See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322319293

Propagation of Regularity and Positive Definiteness: a Constructive Approach

Article in Zeitschrift für Analysis und ihre Anwendungen · January 2018

DOI: 10.4171/ZAA/1599

CITATIONS READS

4 102

3 authors:

Jorge Buescu A. C. Paixão

University of Lisbon Instituto Politécnico de Lisboa
68 PUBLICATIONS 1,502 CITATIONS 17 PUBLICATIONS 165 CITATIONS

SEE PROFILE SEE PROFILE

Claudemir Pinheiro Oliveira

Universidade Federal de Itajubá (UNIFEI)
23 PUBLICATIONS 124 CITATIONS

SEE PROFILE

All content following this page was uploaded by Jorge Buescu on 28 March 2018.

The user has requested enhancement of the downloaded file.

 PROPAGATION OF REGULARITY AND POSITIVE
DEFINITENESS: A CONSTRUCTIVE APPROACH

J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

Abstract. We show that, for positive definite kernels, if specific forms

arXiv:1802.07092v1 [math.CV] 20 Feb 2018

of regularity (continuity, Sn -differentiability or holomorphy) hold locally

on the diagonal, then they must hold globally on the whole domain
of positive-definiteness. This local-to-global propagation of regularity
is constructively shown to be a consequence of the algebraic structure
induced by the non-negativity of the associated bilinear forms up to
order 5. Consequences of these results for topological groups and for
positive definite and exponentially convex functions are explored.

1. Positive definite kernels and functions

1.1. Introduction. The phenomenon of regularity propagation for positive
definite functions has recently attracted renewed attention. If an appro-
priate form of regularity (continuity, C 2n differentiability in the real case,
holomorphy in the complex case) holds in a neighborhood of the origin or the
imaginary axis, then positive definiteness effectively sweeps it along horizon-
tal lines, producing a local-to-global propagation of regularity to horizontal
tubes in the whole domain. Regularity propagation is known in different con-
texts: in Rn , where it follows directly [8] or by properties of Fourier trans-
forms [15]; for complex functions, for which it follows by Fourier-Laplace
transforms [5], [11]; for complexifications of finite-dimensional real vector
spaces [17]; and for involutive topological groups [29]. Proofs of regularity
propagation usually rely on an appropriate integral representation deter-
mined by Bochner’s theorem.
In this paper we show that regularity propagation holds in the much
more general setting of positive definite kernels, for which no analog of the
Bochner integral representation exists. Regularity propagation is thus a
deeper and more fundamental property than previously considered, since
it is a direct algebraic-analytic consequence of the definition of positive-
definiteness. More specifically, we show constructively that, if the appro-
priate kind of regularity (continuity, C 2n -type regularity for real-variable
kernels, holomorphy for complex-variable kernels) holds in a neighborhood

Date: 2016-10-19.
2010 Mathematics Subject Classification. Primary: 42A82. Secondary: 30A10, 30C40,
26B35.
Key words and phrases. Positive definite kernels, positive definite functions, differen-
tiability, holomorphy, constructive approximation, exponentially convex functions.
The first author acknowledges partial support by Fundação para a Ciência e Tecnologia,
UID/MAT/04561/2013. The third author was supported by CAPES -Brazil, grant 10884 -
13 -0. Part of this research was done while in a post doctoral program at the Universidade
de Lisboa.
1
2 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

of the diagonal, then the positive-definiteness condition “sweeps” the whole

domain, forcing regularity to propagate throughout it.
The key property for regularity propagation is the basic positive-defi-
niteness condition. Indeed, our results follow constructively from matrix
positive definiteness properties of the relevant bilinear forms. While it is
quite possible that more abstract representations such as those associated
with the concept of reproducing kernel Hilbert space (RKHS) as introduced
by Aronszajn [2] and Krein [18] may conceivably be used to provide an
alternative derivation of our results, our methods provide purely algebraic
and analytical proofs which are constructive from first principles and do not
rely on the use of those more sophisticated functional analytic methods. In
fact, it is possible that replicating these results relying solely on RKHS the-
ory may involve technical subtleties which may render fully rigorous proofs
quite involved. For instance, in §3.2 it becomes necessary to relate different
positive definite kernels in distinct Sn classes to study C 2n propagation. In
the RKHS context this would imply relating different RKHS consisting of
functions of different finite order of differentiability equipped with with un-
known norms, which seems a highly non-trivial task. In the same vein, the
RKHS approach would not be able to circumvent use of (some version of)
the constructive methods we develop in §3 to attack differentiability, leaving
the extra problem of showing, by density or more sophisticated functional
analytic arguments, the existence of the relevant derivatives in the RKHS.
Regularity propagation for positive definite functions is not to be confused
with the classical extension problem, which has a long and rich history
[19, 20, 22, 23, 24] but a quite different nature. In this problem one seeks to
extend a positive definite function beyond its domain of definition preserving
positive definiteness and, if possible, regularity. In regularity propagation
one starts with a positive definite function in a fixed domain and studies how
specific types of regularity extend from local to global sets as a consequence
of positive definiteness. There are however some connections between the
two problems, see remark 3.23.
The paper is structured as follows. In the remainder of §1 we set up the
relevant basic definitions. In §2 we provide necessary and sufficient condi-
tions for propagation of continuity from a neighborhood of the diagonal. As
an application we characterize a naturally generalized class of shift-invariant
kernels on topological groups. In §3 we characterize regularity propagation
for differentiable and holomorphic positive definite kernels. We conclude
with an application to positive definite functions by means of the associated
kernels.
1.2. Basic definitions and properties.
Definition 1.1. Let Ω be a nonempty set. A function k : Ω × Ω → C is
a positive definite kernel on Ω if, for every n ∈ N and all finite collections
{xj }nj=1 ⊂ Ω and {ξj }nj=1 ⊂ C,
n
X
k(xi , xj ) ξi ξj ≥ 0 (1.1)
i,j=1

that is, if every square matrix [k(xi , xj )]ni,j=1 is positive semidefinite.

 PROPAGATION OF REGULARITY 3

Consideration of order 1 and order 2 matrices easily leads to the following

basic properties of positive definite kernels.
(P1 ) k(x, x) = k(x, x) ≥ 0 ∀x ∈ Ω;
(P2 ) k(x, y) = k(y, x) ∀(x, y) ∈ Ω2 ;
(P3 ) |k(x, y)|2 ≤ k(x, x) k(y, y) ∀(x, y) ∈ Ω2 .
Given a nonempty set Ω, we define the diagonal of Ω2 as the set
D(Ω2 ) = {(x, x) : x ∈ Ω}.
Properties P1 –P3 illustrate the central rôle played by the values assumed
by k along the diagonal D(Ω2 ) on the global behaviour of the kernel k.
Consideration of higher order matrices in the appropriate topological and
differentiable contexts has been shown to imply a number of additional prop-
erties of positive definite kernels (see [6]) which further stress the importance
of the diagonal in the control of regularity.
In this paper we explore the consequences of the properties of 3rd , 4th and
5th order bilinear forms resulting from definition 1.1 for the propagation
of regularity of the kernel k from the diagonal D(Ω2 ) to Ω2 . Convenient
assumptions will be made about the nature of the set Ω in order to derive
conclusions for the continuous, C 2n -differentiable and holomorphic cases.
Natural connections exist between positive definite kernels and positive
definite functions, allowing the identification of the latter with a subclass
of the former. This fact is clear from the following definition and will be
used later in this paper to derive regularity propagation for positive definite
functions as a consequence of those obtained for positive definite kernels.
Definition 1.2. Let (G, +) be an Abelian group and suppose x → x∗ is a
map in G. Consider any nonempty subset Ω of G and define S = Ω + Ω∗ =
{z ∈ G : z = x + y ∗ , x, y ∈ Ω}. Then a function f : S → C is said to be
positive definite if k(x, y) ≡ f (x+y ∗ ) is a positive definite kernel k : Ω2 → C,
that is, if
Xn
f (xi + x∗j ) ξi ξj ≥ 0
i,j=1

for every n ∈ N and all finite collections {xj }nj=1 ⊂ Ω and {ξj }nj=1 ⊂ C.
This definition of a positive definite function provides a natural general-
ization of the notion of shift-invariant positive definite kernel on Ω. Given
a nonempty set Ω, it is convenient to define the diagonal of S = Ω + Ω∗ as
the set
DΩ (S) = {x + x∗ : x ∈ Ω}.
Remark 1.3. DΩ (S) has zero as its unique element in examples (i) through
(iii) below. However, example (iv) shows that this is not, in general, the
case.
We list a few concrete examples of this abstract setting where the map
x 7→ x∗ is an involution. Recall that, in a group (G, ◦), an involution
∗ : G → G is a map whose second power is the identity and which satisfies
(g1 ◦ g2 )∗ = g2∗ ◦ g1∗ (see e.g. [3]).
4 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

(i) Taking Ω = G and x∗ = −x, we have S = G − G = G. Then

f : G → C is a positive definite function if k(x, y) = f (x − y) is a
positive definite kernel on G.
(ii) In the special case where Ω = Rn and x∗ = −x, we obtain the
standard definition of positive definite function f : Rn → C as-
sociated with the positive definite kernel k : R2n → C defined by
k(x, y) = f (x − y).
(iii) If Ω ⊂ Rn and x∗ = −x, we have S = Ω − Ω. Then f : S → C is a
positive definite function if k(x, y) = f (x − y) is a positive definite
kernel on Ω.
(iv) Examples (i) to (iii) admit a natural counterpart where the ∗−map
is replaced with the identity x∗ = x, S = Ω + Ω∗ = Ω + Ω and
k(x, y) = f (x + y). In this special case we will refer to k as a co-
positive definite function [11]. Co-positive definite functions are also
known as exponentially convex functions in the context of example
(ii). In the real-variable context, these functions were introduced by
Bernstein in 1929 [4] and have been studied by Widder [27], Devinatz
[13], [14] and other authors.
(v) Suppose that Ω ⊂ Cn , x∗ = −x and S = {z = x − y : x, y ∈ Ω} =
codiff(Ω) (for the general definition of codifference sets see [9], [10]).
Then f : S → C is a positive definite function iff k(x, y) = f (x − y)
is a positive definite kernel on Ω.

2. Propagation of continuity for positive definite kernels

Suppose that Ω is a topological space and that k is a positive definite
kernel on Ω, where Ω2 is endowed with the product topology. For conve-
nience, we define ky : Ω → C by ky (x) = k(x, y) for each fixed y ∈ Ω and
kx (y) = k(x, y) for each fixed x ∈ Ω.
The following lemmas will be used in the proof of this and next section’s
main results.
Lemma 2.1. Suppose that K : Ω2 → C is a separately continuous function
and that, for each x0 ∈ Ω, there exists a neighborhood V (x0 ) such that
{Kx : x ∈ V (x0 )} is equicontinuous in Ω. Then K is continuous in Ω2 .
Proof. Fix (x0 , y0 ) in Ω2 . For any (x, y) ∈ Ω2 we have
|K(x, y) − K(x0 , y0 )|≤ |K(x, y) − K(x, y0 )|+|K(x, y0 ) − K(x0 , y0 )|.
Observe that, for any fixed y0 , Ky0 (x) is by hypothesis continuous in Ω and,
in particular, at x = x0 . Hence, for every δ > 0 there exists a neighborhood
V δ (x0 ) such that
|Ky0 (x) − Ky0 (x0 )|= |K(x, y0 ) − K(x0 , y0 )|< δ/2
whenever x ∈ V δ (x0 ). Now, by hypothesis, if x ∈ V (x0 ) there exists a
neighborhood W (y0 ), independent of x, such that
|Kx (y) − Kx (y0 )|= |K(x, y) − K(x, y0 )|< δ/2
for every y ∈ W (y0 ).
Then, if (x, y) ∈ U (x0 , y0 ) = V (x0 ) ∩ V δ (x0 ) × W (y0 ), it follows that

|K(x, y) − K(x0 , y0 )|< δ. This finishes the proof. 

 PROPAGATION OF REGULARITY 5

Lemma 2.2. Let Ω be a set and suppose k : Ω2 → C is a positive definite

kernel. Then for every x1 , x2 , x3 in Ω, we have
 
2 k(x2 , x2 ) + k(x3 , x3 )
|k(x1 , x2 ) − k(x1 , x3 )| ≤ k(x1 , x1 ) 2 ℜ − k(x2 , x3 ) .
2
(2.1)
Proof. Let {x1 , x2 , x3 } ⊂ Ω. Since k is a positive definite kernel, we have
3
X
k(xi , xj ) ξi ξj ≥ 0 (2.2)
i,j=1

for any {ξj }3j=1

⊂ C. Choosing ξ1 = η, ξ2 = 1 and ξ3 = −1 and using
properties P1 and P2 , condition (2.2) may be written in the form
−k(x1 , x1 )|η|2 − 2ℜ(η(k(x1 , x2 ) − k(x1 , x3 ))
(2.3)
 
k(x2 , x2 ) + k(x3 , x3 )
≤ 2ℜ − k(x2 , x3 )
2
for every η ∈ C.
If k(x1 , x1 ) = 0 then (2.1) holds trivially since, in this case, it follows from
property P3 that the left hand side of (2.3) is zero.
−(k(x1 , x2 ) − k(x1 , x3 ))
Suppose then that k(x1 , x1 ) 6= 0. Choosing η = ,
k(x1 , x1 )
we obtain from (2.3)

|k(x1 , x2 ) − k(x1 , x3 )|2 |k(x1 , x2 ) − k(x1 , x3 )|2

− +2
k(x1 , x1 ) k(x1 , x1 )
 
k(x2 , x2 ) + k(x3 , x3 )
≤ 2ℜ − k(x2 , x3 ) (2.4)
2
which implies (2.1). 
Theorem 2.3. Let Ω be a topological space. Suppose that k : → C is a Ω2
positive definite kernel whose real part ℜ(k) is continuous on the diagonal
D(Ω2 ). Then k is continuous in Ω2 .
Proof. Let {x1 , x2 , x3 } ⊂ Ω. Then acording to lemma 2.2, we have
 
2 k(x2 , x2 ) + k(x3 , x3 )
|k(x1 , x2 ) − k(x1 , x3 )| ≤ k(x1 , x1 ) 2 ℜ − k(x2 , x3 ) .
2
Taking limits when x3 → x2 and using continuity of ℜ(k) on the diagonal
we conclude that k is separately continuous on the second variable. From
property P2 we then derive that k is separately continuous in Ω2 . Since,
according to the hypothesis, the real part of k is continuous on D(Ω2 ), it
follows that k(x, x) is bounded on some neighborhood V (x0 ) of any x0 ∈ Ω.
Then there exists M > 0 such that k(x1 , x1 ) = |ℜk(x1 , x1 )|≤ M for every
x1 ∈ V (x0 ) and we derive from (2.1) that
 
2 k(x2 , x2 ) + k(x3 , x3 )
|k(x1 , x2 ) − k(x1 , x3 )| ≤ 2 M ℜ − k(x2 , x3 ) .
2
6 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

Since the right hand side of the equation does not depend on x1 we con-
clude, by taking limits when x3 → x2 and using continuity of ℜ(k) on the
diagonal, that {kx : x ∈ V (x0 )} is equicontinuous in Ω. Since this is true
for any x0 ∈ Ω, we conclude by lemma 2.1 that k is continuous in Ω2 , as
stated. 

Remark 2.4. Although a different proof of this result may be derived in the
context of the theory of reproducing kernels, our direct approach has the
advantage of clarifying the precise analytical rôle of each hypothesis. In fact,
this result is frequently stated and proved, in the RKHS approach, by re-
quiring cumulatively separate continuity of the kernel in both variables and
continuity on the diagonal (see e.g. Schwartz [25], Prop. 24, or Steinwart
and Christmann [26], Lemma 4.29). Our proof shows that the hypothesis
of separate continuity is unnecessary, since it follows from continuity on the
diagonal and joint continuity implied by continuity on the diagonal alone.
The fact that continuity on the diagonal is, in this context, the more fun-
damental property may be judged from a result by Lehto [21] showing that
separately continuous, bounded positive definite kernels are not necessarily
continuous. It is worth noting, however, that in the original paper of Krein
[18] on RKHS this result is stated in a rigorously equivalent form to theorem
2.3.

2.1. Application to shift-invariant kernels on groups. A first and

straightforward consequence of definition 1.2 and theorem 2.3 may be stated
as follows.

Corollary 2.5. Suppose (G, +) is a topological group and that x → x∗ is

a continuous map. Let Ω be an open subset of G, S = Ω + Ω∗ and suppose
f : S → C is a positive definite function. Then f is continuous in S if and
only if ℜ(f ) is continuous in DΩ (S).

Proof. Since the “only if” part of the assertion is trivial, we concentrate on
the sufficiency assertion. Define the positive definite kernel k : Ω2 → C by
k(x, y) = f (x + y ∗ ). Observe that ℜ(k(x, y)) = ℜ(f (x + y ∗ )) is continuous in
D(Ω2 ) since ℜ(f ) is continuous in DΩ (S) by hypothesis. Since k is a positive
definite kernel, we conclude by theorem 2.3 that k is continuous in Ω2 .
For any fixed y ∈ Ω, define Ωy = Ω + {y ∗ } and ly : Ωy → Ω by ly (z) =
z − y ∗ . Notice that Ωy is an open set and that ly is continuous for the
topologies induced by G on Ω and Ωy . Now observe that k(z − y ∗ , y) =
ky ◦ ly (z) = f|Ωy (z) is continuous in Ωy since k is continuous in Ω2 . Notice,
furthermore, thatS since Ωy is open, any open set of Ωy is an open set of
S = Ω + Ω∗ = y∈Ω Ωy . For any open set U ⊂ C, this imples that f −1 (U ) =
S −1
y∈Ω f|Ωy (U ) is an open set of S, showing that f is continuous in S and
concluding the proof. 
The next result follows from direct application of corollary 2.5 to example
(iii) in §1 and is a standard result, readily found in the literature, for the
case where Ω = Rn .

Corollary 2.6. Suppose Ω is an open subset of Rn , S = Ω − Ω, and let

f : S → C be a positive definite function in the usual Rn variable sense
 PROPAGATION OF REGULARITY 7

(that is, taking x∗ = −x). Then f is continuous in S if and only if ℜ(f ) is

continuous at the origin.
Proof. The result follows immediately from corollary 2.5 by observing that,
in this case, DΩ (S) = {0}. 
In a similar way, direct application of corollary 2.5 the case of example
(iv) leads to the following result.
Corollary 2.7. Suppose Ω is an open subset of Cn , S = {z = x − y, x, y ∈
Ω} ≡ codiff(Ω), and let f : S → C be a positive definite function on S in the
usual Cn variable sense (i.e. with x∗ = −x). Then f is continuous in S if
and only if ℜ(f ) is continuous on DΩ (S).

3. Propagation of differentiability for positive definite

kernels and functions
We now focus on differentiability properties of positive definite kernels
of two real or complex variables. Results will be derived in a constructive
way from direct consideration of 4th and 5th order square matrices associ-
ated with the corresponding bilinear forms on definition 1.1 under suitable
hypotheses on the regularity of the kernel on the diagonal. The main results
will be established in §3.2. Consequences for real and complex variable pos-
itive definite functions are explored in §3.3. Notation and terminology will
be fixed in the next section along with some essential but not necessarily
original results. Proofs are nevertheless provided when they shed light on
the methods used in other sections.

3.1. Differentiation of one or two real or complex variable func-

tions. A real or complex variable function f : D → C is said to be differ-
entiable at a point a ∈ int(D) if its derivative

df f (a + h) − f (a)
(a) = lim (3.1)
dx h→0 h
is a complex number.
In the complex variable case, differentiability of f at every point of an
open set Ω imples analiticity in Ω and is usually referred to as holomorphy
in this set (see [6] for further details). The next definition is also very useful
in this context. Given a set U , we denote by U ∗ its complex conjugate, that
is, U ∗ = {z : z ∈ U }.
Definition 3.1. Let D ⊂ C. A function f : D → C is said to be anti-
holomorphic on an open subset U ⊂ D if there exists a holomorphic function
g : U ∗ → C such that f (x) = g(x) for all x ∈ U . We define, for every a ∈ U ,

df dg
= . (3.2)
dx |x=a dx |x=a
As a consequence, the corresponding higher-order differential operators
m dm g
are given by ddxmf |x=a = dx m |x=a .

A straightforward consequence of definition 3.1 is that, for a ∈ int(D),

8 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

df dg g(a + h) − g(a) f (a + h) − f (a)

(a) = (a) = lim = lim . (3.3)
dx dx h→0 h h→0 h
The following characterization of the existence of the derivatives defined
in (3.1) and (3.3) will be essential for what follows.
Lemma 3.2. (i) Suppose f : D → C is a real or complex variable
df
function and a ∈ int(D). Then dx (a) = limh→0 f (a+h)−f
h
(a)
is a
complex number if and only if

 
f (a + h) 1 1 f (a + l)
lim + f (a) − − = 0.
(h,l)→(0,0) h l h l
(ii) Suppose f : D → C is a complex variable function and a ∈ int(D).
Then dxdf
(a) = limh→0 f (a+h)−f
h
(a)
is a complex number if and only if
 
f (a + h) 1 1 f (a + l)
lim + f (a) − − = 0.
(h,l)→(0,0) h l h l
Proof. Since the “only if” part of the assertions is trivial, we focus on the
sufficiency of the conditions given in (i) and (ii). Let G(h) = f (a+h)−f h
(a)

in the first case and G(h) = f (a+h)−f h

(a)
in the second. In either case write
P(h, l) = G(h) − G(l). Both statements will be simultaneously proved if
we show that if lim(h,l)→(0,0) P(h, l) = 0, then limh→0 G(h) is necessarily a
complex number.
We begin by showing that G must be bounded on some (real or complex,
according to the relevant case) neighborhood of zero. In order to establish a
contradiction, suppose that this is not the case. Then we can find a sequence
ln convergent to zero such that |G(ln+1 )−G(ln )|> 1 for every n ∈ N. Defining
hn = ln+1 , we have that (hn , ln ) → (0, 0) (in R2 or C2 , according to the case)
and that |P(hn , ln )|= |G(hn ) − G(ln )|= |G(ln+1 ) − G(ln )|> 1 for every n ∈ N.
Hence lim(h,l)→(0,0) P(h, l) cannot be zero, contradicting the hypothesis.
It therefore follows that G is bounded in some neighborhood of zero.
Then, for any sequence ln → 0, we conclude that G(ln ) is bounded in C
and therefore admits a subsequence G(lbn ) converging to α ∈ C. Since
(ln , lbn ) → (0, 0) (respectively in R2 or C2 ) and lim(h,l)→(0,0) P(h, l) = 0 by
hypothesis, it follows that G(ln ) − G(lbn ) → 0 as n → ∞. Writing G(ln ) =
G(ln ) − G(lbn ) + G(lbn ), it follows that limn→∞ G(ln ) = limn→∞ G(lbn ) = α.
Arbitrariness of ln implies that limh→0 G(h) = α ∈ C, finishing the proof. 
We now turn our attention to functions of two real or complex variables.
We first concentrate on this last case, where particularly significant phenom-
ena take place. It is convenient to define, given U ⊂ C2 , the sesquiconjugate
set U ∗ = {(x, y) ∈ C2 : (x, y) ∈ U }.
Definition 3.3. Let U ⊂ C2 be an open set and k : U → C. We say that
k(u, v) is sesquiholomorphic in U if there exists a separately holomorphic
function g : U ∗ → C such that k(u, v) = g(u, v) for all (u, v) ∈ U , i.e., if
k(u, v) is separately holomorphic in u and anti-holomorphic in v.
 PROPAGATION OF REGULARITY 9

Proposition 3.4. Let U ⊂ C2 be an open set and suppose k(u, v) is sesqui-

holomorphic in U . Then, for all m1 , m2 ∈ N, k has continuous partial
m +m
derivatives ∂v∂ m21 ∂um2 1 k(u, v) of all orders with respect to the variables u and
v and the order of differentiation is immaterial.
A proof of this fact may be found, after minor adaptations, in [6], prop.
1.4.
The following lemma will be relevant for the use of finite increments in
the context of differentiation. It is convenient to define
∆h,l k(u, v) = k(u + h, v + l) − k(u, v + l) − k(u + h, v) + k(u, v) (3.4)
whenever the right-hand side is defined for u, v, h, l ∈ C.
Lemma 3.5. Let U ⊂ C2 be an open set and suppose k(u, v) is sesquiholo-
morphic in U . Then, for (u, v) ∈ U ,
∂2 ∂2 ∆h,l k(u, v)
k(u, v) = k(u, v) = lim . (3.5)
∂u∂v ∂v∂u (h,l)→(0,0) hl
Proof. We will show that, for separately holomorphic g(u, v), the identity
∂2 ∂2 ∆h,l g(u, v)
g(u, v) = g(u, v) = lim (3.6)
∂u∂v ∂v∂u (h,l)→(0,0) hl
holds. This will imply that, for g such that k(u, v) = g(u, v), we have, by
virtue of (3.2) and (3.3),
∂2 ∂2
k(u, v) = g(u, v)
∂u∂v ∂u∂v
and
∂2 ∂2 ∆h,l g(u, v) ∆h,l k(u, v)
k(u, v) = g(u, v) = lim = lim
∂v∂u ∂v∂u (h,l)→(0,0) hl (h,l)→(0,0) hl
implying the statement of the lemma.
In order to establish identity (3.6), it is convenient to recall the following
version of the finite increment formula for holomorphic functions (see e.g.
[1], pg. 125). Suppose f is analytic on an open set U ⊂ C, D is a topological
disc whose closure is contained in U , C = ∂D, z and h are complex numbers
such that z and z + h are in D. Then
f (z + h) − f (z) 1 f (ζ)
Z
= dζ. (3.7)
h 2πi C (ζ − z)(ζ − z − h)
Suppose g : U ∗ → C is separately holomorphic and that (u, v) ∈ U ∗ . Let
D1 (resp. D2 ) be a disc centered at U (resp. v such that D1 × D2 ⊂ U ∗ ,
C1 = ∂D1 , C2 = ∂D2 . Successive applications of formula (3.7) yield

∆h,l g(u, v)
hl
1 g(ζ1 , ζ2 )
Z Z
= dζ1 dζ2 .
(2πi)2 C2 C1 (ζ1 − u)(ζ1 − u − h)(ζ2 − v)(ζ2 − v − l)
(3.8)
Continuity of g and the integral representation (3.8) imply that
10 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

∆h,l g(u, v) 1 g(ζ1 , ζ2 )

Z Z
lim = dζ1 dζ2 =
(h,l)→(0,0) hl (2πi)2 C2 C1 (ζ1 − u)2 (ζ2 − v)2
(3.9)
∂2 ∂2
= g(u, v) = g(u, v),
∂v∂u ∂u∂v
concluding the proof. 
It will be essential for what follows to define the following differentiability
class; for further details see e.g. [7].
Definition 3.6. Let U ⊂ R2 be an open set. A function k : U → C is said
to be of class Sn (U ) if, for every m1 = 0, 1, . . . , n and m2 = 0, 1, . . . , n, the
∂ m1 +m2
partial derivatives k(u, v) are continuous in U .
∂v m2 ∂um1
The special significance of this class of kernels in the present context is
due to the following result.
Theorem 3.7. Let Ω ⊆ R be an open set and k(x, y) be a positive definite
kernel of class Sn (Ω2 ). Then, for all 0 ≤ m ≤ n,
∂ 2m
km (x, y) ≡ k(x, y)
∂y m ∂xm
is a positive definite kernel of class Sn−m (Ω2 ).
For a proof in the case where Ω is an interval, see theorem 3.6 in [7];
extension to general open sets is immediate. Generalizations to several real
and complex variables exist (see e.g. [6]) but will not be necessary in this
paper.
We are now ready to construct a version of lemma 3.5 for functions of
two real variables using definition (3.4) in the appropriate way.
Lemma 3.8. Let U ⊂ R2 be an open set and suppose that k : U → C is of
class S1 (U ). Then, for (u, v) ∈ U ,
∂2 ∂2 ∆h,l k(u, v)
k(u, v) = k(u, v) = lim . (3.10)
∂u∂v ∂v∂u (h,l)→(0,0) hl
Proof. Suppose that I1 and I2 are open intervals centered at u and v, re-
spectively, such that I1 × I2 ⊂ U . For any h, l such that u + h ∈ I1 and
v + l ∈ I1 define
φ1 (t) = k(u + th, v + l) − k(u + th, v),
observe that φ1 is differentiable in an open interval containing [0, 1] and that
∆h,l k(u, v) = φ1 (1) − φ1 (0). By the mean value theorem for functions of one
real variable there exists t1 ∈ ]0, 1[ such that
 
′ ∂k ∂k
φ1 (1) − φ1 (0) = φ1 (t1 ) = (u + t1 h, v + l) − (u + t1 h, v) h.
∂u ∂u
A second application of the mean value theorem to the function
∂k
φ2 (t) = (u + t1 h, v + tl)
∂u
 PROPAGATION OF REGULARITY 11

yields
∂2k
∆h,l k(u, v) = (u + t1 h, v + t2 l) h l
∂u∂v
∂2k
for some t2 ∈]0, 1[. Continuity of ∂v∂u then implies that
∆h,l k(u, v) ∂2k
lim = (u, v).
(h,l)→(0,0) hl ∂v∂u
2
∂ k
The corresponding equality for the symmetric mixed derivative ∂u∂v is es-
tablished in a completely analogous way. 
For the purposes of the next section, it will be convenient to establish the
following consequences from lemmas 3.5 and 3.8.
Corollary 3.9. In the conditions of lemma 3.5 (resp. lemma 3.8) define
 
∆h,l k(u, v) ∆h,l k(u, v)
ψ(h, l) = resp.
hl hl
and
φ(h, l) = ψ(h, h) + ψ(l, l) − ψ(h, l) − ψ(l, h).
Then
lim φ(h, l) = 0.
(h,l)→(0,0)

Proof. Observe that, under the conditions of lemma 3.5 (resp. 3.8), we have

∂2
k(u, v) = lim ψ(h, l) = lim ψ(l, h)
∂v∂u (h,l)→(0,0) (h,l)→(0,0) (3.11)
= lim ψ(h, h) = lim ψ(l, l).
(h,l)→(0,0) (h,l)→(0,0)


Remark 3.10. Notice that, if u = v, property P2 yields ψ(h, l) = ψ(l, h),
implying, in this case, that φ(h, l) is a real-valued function.
3.2. Propagation of differentiability for positive definite kernels.
We first consider the case of complex variable positive definite kernels.
Let Ω ⊂ C be an open set and k : Ω2 → C a positive definite kernel
on Ω. We say that k is a holomorphic positive definite kernel in Ω if k is
sesquiholomorphic on Ω2 .
Remark 3.11. Holomorphic positive definite kernels in Ω, also known as
holomorphic reproducing kernels, are not defined as holomorphic functions
on Ω2 . In fact, imposing separate holomorphy on these kernels would lead
to a trivialization of the concept; see [6] for further details.
Remark 3.12. As previously observed, the theory of reproducing kernels may
provide an alternative approach to the issues under discussion. For instance,
results have been stated in Krein [18] for holomorphic functions that may be
adapted to address the much more significant sesquiholomorphic case. Our
proof of the following results does not rely on any of these functional-analytic
constructions.
We may now state our main theorem for the complex context.
12 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

Theorem 3.13 (Propagation of regularity, complex context). Suppose Ω is

an open subset of C and let k : Ω2 → C be a positive definite kernel. If k is
sesquiholomorphic on an open set U containing the diagonal D(Ω2 ), then k
is a holomorphic reproducing kernel in Ω.
Proof. We will show that k is separately holomorphic in the first variable and
anti-holomorphic in the second variable for (u, v) ⊂ Ω2 . For this purpose it
∂k
will be sufficient to see that exists for (u0 , v0 ) ∈ Ω2 since, using property
∂v
P2 , definition 3.1 and formula (3.3), we have:

 
∂k k(u0 + h, v0 ) − k(u0 , v0 ) k(v0 , u0 + h) − k(v0 , u0 )
(u0 , v0 ) = lim = lim
∂u h→0 h h→0 h
∂k
=(v0 , u0 ).
∂v
Now consider an arbitrary collection {xj }4j=1 ⊂ Ω and the corresponding
bilinear form in definition 1.1. Since k is a positive definite kernel, it follows
that, for any collection {ξj }4j=1 ⊂ C,
4
X
k(xi , xj ) ξi ξj ≥ 0. (3.12)
i,j=1

For (u, v) ∈ Ω2 , fix some open disc D ⊂ C with center v such that
D × D ⊂ U and {u} × D ⊂ Ω2 . For any nonzero h, l ∈ C such that v + h
and v + l are in D, define
x1 = u, x2 = v + h, x3 = v, x4 = v + l.
Choosing for the ξi the values
1 1 1 1
ξ1 = η, ξ2 = , ξ3 = − , ξ4 = − , η ∈ C,
h l h l
and observing that by property P2 we may write
k(xi , xj ) = k(xj , xi ), i, j = 1, . . . , 4 (3.13)
we derive from (3.12) that
k(u, u) |η|2 +2ℜ[η β0 (h, l)] + γ(h, l) ≥ 0, (3.14)
where we have defined
 
k(u, v + h) 1 1 k(u, v + l)
β0 (h, l) = + k(u, v) − − (3.15)
h l h l
and
k(v + h, v + h) 1 1 2 k(v + l, v + l)
γ(h, l) = + k(v, v) − +
|h|2 l h |l|2
    
1 1 1 1 1
+ 2 ℜ k(v + h, v) − + k(v + h, v + l) − (3.16)
h l h h l
   
1 1 1
+k(v, v + l) − − .
l h l
 PROPAGATION OF REGULARITY 13

Next we show that the inequality

|β0 (h, l)|2 ≤ k(u, u) γ(h, l). (3.17)
is a consequence of (3.14). Indeed, if k(u, u) = 0, this is a trivial consequence
of the fact that, by property P3 , β0 (h, l) = 0. Suppose that k(u, u) 6= 0.
1
Then (3.17) follows from (3.14) by choosing η = − k(u,u) β0 (h, l).
Finally observe that we only need to show that lim(h,l)→(0,0) γ(h, l) = 0
in order to finish the proof. Indeed, according to (3.17), the aforementioned
condition implies that lim(h,l)→(0,0) β0 (h, l) = 0. This, in turn, will lead to
∂k
the conclusion that (u, v) is a complex number, as stated in the theorem
∂v
by direct application of lemma 3.2 (ii) with the appropriate identification of
k(u, v) with f (v).
It will prove convenient to rewrite (3.16) in the following form:

−hl − lh + |h|2 +|l|2 k(v + l, v + l)


k(v + h, v + h)
γ(h, l) =ℜ + k(v, v) +
|h|2 |h|2 |l|2 |l|2
 
1 1 1
+2k(v + h, v) − 2 − 2k(v + h, v + l)
hl |h| hl
 
1 1
+2k(v, v + l) − 2 .
h l |l|
(3.18)
Now recall the contents of corollary 3.9 and set
∆h,l k(v, v)
ψ(h, l) = ,
hl
φ(h, l) = ψ(h, h) + ψ(l, l) − ψ(h, l) − ψ(l, h).
Then the conclusions of the corollary together with the fact that k is, by
hypothesis, sesquiholomorphic on an open set U containing (v, v), imply that
lim φ(h, l) = 0. (3.19)
(h,l)→(0,0)

A straightforward calculation now reveals, through the use of formulas

(3.13) and direct comparison with (3.18), the following identification:

γ(h, l) = φ(h, l) (3.20)

(observe that this expression is real by remark 3.10). From (3.19) and (3.20)
we finally derive that lim(h,l)→(0,0) γ(h, l) = 0, which finishes the proof. 
We now focus on establishing the real variable counterpart of theorem
3.13. The constructive arguments used in the proof of this theorem will be
adapted in order to prove the following lemmas.
Lemma 3.14. Suppose Ω is an open subset of R and let k : Ω2 → C be
a positive definite kernel. If k is of class S1 (U ) for some open subset U
containing the diagonal D(Ω2 ), then k is separately differentiable in Ω2 .
14 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

Proof. The proof follows very closely that of theorem 3.13. Since by property
∂k ∂k
P2 and definition 3.1 we have that (u0 , v0 ) = (v0 , u0 ) it is sufficient to
∂u ∂v
∂k
show that (u0 , v0 ) exists in C for all (u0 , v0 ) ∈ Ω2 .
∂v
For (u, v) ∈ Ω2 we fix some open interval I centered at v such that
I × I ⊂ U and {u} × I ⊂ Ω2 and consider nonzero h, l ∈ R such that v + h
and v + l belong to I. Since k is a positive definite kernel, we may write
definition 1.1 as (3.12) with the following choices:

x1 = u, x2 = v + h, x3 = v, x4 = v + l
and
1 1 1 1
ξ1 = η, ξ2 = , ξ3 = − , ξ4 = − , η ∈ C.
h l h l
Now formulas (3.14) through (3.18) follow in the exact same way as in
the complex variable case, with the obvious particularity that now h =
h and l = l. The arguments leading to the conclusion (3.17) also carry
through. Thus, if γ(h, l) may be proved to have zero limit at the origin,
direct application of lemma 3.2 (i) with the identification k(u, v) = f (v)
yields the conclusions of the theorem as a consequence of the fact that, in
that case, lim(h,l)→(0,0) β0 (h, l) = 0. To show that lim(h,l)→(0,0) γ(h, l) = 0
we recall once again the definitions of ψ and φ in corollary 3.9 and set
∆h,l k(v, v)
ψ(h, l) =
hl
in this case. Since k is, by hypothesis, of class S1 (U ) on an open set U
containing (v, v), the conclusion of the corollary will yield formula (3.19)
as a consequence. The proof is finished by establishing formula (3.20) and
using it together with (3.19) to conclude that lim(h,l)→(0,0) γ(h, l) = 0. 
∂k ∂k
Lemma 3.15. In the conditions of lemma 3.14, (u, v) (resp. (u, v))
∂v ∂u
is continuous in v ∈ Ω (resp. is continuous in u ∈ Ω) for every fixed value
of u ∈ Ω (for every fixed value of v ∈ Ω). Furthermore, for any (u0 , v0 ) ∈
Ω2 there exists a neighborhood U (u0 ) (resp. V (v0 )) such that the family
{ ∂k ∂k
∂v (u, v), u ∈ U (u0 )} of fixed-u functions of v (resp. { ∂u (u, v), v ∈ V (v0 )}
of fixed-v functions of u) is equicontinuous in Ω.
Proof. As in the proof of lemma 3.14, observe that by property P2 we have
∂k ∂k
that (u0 , v0 ) = (v0 , u0 ) for all (u0 , v0 ) ∈ Ω2 and therefore it suffices to
∂u ∂v
∂k
prove the assertions for .
∂v
2
For (u, v) ∈ Ω we fix some open interval I centered at v such that
I × I ⊂ U and {u} × I ⊂ Ω2 and consider nonzero h, l ∈ R such that v + h
and v + h + l belong to I. Since k is a positive definite kernel, we may write
definition 1.1 as (3.12) with the following choices:

x1 = u, x2 = v, x3 = v + h, x4 = v + h + l
 PROPAGATION OF REGULARITY 15

and
1 1 1 1
, ξ3 = − , ξ4 = − .
ξ1 = η, ξ2 =
h l h l
We formally rewrite inequality (3.14) as
k(u, u) |η|2 +2ℜ[η β 0 (h, l)] + γ 0 (h, l) ≥ 0, (3.21)
with the new definitions

k(u, v + h) − k(u, v) k(u, v + h + l) − k(u, v + h)

β 0 (h, l) = − (3.22)
h l
and

k(v, v) 1 1 2 k(v + h + l, v + h + l)
γ 0 (h, l) = + k(v + h, v + h) + +
|h|2 l h |l|2
   
1 1 1 1
+ 2 ℜ k(v, v + h) − + + k(v, v + h + l)
h l h hl
  
1 1 1
+k(v + h, v + h + l) − + .
l h l
(3.23)
The formula corresponding to (3.17) is obtained in the exact same way,
leading to
|β 0 (h, l)|2 ≤ k(u, u) γ 0 (h, l). (3.24)
Now, defining
∆h,l k(v, v + h)
ψ 0 (h, l) = (3.25)
hl
and
φ0 (h, l) = ψ 0 (h, h) + ψ 0 (l, l) − ψ 0 (h, l) − ψ 0 (l, h) (3.26)
the identification
γ 0 (h, l) = φ0 (h, l) (3.27)
may be established by direct comparison of (3.26) and (3.23). On the other
hand, by adapting the procedures of the proof of lemma 3.8 it is possible to
derive the identity
∂2
k(v, v) = lim ψ 0 (h, l) (3.28)
∂u∂v (h,l)→(0,0)

from the hypothesis that k is of class S1 (U ) on an open set containing (v, v).
From (3.28), (3.26) and (3.27) we may now conclude that
lim γ 0 (h, l) = lim φ0 (h, l) = 0
(h,l)→(0,0) (h,l)→(0,0)

and also that, according to (3.24),

lim β 0 (h, l) = 0. (3.29)
(h,l)→(0,0)

Finally, observe that condition (3.29) implies

 
0
lim lim β (h, l) = 0. (3.30)
h→0 l→0
16 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

k(u, v + h) − k(u, v) ∂k
Since liml→0 β 0 (h, l) = − (u, v + h), it follows from
h ∂v
(3.30) that
∂k ∂k
(u, v) = lim (u, v + h),
∂v h→0 ∂v
∂k
showing that (u, v) is continuous on the second variable v ∈ Ω for any
∂v
fixed u ∈ Ω, as asserted.
To finish the proof we finally observe that, for any u0 ∈ Ω, continuity of
k(u, u) implies that there exists a neighborhood U (u0 ) such that k(u, u) ≤ M
for some positive M and all u ∈ U (u0 ). Hence we may rewrite formula (3.24)
as
|β 0 (h, l)|2 ≤ M γ 0 (h, l). (3.31)
for u ∈ U (u0 ). Since the right hand side of (3.31) does not depend on u, we
conclude from (3.30) that the family of fixed-u functions of v { ∂k
∂v (u, v), u ∈
U (u0 )} is equicontinuous in Ω. 
The two following results will be essential in the proof of lemma 3.18
below. The first is a corollary of lemma 3.2.
Corollary 3.16. Let Ω be an open subset of R and k : Ω2 → C be a
∂k
complex function. Suppose that (u, v) exists for every (u, v) ∈ Ω2 and
∂u
write, whenever meaningful for λ, h, l ∈ R,
 
k(u + λ, v + h) 1 1 k(u + λ, v + l)
βλ (h, l) = + k(u + λ, v) − − .
h l h l
∂2k
Then (u, v) is a complex number if and only if
∂v∂u
βλ (h, l) − β0 (h, l)
lim lim = 0.
(h,l)→(0,0) λ→0 λ
∂k
Proof. Applying lemma 3.2 to (u, v) as a function of v, we have that
∂u
∂2k
(u, v) is a complex number if and only if
∂v∂u
∂k ∂k
∂u (u, v + h) ∂u (u, v + l)
 
∂k 1 1
lim + (u, v) − − = 0. (3.32)
(h,l)→(0,0) h ∂u l h l
∂k k(u + λ, v) − k(u, v)
Writing (u, v) = limλ→0 and using the definition of
∂u λ
βλ , we conclude that condition (3.32) may be written in the form
βλ (u, v) − β0 (u, v)
lim lim = 0,
(h,l)→(0,0) λ→0 λ
as asserted. 
Proposition 3.17. Let T be a square matrix of order r1 + r2 partitioned in
the block form
 
A B
T =
D C
 PROPAGATION OF REGULARITY 17

where A = [aij ], B = [biq ], C = [cpq ], D = [dpj ] with i, j = 1 . . . , r1 and

p, q = 1, . . . , r2 and let z = (z1 , . . . , zr1 ) ∈ Cr1 , w = (w1 , . . . , wr2 ) ∈ Cr2 .
Then, if T is positive semidefinite, we have
|z T B w|2 ≤ (z T A z) (wT C w).
References and a proof of proposition 3.17 may be found in [16] or [6].
Lemma 3.18. Suppose Ω is an open subset of R and let k : Ω2 → C be a
positive definite kernel. If k is of class S1 (U ) for some open subset U con-
taining the diagonal D(Ω2 ), then the second order mixed partial derivatives
∂2k ∂2k 2
∂v∂u (u, v), ∂u∂v (u, v) exist for all (u, v) ∈ Ω .
∂ k 2
Proof. We will prove the result only for the mixed derivative ∂v∂u (u, v); the
corresponding result for the mixed partial in the reverse order then follows
immediately either by an analogous argument or simply by invoking the
Hermitian property P2 .
For (u, v) ∈ Ω2 we fix some open intervals Iu and Iv centered at u and
v respectively, such that Iu × Iv ⊂ U and Iu × Iv ⊂ Ω2 . Consider nonzero
λ, h, l such that u + λ ∈ Iu and v + l, v + λ ∈ Iv and observe that, since k
is a positive definite kernel, we may write definition 1.1 for n = 5 with the
following choices:
x1 = u, x2 = u + λ, x3 = v + h, x4 = v, x5 = v + l
and
1 1 1 1 1 1
ξ1 = − , ξ2 = , ξ3 = , ξ4 = − , ξ5 = − .
λ λ h l h l
Fix r1 = 2, r2 = 3 and consider the order 5 square matrix [k(xi , xj )]i,j=1,...5
partitioned in the block form given in proposition 3.17. This will lead to the
following identifications for i, j = 1, 2 and p, q = 1, 2, 3:
aij = k(xi , xj ), biq = k(xi , xq+2 ), dpj = k(xp+2 , xj ), cpq = k(xp+2 , xq+2 ),

zi = ξi , wp = ξp+2 .
We now recall formula (3.4) and the definitions of βλ and γ in corollary 3.16
and lemma 3.14, respectively, to obtain
∆λλ k(u, u)
zT A z = ,
λ2
βλ (h, l) − β0 (h, l)
zT B w = ,
λ
wT C w = γ(h, l).
Then, according to the conclusions of proposition 3.17, we have:
2
βλ (h, l) − β0 (h, l) ∆λλ k(u, u)
≤ γ(h, l)
λ λ2
Applying limits to both sides, we obtain:
βλ (h, l) − β0 (h, l) ∆λλ k(u, u)
lim lim ≤ lim lim γ(h, l).
(h,l)→(0,0) λ→0 λ λ→0 λ2 (h,l)→(0,0)
18 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

By using lemma 3.8, we recognize that limλ→0 ∆λλλk(u,u) 2 must coincide with
∂2k
(u, u) which, according to the hypothesis, exists for every u ∈ Ω. Since
∂v∂u
lim(h,l)→(0,0) γ(h, l) = 0 as observed in the proof of lemma 3.14, we have that
βλ (h, l) − β0 (h, l)
lim lim = 0.
(h,l)→(0,0) λ→0 λ
∂2k
Hence, according to the conclusions of corollary 3.16, (u, v) is a com-
∂v∂u
plex number, as asserted. 
Remark 3.19. It may be of interest to observe that proposition 3.17 might
also have been used, as in the proof above, to establish the relevant inequal-
ities in the proofs of theorem or lemma 3.13 and lemmas 3.14 and 3.15.
We are now ready to prove our main result in the real context.
Theorem 3.20 (Propagation of regularity, real context). Suppose Ω is an
open subset of R and let k : Ω2 → C be a positive definite kernel. If k is of
class Sn (U ) for some open subset U containing the diagonal D(Ω2 ), then k
is of class Sn (Ω2 ).
Proof. We first concentrate on the case n = 1. Since k is of class S1 (U ), we
∂k ∂k ∂ 2 k ∂2k
derive from lemmas 3.15 and 3.18 that , , and exist for
∂u ∂v ∂v∂u ∂u∂v
2
all (u, v) ∈ Ω . Continuity of these functions will now ensure that k is in
class S1 (Ω2 ).
∂k
We first consider . By lemma 3.9, for every u0 ∈ Ω there exists a
∂v  
∂k
neighborhood U (u0 ) such that the family (u, v), u ∈ U (u0 ) of fixed-
∂v
u functions of v is equicontinuous in Ω. On the other hand, continuity of
∂k ∂2k
(u, v) in the first variable is implied by the existence of for all
∂v ∂u∂v
∂k
u ∈ Ω. Hence is in the conditions of lemma 2.1 and we conclude that it
∂v
is continuous in Ω2 . A similar procedure, or the simple observation that k
∂k
satisfies property P2 , leads to the corresponding conclusion for .
∂u
2
∂ k
In order to show that is continuous in Ω2 , we first recall from
∂u∂v
theorem 3.7 that this function is itself a positive definite kernel. Since it
is, by hypothesis, continuous on U , we conclude from theorem 2.3 that it is
∂2k
continuous on Ω2 and coincides (by Schwarz’s theorem) with and is
∂v∂u
of class S1 (Ω2 ).
For the induction step, suppose the statement holds for n − 1. In order
to prove that it also holds for n, suppose k is in class Sn (U ). Then it is of
class Sn−1 (U ) and therefore of class Sn−1 (Ω). Hence
∂ 2(n−1) k
kn−1 ≡ (u, v)
∂un−1 ∂v n−1
 PROPAGATION OF REGULARITY 19

exists for all (u, v) ∈ Ω2 and, according to [7], is a positive definite kernel.
Writing kn−1 in place of k and repeating the arguments of the first part
of the proof, we conclude that kn−1 has continuous first and second mixed
derivatives in Ω2 . Therefore k is of class Sn (Ω2 ), and the conclusions of the
theorem now follow by induction on n. 
3.3. Propagation of differentiability for positive definite functions.
We now derive consequences from theorems 3.13 and 3.20 for positive and
co-positive definite functions. In the complex variable case, we have:
Theorem 3.21. Suppose Ω ⊂ C is an open set, S = codiff Ω = Ω − Ω∗ , and
f : S → C is a positive definite function. If f is holomorphic on an open
set U containing DΩ (S), then f is holomorphic in S.
Proof. Define k : Ω2 → C by k(u, v) = f (u − v). Then k is a positive definite
kernel. Continuity of the mapping s(u, v) = u − v implies that U = s−1 (U )
is an open subset of C2 . Moreover, it is clear that D(Ω2 ) ⊂ U . Since, for
every (x, y) ∈ U , we have u − v ∈ U , the fact that f is holomorphic in U
implies that k is sesquiholomorphic in U , with
∂k ∂k
(u, v) = f ′ (u − v), (u, v) = −f ′ (u − v).
∂u ∂v
Then, according to theorem 3.13, we conclude that k is a holomorphic pos-
itive definite kernel.
Now, for any v ∈ Ω, write z = u − v for all u ∈ Ω and consider the
mapping f : Ω − {v} → C defined by
f (z) = k(z + v, v).
This mapping is clearly holomorphic in Ω − {v} since k(u, v) is holomorphic
in the first variable u ∈ Ω for any v ∈ Ω. Since v is arbitrary, we conclude
that f is holomorphic in S = Ω − Ω∗ , as asserted. 
Theorem 3.22. Suppose Ω ⊂ R is an open set, S = Ω−Ω (resp. S = Ω+Ω),
and f : S → C is a positive definite function (resp. co-positive definite
function). If f is of class C 2n in an open set U containing DΩ (S) = {0}
(resp. DΩ (S)), then f is of class C 2n in S.
Proof. Define k : Ω2 → C by k(u, v) = f (u + v ∗ ), where v ∗ = −v (resp.
v ∗ = v). Then k is a positive definite kernel. Continuity of the map s(u, v) =
u + v ∗ implies that U = s−1 (U ) is an open subset of R2 . Moreover, it is clear
that D(Ω2 ) ⊂ U . Since for every (u, v) ∈ U we have u + v ∗ ∈ U , the fact
that f is of class C 2n in U implies that k is of class Sn (U ). Then theorem
3.20 implies that k is of class Sn (Ω2 ). Let z0 = u0 + v0∗ , where z0 ∈ S and
(u0 , v0 ) ∈ Ω2 and suppose I is an interval containing the origin such that
(u0 + I) × (v0 + I ∗ ) ⊂ Ω2 , where I ∗ = −I (resp. I ∗ = I). Then
J = (u0 + I) + (v0 + I ∗ )∗ = z0 + 2I ⊂ S
is an interval containing z0 . For any z ∈ J, write (u, v) = (u0 + z−z 2 , v0 +
0

z−z0 ∗


2 ) and observe that f (z) = k(u, v) to conclude that f is of class
C 2n (J) with
 i  ∗ j
(i+j) 1 1 ∂ i+j k
f (z) = (u, v), i, j = 1, . . . , n.
2 2 ∂ui ∂v j
20 J. BUESCU, A. C. PAIXÃO, AND C. P. OLIVEIRA

Since z0 is arbitrary, it follows that f is of class C 2n in S, as asserted. 

Remark 3.23. Theorem 3.21 is particularly significant: its results may be
obtained via Fourier-Laplace integral transforms [5], [24], [11] in horizontal
strips of the complex plane but not in more general codifference sets. The
reason our result is more general is that it does not require the existence of
integral representations, building directly from the algebraic positive defi-
niteness condition.
Analogously, the proof of theorem 3.22 does not require the existence of
integral representations. For positive definite functions, in the case Ω = R
the Bochner integral representation yields equivalent results to our theorem
3.22 (see e.g. Donoghue [15]). If S does not coincide with R the situation
is more delicate. First note that an open codifference set always contains a
neighborhood V of the origin. Then, given a C 2n (n ≥ 0) positive definite
function in S, the classical extension theorems [19], [22] together with theo-
rem 3.7 imply that there exists a positive definite C 2n extension of f|V to R.
This extension is, however, in general not unique, and thus does not allow
the reconstruction of the original f in S.
When f is co-definite positive and DΩ (S) = S = Ω + Ω the statements
in the theorem are obviously trivial. However, this does not happen in
general (consider, for instance, the case where Ω is the union of disjoint
open intervals).
Remark 3.24. It is worth observing that regularity propagation does not
occur in the odd (i.e. C 2n+1 ) case. Wolfe [28] shows constructively that,
if k is a positive odd integer, there exists a characteristic function f such
that f (k) (0) exists but f (k)(tm ) does not exist for a sequence of numbers
{tm } such that tm → 0 as m → ∞. Characteristic functions are Fourier
transforms of probability measures on R, and thus, by Bochner’s theorem,
coincide with the real-variable continuous positive definite functions up to a
normalization factor. Thus this negative regularity propagation result holds
for real-variable positive definite functions.

References
[1] L. Ahlfors, Complex Analysis, 3rd ed. McGraw-Hill, 1979.
[2] N. Aronszajn, Theory of reproducing kernels. Trans. Amer. Math. Soc. 68 (1950),
337–404.
[3] C. Berg, J. Christensen, P. Ressel, Harmonic analysis on semigroups. Springer-Verlag,
GTM 100, New York, 1984.
[4] S. Bernstein, Sur les fonctions absolument monotones. Acta Math. 52 (1929), 1–66.
[5] T. Bisgaard and Z. Sasvári, Characteristic functions and moment problems. Nova
Science Publishing, New York, 2000.
[6] J. Buescu, A. Paixão, A linear algebraic approach to holomorphic reproducing kernels
in Cn . Linear Algebra Appl. 412 (2006), no. 2-3, 270–290.
[7] J. Buescu, A. Paixão, Positive definite matrices and differentiable reproducing kernel
inequalities. J. Math. Anal. Appl. 320 (2006), 279–292.
[8] J. Buescu, A. Paixão, On differentiability and analyticity of positive definite functions.
J. Math. Anal. Appl. 375 (2011), no. 1, 336–341.
[9] J. Buescu, A. Paixão, Real and complex variable positive definite functions. São Paulo
Journal of Mathematical Sciences 6, 2 (2012), 155–169.
[10] J. Buescu, A. Paixão, Complex variable positive definite functions. Complex Anal.
Oper. Theory 8 (2014), no. 4, 937–954.
 PROPAGATION OF REGULARITY 21

[11] J. Buescu, A. Paixão, A. Symeonides, Complex positive definite functions on strips.

Complex Anal. Oper. Theory 11 (2017), no. 3, 627–649.
[12] A. Devinatz, On the extensions of positive definite functions. Acta Math. 102 (1959),
109–134.
[13] A. Devinatz, Integral representations of positive definite functions. Trans. Amer.
Math. Soc. 74 (1953), 56–77.
[14] A. Devinatz, Integral representations of positive definite functions II. Trans. Amer.
Math. Soc. 77, (1954), 455–480.
[15] W. Donoghue, Distributions and Fourier transforms. Academic Press, New York,
1969.
[16] C. Fitzgerald, R. Horn, On the structure of Hermitian-symmetric inequalities. J.
Lond. Math. Soc. (2), 15 (1977), 419–430.
[17] P. Graczyk, J. Loeb, Bochner and Schoenberg theorems on symmetric spaces in the
complex case. Bull. Soc. Math. France 122 (1994), no. 4, 571–590.
[18] M. Krein, Hermitian positive kernels on homogeneous spaces I. Amer. Math. Soc.
Transl. (2), 34 (1963), 69–108.
[19] M. Krein, Sur le problème du prolongement des fonctions hermitiennes positives et
continues. C. R. (Doklady) Acad. Sci. URSS (N.S.) 26 (1940). 17–22.
[20] P. Jorgensen, R. Niedzialomski, Extension of positive definite functions. J. Math.
Anal. Appl. 422 (2015), no. 1, 71–740.
[21] O. Lehto, Some remarks on the kernel functions in Hilbert spaces. Ann. Acad. Sci.
Fenn., Ser. A I, 6 (1952), 109.
[22] W. Rudin, The extension problem for positive-definite functions. Illinois J. Math. 7
(1963) 532–539.
[23] Z. Sasvári, Positive definite and definitizable functions. Mathematical Topics, 2.
Akademie Verlag, Berlin, 1994.
[24] Z. Sasvári, Multivariate characteristic and correlation functions. De Gruyter Studies
in Mathematics, 50. Walter de Gruyter & Co., Berlin, 2013.
[25] L. Schwartz, Sous-espaces hilbertiens d´espaces vectoriels topologiques et noyaux
associés (noyaux reproduisants). J. Analyse Math., 13 (1964), 115–256.
[26] I. Steinwart and A. Christmann, Support vector machines. Information Science and
Statistics, Springer, New York, 2008.
[27] D. Widder, Necessary and sufficient conditions for the representation of a function by
a doubly infinite Laplace integral. Bull. Amer. Math. Soc. 40 (1934), no. 4, 321–326.
[28] S. Wolfe, On derivatives of characteristic functions. Ann. Probab. 3 (1975), 4, 737–
738.
[29] E. Youssfi, Harmonic analysis on conelike bodies and holomorphic functions on tube
domains. J. Funct. Anal. 155 (1998), no. 2, 381–435.

Dep. Matemática, FCUL and CMAFCIO, Portugal

E-mail address: [email protected]

Área Departamental de Matemática, ISEL, Portugal

E-mail address: [email protected]

UNIFEI, Instituto de Matemática e Computação, Brazil

E-mail address: [email protected]

View publication stats