1

Approximation of Univariate Functions

1.1 Introduction
The primary problem in approximation theory is the choice of a successful method
of approximation. In this chapter and in Chapter 2 we test various approaches,
based on the concept of width, to the evaluation of the quality of a method of
approximation. We take as an example the approximation of periodic functions of
a single variable. The two main parameters of a method of approximation are its
accuracy and complexity. These concepts may be treated in various ways depend-
ing on the particular problems involved. Here we start from classical ideas about the
approximation of functions by polynomials. After Fourier’s 1807 article the repre-
sentation of a 2π -periodic function by its Fourier series became natural. In other
words, the function f (x) is approximately represented by a partial sum Sn ( f , x) of
its Fourier series:
n
Sn ( f , x) := a0 /2 + ∑ (ak cos kx + bk sin kx),
k=1
 π  π
1 1
ak := f (x) cos kx dx, bk := f (x) sin kx dx.
π −π π −π

We are interested in the approximation of a function f by a polynomial Sn ( f ) in

some L p -norm, 1 ≤ p ≤ ∞. In the case p = ∞ we assume that we are dealing with
the uniform norm. As a measure of the accuracy of the method of approximating
a periodic function by means of its Fourier partial sum we consider the quantity
 f − S( f ) p . The complexity of this method of approximation contains the fol-
lowing two characteristics. The order of the trigonometric polynomial Sn ( f ) is the
quantitative characteristic. The following observation gives us the qualitative char-
acteristic. The coefficients of this polynomial are found by the Fourier formulas,
which means that the operator Sn is the orthogonal projector onto the subspace of
trigonometric polynomials of order n.
2 Approximation of Univariate Functions

In 1854 Chebyshev suggested representing continuous function f by its polyno-

mial of best approximation, namely, by the polynomial tn ( f ) such that
 
 n 
 
 f − tn ( f )∞ = En ( f )∞ := inf  f (x) − ∑ (αk cos kx + βk sin kx) .
αk ,βk  k=0

∞

He proved the existence and uniqueness of such a polynomial. We consider this

method of approximation not only in the uniform norm but in all L p -norms, 1 ≤
p < ∞. The accuracy of the Chebyshev method can be easily compared with the
accuracy of the Fourier method:
 
En ( f ) p ≤  f − Sn ( f ) p .

However, it is difficult to compare the complexities of these two methods. The

quantitative characteristics coincide but the qualitative characteristics are different
(for example, it is not difficult to understand that for p = ∞ the mapping f → tn ( f )
is not a linear operator).
 The Du Bois–Reymond 1873 example of a continuous
function f such that f − Sn ( f )∞ → ∞ when n → ∞, and the Weierstrass theorem

which says that for each continuous function f we have En ( f )∞ → 0 as n → ∞,
showed the advantage of the Chebyshev method over the Fourier method from the
point of view of accuracy.
The desire to construct methods of approximation which have the advantages of
both the Fourier and Chebyshev methods has led to the study of various methods
of summation of Fourier series. The most important among them from the point of
view of approximation are the de la Vallée Poussin, Fejér, and Jackson methods,
which were constructed early in the twentieth century. All these methods are linear.
For example, in the de la Vallée Poussin method a function f is approximated by
the polynomial

1 2n−1
Vn ( f ) := ∑ Sl ( f )
n l=n

of order 2n − 1.
From the point of view of accuracy this method is close to the Chebyshev
method; de la Vallée Poussin proved that
 
 f −Vn ( f ) ≤ 4En ( f ) p , 1 ≤ p ≤ ∞.
p

From the point of view of complexity it is close to the Fourier method, and the
property of linearity essentially distinguishes it from the Chebyshev method.
We see that common to all these methods is approximation by trigonometric
polynomials. However, the methods of constructing these polynomials differ: some
 1.1 Introduction 3

methods use orthogonal projections on to the subspace of trigonometric polynomi-

als of fixed order, some use best-approximation operators, and some use linear
operators.
Thus, the approximation of periodic functions by trigonometric polynomials
is natural and this problem has been thoroughly studied. The approximation of
functions by algebraic polynomials has been studied in parallel with approximation
by trigonometric polynomials. We now point out some results, which determined
the style of investigation of a number of problems in approximation theory. These
problems are of interest even today.
It was proved by de la Vallée Poussin (1908) that, for best approximation of the
function |x| in the uniform norm on [−1, 1] by algebraic polynomials of degree n,
the following upper estimate or bound holds:
 
en |x| ≤ C/n.
He raised the question of the possibility of an improvement of this estimate in the
sense of order. In other words, could the function C/n be replaced by a function
that decays faster to zero? Bernstein (1912) proved that this order estimate is sharp.

Moreover, he then established the asymptotic behavior of the sequence en |x|
(see Bernstein, 1914):
 
en |x| = µ /n + o(1/n), µ = 0.282 ± 0.004.
These results initiated a series of investigations into best approximations of
individual functions having special singularities.
At this stage of investigation the natural conjecture arose that the smoother a
function, the more rapidly its sequence of best approximations decreases.
In 1911 Jackson proved the inequality
En ( f )∞ ≤ Cn−r ω ( f (r) , 1/n)∞ .
The relations which give upper estimates for the best approximations of a func-
tion in terms of its smoothness are now called the Jackson inequalities, and in a
wider sense such relations are called direct theorems of approximation theory.
As a result of Bernstein’s (1912) and de la Vallée Poussin’s (1908, 1919) inves-
tigations we can formulate the following assertion, which is now called the inverse
theorem of approximation theory. If
En ( f )∞ ≤ Cn−r−α , 0 ≤ r integer, 0 < α < 1,
then f has a continuous derivative of order r which belongs to the class Lip α ; that
is, f ∈ W r H α (in the notation of this book it is the class H∞r+α ). Thus, the results
of Jackson, Bernstein, and de la Vallée Poussin show that functions from the class
W r H α , 0 < α < 1, can be characterized by the order of decrease of its sequences
of best approximations.
4 Approximation of Univariate Functions

We remark that at that time, early in the twentieth century, classes similar to
W r H α were used in other areas of mathematics for obtaining the orders of decrease
of various quantities. As an example we formulate a result of Fredholm (1903). Let
f (x, y) be continuous on [a, b] × [a, b] and
 
max f (x, y + t) − f (x, y) ≤ C|t|α , 0 < α ≤ 1.
x,y

Then for eigenvalues λ (J f ) of the integral operator

 b
(J f ψ )(x) = f (x, y)ψ (y)dy
a

the following relation is valid for any ρ > 2/(2α + 1):

∞  ρ
∑ λn (J f ) < ∞.
n=1

The investigation of the upper bounds or estimates of errors of approximation of

functions from a fixed class by some method of approximation began with an article
by Lebesgue (1910). In particular, Lebesgue proved that
 
Sn (Lip α )∞ := sup  f − Sn ( f )∞ ≍ n−α ln n.
f ∈Lip α

Here and later we write an ≍ bn for two sequences an and bn if there are two positive
constants C1 and C2 such that C1 bn ≤ an ≤ C2 bn for all n.
The problem of approximation of functions in the classes W r H α by trigonomet-
ric polynomials was so natural that a tendency to find either asymptotic or exact
values of the following quantities appeared:
Sn (W r H α )∞ := sup  f − Sn ( f )∞ , En (W r H α )∞ := sup En ( f )∞ .
f ∈W r H α f ∈W r H α

We now formulate the first results in this direction. Kolmogorov (1936) proved the
relation (in our notation W r = W∞,r
r , see §1.4)

4 ln n
Sn (W r )∞ = + O(n−r ), n → ∞.
π 2 nr
Independently, Favard (1937) and Akhiezer and Krein (1937) proved the equality
En (W r )∞ = Kr (n + 1)−r ,
where Kr is a number depending on the natural number r.
In 1936 Kolmogorov introduced the concept of the width dn of a class F in a
space X:
 
 n 
 
dn (F, X) := infn sup infn  f − ∑ c j φ j  .
{φ j } j−1 f ∈F {c j } j−1  j=1

X
 1.1 Introduction 5

This concept allows us to find, for a fixed n and for a class F, a subspace of
dimension n that is optimal with respect to the construction of a best approxi-
mating element. In other words, the concept of width allows us to choose from
among various Chebyshev methods having the same quantitative characteristic of
complexity (the dimension of the approximating subspace) the one which has the
greatest accuracy.
The first result about widths (Kolmogorov, 1936), namely
d2n+1 (W2r , L2 ) = (n + 1)−r ,
showed that the best subspace of dimension 2n + 1 for the approximation of classes
of periodic functions is the subspace of trigonometric polynomials of order n. This
result confirmed that the approximation of functions in the class W2r by trigonomet-
ric polynomials is natural. Further estimates of the widths d2n+1 (Wq,r α , L p ), 1 ≤ q,
p ≤ ∞, some of which are discussed in §2.1 below, showed that, for some values
of the parameters q, p, the subspace of trigonometric polynomials of order n is
optimal (in the sense of the order of decay) but for other values of q, p this subspace
is not optimal.
The Ismagilov (1974) estimate for the quantity dn (W1r , L∞ ) gave the first exam-
ple, where the subspace of trigonometric polynomials of order n is not optimal.
This phenomenon was thoroughly studied by Kashin (1977).
In analogy to the problem of the Kolmogorov width, that is, to the problem con-
cerning the best Chebyshev method, problems concerning the best linear method
and the best Fourier method were considered.
Tikhomirov (1960b) introduced the linear width:
λn (F, L p ) := inf sup  f − A f  p ,
A:rank A≤n f ∈F

and Temlyakov (1982a) introduced the orthowidth (Fourier width):

 
 n 
 
ϕn (F, L p ) := inf sup 
orthonormal system {ui }ni=1 f ∈F 
f − ∑ i i  .
 f , u u
i=1 p

A discussion and comparison of results concerning dn (Wqr , L p ),

λn (Wqr , L p ) and
r
ϕn (Wq , L p ) can be found in §2.1. Here we remark that, from the point of view of the
orthowidth, the Fourier operator Sn is optimal (in the sense of order) for all 1 ≤ q,
p ≤ ∞ with the exception of the two cases (1, 1) and (∞, ∞).
Keeping in mind the primary question about the selection of an optimal sub-
space of approximating functions, we now draw some conclusions from this brief
historical survey.
(1) The trigonometric polynomials have been considered as a natural means of
approximation of periodic functions during the whole period of development
of approximation theory.
6 Approximation of Univariate Functions

(2) In approximation theory (as well as in other fields of mathematics) it has turned
out that it is natural to unite functions with the same smoothness into a class.
(3) The subspace of trigonometric polynomials has been obtained in many cases as
the solution of problems regarding the most precise method for the classes of
smooth functions: the Chebyshev method (which uses the Kolmogorov width),
the linear method (which uses the linear width), or the Fourier method (which
uses the orthowidth).
On the basis of these remarks we may formulate the following general strategy
for investigating approximation problems; we remark that this strategy turns out to
be most fruitful in those cases where we do not know a priori a natural method
of approximation. First, we solve the width problem for a class of interest in the
simplest case, that of approximation in Hilbert space, L2 . Second, we study the
system of functions obtained and apply it to approximation in other spaces L p .
This strategy will be used in Chapters 3, 4, and 5.

1.2 Trigonometric Polynomials

Functions of the form
n
t(x) = ∑ ck eikx = a0 /2 + ∑ (ak cos kx + bk sin kx)
|k|≤n k=1

(ck , ak , bk are complex numbers) will be called trigonometric polynomials of order

n. We denote the set of such polynomials by T (n) and the subset of T (n) of real
polynomials by RT (n).
We first consider a number of concrete polynomials that play an important role
in approximation theory.

1.2.1 The Dirichlet Kernel of Order n

The classical univariate Dirichlet kernel of order n is defined as follows:
Dn (x) := ∑ eikx = e−inx (ei(2n+1)x − 1)(eix − 1)−1
|k|≤n
sin(n + 1/2)x
= . (1.2.1)
sin(x/2)
The Dirichlet kernel is an even trigonometric polynomial with the majorant
   
Dn (x) ≤ min 2n + 1, π /|x| , |x| ≤ π . (1.2.2)
The estimate
Dn 1 ≤ C ln n, n = 2, 3, . . . , (1.2.3)
follows from (1.2.2).
 1.2 Trigonometric Polynomials 7

We mention the well-known relation (see Dzyadyk, 1977, p. 112)

4
Dn 1 = ln n + Rn , |Rn | ≤ 3, n = 1, 2, 3, . . .
π2
For any trigonometric polynomial t ∈ T (n) we have

−1
Dn ∗ t := (2π ) Dn (x − y)t(y)dy = t.
T

Denote
xl := 2π l/(2n + 1), l = 0, 1, . . . , 2n.

Clearly, the points xl , l = 1, . . . , 2n, are zeros of the Dirichlet kernel Dn on [0, 2π ].
For any |k| ≤ n we have
2n 2n
l l
∑ eikx Dn (x − xl ) = ∑ eimx ∑ ei(k−m)x = eikx (2n + 1).
l=0 |m|≤n l=0

Consequently, for any t ∈ T (n),

2n
t(x) = (2n + 1)−1 ∑ t(xl )Dn (x − xl ). (1.2.4)
l=0

Further, it is easy to see that for any u, v ∈ T (n) we have

 π 2n
−1
u, v := (2π ) u(x)v(x)dx = (2n + 1)−1 ∑ u(xl )v(xl ) (1.2.5)
−π l=0

and, for any t ∈ T (n),

2n  2
t22 = (2n + 1)−1 ∑ t(xl ) . (1.2.6)
l=0

For 1 < q ≤ ∞ the estimate

Dn q ≤ C(q)n1−1/q (1.2.7)

follows from (1.2.2). Applying the Hölder inequality (see (A.1.1) in the Appendix)
for estimating Dn 22 we get

2n + 1 = Dn 22 ≤ Dn q Dn q′ . (1.2.8)

Relations (1.2.7) and (1.2.8) imply for 1 < q < ∞ the relation

Dn q ≍ n1−1/q . (1.2.9)

Relation (1.2.9) for q = ∞ is obvious.

8 Approximation of Univariate Functions

We denote by Sn the operator taking a partial sum of order n. Then for f ∈ L1

we have
 π
Sn ( f ) := Dn ∗ f = (2π )−1 Dn (x − y) f (y)dy.
−π

Theorem 1.2.1 The operator Sn does not change polynomials from T (n) and for
p = 1 or ∞ we have

Sn  p→p ≤ C ln n, n = 2, 3, . . . ,

and for 1 < p < ∞ for all n we have

Sn  p→p ≤ C(p).

This theorem follows from (1.2.3) and the Marcinkiewicz multiplier theorem
(see Theorem A.3.6).
For t ∈ T (n),
n
t(x) = a0 /2 + ∑ (ak cos kx + bk sin kx),
k=1

we call the polynomial t˜ ∈ T (n), where

n
t˜(x) := ∑ (ak sin kx − bk cos kx)
k=1

the polynomial conjugate to t.

Corollary 1.2.2 For 1 < p < ∞ and all n we have

t˜ p ≤ C(p)t p .

Proof Let t ∈ T (n). It is not difficult to see that t˜ = t ∗ Dn , where

n
Dn (x) := 2 ∑ sin kx.
k=1

Clearly, it suffices to consider the case of odd n. Let this be the case and set m :=
(n + 1)/2, l := (n − 1)/2. Representing D̃n (x) in the form
n −1
1 1  imx 
D̃n (x) = ∑ eikx − ∑ eikx = e Dl (x) − e−imx Dl (x) ,
i k=1 k=−n i

we obtain the corollary.

A trigonometric conjugate operator maps a function f (x) to a function

∑(sign k) fˆ(k)eikx .
k
 1.2 Trigonometric Polynomials 9

The Marcinkiewicz multiplier theorem A.3.6 implies that this operator is bounded
as an operator from L p to L p for 1 < p < ∞. We denote by f˜ the conjugate function.

1.2.2 The Fejér Kernel of Order n − 1

The classical univariate Fejér kernel of order n − 1 is defined as follows:
n−1  
Kn−1 (x) := n−1 ∑ Dm (x) = ∑ 1 − |m|/n eimx
m=0 |m|≤n
 2
sin(nx/2)
=  2 .
n sin(x/2)
The Fejér kernel is an even nonnegative trigonometric polynomial in T (n − 1)
with majorant
   
Kn−1 (x) = Kn−1 (x) ≤ min n, π 2 /(nx2 ) , |x| ≤ π . (1.2.10)
From the obvious relations
Kn−1 1 = 1, Kn−1 ∞ = n
and the inequality, see (A.1.6),
1/q 1−1/q
 f q ≤  f 1  f ∞
we get in the same way as we obtained (1.2.9),
Cn1−1/q ≤ Kn−1 q ≤ n1−1/q , 1 ≤ q ≤ ∞. (1.2.11)

1.2.3 The de la Vallée Poussin Kernels

The classical univariate de la Vallée Poussin kernel with parameters m, n is defined
as follows:
n−1
Vm,n (x) := (n − m)−1 ∑ Dl (x), n > m.
l=m

It is convenient to represent this kernel in terms of Fejér kernels:

 
Vm,n (x) = (n − m)−1 nKn−1 (x) − mKm−1 (x)
  2 −1
= (cos mx − cos nx) 2(n − m) sin(x/2) .
The de la Vallée Poussin kernels Vm,n are even trigonometric polynomials of order
n − 1 with majorant
    
Vm,n (x) ≤ C min n, 1/|x|, 1/ (n − m)x2 ) , |x| ≤ π . (1.2.12)
10 Approximation of Univariate Functions

Relation (1.2.12) implies the estimate

 
Vm,n 1 ≤ C ln 1 + n/(n − m) .
We often use the de la Vallée Poussin kernel with n = 2m and denote it by
Vm (x) := Vm,2m (x), m ≥ 1, V0 (x) := 1.
Then for m ≥ 1 we have
Vm = 2K2m−1 − Km−1 ,
which, with the properties of Kn , implies
Vm 1 ≤ 3. (1.2.13)
In addition,
Vm ∞ ≤ 3m.
Consequently, in the same way as above, see (1.2.9) and (1.2.11), we get
Vm q ≍ m1−1/q , 1 ≤ q ≤ ∞. (1.2.14)
Denote
x(l) := π l/2m, l = 1, . . . , 4m.
Then, analogously to (1.2.4), for each t ∈ T (m) we have
4m    
t(x) = (4m)−1 ∑ t x(l) Vm x − x(l) . (1.2.15)
l=1

The operator Vm defined on L1 by the formula

Vm ( f ) := f ∗ Vm
is called the de la Vallée Poussin operator.
The following theorem is a corollary of the definition of the kernels Vm and the
relation (1.2.13).
Theorem 1.2.3 The operator Vm does not change polynomials from T (m), and
for all 1 ≤ p ≤ ∞ we have
Vm  p→p ≤ 3, m = 1, 2, . . .
In addition, we formulate two properties of the de la Vallée Poussin kernels.
(1) Relation (1.2.12) with n = 2m implies the inequality
   
Vm (x) ≤ C min m, 1/(mx2 ) , |x| ≤ π .
It is easy to derive from this inequality the following property.
 1.2 Trigonometric Polynomials 11

(2) For h satisfying the condition C1 ≤ mh ≤ C2 we have

 
∑ Vm (x − lh) ≤ Cm.
0≤l≤2π /h

We remark that property (2) is valid for the Fejér kernel Km .

1.2.4 The Jackson Kernel

The classical univariate Jackson kernel with parameters n, a is defined as follows:
2a
−1 sin(nx/2)
Jna (x) := γa,n , a ∈ N,
sin(x/2)
where γa,n is selected in such a way that

Jna 1 = 1. (1.2.16)

Let us estimate γa,n from below. We have

 π 2a
sin(nx/2)
γa,n = (2π )−1 dx
−π sin(x/2)
 π /n 2a
−1 nx/π
≥π dx ≥ Cn2a−1 . (1.2.17)
0 x/2
The Jackson kernel is an even nonnegative trigonometric polynomial of order
a(n − 1). It follows from (1.2.17) that

Jna (x) ≤ C min(n, n1−2a x−2a ), |x| ≤ π . (1.2.18)

Relation (1.2.18) implies that for 0 ≤ r < 2a − 1,

 π
Jna (x)xr dx ≤ C(r)n−r . (1.2.19)
0

1.2.5 The Rudin–Shapiro Polynomials

We define recursively pairs of trigonometric polynomials Pj (x) and Q j (x) of order
2 j − 1:

P0 := Q0 := 1,
i2 j x j
Pj+1 (x) := Pj (x) + e Q j (x), Q j+1 (x) := Pj (x) − ei2 x Q j (x).
12 Approximation of Univariate Functions

Then at each point x we have

j j
|Pj+1 |2 + |Q j+1 |2 = (Pj + ei2 x Q j )(P j + e−i2 x Q j )
j j
+ (Pj − ei2 x Q j )(P j − e−i2 x Q j )

= 2 |Pj |2 + |Q j |2 ).

Therefore, for all x

   
Pj (x)2 + Q j (x)2 = 2 j+1 .

Thus, for example,

Pn ∞ ≤ 2(n+1)/2 . (1.2.20)

It is clear from the definition of the polynomials Pn that

2n −1
Pn (x) = ∑ εk eikx , εk = ±1, ε0 = 1.
k=0

Let N be a natural number and

m
N= ∑ 2n , j
n1 > n2 > · · · > nm ≥ 0,
j=1

its binary representation. We set

m n j−1
n
R′N (x) := Pn1 (x) + ∑ Pn j (x)ei(2 1 +···+2 )x
,
j=2

RN (x) := R′N (x) + R′N (−x) − 1.

Then RN (x) has the form

RN (x) = ∑ εk eikx , εk = ±1,

|k|<N

and for this polynomial the estimate

RN ∞ ≤ CN 1/2 (1.2.21)

holds.

1.2.6 A Modification of the Fejér Kernel

We consider the polynomials
 1/2 imx
Gn (x) := ∑ 1 − |m|/n e .
|m|<n
 1.2 Trigonometric Polynomials 13

These are even trigonometric polynomials of order n − 1 with the properties

 2π
(2π )−1 Gn (x − a)Gn (x − b)dx = Kn−1 (a − b), (1.2.22)
0

Gn 1 ≤ C. (1.2.23)

Relation (1.2.22) is obvious. It implies that the system of polynomials

Gn (x − 2π l/n), l = 1, . . . , n is an orthogonal system in T (n − 1).
Let us prove the relation (1.2.23). Denote
 1/2
φ (u) := 1 − |u| , |u| ≤ 1.

Then we have on [−1, 1]

φ (u) = ∑ al eiπ lx
l

and it is not hard to prove that

 −3/2
|al | ≤ C |l| + 1 . (1.2.24)

Further,

Gn (x) = ∑ φ (m/n)eimx
|m|≤n

= ∑ al ∑ eim(x+π l/n) = ∑ al Dn (x + π l/n). (1.2.25)

l |m|≤n l

Let us consider the function

gn,l (x) := Dn (x + π l/n) − (−1)l Dn (x).

Using the representation (1.2.1) one can obtain the estimate

 
gn,l 1 ≤ C ln |l| + 2 . (1.2.26)

Further, owing to the equality

φ (1) = ∑ al (−1)l = 0,
l

relation (1.2.25) can be rewritten in the form

Gn (x) = ∑ al gn,l (x).

l

Using relations (1.2.24) and (1.2.26) we then obtain relation (1.2.23).

14 Approximation of Univariate Functions

1.2.7 A Generalization of the Rudin–Shapiro Polynomials

The trigonometric polynomials considered above (see §§1.2.1–1.2.6) were con-
structively obtained: either they are given by a formula (§§1.2.1–1.2.4, 1.2.6) or
a method of construction is supplied (§1.2.5). In this subsection we formulate a
theorem that proves the existence of polynomials with given properties.

Theorem 1.2.4 Let ε > 0, and let a subspace Ψ ⊂ T (n) be such that dim Ψ ≥
ε (2n + 1). Then there exists a t ∈ Ψ such that

t∞ = 1,

and
t2 ≥ C(ε ) > 0.

An analogous statement is valid for the multivariable trigonometric polynomials,

will be proved in Chapter 3 (see Theorem 3.2.1).
We remark that the polynomial t from Theorem 1.2.4, by virtue of the inequality

t22 ≤ t1 t∞ ,

satisfies the condition

t1 ≥ C(ε )2 > 0. (1.2.27)

1.2.8 An Application of the Gaussian Sums

In this subsection we construct polynomials that we will use in studying linear
widths. This construction is based on properties of the Gaussian sums:
q
2
S(q, l) := ∑ ei2π l j /q ,
j=1

where q is a natural number and l, q are coprime; that is, (l, q) = 1. We confine
ourselves to the case where q is an odd prime.

Theorem 1.2.5 Let q > 2 be a prime, l = 0 an integer, and k an integer. Then, for
q
2
S(q, l, k) := ∑ ei2π (l j +k j)/q ,
j=1

the following equality is true:

 
S(q, l, k) = q1/2 .
 1.2 Trigonometric Polynomials 15

Proof We first consider the case k = 0. Note that the quantity S(q, l) does not
change if we sum over the complete system of remainders modulo q instead of the
segment [1, q]. Consequently, for any integer h,
q
2
S(q, l) = ∑ ei2π l( j+h) /q . (1.2.28)
j=1

Further,
  q q
S(q, l)2 = ∑ e−i2π lh
2 /q 2
∑ ei2π l j /q .
h=1 j=1

Using (1.2.28), we see that this is equal to

q q q q
−i2π lh2 /q i2π l( j+h)2 /q 2
∑e ∑e = ∑ ∑ ei2π l( j +2 jh)/q . (1.2.29)
h=1 j=1 h=1 j=1

Taking into account that


q
i2π l2 jh/q q for j = q,
∑e =
0 for j ∈ [1, q),
h=1

we get from (1.2.29),

 
S(q, l)2 = q. (1.2.30)

Now let k be nonzero. Since q is a prime different from 2, the numbers 2lb,
b = 1, . . . , q, run through a complete system of remainders modulo q. Consequently,
there is a b such that
2lb ≡ k (mod q).

Then
l j2 + k j ≡ l( j + b)2 − lb2 (mod q)

and, consequently,
   
S(q, l, k) = S(q, l) = q1/2 .

The theorem is proved.

Theorem 1.2.6 Let q be a prime and q = 2a + 1. For any n ∈ [1, a] there is a

trigonometric polynomial tn ∈T (a)  such that only n Fourier coefficients of tn are
nonzero and for all k we have t (k) ≤ 1 and in addition
 ˆ
 
tn (0) ≥ (n + 1)/2, t(2π l/q) ≤ Cq1/2 , l = 1, . . . , 2a.
16 Approximation of Univariate Functions

Proof The proof of this theorem can easily be derived from a deep number the-
oretical result due to Hardy and Littlewood about estimating incomplete Gaussian
sums: for any n ∈ [1, q]
 
 n 
 i2π l j2 /q 
∑ e  ≤ Cq1/2 , (l, q) = 1. (1.2.31)
 j=1 

Indeed, let k j denote the smallest nonnegative remainder of the number j2

modulo q, j = 1, . . . , n, and let

G := {k j − a, j = 1, . . . , n}.

We set
tn (x) := ∑ eikx .
k∈G

Then
   n 
   π
  π 2 /q 
tn (2π l/q) =  ∑ ei2 lk/q  = ∑ ei2 l j ,
   
k∈G j=1

which by (1.2.31) implies the required estimates for tn (2π l/q). The bound
tn (0) = n ≥ (n + 1)/2 is obvious.

For the sake of completeness we will prove Theorem 1.2.6 using Theorem 1.2.5.
Instead of (1.2.31) we prove the inequality
 
   i2π l j2 /q 
 
∑ (1 − | j − a|/n + e  ≤ q1/2 , (l, q) = 1. (1.2.32)
 j 

Let l ∈ [1, q − 1]. Consider the trigonometric polynomial

q−1
2
t(x) := ∑ ei2π l j /q ei( j−a)x .
j=0

Then at the points xk = 2π k/(2a + 1) = 2π k/q we have

 k   
t(x ) = S(q, l, k) = q1/2 , k = 0, . . . , 2a. (1.2.33)

We set
un (x) := t(x) ∗ Kn−1 (x).

Then by (1.2.5),
2a
un (x) = q−1 ∑ t(xk )Kn−1 (x − xk ),
k=0
 1.3 The Bernstein–Nikol’skii Inequalities. The Marcienkiewicz Theorem 17

and, using (1.2.33) we find that

  2a
un (x) ≤ q−1/2 ∑ Kn−1 (x − xk ) = q1/2 . (1.2.34)
k=0

Further,
  2
un (0) = ∑ 1 − | j − a|/n + ei2π l j /q ,
j

where (a)+ := max(a, 0). By (1.2.34) this implies (1.2.32).

Setting
 
tn (x) := ∑ 1 − | j − a|/n + ei(k j −a)x ,
j

where the k j are the same as in the beginning of the proof of this theorem, we get
   
     i2π lk /q     i2π l j2 /q 
tn (2π l/q) = ∑ 1 − | j − a|/n e j
 = ∑ 1 − | j − a|/n + e ,
 j +   j 

which by (1.2.32) implies the conclusion of Theorem 1.2.6, with 2n − 1 nonzero

Fourier coefficients instead of n.

1.3 The Bernstein–Nikol’skii Inequalities. The Marcienkiewicz Theorem

The Bernstein–Nikol’skii inequalities connect the L p -norms of a derivative of some
polynomial with the Lq -norm, 1 ≤ q ≤ p ≤ ∞, of this polynomial. We obtain here
inequalities for a derivative that is slightly more general than the Weyl fractional
derivative. We first make some auxiliary considerations.
For a sequence {aν }ν∞=0 we write
Δaν := aν − aν +1 ; Δ2 aν := Δ(Δaν ) = aν − 2aν +1 + aν +2 .
Theorem 1.3.1 We have
 
 π  n  n
 
(π )−1 a0 /2 + ∑ aν cos ν x dx ≤ ∑ (ν + 1)|Δ2 aν |.
−π  ν =1
 ν =0

Proof Applying twice the Abel transformation (see (A.1.18) in the Appendix)
with aν = 0 for ν > n, we obtain
n n
t(x) := a0 + ∑ aν 2 cos ν x = ∑ Dν (x)Δaν
ν =1 ν =0
n ν n
= ∑ ∑ Dµ (x) Δ2 aν = ∑ (ν + 1)Kν (x)Δ2 aν . (1.3.1)
ν =0 µ =0 ν =0
18 Approximation of Univariate Functions

From (1.3.1), using Kν 1 = 1 we find

n
t1 ≤ ∑ (ν + 1)|Δ2aν |,
ν =0

as required.

1.3.1 The Bernstein inequality

We first prove the Bernstein inequality. Let us consider the following special
trigonometric polynomials. Let s be a nonnegative integer. We define
A0 (x) := 1, A1 (x) := V1 (x) − 1, As (x) := V2s−1 (x) − V2s−2 (x), s ≥ 2,
where the Vm are the de la Vallée Poussin kernels (see §1.2.3). Then As ∈ T (2s )
and by (1.2.13),
As 1 ≤ 6. (1.3.2)
Let r ≥ 0 and α be real numbers. We consider the polynomials
n
Vnr (x, α ) := 1 + 2 ∑ kr cos(kx + απ /2)
k=1
2n−1  
+2 ∑ kr 1 − (k − n)/n cos(kx + απ /2).
k=n+1

Let us prove that, for all r > 0 and α ,

 r 
Vn (x, α ) ≤ C(r)nr , n = 1, 2, . . . (1.3.3)
1

Since for an arbitrary α

   
Vnr (x, α ) − 1 = Vnr (x, 0) − 1 cos(απ /2) + Vnr (x, 1) − 1 sin(απ /2),
it suffices to prove (1.3.3) for α = 0 and for α = 1. We first consider the case
α = 0. Let vk be the Fourier cosine coefficients of the function Vnr (x, 0). Then, by
Theorem 1.3.1,
 r  2n−1
Vn (x, 0) ≤ ∑ (k + 1)|Δ2 vk |. (1.3.4)
1
k=0

It is easy to see that, for 1 ≤ k ≤ n − 2,

|Δ2 vk | ≤ C(r)kr−2 . (1.3.5)
By the identity
Δ2 (ak bk ) = (Δ2 ak )bk + 2(Δak+1 )(Δbk ) + ak+2 (Δ2 bk )
 1.3 The Bernstein–Nikol’skii Inequalities. The Marcienkiewicz Theorem 19

with ak = kr and bk = 1 − (k − n)/n, we see that the inequality (1.3.5) will be valid
for n ≤ k ≤ 2n − 3 too. For the remaining values of k = 0 we have
|Δ2 vk | ≤ |Δvk | + |Δvk+1 | ≤ C(r)nr−1 . (1.3.6)
From the inequality |Δ2 v0 | ≤ C(r) and relations (1.3.4)–(1.3.6) we get the rela-
tion (1.3.3) for r > 0 and α = 0.
Let α = 1 and let A˜s (x) denote the polynomial which is the trigonometric con-
jugate to As (x), which means that in the expression for As (x) the functions cos kx
are substituted by sin kx. We prove that
A˜s 1 ≤ C. (1.3.7)
Clearly, it suffices to consider s ≥ 3. It is not difficult to see that the equality
   s−1 s−3 
A˜s (x) = 2 Im As (x) ∗ 4K2s−1 −1 (x) − 3K2s−1 −2s−3 −1 (x) ei(2 +2 )x ,
holds. From this equality, by virtue of the Young inequality with p = q = a = 1
(see A.1.16)) and the properties of the functions Kn and As , we obtain (1.3.7).
Further, for n = 2m , we have
 
Vnr (x, 1) − 1 = V2nr (x, 0) − 1 ∗ Vn0 (x, 1)
m+1 m+1
=− ∑ V2nr (x, 0) ∗ A˜s(x) = − ∑ V2r (x, 0) ∗ A˜s (x).
s (1.3.8)
s=1 s=1

From (1.3.8) by means of the Young inequality and using (1.3.7) and relation
(1.3.3), which has been proved for α = 0, we get
 r m+1
Vn (x, α )1 ≤ C(r) ∑ 2rs ≤ C(r)nr . (1.3.9)
s=0

Now let 2m−1 ≤ n < 2m ; then

Vnr (x, 1) = V2rm+1 (x, 1) ∗ Vn (x),
which by (1.3.9) and the Young inequality gives the required estimate for all n.
Relation (1.3.3) is proved.
We define the operator Dαr , r ≥ 0, α ∈ R, on the set of trigonometric polynomials
as follows. Let f ∈ T (n); then
Dαr f := f (r) (x, α ) := f (x) ∗ Vnr (x, α ), (1.3.10)
and f (r) (x, α ) is called the (r, α ) derivative. It is clear that for f (x) such that
fˆ(0) = 0 we have for natural numbers r,
dr
Drr f = f.
dxr
20 Approximation of Univariate Functions

The operator Dαr is defined in such a way that it has an inverse for each T (n).
This property distinguishes Dαr from the differential operator and it will be conve-
nient for us. On the other hand it is clear that
dr f
= Drr f − fˆ(0).
dxr
Theorem 1.3.2 For any t ∈ T (n) we have, for r > 0, α ∈ R, 1 ≤ p ≤ ∞,
 (r) 
t (x, α ) ≤ C(r)nr t p , n = 1, 2, . . .
p

Proof By the definition (1.3.10),

t (r) (x, α ) = t(x) ∗ Vnr (x, α ).
Therefore, by the Young inequality (A.1.16) with p = q, a = 1 for all 1 ≤ p ≤ ∞
and r we have
 (r)   
t (x, α ) ≤ t p Vnr (x, α ) .
p 1

To conclude the proof we just use inequality (1.3.3).

Let us discuss the case r = 0, which is excluded from Theorem 1.3.2. In the case
where r = 0 and α is an even integer we have
 (0)  
t (x, α )| = t(x)

and, consequently,
 (0) 
t (x, α ) = t p , 1 ≤ p ≤ ∞. (1.3.11)
p

To investigate the general case it suffices to study the trigonometric conjugate

operator. Theorem 1.2.1 and its corollary show that for all α and 1 < p < ∞ the
inequality
 (0) 
t (x, α ) ≤ C(p)t p

holds.
It remains to consider the cases p = 1, ∞. It is sufficient to consider α = 1.
We have for t ∈ T (n),
t (0) (x, 1) = tˆ(0) − t˜(x) = tˆ(0) − t(x) ∗ D2n+1 (x).
Further,
2n+1
D̃2n+1 (x) = 2 ∑ sin kx = 2 Im Dn (x)ei(n+1)x ;
k=1

consequently,
D2n+1 1 ≤ C ln(n + 2).
 1.3 The Bernstein–Nikol’skii Inequalities. The Marcienkiewicz Theorem 21

Thus, for t ∈ T (n),

 (0) 
t (x, 1) ≤ C ln(n + 2)t p , p = 1, ∞. (1.3.12)
p

The relation (1.3.11) with α = 0 and (1.3.12) imply for all α the inequality
 (0) 
t (x, α ) ≤ C ln(n + 2)t p , p = 1, ∞. (1.3.13)
p

Remark 1.3.3 We have the relation

 
sup t (0) (x, 1) p t p ≍ ln(n + 2), p = 1, ∞.
t∈T (n)

The upper estimate follows from (1.3.12). Let us prove the lower estimate. We
first consider the case p = ∞. Let f (x) = (π − x)/2, 0 < x < 2π , be a 2π -periodic
function; then
∞
f (x) = ∑ (sin kx)/k.
k=1

Let m = [n/2]. Then

t(x) := f (x) ∗ Vm (x)

has the following properties: t ∈ T (n),

m
t∞ ≤ 3π /2, t (0) (0, 1) ≥ ∑ 1/k ≥ C ln(m + 2), (1.3.14)
k=1

which imply the required lower estimate in the case p = ∞.

Let p = 1 and m = [n/2]. Then the function Vm ∈ T (n) has the following
properties:

Vm 1 ≤ 3, (1.3.15)
 (0) 
Vm (x, 1) ≥ C ln(m + 2). (1.3.16)
1

Let us prove (1.3.16). For t we have from the above consideration for p = ∞,
(0)  (0) 
σ = |Vm (x, 1),t| ≤ Vm (x, 1)1 t∞ (1.3.17)

and
m
σ≥ ∑ 1/k ≥ C ln(m + 2). (1.3.18)
k=1

From relations (1.3.14), (1.3.17) and (1.3.18) we obtain (1.3.16). Then (1.3.15) and
(1.3.16) give the required lower estimate for p = 1.
22 Approximation of Univariate Functions

1.3.2 The Nikol’skii Inequality

Let us now prove the Nikol’skii inequality.
Theorem 1.3.4 For any t ∈ T (n), n > 0, we have the inequality
t p ≤ Cn1/q−1/p tq , 1 ≤ q < p ≤ ∞.
Proof First let p = ∞; then
t = t ∗ Vn
and by the Hölder inequality (A.1.1) we have
t∞ ≤ tq Vn q′ ,
which, by (1.2.14), implies that
t∞ ≤ Ctq n1/q . (1.3.19)
Further, let q < p < ∞. Then by (A.1.6),
q/p 1−q/p
t p ≤ tq t∞ . (1.3.20)
The theorem follows from relations (1.3.19) and (1.3.20).
We now formulate a corollary of Theorems 1.3.2 and 1.3.4.
Corollary 1.3.5 (The Bernstein–Nikol’skii inequality) For t ∈ T (n) and arbi-
trary r > 0, α , 1 ≤ q ≤ p ≤ ∞, we have the inequality
 (r) 
t (x, α ) ≤ C(r)nr+1/q−1/p tq , n = 1, 2, . . .
p

1.3.3 The Marcinkiewicz Theorem

The set T (n) of trigonometric polynomials is a space of dimension 2n
 + 1.Each
polynomial t ∈ T (n) is uniquely defined by its Fourier coefficients tˆ(k) |k|≤n ,
and by the Parseval identity we have
 2
t22 = ∑ tˆ(k) , (1.3.21)
|k|≤n

which means that the set T (n) as a subspace of L2 is isomorphic to ℓ22n+1 . Relation
(1.2.6) shows that a similar isomorphism can be
 set up in another way: by mapping
2n
a polynomial t ∈ T (n) to the vector m(t) := t(xl ) l=0 of its values at the points
xl := 2π l/(2n + 1), l = 0, . . . , 2n.
Relation (1.2.6) gives
 
t2 = (2n + 1)−1/2 m(t)2 .
The following statement is the Marcinkiewicz theorem.
 1.3 The Bernstein–Nikol’skii Inequalities. The Marcienkiewicz Theorem 23

Theorem 1.3.6 Let 1 < p < ∞; then for t ∈ T (n), n > 0, we have the relation
 
C1 (p)t p ≤ n−1/p m(t) p ≤ C2 (p)t p .

Proof We first prove a lemma.

Lemma 1.3.7 Let 1 ≤ p ≤ ∞; then, for n > 0,

 
 2n 
 
 ∑ al Vn (x − xl ) ≤ Cn1−1/p aℓ2n+1 , a := (a0 , . . . , a2n ).
l=0  p
p

Proof Let V be an operator on ℓ2n+1

p defined as follows:
2n
V (a) := ∑ al Vn (x − xl ).
l=0

It is obvious that (see (1.2.13))

V ℓ2n+1 →L1 ≤ 3. (1.3.22)

1

Using the estimate (see (1.2.12))

  
Vn (x) ≤ C min n, (nx2 )−1 )

it is not hard to prove that

V ℓ2n+1
∞ →L∞ ≤ Cn. (1.3.23)

From relations (1.3.22) and (1.3.23), using the Riesz–Torin theorem (see Theorem
A.3.2) we find that
1−1/p
V ℓ2n+1
p →L p ≤ Cn ,

which implies the lemma.

We now continue the proof of Theorem 1.3.6. Let Sn be the operator that takes
the partial Fourier sum of order n. Using Theorem 1.2.1 we derive from Lemma
1.3.7 the upper estimate (the first inequality in Theorem 1.3.6):
2n
t(x) = (2n + 1)−1 ∑ t(xl )Dn (x − xl )
l=0
2n
= Sn (2n + 1)−1 ∑ t(xl )Vn (x − xl ) .
l=0

Consequently,
 
t p ≤ C(p)n−1/p m(t) p .
24 Approximation of Univariate Functions

We now prove the lower estimate (the second inequality in Theorem 1.3.6) for
1 ≤ p < ∞. We have
  2n   2n  
m(t) p = ∑ t(xl ) p = ∑ t(xl )εl t(xl ) p−1
p
l=0 l=0
 2π 2n   p−1
= (2π )−1 t(x) ∑ εl t(xl ) Vn (x − xl )dx
0 l=0
 
 2n   p−1 
 
≤ t p  ∑ εl t(x ) Vn (x − x ) ,
l l
l=0  ′
p

using Lemma 1.3.7 we see that the last expression is

  p−1
≤ Ct p n1/p m(t) , p

which implies the required lower estimate and the theorem is proved.

Remark 1.3.8 In the proof of Theorem 1.3.6 we also proved the inequality

m(t)1 ≤ Cnt1 .

We now prove a statement that is analogous to Theorem 1.3.6 but, in contrast to

it, includes the cases p = 1 and p = ∞. Instead of the vector m(t) we now consider
the vector
   
M(t) := t x(1) , . . . ,t x(4n) , x(l) := π l/(2n), l = 1, . . . , 4n.

Theorem 1.3.9 For an arbitrary t ∈ T (n), n > 0, 1 ≤ p ≤ ∞, we have

 
C1 t p ≤ n−1/p M(t) p ≤ C2 t p .

Proof In the same way as for Lemma 1.3.7 one can prove:

Lemma 1.3.10 Let 1 ≤ p ≤ ∞, then, for n > 0,

 
 4n  
 
 ∑ al Vn x − x(l)  ≤ Cn1−1/p aℓ4n .
l=1  p
p

Lemma 1.3.10 with a = M(t) and relation (1.2.15) implies the upper estimate
 
t p ≤ Cn−1/p M(t) p .

The corresponding lower estimate for 1 ≤ p < ∞ can be proved in the same way as
above for m(t), substituting xl by x(l).
The lower estimate for p = ∞ is obvious.
 1.4 Approximation of Functions in the Classes Wq,r α and Hqr 25

1.4 Approximation of Functions in the Classes Wq,r α and Hqr

1.4.1 Some Properties of the Bernoulli Kernels
For r > 0 and α ∈ R the functions
∞
Fr (x, α ) = 1 + 2 ∑ k−r cos(kx − απ /2)
k=1

are called Bernoulli kernels.

We define the following operator in the space L1 ,
 2π
−1
(Iαr φ )(x) := (2π ) Fr (x − y, α )φ (y)dy. (1.4.1)
0

Let us prove that the definition of this operator is reasonable. To establish this it
suffices to prove that Fr ∈ L1 .

Theorem 1.4.1 For r > 0, α ∈ R we have

Fr ∈ L1 , En (Fr )1 ≤ C(r)(n + 1)−r , n = 0, 1, . . .

Proof Let us consider the functions

2s
f r
s (x, α ) := As (x) ∗ 1 + 2 ∑ k−r cos(kx − απ /2) ,
k=1

where the As are defined in §1.3.

We first consider the case α = 0. Using Theorem 1.3.1 in the same way as in the
proof of inequality (1.3.3) we get
 r 
 f s (x, 0) ≤ C(r)2−rs . (1.4.2)
1

Further,
fsr (x, α ) = Dr−α fs2r (x, 0),

and, consequently, from (1.4.2) and Theorem 1.3.2 we find that

 r 
 f s (x, α ) ≤ C(r)2−rs . (1.4.3)
1

Thus the series

∞
∑ f rs (x, α )
s=0

converges in L1 to some function f (x) and

 ∞ 
 
 ∑ f rs (x, α ) ≤ C(r)2−rm . (1.4.4)
 
s=m 1
26 Approximation of Univariate Functions

From the definition of the function f rs (x, α ) we get

n
Sn ( f ) = 1 + 2 ∑ k−r cos(kx − απ /2)
k=1
and
  ∞   
 f − Sn ( f )  ≤ ∑  f rs (x, α ) − Sn f rs (x, α ) 1
1
s=0
 r 
≤ ∑  f s − Sn ( f rs ) ≤ C ln(n + 2) ∑  f rs 1
1
s:2s >n 2s >n
−r
≤ C(r)n ln(n + 2). (1.4.5)
Here we have used Theorem 1.2.1 and relation (1.4.3). Relation (1.4.5) shows that
the series defining the function Fr (x, α ) converges in L1 to f (x). The first part
of the theorem is proved. The second part of the theorem follows from relation
(1.4.4).
We now proceed to formulate some properties of the operators Drα and Iαr . From
the equality (φ ∈ L1 )
 2π  2π    
−1
π φ (u) cos k(y − u) + απ /2 cos k(x − y) + β π /2 dy du
0 0
 2π 
= φ (u) cos k(x − u) + (α + β )π /2)du,
0
which is valid for any nonzero k, the equalities
Drα11 Drα22 = Dαr11+r
+ α2 ,
2
(1.4.6)
Iαr11 Iαr22 = Iαr11+r
+ α2 ,
2
(1.4.7)
Dαr Iαr = Iαr Drα = I (1.4.8)
follow (we assume that the operators act on a set of trigonometric polynomials).
Denote by Wq,r α B, r > 0, α ∈ R, 1 ≤ q ≤ ∞, the class of functions f (x) repre-
sentable in the form
f = Iαr φ , φ q ≤ B. (1.4.9)
For such functions, with some q and B. we define (see (1.4.8))
Dαr f = φ .
Let 1 < q < p < ∞, β := 1/q−1/p. From Corollary A.3.8 of the Hardy–Littlewood
inequality (see the Appendix) and the boundedness of the trigonometric conjugate
operator as an operator from L p to L p for 1 < p < ∞ (see Corollary 1.2.2), it follows
that
β
Iα q→p ≤ C(q, p). (1.4.10)
 1.4 Approximation of Functions in the Classes Wq,r α and Hqr 27

Relations (1.4.7) and (1.4.10) imply the following embedding theorem.

Theorem 1.4.2 Let 1 < q < p < ∞, β = 1/q − 1/p, r > β ; then
r−β
Wq,r α1 ⊂ Wp,α2 B, α1 , α2 ∈ R.

1.4.2 Approximation for Smoothness Classes

Let us define the classes Hqr B, r > 0, 1 ≤ q ≤ ∞ as follows:
 
 a 
Hq B := f ∈ Lq :  f q ≤ B, Δt f (x)q ≤ B|t| , a = [r] + 1 ,
r r

Δt f (x) := f (x) − f (x + t), Δta := (Δt )a .

For the case B = 1 we simply write Hqr := Hqr 1, i.e., we drop the constant B.
Let us study these classes from the point of view of their approximation by
trigonometric polynomials.

Theorem 1.4.3 Let r > 0, 1 ≤ q ≤ ∞, then

En (Hqr )q ≍ (n + 1)−r , n = 0, 1, . . .

Proof Let us prove the upper estimate. Clearly, it suffices to consider the case
n > 0. Let f ∈ Hqr . We consider (see §1.2.4)
 π 
t(x) := (2π )−1 f (x) − Δay f (x) Jna (y)dy.
−π

Then t ∈ T (an) and

 π
f (x) − t(x) = (2π )−1 Δay f (x)Jna (y)dy.
−π

By a generalization of the Minkowskii inequality, (A.1.9), we have

 π 
 f − tq ≤ (2π )−1 Δay f (x) Jna (y)dy,
q
−π

which by the definition of the class Hqr and relation (1.2.19) implies that

 f − tq ≤ C(r)n−r .

The upper estimate is proved.

We now prove the lower estimate. We construct functions which will be used in
the proof of the more general Theorem 1.4.9. Let n > 0 be given and s be such that

4n ≤ 2s ≤ 8n.
28 Approximation of Univariate Functions

We consider
f (x) := 2−(r+1−1/q)s As (x) (1.4.11)
and remark that to prove the theorem it suffices to consider the simpler function
f (x) = (n + 1)−r ei(n+1)x . Then, for any t ∈ T (n), we have on the one hand
 f − t, As  =  f , As  = 2−(r+1−1/q)s As 22 ≥ C2−(r−1/q)s . (1.4.12)
On the other hand using the definition of As and (1.2.14) we get
 f − t, As  ≤  f − tq As q′ ≤ C2s/q  f − tq . (1.4.13)
From relations (1.4.12) and (1.4.13) we obtain
En ( f )q ≥ C2−rs ≥ Cn−r .
To show that f ∈ Hqr B, we prove the following auxiliary statement.
Lemma 1.4.4 Let g(x) be an a-times continuously differentiable 2π -periodic
function. Then for all 1 ≤ q ≤ ∞ we have
 a   
Δy g(x) ≤ | y|a g(a) (x) .
q q

Proof Clearly it suffices to consider the case a = 1. We have

 x+y   y 
     
Δy g(x) =  g (u)du =  g (x + u)du
′   ′ ′
q  x 0
 ≤ |y|g q ,
q q

as required.
From (1.4.11), (1.2.14), and the Bernstein inequality (Theorem 1.3.2) we get
 f (a) q ≤ C(a)2(a−r)s . (1.4.14)
Using Lemma 1.4.4 and the simple inequality
 a 
Δy f (x) ≤ 2a  f q ,
q

we obtain
 a   
Δy f (x) ≤ C(a) min |y|a na−r , n−r , (1.4.15)
q

which implies that f ∈ Hqr B with some B that is independent of n, and this proves
the lower estimate.
Let us now prove a representation theorem for the class Hqr B. Let
As ( f ) := As ∗ f
and denote the value of As ( f ) at a point x by As ( f , x).
 1.4 Approximation of Functions in the Classes Wq,r α and Hqr 29

Theorem 1.4.5 Let f ∈ Lq , 1 ≤ q ≤ ∞,  f q ≤ 1. For Δta f q ≤ |t|r , a = [r] + 1

it is necessary and sufficient that the following conditions be satisfied:
 
As ( f ) ≤ C(r, q)2−rs , s = 0, 1, . . . .
q

(The constants C(r, q) may be different for the cases of necessity and sufficiency.)
Proof

Necessity. Let f ∈ Hqr ; then for any ts ∈ T (2s−2 ), s ≥ 2 we have

A s ( f ) = A s ( f − ts )
and

As ( f )q ≤ As 1  f − ts q .

Applying Theorem 1.4.3 and using relation (1.3.2) we get

 
As ( f ) ≤ C(r, q)2−rs .
q

Sufficiency. Let
 
As ( f ) ≤ γ 2−rs , (1.4.16)
q

then using Corollary 2.2.7 we get

∞
f= ∑ As ( f ),
s=0

in the sense of convergence in Lq , and

∞  
Δta f q ≤ ∑ Δta As ( f )q . (1.4.17)
s=1

From Lemma 1.4.4 we find, in the same way as in (1.4.15),

 a     
Δt As ( f ) ≤ C(a)2−rs min 1, |t|2s a . (1.4.18)
q

From (1.4.17) and (1.4.18) we obtain

Δta f q ≤ C(r)γ |t|r ,
which concludes the proof of the theorem if we take γ < 1/C(r).
Denote
δ0 ( f ) := S0 ( f ), δs ( f ) := S2s −1 ( f ) − S2s−1 −1 ( f ), s = 1, 2, . . .
Corollary 1.4.6 In the case 1 < q < ∞ the functions As ( f ) in Theorem 1.4.5 can
be replaced by δs ( f ).
30 Approximation of Univariate Functions

Proof For 1 < q < ∞ the conditions

 
(1) As ( f )q ≤ C(q)2−rs ,
 
(2) δs ( f ) ≤ C(q)2−rs
q

are equivalent for all s. Indeed,

 
As ( f ) = As ∗ δs−1 ( f ) + δs ( f ) ,
 
δs ( f ) = δs As ( f ) + As+1 ( f ) ,
which by (1.3.2) and the boundedness of the operator δs as an operator from Lq
to Lq , 1 < q < ∞ (see Corollary A.3.4) implies the equivalence of conditions (1)
and (2).
Corollary 1.4.7 Let 1 ≤ q ≤ ∞,  f q ≤ 1 and
En ( f )q ≪ (n + 1)−r , n = 0, 1, . . . ;
then f ∈ Hqr B for some B.
Indeed, in the same way as in the proof of the necessity in Theorem 1.4.5 we get

As ( f )q ≪ 2−rs ,

which by Theorem 1.4.5 (regarding the sufficiency) implies that f ∈ Hqr B.

Statements of the type of Theorem 1.4.3 are called direct theorems of approx-
imation theory, and statements of the type of Corollary 1.4.7 are called inverse
theorems of approximation theory.
Theorem 1.4.1 and Corollary 1.4.7 imply that
Fr (x, α ) ∈ H1r B. (1.4.19)
Consequently, for f ∈ Wq,r α we have
 a   
Δt f (x) ≤ Δta Fr (x, α ) Dαr f q ≤ B|t|r ;
q 1

that is, f ∈ Hqr B.

Thus, we have proved that
Wq,r α ⊂ Hqr B. (1.4.20)

Let us prove an embedding theorem for the H classes.

Theorem 1.4.8 Let 1 ≤ q ≤ p ≤ ∞, β := 1/q − 1/p, r > β . We have the inclusion
Hqr ⊂ Hpr−β B
r−β
(in the case p = ∞ this means that for any f ∈ Hqr there is an equivalent g ∈ H∞ B).
 1.4 Approximation of Functions in the Classes Wq,r α and Hqr 31

Proof Let f ∈ Hqr . By Theorem 1.4.5

 
As ( f ) ≤ C(r, q)2−rs .
q

Therefore, by the Nikol’skii inequality (Theorem 1.3.4) we have

 
As ( f ) ≤ C(r, q)2−(r−β )s . (1.4.21)
p

Let g(x) denote the sum of the series ∑∞s=0 As ( f , x) in the sense of convergence in
L p . From Corollary 2.2.7 below it follows that f and g are equivalent. From (1.4.21)
r−β
and the equality As ( f ) = As (g), by Theorem 1.4.5 we obtain g ∈ Hp B.
The theorem is proved.
With the aid of Theorem 1.4.8 we can prove the following statement.
Theorem 1.4.9 Let 1 ≤ q, p ≤ ∞, r > (1/q − 1/p)+ . Then
En (Wq,r α ) p ≍ En (Hqr ) p ≍ n−r+(1/q−1/p)+ .
Proof By relation (1.4.20) it suffices to prove the upper estimate for the H classes
and the lower estimate for the W classes. We first prove the upper estimate. Let
1 ≤ q ≤ p ≤ ∞. Then Theorems 1.4.8 and 1.4.3 give
En (Hqr ) p ≪ n−r+1/q−1/p . (1.4.22)
For 1 ≤ p < q ≤ ∞ we have, by the monotonicity of the L p -norms and Theorem
1.4.3,
En (Hqr ) p ≤ En (Hqr )q ≪ n−r .
From this and relation (1.4.22) the required upper estimates follow.
Let us prove the lower estimate. Let n and s be the same as in the proof of the
lower estimate in Theorem 1.4.3 and let f be defined by (1.4.11). Then by the
Bernstein inequality,
Drα f q ≤ C(r),
and f ∈ Wq,r α C(r).
Let 1 ≤ q ≤ p ≤ ∞. From relation (1.4.12) and relation (1.4.13) with p instead
of q we get
En ( f ) p ≥ Cn−r+1/q−1/p . (1.4.23)
For 1 ≤ p ≤ q ≤ ∞ it suffices to consider as an example
f (x) = 2(n + 1)−r cos(n + 1)x.
r and, for any t ∈ T (n),
Then f ∈ W∞, α

σ =  f (x) − t(x), cos(n + 1)x = (n + 1)−r , σ ≤  f − t1 ,

32 Approximation of Univariate Functions

which implies the estimate

r −r
En (W∞, α )1 ≥ (n + 1) . (1.4.24)
The required lower estimates follow from (1.4.23) and (1.4.24) and the theorem is
proved.
Remark 1.4.10 Theorem 1.2.3 implies that for any f ∈ L p the de la Vallée
Poussin inequality holds:
 
 f −Vn ( f ) ≤ 4En ( f ) p , 1 ≤ p ≤ ∞. (1.4.25)
p

This inequality and Theorem 1.4.9 show that, for all 1 ≤ q, p ≤ ∞,

 
Vn (Hqr ) p := sup  f −Vn ( f ) p ≍ E2n (Hqr ) p , (1.4.26)
f ∈Hqr

and an analogous relation is valid for the W classes.

Thus, for the classes Wq,r α and Hqr there exist linear methods giving an approxi-
mation of the same order as the best approximation.
Remark 1.4.11 From Theorem 1.2.1 it follows that for all 1 < p < ∞ and f ∈ L p ,
 
 f − Sn ( f ) ≤ C(p)En ( f ) p . (1.4.27)
p

Consequently, if we are interested only in the dependence of the approximation

of a function f ∈ L p on n then it suffices, in the case 1 < p < ∞, to consider the
simplest method of approximation, namely, the Fourier method.
This remains true for the classes Wq,r α and Hqr for all 1 ≤ q, p ≤ ∞, excepting the
cases q = p = 1 and q = p = ∞. For the function class F let us denote
 
Sn (F) p := sup f − Sn ( f ) p .
f ∈F

Theorem 1.4.12 Let 1 ≤ q, p ≤ ∞, (q, p) = (1, 1) or (∞, ∞), and r > (1/q −
1/p)+ . Then
Sn (Wq,r α ) p ≍ Sn (Hqr ) p ≍ n−r+(1/q−1/p)+ .
Proof In the case 1 < p < ∞ the theorem follows from Theorem 1.4.9 and relation
(1.4.27). It remains to consider the cases p = 1, q > 1 and 1 ≤ q < p = ∞. In the
case p = 1, q > 1 we have
Sn (Hqr )1 ≤ Sn (Hqr∗ )q∗ ≪ n−r ,
where q∗ = min(q, 2).
Now let 1 ≤ q < p = ∞. In the case 1 ≤ q < 2, by Theorem 1.4.8 we have
r−(1/q−1/2)
Hqr ⊂ H2 B,
 1.4 Approximation of Functions in the Classes Wq,r α and Hqr 33

which indicates that it suffices to consider the case 2 ≤ q < ∞. In this case by
Theorem 1.2.1 and Corollary 1.4.6 we have for s > sn , where sn is such that 2sn −1 ≤
n < 2sn ,
 
δs ( f ) ≤ C(r, q)2−rs ,
q
  
δs ( f ) − Sn δs ( f )  ≤ C(r, q)2−rsn .
n n q

From these inequalities, using the Nikol’skii inequality, we get

      
 f − Sn ( f ) ≤ δs ( f ) − Sn δs ( f )  + ∑ δs ( f )
∞ n n ∞ ∞
s>sn
−(r−1/q)s
≤ C(r, q) ∑2 ≤ C(r, q)n−r+1/q ,
s≥sn

which concludes the proof of the theorem.

We proceed to the cases q = p = 1 or ∞, which were excluded in Theorem 1.4.12.

For these cases we obtain from Theorem 1.2.1 the following Lebesgue inequality:
for f ∈ L p , p = 1, or ∞,
 
 f − Sn ( f ) ≤ C(ln n)En ( f ) p , n = 2, 3, . . . (1.4.28)
p

Theorem 1.4.13 Let p = 1, or ∞ and r > 0; then

Sn (Wp,r α ) p ≍ Sn (Hpr ) p ≍ n−r ln n, n = 2, 3, . . .

Proof The upper estimates follow from Theorem 1.4.9 and the inequality
(1.4.28). Owing to (1.4.20) it suffices to prove the lower estimates for the W
classes. We first remark that

Sn (W1,r α )1 = Sn (W∞,−
r
α )∞ . (1.4.29)

Indeed (see Theorem A.2.1),

  
Sn (W1,r α )1 = sup Fr (x, α ) ∗ φ − Sn (φ ) 1
φ 1 ≤1
 
= sup sup |Fr (x, α ) ∗ φ − Sn (φ ) , ψ |
φ 1 ≤1 ψ ∞ ≤1
 
= sup sup |φ , Fr (x, −α ) ∗ ψ − Sn (ψ ) |
φ 1 ≤1 ψ ∞ ≤1
r
= Sn (W∞,− α )∞ .

Therefore, to obtain the lower estimate it suffices to consider the case p = 1. Let n
be given. We consider
f (x) := einx Kn−1 (x);
34 Approximation of Univariate Functions

then, by the Bernstein inequality,

Dαr f 1 ≤ C(r)nr Kn−1 1 = C(r)nr . (1.4.30)

Further (see the analogous reasoning in the proof of (1.3.16)),

 n   n 
     
 f − Sn ( f ) =  ∑ (1 − k/n)eikx  ≥  ∑ (1 − k/n) sin kx
1    
k=1 1 k=1 1
n
≥ ∑ (1 − k/n)k−1 π − x−1
∞ ≥ C ln n. (1.4.31)
k=1

Relations (1.4.29)–(1.4.31) imply the theorem.

1.5 Historical Remarks

In §1.1, along with classical results of Fourier, Du Bois-Reymond, and Weierstrass,
which are usually included in a standard course of mathematical analysis, the
following papers are cited: Chebyshev (1854), de la Vallée Poussin (1908, 1919),
Bernstein (1912, 1914), Jackson (1911), Fredholm (1903), Lebesgue (1910),
Kolmogorov (1936, 1985), Favard (1937), Akhiezer and Krein (1937), Ismagilov
(1974), Kashin (1977), Tikhomirov (1960b), and Temlyakov (1982a).
Theorem 1.2.1 and its corollary were obtained by Riesz (see Zygmund, 1959,
vol. 1). A more detailed treatment of properties of the kernels of Dirichlet, Fejér,
de la Vallée Poussin, and Jackson can be found in Dzyadyk (1977). The Rudin–
Shapiro polynomials were constructed in Shapiro (1951) and Rudin (1952). The
polynomials Gn (x) were considered in Temlyakov (1989b). The proof of relation
(1.2.23) is analogous to reasoning from Trigub (1971). Theorem 1.2.5 is a classical
result of Gauss. Relation (1.2.31) was obtained by Hardy and Littlewood (1966).
Theorem 1.3.1 was obtained by Kolmogorov (1985), vol. 1, pp. 12–14. Theorem
1.3.2 in the case p = ∞, r = 1, α = r was proved by Bernstein (1952), vol. 1,
pp. 11–104. After this paper appeared, inequalities of this type began to be known
as Bernstein inequalities. Today in a number of cases the Bernstein inequalities are
known with explicit constants C(r). Theorem 1.3.4 in the case p = ∞ was obtained
by Jackson (1933) and in the general case by Nikol’skii (1951). Such inequalities
are known as Jackson–Nikol’skii or simply Nikol’skii inequalities. Theorem 1.3.6
was obtained by Marciekiewicz (see Zygmund, 1959, vol. 2).
In a number of cases of Theorem 1.4.1 the exact values are known (see the survey
Telyakovskii, 1988). Theorem 1.4.2 was proved by Hardy and Littlewood (1928).
The classes Hqr coincide with the Lipschitz classes for 0 < r < 1 and with the
Zygmund classes for r = 1. For r non-natural, the classes Hqr are analogous to the
 1.5 Historical Remarks 35
r−[r]
classes W [r] Hq . This statement follows from both direct and inverse theorems
for these classes because these theorems have the same form (see Theorem 1.4.3
and Corollary 1.4.6 as well as the survey Telyakovskii, 1988). Theorem 1.4.3 for
q = ∞ is a simple consequence of the results of Stechkin (1951). The proof in the
general case 1 ≤ q ≤ ∞ is carried out in the same way as in the case q = ∞. In fact,
Theorem 1.4.5 includes both the direct and inverse theorems for the approximation
of the classes Hqr B. Theorem 1.4.8 was obtained by Nikol’skii (see his 1969 book).
Theorem 1.4.9 is well known but it is not easy to assign priority; the situation is
similar for Theorem 1.4.12. Theorem 1.4.13 is due to Lebesgue (1910) for p = ∞
and to Nikol’skii for p = 1 (see the survey Telyakovskii, 1988).