
kk j(2− j·)C M(Rn)k ≤ C · 2 jn
(cid:229)
(cid:229)
(cid:229)
(cid:229)
10
BENJAMIN SCHARF
with M ≥ L and that for every y ∈ C N(Rn) it holds
(15)
≤ C · 2− jN · ky C N(Rn)k
with N ≥ K because the atomic conditions (4) and (5) are ordered in N and M, see Remark 2.4. As before,
condition (15) can be strengthened by
Zd·Q j,0
(cid:12)(cid:12)(cid:12)(cid:12)
y (x)k j(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
Z x
b
k j(x) dx = 0 for all b < N.
Through these considerations the idea arises how to prove a counterpart of Theorem 2.12 for the local mean
characterization in [32, Theorem 1.15] without further substantial efforts. This is done in the following
Corollary, including some technical issues concerning the definition of a dual pairing (see [32, Remark
1.14]). It is obvious that the original version of Theorem 1.15 in [32] is just some kind of modification of
this Corollary.
, 0 < q ≤ ¥ and s ∈ R. Let M,N ∈ R, M,N ≥ 0, M > s p − s and N > s.
Corollary 2.15. (i) Let 0 < p ≤ ¥
Assume that for all j ∈ N0 it holds that k j ∈ C M(Rn), supp k j ⊂ c · Q j,0 and k j fulfils (14) and (15). Then
there is a constant c such that
k f Bs
p,q(Rn)kk := kk0 ∗ f Lp(Rn)k + ¥
j=1
2 jsqkk j ∗ f Lp(Rn)kq!
1
q
≤ c · k f Bs
p,q(Rn)k
(modified for q = ¥
) for all f ∈ Bs
(ii) Let 0 < p < ¥
p,q(Rn).
and s ∈ R. Let M,N ∈ R, M,N ≥ 0, M > s p,q − s and N > s. Assume
that for all j ∈ N0 it holds that k j ∈ C M(Rn), supp k j ⊂ c · Q j,0 and k j fulfils (14) and (15). Then there is a
constant c such that
, 0 < q ≤ ¥
k f F s
p,q(Rn)kk : = kk0 ∗ f Lp(Rn)k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)
p,q(Rn)k
≤ c · k f Fs
p,q(Rn).
j=1
2 jsq(cid:12)(cid:12)(k j ∗ f )(·)(cid:12)(cid:12)
q(cid:17)
1
q(cid:12)(cid:12)(cid:12)
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)
) for all f ∈ F s
(modified for q = ¥
Proof. There is nearly nothing left to prove because the crucial steps were done in the proof before: Let
f ∈ Bs
p,q(Rn) (analogously for f ∈ F s
p,q(Rn) by
an "optimal" atomic decomposition
p,q(Rn)) be given. By Theorem 2.12 we can represent f ∈ Bs
f =
n =0
m∈Zn
l n ,man ,m,
where an ,m is an (s, p)N,M-atom located at Qn ,m and k f Bs
of f ).
p,q(Rn)k ∼ kl bp,qk (with constants independent
But, by the second step of the proof of Theorem 2.12 and the considerations in the succeeding remark
we have
(16)
for all n 0,m0 ∈ N0 with a constant C independent of n 0 and m0.
n ≤n 0
m≤m0
(cid:13)(cid:13)(cid:13)
l n ,man ,m(cid:12)(cid:12)
Bs
p,q(Rn)(cid:13)(cid:13)(cid:13)k
≤ C · kl bp,qk ∼ k f Bs
p,q(Rn)k
Finally, we use a similar duality argument as in [32, Remark 1.14] or [31, Section 5.1.7] to justify the
dual pairing of k j and f . Looking into the proof of Lemma 2.11, we see that
(17)
n
m (cid:12)(cid:12)(cid:12)(cid:12)
ZRn
l n ,man ,m(x)j (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
for j ∈ C M(Rn) with compact support, M − e ≥ 0 and M − e > s p − s, where C′ depends on the support
of j . This includes the functions k j for j ∈ N0. Because of this absolut convergence the dual pairing of f
and j
is given by
≤ C′ · kj C M−e
(Rn)k · kl bp,¥ k
lim
m0,n 0→¥
m≤m0ZRn
n ≤n 0
l n ,man ,m(x)j (x) dx.
(cid:229)
¥
(cid:229)
¥
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH11
Furthermore, for two different atomic decompositions of f these limits are the same: By definition of
a distribution f ∈ S ′(Rn) and Lemma 2.11 this is valid for j ∈ S (Rn). For arbitrary j ∈ C M(Rn) with
compact support this follows by (17) and density arguments because C¥
(Rn) is dense in C M(Rn) with
respect to the norm of C M−e
C M(Rn) ֒→ BM¥
(Rn). For instance, this can be seen using
,q (Rn) ֒→ BM−e
p,q(Rn) if q < ¥
,¥ (Rn) ֒→ BM−e
(Rn) is dense in Bs
.
,¥ = C M−e
(Rn)
for M − e /∈ N0 and the fact that C¥
Hence we have
l n ,m (an ,m ∗ k j) (x) → ( f ∗ k j)(x) for n 0,m0 → ¥
n ≤n 0
m≤m0
for all x ∈ Rn. Using the standard Fatou lemma and (16) we finally get
k f Bs
p,q(Rn)kk ≤ C · kl bp,qk ∼ k f Bs
p,q(Rn)k.
3. KEY THEOREMS
(cid:3)
3.1. Pointwise multipliers. Triebel proved in Section 4.2 of [28] the following assertion.
Theorem 3.1. Let s ∈ R and 0 < q ≤ ¥
(i) Let 0 < p ≤ ¥ and r > max(s,s p − s). Then there exists a positive number c such that
.
p,q(Rn)k ≤ ckj C
r
(Rn)k · k f Bs
p,q(Rn)k
for all j ∈ C
r
(Rn) and all f ∈ Bs
(ii) Let 0 < p < ¥ and r > max(s,s p,q − s). Then there exists a positive number c such that
kj
kj
f Bs
p,q(Rn).
f F s
p,q(Rn).
for all j ∈ C
r
(Rn) and all f ∈ F s
p,q(Rn)k ≤ ckj C
r
(Rn)k · k f Fs
p,q(Rn)k
is still an (s, p)K,L-atom.
He excluded the cases r ∈ N. This is not necessary in our considerations.
The very first idea to prove this result is to take an atomic decomposition of f , to multiply it by j and
to prove that the resulting sum is again a sum of atoms. Hence one has to check whether a product of an
(s, p)K,L-atom and a function j
But there was a problem: Moment conditions like (7) are (in general) destroyed by multiplication with
j . So the atomic approach in [28] only worked when no moment conditions were required, hence if s > s p
resp. s > s p,q, and the full generality of Theorem 3.1 had to be obtained by an approach via local means.
Looking at condition (5) instead the situation when multiplying by j
Furthermore, the atomic approach only worked for j ∈ Ck(Rn) with k ∈ N and k > s having in mind
condition (6). Now we are able to give weaker conditions using the new atomic approach with condition
(4).
is now different.
We start with a first standard analytical observation.
Lemma 3.2. Let s ≥ 0. There exists a constant c > 0 such that for all f ,g ∈ C s(Rn) the product f · g
belongs to C s(Rn) and it holds
k f · gC s(Rn)k ≤ c · k f C s(Rn)k · kgC s(Rn)k.
Proof. This can be proven using standard arguments, especially Leibniz formula.
(cid:3)
Now we are ready to prove Theorem 3.1. This is done by the following lemma together with Theorem
2.12 using the mentioned technique of atomic decompositions. For some further technicalities see the
upcoming Remark 3.5 or [28, 4.2.2, Remark 1]. This covers also the well-definedness of the product.
Lemma 3.3. There exists a constant c with the following property: For all n ∈ N0, m ∈ Z, all (s, p)K,L-
atoms an ,m with support in d · Qn ,m and all j ∈ C
r
c · kj C
(Rn) with r ≥ max(K,L) the product
(Rn)k−1 · j
r
· an ,m
is an (s, p)K,L-atom with support in d · Qn ,m.
¥
¥
(cid:229)
(cid:229)
12
BENJAMIN SCHARF
Proof. Regarding the conditions (4) on the derivatives Lemma 3.2 gives
k(j
· a)(2−n
·)C K(Rn)k ≤ c · kj (2−n
·)C K(Rn)k · ka(2−n
·)C K(Rn)k
≤ c′ · kj C K(Rn)k · 2−n (s− n
p ).
Now we come to the preservation of the moment conditions (5). By our assumptions there exists a constant
C > 0 such that for every y ∈ C L(Rn) it holds
Using this inequality now for y
instead of y
(cid:12)(cid:12)(cid:12)(cid:12)
y (x)a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
· j
=(cid:12)(cid:12)(cid:12)(cid:12)
y (x)(cid:0)j (x) · a(x)(cid:1) dx(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
≤ C · 2−n κLky C L(Rn)k.
together with Lemma 3.2 it follows
Zd·Qn ,m(cid:0)y (x) · j (x)(cid:1)a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · 2−n κLky
≤ C′ · 2−n κLky C L(Rn)k · kj C L(Rn)k.
· j C L(Rn)k
Hence our lemma is shown.
(cid:3)
Remark 3.4. This is the more general version of part 1 of Lemma 1 in [25] using now the wider atomic
approach from 2.1 which yields a stronger result than in [25].
Remark 3.5. As at the end of Corollary 2.15 we have to deal with some technicalities. We concentrate on
the Bs
p,q(Rn)-case is nearly the same. In principle, Lemma 3.3 shows that
p,q(Rn)-case, the F s
(18)
converges unconditionally in S ′(Rn) where
f = (cid:229)
n
l n ,m(an ,m · j )
l n ,man ,m in S ′(Rn)
n
m
m
and the limit belongs to Bs
p,q(Rn) if f belongs to Bs
p,q(Rn).
To define the product of j and f as this limit, we have to show that the limit does not depend on the
atomic decomposition we chose for f .
Hence we are pretty much in the same situation as at the end of Corollary 2.15: Let at first be j ∈ C¥
.
is a continuous operator mapping S ′(Rn) to S ′(Rn). So (18) converges to
· f for all choices of atomic decompositions of f . Using Lemma 3.3 and the standard Fatou lemma we get
Then the multiplication with j
j
for all f ∈ Bs
p,q(Rn).
For arbitrary j ∈ C
kj
· f Bs
p,q(Rn)k ≤ c · kj C
r
(Rn)k · k f Bs
p,q(Rn)k
r
(Rn) we use a density argument similar to that at the end of Corollary 2.15. We
know
kj ∗ · f Bs
p,q(Rn)k ≤ c · kj ∗C
r −e
(Rn)k · k f Bs
p,q(Rn)k
(Rn), r as in Lemma 3.3 and e small enough. Now using the density of C¥
(Rn) in C
r
(Rn)
for j ∗ ∈ C¥
with respect to the norm of C
r −e
· f Bs
p,q(Rn)k ≤ c · kj C
(Rn) the uniqueness of the product and
p,q(Rn)k ≤ c · kj C
(Rn)k · k f Bs
r −e
r
(Rn)k · k f Bs
p,q(Rn)k
kj
follows.
Remark 3.6. Since
C L(Rn) ֒→ BL¥
(Rn)k by kj B
r
,¥ (Rn) ֒→ C L−e
r
,¥ (Rn)k, even by kj B
(Rn)
r
for L −e ≥ 0, we can replace kj C
in Lemma 3.1.
,q(Rn)k for arbitrary 0 < q ≤ ¥
p,q(Rn)-spaces in Theorem 3.1 can be replaced by r >
max(s,s p − s). This is a matter of complex interpolation, see the proof of the corollary in Section 4.2.2 of
[28].
The condition r > max(s,s p,q − s) for the F s
(cid:229)
(cid:229)
(cid:229)
¥
¥
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH13
r
,¥ (Rn) for r > 0 and r /∈ N. So, let f ∈ Bs
Remark 3.7. Our Theorem 3.1 is a special case of Theorem 4.7.1 in [19]: By Remark 1.6 it holds C
B
Then j ∈ B
p,q(Rn) as well as j ∈ C
p,q(Rn) or f ∈ F s
r ′
,¥ (Rn) for s < r ′ < r . By Theorem 4.7.1 of [19] it holds
p,q(Rn) · B
r ′
,¥ (Rn) ֒→ Bs
r ′
,¥ (Rn) ֒→ F s
p,q(Rn)
Bs
p,q(Rn) · B
resp. F s
p,q(Rn)
r
r
(Rn) =
(Rn) with r > s.
if
r ′ > s
and r ′ > s p − s.
p,q(Rn) these are the same conditions as in Theorem 3.1 - in case of F s
and s + r ′ > s p ⇔ r ′ > s
p,q(Rn) these are even
In case of Bs
better (no dependency on q).
It was not the idea of this paper to give such a detailed and comprehensive treatise as in Runst' and
Sickel's book [19] but to show an application of the more general atomic decompositions where the proof
is easy to follow (see Triebel [28, Section 4.1]).
3.2. Diffeomorphisms. We want to study the behaviour of the mapping
Dj
where f is an element of the function space Bs
map.
: f 7→ f (j (·)),
p,q(Rn) resp. Fs
p,q(Rn) and j
: Rn → Rn is a suitably smooth
One would like to deal with this problem analogously to the pointwise multiplier problem in Section 3.1.
Hence we start with an atomic decomposition of f and composed with j . Then we are confronted with
functions of the form an ,m ◦j originating from the atoms an ,m. This was the idea of Section 4.3.1 in Triebel
[28]. But in general, moment conditions of type (7) are destroyed by this operator. So s > s p resp. s > s p,q
was necessary. As we will see, conditions like (5) behave more friendly under diffeomorphisms.
Furthermore, we are confronted with more difficulties than in section 3.1 because the support of an
atom changes remarkably. In particular, after composing with j
two or more atoms can be associated with
the same cube Qn ,m which is not possible in the atomic representation theorem 2.12. This has not been
considered in detail in Section 4.3.1 by Triebel [28] while there is some work done in the proof of Lemma
3 by Skrzypczak [25].
The special case of bi-Lipschitzian maps, also called Lipschitz diffeomorphisms, is treated in Section 4.3
by Triebel [30]. The main theorem there is used to obtain results for characteristic functions of Lipschitz
domains as pointwise multipliers in Bs
Definition 3.8. Let r ≥ 1.
p,q(Rn) and F s
p,q(Rn).
(i) Let r = 1. We say that the map j
: Rn → Rn is a r -diffeomorphism if j
is a bi-Lipschitzian map, i.e.
that there are constants c1,c2 > 0 such that
(19)
c1 ≤
j (x) − j (y)
x − y
≤ c2.
for all x,y ∈ Rn with 0 < x − y ≤ 1.
(ii) Let r > 1. We say that the one-to-one map j
: Rn → Rn is a r -diffeomorphism if the components
r −1(Rn) for all
j i of j (x) = (j 1(x), . . . ,j n(x)) have classical derivatives up to order ⌊r ⌋ with
i, j ∈ {1, . . . ,n} and if detJ(j )(x) ≥ c for some c > 0 and all x ∈ Rn. Here J(j )(x) stands for the Jacobian
matrix of j at the point x ∈ Rn.
Remark 3.9. It does not matter, whether we assume (19) for all x,y ∈ Rn with x 6= y or for all x,y ∈ Rn
with 0 < x − y < c for a constant c > 0. This is obvious for the upper bound. For the lower bound we have
to use the upper bound of the bi-Lipschitzian property of the inverse j −1 of j . Its existence independent of
the given exact definition of a bi-Lipschitzian map is shown in the following lemma.
Lemma 3.10. Let r ≥ 1.
¶j
i
¶ x j
∈ C
(i) If j
(ii) Let r > 1. If j
(iii) If j
is a 1-diffeomorphism, then j
is bijective and j −1 is a 1-diffeomorphism, too.
is a r -diffeomorphism, then j
is a r -diffeomorphism, then its inverse j −1 is a r -diffeomorphism as well.
is a r ′-diffeomorphism for 1 ≤ r ′ ≤ r . Hence j
is a bi-
Lipschitzian map.
Proof. To prove part (i) we use Brouwer's invariance of domain theorem (see [2]): Since j
: Rn → Rn
is continuous and injective, the image j (U) of U is an open set if U is open. Otherwise, if U is closed,
then also j (U) is closed: If j (xn) → y with xn ∈ U, then xn converges to some x ∈ U by (19) and hence
¥
¥
¥
¥
14
BENJAMIN SCHARF
j (xn) → j (x) = y. Thus j maps Rn to Rn. The inverse j −1 is automatically a bi-Lipschitzian map, see
(19).
The proof of observation (iii) for r ′ > 1 is trivial. Hence, we have to show that every r -diffeomorphism
is a bi-Lipschitzian map for r > 1. The estimate
j (x) − j (y)
x − y
≤ c2
follows from the fact that the derivatives
¶j
i
¶ x j
are bounded for all i, j ∈ {1, . . . ,n}. The formula
(20)
J(j −1)(j (x)) = (J(j )(x))−1
and detJ(j )(x) ≥ c together show that the derivatives of the inverse
are bounded for all i, j ∈
{1, . . . ,n}, for instance using the adjugate matrix formula. By the mean value theorem there exists a c > 0
such that
¶ x j
¶ (j −1)i
j −1(x) − j −1(y)
x − y
≤ c
and so part (iii) is shown.
Finally, for (ii) we have to show that
¶ (j −1)i
¶ x j
part follows from (20) and the boundedness of
way as in the inverse function theorem, starting with
r −1(Rn) and detJ(j −1)(x) ≥ c for r > 1. The latter
. For the first we have to argue inductively in the same
∈ C
¶j
i
¶ x j
It is well known that
J(j −1)(x) =(cid:0)J(j )(j −1(x))(cid:1)−1
A → A−1
is a C¥
(Rn×n)-mapping for invertible A. Together with the upcoming Lemma 3.11 this shows: If the
r −1(Rn) and j −1 is an l-diffeomorphism, then the components of J(j −1)
components of J(j ) belong to C
belong to C min(r −1,l)(Rn) and hence j −1 is a min(l + 1,r )-diffeomorphism. This inductive argument and
the induction starting point that j −1 is a 1-diffeomorphism (by part (i) and (iii)) prove that j −1 is a r -
diffeomorphism. Thus the lemma is shown.
(cid:3)
We go on with a second standard analytical observation.
Lemma 3.11. Let j be a r -diffeomorphism and let max(1,s) ≤ r . Then there exists a constant C depending
on r such that for all f ∈ C s(Rn) it holds
k f ◦ j C s(Rn)k ≤ Cj · k f C s(Rn)k.
Proof. By definition
k f ◦ j C s(Rn)k = k f ◦ j C⌊s⌋(Rn)k + (cid:229)
a =⌊s⌋
a
kD
[ f ◦ j ] lip{s}(Rn)k.
The lemma follows now by using the chain rule and Leibniz rule for spaces of differentiable functions and
for Hölder spaces C s(Rn).
(cid:3)
Remark 3.12. As one can easily see, the constant in Lemma 3.11 depends on
we have a sequence of functions {j m}m∈N and
n(cid:229)
¶j m
i
n(cid:229)
C
< ¥
,
sup
m∈N
i=1
j=1(cid:13)(cid:13)(cid:13)(cid:13)
¶ x j (cid:12)(cid:12)
r −1(Rn)(cid:13)(cid:13)(cid:13)(cid:13)
then there is a universal constant C with Cj m ≤ C, i.e. for all m ∈ N it holds
k f ◦ j mC s(Rn)k ≤ C · k f C s(Rn)k.
¶j
i
¶ x j
C
n(cid:229)
i=1
n(cid:229)
j=1(cid:13)(cid:13)(cid:13)
. If
r −1(Rn)(cid:13)(cid:13)(cid:13)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH15
Lemma 3.13. Let 0 < p ≤ ¥
. Let j
: Rn → Rn be bijective and let there be a constant c > 0 such that
(21)
c ≤
j (x) − j (y)
x − y
for x,y ∈ Rn with x 6= y. Then there is a constant C > 0 such that
(22)
Proof. If p < ¥
, it suffices to prove (22) for
k f ◦ j Lp(Rn)k ≤ C · k f kLp(Rn)k.
where a j ∈ C, A j are pairwise disjoint rectangles in Rn and c A j is the characteristic function of A j. We have
N(cid:229)
f =
a jc A j ,
j=1
Z ( f ◦ j )(x)p dx =Z (cid:12)(cid:12)(cid:12)
N(cid:229)
j=1
a jc j −1(A j)(x)(cid:12)(cid:12)(cid:12)
p
dx =
N(cid:229)
j=1
a jpm (j −1(A j))
because the preimages j −1(A j) are also pairwise disjoint. Hence we have to show:
There is a constant C > 0 such that for all rectangles A it holds
m (j −1(A)) ≤ C · m (A).
(23)
To prove this let Br(x0) = {x ∈ Rn : x − x0 < r} be the open ball around x0 ∈ Rn with radius r > 0. Then
by (21) we have
(24)
Hence there is a constant C > 0 such that
j −1(Br(x0)) ⊂ B r
(j −1(x0)).
c
for all x0 ∈ Rn, r > 0.
Now, we cover a given rectangle A with finitely many open balls {B j}M
j=1 such that
m (j −1(Br(x0))) < C · m (Br(x0))
(25)
m M
[j=1
B j! ≤ 2m (A).
Afterwards we make use of the following Vitali covering lemma: There exists a subcollection B j1, . . . ,B jm
of these balls which are pairwise disjoint and satisfy
Using this, (25) and (24) for the balls 3 · B jk finally gives
M
m
B j ⊂
3 · B jk.
[j=1
[k=1
m (j −1(A)) ≤ m j −1 M
[j=1
B j!! ≤ m j −1 M
[k=1
m (j −1(3 · B jk)) ≤ C ·
m (3 · B jk) ≤ C · 3n ·
M(cid:229)
k=1
M(cid:229)
k=1
=
≤ 2C · 3n · m (A).
.
This proves the result for 0 < p < ¥
For p = ¥ we have to show
j −1(3 · B jk)!
3 · B jk!! = m M
[k=1
m (B jk )
M(cid:229)
k=1
k f ◦ j L¥ (Rn)k ≤ k f L¥ (Rn)k.
This follows from: If m ({x ∈ Rn : f (x) > a)}) = 0, then also m ({x ∈ Rn : f (j (x)) > a}) = 0, which is a
consequence of (23):
Let M be a measurable set with m (M) = 0. Then also m (j −1(M)) = 0.
Hence the lemma is shown for p = ¥
, too.
Remark 3.14. A proof of a more general observation using the Radon-Nikodym derivative and the Lebesgue
point theorem can be found in Corollary 1.3 and Theorem 1.4 of [35] - but here we wanted to give a direct,
more instructive proof for our special situation.
(cid:3)
16
BENJAMIN SCHARF
Remark 3.15. By the previous proof it is obvious that Condition (23) is equivalent to (22) for 0 < p < ¥
.
Condition (23) does not depend on p. For Condition (23) it is necessary that the measure m with m(A) :=
m (j −1(A)) is absolutely continuous with respect to the Lebesgue measure m . In case of p = ¥
this condition
is also sufficient for (23) by the previous proof.
Now we are ready for the main theorem of this section.
Theorem 3.16. Let s ∈ R, 0 < q ≤ ¥ and r ≥ 1.
(i) Let 0 < p ≤ ¥ and r > max(s,1 + s p − s). If j
such that
is a r -diffeomorphism, then there exists a constant c
k f (j (·))Bs
p,q(Rn)k ≤ c · k f Bs
p,q(Rn)k
for all f ∈ Bs
p,q(Rn) onto Bs
(ii) Let 0 < p < ¥ and r > max(s,1 + s p,q − s). If j
p,q(Rn). Hence Dj maps Bs
p,q(Rn).
is a r -diffeomorphism, then there exists a constant
c such that
for all f ∈ F s
p,q(Rn). Hence Dj maps Fs
p,q(Rn) onto Fs
p,q(Rn).
k f (j (·))F s
p,q(Rn)k ≤ c · k f Fs
p,q(Rn)k
Proof. At first, beside the two conditions (4) and (5) we need to take a closer look at the centres and supports
of the atoms. Briefly speaking, the decisive local properties of the set of atoms an ,m are maintained by a
superposition with the diffeomorphism j .
To be more specific: Let Mn = {x ∈ Rn : x = 2−n m,m ∈ Zn}. Having in mind Lemma 3.10 there is a
c2 > 0 with
(26)
By a simple volume argument for Qn ,m and by 2−n m − 2−n m′ ≥ c · 2−n for m 6= m′ there is a constant
M ∼ cn
x − y ≤ c2j −1(x) − j −1(y).
2 such that
for all n ∈ N0,m ∈ Z. Hence we can take our atomic decomposition and split it into M disjunct sums, i.e.
j −1(Mn ) ∩ Qn ,m ≤ M
with
M
f =
M(cid:229)
j=1
n ∈N0
m∈Mn , j
l n ,man ,m
Mn , j = Zn, Mn , j ∩ Mn , j′ = /0 for j 6= j′
[j=1
so that for all n ∈ N0, m ∈ Zn and j ∈ {1, . . . ,M}
(27)
Therefore, not more than one function an ′,m′ ◦ j
(cid:8)m′ ∈ Zn : m′ ∈ Mn , j and j −1(2−n
The support of a function an ,m ◦ j
c1 > 0 with
is located at the cube Qn ,m for each of the M sums.
is contained in j −1(d · Qn ,m) by (3). By Lemma 3.10 there exists a
m′) ∈ Qn ,m(cid:9) ≤ 1.
Hence we get
j −1(x) − j −1(y) ≤
1
c1
x − y.
j −1(d · Qn ,m) ⊂ c ·
· B2−n (j −1(2−n
m)),
d
c1
where Br(x0) = {x ∈ Rn : x − x0 ≤ r}. Hence, together with (27) it follows: There is a constant d′ depend-
ing on c1 such that for every n ∈ N0 and every j ∈ {1, . . . ,M} there is an injective map F
n , j : Mn , j → Zn
with
supp (an ,m ◦ j ) ⊂ d′ · Qn ,F
(28)
for all m ∈ Mn , j. The constant d′ does not depend on n or m.
n , j (m).
(cid:229)
(cid:229)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH17
Thus, if we take the derivative conditions (4) and the moment conditions (5) for an ,m ◦j now for granted
(which will be shown later), then
f j ◦ j = (cid:229)
n ∈N0
m∈Mn , j
l n ,m(an ,m ◦ j )
is an atomic decomposition of the function f j ◦ j . Finally, we have to look at the sequence space norms,
see Definition 2.5.
We will concentrate on the F s
p,q(Rn)-case is easier because it does not matter if
one changes the order of summation over m. By the atomic representation theorem and (28) we will have
p,q(Rn)-case since the Bs
k f j ◦ j Fs
p,q(Rn)k ≤ c(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
l n ,mc (p)
n ,F
1
q
n , j(m)(·)q!
m∈Mn , j
n =0
.
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)
To transfer this into the usual sequence space norm we make use of
n , j (m) ⊂ j −1(c · Qn ,m)
(29)
with a constant c depending on c2 from (26), but independent of n and m. This follows from j −1(2−n m) ∈
Qn ,F
n , j (m). Hence assuming that an ,m ◦ j
fulfil (4) and (5) we obtain
Qn ,F
(30)
k f j ◦ j Fs
p,q(Rn)k ≤ c(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
≤ c′(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
≤ c′(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
¥
l n ,mc (p)
n =0
m∈Mn , j
l n ,mc (p)
n =0
m∈Mn , j
1
q
n ,m(j (·))q!
n ,m(·)q!
n ,m(·)q!
1
q
1
q
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)
Lp(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:12)(cid:12)
l n ,mc (p)
m∈Zn
n =0
p,q(Rn)k.
≤ c′′k f F s
In the first step we used (29), in the second step we used Lemma 3.13 and part (iii) of Lemma 3.10 and in
the last step we applied the atomic decomposition theorem for f . As done in the first step, one can replace
the characteristic function of c · Qn ,m by the characteristic function of Qn ,m in the sequence space norm
getting equivalent norms, see [32, section 1.5.3]. This can be proven using the Hardy-Littlewood maximal
function.
Finally, we have to take a look at the derivative conditions (4) and the moment conditions (5). The latter
part is also considered in Lemma 5 of [25] using the atomic approach with condition (6).
Let an ,m be an (s, p)K,L-atom and let r ≥ max(K,L + 1). If we can show that j ◦ an ,m is an (s, p)K,L-
atom as well, we are done with the proof since we can choose K and L suitably small enough by the atomic
decomposition theorem 2.12. Let Tn (x) := 2−n x and Tn (j ) = T −1n
◦ j ◦ Tn . Then
k (an ,m ◦ j ) (2−n
·)C K(Rn)k = kan ,m ◦ j ◦ Tn C K(Rn)k = kan ,m ◦ Tn ◦ Tn (j )C K(Rn)k.
By a simple dilation argument for the Hölder spaces C
r −1(Rn) it holds
¶ (Tn (j ))i
¶ x j
(cid:13)(cid:13)(cid:13)(cid:13)
C
(cid:12)(cid:12)
r −1(Rn)(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)(cid:13)
C
¶j
i
¶ x j(cid:12)(cid:12)
r −1(Rn)(cid:13)(cid:13)(cid:13)(cid:13)
for all i, j ∈ {1, . . . ,n} and n ∈ N0. Hence by Lemma 3.11 and Remark 3.12 we find a constant C indepen-
dent of n and m such that
k (an ,m ◦ j )(2−n
·)C K(Rn)k = kan ,m ◦ Tn ◦ Tn (j )C K(Rn)k ≤ C · kan ,m(2−n
·)C K(Rn)k
So the derivative condition (4) is shown.
Regarding the moment condition (5) of an ,m ◦j we consider two cases: At first, let j be a r -diffeomorphism
with r > 1. Then j and j −1 are differentiable. We use the moment condition of an ,m itself and Lemma
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
18
3.11 to get
BENJAMIN SCHARF
Z
d′·Qn ,F
n , j (m)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
y (x) · a(j (x)) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
We used the transformation formula for integrals and
Z
j −1(d·Qn ,m)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
y (x) · a(j (x)) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
y (cid:0)j −1(x)(cid:1) · detj −1(x) · a(x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Zd·Qn ,m
≤ C · 2−n κL · kdetj −1(x) ·(cid:0)y ◦ j −1(cid:1) C L(Rn)k
det J(cid:0)j −1(cid:1) ∈ C L(Rn)
≤ C′ · 2−n κL · ky C L(Rn)k.
since j
If r = 1, then L = 0 by our choice of r . This means, that no moment conditions are needed. Hence we
is a r -diffeomorphism with r ≥ L + 1. Furthermore, the sign of detJ(cid:0)j −1(cid:1) is constant.
have nothing to prove. The choice of r = 1 is only allowed if s p < s < 1 resp. s p,q < s < 1.
For some further technicalities similar as in Remark 3.5 see Remark 3.20.
(cid:3)
Remark 3.17. This has been proven (in a sketchy way) in Lemma 3 in [25] for the more special atomic
definition there.
Remark 3.18. If s p < s < 1 resp. s p,q < s < 1, then the choice of r = 1 is possible for these values of s.
This gives the same result as in Proposition 4.1 in [30], where the notation of Lipschitz diffeomorphisms as
in Definition 3.8 is used. This results in
Theorem 3.19. Let 0 < q ≤ ¥
.
(i) Let 0 < p ≤ ¥ and s p < s < 1. If j
: Rn → Rn is a bi-Lipschitzian map, then there exists a constant
c such that
k f (j (·))Bs
p,q(Rn)k ≤ c · k f Bs
p,q(Rn)k.
for all f ∈ Bs
p,q(Rn). Hence Dj maps Bs
(ii) Let 0 < p < ¥ and s p,q < s < 1. If j
p,q(Rn) onto Bs
p,q(Rn).
: Rn → Rn is a bi-Lipschitzian map, then there exists a constant
c such that
for all f ∈ F s
p,q(Rn). Hence Dj maps Fs
p,q(Rn) onto Fs
p,q(Rn).
k f (j (·))F s
p,q(Rn)k ≤ c · k f Fs
p,q(Rn)k.
Remark 3.20. We have to deal with some technicalities of the proof of Theorem 3.16. We concentrate on
the Bs
p,q(Rn)-case is nearly the same.
p,q(Rn)-case, the F s
Let at first be r > 1. In principle, Theorem 3.16 and Lemma 2.11 show that
(31)
n
converges unconditionally in S ′(Rn), where
l n ,m(an ,m ◦ j )
m
f = (cid:229)
l n ,man ,m in S ′(Rn),
n
m
To define the superposition of f and j as this limit, we have to show that the limit does not depend on
p,q(Rn).
and the limit belongs to Bs
p,q(Rn) if f belongs to Bs
the atomic decomposition we chose for f . Let y ∈ C¥
ZRn
l n ,m (an ,m ◦ j ) (x)y (x) dx(cid:12)(cid:12)(cid:12)(cid:12)
m (cid:12)(cid:12)(cid:12)(cid:12)
m (cid:12)(cid:12)(cid:12)(cid:12)
ZRn
= (cid:229)
n
n
(Rn) with compact support be given. Then
l n ,man ,m(x)(cid:2)y (cid:0)j −1(x)(cid:1) · detj −1(x)(cid:3) dx(cid:12)(cid:12)(cid:12)(cid:12)
makes sense, see (17), because by Lemma 3.11 the function y (cid:0)j −1(x)(cid:1) · detj −1(x) has compact support
and belongs to C M(Rn) for a suitable M > 0 with M > s p −s. Now the achievements at the end of Corollary
2.15 show that this integral limit does not depend on the choice of the atomic decomposition for f . Hence
we obtain that the limit in (31) (considered as an element in S ′(Rn)) is the same for all choices of atomic
decompositions.
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
(cid:229)
ATOMIC REPRESENTATIONS IN FUNCTION SPACES AND APPLICATIONS TO POINTWISE MULTIPLIERS AND DIFFEOMORPHISMS, A NEW APPROACH19
If the choice of r = 1 is allowed, then automatically s > s p and Bs
p,q(Rn) consists of regular distributions
p,q(Rn) ⊂ Lp(Rn) for 1 ≤ p ≤ ¥
by Sobolev's embedding. Hence the superposition of f ∈ Bs
f ∈
p,q(Rn) ⊂ L1(Rn) for 0 < p ≤ 1 with a 1-diffeomorphism j
Bs
is defined as the superposition of a regular
distribution with a 1-diffeomorphism and is continuous as an operator from Lp(Rn) resp. L1(Rn) to Lp(Rn)
resp. L1(Rn) by Lemma 3.13.
, then atomic decompositions of f ∈ Bs
p,q(Rn) converge to f with respect to the norm of Lp(Rn)
for 1 ≤ p < ¥
resp. with respect to the norm of L1(Rn) for 0 < p < 1, see [31, Section 2.12]. Hence the
limit does not depend on the choice of the atomic decomposition and is equal to the usual definition of the
superposition of a regular distribution f and the 1-diffeomorphism j .
If p < ¥
resp.
, we use the local convergence of the atomic decompositions of f in L¥ (Rn), i.e. we restrict
f and its atomic decomposition to a compact subset K of Rn. Then this restricted atomic decomposition
converges to the restricted f with respect to the norm of L¥ (K). This suffices to prove uniqueness of the
limit which is an L¥ (Rn)-function.
Remark 3.21. For fixed s, p and q the constant c in Theorem 3.16 depends on the r -diffeomorphism j .
Looking into the proof of Theorem 3.16 and Remark 3.12 the following definition is useful:
Definition 3.22. Let r ≥ 1. We call {j m}m∈N a bounded sequence of r -diffeomorphisms if every j m is a
r -diffeomorphism, if there are universal constants c1,c2 > 0 with
If p = ¥
c1 ≤
j m(x) − j m(y)
x − y
≤ c2
for m ∈ N, x,y ∈ Rn with 0 < x − y ≤ 1 and if - for r > 1 - there is a universal constant c with
n(cid:229)
i=1
¶j m
i
¶ x j (cid:12)(cid:12)
n(cid:229)
j=1(cid:13)(cid:13)(cid:13)(cid:13)
C
r −1(Rn)(cid:13)(cid:13)(cid:13)(cid:13)
< c.
for m ∈ N.
Remark 3.23. If {j m}m∈N is a bounded sequence of r -diffeomorphisms, then (j m)−1 exists for all m ∈ N
and {(j m)−1}m∈N is a bounded sequence of r -diffeomorphisms, too. This follows by the arguments of
Lemma 3.10.
Now, by going through the proof of Theorem 3.16 and Remark 3.12 it follows
Corollary 3.24. Let s ∈ R, 0 < q ≤ ¥ and r ≥ 1.
(i) Let 0 < p ≤ ¥ and r > max(s,1 +s p − s). If {j m}m∈N is a bounded sequence of r -diffeomorphisms,
then there exists a constant C such that
k f (j m(·))Bs
p,q(Rn)k ≤ c · k f Bs
p,q(Rn)k
for all f ∈ Bs
p,q(Rn) and m ∈ N.
then there exists a constant C such that
(ii) Let 0 < p < ¥ and r > max(s,1+s p,q −s). If {j m}m∈N is a bounded sequence of r -diffeomorphisms,
k f (j m(·))F s
p,q(Rn)k ≤ c · k f Fs
p,q(Rn)k
for all f ∈ Fs
p,q(Rn) and m ∈ N.
REFERENCES
[1] G. Bourdaud and W. Sickel. Changes of variable in Besov spaces. Math. Nachr., 198:19 -- 39, 1999.
[2] L. E. J. Brouwer. Zur Invarianz des n-dimensionalen Gebiets. Mathematische Annalen, 72(1):55 -- 56, 1912.
[3] W. Dahmen. Wavelet methods for PDEs - some recent developments. Journal of Computational and Applied Mathematics,
128(1-2):133 -- 185, 2001.
[4] J. Franke. On the spaces of Triebel-Lizorkin type: pointwise multipliers and spaces on domains. Mathematische Nachrichten,
125:29 -- 68, 1986.
[5] M. Frazier and B. Jawerth. Decomposition of Besov spaces. Indiana Univ. Math. Journ., 34:777 -- 799, 1985.
[6] M. Frazier and B. Jawerth. A discrete transform and decompositions of distribution spaces. J. Funct. Analysis, 93:34 -- 170, 1990.
[7] V. M. Gol'dshtein and Y. G. Reshetnyak. Quasiconformal mappings and Sobolev Spaces. Kluwer Academic Publishers, Dor-
drecht, 1990.
[8] V. M. Gol'dshtein and A. S. Romanov. Transformations that preserve Sobolev spaces. Sib. Math. Zh., 25(3):55 -- 61, 1984.
[9] V. M. Gol'dshtein and S. K. Vodop'yanov. Lattice isomorphisms of the spaces L1
n and quasiconformal mappings. Sib. Math. Zh.,
16(2):224 -- 246, 1975.
[10] V. M. Gol'dshtein and S. K. Vodop'yanov. Functional characteristics of quasiisometric mappings. Sib. Math. Zh., 17(4):768 -- 773,
1976.
20
BENJAMIN SCHARF
[11] J. Kahane and P. Lemarie-Rieusset. Fourier Series and Wavelets. Gordon and Breach Publ., Amsterdam, 1995.
[12] H. Koch and W. Sickel. Pointwise multipliers of Besov spaces of smoothness zero and spaces of continuous functions. Rev. Mat.
Iberoamericana, 18(3):587 -- 626, 2002.
[13] I. G. Markina. A change of variable that preserves the differential properties of functions. Sib. Math. Zh., 31(3):422 -- 432, 1990.
[14] V. G. Maz'ya. Weak solutions of the Dirichlet and Neumann problems. Tr. Mosk. Mat. O-va., 20:137 -- 172, 1969.
[15] V. G. Maz'ya and T. O. Shaposhnikova. Theory of Multipliers in Spaces of Differentiable Functions. Pittman, Boston-London,
1985.
[16] V. G. Maz'ya and T. O. Shaposhnikova. Theory of Sobolev Multipliers. Springer, Berlin-Heidelberg, 2009.
[17] V. Netrusov. Theorems on traces and multipliers for functions in Lizorkin-Triebel spaces. Zapiski Nauchnykh Seminarov POMI,
200:132 -- 138, 1992.
[18] J. Peetre. New Thoughts on Besov Spaces. Duke Univ. Math. Ser., Durham, 1976.
[19] T. Runst and W. Sickel. Sobolev spaces of fractional order, Nemytskij operators, and nonlinear partial differential equations. de
Gruyter, Berlin.
[20] V. S. Rychkov. On a theorem of Bui, Paluszy´nski, and Taibleson. Tr. Mat. Inst. Steklova, 227:286 -- 298, 1999.
[21] B. Scharf. Local means and atoms in vector-valued function spaces. Jenaer Schriften zur Mathematik und Informatik,
Math/Inf/05/10, 2010. On ArXiv.org: http://arxiv.org/abs/1103.6159v1.
[22] C. Schneider and J. Vybiral. Non-smooth atomic decompositions, traces on Lipschitz domains, and pointwise multipliers in
function spaces. submitted, 2012.
p,q(Rn) in case s p,q < s < n/p. Ann. Mat. Pura Appl., 174:209 -- 250, 1999.
[23] W. Sickel. On pointwise multipliers for Fs
[24] W. Sickel. Pointwise multipliers of Lizorkin-Triebel spaces. Operator Theory, Advances Appl., 110:295 -- 321, 1999.
[25] L. Skrzypczak. Atomic decompositions on manifolds with bounded geometry. Forum Math., 10(1):19 -- 38, 1998.
[26] R. Strichartz. Multipliers on fractional Sobolev spaces. Journal of Mathematics and Mechanics, 16:1031 -- 1060, 1967.
[27] H. Triebel. Theory of Function Spaces. Birkhäuser, Basel, 1983.
[28] H. Triebel. Theory of Function Spaces II. Birkhäuser, Basel, 1992.
[29] H. Triebel. Fractals and Spectra. Birkhäuser, Basel, 1997.
[30] H. Triebel. Function spaces in Lipschitz domains and on Lipschitz manifolds. Characteristic functions as pointwise multipliers.
Rev. Mat. Complut., 15(2):475 -- 524, 2002.
[31] H. Triebel. Theory of Function Spaces III. Birkhäuser, Basel, 2006.
[32] H. Triebel. Function spaces and wavelets on domains. Publishing House European Math. Soc., Zürich, 2008.
[33] H. Triebel and H. Winkelvoss. Intrinsic atomic characterizations of function spaces on domains. Math. Z., 221(1):647 -- 673, 1996.
[34] A. D. Ukhlov and S. K. Vodop'yanov. Superposition operators in Sobolev spaces. Izv. Vyssh. Uchebn. Zaved. Mat., 10:11 -- 33,
2002.
[35] S. K. Vodop'yanov. Mappings of homogeneous groups and imbeddings of functional spaces. Sib. Math. Zh., 30:25 -- 41, 1989.
[36] A. Zygmund. Smooth functions. Duke Math. J., 12(1):47 -- 76, 1945.
E-mail address: [email protected]
MATHEMATISCHES INSTITUT, FAKULTÄT FÜR MATHEMATIK UND INFORMATIK, FRIEDRICH-SCHILLER-UNIVERSITÄT JENA,
D-07737 JENA, GERMANY
|
1201.2376 | 2 | 1201 | 2012-12-11T17:07:07 | Surfaces Meeting Porous Sets in Positive Measure | [
"math.FA",
"math.CA",
"math.MG"
] | Let n>2 and X be a Banach space of dimension strictly greater than n. We show there exists a directionally porous set P in X for which the set of C^1 surfaces of dimension n meeting P in positive measure is not meager. If X is separable this leads to a decomposition of X into a countable union of directionally porous sets and a set which is null on residually many C^1 surfaces of dimension n. This is of interest in the study of certain classes of null sets used to investigate differentiability of Lipschitz functions on Banach spaces. | math.FA | math |
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
GARETH SPEIGHT
Abstract. Let X be a Banach space and 2 < n < dim X. We show there
exists a directionally porous set P in X for which the set of C 1 surfaces of
dimension n meeting P in positive measure is not meager. If X is separable
this leads to a decomposition of X into the union of a σ-directionally porous
set and a set which is null on residually many C 1 surfaces of dimension n. This
is of interest in the study of Γn-null and Γ-null sets and their applications to
differentiability of Lipschitz functions.
1. Introduction
We investigate the extent to which C1 surfaces meet (directionally) porous sets
(Definition 1.1) in positive measure. By definition each point in a porous set sees
nearby holes in the set of size proportional to their distance away. It is thus intu-
itively clear that porous sets are somehow small. It follows easily from the definition
that porous sets are nowhere dense and hence σ-porous sets are meager. Further, if
P is a porous (directionally porous) set in a Banach space then the Lipschitz map
x 7→ dist(x, P ) is Fr´echet (Gateaux) differentiable at no point of P . If the Banach
space is finite dimensional it follows by the classical Rademacher theorem that P
has Lebesgue measure zero.
Porous sets have been widely studied (see [1] and [2] for surveys of the area) and
recently have been used in the study of differentiability. It has been of much inter-
est to what extent an analogue of Rademacher's theorem, either with Gateaux or
Fr´echet differentiability, holds for Lipschitz functions defined on infinite dimensional
Banach spaces (see [3] for an introduction and [4], [5] for some recent developments
relevant to us).
Since Lebesgue measure is unavailable in infinite dimensional Banach spaces
some other notion of null set is needed to say the set of points of non differentiability
of a Lipschitz function is small. The classes of most interest to us are the Γn-null
sets (Definition 1.2) and Γ-null sets introduced in [4] and [5] respectively. These
are sets which are null on typical n dimensional C1 surfaces (or suitably defined
infinite dimensional surfaces for the Γ-null case).
In separable Banach spaces there are several well known classes of null sets (for
example, Aronszajn null, Haar null and Gauss null sets) with respect to which real
valued Lipschitz functions are Gateaux differentiable almost everywhere (Theorem
6.42 [3], also see [6], [7] and [8]). The Γ-null sets also have this property (Theorem
2.5 [5]). These classes of null sets form σ-ideals (in other words they are closed
under taking subsets and countable unions) so give a reasonable notion of null set.
Since x 7→ dist(x, P ) is Gateaux differentiable at no point of a directionally
porous set P it follows σ-directionally porous sets must be null in all of the above
senses. A Borel set E ⊂ X is Haar null if there exists a Borel probability measure
µ on X such that µ(x + E) = 0 for all x ∈ X. It follows from this definition that
any σ-directionally porous set is null on many lines in X. On the other hand, it can
This work was done while the author was a PhD student of David Preiss and supported by
EPSRC funding. I also thank Jaroslav Tiser for suggesting improvements to the presentation.
1
2
GARETH SPEIGHT
be shown (Theorem 6.39 [3], see also [9]) that any infinite dimensional separable
Banach space contains a σ-porous set whose complement is null on all lines. Such
a σ-porous set is not Haar null. This illustrates that in infinite dimensional spaces
directionally porous sets are much smaller than porous sets. In finite dimensions,
as one might expect, it follows from compactness of the unit sphere that porous
and directionally porous sets coincide.
Proving existence of Fr´echet derivatives is much more difficult. The main known
result is if a Banach space X has separable dual then every real valued Lipschitz
function on X is Fr´echet differentiable on a dense set [10]. However, this is not an
'almost everywhere' type result. Indeed, it is not even known if three real valued
Lipschitz functions on a separable Hilbert space have a common point of Fr´echet
differentiability [4].
It has recently been shown if X ∗ is separable then every real valued Lipschitz
map on X is Fr´echet differentiable outside a Γ-null set if and only if every σ-porous
subset of X is Γ-null (Corollary 3.12 [5]). This implies, for example, that any
real valued Lipschitz function on c0 or C(K) (for K countable compact) is Fr´echet
differentiable outside a Γ-null set (Theorem 4.6 [5]). It is known (Theorem 5.4.2
[4]) that Gδ sets are Γ-null if they are Γn-null for infinitely many n. Thus it is
desirable to understand when porous sets are either Γn-null or Γ-null.
It can be shown if n ≥ dim X then Γn-null and Lebesgue null sets coincide
(Theorem 5.3.8 [4]). If n < dim X the situation is much more interesting. Every
σ-porous subset of a Banach space with separable dual is Γ1-null (Theorem 10.4.1
[4]) and every σ-directionally porous subset of a separable Banach space is Γ1-
null and Γ2-null (Theorem 10.4.2 [4]). We show the situation is very different for
higher dimensional surfaces - even directionally porous sets need not be Γn null
when 2 < n < dim X (Theorem 1.4) and if X is separable the complement of a
σ-directionally porous set may be Γn-null (Theorem 1.5).
One might ask if there is any differentiability result using only the notion of
Γn-null sets rather than Γ-null sets. It has been shown (Theorem 11.3.6 [4]) that
if X ∗ is separable and every porous set in X can be decomposed into the union
of a σ-directionally porous set and a Γn-null set of class Gδ, then Lipschitz maps
on X into Banach spaces of dimension not exceeding n have points of ε-Fr´echet
differentiability for every ε > 0. At the time it was not yet known if porous sets
in some infinite dimensional spaces were necessarily Γn-null. By showing even
directionally porous sets in infinite dimensional Banach spaces need not be Γn-null
we answer a question posed in [4] (pages 186 and 203) and show that for the theorem
mentioned to be meaningful the more complicated hypothesis is necessary. That
is, it could not instead be more simply assumed porous sets are Γn-null.
We now give the formal definitions that will be relevant for us. In what follows
B(x, r) will denote the open ball in X with centre x ∈ X and radius r > 0.
Definition 1.1. A set P ⊂ X is called porous if there exists 0 < ρ < 1 such that
for all x ∈ P and δ > 0 there exists y ∈ X with ky − xk < δ such that
B(y, ρky − xk) ∩ P = ∅.
We refer to the ball B(y, ρky − xk) as a hole in P .
A set P ⊂ X is called directionally porous if there exists 0 < ρ < 1 such that
for all x ∈ P there exists v ∈ X with kvk = 1 such that for any δ > 0 there exists
t ∈ R with t < δ such that
We refer to the constant ρ appearing in the above definitions as a porosity
constant of P .
B(x + tv, ρt) ∩ P = ∅.
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
3
A set is called σ-porous (σ-directionally porous) if it is a union of countably
many porous (directionally porous) sets.
Recall a subset of a metric space is called typical, or residual, if its complement
is meager.
Definition 1.2. A C1 surface of dimension n in X is a C1 map f : [0, 1]n → X.
We define the C1 norm by
kfkC 1 = max(kfk∞,k∂f/∂x1k∞, . . . ,k∂f/∂xnk∞)
where k · k∞ denotes the supremum norm. Denote the space of C1 surfaces of
dimension n in X with C1 norm by Γn(X).
A set N ⊂ X is called Γn-null if
Hn(f ([0, 1]n) ∩ N ) = 0
for residually many f ∈ Γn(X).
Remark 1.3. In [4] a set N ⊂ X is defined to be Γn-null if
Ln{t ∈ [0, 1]n : f (t) ∈ N} = 0
for residually many f ∈ Γn(X).
It is easy to see a set which is Γn-null in this
sense is also Γn-null in the sense of Definition 1.2. In fact the two definitions are
equivalent for n ≤ dim X. This follows from Lemma 5.3.5 [4] which states that for
a typical surface f ∈ Γn(X) the derivative df has rank equal to min(n, dim X) at
Lebesgue almost every point of [0, 1]n.
We can now state precisely the results of this paper.
Theorem 1.4. Let 2 < n < dim X. Then there exists a directionally porous set
P ⊂ X which is not Γn-null.
Theorem 1.5. Let X be separable and 2 < n < dim X. Then there exists a
σ-directionally porous set Q ⊂ X and a Γn-null set N ⊂ X such that X = Q ∪ N .
Intuitively the reason for the difference between the cases n ≤ 2 and n > 2 is
that for n > 2 modifying a surface to go through a nearby hole causes an area
change comparable to the size of the hole (this is stated precisely in Proposition 4.2
which is a corollary of a Poincar´e type inequality). By a careful construction we
exploit this fact to construct a porous set where the size of holes a surface meets is
controlled by the area of the surface.
Fix n > 2 throughout the remainder of the paper. We focus on proving Theorem
1.4 in the case X = Rn+1. Since porous sets and directionally porous sets coincide
in finite dimensions it suffices to construct a porous set in Rn+1 which is not Γn-
null. In the final section we deduce the result for a general Banach space and see
Theorem 1.5 follows.
2. Geometric conditions for the construction
Let A be a subset of a complete metric space M and Uk be a sequence of open
k=1 Uk ⊂ A then U \ A
is meager in M . Hence by the Baire Category Theorem A is not meager in M . We
now investigate what this means in terms of surfaces and porous sets.
sets which are each dense in a fixed open set U ⊂ M . IfT∞
Define the C1 surface p : [0, 1]n → Rn+1 by p(x) = (x, 0) and and let c =
(1/2, . . . , 1/2) ∈ Rn. Fix 0 < r < 1/32 to be chosen small later.
If g : B(c, t) → R let
G(g) = {(x, g(x)) : x ∈ B(c, t)}
denote the graph of g.
It will usually be simpler to work with surfaces represented as graphs.
4
GARETH SPEIGHT
Lemma 2.1. There exists s > 0 (fixed and independent of r) and δ(r) > 0 such
that for all f ∈ BC 1 (p, δ(r)) there exists g : B(c, s) → R of class C1 with kgkC 1 ≤ r
and
f ([0, 1]n) ∩ (B(c, s) × R) = G(g).
Proof. Let e1, . . . , en denote the standard basis of Rn. We apply the Inverse Func-
depend only on bounds on its derivative (this is clear from the proof of the Inverse
Function Theorem given in [11]).
tion Theorem to ef = (f1, . . . , fn) which is a C1 mapping from [0, 1]n to Rn and
satisfies k∂ef /∂xi− eik∞ < δ(r) for i = 1, . . . , n. Note the size of the neighbourhood
on which ef is invertible can be made independent of the particular function f and
Provided δ(r) is sufficiently small (independently of f ) ef is invertible on a region
whose image contains c and has a C1 inverse (ef )−1 : B(c, s) → Rn with s > 0
independent of r, f and k∂(ef )−1/∂xik∞ ≤ 2 for i = 1, . . . , n. Define the C1 map
g : B(c, s) → R by g = fn+1 ◦ (ef )−1. Since kfn+1kC 1 < δ(r) we have kgkC 1 ≤ r for
Let F be a (not necessarily disjoint) collection of open balls in Rn+1 and L > 1.
Intuitively F will correspond to the holes of a porous set. Define, for k ≥ 1,
(2.1)
sufficiently small δ(r).
B,
(cid:3)
Pk = [B∈F
diamB<1/k
LB, H = [B∈F
where LB denotes the ball with the same centre as B but radius enlarged by
factor L.
Notice Pk is open and Pk+1 ⊂ Pk for all k ≥ 1. Further
P = ∞\k=1
Pk! \ H
is porous with porosity constant 1/L. InformallyT∞
Given A ⊂ Rn+1 denote the set of n dimensional surfaces which meet A in n
k=1 Pk consists of points which
have a nearby hole, of radius proportional to its distance from the point, of diameter
less than 1/k and H is the union of all the holes (of any size).
dimensional measure greater than α ≥ 0 by
Sn(A, α) = {f ∈ Γn(X) : Hn(A ∩ f ([0, 1]n)) > α}.
To prove Theorem 1.4 in the case X = Rn+1 it suffices to construct F and L > 1
for which there is α > 0 such that, for sufficiently small r,
Sn((B(c, s) × R) ∩ Pk, α) is dense in BC 1(p, δ(r)) for all k ≥ 1
(2.2)
and
(2.3)
As the sets Pk are decreasing it follows
Hn(f ([0, 1]n) ∩ (B(c, s) × R) ∩ H) ≤ α/4 for all f ∈ BC 1(p, δ(r)).
(B(c, s) × R) ∩ Pk, α/2! .
∞\k=1
Sn((B(c, s) × R) ∩ Pk, α) ⊂ Sn ∞\k=1
Since the sets Pk are open each of the sets Sn((B(c, s) × R) ∩ Pk, α) is also open.
By (2.2) they are also dense in BC 1(p, δ(r)) so it follows, by the discussion at the
start of this section, that the complement of
Sn ∞\k=1
(B(c, s) × R) ∩ Pk, α/2!
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
5
is meager in BC 1 (p, δ(r)). Then by (2.3),
Sn ∞\k=1
(B(c, s) × R) ∩ Pk, α/2! ⊂ Sn(P, α/4)
so the complement of Sn(P, α/4) is meager in BC 1(p, δ(r)). By the Baire Category
Theorem, as in the discussion at the start of this section, this will prove Theorem
1.4 in the case Rn+1.
Thus proving Theorem 1.4 amounts to constructing smaller and smaller open
balls (intuitively the holes of the porous set) whose enlargements mostly cover
surfaces in a countable dense set and so that the intersection of all balls with any
one surface is kept small.
3. Construction of the porous set
One of the requirements on F is that the enlargements of balls in F of arbitrarily
small radii must cover a fixed proportion of each surface in a countable dense subset
of BC 1(p, δ(r)). To do this it is natural to choose our countable dense subset to be
as simple as possible.
The following lemma follows from Corollary 10.2.2 [4], which states surfaces in
which almost every point has a neighbourhood on which the surface is affine are
dense in Γn(Rn+1), and allows us to choose a countable dense subset consisting of
surfaces that are mostly covered by countably many planes.
Let ωn denote the volume of a unit ball in Rn.
Lemma 3.1. There is a countable dense set of surfaces {fl}∞
l=1 in BC 1(p, δ(r)) and
a countable collection of affine planes Al(x) = (x, al(x)), where x ∈ [0, 1]n, ∇al is
constant and ∇al ≤ r, such that
Al([0, 1]n)! < ωnsn/4.
fl([0, 1]n) \
Hn ∞[l=1
∞[l=1
Without loss of generality let f1 = A1 = p.
We will define F and L > 1 so that, if Pk is defined as in (2.1),
Hn(Al([0, 1]n) ∩ (B(c, s) × R) \ Pk) ≤ ωnsn/2l+2 for all k, l ≥ 1.
(3.1)
This will imply
Hn(fl([0, 1]n) ∩ (B(c, s) × R) ∩ Pk) > ωnsn/2 for all k, l ≥ 1
and hence fl ∈ Sn((B(c, s) × R) ∩ Pk, ωnsn/2) for all k, l ≥ 1 which implies (2.2)
with α = ωnsn/2.
In Section 5 we then show that if r is sufficiently small our construction ensures
and 3P∞
Hn(f ([0, 1]n) ∩ (B(c, s) × R) ∩ H) ≤ ωnsn/8 for all f ∈ BC 1(p, δ(r))
which is (2.3) with α = ωnsn/2. Theorem 1.4 hence follows in the case X = Rn+1.
Choose a sequence mk of natural numbers with m1 = 1 in which every natural
number is repeated infinitely many times. Fix a sequence εi > 0 with εi < 1/2i
i=1 εi ≤ 1/64 to be chosen as small as required later.
Let r0 = s. For k ≥ 1 we will inductively define Fk and rk > 0 such that Fk
consists of finitely many balls in Rn+1, of radius less then ε3
krk−1 and greater than
rk, whose enlargements (by a factor independent of k) mostly cover Amk ([0, 1]n).
The family Fk will consist of subfamilies of balls on different levels relative to
Amk constructed so there are relatively few balls on any one level. Each of these
subfamilies of Fk is formed by lifting families of balls Gl
and hence define Fk and rk. Let G0
k = rk−1.
k in Rn to Rn+1.
Fix k ≥ 1 for which rk−1 has been defined. We show how to define Gl
k = ∅ and r0
k inductively
6
GARETH SPEIGHT
k of finitely many balls in Rn
Lemma 3.2. Fix l ≥ 0. Suppose families G0
and r0
k, . . . ,Gl
k > 0 have been constructed so that (if l ≥ 1):
k consists of balls of radius rq
k, . . . , rl
• For each 1 ≤ q ≤ l, Gq
• For each 1 ≤ q ≤ l, the balls in
{(1/ε3
k)B : B ∈ Gq
k}
are disjoint and contained inside B(c, s).
k ≤ ε3
krq−1
k
.
• Any two balls from
{(1/ε3
k)B : B ∈ G1
k ∪ . . . ∪ Gl
k}
are either disjoint or one is contained inside the other.
Then there exists a family Gl+1
ε3n
k /2n+1 proportion of
k
of finitely many balls in Rn which cover at least
B(c, s) \ [B∈G1
k∪...∪Gl
k
B
and 0 < rl+1
krl
k ≤ ε3
• Balls in Gl+1
• The balls in
k
k such that:
have radius rl+1
k
.
k)B : B ∈ Gl+1
k }
are disjoint and contained inside B(c, s).
{(1/ε3
• Any two balls from
{(1/ε3
k)B : B ∈ G1
k ∪ . . . ∪ Gl+1
k }
are either disjoint or one is contained inside the other.
Proof. Find 0 < rl+1
krl
k ≤ ε3
k such that at least half of the points in
B(c, s) \ [B∈G1
k∪...∪Gl
k
B
are of distance at least rl+1
k /ε3
k away from the closed set
(Rn \ B(c, s)) ∪ [B∈G1
k∪...∪Gl
k
∂((1/ε3
k)B).
k from centres of previously chosen balls, we can find a finite family Gl+1
k
, whose centres are of distance at least
of
k } are disjoint, contained
k)B : B ∈ Gl+1
k /ε3
By inductively choosing balls of radius rl+1
2rl+1
balls of radius rl+1
in
k
k
such that balls in {(1/ε3
B(c, s) \ [B∈G1
k∪...∪Gl
k
∂((1/ε3
k)B)
and balls in {(2/ε3
k)B : B ∈ Gl+1
k } cover at least half of
B(c, s) \ [B∈G1
k∪...∪Gl
k
B.
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
7
Then
Ln [B∈Gl+1
k
as required.
B = XB∈Gl+1
k
= (ε3
k)B)
Ln((2/ε3
Ln(B)
k/2)n XB∈Gl+1
k/2)nLnB(c, s) \ [B∈G1
k
≥ (ε3
B /2
(cid:3)
k∪...∪Gl
k
Once Gl
k has been defined if
LnB(c, s) \ [B∈G1
k∪...∪Gl
k
B ≤ ωnsn/2k+3
we stop, set rk = rl
k and let
Gk = G1
Otherwise we continue and define Gl+1
as in Lemma 3.2. Since at each
stage we cover a fixed proportion (independent of l) of the region not yet covered
this process stops in a finite number of steps.
To define Fk from Gk we replace each ball
k ∪ . . . ∪ Gl
k.
and rl+1
k
k
by
B(x, t) ∈ Gk
B((x, amk (x) + 2t), t) ∈ Fk.
Inductively define Fk and rk as above for all k ≥ 1, then let F = ∪∞
L = √10.
Proposition 3.3. If Pk are defined from F , L as in (2.1) then
k=1Fk and
Hn(Amk ([0, 1]n) ∩ (B(c, s) × R) \ Pk) ≤ ωnsn/2k+2
for all k ≥ 1.
Proof. Notice if B(x, t) ∈ Gk then, since ∇amk ≤ 1,
Hence
Amk ([0, 1]n) ∩ (B(x, t) × R) ⊂ B((x, amk (x) + 2t), t√10) ⊂ Pk.
Hn(Amk ([0, 1]n) ∩ (B(c, s) × R) \ Pk) ≤ Hn Amk ([0, 1]n) ∩ B(c, s) \ [B∈Gk
≤ 2Ln B(c, s) \ [B∈Gk
B!
≤ ωnsn/2k+2.
B! × R!
(cid:3)
Since the sequence mk takes the value of every natural number infinitely often
and the sets Pk are decreasing, Proposition 3.3 implies (3.1).
8
GARETH SPEIGHT
4. Area estimates for surfaces
BC 1(p, δ(r)) is small.
It remains to show the intersection of balls in F with any fixed surface in
We now establish several results which later allow us to show how passing through
many holes forces the area of a surface f to increase and to distinguish between
area increments arising from holes of different sizes.
The following lemma is an adaptation of Theorem 1(iii) in Section 5.6.1 of [12].
The proof is essentially the same as in [12] but with the numbers 1 and n/(n − 1)
replaced by 2 and its corresponding Sobolev conjugate 2n/(n − 2). Notice the
assumption n > 2 is essential.
In what follows C will be a positive constant (depending only on n) whose value
may be different in different expressions. We will sometimes write 2C instead of C
if it is important but not obvious an extra constant term has been added.
Lemma 4.1. For each 0 < α ≤ 1 there exists a constant C(α) > 0 such that
for all balls B ⊂ Rn and g ∈ W n,2
kgkL2n/(n−2)(B) ≤ C(α)k∇gkL2(B)
loc (Rn) satisfying
Ln(B ∩ {g = 0})
Ln(B)
≥ α.
≥ α.
Proof. Suppose B = B(x, t) ⊂ Rn and g ∈ W n,2
Ln(B ∩ {g = 0})
loc (Rn) satisfies
By the triangle inequality
Ln(B)
kgkL2n/(n−2)(B) ≤ kg − (g)BkL2n/(n−2)(B) + (g)B(Ln(B))(n−2)/2n
where
(g)B =
1
Ln(B)ZB
g.
We estimate the two terms individually.
Since n > 2, Poincar´e's inequality (Theorem 2 of Section 4.5.2 [12]) states,
(cid:18)−ZB g − (g)B2n/(n−2)(cid:19)(n−2)/2n
≤ Ct(cid:18)−ZB ∇g2(cid:19)1/2
.
As Ln(B) = ωntn this implies
Next notice
kg − (g)BkL2n/(n−2)(B) ≤ Ck∇gkL2(B).
(g)B(Ln(B))(n−2)/2n ≤ (Ln(B))−(n+2)/2nZB gχ{g6=0}
where χA denotes the characteristic function of a set A. Applying Holder's inequal-
ity with exponents 2n/(n − 2) and 2n/(n + 2) implies
(g)B(Ln(B))(n−2)/2n ≤ (Ln(B))−(n+2)/2nkgkL2n/(n−2)(B)(Ln(B∩{g 6= 0}))(n+2)/2n.
The assumption on g then yields
(g)B(Ln(B))(n−2)/2n ≤ (1 − α)(n+2)/2nkgkL2n/(n−2)(B).
Putting the two estimates together implies
kgkL2n/(n−2)(B) ≤ Ck∇gkL2(B) + (1 − α)(n+2)/2nkgkL2n/(n−2)(B).
Rearranging this expression and relabelling constants leads to the desired inequality.
(cid:3)
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
9
Our use of Lemma 4.1 is expressed in the following proposition. The underlying
idea will be that if a surface passes through different vertical levels it creates an
area change comparable to holes at those heights.
Proposition 4.2. Suppose h > 0 and B ⊂ Rn is a ball of radius at least h. Then
ωnhn ≤ CZB∩{g≥h/2} ∇g2
for all g : B → R of class C1 with ∇g ≤ 1 such that
Ln(B ∩ {g ≤ h/2})
≥ 1/2
Ln(B)
and g(ex) ≥ h for some ex ∈ B.
Proof. Let g : B → R be as in the statement of the proposition. Extend g to a
function in W n,2
loc (Rn) and let (g − h/2)+ denote the function which equals g − h/2
if g ≥ h/2 and is zero otherwise. Then (g−h/2)+ belongs to W n,2
loc (Rn) and satisfies
Ln(B ∩ {(g − h/2)+ = 0})
Ln(B)
≥ 1/2.
Hence by Lemma 4.1
k(g − h/2)+kL2n/(n−2)(B) ≤ C(1/2)k∇(g − h/2)+kL2(B).
Since g is C1, ∇g ≤ 1 in B, g(ex) ≥ h and B has radius at least h it follows there
is C > 0 such that g ≥ 3h/4 inside a region of measure at least Chn and contained
in B. Hence
((h/4)2n/(n−2)Chn)(n−2)/2n ≤ C(1/2) ZB∩{g≥h/2} ∇g2!1/2
.
Simplifying this expression and relabelling the constants gives the claimed inequal-
ity.
(cid:3)
Our holes were constructed relative to planes with varying directions. The fol-
lowing proposition will later allow us to use the previous result in this context.
Proposition 4.3. Let B = B(x, t) ⊂ Rn, g : B → R of class C1 with ∇g ≤ 1
and a : B → R of class C1 with ∇a constant and ∇a ≤ 1. Suppose ε > 0 and
g − a ≤ εt in B. Then
Proof. Write g = (g − a) + a and expand the terms to obtain
∂a
∂xi
(cid:12)(cid:12)(cid:12)(cid:12)ZB
(∇g2 − ∇a2 − ∇(g − a)2)(cid:12)(cid:12)(cid:12)(cid:12) ≤ CεLn(B).
(∇g2 − ∇a2 − ∇(g − a)2)(cid:12)(cid:12)(cid:12)(cid:12) = 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZB
nXi=1
nXi=1(cid:12)(cid:12)(cid:12)(cid:12)ZB
≤ C
≤ CεLn(B).
(cid:12)(cid:12)(cid:12)(cid:12)ZB
∂(g − a)
∂xi
(cid:12)(cid:12)(cid:12)(cid:12)
∂(g − a)
∂xi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
using the Divergence Theorem and the assumption g − a ≤ εt.
(cid:3)
In what follows we will need to smooth our surfaces in order to distinguish
between oscillations arising from different sized holes. The following definition
recalls some basic notions [12].
10
GARETH SPEIGHT
Definition 4.4. Define the C∞ function η : Rn → R by η(x) = κ exp(1/(x2 − 1))
for x < 1 and η(x) = 0 for x ≥ 1, where κ > 0 is chosen so that
ZRn
η(x) dx = 1.
For ε > 0 define the standard mollifier ηε : Rn → R by ηε(x) = (1/εn)η(x/ε). If
g ∈ L1
loc(U ) and ε > 0 define
gε(x) =ZB(x,ε)
ηε(x − y)g(y) dy
for x ∈ Uε = {x ∈ U : dist(x, ∂U ) > ε}.
We informally refer to gε as a smoothing of g. The following facts are either well
known [12] or easy to prove using the definition.
Lemma 4.5. Let 0 < ε < t and g : B(x, t) → R be of class C1 and satisfy ∇g ≤ 1
in B(x, t). Then gε is of class C∞(B(x, t − ε)) and gε − g ≤ ε.
The following proposition gives an estimate on the smoothing of a function whose
values are inside a small interval.
Proposition 4.6. Let B = B(x, t) ⊂ Rn and 0 < ε < 1/2. Suppose g : B → R is
of class C1 with ∇g ≤ 1 and g ≤ ε2t in B. Then
Proof. For x ∈ B(x, t − εt), using the formula for the derivative of a convolution,
ZB(x,t−εt) ∇(gεt)2 ≤ Cε2Ln(B).
(x − y)g(y) dy(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(x)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZB(x,εt)
≤ZB(x,εt)
≤ ωnεntn(C/εn+1tn+1)ε2t
= Cε.
(C/εn+1tn+1)ε2t dy
∂ηεt
∂xi
∂gεt
∂xi
(cid:12)(cid:12)(cid:12)(cid:12)
Hence ∇(gεt) ≤ Cε in B(x, t − εt) so the desired inequality follows.
(cid:3)
We will need to smooth locally at different scales in different regions. The follow-
ing proposition gives a function which allows us to interpolate smoothly between
two functions.
Proposition 4.7. Let B(x, t) ⊂ Rn and 0 < ε < 1/2. Then there exists a function
w : Rn → R of class C1 with 0 ≤ w ≤ 1 such that:
• w = 1 inside B(x, t − 2εt).
• w = 0 outside B(x, t − εt).
• ∇w ≤ 3/(εt) in B(x, t − εt) \ B(x, t − 2εt).
Suppose g1 : B(x, t − εt) → R and g2 : B(x, t) → R are C1 with bounded deriva-
Then the function v : B(x, t) → R defined to be wg1 + (1 − w)g2 in B(x, t − εt)
tives and g1 − g2 ≤ ε2t in B(x, t − εt) \ B(x, t − 2εt).
and g2 in B(x, t) \ B(x, t − εt) is C1 with
∇v ≤ max(∇g1,∇g2) + 3ε
in B(x, t − εt).
Proof. Define ew : Rn → R such that:
• ew = 1 inside B(x, t − 5εt/3).
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
11
• ew = 0 outside B(x, t − 4εt/3).
• ew interpolates between 0 and 1 as a linear function of distance to x.
It is easy to see the function w = ewεt/3 then has the properties required.
∇(wg1 + (1 − w)g2) = w∇g1 + (1 − w)∇g2 + ∇w(g1 − g2)
Clearly
in B(x, t − εt) so
in B(x, t − εt). Since w is C1 and equal to zero outside B(x, t − εt) it also follows
v is C1 on B(x, t).
(cid:3)
∇v ≤ max(∇g1,∇g2) + 3ε
5. Holes are small on surfaces
We now need to show (2.3) with α = wnsn/2 which intuitively states that the
intersection of surfaces in BC 1(p, δ(r)) with (B(c, s)× R)∩H is small. To do this we
control the intersection of any surface f ∈ BC 1(p, δ(r)), which is the graph of some
function g over B(c, s), with H usingRB(c,s) ∇g2. This integral is approximately
the n dimensional area difference between the image of f over B(c, s) and the flat
disc B(c, s)×{0}. The intuition is thus that passing through holes forces the surface
to oscillate in a way that increases its area.
If x ∈ Rn+1 let x′ ∈ Rn be the point consisting of the first n coordinates of x. If
B ⊂ Rn+1 is a ball of radius t > 0 let B = ωntn be the area of an n dimensional
cross section of B.
Proposition 5.1. There is a constant C ≥ 1 such that
B ≤ CZB(c,s) ∇g2 + C
for all g : B(c, s) → R of class C1 satisfying kgkC 1 ≤ 1/64.
Proof of Proposition 5.1. To prove the proposition we will reformulate it in a way
that allows us to use induction (Lemma 5.5 and Lemma 5.6). At times in the
argument we will interpolate between pieces of the surface which are smoothed
on different scales. Thus it is natural to control the intersection of surfaces with
slightly enlarged holes and begin with weaker bounds on ∇g which we tighten
after each induction step.
For k ≥ 1 let Kk = 2 −Pk
i=1(1/2)i. If B = B(x, t) ∈ Fk for some k ≥ 1 denote
k and B′ = B(x′, t′) ⊂ Rn. Note this definition makes sense because any
t′ = t/ε3
ball can be in Fk for at most one k ≥ 1. Given g : B(c, s) → R of class C1 with
∇g ≤ 1/32 if B ∈ F and G(g) ∩ K1B 6= ∅ define
XB∈F
∞Xi=1
G(g)∩B6=∅
εi
Rk(B) = B′ ∩ {g − amk > t/4} ⊂ B(c, s).
Note it will be clear from the context which map g is meant when Rk(B) appears.
Claim 5.2. If k ≥ 1 is fixed and B1, B2 ∈ Fk are distinct balls with G(g)∩ K1B1 6=
∅ and G(g) ∩ K1B2 6= ∅ then Rk(B1) ∩ Rk(B2) = ∅.
Proof of Claim 5.2. Let B1 = B(x, t1) and B2 = B(y, t2) ∈ Fk.
since {(1/ε3
2 = ∅ so
If there exists l ≥ 1 such that both B(x′, t1) and B(y′, t2) belong to Gl
k} is a disjoint family, B′
k)B : B ∈ Gl
k then,
1 ∩ B′
Rk(B1) ∩ Rk(B2) = ∅.
Suppose not and t1 ≥ t2. Then, since radii of balls in different families Gl
k differ
kt1 ≥ t2. Note, from the definition of Fk in terms
k, necessarily ε3
by factor at least ε3
12
GARETH SPEIGHT
of Gk, that if B(z, t) ∈ Fk then zn+1 − amk (z′) = 2t. Since G(g) ∩ K1B2 6= ∅
there exists ey ∈ B(y′, t2) such that
Hence
g(ey) − amk (ey) ≤ 4t2.
g − amk ≤ 4t2 + 2t′
2/16 ≤ t1/4
in B′
has diameter 2t′
2 since ∇g ≤ 1/32 and ∇amk ≤ 1/32 implies ∇(g − amk ) ≤ 1/16 while B′
2
2 = ∅ which again implies
2. Hence Rk(B1) ∩ B′
as required.
Rk(B1) ∩ Rk(B2) = ∅
For each k ≥ 1 we define two subfamilies of Fk.
and
F u
k = {B ∈ Fk : G(g) ∩ K1B 6= ∅,B ≤ εkLn(Rk(B))}
F d
k = {B ∈ Fk : G(g) ∩ K1B 6= ∅,B > εkLn(Rk(B))}.
Notice {B ∈ Fk : G(g) ∩ K1B 6= ∅} = F u
Intuitively F u
k consists of holes where the surface stays mostly on the same level
around the hole. Since there are relatively few holes on any one level we can show
the holes in this family are small.
k ∪ F d
k .
Claim 5.3. For each k ≥ 1, XB∈F u
k
B ≤ εk.
Proof of Claim 5.3. By definition B ≤ εkLn(Rk(B)) for all B ∈ F u
Claim 5.2,
k . Hence, using
(cid:3)
(cid:3)
(cid:3)
B ≤ εk XB∈F u
k
Ln(Rk(B)) ≤ εkLn [B∈F u
k
Rk(B) ≤ εk.
XB∈F u
Intuitively F d
k
k consists of holes where the surface goes to other levels around the
hole. Using results from Section 4 we can show this forces an increase in relative
area.
Claim 5.4. For each k ≥ 1 and B ∈ F d
k ,
B ≤ CZRk(B) ∇(g − amk )2.
Proof of Claim 5.4. Notice if B = B(x, t) ∈ F d
k then
Ln(B′ ∩ {g − amk ≤ t/4})
= Ln(B′ \ Rk(B))
Ln(B′)
Ln(B′)
≥ 1/2.
Proposition 4.2 with h = t/2,
Since G(g) ∩ B 6= ∅ there existsex ∈ B′ such that g(ex)− amk (ex) ≥ t/2. Hence, by
B ≤ CZRk(B) ∇(g − amk )2.
Lemma 5.5. Fix g : B(c, s) → R of class C1 satisfying ∇g ≤ 1/32. Then
XB∈F1
G(g)∩K1B6=∅
B ≤ CZB(c,s) ∇g2 + ε1.
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
13
Proof of Lemma 5.5. Since Am1 = A1 = p and so am1 = 0 Claim 5.4 implies
B ≤ CZ∪B∈Fd
1
R1(B) ∇g2 ≤ CZB(c,s) ∇g2
XB∈F d
1
which, together with Claim 5.3, proves the lemma.
(cid:3)
Lemma 5.6. Suppose we have shown
G(g)∩KkB6=∅
XB∈F1∪...∪Fk
B ≤ CZB(c,s) ∇g2 + C
kXi=1
for all g : B(c, s) → R of class C1 satisfying ∇g ≤ 1/32 − 3Pk
k+1Xi=1
B ≤ CZB(c,s) ∇g2 + C
XB∈F1∪...∪Fk+1
Then
εi
εi
G(g)∩Kk+1B6=∅
i=1 εi.
i=1 εi.
Further, the constants can be chosen so as to remain bounded as the lemma is
for all g : B(c, s) → R of class C1 satisfying ∇g ≤ 1/32 − 3Pk+1
repeatedly applied for all k ∈ N.
Proof of Lemma 5.6. Fix g : B(c, s) → R of class C1 satisfying ∇g ≤ 1/32 −
3Pk+1
i=1 εi. We show the measure of holes in Fk+1 that the graph of g meets is
controlled by the area difference between g and a smoothing eg of g. Since the
smoothing is only done on small scales the measure of holes in F1 ∪ . . . ∪ Fk that
g meets will be controlled by the area ofeg.
Claim 5.7. There exists eg of class C1 with g −eg ≤ εk+1rk and ∇eg ≤ 1/32 −
3Pk
B ≤ CZB(c,s)
i=1 εi such that
XB∈Fk+1
G(g)∩Kk+1B6=∅
(∇g2 − ∇eg2) + Cεk+1.
Proof of Claim 5.7. Recall from Claim 5.3,
XB∈F u
k+1 of F d
other we can choose a subfamily F s
are disjoint and for each ball B1 ∈ F d
k+1 there exists B ∈ F s
Since balls in {B′ : B ∈ Fk+1} are either disjoint or one is contained inside the
k+1}
1 ⊂ B′.
k+1 such that balls in {B′ : B ∈ F s
k+1 such that B′
B ≤ εk+1.
k+1
Notice, by definition of F d
k+1, for each ball B ∈ F s
k+1
Ln(B′ ∩ {g − amk+1 > ε3
k+1t′/4}) = Ln(B′ ∩ Rk+1(B))
≤ B/εk+1
≤ ωn(ε2
k+1t′)n
which, since ∇(g − amk+1) ≤ 1/16, implies g − amk+1 ≤ ε2
k+1t′/2 in B′.
14
GARETH SPEIGHT
Hence by Claim 5.4 and Proposition 4.3,
XB1∈F d
k+1
1⊂B′
B′
B1 ≤ C XB1∈F d
k+1
1⊂B′
B′
ZRk+1(B1) ∇(g − amk+1 )2
≤ CZB′ ∇(g − amk+1)2
≤ CZB′
Defineeg to be equal to g outside
B′,
[B∈F s
k+1
(∇g2 − ∇amk+12) + Cεk+1Ln(B′).
k+1t′ ≤ εk+1rk.
k+1t′ in B(x′, t′ − εk+1t′).
i=1 εi in B(x′, t′−
i=1 εi.
) ≤ 1/32−3Pk+1
g − amk+1 ≤ ε2
k+1t′ in B(x′, t′ − εk+1t′),
for B = B(x, t) ∈ F s
smoothly as in Proposition 4.7 between g and gεk+1t′
2εk+1t′).
k+1 leteg be equal to gεk+1t′
in B(x′, t′−2εk+1t′) and interpolate
in B(x′, t′− εk+1t′)\ B(x′, t′−
k+1t′/2 in B′ and smoothing leaves an affine plane un-
− amk+1 ≤ ε2
k+1t′/2 in B(x′, t′ − εk+1t′). Hence g −
i=1 εi implies ∇(gεk+1t′
Since g − amk+1 ≤ ε2
changed it follows gεk+1t′
gεk+1t′
≤ ε2
Also ∇g ≤ 1/32−3Pk+1
εk+1t′). Proposition 4.7 hence implieseg is of class C1 with ∇eg ≤ 1/32− 3Pk
Further, g −eg ≤ ε2
By Proposition 4.3 and Proposition 4.6, since eg − amk+1 ≤ εk+1t′ in B′ and
(cid:12)(cid:12)(cid:12)(cid:12)ZB′ ∇eg2 −ZB′ ∇amk+12(cid:12)(cid:12)(cid:12)(cid:12) ≤ZB′ ∇(eg − amk+1)2 + Cεk+1Ln(B′)
≤ZB(x′,t′−2εk+1t′)(cid:12)(cid:12)(cid:12)∇(cid:16)(g − amk+1 )εk+1t′(cid:17)(cid:12)(cid:12)(cid:12)
k+1 XB1∈F d
k+1(cid:18)ZB′
k+1(cid:18)ZB′
(∇g2 − ∇eg2) + Cεk+1
B + XB∈F u
(∇g2 − ∇amk+12) + Cεk+1Ln(B′)(cid:19) + εk+1
(∇g2 − ∇eg2) + 2Cεk+1Ln(B′)(cid:19) + εk+1
Putting these estimates together and using the definition of F s
XB∈Fk+1
B ≤ XB∈F s
≤ C XB∈F s
≤ C XB∈F s
≤ CZB(c,s)
≤ Cεk+1Ln(B′).
k+1 gives
B
k+1
as required.
Kk+1B∩G(g)6=∅
k+1
1⊂B′
B′
(cid:3)
2
+ 2Cεk+1Ln(B′)
Recall balls in F1 ∪ . . . ∪ Fk have radius at least rk. Since
g −eg ≤ εk+1rk ≤ (Kk − Kk+1)rk
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
15
it follows G(g) ∩ Kk+1B 6= ∅ implies G(eg) ∩ KkB 6= ∅ for B ∈ F1 ∪ . . . ∪ Fk.
Applying the induction assumption toeg and Claim 5.7 gives,
XB∈F1∪...∪Fk+1
εi + CZB(c,s)
B + XB∈Fk+1
G(g)∩Kk+1B6=∅
G(g)∩Kk+1B6=∅
B
G(eg)∩KkB6=∅
B ≤ XB∈F1∪...∪Fk
≤ CZB(c,s) ∇eg2 + C
≤ CZB(c,s) ∇g2 + C
(∇g2 − ∇eg2) + Cεk+1
kXi=1
k+1Xi=1
εi.
It is clear from above that the constants remain bounded as Lemma 5.6 is re-
(cid:3)
peatedly applied.
together prove the proposition.
i=1 εi ≤ 1/64 and Kk > 1 for all k ≥ 1 Lemma 5.5 and Lemma 5.6
(cid:3)
Since 3P∞
Proposition 5.1 finally proves Theorem 1.4 in the case X = Rn+1. To see this
notice the proposition implies that for 0 < r ≤ 1/64
Hn(f ([0, 1]n) ∩ (B(c, s) × R) ∩ H) ≤ XB∈F
≤ Csnr2 + C
f ([0,1]n)∩B6=∅
Hn(f ([0, 1]n) ∩ B)
∞Xi=1
εi
for all f ∈ BC 1(p, δ(r)). Provided εi and r are chosen sufficiently small this implies
Hn(f ([0, 1]n) ∩ (B(c, s) × R) ∩ H) ≤ ωnsn/8
for all f ∈ BC 1 (p, δ(r)) which is (2.3) with α = ωnsn/2.
6. Decomposition of Banach spaces
We now establish the general case of Theorem 1.4 and prove Theorem 1.5. We
have shown there exists β > 0, 0 < R < 1 and a directionally porous set P ⊂ Rn+1
such that
{f ∈ BC 1(p, (n + 1)R) : Hn(f ([0, 1]n) ∩ P ) < (n + 1)nβ}
is meager, where p : [0, 1]n → Rn+1 is the plane p(x) = (x, 0).
Let X be a (possibly infinite dimensional) Banach space satisfying dim X > n.
The main idea used to prove Theorem 1.4 and Theorem 1.5 will be that a C1 surface
of dimension n can be locally approximated by n dimensional affine planes. We use
linear maps sending n dimensional affine planes to the plane p in Rn+1 to pull back
our directionally porous set in Rn+1 to directionally porous sets in X then rescale.
Temporarily fix ε > 0, y ∈ [0, 1 − ε]n and w ∈ X. Let v1, . . . , vn+1 be linearly
independent vectors in X and L : Span(v1, . . . , vn+1) → Rn+1 be the corresponding
bijective linear map sending v1, . . . , vn+1 to the standard basis e1, . . . , en+1 of Rn+1.
Using the Hahn-Banach theorem we can extend L to a linear map eL defined on
X with keLk ≤ (n + 1)kLk.
Lemma 6.1. Let eL be as defined above. Then the preimage eL−1(P ) of a direction-
ally porous set P ⊂ Rn+1 is a directionally porous set in X.
16
GARETH SPEIGHT
Proof. Suppose P is directionally porous with porosity constant ρ. Let x ∈ eL−1(P ).
Then eL(x) ∈ P so there exists v ∈ Rn+1 with kvk = 1 such that for any δ > 0 there
exists t ∈ R with t < δ such that
The previous line implies
B(eL(x) + tv, ρt) ∩ P = ∅.
(cid:3)
Hence eL−1(P ) is directionally porous.
Consider the map T : Γn(X) → Γn(Rn+1) defined by
(cid:19)
B(x + tL−1(v), ρt/keLk) ∩eL−1(P ) = ∅.
T (f )(x) =eL(cid:18) f (y + εx) − w
ε
for f ∈ Γn(X) and x ∈ [0, 1]n.
The map f 7→ T (f ) +eL(w)/ε is a continuous linear surjection and hence, by the
open mapping theorem, is open. It follows that T is continuous and open. Hence if
a set A ⊂ Γn(Rn+1) is meager then the preimage T −1(A) ⊂ Γn(X) is also meager.
The following lemma follows easily from the definition of T and basic facts about
Hausdorff measures.
Lemma 6.2. Let q : [0, 1]n → X be the affine plane
q(x) = w + (x1 − y1)v1 + . . . + (xn − yn)vn.
Define M (ε, y, w, v1, . . . , vn+1) to be the set of surfaces f ∈ Γn(X) such that
kf (x) − q(x)k < εR/kLk for x ∈ y + [0, ε]n,
k∂f /∂xi(x) − vik < R/kLk for x ∈ y + [0, ε]n and i = 1, . . . , n,
and
Hn(f (y + [0, ε]n) ∩ (w + εeL−1(P ))) < βεn/kLkn.
Then
M (ε, y, w, v1, . . . , vn+1) ⊂ T −1{f ∈ BC 1(p, (n+1)R) : Hn(f ([0, 1]n)∩P ) < (n+1)nβ}
and hence M (ε, y, w, v1, . . . , vn+1) is meager in Γn(X).
It follows immediately from the lemma that for any choice of v1, . . . , vn+1 linearly
directionally porous this establishes the general case of Theorem 1.4.
independent in X the set Sn(eL−1(P ), 0) is not meager in Γn(X). Since eL−1(P ) is
Now we suppose X is separable with countable dense subset F ⊂ X. Let D
be the set of all bijective linear maps corresponding to all linearly independent
(n + 1)-tuples v1, . . . , vn+1 ∈ F in the way defined earlier. Let
Q =[(w + εeL−1(P ))
where the union is taken over rational ε > 0, w ∈ F and L ∈ D. Each of the sets
show X \ Q is Γn-null.
w + εeL−1(P ) is directionally porous and hence Q is σ-directionally porous. We
It is stated in Lemma 5.3.5 [4] that if
M ′ = {f ∈ Γn(X) : rank df < n on a set of positive measure}
then M ′ is meager.
SURFACES MEETING POROUS SETS IN POSITIVE MEASURE
17
Lemma 6.3. Define
M = M ′ ∪[ M (ε, y, w, v1, . . . , vn+1)
where the union is taken over rational ε > 0, y ∈ [0, 1 − ε]n ∩ Qn, w ∈ F and all
(n + 1)-tuples v1, . . . , vn+1 of linearly independent vectors in F .
Suppose the image of a surface f ∈ Γn(X) meets X \ Q in positive measure.
Then f belongs to the meager set M .
Proof. Since f is Lipschitz it follows Ln([0, 1]n \ f −1(Q)) > 0. If f does not belong
to M ′ ⊂ M then there exists z ∈ (0, 1)n such that rank df (z) = n and z is a
Lebesgue density point of [0, 1]n \ f −1(Q).
If Lw1,...,wn+1 : Span(w1, . . . , wn+1) → Rn+1 denotes the linear map sending
linearly independent vectors w1, . . . , wn+1 ∈ X to e1, . . . , en+1 ∈ Rn+1 it is not
hard to show the map defined by
is continuous on the open set on which it is defined.
(w1, . . . , wn+1) 7→ kLw1,...,wn+1k
Since rank df (z) = n we can fix vn+1 ∈ F such that
∂f /∂x1(z), . . . , ∂f /∂xn(z), vn+1
are linearly independent. Hence there exists η > 0 such that if
kvi − ∂f /∂xi(z)k < η
for i = 1, . . . , n then η ≤ kLv1,...,vn+1k ≤ 1/η.
Fix rational ε > 0 to be small enough so that
kf (x) − f (z) − ∂f /∂x1(z)(x1 − z1) − . . . − ∂f /∂xn(z)(xn − zn)k < εηR/4
and
k∂f /∂xi(x) − ∂f /∂xi(z)k < ηR/2
for x ∈ z + [0, 2ε]n ⊂ [0, 1]n and i = 1, . . . , n and, using the fact that z is a density
point of [0, 1]n \ f −1(Q),
Hn(f (z + [0, 2ε]n) ∩ Q) < βηnεn.
Now fix y ∈ [0, 1 − ε]n ∩ Qn sufficiently close to z so that
kf (x) − f (y) − ∂f /∂x1(z)(x1 − y1) − . . . − ∂f /∂xn(z)(xn − yn)k < εηR/2
for x ∈ y + [0, ε]n ⊂ z + [0, 2ε]n.
There exists w ∈ F and linearly independent v1, . . . , vn ∈ F with
kvi − ∂f /∂xi(z)k < η
for i = 1, 2, . . . , n such that if q : [0, 1]n → X is the affine plane
q(x) = w + (x1 − y1)v1 + . . . + (xn − yn)vn
then
and
kf (x) − q(x)k < εηR
k∂f /∂xi(x) − vik < ηR
We note η ≤ kLv1,...,vn+1k ≤ 1/η and observe
for x ∈ y + [0, ε]n and i = 1, . . . , n.
Hn(f (y + [0, ε]n) ∩ (w + εeL−1
v1,...,vn+1(P ))) ≤ Hn(f (z + [0, 2ε]n) ∩ Q)
< βηnεn
≤ βεn/kLv1,...,vn+1kn.
Hence f ∈ M (ε, y, w, v1, . . . , vn+1) ⊂ M as required.
(cid:3)
18
GARETH SPEIGHT
Lemma 6.3 shows X \ Q is Γn-null. Since Q is σ-directionally porous this con-
cludes the proof of Theorem 1.5.
References
[1] L. Zaj´ıcek, Porosity and σ-porosity, Real Anal. Exchange, Volume 13 (1987/1988), Number 2,
314-350.
[2] L. Zaj´ıcek, On σ-porous sets in abstract spaces, Abstract and Applied Analysis, Volume 2005,
Number 5, 509-534.
[3] Y. Benyamini, Geometric Nonlinear Functional Analysis, Volume 1, American Mathematical
Society, Colloqium Publications, Volume 48, 1999.
[4] J. Lindenstrauss, D. Preiss and J. Tiser, Fr´echet Differentiability of Lipschitz Functions and
Porous Sets in Banach Spaces, Annals of Mathematics Studies, 179, Princeton University
Press, Princeton, NJ, 2012.
[5] J. Lindenstrauss and D. Preiss, On Fr´echet Differentiability of Lipschitz maps between Banach
Spaces, Annals of Mathematics, Volume 157 (2003), 257-288.
[6] N. Aronszajn, Differentiability of Lipschitzian mappings between Banach spaces, Studia Math,
Volume 57 (1976), 147-190.
[7] J. P. R. Christensen, Measure theoretic zero sets in infinite dimensional spaces and applica-
tions to differentiability of Lipschitz mappings II, Publ Dep Math Lyon, Volume 10 (1973),
29-39.
[8] P. Mankiewicz, On the differentiability of Lipschitz mappings in Fr´echet spaces, Studia Math,
Volume 45 (1973), 15-29.
[9] D. Preiss and J. Tiser, Two unexpected examples concerning differentiability of Lipschitz func-
tions on Banach spaces, Operator Theory: Advances and Applications, Volume 77 (1995),
219-238.
[10] D. Preiss, Fr´echet Differentiability of Lipschitz functions, Journal of Functional Analysis,
Volume 91 (1990), 312-345.
[11] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill Science/Engineering/Math,
Third Edition, 1976.
[12] L. Evans and R. Gariepy, Measure Theory and Fine Properties of Functions, Studies in
Advanced Mathematics CRC Press, 1991.
E-mail address: [email protected]
|
1806.02994 | 1 | 1806 | 2018-06-08T07:20:13 | Existence of Large Independent-Like Sets | [
"math.FA"
] | Let $G$ be a compact abelian group and $\Gamma$ be its discrete dual group. For $N \in \mathbb{N}$, we define a class of independent-like sets, $N$-PR sets, as a set in $\Gamma$ such that every $\mathbb{Z}_N$-valued function defined on the set can be interpolated by a character in $G$.
These sets are examples of $\varepsilon$-Kronecker sets and Sidon sets. In this paper we study various properties of $N$-PR sets. We give a characterization of $N$-PR sets, describe their structures and prove the existence of large $N$-PR sets. | math.FA | math | EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
ROBERT (XU) YANG
Abstract. Let G be a compact abelian group and Γ be its discrete
dual group. For N ∈ N, we define a class of independent-like sets, N -PR
sets, as a set in Γ such that every ZN -valued function defined on the set
can be interpolated by a character in G.
These sets are examples of ε-Kronecker sets and Sidon sets. In this
paper we study various properties of N -PR sets. We give a characteri-
zation of N -PR sets, describe their structures and prove the existence
of large N -PR sets.
8
1
0
2
n
u
J
8
]
.
A
F
h
t
a
m
[
1
v
4
9
9
2
0
.
6
0
8
1
:
v
i
X
r
a
1. Introduction
Independence is a property prevalent throughout mathematics. In har-
monic analysis, independence has been used to produce curious examples.
For example, a perfect independent set in R was first constructed by Von
Neumann [10]. Independent Cantor sets in non-discrete locally compact
abelian groups were constructed by Hewitt and Kakutani [8], in part for
showing that M(G) is asymmetric [7]. In these examples, by 'independent'
we mean algebraically independent: a set of (non-trivial) characters E is
i = 1, then
independent if whenever γ1, ..., γN ∈ E, mi ∈ Z, and QN
γmi
i = 1 for all i.
i=1 γmi
The Rademacher functions in the dual of the infinite direct product of
infinitely many copies of Z2 is a set of characters which is both algebraically
and probabilistically independent. These functions have proven to be very
useful in harmonic analysis. In particular, they have the property that every
±1-valued function defined on the set is evaluation at some x in the group.
A similar interpolation property holds for all algebraically independent sets.
A weaker notion than independence is an ε-Kronecker set.
Definition. Let ε > 0. The set E is said to be (weak) ε-Kronecker if for
every ϕ : E → T there exists x ∈ G such that ϕ(γ) − γ(x) < ε for all
γ ∈ E (resp. ϕ(γ) − γ(x) ≤ ε for all γ ∈ E).
2010 Mathematics Subject Classification. Primary 43A25; Secondary 43A46.
Key words and phrases. Sidon Sets, Independent Sets, Kronecker Sets.
Research was partially supported by NSERC grant 2016-03719.
1
2
ROBERT (XU) YANG
This notion was inspired in part by the classical approximation theorem
of Kronecker, with early work done by by Hewitt and Kakutani [8] and
Rudin [13]. The terminology was introduced by Varapolous in [14].
The set of Rademacher functions is clearly an example of a weak √2-
Kronecker set. Kronecker-like sets have been studied intensively, and are
known to have many interesting properties. For example, Hare and Ramsey
[6] proved that every ε-Kronecker set with ε < 2 is a Sidon set. Graham and
Hare in [3] introduced the weaker notion of pseudo-Rademacher sets, sets
of characters where every ±1-valued function is point-wise evaluation, in
order to study the problem of the existence of Kronecker-like sets. Galindo
and Hernandez in [1] and Graham and Lau in [5] both consider interpolation
sets of characters of finite order. For other references and further background
information we refer the reader to [4].
In this paper, we generalize this notion to N-pseudo-Rademacher sets
(or N-PR sets for short), sets of characters with the property that every
ZN -valued function on the set is point-evaluation. Of course, a pseudo-
Rademacher set is a 2-PR set and we will see that N-PR sets are ε-Kronecker
sets for suitable ε = ε(N), which tends to 0 as N → ∞. We give an
algebraic characterization of N-PR sets, compare them with ε-Kronecker
sets, describe their structures and prove existence theorems (Theorem 4.2)
of large N-PR sets. Theorem 4.2 gives a new proof that any uncountable
subset in Γ contains a large ε-Kronecker set.
2. Characterization of N -PR sets
Throughout this paper G is a compact abelian group and Γ is its discrete
dual group.
Definition. Let E ⊂ Γ be a subset and N ∈ N. We define E to be an
"N-pseudo-Rademacher" set (or N-PR set) if for every ϕ : E → ZN ⊂ T,
there exists x ∈ G such that ϕ(γ) = γ(x) for all γ ∈ E.
In this section we give an algebraic characterization of N-PR sets. We
first establish some useful lemmas.
Lemma 2.1 (Quotient and Subgroup Lemma). Let E ⊂ Γ, n ∈ N, γ ∈ Γ
and Λ ⊂ Γ be a subgroup.
(1) Let q : Γ → Γ/Λ be the quotient map. If q is one-to-one on E and
q(E) is n-PR, then E is n-PR.
(2) Suppose E ⊂ Λ. Then E is n-PR as a subset of Γ if and only if E is
n-PR as a subset of Λ.
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
3
Proof. (1) Suppose q : Γ → Γ/Λ is one-to-one on E and q(E) is n-PR. We
will show that E is n-PR. Let ϕ : E → Zn be a function. Because q is
one-to-one on E, for each γ, β ∈ E, if β 6= γ, then γ − β /∈ Λ. Thus, we can
define ϕ′ : q(E) → Zn via ϕ′(γ + Λ) = ϕ(γ) for γ ∈ E. Since q(E) is n-PR,
there exists x ∈ Λ⊥ = dΓ/Λ such that ϕ′(γ + Λ) = x(γ + Λ) for all γ ∈ E.
As x ∈ Λ⊥, ϕ(γ) = γ(x) for all γ ∈ E. This means E is n-PR.
(2) We first suppose E is an n-PR subset of Γ. Let ϕ : E → Zn be a
function. There exists x ∈ G such that ϕ(γ) = γ(x) for all γ ∈ E. Let
x + Λ⊥ ∈ G/Λ⊥ =bΛ. Since E ⊂ Λ, for all γ ∈ E we have
ϕ(γ) = γ(x) = (x + Λ⊥)(γ).
This means E is n-PR as a subset of Λ. The proof of the converse part of
(2) is similar.
(cid:3)
The following proposition is a stronger version of Lemma 3.2 in [9].
Proposition 2.2. Let E ⊂ Γ. The following are equivalent:
(1) For all ϕ : E → T with ϕ(γ) ∈ Range(γ) there exists x ∈ G such
that ϕ(γ) = γ(x) for γ ∈ E.
(2) E is independent.
Proof. Assume E is independent and that ϕ : E → T satisfies ϕ(γ) ∈
Range(γ) for γ ∈ E. By similar arguments to the proof of Lemma 2.1, if
we can find x ∈dhEi such that ϕ(γ) = γ(x) for all γ ∈ E, then there exists
x′ ∈ G such that ϕ(γ) = γ(x′). Thus we may assume Γ = hEi =Lγ∈Ehγi.
For each γ ∈ E, there exists xγ ∈ chγi such that γ(xγ) = ϕ(γ). If E is finite,
we let x =Qγ∈E xγ and x can interpolate ϕ exactly. For the case that E is
infinite, since G is compact, we let x ∈ G be a cluster point of the following
set
(Yγ∈F
xγ : F ⊂ E,F < ∞) ,
and such an x can interpolate ϕ exactly.
1
i
...γmk
k = 1 but γmi
Conversely, if E is not independent, then there exist γ1, ..., γk ∈ E and
m1, ..., mk ∈ Z such that γm1
6= 1 for all 1 ≤ i ≤ k.
Consider the function ϕ : E → T such that ϕ(γi) = 1 for all i > 1,
ϕ(γ1)m1
6= 1 and ϕ(γ) ∈ Range(γ) for all γ ∈ E. Notice that such ϕ
exists, because γm1
1 ) 6= {1} and hence there exists
x ∈ G such that γ1(x)m1 6= 1. Let ϕ(γ1) = γ1(x). This function ϕ cannot be
interpolated by any x ∈ G.
6= 1 means Range(γm1
(cid:3)
1
4
ROBERT (XU) YANG
Corollary 2.3. Let N ∈ N and E ⊂ Γ. If E is independent and ZN ⊂
Range(γ) for all γ ∈ E, then E is N-PR.
Proof. This follows directly from Proposition 2.2.
(cid:3)
Theorem 2.4. The following are equivalent:
(1) E ⊂ Γ is N-PR.
i=1 γmi
i=1 γmi
i = 1 for some
i=1 γi(x)mi, whileQn
cannot be interpolated by any x ∈ G and therefore (1) fails.
Conversely, suppose (2) holds. Let EN = hγN : γ ∈ Ei, the subgroup
mi ∈ Z, then N divides mi for all i.
Proof. Suppose (2) fails. Then there exist distinct γi ∈ E and integers mi ∈
i = 1, while N does not divide m1. Let f : E → ZN be
given by f (γ1) = e2πi/N and f (γ) = 1 for all γ 6= γ1. For any x ∈ G,
i=1 f (γi)mi = f (γ1)m1 6= 1. Thus, this function f
(2) If γi ∈ E for 1 ≤ i ≤ n are distinct and Qn
Z with Qn
1 =Qn
generated by(cid:8)γN : γ ∈ E(cid:9), and π : Γ → Γ/EN be the quotient map. Ele-
distinct π(γi) and mi ∈ Z such thatQn
kj ∈ Z, 1 ≤ j ≤ s and s ≥ n such thatQn
γj ∈ E. Hence,Qn
ments in EN have the form γk1N
for γ1, ...γm ∈ E and k1, ..., km ∈ Z.
We claim that π(E) ⊂ π(Γ) is independent. Indeed, suppose γi ∈ E with
i ∈ ker(π). Then there exists
with distinct
= 1 ∈ Γ. By (2), N mi − kiN
for all 1 ≤ i ≤ n and hence N mi for all 1 ≤ i ≤ n. This means each
γmi
i ∈ ker(π) and therefore the set π(E) is independent.
Moreover, we claim that π is one-to-one on E. Suppose, otherwise, that
2 ∈ ker(π). Then there exists kj ∈ Z,
with distinct γj ∈ E.
= 1. Again, (2) gives N 1 − k1N and
there are γ1 6= γ2 ∈ E and γ1γ−1
1 ≤ j ≤ s and s ≥ 2 such that γ1γ−1
j=3 γ−kjN
We have γ1−k1N
N − 1 − k2N, which are not possible.
Qs
γ−1−k2N
2
2 = Qs
i =Qs
Similar arguments show that elements of π(E) have order N. Thus
π(E) is N-PR from Proposition 2.2. Because π is also one-to-one on E,
Lemma 2.1 gives that E is N-PR, proving (1).
(cid:3)
j=n+1 γ−kjN
j
i=1 γmi−kiN
i
i=1 γmi
i=1 γmi
1
j
...γkmN
m
1
Qs
j=1 γkj N
j
j=1 γkjN
j
Corollary 2.5. Let a, b ∈ N be co-prime. The subset E ⊂ Γ is (ab)-PR if
and only if E is both a-PR and b-PR.
Proof. Since Za, Zb ⊂ Zab, if E is (ab)-PR, E is both a-PR and b-PR. To see
the converse, we assume E is both a-PR and b-PR. Consider γ1, ..., γn ∈ E
i = 1. Since E is both a-PR and b-PR,
ami and bmi for all 1 ≤ i ≤ n. Since a and b are co-prime, abmi for all
1 ≤ i ≤ n. Hence, by Theorem 2.4, E is (ab)-PR.
and m1, ..., mn ∈ Z such thatQn
i=1 γmi
(cid:3)
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
5
Corollary 2.6. Suppose E ⊂ Γ is N-PR and hEi ∩hγi = {1}. Then γE is
also N-PR.
Proof. Let γ1, ...γn ∈ E and m1, ..., mi ∈ Z be such thatQn
Since hEi ∩ hγi = {1}, we have Qn
1 ≤ i ≤ n. This shows γE is N-PR by Theorem 2.4.
i=1(γγi)mi = 1.
i = 1 and therefore Nmi for all
i=1 γmi
(cid:3)
Next we compare N-PR sets with ε-Kronecker sets. We first note that
N-PR sets are weak ε-Kronecker sets for appropriate ε depending on N.
Proposition 2.7. N-PR sets are weak ε-Kronecker for ε = 1 − e
Proof. Assume E ⊂ Γ is N-PR. Let ϕ : E → T. Define ϕN : E → ZN by
ϕN (γ) = t, where t ∈ ZN satisfies t − ϕ(γ) ≤ 1 − e
N . Let x ∈ G be such
that γ(x) = ϕN (γ) for γ ∈ E. We have γ(x) − ϕ(γ) ≤ 1 − e
N .
N .
πi
πi
πi
(cid:3)
Of course, not every ε-Kronecker set in a torsion group is N-PR. Here
is an example.
Example. For every ε > 0 and N ∈ N, there exists an infinite ε-Kronecker
set E with every γ ∈ E having finite order a multiple of N, but E is not
N-PR.
Let N ∈ N and ε > 0. Let (pi)∞
coprime to N. Let Γ = ZN ⊕Li≥1 Zpi. Let
S := {(1, 0, 0, ...), (1, 1, 0, 0, ...), (1, 1, 1, 0, 0, ...), ...} .
i=1 be an increasing sequence of primes
Each element in S has order a multiple of N. It is not hard to see that
if we exclude finitely many elements of small orders, we have a co-finite
ε-Kronecker set E ⊂ S. From Theorem 2.4 no subsets of S other than
singletons are N-PR and therefore E is not N-PR.
Remark. Recall that the Bohr compactification of Γ, denoted by Γ, is defined
as the dual group of Gd, where Gd is the group G equipped with the discrete
topology.
Let E be the closure of E in Γ. In [4], Ex. 1.4.3, it is shown that there
is an ε-Kronecker set E and an integer M ≥ 1 such that (E ∪ E−1)M has
Haar measure 1. That is not the case for N-PR sets. Indeed, if E is any
N-PR set with N ≥ 3, then the Haar measure of (E ∪ E−1)M is 0 for all
M ≥ 1. The proof is very similar to the proof of Theorem 3.1 in [2] and
therefore is omitted.
6
ROBERT (XU) YANG
3. Structure of N -PR sets
In this section, we investigate the structure of N-PR sets. We rely heavily
on the following structure theorem for general abelian groups.
Notation: Let p be a prime number. The group C(p∞) is the group of all
pn-th roots of unity for n ≥ 1.
Theorem 3.1. [11] Every abelian group Γ is isomorphic to a subgroup of
Mα
Qα ⊕Mβ
C(p∞
β ),
where Qα are copies of Q and pβ are prime numbers.
Notation: (1) We let Γ0 be the torsion subgroup of Γ and π0 : Γ →
Γ/Γ0 be the quotient map. In the notation of Theorem 3.1, π0 :Lα Qα ⊕
Lβ C(p∞
β ) →Lα Qα is the quotient map.
(2) Let p ∈ N be a prime number and n ∈ N. We let Γpn be the subgroup
of Γ0 containing elements whose orders are not a multiple of pn, or equiva-
lently, whose orders are not divisible by pn. Notice Γpn is indeed a subgroup
because if γ1 and γ2 have orders m1 and m2, neither a multiple of pn, then
lcm(m1, m2), the order of γ1γ2, is not a multiple of pn. Let πpn : Γ → Γ/Γpn
be the quotient map.
(3) For γ ∈Lα∈A Qα ⊕Lβ∈B C(p∞
the i-th coordinate of γ. In particular, if Γ =Li C(p∞
map πpi can be viewed as πpi = Proji.
β ) and i ∈ A ∪ B, we let Proji(γ) be
i ) with pi distinct, the
Proposition 3.2. Let E ⊂ Γ, n ∈ N and p be a prime number. The follow-
ing are equivalent:
(1) E is pn-PR.
(2) πpk(E) is pn+1−k-PR and πpk is one-to-one on E for all 1 ≤ k ≤ n.
(3) πpk(E) is pn+1−k-PR and πpk is one-to-one on E for some 1 ≤ k ≤ n.
Proof. We first show (1) implies (2). Fix 1 ≤ k ≤ n and we suppose E ⊂
Γ is pn-PR. We first claim that πpk is one-to-one on E. This is because
if γ1, γ2 ∈ E have γ1γ−1
has finite order that is not
divisible by pk. But from Theorem 2.4, this implies E is not even pk-PR
and therefore contradicts that E is pn-PR. Next, we show that πpk(E) is
pn+1−k-PR. We let γ1, ..., γs ∈ E and βi := πpk(γi) for 1 ≤ i ≤ s. Suppose
i ∈ Γpk
and therefore there exists l ∈ N, which is not divisible by pk, such that
i = 1. By Theorem 2.4, pnlmi and hence pn+1−kmi. Theorem 2.4
for some m1, ..., ms ∈ Z we haveQs
Qs
i = 1. This meansQs
2 ∈ Γpk, then γ1γ−1
implies πpk(E) is pn+1−k-PR.
i=1 γlmi
i=1 βmi
i=1 γmi
2
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
7
i=1 γmi
i = 1. Then Qs
Since (2) implies (3) is obvious, it remains to show (3) implies (1). We
assume (3) holds for some 1 ≤ k ≤ n. Let γ1, ..., γs ∈ E be distinct and
i=1 πpk(γi)mi = 1 and the
injectivity of πpk implies πpk (γi) are distinct. Since πpk(E) is pn+1−k-PR, we
have pn+1−kmi for all 1 ≤ i ≤ s.
i=1 γmi/pk−1
m1, ..., ms ∈ Z such that Qs
If n + 1− k ≥ k, we note that order of the productQs
∈ Γpk. This means Qk
pk−1 and therefore Qk
i=1 γmi/pk−1
divides
i=1 πpk(γi)mi/pk−1 =
1 ∈ Γ/Γpk. Hence, by Theorem 2.4, we have pn+1−k divides mi/pk−1 and
this gives pn divides mi for all 1 ≤ i ≤ s. Theorem 2.4 gives (1).
Otherwise, we have n + 1 − k < k. Notice that the order of the product
i=1 γmi/pn+1−k
∈ Γpk.
As above, we have pn+1−k divides mi/pn+1−k, which means p2(n+1−k)mi. We
continue doing this until we reach r(n + 1 − k) ≥ k for some r ∈ N. The
previous case gives (1).
divides pn+1−k. Since n + 1 − k < k,Qk
i=1 γmi/pn+1−k
Qs
(cid:3)
i
i
i
i
Corollary 3.3. Suppose Γ = LC(p∞). Then pn-PR subsets in LC(p∞)
are in correspondence with p-PR subsets inLC(p∞) in the following man-
ner: Given a set E ⊂LC(p∞), we let Epn :=nγpn−1 : γ ∈ Eo and E1/pn :=
{ξγ : γ ∈ E}, where ξpn−1
E. If E is p-PR then E1/pn is pn-PR and E = E1/pn.
Proof. Since onL C(p∞), the map πpn can be identified as πpn(γ) = γpn−1,
= γ. If E is pn-PR then Epn is p-PR and Epn =
this follows from Proposition 3.2.
(cid:3)
γ
1 ...pnk
Corollary 3.4. Let N = pn1
E is N-PR if and only if πp
all 1 ≤ i ≤ k.
Proof. This follows from Corollary 2.5 and Proposition 3.2.
(E) is pi-PR and πp
ni
i
ni
i
for distinct primes pi and ni ∈ N. Then
is one-to-one on E for
(cid:3)
k
We thus have the following result about the structure of N-PR sets in
the torsion subgroup.
Proposition 3.5. Let E ⊂ Γ0 be an N-PR set and N = pn1
k where the
pi are distinct prime numbers. There exist pni
i ) for
1 ≤ i ≤ k, and bijections f2 : E1 → E2, ..., fk : E1 → Ek such that we can
represent E as
i -PR sets Ei ⊂L C(p∞
1 ...pnk
E = {γ ⊕ f2(γ) ⊕ ... ⊕ fk(γ) ⊕ βγ : γ ∈ E1}
for some βγ ∈Lp6=pi ∀1≤i≤k C(p∞).
Proof. For 1 ≤ i ≤ k, we let Ei := πpi(E). Since E is N-PR, from Proposition 3.2,
each Ei is pni
i -PR and the maps πpi : E → Ei are injective. For 2 ≤ i ≤ k,
8
ROBERT (XU) YANG
we define fi on E1 as fi(πp1(γ)) := πpi(γ) for γ ∈ E. The injectivity of
πp1 on E ensures the maps are well-defined. Moreover, the injectivity of πpi
implies fi is injective. Each fi is clearly surjective by its construction and
therefore a bijection. Since each γ ∈ E can be represented as
γ = ⊕k
i=1πpi(γ) ⊕ βγ
for some βγ ∈Lp6=pi ∀1≤i≤k C(p∞), Proposition 3.5 follows.
Next, we will discuss the relation between p-PR sets and independent
(cid:3)
sets.
We first note that not every p-PR set is an independent set. The following
example shows a p-PR set whose only independent subsets are singletons.
Example. Consider E = {γn : n ≥ 2} ⊆ Li≥1 C(p∞), where Proj1(γn) =
1/p2, Projn(γn) = 1/p and Projk(γn) = 0 for all k 6= 1, n. Here we use
additive group operation. E is p-PR by Theorem 2.4, but E does not contain
any independent subsets other than singletons because for all i 6= j, pγi =
pγj 6= 1.
Similar to independent sets, we have the following result about the max-
imum size of a p-PR set inside a product group.
Q ⊕Li∈B2 C(p∞) is p-PR. Then
E ≤ B1 + B2.
Proof. We identify each element in C(p∞) in the form of a/pn for some
Q ⊕
Proposition 3.6. Suppose E ⊆ Li∈B1
a, n ∈ N with additive group operation. Hence, we can identifyLi∈B1
Li∈B2 C(p∞) as a subset of the real vector space RB1+B2.
We claim that if E ⊆ Li∈B1
Q ⊕Li∈B2 C(p∞) is p-PR, E is linearly
independent in RB1+B2. Indeed, suppose that E is not linearly indepen-
dent. Then there exist {γi 1 ≤ i ≤ k} ⊆ E and a1, ..., ak ∈ R such that
a1γ1 + ... + akγk = 0, while not all ai's are zero and each γi ∈ RB1+B2 is
identified as above. Since the entries of each γi are in Q, we may assume
ai ∈ Q for all 1 ≤ i ≤ k. Furthermore, we may assume ai ∈ Z. Notice that if
pai for all i, we may replace ai by ai/p. Hence, we may find a choice of such
ai's, not all divisible by p. This contradicts that E is p-PR by Theorem 2.4.
(cid:3)
Thus, E ≤ B1 + B2.
4. Existence of N -PR Sets
In this section, we show some existence results about N-PR sets and
that large N-PR sets are plentiful. We first prove a lemma.
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
9
Lemma 4.1. Assume E ⊆Lβ∈B Γβ is uncountable.
(1) Let p be a prime number. If Γβ = C(p∞) for all β ∈ B, then E
contains a p-PR subset of the same cardinality.
(2) If Γβ = Q for all β ∈ B, then E contains an independent subset of
the same cardinality.
Proof. Without loss of generality, we further assume for each β ∈ B there
exists γ ∈ E such that Projβ(γ) is non-trivial. (1) We first prove (1) in the
special case that every γ ∈ E has order p. We consider the collection C of
subsets of E defined as A ∈ C if for all finite subsets F ⊆ A there exists an
arrangement F = {γ1, ..., γn} such that for each 1 ≤ k ≤ n there is some
β ∈ B with Projβ(γk) non-trivial, but Projβ(γj) trivial for all 1 ≤ j ≤ k. We
partially order C by inclusion and Zorn's Lemma gives a maximal S ∈ C.
We claim S = E. Indeed, if S < E, we let B1 ⊂ B be given by
β ∈ B1 if there exists γ ∈ S such that Projβ(γ) is non-trivial. We thus
have B1 < B. Let β0 ∈ B\B1 and γ0 ∈ E be such that Projβ0(γ0) is
non-trivial. Since β0 ∈ B\B1, γ0 /∈ S and we form the set S1 := S ∪ {γ0}.
It is easy to see S1 ∈ C and this contradicts the maximality of S.
Moreover, the construction of S, the assumption that every element has
order p and Theorem 2.4 imply S is p-PR, which finishes the proof for the
special case.
For the general case, we let Ek be defined as the subset in E containing
elements of order pk. Then E = ∪k≥0Ek and hence there exists a positive
integer K ∈ N+ such that EK = E. If K = 1, the special case finishes
the proof, and therefore we suppose K > 1.
We have two cases. The first case is that there exists 1 ≤ n0 < K such
that πpn0 (EK) = EK, while πpn0+1(EK) < EK. We let B1 ⊂ B be given
by β ∈ B1 if there exists γ ∈ EK such that Projβ(γ) has order greater
or equal to pn0+1. Since πpn0+1(EK) < EK, we have B1 < B = EK.
For a subset C ⊂ B, we define the projection ProjC : Lβ∈B C(p∞) →
Lβ∈C C(p∞). We have ProjB\B1(πpn0 (EK)) = EK. Furthermore, each α ∈
ProjB\B1(πpn0 (EK)) has the property that for all β ∈ B\B1, Projβ(α) has
order p or 1. Thus, the special case gives a p-PR set F ⊂ ProjB\B1(πpn0 (EK))
such that F = ProjB\B1(πpn0 (EK)) = EK = E and the quotient lemma
finishes the proof for this case.
The other case is that πpK (EK) = EK. Since πpK (EK) satisfies the
special case, the quotient lemma again finishes the proof.
(2) The proof of (2) is similar to the first part of the argument of (1),
(cid:3)
but much simpler.
10
ROBERT (XU) YANG
Theorem 4.2. Let E ⊂ Γ be uncountable. Then there exists a prime number
p such that E contains a p-PR set of the same cardinality.
Q ⊕ ⊕∞
j ), where (Bj)∞
Proof. Embed E ⊂Li∈B0
sets and pj are distinct primes. We assume that for each index i ∈S∞
j=0 are index
j=0 Bj,
there exists γ ∈ E such that Proji(γ) is non-trivial. Since E is uncountable
and the groups Q and C(p∞
j ) are countable, there exists K ∈ N such that
BK = E.
j=1Li∈Bj C(p∞
If K = 0, from Lemma 4.1 (2) we may extract an independent set F ⊂
π0(E) with F = E. If we choose E′ ⊂ E such that π0 is one-to-one on
E′ and π0(E′) = F , then E′ is N-PR for all N ≥ 2 by Lemma 2.1 and
Proposition 2.2.
Similarly, if K ≥ 1, by Lemma 4.1 (1) we extract a pK-PR subset F ⊂
πpK (E) with F = E and therefore obtain a pK-PR subset in E of the
same cardinality.
(cid:3)
Remark. Since any p-PR set is a Sidon set, Theorem 4.2 implies any un-
countable subset in Γ contains a large Sidon set. I0 sets are more special
interpolation sets. A set E ⊂ Γ is I0 if every bounded function defined on E
can be interpolated as a Fourier transform of a discrete measure. If p ≥ 3,
any p-PR set is an I0 set, and therefore in that case any uncountable subset
in Γ contains a large I0 set.
Theorem 4.3. Let E ⊂ Γ be uncountable, p be a prime number and n ∈
N. Then E contains a pn-PR subset of the same cardinality if and only if
πpn(E) = E.
Proof. If E contains a pn-PR subset E1 of the same cardinality, by Proposition 3.2
πpn(E) ≥ πpn(E1) = E1 = E.
To see the converse, we define the projection
π1 :Mi∈B1
Q ⊕Mi∈B2
C(p∞) →Mi∈B2
C(p∞).
Assume πpn(E) = E. We claim that either π0(πpn(E)) = E or π1(πpn(E)) =
E.
If π0(πpn(E)) < E and π1(πpn(E)) < E, then because E is uncount-
able,
πpn(E) ≤ π0(πpn(E))π1(πpn(E)) < E,
which is a contradiction. If π0(πpn(E)) = E, we appeal to (2) in Lemma 4.1
to get a pn-PR set in π0(πpn(E)) and Lemma 2.1 finishes the proof. If
π1(πpn(E)) = E, then (1) in Lemma 4.1 similarly finishes the proof. (cid:3)
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
11
Example. If E is countable, Theorem 4.2 may fail. Let E = C(p∞). E is
countably infinite, but by Proposition 3.6, E only contains p-PR sets that
are singletons.
Proposition 4.4. Suppose E = Γ ⊂Li∈B1
Q⊕Li∈B2 C(p∞) is a countably
infinite group. Then E contains an infinite p-PR set if and only if E contains
an infinite independent set.
k
Proof. Assume that all independent sets in E are finite. Zorn's Lemma gives
a maximal independent set S ⊂ E, where the partial order is inclusion. Our
assumption gives that S is finite and the maximality implies hSi = E. Since
S is independent, we embed S intoLγ∈S Γγ, where Γγ = Q if γ has infinite
order and Γγ = C(p∞) if γ has order some power of p. Since S is maximal, we
may extend the embedding to hSi = E. From Proposition 3.6, this implies
all p-PR sets in E have cardinality at most S. The converse is trivial. (cid:3)
Proposition 4.5. Suppose Γ is an uncountable infinite group and N =
1 ...pmk
pm1
is an integer with prime numbers pi, 1 ≤ i ≤ k. Then Γ contains
(Γ) = Γ for all 1 ≤ i ≤ k.
an N-PR set E with E = Γ if and only if πp
Proof. If Γ contains an N-PR set E with E = Γ, then by Corollary 3.4
πp
(Γ) = Γ for all 1 ≤ i ≤ k.
To see the converse, first of all, if π0(Γ) = Γ, since Γ is uncountable,
the conclusion follows by Lemma 4.1 and Lemma 2.1. Thus we may assume
π0(Γ) < Γ. Hence Γ0 = Γ and we may further assume Γ is a torsion
(Γ) = Γ for all 1 ≤ i ≤ k, then by Lemma 4.1 (1) we let
group. If πp
(Γ) such that Si is pi-PR with Si = Γ, 1 ≤ i ≤ k.
Si be a subset in πp
is one-to-one on Ji and
For 1 ≤ i ≤ k we let Ji ⊂ Γ be such that πp
(Ji) = Si and moreover, we may assume the order of each γ ∈ Ji is
πp
a power of pi. Since Ji = Γ for all 1 ≤ i ≤ k, we let, for 2 ≤ i ≤ k,
fi : J1 → Ji be bijections and we form the set
mi
i
mi
i
mi
i
mi
i
mi
i
mi
i
E := {γf2(γ)...fk(γ) : γ ∈ J1} .
Then E = Γ. Since for all γ ∈ Ji, 1 ≤ i ≤ k, the order of γ is a power of
pi, if γ ∈ J1, fi(γ) ∈ Γp
m1
1
for all 2 ≤ i ≤ k and hence
(γf2(γ)...fk(γ)) = πpm1
(γ).
1
πpm1
1
Similarly,
πp
mi
i
(γf2(γ)...fk(γ)) = πp
(fi(γ)),
mi
i
for all 2 ≤ i ≤ k. Hence, πp
Corollary 3.4, E is N-PR.
mi
i
(E) = πp
mi
i
(Ji) = Si for 1 ≤ i ≤ k. By
(cid:3)
12
ROBERT (XU) YANG
Remark. In [3] the terminology "N-large" sets is introduced. A set E ⊂ Γ
is N-large if QN (E) < E where QN : Γ → Γ/HN is the quotient map and
HN ⊂ Γ is the subgroup of elements of orders divisible by N.
Theorem 2.2 in [3] states that if E ⊂ Γ and N is the smallest integer for
which E is N-large, then for all primes powers pn dividing N there exists a
weak 1 − eπi/pn-Kronecker subset F ⊂ E with F = E.
First, we note that the assumption in Theorem 4.3 is weaker than the
assumption in Theorem 2.2 in [3]; specifically, if E is infinite and N-large for
minimal N = pmpm1
and n ≤ m for some integer n, then πpn(E) =
E. To see this, we argue by contradiction and assume πpn(E) < E. Let
M := pn−1pm1
k < N. If γ ∈ Γpn ∩ HN and k is the order of γ, then k
divides N while pn does not divide k. This implies k divides M and hence
γ ∈ HM . Thus Γpn ∩ HN = HM . As a result, the map
T : QM (E) → QN (E) × πpn(E)
T (QM (γ)) = (QN (γ), πpn(γ)).
1 ...pmk
k
1 ...pmk
is well-defined and injective. Thus, if πpn(E) < E, then
QM (E) ≤ QN (E)πpn(E) < EE = E
for infinite E. Thus E is M-large and this contradicts the assumption that
N is minimal.
Moreover, recall that pn-PR sets are special weak 1 − eπi/pn-Kronecker
sets. This shows Theorem 4.3 improves Theorem 2.2 in [3] when E is un-
countable.
References
[1] J. Galindo and S. Hernandez, The concept of boundedness and the Bohr compactifi-
cation of a MAP Abelian group, Fundamenta Mathematicae, 159 (3):195 -- 217, 2010.
[2] C. C. Graham, K. E. Hare and T. W. Korner, ε-Kronecker and I0 sets in abelian
groups, II: sparseness of products of ε-Kronecker sets, Math. Proc. Cambridge Philos.
Soc., 140(3): 491-508, 2006.
[3] C. C. Graham and K. E. Hare, Existence of large ε-Kronecker and F Z I0(U ) sets in
discrete abelian groups, Colloquium Math: 1-15, 2012.
[4] C. C. Graham and K. E. Hare, Interpolation and Sidon sets for compact groups,
Springer, New York, 2013.
[5] C. C. Graham and A. Lau, Relative weak compactness of orbits in Banach spaces
associated with locally compact groups, Transactions of the American Mathematical
Society, 359 (3):1129-1160, 2007.
[6] K. E. Hare and L. T. Ramsey, The Relationship Between ε-Kronecker Sets and Sidon
Sets, Canad. Math. Bull. 59: 521-527, 2016.
[7] Hewitt, Edwin, The asymmetry of certain algebras of Fourier-Stieltjes transforms,
Michigan Math. J. 5 , no. 2: 149-158, 1958.
[8] E. Hewitt and S. Kakutani, A class of multiplicative linear functionals on the measure
algebra of a locally compact Abelian group, Illinois J. Math. 4: 553-574, 1960.
EXISTENCE OF LARGE INDEPENDENT-LIKE SETS
13
[9] K. Kunen and W. Rudin, Lacunarity and the Bohr topology, Math. Proc. Camb. Phil.
Soc. 126: 117-137, 1999.
[10] J. von Neumann, Ein System algebraisch unabhangiger Zahlen, Ann. of Math. 99:
134-141, 1928.
[11] J. J. Rotman, An Introduction to the Theory of Groups, 4th ed., Grad. Texts in
Math. 148, Springer, New York, 1995.
[12] W. Rudin, Fourier-Stieltjes transforms of measures on independent sets, Bull. Amer.
Math. Soc. 66: 199-202, 1960.
[13] W. Rudin, Fourier Analysis on Groups, Interscience Tracts in Pure and Applied
Math., No. 12. Wiley, New York, 1962.
[14] N. Th. Varopoulus, Tensor algebras over discrete spaces, J. Functional Anal., 3:
321-335, 1969.
Department of Pure Mathematics University of Waterloo Waterloo
ON N2L 3G1, CANADA
E-mail address: yangxu [email protected]
|
1803.10944 | 1 | 1803 | 2018-03-29T07:42:40 | Functional Version for Furuta Parametric Relative Operator Entropy | [
"math.FA"
] | Functional version for the so-called Furuta parametric relative operator entropy is here investigated. Some related functional inequalities are also discussed. The theoretical results obtained by our functional approach immediately imply those of operator versions in a simple, fast and nice way. | math.FA | math |
FUNCTIONAL VERSION FOR FURUTA PARAMETRIC
RELATIVE OPERATOR ENTROPY
MUSTAPHA RAISSOULI1
,
2 AND SHIGERU FURUICHI3
Abstract. Functional version for the so-called Furuta parametric relative
operator entropy is here investigated. Some related functional inequalities are
also discussed. The theoretical results obtained by our functional approach
immediately imply those of operator versions in a simple, fast and nice way.
1. Introduction
Let H be a complex Hilbert space. We denote by B(H) the C∗-algebra of
bounded linear operators acting on H and by B+∗(H) the open cone of all (self-
adjoint) positive invertible operators in B(H). Let A, B ∈ B+∗(H) and p ∈ [0, 1]
be a real number. The following
A∇pB = (1 − p)A + pB, A!pB =(cid:16)(1 − p)A−1 + pB−1(cid:17)−1
,
A♯pB = A1/2(cid:16)A−1/2BA−1/2(cid:17)p
A1/2
are known in the literature as the weighted arithmetic mean, weighted harmonic
mean and weighted geometric mean of A and B, respectively. If p = 1/2 they
are simply denoted by A∇B, A!B and A♯B, respectively. The previous operator
means satisfy the following relationships
A∇pB = B∇1−pA, A!pB = B!1−pA, A♯pB = B♯1−pA.
(1.1)
It is well-known that the following double inequality
(1.2)
holds for any A, B ∈ B+∗(H) and p ∈ [0, 1]. Here, the notation T ≤ S means
that T, S ∈ B(H) are self-adjoint and S − T is positive semi-definite.
A!pB ≤ A♯pB ≤ A∇pB
Otherwise, the relative operator entropy S(AB) and the Tsallis relative oper-
ator entropy Tp(AB) are, respectively, defined by, see [2, 3, 6]
S(AB) = A1/2 log(cid:0)A−1/2BA−1/2(cid:1)A1/2, Tp(AB) =
The following double inequality is known in the literature:
A♯pB − A
p
, p 6= 0.
A − AB−1A ≤ S(AB) ≤ B − A.
(1.3)
Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz.
2010 Mathematics Subject Classification. Primary 46N10; Secondary 46A20, 47A63, 47N10,
39B62, 52A41.
Key words and phrases. Operator inequalities, functional inequalities, operator entropies,
convex analysis.
1
2
M. RAISSOULI, S. FURUICHI
In [5], Furuta introduced a parametric extension of S(AB) as follows
(1.4)
Sp(AB) = A1/2(cid:0)A−1/2BA−1/2(cid:1)p
log(cid:0)A−1/2BA−1/2(cid:1)A1/2.
In fact, Sp(AB) was introduced in [5] for any real number p, but here we restrict
ourselves to the case p ∈ [0, 1].
As pointed out in [5], it is not hard to see that
S0(AB) = S(AB), S1(AB) = −S(BA) and Sp(AB) = −S1−p(BA).
The fundamental goal of this paper is to give an extension of Sp(AB) when
the operator variables A and B are (convex) functionals. Some functional rela-
tionships and inequalities are provided as well. The related operator versions are
deduced in a fast and nice way.
2. Functional Extensions
The previous operator concepts have been extended from the case that the vari-
ables are positive operators to the case that the variables are convex functionals,
see [10].
Let RH be the extended space of all functionals defined from H into R∪{+∞}.
Let f, g ∈ RH be two given functionals (convex or not) and p ∈ (0, 1). The
following
Ap(f, g) = (1 − p)f + pg,
Hp(f, g) =(cid:16)(1 − p)f ∗ + pg∗(cid:17)∗
,
sin(pπ)
tp−1
(1 − t)p Ht(f, g)dt
Z 1
0
Gp(f, g) =
π
are called, by analogy, the weighted functional arithmetic mean, the weighted
harmonic mean and the weighted geometric mean of f and g, respectively. Here,
the notation f ∗ refers to the Fenchel conjugate of f defined through
∀x∗ ∈ H
f ∗(x∗) = sup
x∈H(cid:8)ℜehx∗, xi − f (x)(cid:9).
(2.1)
For p = 1/2 we will denote the previous functional means by A(f, g), H(f, g)
and G(f, g), respectively. We extend these means on the whole interval [0, 1] by
setting:
A0(f, g) = H0(f, g) = G0(f, g) = f, A1(f, g) = H1(f, g) = G1(f, g) = g.
(2.2)
It is worth mentioning that (2.2) is not immediate from the definition of their
related functional means, since our involved functionals f and/or g can take the
value +∞, with the convention 0.(+∞) = (+∞) − (+∞) = +∞.
With this, analogous relationships of (1.1) for the previous functional means
are also valid, i.e.
Ap(f, g) = A1−p(g, f ), Hp(f, g) = H1−p(g, f ), Gp(f, g) = G1−p(g, f ).
In fact, the two first relations are immediate from their definitions and for the
third one see a detailed proof in [9]. Also, analog of (1.2), i.e.
Hp(f, g) ≤ Gp(f, g) ≤ Ap(f, g)
(2.3)
FUNCTIONAL VERSION
3
holds for any f, g ∈ RH and p ∈ [0, 1]. Here the notation f ≤ g means that
g(x) − f (x) ≥ 0 for all x ∈ H, with the convention +∞ − (+∞) = +∞ as usual
in convex analysis. The double inequality (2.3) implies that the three involved
functional means are with finite values whenever f and g are as well.
In two earlier papers [10] and [11] we extended S(AB) and Tp(AB) from
operators to (convex) functionals, respectively, as follows:
S(f g) =Z 1
0
Ht(f, g) − f
t
dt,
Tp(f g) =
Gp(f, g) − f
p
, p 6= 0.
The previous functional concepts were constructed as extensions of their related
operator versions in the following sense: if O(A, B) is one of the previous operator
concepts, its functional extension F (f, g) is such that
where the notation fT , for any T ∈ B(H), refers to the quadratic function gener-
ated by the operator T , i.e. fT (x) = (1/2)hT x, xi for all x ∈ H.
F(cid:0)fA, fB(cid:1) = fO(A,B),
(2.4)
3. Needed Tools
Let f ∈ RH. We denote by dom f := {x ∈ H, f (x) < +∞} the so-called
effective domain of f . The notation int(dom f ) refers to the topological interior
of dom f in H. The Fenchel conjugate f ∗ of f defined by (2.1) satisfies
f ∗(x∗) := sup
x∈dom f(cid:8)ℜehx∗, xi − f (x)(cid:9)
for any x∗ ∈ H. As supremum of a family of affine (so convex) functions, f ∗
is always convex even if f is not. The conjugate map f 7−→ f ∗ is point-wise
increasing and convex. That is, f ≤ g implies g∗ ≤ f ∗, and the inequality
holds for any f, g ∈ RH and p ∈ [0, 1].
(cid:0)(1 − p)f + pg(cid:1)∗ ≤ (1 − p)f ∗ + pg∗
The sub-differential of f at x ∈ dom f is the set ∂f (x) defined through
∂f (x) =(cid:8)x∗ ∈ H; ∀z ∈ H f (z) ≥ f (x) + ℜehx∗, z − xi(cid:9).
As it is well known, ∂f (x) is a (possibly empty) convex and closed set. If x ∈
int(dom f ) then ∂f (x) 6= ∅.
In the case where ∂f (x) 6= ∅ then we have the
equivalence:
x∗ ∈ ∂f (x) ⇐⇒ f (x) + f ∗(x∗) = ℜehx∗, xi.
As usual we denote by Γ0(H) the cone of all functionals f ∈ RH that are
convex, lower semi-continuous and proper (i.e. not identically equal to +∞). It
is well-known that f ∗∗ := (f ∗)∗ ≤ f for any f ∈ RH and, f ∈ Γ0(H) if and only
if f = f ∗∗ := (f ∗)∗. Moreover, x∗ ∈ ∂f (x) always implies x ∈ ∂f ∗(x∗), with
reversed implication provided that f ∈ Γ0(H).
4
M. RAISSOULI, S. FURUICHI
The function f is called Gateaux-differentiable (in short G-differentiable) at x
if the directional derivative
′
f
(x, d) = lim
t↓0
f (x + td) − f (x)
t
of f at x exists in all direction d ∈ H and the map d 7−→ f ′(x, d) is linear.
In this case we write f ′(x, d) = ∇f (x).d and ∇f (x) is called the G-derivative
of f at x.
It is well known that if f is convex and G-differentiable at x then
∂f (x) = {∇f (x)}.
For the sake of clearness and simplicity for the reader, we state the following
example illustrating the previous concepts.
Example 3.1. Let A ∈ B(H) and let fA be the quadratic function associated to
A, i.e. fA(x) = (1/2)hAx, xi for all x ∈ H.
(i) Assume that A ∈ B+∗(H). Then fA is convex and G-differentiable on H and
so
∀x ∈ H
∂fA(x) = {∇fA(x)} = {Ax}.
The coefficient 1/2 appearing in fA enjoys a symmetry role, in the aim to have
(fA)∗(x∗) = (1/2)hA−1x∗, x∗i for all x∗ ∈ H, or in short (cid:0)fA(cid:1)∗ = fA−1.
(ii) For any A, B ∈ B(H) it is easy to check that fA ± fB = fA±B and fA(Bx) =
fBAB(x) for any x ∈ H.
The following result, which will be needed later, has been proved in [12].
Theorem 3.2. Let f ∈ Γ◦(H) be such that int(dom f ) is nonempty. Then
(i) The following inequality
sup
x∗∈∂f (x)(cid:0)f ∗ − g∗(cid:1)(x∗) ≤ S(f /g)(x) ≤ (g − f )(x)
holds true for all x ∈ int(dom f ).
(ii) If f is moreover G-differentiable at x, then we have
f ∗(cid:0)∇f (x)(cid:1) − g∗(cid:0)∇f (x)(cid:1) ≤ S(f /g)(x) ≤ (g − f )(x).
As explained in [12], (3.1) as well as (3.2) is a functional extension of (1.3) from
positive operators to convex functionals.
For the sake of simplicity for the reader, we need to introduce auxiliary nota-
tion. For f, g ∈ RH and p ∈ [0, 1] we set
(3.1)
(3.2)
T ∗
p (f g) = (cid:0)Gp(f, g)(cid:1)∗ − f ∗
p
, p 6= 0.
(3.3)
We have the following result summarizing the elementary properties of T ∗
p (f g).
Proposition 3.3. The following assertions hold:
(i) For any p ∈ [0, 1) one has
T ∗
1−p(gf ) = (cid:0)Gp(f, g)(cid:1)∗ − g∗
1 − p
.
FUNCTIONAL VERSION
5
(ii) For all p ∈ (0, 1], the left hand-side of the following inequality
(cid:0)Ap(f, g)(cid:1)∗
(x∗) − f ∗(x∗)
p
≤ T ∗
p (f g)(x∗) ≤ g∗(x∗) − f ∗(x∗)
holds for any x∗ ∈ H while the right hand-side holds for x∗ such that g∗(x∗) = +∞
or x∗ ∈ dom f ∗.
Proof. (i) Follows from (3.3) with the relation Gp(f, g) = G1−p(g, f ).
(ii) From (2.3) we obtain by taking the conjugate side to side
Remarking that
we then deduce the desired result.
(cid:0)Ap(f, g)(cid:1)∗ ≤(cid:0)Gp(f, g)(cid:1)∗ ≤(cid:0)Hp(f, g)(cid:1)∗
(cid:0)Hp(f, g)(cid:1)∗ ≤ (1 − p)f ∗ + pg∗
.
(cid:3)
Proposition 3.4. For any A, B ∈ B+∗(H) and p ∈ (0, 1] there holds
Proof. By (3.3) with (2.4) we have
T ∗
p (fAfB) = fTp(A−1B−1).
T ∗
p (fAfB) = (cid:0)Gp(fA, fB)(cid:1)∗ − f ∗
p
A
=
f ∗
A♯pB − f ∗
A
p
=
f(A♯pB)−1 − fA−1
p
=
fA−1♯pB−1 − fA−1
p
.
This, with the fact that αfT = fαT and fT − fS = fT −S for any α ∈ R and
T, S ∈ B(H), immediately yields the desired result.
(cid:3)
4. Functional Version of Sp(AB)
As already pointed before, our aim here is to give an analog of Sp(AB) when
the operator arguments A and B are (convex) functionals f and g, respectively.
Such analog seems to be hard to define from (1.4), since (1.4) involves product
of operators whose analogues for functionals is not known yet. For this, we need
to state the following result.
Theorem 4.1. The following equalities
Sp(AB) = −
S(cid:0)A♯pBA(cid:1)
p
=
S(cid:0)A♯pBB(cid:1)
1 − p
.
(4.1)
hold for any A, B ∈ B+∗(H) and p ∈ (0, 1).
Proof. Indeed, we have the property
T ∗S(AB)T = S(T ∗AT T ∗BT )
for any A, B ∈ B∗(H) and any invertible operator T ∈ B(H) by using Kubo-Ando
theory [7] and the integral form
S(AB) =Z 1
0
A!tB − A
t
dt.
6
M. RAISSOULI, S. FURUICHI
We thus have the first equality as
S (A♯pBA) = A1/2S(cid:0)I♯pA−1/2BA−1/2I(cid:1) A1/2 = A1/2S(cid:0)(A−1/2BA−1/2)pI(cid:1) A1/2
= −A1/2(A−1/2BA−1/2)p log(A−1/2BA−1/2)pA1/2 = −pSp(AB),
since S(AI) = −A log A for any A ∈ B+∗(H).
The second equality can be proved in a similar manner.
(cid:3)
Now, to give a functional version of Sp(AB) we use (4.1) which is more appro-
priate for our aim, since (4.1) involves only operator concepts (relative operator
entropy and operator geometric mean) whose functional extensions were already
done. Taking into account to have a symmetric character between p and 1 − p in
our desired definition, we then put the following.
Definition 4.2. Let f, g ∈ RH and p ∈ [0, 1]. We set
Sp(cid:0)f g(cid:1) =
with
S(cid:16)(cid:0)Gp(f, g)(cid:1)g(cid:17)
2(1 − p)
−
S(cid:16)(cid:0)Gp(f, g)(cid:1)f(cid:17)
2p
,
S0(f g) = S(f g) and S1(f g) = −S(gf ).
As first result we state the following.
Proposition 4.3. Let f, g ∈ RH. Then we have
Further, if dom f = dom g = H then the equality
S1/2(f g) = S(cid:16)(cid:0)G(f, g)(cid:1)g(cid:17) − S(cid:16)(cid:0)G(f, g)(cid:1)f(cid:17).
Sp(f g) = −S1−p(gf )
holds for any p ∈ (0, 1).
(4.2)
(4.3)
(4.4)
Proof. The equality (4.3) is immediate from (4.2). However, we mention that
(4.4) is not immediate from (4.2), since our involved functionals could take the
value +∞. Indeed, we pay attention to the fact that, if φ, ψ ∈ RH, the equality
φ−ψ = −(ψ−φ) is not always true, unless dom φ∪dom ψ = H. For this reason we
have assumed in our statement that dom f = dom g = H in the aim to guarantee
that Ht(cid:0)Gp(f, g), g(cid:1) or Ht(cid:0)Gp(f, g), f(cid:1) is with finite values. With this, (4.4) can
be deduced from (4.2) when we refer to the relationship Gp(φ, ψ) = G1−p(ψ, φ)
valid for any φ, ψ ∈ RH and p ∈ [0, 1].
(cid:3)
A connection relationship between the functional parametric entropy Sp(f g)
and the operator parametric entropy Sp(AB) is expressed by the following result.
Proposition 4.4. Let A, B ∈ B+∗(H) and p ∈ [0, 1]. Then we have:
Proof. By (4.2), with (2.4) and (4.1), we have
Sp(cid:0)fAfB(cid:1) = fSp(AB).
Sp(cid:0)fAfB(cid:1) =
S(cid:0)fA♯pBfB(cid:1)
2(1 − p)
−
S(cid:0)fA♯pBfA(cid:1)
2p
=
f
S(cid:0)A♯pBB(cid:1)
2(1 − p)
−
(4.5)
.
f
S(cid:0)A♯pBA(cid:1)
2p
FUNCTIONAL VERSION
7
This, with similar arguments as in the proof of Proposition 3.4, implies the desired
result.
(cid:3)
Relationship (4.5) justifies that Sp(f g) is a reasonable extension of Sp(AB),
from operators to functionals, in the sense of (2.4).
For the sake of simplicity, we use in the next theorem and in its proof the
following notations:
Gp := Gp(f, g), G∗
We now are in a position to state the following main result.
p :=(cid:0)Gp(f, g)(cid:1)∗, ∇Gp := ∇(cid:0)Gp(f, g)(cid:1), ∂Gp := ∂(cid:0)Gp(f, g)(cid:1).
Theorem 4.5. Let f, g ∈ RH be such that int(cid:0)dom Gp(f, g)(cid:1) 6= ∅. Then the
2 sup
following double inequality
x∗∈∂Gp(x)
T ∗
1
1−p(gf )(x∗) + Tp(f g)(x)! ≤ Sp(f g)(x)
2 −T1−p(gf )(x) − sup
≤
1
T ∗
p (f g)(x∗)! (4.6)
x∗∈∂Gp(x)
sup
holds for any x ∈ int(cid:0)dom Gp(f, g)(cid:1) and p ∈ (0, 1).
Proof. Since int(cid:0)dom Gp(f, g)(cid:1) 6= ∅ then ∂Gp(x) 6= ∅ for any x ∈ int(cid:0)dom Gp(f, g)(cid:1).
Now, according to Theorem 3.2 we have, for x ∈ int(cid:0)dom Gp(f, g)(cid:1)
p − f ∗(cid:1)(x∗) ≤ S(cid:0)Gpf(cid:1)(x) ≤(cid:0)f − Gp(cid:1)(x),
p − g∗(cid:1)(x∗) ≤ S(cid:0)Gpg(cid:1)(x) ≤(cid:0)g − Gp(cid:1)(x).
x∗∈∂Gp(x)(cid:0)G∗
x∗∈∂Gp(x)(cid:0)G∗
Multiplying (4.7) by −1/p and (4.8) by 1/(1 − p) and then summing side to side
we obtain the desired inequalities, after simple manipulations with the help of
Proposition 3.3. Details are simple and therefore omitted here.
(cid:3)
(4.7)
(4.8)
and
sup
Remark 4.6. It is worth mentioning that the condition int(cid:0)dom Gp(f, g)(cid:1) 6= ∅ is
satisfied if int(cid:0)dom f ∩ dom g(cid:1) 6= ∅, since dom f ∩ dom g ⊂ dom Gp(f, g).
Corollary 4.7. Let f, g ∈ RH be such that Gp(f, g) is G-differentiable at x ∈ H.
Then the following inequalities (in the point-wise order sense)
1
2(cid:16)T ∗
1−p(gf )(cid:0)∇Gp(f, g)(cid:1) + Tp(f g)(cid:17) ≤ Sp(f g)
1
hold for any p ∈ (0, 1).
≤
2(cid:16) − T1−p(gf ) − T ∗
p (f g)(cid:0)∇Gp(f, g)(cid:1)(cid:17)
Proof. Since Gp(f, g) is G-differentiable at x then ∂Gp(f, g)(x) = {∇Gp(f, g)(x)}.
Substituting this in (4.6) and using the definition of the point-wise order we
immediately obtain the desired inequalities.
(cid:3)
The operator version of the above theorem (and corollary) reads as follows.
8
M. RAISSOULI, S. FURUICHI
Corollary 4.8. Let A, B ∈ B+∗(H) and p ∈ (0, 1). Then we have
1
2(cid:16)(A♯pB)T1−p(B−1A−1)(A♯pB) + Tp(AB)(cid:17) ≤ Sp(AB)
1
≤
2(cid:16) − T1−p(BA) − (A♯pB)Tp(A−1B−1)(A♯pB)(cid:17).
Proof. Combining Corollary 4.7, Proposition 3.4 and Example 3.1,(ii) we obtain
the desired operator inequalities after simple manipulations. Details are simple
and therefore omitted here.
(cid:3)
Corollary 4.8 gives the relation between Furuta parametric relative operator
entropy and Tsallis relative operator entropy, in more general setting than the
result in [4, Theorem 2.3].
Acknowledgement
The author (S.F.) was partially supported by JSPS KAKENHI Grant Number
16K05257.
References
[1] I. Ekeland and R. Temam, Convex Analysis and Variational Problems, SIAM, 1999.
[2] J.I. Fujii and E. Kamei, Relative Operator Entropy in noncommutative Information Theory,
Math. Japonica, 34 (1989), 341-348.
[3] S. Furuichi, Inequalities for Tsallis Relative Entropy and Generalized Skew Information,
Linear Multilinear Alg., 59/10 (2011), 1143-1158.
[4] S. Furuichi and N. Minculete, Inequalities for relative operator entropies and operator means,
Acta. Math. Vietnam (2018). https://doi.org/10.1007/s40306-018-0250-7.
[5] T. Furuta, Parameteric extensions of Shannon inequality and its reverse one in Hilbert space
operators, Linear Algebra Appl., 381 (2004), 219-235.
[6] H. Isa, M. Ito, E. Kamei, H. Tohyama and M. Watanabe, Shannon type inequalities of a
relative operator entropy including Tsallis and R´enyi ones, Ann. Funct. Anal., 6/4 (2015),
289-300.
[7] F. Kubo and T. Ando, Means of positive linear operators, Math. Ann. 246 (1980), 205-224.
[8] P.J. Laurent, Approximation et Optimisation, Hermann, 1972.
[9] M. Raıssouli and H. Bouziane,Arithmetico-geometrico-harmonic functional mean in convex
analysis, Ann. Sc. Math. Qu´ebec, 30/1 (2006), 79-107.
[10] M. Raıssouli, Tsallis relative entropy for convex functionals, Int. J. Pure Appl. Math.,
51/4 (2009), 555-563.
[11] M. Raıssouli, Relative functional entropy in convex analysis, J. Appl. Anal., 17 (2011),
231-239.
[12] M. Raıssouli, Functional versions for some operator entropy inequalities, Ann. Funct. Anal.,
6/2 (2015), 204-211.
[13] E.Zeidler, Nonlinear Functional Analysis and its Applications III, Springer-Verlag, 1984.
1 Department of Mathematics, Science Faculty, Taibah University, Al Madi-
nah Al Munawwarah, P.O.Box 30097, Zip Code 41477, Saudi Arabia.
2 Department of Mathematics, Science Faculty, Moulay Ismail University,
Meknes, Morocco.
FUNCTIONAL VERSION
9
3Department of Information Science, College of Humanities and Sciences, Ni-
hon University, 3-25-40, Sakurajyousui, Setagaya-ku, Tokyo, 156-8550, Japan.
E-mail address: [email protected]
E-mail address: [email protected]
|
Subsets and Splits