1

1 Polynomial approximation and interpolation

1.1 Jackson theorems
1.1.1 Polynomials Pn and trigonometric polynomials Tn
In order to state the approximation problem we define the functions which we want to approx-
imate and the functions which we want to use for approximation:

Definition 1.1 We denote by C k [a, b] for k = 0, 1, 2, . . . the space of functions which have
derivatives f (1) ,. . . ,f (k) that are continuous on the closed interval [a, b].
We denote by
P n = { c0 + c1 x + · · · + cn xn | ck ∈ C
/ }

the space of polynomials of degree less or equal to n (n = 0, 1, 2, . . .).

Then the approximation problem is: Given f ∈ C k [a, b], what is the rate with which error
of the best approximation
inf kf − pn k∞
pn ∈Pn

converges to zero as n goes to infinity?

The so-called Jackson theorems shows that the decay rate of the error depends on the
smoothness of the function f . E.g. for f ∈ C 1 [a, b] we will prove an approximation rate of
O(1/n), and for f ∈ C 2 [a, b] we will obtain an approximation rate of O(1/n2 ).
The problem of approximating functions on intervals by polynomials is closely related to
the problem of approximating periodic functions by trigonometric polynomials:

Definition 1.2 The space C2π of 2π-periodic functions consists of all functions f ∈ C(IR)
which satisfy
∀x ∈ IR f (x) = f (x + 2π).
k
The space of k times continuously differentiable 2π-periodic functions is defined as C2π =
k
C (IR) ∩ C2π .
We denote by Tn the space of trigonometric polynomials of degree less or equal to n (n =
0, 1, 2, . . .):
n
X
Tn = { a0 + (ak cos kx + bk sin kx) | ak , bk ∈ C
/ }
k=1

1.1.2 Interpolation with trigonometric polynomials

Note that the space Tn has dimension 2n + 1 (whereas Pn has dimension n + 1). Therefore we
can ask whether we can always find a trigonometric interpolation polynomial pn ∈ Tn which
passes through 2n + 1 given points (xj , yj ), j = 0, . . . , 2n. We first note that Tn has a property
similar to Pn :

Lemma 1.3 A nonzero function f ∈ Tn has at most 2n zeros in [0, 2π).

1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 2

Proof: Assume that f ∈ Tn has 2n + 1 zeros θ0 , . . . , θ2n in [0, 2π). Writing sin kθ and cos kθ in
terms of eikθ and e−ikθ and with z := eiθ we have that
n
X n
X
f (θ) = ck eikθ = ck z k
k=−n k=−n
P
equals zero for zj := eiθj . Multiplying by z n we obtain that nk=−n ck z k+n is a polynomial of
degree ≤ 2n in z which has at least 2n + 1 zeros zj ∈ C/ . Hence it must be the zero polynomial

and all ck = 0.

Corollary 1.4 Let x0 , . . . , x2n be distinct values in [0, 2π). Then for any given values
y0 , . . . , y2n there exists a unique interpolating trigonometric polynomial pn ∈ Tn which satis-
fies pn (xj ) = yj , j = 0, . . . , 2n.
Proof: The interpolation problem leads to a linear system of 2n + 1 equations for the 2n + 1
unknowns a0 , . . . , an , b1 , . . . , bn with the right hand side vector (y0 , . . . , y2n )⊤ . This system has
a unique solution for every right hand side if the matrix is nonsingular. To show that the matrix
is nonsingular consider the problem with the zero right hand side vector. Any solution of this
linear system corresponds to a function pn ∈ Tn which is zero in x0 , . . . , x2n . By Lemma 1.3 pn
must be zero. Hence the homogeneous linear sytem has only the zero solution and the matrix
is nonsingular.

1.1.3 An auxiliary approximation problem

As a first step toward proving the Jackson theorems let us consider the 2π periodic function
f with f (x) = x for x ∈ (−π, π]. In order to approximate it by a function in Tn we can use
π
the interpolation pn through the 2n + 1 nodes k n+1 , k = −n, . . . , n which exists and is unique
due to Corollary 1.4. Since the function f is odd, i.e., f (−x) = −f (x) (for x 6= kπ), the
function −pn (−x) ∈ Tn is also an interpolation. By the uniqueness of the interpolation, pn (x)
and −pn (−x) must have the same coefficients and so we have that
n
X
pn (x) = bk sin kx. (1.1)
k=1

Now we consider the interpolation error e(x) := f (x) − pn (x). Since pn interpolates f in
2n + 1 points in (−π, π) , e(x) has at least 2n + 1 simple zeros in (−π, π). There cannot be
more zeros in (−π, π): If e(x) has 2n + 2 zeros in (−π, π), then e′ (x) = 1 − p′n (x) ∈ Tn has
at least 2n + 1 zeros by Rolle’s theorem. By Corollary 1.3 we then have e′ (x) = 0 which is a
contradiction. The same argument also shows that the 2n + 1 zeros of e in (−π, π) are simple,
i.e., e′ (x) 6= 0 in those points. Hence the function e(x) changes its sign in (−π, π) only at the
π
interpolation points k n+1 , k = −n, . . . , n. Let s(x) denote the function which is alternatingly
1 and −1 between the nodes, i.e.,
s(x)|[k π π
,(k+1) n+1 ) = (−1)k ,
n+1

then we have
Zπ  Zπ   Zπ Zπ 
  
|f (x) − pn (x)| dx =  (f (x) − pn (x))s(x) dx =  f (x)s(x) dx − pn (x)s(x) dx.
−π −π −π −π
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 3

Let us show that the second term on the right hand side must be 0: Consider
Zπ Zπ Z π
π π
A := s(x) sin kx dx = − s(x + ) sin kx dx = − s(x) sin k(x − ) dx
n+1 n+1
−π −π −π
Zπ Zπ
π π
= − s(x) sin kx cos(−k ) dx − s(x) cos kx sin(−k ) dx
n+1 | {z } | {z } n+1
−π −π odd even
π
= − cos(k )A + 0
n+1
R
π
Since cos(k n+1 ) 6= −1
Rπ
we obtain A = 0 and also ππ pn (x)s(x) dx = 0 because of (1.1). It
remains to evaluate π f (x)s(x) dx: It can be easily seen geometrically that we have

Zπ  Zπ  
π
2 
 1 3 1 

π
 2
n+1 π2
 
 xs(x) dx = 2 xs(x) dx =2 − + − · · · ± (n + ) = 2 = .
π
n+1 
2 2 2 n+1 2 n+1
0

Hence we obtain the following result for the interpolation error:

Zπ
π2
|f (x) − pn (x)| dx = .
π
n+1

1.1.4 Jackson theorems for periodic functions

1
Now we can prove the Jackson theorem for f ∈ C2π : Using integration by parts we see that
Zπ Zπ Zπ
′
θf (θ + x + π) dθ = [θf (θ + x + π)]π−π − 1f (θ + x + π) dθ = 2πf (x) − f (θ) dθ
−π −π −π

yielding
π π
1 Z 1 Z
f (x) = f (θ) dθ + θf ′ (θ + x + π) dθ (1.2)
2π 2π
−π −π
| {z }
=:a0 =const.

In order to approximate f (x) we replace θ in the second integral by the interpolation pn ∈ Tn

from above and obtain
π π
1 Z ′ 1 Z
q(x) = a0 + pn (θ)f (θ + x + π) dθ = a0 + pn (θ − π − x)f ′ (θ) dθ (1.3)
2π 2π
−π −π

where
n
X n
X  
pn (θ − π − x) = bk sin k(θ − π − x) = bk sin k(θ − π) |cos{zkx} − cos k(θ − π) |sin{zkx} . (1.4)
k=1 k=1

If we insert this expression for pn (θ−π−x) into the integral in (1.3) we see that the underbraced
terms can be pulled out of the integral and that q ∈ Tn .
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 4

It remains to estimate the approximation error:

π π
 1 Z ′
 1 Z
|θ − pn (θ)| |f ′ (θ + x + π)| dθ
 
kf − qk∞ = sup (θ − pn (θ))f (θ + x + π) dθ ≤
x 2π 2π
−π −π
Zπ
1 1 π2
≤ kf ′ k∞ |θ − pn (θ)| dθ = kf ′ k∞
2π 2π n+1
−π

We have therefore proved the following theorem:

1 1
Theorem 1.5 (Jackson theorem for C2π ) For f ∈ C2π there holds
π
inf kf − pk∞ ≤ kf ′ k∞ (1.5)
p∈Tn 2(n + 1)
π R
If the function f (x) has mean 0, i.e., −π f (x) dx = 0, then the approximating function q(x)
has the same property: The constant a0 in (1.2) is 0, and (1.3), (1.4) show that q(x) is of the
P
form nk=1 (ak cos kx + bk sin kx). Let us define TN0 as the space of trigonometric polynomials of
degree less or equal n with mean 0:
Zπ n
X
TN0 := { p ∈ TN | p(x) dx = 0 } = { (ak cos kx + bk sin kx) | ak , bk ∈ C
/ }

−π k=1

We then have
1
Rπ
Lemma 1.6 Let f ∈ C2π with −π f (x) dx = 0. Then
π
inf0 kf − pk∞ ≤ kf ′ k∞ . (1.6)
p∈TN 2(n + 1)
Next, we would like to investigate the approximation rates for functions with more smoothness,
3 1
e.g., f ∈ C2π . We can use the following “bootstrap” argument: Assume that f ∈ C2π . Let p̃
be an arbitrary trigonometric polynomial in Tn . Then we have obviously
π
inf kf − pk∞ = inf k(f − p̃) − pk∞ ≤ kf ′ − p̃′ k∞ (1.7)
p∈Tn p∈Tn 2(n + 1)
using (1.5) for (f − p̃). Note that (1.7) holds for any function p̃ ∈ Tn , so we can choose it in
such a way that the right hand side becomes as small as possible. Since
{ p̃′ | p̃ ∈ Tn } = Tn0 ,
equation (1.7) implies
π
inf kf − pk∞ ≤ inf kf ′ − pk∞ (1.8)
p∈Tn 2(n + 1) p∈Tn0
For
Rπ
approximation errors in Tn0 1
we can use a similar argument: Assume f ∈ C2π with
−π f (x) dx = 0. Using (1.6) and

{ p̃′ | p̃ ∈ Tn0 } = Tn0 ,

we obtain analogously
π
inf0 kf − pk∞ ≤ inf kf ′ − pk∞ . (1.9)
p∈Tn 2(n + 1) p∈Tn0
This allows us to prove
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 5

k k
Theorem 1.7 (Jackson theorem for C2π ) Let k ∈ {1, 2, . . .} and f ∈ C2π . Then
!k
π
(k)
inf kf − pk∞ ≤ f (1.10)
p∈Tn 2(n + 1) ∞

Proof: Note that f ′ ,. . . ,f (k−1) have mean 0 since f is periodic. Applying (1.8), (1.9) (k − 2
times) and finally (1.6) we get
!2
(1.8) π (1.9) π
inf kf − pk∞ ≤ inf0 kf ′ − pk∞ ≤ inf kf ′′ − pk∞
p∈Tn 2(n + 1) p∈Tn 2(n + 1) p∈Tn0
!k−1 !k
(1.9) (1.9) π
(k−1)

(1.6) π
(k)
≤ ··· ≤ inf0 f − p ≤ f .
2(n + 1) p∈Tn ∞ 2(n + 1) ∞

1.1.5 Jackson theorems on an interval

For simplicity we first consider the interval [−1, 1] and consider a function f ∈ C 1 [−1, 1]. We
can transform this function to a 2π-periodic function g by the change of variables

g(θ) = f (cos θ) (1.11)

Note that the function g is even. This transformation is useful since polynomials f (x) of
degree less or equal n are transformed into even trigonometric polynomials g(θ) and vice versa.
If g ∈ Tm and h ∈ Tn then the product g · h is in Tm+n as can be seen, e.g., by writing
cos x and sin x in terms of e±ix . If f ∈ Pn then we see that g given by (1.11) is in Tn and
even. It is also clear that g can only be the zero function if f is the zero function. Since
dim Tneven = dim Pn = n + 1 it follows that the transformation (1.11) gives a one-to-one linear
mapping between Pn and Tneven .
The functions cos kx, k = 0, . . . , n form a basis of Tneven . These functions are transformed
by (1.11) to certain polynomials Tk (x) with

cos kθ = Tk (cos θ). (1.12)

These are the so-called Chebyshev polynomials. Obviously, T0 (x) = 1 and T1 (x) = x. If we add
the formulae cos(k ± 1)x = cos kx cos x ∓ sin kx sin x we obtain cos(k + 1)x = 2 cos kx cos x −
cos(k − 1)x which gives the recursion formula

Tk+1 (x) = 2Tk (x)x − Tk−1 (x). (1.13)

We can now prove an approximation result for polynomials on an interval:

Theorem 1.8 (Jackson theorem for C 1 [ − 1, 1]) For f ∈ C 1 [−1, 1] there holds
π
inf kf − pk∞ ≤ kf ′ k∞ (1.14)
p∈Pn 2(n + 1)
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 6

Proof: Define g by (1.11). Then g ′ (θ) = −f ′ (cos θ) sin θ. Therefore the limits limθ→±π g ′ (θ)
1
exist and are equal (both are 0), hence g ∈ C2π . We also see that

kg ′ k∞ ≤ kf ′ k∞ . (1.15)

Consider an approximation p ∈ Tn for g. Since p need not be even we consider the symmetrized
function p̃(θ) = (p(θ) + p(−θ))/2. Then we see

kg(θ) − p̃(θ)k∞ = 21 kg(θ) + g(−θ) − p(θ) − p(−θ)k∞

≤ 12 (kg(θ) − p(θ)k∞ + kg(−θ) − p(−θ)k∞ ) = kg(θ) − p(θ)k∞

and this implies

(1.5) π
infeven kg − pk∞ = inf kg − pk∞ ≤ kg ′ k∞ (1.16)
p∈Tn p∈Tn 2(n + 1)
The change of variables q(cos θ) = p(θ) defines for p ∈ Tneven a function q ∈ Pn such that

max |f (x) − q(x)| = max |g(θ) − p(θ)| . (1.17)

x∈[−1,1] θ∈[−π,π)

Equations (1.17), (1.16), and (1.15) together yield (1.14).

In the same way as we proved Theorem 1.7 using Theorem 1.5 we can use Theorem 1.8 to
prove an approximation result for f ∈ C k [−1, 1]:

Theorem 1.9 (Jackson theorem for C k [ − 1, 1]) Let n, k be integers with n ≥ k − 1 ≥ 0

and f ∈ C k [−1, 1]. Then
 k
π 1
(k)
inf kf − pk∞ ≤ f (1.18)
p∈Pn 2 (n + 1)n · · · (n − k + 2) ∞

1.2 Polynomial interpolation

The proofs for the Jackson theorems are constructive. But the constructions of the approxi-
mating functions are of no practical value since they are very complicated, involving various
differentiations and integrations (especially for higher values of k) and are therefore not very
efficient.
In this section we will see that we can achieve an approximation which is almost as good as
the optimal approximation using interpolation, provided the interpolation nodes are correctly
chosen.

1.2.1 The error formula and Chebyshev nodes

Theorem 1.10 Let f ∈ C n+1 [a, b] and let pn ∈ Pn be the interpolating polynomial through the
distinct interpolation nodes x0 , . . . , xn ∈ [a, b]. Then for t ∈ [a, b] there exists ξ ∈ [a, b] such
that
f (n+1) (ξ)
f (t) − pn (t) = (t − x0 ) · · · (t − xn ) (1.19)
(n + 1)!
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 7

Proof: Let ω(x) := (x − x0 ) · · · (x − xn ). Fix t ∈ [a, b] with t 6= xj , j = 0, . . . , n (otherwise (1.19)

is trivial). Then ω(t) 6= 0 and the definitions

At := (f (t) − pn (t))/ω(t), g(x) := f (x) − pn (x) − At ω(x)

imply that g(t) = 0. Since also g(xj ) = 0 (pn interpolates f and ω(xj ) = 0) we see that g(x)
has at least n + 2 zeros in [a, b]. By Rolle’s theorem, g ′ has at least n + 1 zeros,. . . , g (n+1) has
at least 1 zero which we will call ξ. Therefore

0 = g (n+1) (ξ) = f (n+1) (ξ) − p(n+1)

n (ξ) − At ω (n+1) (ξ) = f (n+1) (ξ) − 0 − At (n + 1)!

since pn ∈ Pn and ω(x) = xn+1 +q(x), q ∈ Pn . Therefore At = f (n+1) (ξ)/(n+1)! which together
with 0 = g(t) = f (t) − pn (t) − At ω(t) yields (1.19).
Estimate (1.19) implies
1
(n+1)
kf − pn k∞ ≤ f kω(x)k∞ (1.20)
(n + 1)! ∞

where ω(x) denotes the node polynomial

ω(x) := (x − x0 ) · · · (x − xn ).

We see that only the last factor in (1.20) depends on the choice of nodes. Therefore it is a good
idea to choose the n + 1 nodes in [a, b] in such a way that kω(x)k∞ becomes minimal.
Let us first consider the interval [−1, 1]. The Chebyshev polynomial Tk (x) has by its defi-
nition (1.12) the property that it oscillates between 1 and −1 and has all its k zeros
!
j + 1/2
xj = cos π , j = 0, . . . , k − 1
k

in the interval [−1, 1]. Furthermore, by the recursion formula (1.13) the polynomial Tk has the
leading coefficient 2k−1 for k ≥ 1. Hence

Tk (x) = 2k−1 (x − x0 ) · · · (x − xk ). (1.21)

If we choose the n + 1 interpolation nodes as the zeros of the polynomial Tn+1 we have
therefore
ω(x) = 2−n Tn+1 (x) (1.22)
and because of kTk k∞ = 1 we obtain

kωk∞ = 2−n .

One cannot achieve a smaller value of kωk∞ than 2−n with any arrangement of nodes x0 , . . . , xn :
Assume kω̃k∞ < 2−n . Since ω alternatingly assumes n + 2 extrema 2−n and −2−n , the polyno-
mial q := ω − ω̃ must have a zero between two subsequent extrema of ω by the intermediate
value theorem. Thus q = ω − ω̃ has at least n + 1 zeros. But since both ω and ω̃ are of the form
xn+1 + lower order terms we see that q ∈ Pn and hence q = 0 which contradicts the assumption.
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 8

 
Therefore using the Chebyshev nodes xk = cos k+1/2 n+1
π , k = 0, . . . , n leads to the best
estimate in (1.20). Using a linear transformation, we see that the best choice of nodes for the
interval [a, b] is !
a+b k + 1/2 a−b
+ cos π , k = 0, . . . , n (1.23)
2 n+1 2
 n+1  n+1
b−a b−a
yielding kωk∞ = 2−n 2
=2 4
and

!n+1 (n+1)
b−a f
∞
kf − pn k∞ ≤ 2 (1.24)
4 (n + 1)!
For an arbitrary choice of nodes we have using (1.20) and |x − xj | ≤ (b − a) the estimate

(n+1)
f
kf − pn k∞ ≤ (b − a)n+1 ∞
. (1.25)
(n + 1)!
These formulas show that the interpolation error converges to 0 if the derivatives do not grow
too fast. For example, assume that the function f is analytic on [a, b], i.e. the Taylor series
converges for every x ∈ [a, b] for all y with |y − x| < ρ with some ρ > 0. Equivalently, the
function f can be extended to a holomorphic function in all points in C / which have a distance

of less than ρ to [a, b] (ρ can be chosen as the distance of the closest singularity of f in C
/ to
(n)
the interval [a, b]). Then we know (e.g., expressing f with the Cauchy theorem) that for any
ρ̃ < ρ there exists C such that
1   1
max f (n) (x) ≤ C n .
 
for all n ≥ 0
n! x∈[a,b] ρ̃
If (b − a) < ρ then we see that we have exponential convergence for any choice of nodes.
But the formulae (1.24) and (1.25) do not give any information about the convergence if f
is less smooth. In fact, interpolation with equally spaced nodes can diverge even if f is analytic
1
on [a, b], as the Runge example f (x) = 1+x 2 on [−5, 5] shows. (Here we have singularities at ±i

and hence ρ = 1.) On the other hand, interpolation at the Chebyshev nodes converges if f is
only marginally smoother than a continuous function: The so-called Dini-Lipschitz condition
log ε max |f (x) − f (y)| → 0 for ε → 0
|x−y|≤ε
x,y∈[a,b]

is sufficient.
q This includes continuous, piecewise differentiable functions, as well as functions
like |x|. However, there exist continuous functions for which the Chebyshev approximations
do not uniformly converge. But one can show that this is true for any sequence of interpolation
nodes.

1.2.2 The Lagrange form and the Lebesgue function

Let x0 ,. . . ,xn be distinct nodes in the interval [a, b]. For a function f ∈ C[a, b] let us denote the
interpolating polynomial in Pn by Pn f . We can give an explicit formula for this polynomial:
The Lagrange polynomials
Y x − xj
lk (x) :=
j=0,...,n xk − xj
j6=k
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 9

obviously satisfy 
0 for k 6= j
l k ∈ Pn , lk (xj ) = . (1.26)
1 for k = j
Therefore we can write the interpolating polynomial in the so-called Lagrange form
n
X
Pn f = f (xk )lk .
k=0

Then we get the following estimate

 
Xn  n
X
 
|(Pn f )(x)| =  f (xk )lk (x) ≤ ( max |f (xk )|) |lk (x)| ≤ kf k∞ λn (x)
  k=0,...,n
k=0 k=0
Pn
where λn (x) := k=0 |lk (x)| is the so-called Lebesgue function. We also have

kPn f k∞ ≤ kλn k∞ kf k∞ (1.27)

and this estimate is sharp, i.e., there exist f ∈ C[a, b] where we have equality in (1.27). Since
λn (xk ) = 1 by (1.26) there holds kλn k∞ ≥ 1. In order to have a good approximation we would
like to choose the interpolation points in such a way that kλn k∞ does not become too large.
Otherwise there exist functions with small values which have interpolating polynomials with
very large values.
The size of the Lebesgue constant kλn k∞ also characterizes the relation between the interpo-
lation error and the best possible approximation error: Let qn ∈ Pn be an arbitrary polynomial,
then Pn q = q and

kf − Pn f k∞ ≤ kf − qk∞ +kq − Pn f k = kf − qk∞ +kPn (q − f )k∞ ≤ kf − qk∞ +kλn k∞ kf − qk∞

Therefore
kf − Pn f k∞ ≤ (1 + kλn k∞ ) inf kf − qk∞ (1.28)
q∈Pn

1.2.3 Estimates for the Lebesgue constant

For uniformly spaced interpolation points one can prove

kλn k∞ ≥ Cen/2

This is illustrated by the Runge example where kPn f k∞ grows exponentially. Unfortunately
one can never achieve a bounded Lebesgue constant: For any sequence of interpolation nodes
there holds
2
kλn k∞ ≥ log n − c.
π
However, the Chebyshev nodes give a Lebesgue constant which has the same growth rate as
this lower bound: For Chebyshev nodes we have
Theorem 1.11 The Lebesgue function for Chebyshev nodes (1.23) satisfies
2 4
kλn k∞ ≤ log(n + 1) + (1.29)
π π
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 10

Proof: Note that we can write the Lagrange polynomial lk in terms of the node polynomial
ω(x) = (x − x0 ) · · · (x − xn ) as
ω(x)
lk (x) =
(x − xk )ω ′ (xk )
Using (1.22) and (1.12) and the change of variables x = cos θ, xk = cos θk with θ, θk ∈ [0, π] we
have
cos(n + 1)θ
lk (cos θ) = 
dθ 
(cos θ − cos θk )(n + 1) sin(n + 1)θk dx 
x=xk
dx
Because of sin(n + 1)θk = ±1 and dθ
= − sin θ this yields
 
n 
|cos(n + 1)θ| X  sin θk 

µ(θ) := λn (cos θ) =   (1.30)
n+1 k=0
 cos θ − cos θk 

or, using the trigonometric identity sin a/(cos a − cos b) = (cot a−b
2
− cot a+b
2
)/2
n 
 
|cos(n + 1)θ| X  θ + θk θ − θk 
µ(θ) = cot − cot  (1.31)
2(n + 1) k=0  2 2 

We want to show that maxx∈[−1,1] λn (x) = maxθ∈[0,π] µ(θ) attains its maximal value in the
π
interval [0, θ0 ] = [0, 2(n+1) ]. Let φ ∈ [0, π] arbitrary, then we can write φ = φ0 + lπ/(n + 1) with
π π
φ0 ∈ [− 2(n+1) , 2(n+1) ] and integer l. Since |cos t| is a π-periodic function we have |cos(n + 1)φ| =
|cos(n + 1)φ0 |. For the sum we claim

n 
    !
n
X  φ + θk φ − θk  X 
 φ + θk   φ − θk 
cot − cot ≤ cot  + cot 

k=0 2 2  k=0  2   2 
   !  
n n 
X 
 φ0 + θk   φ0 − θk  X  φ 0 + θk φ0 − θk 
= cot  + cot  = cot − cot  (1.32)
k=0
 2   2  
k=0 2 2 

The 2(n+1) summands in the second and third sum are the same, but in different combinations
since cot is π-periodic and
π π
{ φ + θk , φ − θk | k = 0, . . . , n } = { φ + +k | k = −n − 1, . . . , n }.
2(n + 1) n+1
The last equality in (1.32) follows from
φ 0 + θk π π φ 0 − θk φ 0 + θk φ 0 − θk
0≤ ≤ , − ≤ ≤0 =⇒ cot ≥ 0, cot ≤0
2 2 2 2 2 2
Therefore we have (using that µ is an even function)
µ(φ) ≤ µ(φ0 ) = µ(|φ0 |) =⇒ max µ(φ) = max π
µ(φ0 ).
φ∈[0,π] φ0 ∈[0, 2(n+1) ]

π
Now we consider θ ∈ [0, 2(n+1) ]. Then cos(n + 1)θ ≥ 0, cos θ − cos θk ≥ 0 and hence we obtain
n n
1 X cos(n + 1)θ 1 X Y
µ(θ) = λn (cos θ) = sin θk = (sin θk ) 2n (cos θ − cos θj )
n + 1 k=0 cos θ − cos θk n + 1 k=0 j=0...n
j6=k
1 POLYNOMIAL APPROXIMATION AND INTERPOLATION 11

using (1.21). Since cos θ − cos θj ≥ 0, each term in the sum is a nonnegative decreasing function
of θ, therefore
n
(1.31) 1 X θk
kλn k∞ = maxπ µ(θ) = µ(0) = 2 cot (1.33)
θ∈[0, 2(n+1) ] 2(n + 1) k=0 2
We see that the last expression is exactly the composite midpoint rule applied to the integral
1 Rπ
π 0
cot(t/2) dt (which is infinite). We split off the first term and note that the integrand is
convex (f ′′ ≥ 0), hence the integral is larger than the midpoint rule approximation:
n Zπ
1 θ0 1 X θk 1 π 1 t
kλn k∞ = cot + cot ≤ cot + cot dt
n+1 2 n + 1 k=1 2 n+1 4(n + 1) π 2
π/(n+1)
  π !
1 π 1 t 1 π 2 π
= cot + 2 log sin = cot − log sin
n+1 4(n + 1) π 2 π/(n+1) n+1 4(n + 1) π 2(n + 1)

Finally the inequalities tan t ≥ t (0 ≤ t < π/2) and sin t ≥ π2 t (0 ≤ t ≤ π/2) give

4 2
kλn k∞ ≤ + log(n + 1)
π π
Table 1 shows the values of kλn k∞ computed directly from (1.33) and the upper bounds

n 5 10 15 20 25
kλn k∞ 2.104 2.489 2.728 2.901 3.037
2 4
π
log(n + 1) + π
2.414 2.800 3.038 3.211 3.347

Table 1: Lebesgue constants for Chebychev nodes

from (1.29).
Because of (1.28) the interpolation Pn f at Chebyshev nodes satisfies for n ≤ 20

kf − Pn f k∞ ≤ 4 inf kf − qk∞ .
q∈Pn

Therefore the Chebyshev interpolation gives almost the best possible approximation. Trying
to find the best possible approximation (which can only be done approximately by an iterative
procedure) can at most decrease the error by a factor of 4.
One can do even better by using the expanded Chebyshev nodes

x̃j := xj /x0

which stretch the Chebyshev nodes so that the endpoints x̃0 = 1 and x̃n = −1 are nodes. In
this case we have kλn k∞ ≤ 2.01 for n ≤ 9 and kλn k∞ ≤ 3 for n ≤ 47. Furthermore one can
show that kλn k∞ is never more than 0.02 larger than the best possible value of the Lebesgue
constant with n + 1 nodes.