A general sampling theory in the functional

Hilbert space induced by a Hilbert space valued

kernel
Antonio G. Garca

Alberto Portal

* Departamento de Matematicas, Universidad Carlos III de Madrid, Avda. de la Uni-

versidad 30, 28911 Leganes-Madrid, Spain. E-mail:agarcia@math.uc3m.es
Departamento de Matematicas, Universidad Carlos III de Madrid, Avda. de la Uni-
versidad 30, 28911 Leganes-Madrid, Spain. E-mail:aportal@math.uc3m.es
Abstract
Let H be a separable Hilbert space and a xed subset of R. Consider an H-
valued function K : H and x H. Then, the function f
x
: C given by
f
x
(t) :=

x, K(t)
_
H
is well-dened. Denote by H
K
the set of functions obtained in this
way. Although a variety of sampling results for H
K
is known in the literature, there
exist some simple examples where they do not apply because the implicit interpolation
condition that appears does not adjust the former pattern we need. The main aim of
this paper is to obtain a more general sampling result including most of these special
cases. In this way, the concept of interpolation condition is redened and we study how
to combine two of them in order to obtain a new sampling result. Some examples are
obtained in this new framework.
Keywords: Reproducing kernel Hilbert spaces; Sampling formulas; Biorthonormal Riesz
bases.
AMS: 94A20; 44A05; 46E22.
1 Statement of the problem
For the past few years a signicant mathematical literature on the topic of sampling theorems
associated with dierential or dierence problems has ourished [1, 4, 5, 6, 7, 8]. See also [15]
and the references therein. In turn, we might consider the Weiss-Kramer sampling theorem
as the leitmotiv of all these results [11, 13]. Roughly speaking, the common situation for
these sampling problems is the following:
Let f be a function dened on C by f(t) =
_
I
F(x) K(x, t) dx, F L
2
(I), (or f(t) =

n
F(n) K(n, t), F
2
). The kernel K, which belongs to L
2
(I) (or
2
) for each xed t C,
satises the dierential (dierence) equation appearing in a dierential (dierence) problem
(P) which has the sequence of eigenvalues {t
n
}. Moreover, whenever we substitute in K the
spectral parameter t by {t
n
} we obtain the sequence of orthogonal eigenfunctions associated
with (P) which constitutes an orthogonal basis for L
2
(I) (
2
). Under these circumstances,
299
JOURNAL OF APPLIED FUNCTIONAL ANALYSIS,VOL.3,NO.3,299-313,COPYRIGHT 2008 EUDOXUS PRESS, LLC
f is an entire function which can be recovered from its samples {f(t
n
)} by means of a
sampling formula f(t) =

n
f(t
n
) S
n
(t), where the sampling functions {S
n
} are given by
S
n
(t) = K(, t
n
)
2
K(, t), K(, t
n
) (the inner product in L
2
(I) or
2
).
All the results above can be formulated in an abstract way following the approach in
Saitohs book [12]. Namely, let H be a separable Hilbert space, and a xed subset of
R. Given an H-valued function K : H, for x H, the function f(t) := x, K(t)
H
is
well-dened as a function f : C. We denote by H
K
the set of functions obtained in
this way and by T the linear transform
T : H H
K
x f
(1)
Hereafter we refer the function K as the kernel of the transform T. Note that the continuity
of the kernel K implies that the functions in H
K
are continuous in , a natural setting for
sampling purposes. If we dene in H
K
the norm f
H
K
= inf{x
H
: f = T(x)} we obtain
a reproducing kernel Hilbert space (RKHS hereafter) whose reproducing kernel is given by
k(t, s) := K(s), K(t)
H
i.e., for each s the function k
s
dened as k
s
(t) := k(t, s) belongs
to H
K
, and the reproducing property
f(s) = f, k
s

H
K
= f, k(, s)
H
K
, s , f H
K
(2)
holds. Recall that the Moore-Aronszajn procedure [2] leads to the same RKHS via the
positive denite (or positive matrix) function k. Under these circumstances it is known that
the linear operator T is one-to-one if and only if T is an isometry between H and H
K
,
or, equivalently, if and only if the set of functions {K(t)}
t
is complete in H [12]. An
important property of H
K
is that convergence in its norm implies pointwise convergence.
In fact, by the reproducing property, we have that
|f(t) f
n
(t)| = |

f f
n
, k(, t)
_
| f f
n

H
K
K(t)
H
.
Notice that the last inequality also implies uniform convergence in subsets of where the
function k(t, t) = K(t)
2
is bounded. The RKHS H
K
has been largely studied in the
mathematical literature (see the superb monograph [12] and references therein).
A sampling result for H
K
can be easily established (see [9] or [10]). Namely, let {x
n
}

n=1
and {x

n
}

n=1
be a pair of biorthonormal Riesz bases for a Hilbert space H. Assume that, for
each xed t , K(t) can be written as K(t) =

n=1
S
n
(t) x

n
, where the functions S
n
H
K
satisfy, for some xed sequence {t
n
}

n=1
in , the interpolation property: S
n
(t
m
) = a
n

n,m
for some constants {a
n
}

n=1
C \ {0}. Then, any function f H
K
can be expanded as
f(t) =

n=1
f(t
n
)
S
n
(t)
a
n
, t , (3)
where the convergence of the series is absolute and uniform on subsets of where the
function K(t) is bounded.
Recall that a Riesz basis {w
n
}

n=1
for H is the image of an orthonormal basis by means of
a bounded invertible operator on H. Any Riesz basis {w
n
}

n=1
has a unique biorthonormal
(dual) Riesz basis {w

n
}

n=1
, i.e., such that w
n
, w

H
=
n,m
, and the expansions
x =

n=1
x, w

H
w
n
=

n=1
x, w
n

H
w

n
GARCIA-PORTAL
300
hold for every x H (see [3] or [14] for more details and proofs).
Perhaps the most important examples of RKHS H
K
that verify the mentioned result are
the classical Paley-Wiener spaces of bandlimited functions, i.e., square integrable functions
in R such that their Fourier transforms are zero outside a bounded set in R. For instance,
any function of the form
f(t) =
1

2
_

F(x)e
itx
dx, t R,
where F L
2
[, ], can be expanded as the cardinal series
f(t) =

n=
f(n) sinc(t n) , t R,
where sinc stands for cardinal sine function (or sinc function) dened as sinc t = sint/t
for t = 0 and sinc 0 = 1.
The sampling series (3) might also contain samples from functions which are related to
f in some sense. Thus, this sampling result can be generalized in the following way: Let
{x
n
}

n=1
{y
n
}

n=1
and {x

n
}

n=1
{y

n
}

n=1
be a pair of biorthonormal Riesz bases for the
Hilbert space H. For each xed t , K(t) can be written as
K(t) =

n=1
_
S
n
(t)x

n
+ T
n
(t)y

,
where S
n
(t) and T
n
(t) denote the evaluation at t of the functions S
n
= T(x
n
) H
K
and T
n
= T(y
n
) H
K
obtained by means of the linear transform (1) .
Assume that there exist two kernels K
1
, K
2
: H each dening a function in the
way K does, i.e., f
j
(t) := x, K
j
(t)
H
, j = 1, 2. Let T
1
and T
2
be the corresponding linear
transforms. These kernels can be written as
K
j
(t) =

n=1
_
S
j
n
(t)x

n
+ T
j
n
(t)y

, j = 1, 2 ,
where S
j
n
(t) = T
j
(x
n
)[t] and T
j
n
(t) = T
j
(y
n
)[t] for t , j = 1, 2. Suppose there exist two
sequences {s
n
}

n=1
and {t
n
}

n=1
in such that the functions {S
j
n
}

n=1
and {T
j
n
}

n=1
, j = 1, 2,
satisfy the interpolation conditions
S
1
n
(s
m
) = a
n

n,m
; T
1
n
(s
m
) = b
n

n,m
, {a
n
}

n=1
, {b
n
}

n=1
C
S
2
n
(t
m
) = c
n

n,m
; T
2
n
(t
m
) = d
n

n,m
, {c
n
}

n=1
, {d
n
}

n=1
C,
where
n
:= a
n
d
n
b
n
c
n
= 0 for all n N. Suppose that f and the functions f
1
and f
2
are
related in the sense that ker T ker T
1
ker T
2
. This implies that ker T = {0} so that H
K
becomes a RKHS under the inner product f, g
H
K
:= x, y
H
where Tx = f and Ty = g.
Under these conditions, the technique used in [9] gives the following result:
Theorem 1 The sequence {S
n
}

n=1
{T
n
}

n=1
is a Riesz basis for H
K
and, expansions with
respect to this basis allow to recover any function f in H
K
from the samples {f
1
(s
n
)}

n=1
GENERAL SAMPLING THEORY
301
and {f
2
(t
n
)}

n=1
of f
1
and f
2
, by means of the sampling formula
f(t) =

n=1
_
d
n
f
1
(s
n
) b
n
f
2
(t
n
)

n
S
n
(t) +
a
n
f
2
(t
n
) c
n
f
1
(s
n
)

n
T
n
(t)
_
=

n=1
_
f
1
(s
n
)
d
n
S
n
(t) c
n
T
n
(t)

n
+ f
2
(t
n
)
a
n
T
n
(t) b
n
S
n
(t)

n
_
, t .
(4)
The convergence of the series is absolute and uniform on subsets of where the function
K(t) is bounded.
Using a matrix notation, formula (4) can be written as
f(t) =

n=1
_
f
1
(s
n
) f
2
(t
n
)
_
_
a
n
c
n
b
n
d
n
_
1
_
S
n
(t)
T
n
(t)
_
, (5)
from which it is not dicult to derive a more general result involving the samples of N
functions related to f.
Multi-channel sampling in Paley-Wiener spaces can be easily derived from this approach
[9]. As a dierent example of Theorem 1 we can obtain a Hermite-type interpolation series
for H
K
, i.e., a sampling series which involves samples of any function f H
K
and its rst
derivative, and in addition, the sampling functions generalize the classical Hermite inter-
polation polynomials. Indeed, let {t
n
}

n=1
be a sequence of distinct nonzero real numbers
such that

n=1
|t
n
|
2
< . There exists an entire function P(t) with simple zeros at the
sequence {t
n
}

n=1
. Specically, the function P(t) is given by the canonical product
P(t) =
_

n=1
_
1
t
t
n
_
exp(t/t
n
) if

n=1
|t
n
|
1
=

n=1
_
1
t
t
n
_
if

n=1
|t
n
|
1
< .
Consider {x
n
}

n=1
{y
n
}

n=1
and {x

n
}

n=1
{y

n
}

n=1
a pair of biorthonormal Riesz bases
for H and Q(t) := P(t)
2
, which has double zeros at {t
n
}

n=1
. Take the functions
S
n
(t) =
Q(t)
(t t
n
)
2
and T
n
(t) =
Q(t)
t t
n
and dene the kernels K(t), K
1
(t) and K
2
(t), t R, by
K(t) =

n=1
_
Q(t)
(t t
n
)
2
x

n
+
Q(t)
t t
n
y

n
_
,
K
1
(t) = K(t) and K
2
(t) = K

(t). It is easy to check the interpolation conditions:

S
n
(t
m
) =
Q

(t
n
)
2

n,m
; T
n
(t
m
) = 0
S

n
(t
m
) =
Q

(t
n
)
6

n,m
; T

n
(t
m
) =
Q

(t
n
)
2

n,m
.
GARCIA-PORTAL
302
Taking into account that Q

(t
n
) = 0 for all n N, any function f H
K
can be expanded
as the series
f(t) =

n=1
_
f(t
n
)
_
1
Q

(t
n
)
3Q

(t
n
)
(t t
n
)
_
+ f

(t
n
)(t t
n
)
_
2Q(t)
Q

(t
n
)(t t
n
)
2
, t R.
1.1 An easy anomalous example
Although Theorem 1 is a quite general sampling result, the following example exhibits a
situation where it does not work. Indeed, consider the usual orthonormal basis for L
2
[, ]
given by
_
1

2
_

_
1

cos nx
_

n=1

_
1

sinnx
_

n=1
and the kernels
K(t) = K
1
(t) = cos tx + sintx andK
2
(t) = cos tx,
from which we dene, for each L
2
[, ], the transforms
f(t) = T(t) =

, K(t)
_
and f
i
(t) = T
i
(t) =

, K
i
(t)
_
, i {1, 2} .
Notice that, if L
2
[, ], we obtain that

, cos(tx)
_
is an even function of t and that

, sin(tx)
_
is an odd one. Consequently, ker T implies that T
2
(t) =

, cos(tx)
_
=

, sin(tx)
_
is both an odd and an even function of t. Therefore, = 0 so that T is
one-to-one.
The corresponding sampling functions are given by
S
0
(t) =
_
1

2
, K(t)
_
=

2 sinc t
S
n
(t) =
_
1

cos nx, K(t)

_
=
2t(1)
n
sint

(t
2
n
2
)
(n N)
T
n
(t) =
_
1

sinnx, K(t)
_
=
2n(1)
n
sint

(t
2
n
2
)
(n N) .
In this case, S
1
0
= S
2
0
= S
0
, S
1
n
= S
2
n
= S
n
, T
1
n
= T
n
and T
2
n
= 0 for n N. For n, m N,
the following interpolation condition holds:
_
S
1
n
(m) T
1
n
(m)
S
2
n
(m) T
2
n
(m)
_
=
m,n
_

0
_
.
However, what is happening with S
1
0
and S
2
0
? Even if we decide to dene T
1
0
= T
2
0
:= 0, the
interpolation matrix that we used in (5) will be singular and we will not be able to apply
Theorem 1.
This example gives us a suitable generalization that, in practice, can be very useful. As
we only need one coecient for S
0
(T
0
is not dened), we can consider the matrix
_
S
1
0
(m)
S
2
0
(m)
_
=
m,0
_
2

2
_
(6)
where m N
0
:= N{0}. The coecient of S
0
(t) can be chosen in two ways for the sampling
formula, f
1
(0)/

2 or f
2
(0)/

2, the choice being at our disposal; we might choose the

simplest, or perhaps the only one available, etc.
GENERAL SAMPLING THEORY
303
2 A general sampling result in H
K
The aim in this section is to prove a new sampling result which also applies for anomalous
examples. To this end, we need to introduce new interpolation conditions which will be
combined in an appropriate way to obtain the desired sampling result. These topics are the
subject of the next three subsections:
2.1 Interpolation condition of type (L, M)
Consider L, M N such that L M and dene M := {1, 2, . . . , M} and L := {1, 2, . . . , L}.
Consider a linear independent system for H written as
{x
1,n
}
nI
{x
2,n
}
nI
{x
M,n
}
nI
where I N can be a nite set. Suppose that we have L linear transforms T
1
, T
2
, . . . , T
L
with associate kernels K
1
(t), K
2
(t), . . . , K
L
(t) and dened as follows
f

(t) = T

(x)[t] =

x, K

(t)
_
, L .
Denote
S

m,n
(t) = T

(x
m,n
)[t] , t ,
where L, m M and n I.
Denition 1 We say that an interpolation condition of type (L, M) is satised by these
elements if there exist L sets of points {t

n
}
nI
in , L, such that, for any xed L
and m M,
S

m,n
(t

k
) = a
n
,m

n,k
, n, k I ,
holds, and the coecients a
n
,m
C verify that the rank of the matrices
A
n
:=
_
_
_
_
_
a
n
1,1
a
n
1,2
a
n
1,M
a
n
2,1
a
n
2,2
a
n
2,M
.
.
.
.
.
.
.
.
.
.
.
.
a
n
L,1
a
n
L,2
a
n
L,M
_
_
_
_
_
is just M for all n I.
Denition 2 Denote by any increasing function from M into L, which we shall call a
choice function. For any matrix B with L rows b(1), b(2), . . . , b(L), we dene the choice of
M rows of B by means of as
B

:=
_
b[(1)] b[(2)] b[(M)]
_

.
As we can see, B

is a matrix obtained from B by choosing the rows of B given by

(1), (2), . . . , (M).
In these terms, the rank of A
n
is M if and only if there exists a choice function
n
such
that A
n
n
is not singular.
GARCIA-PORTAL
304
Denition 3 We shall call such a
n
, an A
n
-regular choice function.
Any A
n
-regular choice function is related to A
n
in such a way that A
n

is regular.
However, this choice function can be applied to any matrix with L rows and any number of
columns.
If we have a vector v =
_
v
1
v
2
v
L
_

, then we can choose the same rows from

both A
n
and v. The product of the matrix and the vector obtained in such a way is given
by
A
n

=
_
_
_
_
_
_
a
n
(1),1
a
n
(1),2
a
n
(1),M
a
n
(2),1
a
n
(2),2
a
n
(2),M
.
.
.
.
.
.
.
.
.
.
.
.
a
n
(M),1
a
n
(M),2
a
n
(M),M
_
_
_
_
_
_
_
_
_
_
_
v
(1)
v
(2)
.
.
.
v
(M)
_
_
_
_
_
.
2.2 Compatibility
We began this article by showing an example in which two interpolation conditions were
used. For the rst one, I
1
= {0}, the linear independent system was just one vector and we
had only one transform, i.e., it was an interpolation condition of type (1, 1). In the second
one, I
2
= N, the partition of the linear independent system had two elements and there were
two transforms, so it was an interpolation condition of type (2, 2). The goal of this section
is to establish which properties must be veried by two interpolation conditions in order for
a sampling theorem to be possible. It is the topic compatible interpolation conditions.
The fact of working with two interpolation conditions at least makes the notation used
hard. For the sake of clarity, hereafter, an index k
j
denotes that the indexed element
corresponds to the j-th interpolation condition of those we are using.
Consider two interpolation conditions of types (L
1
, M
1
) and (L
2
, M
2
), respectively. This
means that we have two linear independent systems in H:
S
1
:={x
1
1,n
1
}
n
1
I
1
{x
1
2,n
1
}
n
1
I
1
{x
1
M
1
,n
1
}
n
1
I
1
S
2
:={x
2
1,n
2
}
n
2
I
2
{x
2
2,n
2
}
n
2
I
2
{x
2
M
2
,n
2
}
n
2
I
2
where I
1
, I
2
N (possibly nite) and M
1
, M
2
N. Moreover, we suppose the j-th interpo-
lation condition, j {1, 2}, has L
j
N transforms T

j
,
j
L
j
and M
j
L
j
, dened from
each kernel K

j
(t),
j
L
j
, by
f

j
(t) = T

j
(x)[t] =

x, K

j
(t)
_
,
j
L
j
.
Denote
S

i
m
j
,n
j
= T

i
(x
j
m
j
,n
j
)
for
i
L
i
, m
j
M
j
, n
j
I
j
and i, j {1, 2}, i.e., the image of S
1
S
2
by each of the
transforms of both the interpolation conditions is calculated.
As they are interpolation conditions, there exist sequences {t

j
k
j
}
k
j
I
j
, where
j
L
j
and
j {1, 2}, such that
S

j
m
j
,n
j
(t

j
k
j
) = a
n
j

j
,m
j

n
j
,k
j
GENERAL SAMPLING THEORY
305
holds for m
j
M
j
,
j
L
j
, n
j
, k
j
I
j
and j {1, 2} in such a way the coecients a
n
j

j
,m
j
C
satisfy that the rank of the matrix
A
n
j
:=
_
_
_
_
_
_
a
n
j
1,1
a
n
j
1,2
a
n
j
1,M
j
a
n
j
2,1
a
n
j
2,2
a
n
j
2,M
j
.
.
.
.
.
.
.
.
.
.
.
.
a
n
j
L
j
,1
a
n
j
L
j
,2
a
n
j
L
j
,M
j
_
_
_
_
_
_
is just M
j
for all n
j
I
j
, j {1, 2}.
Denition 4 Two interpolation conditions are compatible if the following compatibility
condition holds:
S

2
m
1
,n
1
(t

2
k
2
) = 0 , S

1
m
2
,n
2
(t

1
k
1
) = 0 ,
for
j
L
j
, m
j
M
j
, n
j
, k
j
I
j
, j {1, 2}.
Observe that this condition implies that S
1
S
2
is a linear independent system. Indeed,
suppose there exists x
2
m
2
,n
2
in S
2
such that
x
2
m
2
,n
2
=
M
1

m
1
=1

m
1
x
1
m
1
,n
1
+
N

k=1

k
x
k
where x
k
S
1
S
2
is not in {x
2
m
2
,n
2
, x
1
1,n
1
, x
1
2,n
1
, . . . , x
1
M
1
,n
1
} and there exists an index
k
1
M
1
= {1, 2, . . . , M
1
} such that
k
1
= 0. For all
1
L
1
, we obtain that
0 = [T

1
x
2
m
2
,n
2
](t

1
n
1
) =
M
1

m
1
=1

m
1
a
n
1

1
,m
1
,
i.e., A
n
1
has a zero column or it has at least two linear dependent columns so that its rank
cannot be M
1
.
2.3 The sampling result
Consider the pair of dual Riesz bases for H given by
R
_
r=1
Mr
_
mr=1
{x
mr,nr
}
nrIr
and
R
_
r=1
Mr
_
mr=1
{x

mr,nr
}
nrIr
(7)
and suppose that R N interpolation conditions are satised, and each is compatible with
every other (see Denition 1 and Denition 4).
Suppose that the following condition is satised, as well:
ker T
R

r=1
Lr

r=1
ker T
r
. (8)
GARCIA-PORTAL
306
We show the transform T is one-to-one. First, observe that the kernel K can be written
as
K(t) =
R

r=1
Mr

mr=1

nrIr
S
mr,nr
(t)x

mr,nr
where S
mr,nr
(t) =

x
mr,nr
, K(t)
_
. Analogously, for
q
L
q
and 1 q R, the kernel K
q
can be expanded as
K
q
(t) =
R

r=1
Mr

mr=1

nrIr
S
q
mr,nr
(t)x

mr,nr
.
Suppose that x H veries Tx = 0. Then, for all q {1, 2, . . . , R} and
q
L
q
, we have
that
f
q
= T
q
x = 0
and therefore, for p
q
I
q
,
0 = f
q
(t
q
pq
) =
_
x, K
q
(t
q
pq
)
_
=
_
x,
R

r=1
Mr

mr=1

nrIr
S
q
mr,nr
(t
q
pq
)x

mr,nr
_
.
Thus, compatibility implies
0 =
_
x,
Mq

mq=1

nqIq
S
q
mq,nq
(t
q
pq
)x

mq,nq
_
,
and by using the q-th interpolation condition, we have
0 =
_
x,
Mq

mq=1
a
pq
q,mq
x

mq,pq
_
=
Mq

mq=1
a
pq
q,mq

x, x

mq,pq
_
, 1
q
L
q
.
Thus, we have a homogeneous linear system with L
q
equations and M
q
unknowns whose
unique solution is the trivial one. Consequently, we obtain that

x, x

mq,pq
_
= 0 for all
q {1, 2, . . . R}, m
q
M
q
and p
q
I
q
. Since {x

mr,nr
: 1 r R, m
r
M
r
, n
r
I
r
} is a
Riesz basis, x = 0. Observe that the following sequences
R
_
r=1
Mr
_
mr=1
{S
mr,nr
}
nrIr
and
R
_
r=1
Mr
_
mr=1
{S

mr,nr
}
nrIr
(9)
are dual Riesz bases of H
K
.
If we do the same for any x H such that f = Tx, we obtain the consistent linear
system
f
q
(t
q
pq
) =
Mq

mq=1
a
pq
q,mq

x, x

mq,pq
_
, 1
q
L
q
,
which has a unique solution. For each p
q
I
q
we can nd a choice function
pq
such that
A
pq
pq
is regular. Thus, we can write the coecients

x, x

mq,pq
_
with respect to the samples
f
q
(t
q
pq
) by means of
_

x, x

1,pq
_
x, x

2,pq
_

x, x

Mq,pq
_
_

=
_
A
pq
pq
_
1
F
pq
pq
,
GENERAL SAMPLING THEORY
307
where F
pq
pq
is obtained from the vector
F
pq
:=
_
f
1
(t
1
pq
) f
2
(t
2
pq
) f
Lq
(t
Lq
pq
)
_

(10)
by using the choice function
pq
.
Now, expanding f by using the Riesz basis given by {S
mr,nr
: n
r
I
r
}
Mr
mr=1
and taking
into account that T is an isometry, we have that
f(t) =
R

r=1
Mr

mr=1

nrIr

f, S

mr,nr
_
H
K
S
mr,nr
(t)
=
R

r=1

nrIr
Mr

mr=1

x, x

mr,nr
_
H
S
mr,nr
(t)
=
R

r=1

nrIr
_
F
nr
nr
_

_
_
A
nr
nr
_
1
_

S
nr
(t) ,
where F
nr
is given by (10) and
S
nr
(t) =
_
S
1,nr
(t) S
2,nr
(t) S
Mr,nr
(t)
_

, (11)
which is just the sampling formula we are looking for. Convergence is, as we know, in the
H
K
-norm sense and, also, absolute in and uniform in those subsets of where K(t) is
bounded. Consequently, we have proved the next result:
Theorem 2 Consider the dual Riesz bases given by (7). Suppose we have R N interpo-
lation conditions of types (L
r
, M
r
), where 1 r R, each two of them being compatible (in
the sense of Denition 4). Assume condition (8) is satised. Then, the sets in (9) are dual
Riesz bases for H
K
and for each set of A
nr
-regular choice functions
_

nr
: M
r
L
r
| n
r

I
r
, 1 r R
_
, we have that any function f H
K
can be recovered from its samples
_
f
r
(t
r
nr
) : n
r
I
r
,
r
L
r
_
R
r=1
by the following sampling formula
f(t) =
R

r=1

nrIr
_
F
nr
nr
_

_
_
A
nr
nr
_
1
_

S
nr
(t) , t ,
where F
nr
and S
nr
are given by (10) and (11), respectively. Convergence is absolute and,
also, uniform in those subsets of where K(t) is bounded.
Notice that the number of transforms we have for each interpolation condition of type
(L, M) is M at least. However, it is not important how many of them we have at most. In
fact, only the possibility of nding a regular choice of rows of A
n
is needed. Thus, if L is
a nite or innite set of indices, we have an interpolation condition of type (cardL, M) if
there exist some points {t

n
: n I}
L
such that, for any xed L and m M,
S

m,n
(t

k
) = a
n
,m

n,k
, n, k I ,
GARCIA-PORTAL
308
holds, and for each n I there exists an A
n
-regular choice function
n
where
A
n
=
_
a
n
,1
a
n
,2
a
n
,M
_
L
represents a function from L into C
M
for each n I. If these more general interpolation
conditions are used, Theorem 2 still remains valid.
3 Sampling by using linear operators in H
Let H be a separable Hilbert space. Consider two dual Riesz bases of H written as
{x
1,n
}

n=1
{x
2,n
}

n=1
{x
M,n
}

n=1
,
{x

1,n
}

n=1
{x

2,n
}

n=1
{x

M,n
}

n=1
,
where M N. Given a kernel K : R H, we dene the linear transform T as in
(1). Suppose we have a family of bounded linear operators {L

: H H}

. Related to
f = Tx H
K
we dene the functions f

(t) :=

L

x, K(t)
_
H
, where . For any xed
n N and 1 m M, we denote S
n
,m
(t) :=

x
m,n
, K(t)
_
H
and, for any n N, we write
S
m,n
(t) :=

x
m,n
, K(t)
_
H
.
In the sequel, we assume that x ker T implies L

x ker T for all (which

occurs when T commutes with L

for all ), and that there exists a family of sequences

_
{t

k
}
kN
:
_
such that S
n
,m
(t

k
) = a
n
,m

n,k
for n, k N, 1 m M and .
For each n N, dene the function
A
n
: C
M
(a
n
,1
, a
n
,2
, . . . , a
n
,M
)
and suppose that there exists a sequence {
n
: M }
nN
of A
n
-regular choice functions.
Then, the following result holds:
Theorem 3 Under the hypotheses as above, any f H
K
can be recovered from its samples
_
{f

(t

n
)}
nN
:
_
by means of the sampling formula
f(t) =

nN
_
F
n
n
_

_
_
A
n
n
_
1
_

S
n
(t) , t ,
where F
n
() := f

(t

n
) and S
n
(t) =
_
S
1,n
(t) S
2,n
(t) S
M,n
(t)
_

. Convergence is ab-
solute and uniform in subsets of where K(t) is bounded.
Proof: Dening K

(t) := L

[K(t)] for , where L

denotes the adjoint operator of L

,
we have
f

(t) := T

(x)[t] =

x, K(t)
_
H
=

x, K

(t)
_
H
, t .
If x ker T then, by assumption, L

x ker T for all , i.e., 0 =

L

x, K(t)
_
H
=

x, K

(t)
_
H
for all . As a consequence,
ker T

ker T

,
and Theorem 2 implies the desired result.
GENERAL SAMPLING THEORY
309
4 Some illustrative examples
In this section we go back to our anomalous example in order to handle it into the new
setting. We also give another example which involves two interpolation conditions of type
(2, 2).
4.1 The introductory example revisited
Consider the orthonormal basis of L
2
[, ] given by
_
1

2
_

_
1

cos nx
_

n=1

_
1

sinnx
_

n=1
and the kernels K
1
(t) = K(t) = cos tx + sintx, K
2
(t) = cos tx sintx and K
3
(t) = cos tx.
Notice that, if L
2
[, ], we obtain that

, cos(tx)
_
is an even function of t and that

, sin(tx)
_
is an odd one. Consequently, ker T implies that T
3
(t) =

, cos(tx)
_
=

, sin(tx)
_
is both an odd and an even function of t. Thus, T is injective and trivially
{0} = ker T ker T
1
ker T
2
ker T
3
.
Sampling functions are given by
S
0
(t) =
_
1

2
, K(t)
_
=

2 sinc t
S
n
(t) =
_
1

cos nx, K(t)

_
=
2t(1)
n
sint

(t
2
n
2
)
(n N)
T
n
(t) =
_
1

sinnx, K(t)
_
=
2n(1)
n
sint

(t
2
n
2
)
(n N)
Easy calculations show that S
1
n
= S
2
n
= S
3
n
= S
n
, that T
3
n
= 0 and that T
2
n
= T
1
n
= T
n
for n N and S
1
0
= S
2
0
= S
3
0
= S
0
. Thus, we have two interpolation conditions of types
(3, 1) and (3, 2), respectively. The rst one veries that:
_
_
S
1
0
(m)
S
2
0
(m)
S
3
0
(m)
_
_
=
_
_

2
_
_

0,m
m N {0} ,
and the second one, that:
_
_
S
1
n
(m) T
1
n
(m)
S
2
n
(m) T
2
n
(m)
S
3
n
(m) T
3
n
(m)
_
_
=
_
_

0
_
_

n,m
m N {0} ,
so we have a couple of compatible interpolation conditions.
We dene f(t) :=

F, K(t)
_
and f
k
(t) :=

F, K
k
(t)
_
for k = 1, 2, 3 and F L
2
[, ].
Finally, Theorem 2 yields the following sampling result:
Corollary 4 Any function f dened as
f(t) =
1

F(x)[cos tx + sintx]dx, t R,
GARCIA-PORTAL
310
where F L
2
[, ], can be recovered from the samples f(0) and {f
k
(n)}

n=1
, k = 1, 2, 3,
of its related functions f
1
, f
2
, f
3
by means of the following sampling formula:
f(t) = f(0) sinc t +
1

n=1
_
G
1
(n)S
n
(t) + G
2
(n)T
n
(t)

, t R,
where
_
G
1
(n) G
2
(n)
_
is any row of the matrix
_
_
_
_
_
_
f
3
(n) f
3
(n) f
2
(n)
f
3
(n) f
1
(n) f
3
(n)
f
2
(n)+f
1
(n)
2
f
1
(n)f
2
(n)
2
_
_
_
_
_
_
. (12)
Convergence is absolute and uniform in subsets of where K(t) is bounded.
This result allows us to recover any classical band-limited function to [, ] by means
of the samples of the related functions f
1
, f
2
, f
3
since band-limited functions can be written
as
f(t) =
1

2
_

F(x)[cos tx + sintx]dx, t R,
where F L
2
[, ]. Notice that the above integral representation involves the Hartley
transform of F (see [16]).
4.2 Another example
In this example, we use two compatible interpolation conditions of type (2, 2). To this end,
consider {x
n
, y
n
}

n=1
{ x
n
, y
n
}

n=1
, a Riesz basis for H whose dual Riesz basis is given by
{x

n
, y

n
}

n=1
{ x

n
, y

n
}

n=1
. Suppose we have ve kernels K, K
1
, K
2
,

K
1
,

K
2
each of them
dening a transform, as usual: f(t) = (Tx)[t] :=

x, K(t)
_
, f
j
(t) = (T
j
x)[t] :=

x, K
j
(t)
_
,
and

f
j
(t) = (

T
j
x)[t] :=

x,

K
j
(t)
_
where t , x H and j {1, 2}, denoting a tilde
over an element that it is related to the second interpolation condition. Assume that the
following condition holds:
ker T ker T
1
ker T
2
ker

T
1
ker

T
2
.
Denote
S
n
:= Tx
n
, T
n
:= Ty
n
,

S
n
:= T x
n
,

T
n
:= T y
n
,
and suppose there exist sequences {s
n
}

n=1
, {t
n
}

n=1
, { s
n
}

n=1
and {

t
n
}

n=1
such that
_
(T
1
(x
n
))(s
m
) (T
1
(y
n
))(s
m
)
(T
2
(x
n
))(t
m
) (T
2
(y
n
))(t
m
)
_
=
_
a
n
1,1
a
n
1,2
a
n
2,1
a
n
2,2
_

m,n
=: A
n

m,n
for the rst interpolation condition, and
_
(

T
1
( x
n
))(s
m
) (

T
1
( y
n
))(s
m
)
(

T
2
( x
n
))(t
m
) (

T
2
( y
n
))(t
m
)
_
=
_
a
n
1,1
a
n
1,2
a
n
2,1
a
n
2,2
_

m,n
=:

A
n

m,n
GENERAL SAMPLING THEORY
311
for the second one, being the matrices A
n
and

A
n
invertible. For these interpolation condi-
tions the compatibility condition reads:
_
(

T
1
(x
n
))( s
m
) (

T
1
(y
n
))( s
m
)
(

T
2
(x
n
))(

t
m
) (

T
2
(y
n
))(

t
m
)
_
=
_
(T
1
( x
n
))(s
m
) (T
1
( y
n
))(s
m
)
(T
2
( x
n
))(t
m
) (T
2
( y
n
))(t
m
)
_
=
_
0 0
0 0
_
.
Finally, as a consequence of Theorem 2 we deduce the following sampling result for the
preceding compatible interpolation conditions:
Corollary 5 Any function f in the Hilbert space H
K
can be recovered from the sequences of
samples {f
1
(s
n
)}

n=1
, {f
2
(t
n
)}

n=1
, {

f
1
( s
n
)}

n=1
and {

f
2
(

t
n
)}

n=1
by means of the following
sampling formula
f(t) =

n=1
_
f
1
(s
n
)
a
n
2,2
S
n
(t) a
n
2,1
T
n
(t)
a
n
1,1
a
n
2,2
a
n
1,2
a
n
2,1
+ f
2
(t
n
)
a
n
1,1
T
n
(t) a
n
1,2
S
n
(t)
a
n
1,1
a
n
2,2
a
n
1,2
a
n
2,1
_
+
+

n=1
_

f
1
(s
n
)
a
n
2,2

S
n
(t) a
n
2,1

T
n
(t)
a
n
1,1
a
n
2,2
a
n
1,2
a
n
2,1
+

f
2
(t
n
)
a
n
1,1

T
n
(t) a
n
1,2

S
n
(t)
a
n
1,1
a
n
2,2
a
n
1,2
a
n
2,1
_
.
The convergence of the series above is absolute and uniform in subsets of where K(t)
is bounded.
4.3 A comment on the choice of the samples
Theorem 2 allows us to combine several interpolation conditions whose types are not neces-
sarily equal. Corollary 4 gives a sampling formula for the example of subsection 1.1 which
shows some advantages of our approach. Indeed, for each n N, we can choose any row of
the matrix in (12). Thus, if the samples of f
2
are lost for n N N, we can still recover f
by using the second row of that matrix for n N.
On the other hand, xed n N, the rows of (12) are obtained as solutions of a consistent
linear system with three equations and two unknowns, so we can choose any two of them
in order to solve the system. This means that every row of (12) is equal to every other.
As a consequence, we can choose any element of the rst column and any element (not
necessarily in the same row) of the second one. This remark allows us to avoid cancellation
errors by choosing the appropriate elements of (12). For instance, suppose we have that
f
1
(n
0
)f
2
(n
0
) > 0 and that f
1
(n
0
)f
3
(n
0
) < 0 for n
0
N. Then, we can choose the third
element of the rst column,
1
2
[f
1
(n
0
) + f
2
(n
0
)], and the second element of the second one,
f
1
(n) f
3
(n).
Acknowledgments: This work has been supported by the grant BFM200301034 from
the D.G.I. of the Spanish Ministerio de Ciencia y Tecnologa.
References
[1] M. H. Annaby. On sampling theory associated with the resolvent of singular Sturm-
Liouville problems. Proc. Amer. Math. Soc., 131:18031812, 2003.
GARCIA-PORTAL
312
[2] N. Aronszajn. Theory of Reproducing Kernels. Trans. Amer. Math. Soc., 68:337404,
1950.
[3] O. Christensen. An Introduction to Frames and Riesz Bases. Birkh auser, Boston, 2003.
[4] W.N. Everitt and G. Nasri Roudsari. Interpolation and sampling theories, and linear
ordinary boundary value problems. In J.R. Higgins and R.L. Stens, editors, Sampling
Theory in Fourier and Signal Analysis, chapter 5, pages 5677. Oxford University Press,
Oxford, 1999.
[5] W.N. Everitt, G. Schottler, and P. L. Butzer. Sturm-Liouville boundary value problems
and Lagrange interpolation series. Rendiconti di Matematica, 14:87126, 1994.
[6] A. G. Garca and M. A. Hernandez-Medina. The discrete Kramer sampling theorem
and indeterminate moment problems. J. Comp. Appl. Math., 134:1322, 2001.
[7] A. G. Garca and M. A. Hern andez-Medina. On an integral transform associated with
the regular Dirac operator. Proc. Roy. Soc. Edin., 131:13571370, 2001.
[8] A. G. Garca and M. A. Hern andez-Medina. Discrete Sturm-Liouville problems, Jacobi
matrices and Lagrange interpolation series. J. Math. Anal. Appl., 280(2):221231, 2003.
[9] A. G. Garca and A. Portal. Sampling in the functional Hilbert space induced by a
Hilbert space valued kernel. Applicable Analysis, 82(12):11451158, 2003.
[10] J. R. Higgins. A sampling principle associated with Saitohs fundamental theory of
linear transformations. In S. Saitoh et al., editor, Analytic extension formulas and
their applications. Kluwer Academic, 2001.
[11] H. P. Kramer. A generalized sampling theorem. J. Math. Phys., 63:6872, 1957.
[12] S. Saitoh. Integral transforms, reproducing kernels and their applications. Longman,
Essex, England, 1997.
[13] P. Weiss. Sampling theorems associated with Sturm-Liouville systems. Bull. Amer.
Math. Soc., 63:242, 1957.
[14] R. M. Young. An Introduction to Nonharmonic Fourier Series. Academic Press, New
York, 1980.
[15] A. I. Zayed. Advances in Shannons Sampling Theory. CRC Press, Boca Raton FL,
1993.
[16] A. I. Zayed. Function and Generalized Function Transformations. CRC Press, Boca
Raton FL, 1996.
GENERAL SAMPLING THEORY
313