Approximation processes by multidimensional

arXiv:2405.16935v1 [math.CA] 27 May 2024

Bernstein-type exponential polynomials on the

hypercube
Laura Angeloni∗, Danilo Costarelli, Chiara Darielli

Department of Mathematics and Computer Science

University of Perugia
1, Via Vanvitelli, 06123 Perugia, Italy

[email protected] - [email protected]
[email protected]

Abstract
In this paper we introduce a new family of Bernstein-type exponential
polynomials on the hypercube [0, 1]d and study their approximation prop-
erties. Such operators fix a multidimensional version of the exponential
function and its square. In particular, we prove uniform convergence, by
means of two different approaches, as well as a quantitative estimate of
the order of approximation in terms of the modulus of continuity of the
approximated function.
Keywords: Exponential polynomials, modulus of continuity, Korovkin
theorems, constructive approximation, Lipschitz classes

AMS Subjclass: 41A25, 41A30

1 Introduction
Polynomial approximation is a classical topic of Approximation Theory, arising
from the well-known fundamental Weierstrass theorem ([37]), asserting that any
continuous function within a closed and bounded interval can be approximated
through polynomials. One of the most famous proofs of such result is given by
means of Bernstein operators, a family of polynomials introduced by Bernstein
∗ Corresponding author

1
in 1912. Such operators, defined as
n
X   
k n k
Bn f (x) = Bn (f, x) := f x (1 − x)n−k , x ∈ [0, 1],
n k
k=0

for f ∈ C([0, 1]), n ∈ N, provide a powerful methodology for approximation,

and their properties play a crucial role in this context (see, e.g., [36, 33, 26, 24,
21, 25]).
Starting from the above operators, several generalizations of such polynomi-
als in many different directions have been considered in the literature. Among
them, we mention the following remarkable examples of linear and nonlinear
operators: the sampling-type operators ([15, 9, 16, 10, 7]), the max-product
operators ([17, 18]), the neural network-type operators ([22, 30]), the Szasz-
Mirakjan operators ([1, 2]) and many others.

In the direction of the multidimensional generalization, the definition of

Bernstein polynomials was extended to functions f defined on the hypercube
Qd = [0, 1]d as follows:
n
X n
X     
en f (x) = B
en (f, x) := h1 hd n n
B ... f ,..., ...
n n h1 hd
h1 =0 hd =0

xh1 1 (1 − x1 )n−h1 . . . xhd d (1 − xd )n−hd , (1)

en f )n were initially
x = (x1 . . . . , xd ) ∈ Qd . The approximation properties of (B
explored by Hildebrandt and Shoenberg ([29]), as well as by Butzer ([20]).
A remarkable development in Approximation Theory is provided by the
Korovkin-type theorems, where the concept of polynomial approximation is gen-
eralized to sequences of positive linear operators. Indeed, this theory provides
more flexible criteria, allowing to consider broader classes of approximable func-
tions.
Recently, the study of approximation results by means of exponential-type
operators has been deeply developed by several authors (see, e.g., [4, 1, 23, 28,
11, 12, 27, 3, 8]). In particular, in [32] Morigi and Neamtu introduced and
studied the following exponential generalization of the Bernstein polynomials:
n
X  
k
Gn f (x) = Gn (f, x) := f e−µk/n eµx pn,k (an (x)), n ∈ N,
n
k=0

for f ∈ C([0, 1]), where

eµx/n − 1
an (x) = ,
eµ/n − 1


µ > 0, and pn,k (x) = nk xk (1 − x)n−k . Besides the uniform convergence, they
proved shape preserving properties and a Voronovskaja formula.

2
 Following the above mentioned path for the classical Bernstein polynomials,
we introduce in this paper a multidimensional version of the operators Gn for
functions defined on the hypercube Qd and study their approximation proper-
ties. In particular, for such operators we prove the uniform convergence in case
of continuous functions by means of the concept of Korovkin subset. More-
over, we also furnish a proof of the same result by means of a constructive
approach, since this opens the way to the achievement of a quantitative esti-
mate employing the modulus of continuity of the approximated function. As a
direct consequence it is possible to obtain the corresponding qualitative order
of approximation in case of functions belonging to suitable Lipschitz classes.

2 Preliminaries
Let us provide some notations, definitions and properties that will be used
throughout the paper.
Given a metric space (X, d), we denote by F (X) the linear space of all the
real-valued functions defined on X and by C(X) the subspace of F (X) of all
the continuous functions on X.
Definition 2.1. A linear subspace E of F (X) is said to be a lattice subspace if
|f | ∈ E f or every f ∈ E.
Definition 2.2. A Banach lattice E is a vector space endowed with a norm k · k
and an order relation ≤ on E such that
(i) (E, k · k) is a Banach space;
(ii) (E, ≤) is a vector lattice;
(iii) if f, g ∈ E and |f | ≤ |g|, then kf k ≤ kgk.
Definition 2.3. Given a metric space (Y, d′ ), a linear operator T : E → F (Y )
is said to be positive if
T (f ) ≥ 0 f or every f ∈ E, f ≥ 0.
We recall the celebrated Korovkin theorem, that provides a very useful and
simple criterion in order to prove that a given sequence (Tn )n≥1 of positive linear
operators on C([0, 1]) is an approximation process, i.e., Tn (f ) → f uniformly
on [0, 1], as n → +∞, for every f ∈ C([0, 1]).
Theorem 2.1 (Korovkin, 1953). Let (Tn )n≥1 be a sequence of positive linear
operators from C([0, 1]) into C([0, 1]) such that, for every g ∈ {e0 , e1 , e2 }, with
ei (t) = ti , i = 0, 1, 2,
lim Tn (g) = g unif ormly on [0, 1].
n→+∞

Then, for every f ∈ C([0, 1]),

lim Tn (f ) = f unif ormly on [0, 1].
n→+∞

3
 In [19], H. Bohman showed a result analogous to Theorem 2.1 by considering
sequences of positive linear operators on C([0, 1]) of the form
X
T (f, x) = f (ai )ϕi (x), 0 ≤ x ≤ 1,
i∈I

where (ai )i∈I is a finite family in [0, 1], ϕi ∈ C([0, 1]) and I is a set of indices.
For this reason, Theorem 2.1 is often called the Bohman-Korovkin theorem.
The above theorem can be generalized introducing the notion of Korovkin
subset (see, e.g., [6, 5]).
Definition 2.4. Let E and F be Banach lattices and consider a positive linear
operator T : E → F. A subset M of E is said to be a Korovkin subset of E
for T if for every sequence (Tn )n≥1 of positive linear operators from E into F
satisfying
(i) sup kTn k < +∞
n≥1

and
(ii) lim Tn (g) = T (g) for every g ∈ M ,
n→+∞

it turns out that

lim Tn (f ) = T (f ) f or every f ∈ E.
n→+∞

By means of such definition, as a consequence of Theorem 2.1, we can assert

that ei (t) = ti , i = 0, 1, 2 form a Korovkin subset of C([0, 1]). A useful criterion
to find Korovkin subsets is furnished by the following proposition.
Proposition 2.1. Let M be a subset of C(X) and let L(M ) be the linear sub-
space of C(X) generated by M . Assume that for every x, y ∈ X, x 6= y, there
exists h ∈ L(M ), h ≥ 0, such that h(x) = 0 and h(y) > 0. Then M is a
Korovkin subset of C(X).
Remark 2.1. Proposition 2.1 also holds for any compact metric space X ⊆ Rd .
Applying the previous proposition it is easy to find other Korovkin sub-
sets of C([0, 1]), besides the power functions {1, x, x2 }, such as, for example,
{1, exp(λ1 x), exp(λ2 x)}, x ∈ [0, 1], where 0 < λ1 < λ2 .
In case of functions of several variables, it is possible to prove that
( d
)
X
2
e0 , pr1 , . . . , prd , pri ,
i=1

where e0 (x) := 1 and pri (x) := xi , for every x := (x1 , . . . , xd ) ∈ X, X compact

subset of Rd , i = 1, . . . , d, is a Korovkin subset of C([0, 1]d ) (see [5]).
For details and insights on Korovkin-type approximation theory, see, e.g.,
[19, 31, 6, 5] and the references therein.

4
3 Bernstein-type Exponential Polynomials
In [32] the authors introduced a generalization of the Bernstein operators asso-
ciated to an exponential function. For a fixed real parameter µ > 0, by expµ we
denote expµ (x) = eµx , x ∈ R.
The above mentioned generalization is defined for f ∈ C([0, 1]) and for n ∈ N
by
n
X  
k
Gn f (x) = Gn (f, x) := f e−µk/n eµx pn,k (an (x)), (2)
n
k=0

where
eµx/n − 1
an (x) = .
eµ/n − 1
The connection between the operators (2), called Bernstein-type exponential
operators, and the classical Bernstein operators is given by
 
f
Gn (f, x) = expµ (x)Bn , an (x) . (3)
expµ

We note that, for each n ∈ N, an (x) is an increasing and convex continuous

function that satisfies an (0) = 0, an (1) = 1 and an (x) > 0 for every x ∈]0, 1].
Consequently, Gn is a positive operator that interpolates continuous functions
at the endpoints of the interval [0, 1]. Moreover we note that the operators Gn
reproduce expµ and exp2µ , i.e.,

Gn (expµ , x) = eµx , Gn (exp2µ , x) = e2µx . (4)

We point out that the second relation is a consequence of the equality

n
X µk
e n pn,k (an (x)) = eµx
k=0

(see [13]).
In addition, we recall the following lemma showing some basic identities of
the operators Gn that will be used later.
Lemma 3.1 ([11]). For each n ∈ N and x ∈ [0, 1], the following identities hold:

n
(i) Gn (e0 , x) = eµ(x−1) eµ/n + 1 − eµx/n ,

n
(ii) Gn (exp3µ , x) = eµx eµ(x+1)/n + eµx/n − eµ/n ,

n
(iii) Gn (exp4µ , x) = eµx eµ(x+2)/n + eµ(x+1)/n + eµx/n − eµ/n − e2µ/n .

5
 For each x ∈ (0, 1), we consider the function expµ,x defined for every t ∈ [0, 1]
by

expµ,x (t) := eµt − eµx .

Using the expression in (3) and Lemma 3.1, we can explicitly compute the
operators Gn in exp2µ,x , for every x ∈ [0, 1]:

Gn (exp2µ,x , x) = e2µx Gn (eo , x) − e2µx = e2µx (Gn (eo , x) − 1)

  n 
= e2µx eµ(x−1) eµ/n + 1 − eµx/n − 1 . (5)

In [11] the authors prove the following theorem of uniform convergence for
the operators (Gn f )n .
Theorem 3.1. If f ∈ C([0, 1]), then Gn f converges to f uniformly on [0, 1] as
n → +∞.
Furthermore, also a Voronovskaja formula has been proved. In particular,
in the case of e0 , one has:
  n 
lim n(Gn (e0 , x) − 1) = lim n eµ(x−1) eµ/n + 1 − eµx/n − 1
n→∞ n→∞
= µ2 x(1 − x), x ∈ [0, 1].

In particular, from the above limit it is easy to deduce the following useful (and
uniform) inequality:

µ2
|Gn (e0 , x) − 1| ≤ , x ∈ [0, 1], n ∈ N. (6)
n
For additional information regarding shape preserving characteristics, an anal-
ysis of Bernstein polynomials applied to convex functions and a study of the
problem of simultaneous approximation see [14, 35, 34, 38, 11, 3].

4 Multidimensional Bernstein-type Exponential

Polynomials
We now extend the definition of the operators Gn to obtain a new tool to re-
construct multivariate functions. The idea is to define the operators Gn in the
multidimensional case, with the purpose of studying some approximation prop-
erties in this setting.
We consider the hypercube Qd = [0, 1]d and the vector x = (x1 , . . . , xd ) ∈
d
R , where d ≥ 1. Additionally, we let
Pd
expµ (x) := eµx1 . . . eµxd = eµ i=1 xi
, (7)

that is, the multidimensional exponential function defined on the space Rd .

6
 For every n ≥ 1, f ∈ C(Qd ) and x ∈ Qd , we can define the multidimensional
Bernstein-type exponential operators as follows:
n
X n
X  
k1 kd
Gen f (x) = Gen (f, x) := ... f ,..., e−µk1 /n . . . e−µkd /n
n n
k1 =0 kd =0

eµx1 . . . eµxd pn,k1 (an (x1 )) . . . pn,kd (an (xd )). (8)

The definition of the multidimensional Bernstein polynomial in the variables

x1 , . . . , xd , corresponding to the function f ∈ C(Qd ), is provided in (1).
Even in the multidimensional case, there exists a connection between the
operators defined in (8) and those ones defined in (1), expressed as
n
X n
X  
k1 kd
Gen (f, x) = ... f ,..., e−µk1 /n . . . e−µkd /n
n n
k1 =0 kd =0
µx1 µxd
e ... e pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
Xn Xn  
µx1 µxd k1 kd
=: e ... e ... fµ ,..., pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
n n
k1 =0 kd =0

= e µx1
... e µxd en fµ (an (x1 ), . . . , an (xd )).
B (9)

where
   
k1 kd k1 kd
fµ ,..., := f ,..., e−µk1 /n . . . e−µkd /n .
n n n n

While Bn fixes the functions e0 and e1 , the operator Gn reproduces expµ and
exp2µ and this property holds true in the multidimensional case as well. Indeed,
the operator Gen , defined in (8), fixes the multidimensional exponential functions
expµ and exp2µ . In particular, as in (9), we have that
n
X n
X
Gen (expµ , x) = expµ (x) ... pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
k1 =0 kd =0
n n
!
X X
= expµ (x) pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
k1 =0 kd =0

= expµ (x),
n
X
taking into account that pn,k (an (x)) = 1, for every x ∈ [0, 1]. For the other
k=0

7
multidimensional exponential function, there holds
n
X n
X  
k1 kd
Gen (exp2µ , x) = ... exp2µ , . . ., e−µk1 /n . . . e−µkd /n
n n
k1 =0 kd =0
µx1
. . . eµxd pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
e
n
X n
X
= ... e2µk1 /n . . . e2µkd /n e−µk1 /n . . . e−µkd /n
k1 =0 kd =0
µx1
. . . eµxd pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
e
n
X Xn
µk1 /n µx1
= e e pn,k1 (an (x1 )) . . . eµkd /n eµxd pn,kd (an (xd ))
k1 =0 kd =0

= Gn (exp2µ , x1 ) ... Gn (exp2µ , xd ) = exp2µ (x1 ) . . . exp2µ (xd )

= exp2µ (x), (10)

by (4) and (7). In conclusion, these operators fix the multidimensional expo-
nential functions, i.e., Gen (expµ , x) = expµ (x) and Gen (exp2µ , x) = exp2µ (x).

It is clear that, similarly to (10), the operators Gen applied to expm

µ reduce to

d
Y
Gen (expm
µ , x) = Gn (expm
µ , xi ), (11)
i=1

for every positive integer m and x ∈ Qd . Then the following result can be easily
derived by Lemma 3.1.
Lemma 4.1. For x ∈ Qd , n ∈ N, there holds
(a) Gen (e0 , x) = eµ(x1 −1) (eµ/n + 1 − eµx1/n )n . . . eµ(xd −1) (eµ/n + 1 − eµxd/n )n ,

n
(b) Gen (exp3µ , x) = eµx1 eµ(x1 +1)/n + eµx1 /n − eµ/n ...

n
eµxd eµ(xd +1)/n + eµxd /n − eµ/n ,

n
(c) Gen (exp4µ , x) = eµx1 eµ(x1 +2)/n + eµ(x1 +1)/n + eµx1 /n − eµ/n − e2µ/n

n
. . . eµxd eµ(xd +2)/n + eµ(xd +1)/n + eµxd /n − eµ/n − e2µ/n .

5 Approximation Properties
We now discuss and prove convergence results for the operators Gen . We first
study the uniform convergence of our operators. In order to do it we prelimi-
narily prove the uniform convergence for the function e0 .
Theorem 5.1. Gen e0 converges to e0 uniformly on Qd , as n → +∞.

8
Proof. From the equation (11), we obtain
d
Y d
Y
Gen (e0 , x) − e0 (x) ≤ Gn (e0 , xi ) Gn (e0 , x1 ) − 1 + Gn (e0 , xi ) Gn (e0 , x2 ) − 1
i=2 i=3

+ . . . + Gn (e0 , xd ) Gn (e0 , xd−1 ) − 1 + Gn (e0 , xd ) − 1

 
d
X Yd
=  Gn (e0 , xi ) Gn (e0 , xj−1 ) − 1  + Gn (e0 , xd ) − 1 .
j=2 i=j

By (i) of Lemma 3.1, we have that

Gn (e0 , xi ) ≤ eµ ,

for every xi ∈ [0, 1], with i = 1, . . . , d, and for every n ∈ N. In addition, we

note that, recalling the consequence of the Voronovskaja formula stated in (6),
we finally obtain that

Gen (e0 , x) − e0 (x) ≤ eµ(d−1) Gn (e0 , x1 ) − 1 + eµ(d−2) Gn (e0 , x2 ) − 1

+ . . . + eµ Gn (e0 , xd−1 ) − 1 + Gn (e0 , xd ) − 1

d−1
µ2 X µj
≤ e ,
n j=0

for n large enough and this completes the proof.

Now, we can establish the following result thanks to the use of a Korovkin
theorem.
Theorem 5.2. If f ∈ C(Qd ), then Gen f converges to f uniformly on Qd , as
n → +∞.
2 2
Proof. We notice that the set M := {e0 , prµ,1 , . . . , prµ,d , prµ,1 , . . . , prµ,d }, where

prµ,i (x) := expµ (xi ), i = 1, . . . , d,

with x ∈ Qd , is a Korovkin subset of C(Qd ), by Proposition 2.1. Indeed, if we

fix x
e ∈ Qd , it is easy to see that the function
d
X d
X
h(x) := d − 2 eµ(xi −exi ) + e2µ(xi −exi )
i=1 i=1

is such that h(x) ≥ 0, for every x ∈ Qd , and h(e x) = 0, while h(x) > 0, for every
x 6= x
e . Therefore, the thesis is fulfilled if we check the uniform convergence for
the functions of M .

9
The uniform convergence on e0 is given by Theorem 5.1. Instead, for the func-
tions prµ,j , with j = 1, . . . , d, we note that
n
X n
X
Gen (prµ,j , x) = ... eµkj /n e−µk1 /n . . . e−µkd /n
k1 =0 kd =0

µx1 µxd
e pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
...e
 
n
X Yd X n
 
= eµkj /n e−µkj /n eµxj pn,kj (an (xj ))  eµxi e−µki /n pn,ki (an (xi ))
kj =0 i=1 ki =0
i6=j

d
Y d
Y
= Gn (expµ , xj ) Gn (e0 , xi ) = expµ (xj ) Gn (e0 , xi ),
i=1 i=1
i6=j i6=j

taking into account that Gn (expµ , xj ) = expµ (xj ) from the first equation of (4).
Hence, we obtain
d
Y
Gen (prµ,j , x) = expµ (xj ) Gn (e0 , xi ),
i=1
i6=j

that tends to expµ (xj ) = prµ,j (x), as n → +∞, since Theorem 3.1 holds.
2
Now, for the remaining functions prµ,j , with j = 1, . . . , d, we do similar reason-
ings, i.e.,
n
X n
X
Gen (prµ,j
2
, x) = ... e2µkj /n e−µk1 /n . . . e−µkd /n
k1 =0 kd =0

eµx1 . . . eµxd pn,k1 (an (x1 )) . . . pn,kd (an (xd ))

n
X d
Y n
X
= e2µkj /n e−µkj /n eµxj pn,kj (an (xj )) eµxi e−µki /n pn,ki (an (xi )),
kj =0 i=1 ki =0
i6=j

where the first sum is equal to Gn (exp2µ , xj ) = exp2µ (xj ) from the second equation
of (4). So we obtain
d
Y
Gen (prµ,j
2
, x) = exp2µ (xj ) Gn (e0 , xi ),
i=1
i6=j

that tends to exp2µ (xj ) = prµ,j

2
(x), as n → +∞, from Theorem 3.1. This
completes the proof.

10
 Now, we will provide an alternative proof of the previous result using a dif-
ferent technique, that is, my means of a constructive approach. This will be the
starting point to achieve quantitative estimates for the order of approximation.
Proof. Without any loss of generality, we will prove the result in the bi-dimensional
case, that is, d = 2, for the sake of clearness. Of course, the general case is com-
pletely analogous.
Taking x = (x1 , x2 ), we get

Gen (f, x) − f (x) ≤ Gen (f, x) − f (x) Gen (e0 , x) + f (x) Gen (e0 , x) − f (x)
=: I1 + I2 .

Now let us fix ε > 0 and we denote by γ > 0 the parameter of the uniform
convergence of f , i.e.,

|f (x1 , x2 ) − f (y1 , y2 )| ≤ ε, ∀ (x1 , x2 ), (y1 , y2 ) ∈ Qd ,

with k(x1 , x2 ) − (y1 , y2 )k2 ≤ γ, where k · k2 denotes the usual Euclidean norm.
Then we have
Xn X n  
k1 k2
I1 ≤ f , − f (x1 , x2 ) e−µk1 /n e−µk2 /n
n n
k1 =0 k2 =0

eµx1 eµx2 pn,k1 (an (x1 ))pn,k2 (an (x2 ))

X X X X
= + +
k1 k2 k1 k2
| n |
−x1 ≤ √γ2 | n |
−x2 ≤ √γ2 | n |
−x1 ≤ √γ2 | n |
−x2 > √γ2

!
X X X X
+ +
k1 k2
| n |
−x1 > √γ2 | n |
−x2 ≤ √γ2 | kn1 −x1 |> √γ2 | kn2 −x2 |> √γ2
 
k1 k2
f , − f (x1 , x2 ) e−µk1 /n e−µk2 /n eµx1 eµx2
n n
pn,k1 (an (x1 )) pn,k2 (an (x2 )) =: I1,1 + I1,2 + I1,3 + I1,4 .

We note that
n
X n
X
2
I1,1 ≤ ε (eµ ) pn,k1 (an (x1 )) pn,k2 (an (x2 ) ≤ ε e2µ .
k1 =0 k2 =0

11
Moreover, we also have that
X X
I1,2 ≤ e2µ 2kf k∞ pn,k1 (an (x1 )) pn,k2 (an (x2 ))
k1 k2
| n |
−x1 ≤ √γ2 | n |
−x2 > √γ2

X
≤ e2µ 2kf k∞ pn,k2 (an (x2 )).
| kn2 −x2 |> √γ2

Now, we can write what follows:

X X ki
n − xi
pn,ki (an (xi )) = ki
pn,ki (an (xi ))
ki ki n − xi
n −xi >δ n −xi >δ

X ki
≤ δ −1 − xi pn,ki (an (xi ))
ki
n
n −xi >δ

X ki
≤ δ −1 − an (xi ) pn,ki (an (xi ))
ki
n
n −xi >δ
X
+ δ −1 |an (xi ) − xi | pn,ki (an (xi ))
ki
n −xi >δ

1
≤ δ −1 √ + δ −1 max |an (x) − x| , (12)
2 n x∈[0,1]

δ > 0, i = 1, 2, where in the above computations we used a well-known inequality

(see, e.g., [6]) concerning the first order moment of the functions pn,ki , i.e.,
n
X ki 1
− x pn,ki (x) ≤ √ , x ∈ [0, 1]. (13)
n 2 n
ki =0

Observing that

γn : = max |an (x) − x|

x∈[0,1]


n  n eµ/n − 1 µn − 1
µ/n
= ln e −1 − −→ 0, as n → +∞, (14)
µ µ eµ/n − 1

and using (12), then we finally obtain that

ε
I1,1 ≤ e2µ 4kf k∞ ,
γ2
for n sufficiently large. Since the reasoning is similar for I1,3 and I1,4 , the thesis
immediately follows.

12
 For each x ∈ Rd , we consider the function expµ,x defined for every t =
(t1 , . . . , td )∈ Rd by
Pd Pd
expµ,x (t) := expµ (t) − expµ (x) = eµ i=1 ti − eµ i=1 xi
.
Now, using Lemma 4.1 and the expressions in (5) and (9), we can explicitly
compute the operator Gen for the square of expµ,x , for every x ∈ Qd :

Gen (exp2µ,x , x) = Gen (exp2µ , x) − 2 expµ (x)Gen (expµ , x) + exp2µ (x)Gen (exp0 , x)

= exp2µ (x) − 2 expµ (x) expµ (x) + exp2µ (x)Gen (exp0 , x)

= exp2µ (x) − 2 exp2µ (x) + exp2µ (x)Gen (exp0 , x)

 
= exp2µ (x) Gen (e0 , x) − 1
d
! !
Pd Y
= e2µ i=1 xi Gn (e0 , xi ) − 1 .
i=1

In order to obtain a quantitative estimate for the approximation error of the

operator Gen , we recall the definition of the modulus of continuity in the mul-
tidimensional case. For f ∈ C(Qd ), the modulus of continuity of f is given
by
ω(f, δ) := sup |f (t) − f (x)| ,
kt−xk2 ≤δ
x,t∈Qd

where δ > 0.
By means of a different approach, it is possible to achieve the following
quantitative estimate for the order of approximation.
Theorem 5.3. For every f ∈ C([0, 1]), n ∈ N, there holds:
     
e d µd 1 µd µ2 µd 1
k Gn f − f k ∞ ≤ 1 + e ω f, √ + e kf k∞ + e (1 + µd) ω f,
2 n n n
Proof. We can write what follows:
Gen (f, x) − f (x)

= Gen (f, x) − Gen (e0 , x)f (x) + Gen (e0 , x)f (x) − f (x)

≤ Gen (f, x) − f (x)Gen (e0 , x) + f (x) Gen (e0 , x) − 1

n
X Xn  
k1 kd
≤ ... f ,..., − f (x) e−µk1 /n . . . e−µkd /n eµx1 . . . eµxd
n n
k1 =0 kd =0

pn,k1 (an (x1 )) . . . pn,kd (an (xd )) + kf k∞ Gen (e0 , x) − 1 =: J1 + J2 .

13
Now, the estimate for J2 follows immediately since, using (6), we can obtain:
µ2
J2 ≤ kf k∞
,
n
for n sufficiently large. About J1 , using the subadditivity of the modulus of
continuity, we have
Xn Xn    
k1 kd
J1 ≤ ... ω f, ,..., −x e−µk1 /n . . . e−µkd /n eµx1 . . . eµxd
n n 2
k1 =0 kd =0

pn,k1 (an (x1 )) . . . pn,kd (an (xd ))

Xn X n     
k1 kd
≤ ... ω f, ,..., − (an (x1 ), . . . , an (xd ))
n n 2
k1 =0 kd =0

+ ω (f, k(an (x1 ), . . . , an (xd )) − xk2 )}

e−µk1 /n . . . e−µkd /n eµx1 . . . eµxd pn,k1 (an (x1 )) . . . pn,kd (an (xd )) =: J1,A + J1,B .
Now, applying the following well-known property
ω(f, λδ) ≤ (1 + λ) ω(f, δ), λ > 0,
we get:
 X
n n 
X   
1 √ k1 kd
J1,A ≤ ω f, √ ... 1+ n ,..., − (an (x1 ), . . . , an (xd ))
n n n 2
k1 =0 kd =0
−µk1 /n −µkd /n µx1 µxd
e ...e e ...e pn,k1 (an (x1 )) . . . pn,kd (an (xd ))
 " Xn Xn  
µd 1 √ k1 kd
≤ e ω f, √ 1+ ... n ,..., − (an (x1 ), . . . , an (xd ))
n n n 2
k1 =0 kd =0

pn,k1 (an (x1 )) . . . pn,kd (an (xd ))]

  
  Xn n
X  Xd 
1  √ kj
≤ eµd ω f, √ 1+ ... n − an (xj )
n  n 
k1 =0 j=1
kd =0

pn,k1 (an (x1 )) . . . pn,kd (an (xd ))]

  
  Xd  Y X n
1  √
≤ eµd ω f, √ 1+  pn,ki (an (xi )) n
n 
j=1 i6=j ki =0

Xn
kj 
− an (xj ) pn,kj (an (xj ))  .
n 
kj =0

Now, using (13) we can finally write

 
  X d Y X n    
µd 1  1  d µd 1
J1,A ≤ e ω f, √ 1+ pn,ki (an (xi )) ≤ 1 + e ω f, √ .
n 2 j=1 2 n
i6=j ki =0

14
Concerning J1,B , we notice that (possibly with the help of some mathematical
software) limn→+∞ nγn = µ8 , where γn is defined in (14). Therefore, for exam-
ple, we can deduce that γn ≤ nµ for n large enough, and so, for every x ∈ Qd ,
P
k(an (x1 ), . . . , an (xd )) − xk2 ≤ dj=1 |an (xj ) − xj | ≤ dγn ≤ dµn . From this we
deduce that
   
dµ 1
ω(f, k(an (x1 ), . . . , an (xd )) − xk2 ) ≤ ω f, ≤ (1 + dµ) ω f, ,
n n

and so, for sufficiently large n ∈ N,

  n n
1 X X
J1,B ≤ (1 + dµ) ω f, ... e−µk1 /n . . . e−µkd /n eµx1 . . . eµxd
n
k1 =0 kd =0

pn,k1 (an (x1 )). . .pn,kd (an (xd ))

  n n
µd 1 X X
≤ e (1 + dµ) ω f, pn,k1 (an (x1 )). . . pn,kd (an (xd ))
n
k1 =0 kd =0
 
µd 1
= e (1 + dµ) ω f, .
n

This completes the proof.

Now, in order to obtain a qualitative estimate for the approximation error
of the operators Gen , we first recall the definition of the Lipschitz class. In
particular, for α ∈ (0, 1], the Lipschitz class of f ∈ C(Qd ) is given by

Lip(α) = {f ∈ C(Qd ) | ω(f, δ) = O(δ α ), δ → 0+ }.

Furthermore, using Theorem 5.3, we can derive the following desired estimate:

kGen f − f k∞ = O(1/nα/2 ), n → +∞,

for any f belonging to Lip(α).

Acknowledgments
The first and the second authors are members of the Gruppo Nazionale per l’Analisi
Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale
di Alta Matematica (INdAM), of the network RITA (Research ITalian network on
Approximation), and of the UMI (Unione Matematica Italiana) group T.A.A. (Teoria
dell’Approssimazione e Applicazioni).

Funding
The authors L. Angeloni and D. Costarelli have been partially supported within the
projects: (1) 2023 GNAMPA-INdAM Project “Approssimazione costruttiva e astratta

15
mediante operatori di tipo sampling e loro applicazioni” CUP E53C22001930001,
(2) 2024 GNAMPA-INdAM Project “Tecniche di approssimazione in spazi funzion-
ali con applicazioni a problemi di diffusione”, CUP E53C23001670001, and (3) PRIN
2022 PNRR: “RETINA: REmote sensing daTa INversion with multivariate functional
modeling for essential climAte variables characterization”, funded by the European
Union under the Italian National Recovery and Resilience Plan (NRRP) of NextGen-
erationEU, under the Italian Ministry of University and Research (Project Code:
P20229SH29, CUP: J53D23015950001). L. Angeloni has been also partially supported
within the project PRIN 2022 “EXPANSION - EXPlainable AI through high eNergy
physicS for medical Imaging in ONcology”, funded by the Italian Ministry of Univer-
sity and Research (CUP: J53D23002530006).

Conflict of interest/Competing interests

The authors declare that they have no conflict of interest and competing interest.

Availability of data and material and Code avail-

ability
Not applicable.

References
[1] T. Acar, A. Aral, D. Cárdenas-Morales, and P. Garrancho. Szász–Mirakyan
type operators which fix exponentials. Results in Mathematics, 72:1393–
1404, 2017.
[2] T. Acar, A. Aral, and H. Gonska. On Szász–Mirakyan operators preserving
e2ax , a > 0. Mediterranean Journal of Mathematics, 14:1–14, 2017.
[3] A. M. Acu, A. Aral, and I. Raşa. Generalized Bernstein Kantorovich op-
erators: Voronovskaya type results, convergence in variation. Carpathian
Journal of Mathematics, 38:1–12, 2022.
[4] J. M. Aldaz, O. Kounchev, and H. Render. Bernstein operators for expo-
nential polynomials. Constructive Approximation, 29:345–367, 2009.
[5] F. Altomare. Korovkin-type Theorems and approximation by positive lin-
ear operators. Surveys in Approximation Theory, 6:92–164, 2010.
[6] F. Altomare and M. Campiti. Korovkin-type Approximation Theory and
its applications. De Gruyter Studies in Mathematics. Walter de Gruyter &
Co., Berlin, New York, 1994.
[7] L. Angeloni, N. Çetin, D. Costarelli, A. R. Sambucini, and G. Vinti. Mul-
tivariate sampling Kantorovich operators: quantitative estimates in Orlicz
spaces. Constructive Mathematical Analysis, 4(2):229–241, 2021.

16
 [8] L. Angeloni and D. Costarelli. Approximation by exponential-type poly-
nomials. Journal of Mathematical Analysis and Applications, 532:127927,
2024.
[9] L. Angeloni, D. Costarelli, and G. Vinti. A characterization of the absolute
continuity in terms of convergence in variation for the sampling Kantorovich
operators. Mediterranean Journal of Mathematics, 16:1–11, 2019.
[10] L. Angeloni, D. Costarelli, and G. Vinti. Approximation properties of mixed
sampling-Kantorovich operators. Revista de la Real Academia de Ciencias
Exactas, Fı́sicas y Naturales - Serie A: Matemáticas, 115:1–14, 2021.
[11] A. Aral, D. Cárdenas-Morales, and P. Garrancho. Bernstein-type operators
that reproduce exponential functions. Journal of Mathematical Inequalities,
12:861–872, 2018.
[12] A. Aral, D. Inoan, and I. Raşa. Approximation properties of Szász–
Mirakyan operators preserving exponential functions. Positivity, 23:233–
246, 2019.
[13] A. Aral, D. Otrocol, and I. Raşa. On approximation by some Bernstein–
Kantorovich exponential-type polynomials. Periodica Mathematica Hun-
garica, 79:236–254, 2019.
[14] A. O. Arama. Properties concerning the monotonicity of the polynomials
(Russian). Studii şi Cercetări de Matematică (Cluj), 2:25–40, 1960.
[15] C. Bardaro, L. Faina, and I. Mantellini. A generalization of the exponential
sampling series and its approximation properties. Mathematica Slovaca,
67(6):1481–1496, 2017.
[16] C. Bardaro, I. Mantellini, and G. Schmeisser. Exponential sampling series:
convergence in Mellin–Lebesgue spaces. Results in Mathematics, 74:1–20,
2019.
[17] B. Bede, L. Coroianu, and S. G. Gal. Approximation by max-product type
operators. Springer, 2016.
[18] L. Boccali, D. Costarelli, and G. Vinti. Convergence results in Orlicz spaces
for sequences of max-product sampling Kantorovich operators. Journal of
Computational and Applied Mathematics, page 115957, 2024.
[19] H. Bohman. On appoximation of continuous and analytic functions. Arkiv
för Matematik, 2:43–46, 1952-54.
[20] P. L. Butzer. On two-dimensional Bernstein polynomials. Canadian Jour-
nal of Mathematics, 5:107–113, 1953.
[21] N. L. Carothers. A short course on approximation theory. Technical report,
Department of Mathematics and Statistics, Bowling Green State Univer-
sity, OH, 1998.

17
[22] D. Costarelli. Density results by deep neural network operators with integer
weights. Mathematical Modelling and Analysis, 27(4):547–560, 2022.
[23] E. Deniz, A. Aral, and V. Gupta. Note on Szász–Mirakyan–Durrmeyer
operators preserving e2ax , a > 0. Numerical Functional Analysis and Op-
timization, 39(2):201–207, 2018.
[24] R. A. DeVore and G. G. Lorentz. Constructive approximation. Springer-
Verlag Berlin, Heidelberg, New York, 1 edition, 1993.
[25] R. T. Farouki. The Bernstein polynomial basis: A centennial retrospective.
Technical report, Department of Mechanical and Aerospace Engineering,
University of California, 2012.
[26] G. Freud. On approximation by positive linear methods II. Studia Scien-
tiarum Mathematicarum Hungarica, 3:365–380, 1968.
[27] V. Gupta. Approximation with certain exponential operators. Revista
de la Real Academia de Ciencias Exactas, Fı́sicas y Naturales - Serie A:
Matemáticas, 114(2):51, 2020.
[28] V. Gupta and G. Tachev. On approximation properties of Phillips operators
preserving exponential functions. Mediterranean Journal of Mathematics,
14:1–12, 2017.
[29] T. H. Hildebrandt and I. J. Schoenberg. On linear functional operations
and the moment problem for a finite interval in one or several dimensions.
Annals of Mathematics, 34:317–328, 1933.

[30] U. Kadak. Multivariate neural network interpolation operators. Journal of

Computational and Applied Mathematics, 414:114426, 2022.
[31] P. P. Korovkin. Convergence of linear positive operators in the spaces
of continuous functions (Russian). Proceedings of the USSR Academy of
Sciences, 90:961–964, 1953.

[32] S. Morigi and M. Neamtu. Some results for a class of generalized polyno-
mials. Advances in Computational Mathematics, 12:133–149, 2000.
[33] T. Popoviciu. Sur l’approximation des fonctions convexes d’ordre supérieur.
Studii şi Cercetări de Matematică (Cluj), 10:49–54, 1935.
[34] I. J. Schoenberg. On variation diminishing spline approximation methods.
In On Numerical Approximation, pages 249–271. University of Wisconsin
Press, 1959.
[35] O. Shisha and B. Mond. The degree of convergence of sequences of lin-
ear positive operators. Proceedings of the National Academy of Sciences
(PNAS), 60:1196–1200, 1968.

18
[36] E. Voronovskaja. Détermination de la forme asymptotique d’approximation
des functions par les polynômes de M. Bernstein. Proceedings of the USSR
Academy of Sciences, pages 79–85, 1932.
[37] K. Weierstrass. Über die analytische Darstellbarkeit sogenannter
willkürlicher Functionen einer reellen veränderlichen. Sitzungsberichte der
Königlich Preussischen Akademie der Wissenschaften zu Berlin, XXII:591–
612, 1885.
[38] Z. Ziegler. Linear approximation and generalized convexity. Journal of
Approximation Theory, 1:420–443, 1968.

19