Professional Documents
Culture Documents
51 (2003)
1. Introduction
The shift-invariant spaces are closed subspaces of L2 ( R n ) that are invariant under
all shifts (i.e., integer translations). The theory of shift-invariant subspaces of
L2 ( R n ) plays an important role in many areas, most notably in the theory of wave-
lets, spline systems, Gabor systems, and approximation theory [BMM; BDR1;
BDR2; BL; Bo1; HLPS; Ji; RS1; RS2; Rz2]. The study of analogous spaces for
L2 ( T, H) with values in a separable Hilbert space H in terms of range func-
tion, often called doubly invariant spaces, is quite classical and goes back to Hel-
son [He1].
The general structure of shift-invariant (SI) spaces was revealed in the work of
de Boor, DeVore, and Ron [BDR1] with the use of fiberization techniques based
on range function. In particular, conditions under which a finitely generated SI
space has a generating set satisfying some desirable properties (e.g., stability, or-
thogonality or quasi-orthogonality) were given. This has been further developed in
the work of Ron and Shen [RS1] with the introduction of the technique of Grami-
ans and dual Gramians. The general properties of SI spaces and shift-preserving
operators have also been studied by the first author [Bo1].
The contribution of this paper is a systematic study of yet another tool in SI
spaces, apparently overlooked in the previous research, which we call the spectral
function. This function was introduced by the second author in his Ph.D. the-
sis. It was motivated by [BDR1] and is similar to the multiplicity function studied
by Baggett, Medina, and Merrill [BMM]. More precisely, to every SI subspace
of L2 ( R n ) we associate a function on R n that contains much useful information
about that space.
Although [BDR1] and Helson’s range function [He1] is the origin of this ap-
proach, it is thanks to Weiss (see [WW]) that the spectral function has a very
elementary definition. Namely, for every SI space V ⊂ L2 ( R n ), there exists a
countable family of functions 8 whose integer shifts form a tight frame with con-
stant 1 for the space V, and the spectral function of V is defined as the sum of
the squares of the Fourier transforms of the elements of 8 (see Lemma 2.3). It
can be shown that such a function is well-defined, additive on orthogonal sums,
and bounded by 1. Moreover, it behaves nicely under dilations and modulations,
which makes it useful in the study of wavelet and Gabor systems. For example, it
387
388 M a rc i n Bow n i k & Z i e m ow i t R z e s z o t n i k
Finally, we need to recall a few facts about dual Gramian analysis of SI systems
introduced by Ron and Shen [RS1]. Suppose 8 ⊂ L2 ( R n ) is a countable set of
functions such that
X
|ϕ̂(ξ)|2 < ∞ for a.e. ξ ∈ R n. (1.5)
ϕ ∈8
Proposition 2.2. Let S be the set of all SI subspaces of L2 ( R n ). Then the spec-
tral function σV of V ∈ S is determined as the unique mapping
σ : S → L∞ ( R n )
that satisfies
¡P −1
|ϕ̂(ξ)|2 k∈ Z n |ϕ̂(ξ + k)|2 for ξ ∈ supp ϕ̂,
σS(ϕ)(ξ) = (2.3)
0 otherwise,
which is additive with respect to the orthogonal sums; that is,
M X
V = Vi for some Vi ∈ S H⇒ σV = σVi . (2.4)
i∈ N i∈ N
The Spectral Function of Shift-Invariant Spaces 391
In particular, (2.5) does not depend on the choice of 8 as long as E(8) is a tight
frame with constant 1 for V.
Proof. Let J(ξ) be the range function of V and let PJ (ξ) be the corresponding or-
thogonal projection onto J(ξ). By [Bo1, Thm. 2.5(ii)], E(8) is a tight frame with
constant 1 for V if and only if {T ϕ(ξ) : ϕ ∈ 8} is a tight frame with constant 1 for
J(ξ) for a.e. ξ. Therefore, for a.e. ξ ∈ T n,
X
kvk2 = |hv, T ϕ(ξ)i|2 for all v ∈ J(ξ).
ϕ ∈8
Hence X
kP(ξ)vk2 = |hv, T ϕ(ξ)i|2 for all v ∈ ` 2 ( Z n )
ϕ ∈8
Proof of Proposition 2.2. It suffices to show that the spectral function σ satisfies
(2.3) and (2.4). Indeed, any mapping σ : S → L∞ ( R n ) satisfying
L (2.3) and (2.4)
is unique, since any SI space can be decomposed as V = ϕ ∈8 S(ϕ) for some
countable family 8 ⊂ L2 ( R n ) (see e.g. [Bo1, Thm. 3.3] and [Rz2, Thm. 1.2]). To
392 M a rc i n Bow n i k & Z i e m ow i t R z e s z o t n i k
see (2.3), recall from [BL; BDR1] that if S(ϕ) is a PSI space then the function ϕ 0
given by
¡P
2 −1/2
|ϕ̂(ξ)| k∈ Z n |ϕ̂(ξ + k)| for ξ ∈ supp ϕ̂,
ϕ̂ 0 (ξ) =
0 otherwise,
is the quasi-orthogonal generator for S(ϕ), meaning that E(ϕ 0 ) is a tight frame
with constant 1 for S(ϕ).LHence, by Lemma 2.3, we have (2.3).
Next, suppose
L V = i∈ N Vi for some SI spaces Vi . We can decompose each
Vi as Vi = ϕ ∈8i S(ϕ) for some 8i ⊂ Vi such that E(8i ) forms a tight frame
with constant 1Sfor Vi , i ∈ N. Since E(8) forms a tight frame with constant 1 for
V, where 8 = i∈ N 8i , by Lemma 2.3 we have
X XX X
σV (ξ) = |ϕ̂(ξ)|2 = |ϕ̂(ξ)|2 = σVi (ξ).
ϕ ∈8 i∈ N ϕ ∈8i i∈ N
Next we will show that the spectral function behaves nicely with respect to the
action of modulations and dilations. This will be relevant in our study of Gabor
systems and wavelets—or, more generally, framelets. Recall that the modulation
by a vector a ∈ R n of f ∈ L2 ( R n ) is given by
Ma (f )(x) = e 2πiha, xi f (x).
The dilation by an n × n nonsingular matrix A of f ∈ L2 ( R n ) is given by
DA f (x) = |det A|1/2f (Ax).
We restrict our attention to dilations A preserving the lattice Z n because this is ex-
actly when in general we can expect that DAV is SI (with respect to the action of
Z n ) if V is SI.
Since 9d (ξ) ⊥ 9d 0 (ξ) for d 6= d 0 ∈ D and since k9d (ξ)k is either 0 or 1, the
system {9d (ξ) : d ∈ D} is a tight frame with constant 1 for its span. By (2.9),
{8l (ξ) : l ∈ L} is also a tight frame with constant 1 for the space J(ξ) =
span{8l (ξ) : l ∈ L} = span{9d (ξ) : d ∈ D}, where J(ξ) is the range func-
tion of DA S(ϕ). Therefore, for every k ∈ Z n,
X X
σDA S(ϕ)(ξ + k) = kPJ(ξ) ek k2 = |h8l (ξ), ek i|2 = |h9d (ξ), ek i|2
l∈L d∈D
Example 2. There exist distinct SI spaces that have identical spectral functions.
Indeed, let ϕ 0 , ϕ 1 ∈ L2 ( R) be given by ϕ̂ 0 = 2−1/2 (1 (0,1) + 1 (1, 2) ) and ϕ̂ 1 =
2−1/2 (1 (0,1) − 1 (1, 2) ). Then ϕ i is a quasi-orthogonal generator of Vi = S(ϕ i ); that
is, E(ϕ i ) is a tight frame of S(ϕ i ) with constant 1. Hence σV 0 = σV1 = 2−11 (0, 2) ,
but V0 ⊥ V1 .
We can now collect the main properties of the spectral function into a single
proposition.
XZ Z
= σV (ξ + l )1K (ξ + l ) dξ = σV (ξ) dξ,
l∈ Z n Tn K
where 0/0 should be interpreted as 0. De Boor, DeVore, and Ron [BDR2, Thms.
1.6, 1.7] showed that the approximation and density orders of a PSI space S(φ)
are characterized by the behavior of the function ξ 7→ 3 φ (ξ)/|ξ| k at the origin.
Moreover, they also proved the following remarkable result showing that approx-
imation by arbitrary (closed) SI subspaces of L2 ( R n ) can be reduced to the case
of PSI spaces.
words, we characterize all multiplicity functions associated with the core subspace
of a GMRA. This extends a result of Baggett and Merrill [BM], who considered
only locally integrable multiplicity functions. In this work we not only allow non-
integrable multiplicity functions, we also allow functions equal to ∞ on a set of
nonzero measure. As an immediate consequence, we also obtain a characteriza-
tion of wavelet dimension functions shown by Speegle and the authors in [BRS].
We start by recalling the notion of a GMRA, which has been studied by a num-
ber of authors [Ba2; BM; BMM; BL; Bo3; HLPS; LTW].
and V0 is SI. The space V0 is often called a core space of (Vj )j ∈ Z . An (orthonor-
mal) wavelet is a collection 9 = {ψ 1, . . . , ψ L } such that the system
{DAj Tk ψ l : j ∈ Z n, k ∈ Z n, l = 1, . . . , L}
is an orthonormal basis of L2 ( R n ).
There is a close relationship between wavelets and GMRAs. For any orthonormal
wavelet 9 we can associate a GMRA (Vj )j ∈ Z by setting
The key role in the proof of the sufficiency of (D1)–(D3) in Theorem 3.2 is played
by Lemma 3.3.
Lemma 3.3. Suppose that a function D satisfies (D1), (D2), and (D3). Then
there exists a measurable set S ⊂ R n such that
X
D(ξ) = 1S (ξ + k), (3.4)
k∈ Z n
S ⊂ BS, (3.5)
\
B jS = ∅, (3.6)
j ∈Z
[ [
B jS = B j1, (3.7)
j ∈Z j ∈Z
The proof of Lemma 3.3 is an adaptation of the proof of the characterization of di-
mension functions of wavelets that follows the constructive procedure described
in [BRS, Algo. 4.4]. The major difference with [BRS] is that we allow a func-
tion D to be infinite on some set with nonzero measure and that the consistency
equation [BRS, (D3) in Thm. 4.2] is replaced by our consistency inequality (D2).
Assuming that Lemma 3.3 holds, we can complete the proof of Theorem 3.2.
i=1
and let E m+1 ⊂ Ẽ m+1 be any measurable set such S that τ (E m+1) = τ (Ẽ m+1) and
τ | E m+1 is injective. We claim that the set S1 = m∈ N E m satisfies
Q ⊂ S1, S1 ⊂ BS1, τ | S1 is injective and onto A1 . (3.11)
The Spectral Function of Shift-Invariant Spaces 401
This can been seen by repeating verbatim the arguments in the proof of [BRS,
Thm. 4.2] and then using Lemma 3.4(i). We continue defining sets {S k } k∈ N by
induction with the use of a technical Lemma 3.5.
Lemma 3.5. Suppose that there exist sets S1, ..., S k such that Q ⊂ S1 = P 1 and
[i
P i ⊂ BP i for i = 1, ..., k, where P i = Sj ; (3.12)
j =1
where Fm ⊂ F̃m is any measurable set such that τ (Fm+1S ) = τ (F̃m+1) and τ | Fm+1
∞
is injective. We claim that S1, . . . , S k+1, where S k+1 = m=1 Fm satisfy (3.12)–
(3.14).
To see (3.12), it suffices to show that S k+1 ⊂ BP k+1. This follows from the fact
that F1 ⊂ BP k and Fm+1 ⊂ BFm ⊂ BS k+1 ⊂ BP k+1 for m ∈ N.
To see that τ | S k+1 is injective, suppose that ξ1, ξ 2 ∈ S n+1 and τ (ξ1 ) = τ (ξ 2 ).
Therefore, ξ1 ∈ Fj and ξ 2 ∈ Fk for some j, k ∈ N. If j < k (the case j > k is iden-
tical) then ξ 2 ∈/ FjP , which contradicts τ (ξ1 ) = τ (ξ 2 ). Thus j = k and ξ1 = ξ 2 ,
since τ | Fk is injective.
To see (3.14), it is enough to prove that S k+1 ∩ P k = ∅. Since F1 ⊂ BP k \ P k
and Fm ⊂ BFm−1 for m ≥ 2, by induction we obtain Fm ⊂ B mP k \ B m−1P k for all
m ∈ N. By (3.12) with i = k it follows that P k ⊂ B m−1P k for any m ∈ N. Hence
Fm ∩ P k = ∅, that is, S k+1 ∩ P k = ∅.
The proof of the remaining part of (3.13) (i.e., τ (S k+1) = A k+1) is much more
difficult. First note that, since τ (Fm ) ⊂ A k+1, we certainly have τ (S k+1) ⊂ A k+1.
For the reverse inclusion, we will find it useful to prove
APk+1 ∩ BS k+1
P P
⊂ S k+1. (3.16)
Indeed, if ξ ∈ APk+1 ∩ P
BS k+1, −1
then B ξ + l ∈ Fm for some m ∈ N and l ∈
Z n. Since ξ 0 := ξ + Bl ∈ BFm and ξ ∈ APk+1, we obtain ξ 0 ∈ BFm ∩ APk+1. If
Sm P P P
ξ0 ∈/ i=1 Fi , then by (3.15) we have ξ 0 ∈ F̃m+1 ⊂ S k+1 and hence ξ ∈ S k+1 .
0 P 0 P P
However, if ξ ∈ Fi for some i = 1, . . . , m, then ξ ∈ S k+1 and hence ξ ∈ S k+1 .
This shows (3.16).
Continuing with the proof of (3.13), it remains to show that A k+1 ⊂ τ (S k+1).
By Lemma 3.4(ii) there is a set H ⊂ 1 such that H ∩ P k = ∅, τ (H ) = A k+1,
and D(B −j ξ) ≥ 1 for every j ≥ 0 and ξ ∈ H. Therefore, all we have to prove is
P
τ (H ) ⊂ τ (S k+1), that is, H ⊂ S k+1 .
402 M a rc i n Bow n i k & Z i e m ow i t R z e s z o t n i k
P
We split the proof of H ⊂ S k+1 into two cases. First, we consider all ξ ∈ H
that, for every j ≥ 0, D(B −j ξ) ≥ k + 1; that is, we consider the set R :=
such T
H∩ ∞ j P −j
j =0 B A k+1. For a.e. ξ ∈R, it follows from (3.10) that B ξ ∈ Q for some
j ≥ 1. Since Q ⊂ S1 ⊂ P k , we can consider j 0 = min{j ∈ N : B −j ξ ∈ P k }. Since
H ∩ P k = ∅, it follows that B −j 0 +1ξ ∈ BP k \ P k . Moreover, since B −j 0 +1ξ ∈
APk+1, we obtain
B −j 0 +1ξ ∈ (BP k \ P k ) ∩ APk+1 = F̃1 ⊂ S k+1
P
.
S∞ P
Hence R ⊂ j =0 B jS k+1. Therefore,
∞
[ ∞
[ j
\
R= (R ∩ B jS k+1
P
)⊂ H∩ B iAPk+1 ∩ B jS k+1
P
.
j =0 j =0 i=0
By the induction hypothesis (3.13), at least one of the elements in (3.19) must lie
in the complement of P k . Hence there is a p ∈ Z n, and B −j 0 ξ + p ∈ BP k \ P k . In
addition, since B −j 0 ξ ∈ APk+1, we have B −j 0 ξ +p ∈ APk+1. Therefore, B −j 0 ξ +p ∈
P P P
F̃1 ⊂ S k+1 ; that is, ξ ∈ B j 0 S k+1 and hence ξ ∈ S k+1. Since both R and R̃ are con-
P P P
tained in S k+1, it follows that H ⊂ S k+1 and so A k+1 ⊂ S k+1 , which completes
the proof of (3.13) by the induction. We have thus shown the existence of the sets
{S k } k∈ N satisfying (3.12)–(3.14), which completes the proof of Lemma 3.5.
Finally, we are
S ∞ready to continue the proof of Lemma 3.3. By Lemma 3.5, define
the set S = k=1 S k . We claim that S satisfies (3.4)–(3.7). To see (3.4),P
it suffices
to show that the equality in (3.4) holds for a.e. ξ ∈ T n. By (3.14), 1S = ∞ k=1 1S k ,
and by (3.13),
X ∞ X
X
1S (ξ + l ) = 1S k (ξ + l )
l∈ Z n k=1 l∈ Z n
X∞ ∞
X
= 1τ (S k )(ξ) = 1A k (ξ) = D(ξ) for a.e. ξ ∈ T n.
k=1 k=1
For (3.5) it suffices to show that S k ⊂ BS for every k ∈ N, but this is immediate
from (3.12). For (3.7) it suffices to show that limj →∞ 1S (B −j ξ) = 1 for a.e. ξ ∈
1̃, which is an immediate consequence of Q ⊂ S and (3.10). T Finally, it remains
to show (3.6). By the contradiction, suppose that S̃ = j ∈ Z B jS has a nonzero
measure. Since S̃ = B S̃, we can partition S̃ into a countable family of subsets
{S̃ i}i∈ N such that S̃ i = B S̃ i and S̃ i has a nonzero (and hence infinite) measure for
all i ∈ N. Since B induces an ergodic endomorphism of T n (see [BRS, proof of
Thm. 5.11]), we have τ (S̃ i ) = T n. Therefore,
X X
D(ξ) = 1S (ξ + k) ≥ 1S̃ (ξ + k)
k∈ Z n k∈ Z n
XX
= 1S̃ i (ξ + k) = ∞ for a.e. ξ,
k∈ Z n i∈ N
which contradicts our initial assumption that {ξ ∈ T n : D(ξ) < ∞} has a nonzero
measure. This shows (3.6) and completes the proof of Lemma 3.3 under this
assumption.
1 = |S i | = |((B iS 0 ) \ (B i−1S 0 ) ∩ S|
\
≥ ((B iS 0 ) \ (B i−1S 0 )) ∩ B jS = |det B|i−1c.
j ∈Z
Theorem 3.6. Suppose (Vj )j ∈ Z is a GMRA. Then the dimension function of the
core space V0 , D(ξ) := dimV 0 (ξ), satisfies (D1)–(D3) and
(D4) lim infj →∞ D(B −j ξ) ≥ 1 for a.e. ξ ∈ R n.
Conversely, if D satisfies (D1)–(D4) then there exists a GMRA (Vj )j ∈ Z such that
dimV 0 (ξ) = D(ξ) for a.e. ξ ∈ R n.
Proof. Suppose that (Vj )j ∈ Z is a GMRA. Then, by Theorem 3.2, D(ξ) = dimV 0 (ξ)
satisfies (D1)–(D3). To see (D4), by Proposition 2.6(g) and (h) it suffices to show
that
lim inf σVj (ξ) = lim inf σV 0 (B −j ξ) = 1 for a.e. ξ ∈ R n. (3.20)
j →∞ j →∞
Example 3.7. For simplicity, assume that we work in the dimension n = 1 and
that the dilation factor A = 2. Given 0 < δ < 1/2, define a set S ⊂ R by
∞
[
S = [−1/2, 1/2] ∪ [2 j, 2 j + δ).
j =−1
T S
An elementary calculation shows that S ⊂ 2S, j ∈ Z 2 jS = ∅, and j ∈ Z 2 jS =
R . Hence the sequence (Vj )j ∈ Z , where Vj = Ľ2 (2 j S ), is a GMRA. Moreover, the
dimension function of its core space satisfies
∞ for τ (ξ) ∈ (0, δ),
dimV 0 (ξ) =
1 otherwise.
Theorem 4.1. Suppose (Vj )j ∈ Z is a GMRA such that {ξ ∈ R n : dimV 0 (ξ) < ∞}
has a nonzero (Lebesgue) measure. Then
∞
X
σW0 ((A∗ ) j ξ) = 1 for a.e. ξ, (4.1)
j =−∞
However, in general the spaces Wj are not SI (with respect to the standard lattice
Z n ) for j < 0. Indeed, for j < 0 we can be sure only that Wj is SI with respect to
a larger lattice A−j Z n. Nevertheless, this idea can be transformed into a rigorous
proof.
L ¡L ⊥
Proof of Theorem 4.1. Since V0 = j <0 Wj = j ≥0 Wj is SI and since V0 ⊕
L n
W
j ≥0 j = L2
( R ), by Proposition 2.6(b) and (g) we have
∞
X
1 = σL2 ( R n )(ξ) = σV 0 (ξ) + σWj (ξ)
j =0
∞
X
= σV 0 (ξ) + σW0 ((A∗ ) j ξ) for a.e. ξ. (4.4)
j =0
Pick any ξ 0 ∈ R n such that (4.4) holds for any ξ = Aiξ 0 , where i ∈ Z . By applying
(4.4) for ξ = Aiξ 0 and letting i → −∞, we obtain
∞
X
σW0 ((A∗ ) j ξ) ≤ 1 for a.e. ξ. (4.5)
j =−∞
This contradicts our initial hypothesis and so proves (4.1). Equation (4.2) follows
from (4.1) and (4.4).
Proof. By the semi-orthogonality condition, E(9) forms a tight frame with con-
P
stant 1 for W0 = S(9) and thus σW0 (ξ) = ψ∈9 |ψ̂(ξ)|2 by Lemma 2.3. By (4.4),
∞
X
σV 0 (ξ) = 1 − σW0 ((A∗ ) j ξ)
i=0
∞ X
X ∞
XX
= 1− |ψ̂((A∗ ) j ξ)|2 = |ψ̂((A∗ ) j ξ)|2,
j =0 ψ∈9 ψ∈9 j =1
P P∞
where we used the Calderón formula ψ∈9 j =−∞ |ψ̂((A∗ ) j ξ)|2 = 1 for a.e. ξ
(see e.g. [Bo2; HW]). This completes the proof of Theorem 4.2.
As a corollary to Proposition 2.6(h) and Theorem 4.2 we obtain the usual formula
for the wavelet dimension function [BRS; LTW; RS4; We].
Remark. It should be noted that neither Theorem 4.2 nor Corollary 4.3 are true
for general affine tight frame systems with constant 1. Indeed, Theorem 3.1 of
[PSWX] shows that a (dyadic) normalized tight frame wavelet ψ ∈ L2 ( R) is
P P
semi-orthogonal if and only if j∞=1 k∈ Z |ψ̂(2 j(ξ + k))|2 is integer-valued al-
most everywhere. Therefore, one can expect that the right-hand side of (4.7) is a
dimension function of V0 , and thus integer-valued, only if 9 is a semi-orthogonal
wavelet. As a result, Theorem 4.2 and Corollary 4.3 are not valid for more general
tight affine frame systems.
Proof. The standard dilation argument allows us to reduce the general case to
A = Id. Indeed, this follows immediately from DA(gBl, Am ) = (DA g)AT Bl, m ,
where l, m ∈ Z n. We present a proof by contradiction; that is, we shall assume that
i i=1, ..., L
the system {gBl , m } l, m∈ Z n is complete and then prove that |det B| ≤ L follows.
We start by constructing a multiresolution scheme for Gabor systems. Suppose
i i=1, ..., L
that {g 1, . . . , g L } ⊂ L2 ( R n ) and that {gBl , m } l, m∈ Z n is the corresponding Gabor
system. For every l ∈ Z n, define
i n L
Gl = span{gBl, m : m ∈ Z , i = 1, . . . , L} = S({MBl (g ), . . . , MBl (g )}).
1
(5.2)
n
We will use the lexicographic order on ( Z , <) defined by (k1, k 2 , . . . , k n ) <
(l1, l 2 , . . . , l n ) if there exist r = 1, . . . , n such that kj = lj for all j < r and k r <
lr . For any l ∈ Z n, let
]
Vl = span{gBk, m : k < l, k, m ∈ Z n } = Gk . (5.3)
k<l
n
Let S be the closed subspace of L ( R ) spanned by {gBl, m }l, m∈ Z n . Our discussion
2
is general and we do not yet assume that the system is complete. We would expect
to have the following properties of the “multiresolution analysis” (Vl )l∈ Z n :
(G1) Vk ⊂ Vl for any k < l ∈ Z n ;
n
(G2) M SBk (Vl ) = Vl+k for any k, l ∈ Z ;
(G3) Tl∈ Z n Vl is dense in S;
(G4) l∈ Z n Vl = {0};
(G5) V0 is shift-invariant, where 0 = (0, . . . , 0) ∈ Z n.
Indeed, all of these conditions except (G4) follow immediately from (5.3). How-
ever, a serious difficulty arises because (G4) does not hold in general (see Example
5.2). Nevertheless, if we assume in addition that the functions g 1, . . . , g L are band-
limited, then this easily implies that (G4) holds. Indeed, if supp ĝ 1, . . . , supp ĝ L ⊂
K for some bounded set K ⊂ R n, then Gl ⊂ Ľ2 (Bl +K) for any l ∈ Z n. Therefore,
] [
Vk = Gl ⊂ Ľ2 (Bl + K)
l<k l<k
and
\ \ [ \ [
Vk ⊂ Ľ2 (Bl + K) = Ľ2 B (l + B −1K) = {0},
k∈ Z n k∈ Z n l<k k∈ Z n l<k
where the orthogonal sum is taken over all nonempty beginning intervals I ⊂ Z n.
By (G6), the orthogonal sum in (5.7) is effectively taken only over the beginning
intervals of type (i), which clearly implies (5.6).
The decomposition (5.6) allows us to see that the spectral function of S is given
by a simple formula. In fact, we have
W0 = V1 ª V0 = (V0 ] G0 ) ª V0 .
By (5.2), G0 is a SI space with the dimension function dim G0 (ξ) ≤ L for a.e. ξ ∈
R n. Hence, by (G5), W0 is also a SI space with the dimension function that satisfies
dimW0 (ξ) ≤ dim G0 (ξ) ≤ L for a.e. ξ ∈ R n. Therefore, W0 has a quasi-orthogonal
basis 9 ⊂ W0 with cardinality at most L; that is, E(9) is a tight frame with con-
P
stant 1 for W0 (see [BDR1]). The spectral function of W0 is thus σW0 = ψ∈9 |ψ̂|2
by Lemma 2.3. Therefore, by Proposition 2.6, (5.5), and (5.6), we have
X X
σS (ξ) = σWl (ξ) = σMBl (W0 )(ξ)
l∈ Z n l∈ Z n
X X X
= σW0 (ξ − Bl ) = |ψ̂(ξ − Bl )|2 (5.8)
l∈ Z n l∈ Z n ψ∈9
for a.e. ξ ∈ R n.
The system {gBl, m }l, m∈ Z n is complete in L2 ( R n ) if and only if S = L2 ( R n ),
which by Proposition 2.6 is equivalent to having σS = 1 a.e. If we integrate this
equality over the cube B([0, 1] n ), then by (5.8) we have
Z X X
|det B| = |ψ̂(ξ − Bl )|2 dξ
B([0,1] n ) l∈ Z n ψ∈9
X XZ X
= |ψ̂(ξ)|2 dξ = kψk22.
ψ∈9 l∈ Z n B(l+[0,1] n ) ψ∈9
lim kPS r f − f k2 = 0.
r→∞
Integrating this limit equality over the cube B([0, 1] n ) yields |det B| ≤ L. Indeed,
by Fatou’s lemma we have
Z X X
|det B| = lim |ψ̂(ξ − Bl )|2 dξ
B([0,1] n ) j →∞ l∈ Z n ψ∈9
j
Z X X X
≤ lim inf |ψ̂(ξ − Bl )|2 dξ = lim inf kψk22 ≤ L,
j →∞ B([0,1] n ) l∈ Z n ψ∈9 j →∞
j ψ∈9j
The following example shows that condition (G4) does not hold in general.
Let g ∈ L2 ( R) be such that g(x) 6= 0 if and only if x ∈ (0, 1), and also let
R1
0 log|g(x)| dx = −∞. Then, by a classical result of Helson [He1, pp. 13, 21],
the system of functions {e 2πkxg(x) : k ∈ N} is complete in L2 (0, 1) := {f ∈
L2 ( R) : supp f ⊂ (0, 1)}. As an immediate consequence,
span{e 2πkxg(x) : k < l} = L2 (0, 1) for any l ∈ Z .
Therefore, Vl = L2 ( R) for any l ∈ Z , and (5.11) holds. This example shows that
we cannot expect (G4) to hold unless g satisfies some additional hypotheses (e.g.,
that g be band-limited).
References
[Ba1] L. W. Baggett, Processing a radar signal and representations of the discrete
Heisenberg group, Colloq. Math. 60/61 (1990), 195–203.
[Ba2] , An abstract interpretation of the wavelet dimension function using group
representations, J. Funct. Anal. 173 (2000), 1–20.
The Spectral Function of Shift-Invariant Spaces 413
M. Bownik Z. Rzeszotnik
Department of Mathematics Mathematical Institute
University of Michigan University of Wrocław
Ann Arbor, MI 48109 50-384 Wrocław
Poland
marbow@umich.edu
Current address
Department of Mathematics
University of Texas
Austin, TX 78712
zioma@mail.ma.utexas.edu