Advanced Methods in Applied Analysis

Assignment 5

Alberto Castro - Jose Melo - Alejandro Vallejo

December 20, 2024

Definition 1 (Haar subspace). An n-dimensional subspace M of C[a, b] is called a Haar subspace if

0 is the only element of M which has n (or more) roots.
Corollary 1. Let M be a subspace and f be an element of C[a, b]. Let p be an element of M which
coincides with f in no more than a finite number of points. This p is a best L1 -approximant of f from
M , if and only if, sgn (f − p) ⊥ M .
Theorem 1 (D. Jackson). RLet M be a Haar subspace of C[a, b]. Then each f ∈ C[a, b] has a unique
L1 -approximant from M if I w(x) dx > 0 for all subinterval I ⊆ [a, b].

Problem 1. Determine a best approximant of f (x) = sin2 x from M in L1

 π π

− 2 , 2 , w , with
w(x) = cos x and M = span{1, sin x}. Is this best approximation unique?

First of all, we are going to study if M is a Haar subspace with Haar condition in Definition 1.
Let’s consider a general element of M and compute its roots in the interval [− π2 , π2 ].
a
sin x = − , (1)
b
which has only one root since the sine is injective in [− π2 , π2 ]. Consequently, the Haar condition is
satisfied and M is a Haar subspace.

Then, because M is a Haar subspace and the second condition from D. Jackson theorem (Theo-
rem 1) is satisfied since the cosine is always positive in the interval [− π2 , π2 ], the best L1 -approximant
to f from M is unique.

Now, in order to determine this best approximation, we use Corollary 1 to find it. Therefore the
problem to solve is to find a, b ∈ R. Such that
Z π2
sgn(f − p) · q(x) · cos (x) dx = 0 , ∀q ∈ M , where p(x) = a + b sin (x) . (2)
−π
2

Since M = span{1, sin x}, the integrals must be 0 for each element of the basis, so we will have a
systems of equations for a and b.

Z π
2
sgn(f − p) · cos (x) dx = 0 , (3)
−π
2

1
 Z π
2
sgn(f − p) · sin (x) · cos (x) dx = 0 . (4)
−π
2

To evaluate these integrals, first we have to study the function sgn(f − p). If there exist, i.e.,
b2 ≥ −4a, the roots of f − p are
sin2 (x) − b sin (x) − a = 0 ,
√
b ± b2 + 4a
sin (x± ) = t± = , x± = arcsin (t± ) , (5)
2
where t± are roots of P (t) = t2 − bt − a. This roots satisfy t+ ≥ t− and we assume that t± ∈ [−1, 1].
Since the arcsine is increasing in [−1, 1], then x+ ≥ x− .

We do not know the signs of t± but we can see that there exists an alternation of the sign between
the intervals
A1 = [− π2 , x− ) , B = (x− , x+ ) , A2 = (x+ , π2 ] .

Therefore, the integrals in (3) and (4), can be calculated as

π
Z x− Z x+ Z 2
λ cos (x) dx − λ cos (x) dx + λ cos (x) dx , (6)
−π
2 x− x+

π
Z x− Z x+ Z 2
λ sin (x) · cos (x) dx − λ sin (x) · cos (x) dx + λ sin (x) · cos (x) dx , (7)
−π
2 x− x+

where λ is the sgn (f − p) in the interval A1 .

Before solving the equations (3) and (4), we compute the primitives of the integrals,
Z
I1 (x) = cos (x) dx = sin (x) , (8)

sin2 (x)
Z
I2 (x) = sin (x) · cos (x) dx = . (9)
2

With the use of the Fundamental Theorem of Calculus, the equations (3) and (4) can be rewritten
as
−I1 (− π2 ) + 2I1 (x− ) − 2I1 (x+ ) + I1 ( π2 ) = 0 , (10)
−I2 (− π2 ) + 2I2 (x− ) − 2I1 (x+ ) + I2 ( π2 ) = 0 . (11)

If we substitute with the values x± we arrive at the non-linear system for t± ,

( (
2t− − 2t+ = −2 , t− − t+ = −1 ,
−→ (12)
2t2− − 2t2+ = 0 , (t− − t+ )(t− + t+ ) = 0 ,

which has a unique solution t− = − 21 and t+ = 21 . Consequently, another non-linear system arises for
a and b, ( √
b + b2 + 4a = 1 ,
√ (13)
b − b2 + 4a = −1 .
1
The solution to this system is a = 4 and b = 0.

After we have the solution to this system, we can conclude that the best L1 -approximation of f (x)
from M is
1
p(x) = . (14)
4

2
Problem 2. Let M be a subspace of L1 [−1, 1] generated by {1, x2 }. Find the best L1 -approxi-
mation of f (x) = x4 from M .

If we look at subspace M , we can easily see that it is not a Haar subspace, so there exists a best
L1 -approximant, but this is not neccesary to be unique. That being so, in this problem we will look
for any possible solutions of best approximations.

We can still use the Corollary 1 to find all the possible best approximations. Like in Problem 1,
the problem to solve is find a, b ∈ R such that
Z 1
sgn(f − p) · q(x) dx = 0 , ∀q ∈ M , where p(x) = a + bx2 . (15)
−1

Since M = span {1, x2 }, we have the following system of equations:

Z 1
sgn(f − p) dx = 0 , (16)
−1
Z 1
sgn(f − p) · x2 dx = 0 . (17)
−1

The approach of this problem consists in determine the sgn (f − p) for all possible values of a and
b, since the solution is not unique. For that, let’s consider the function

f (x) − p(x) = x4 − bx2 − a , (18)

which has at most 4 roots given by

s √ s √
b − b2 + 4a b+ b2 + 4a
±x1 = ± , ±x2 = ± . (19)
2 2

We will divide the study of the sign in 3 cases in order to find solutions of the best approximation.
The first one is when
b2 + 4a < 0 .

In this case the square roots of (19) do not exist in the real line, so there are not any roots in (18).
Because of this, the integral in (16) is always 2, so the equation is never satisfied, thus there is not a
best approximation in this case.

The second case is in the parabola

b2 + 4a = 0 .

With this constraint, there exists 2 roots at most in (18) given by

r
b
x± = ± . (20)
2

If b < 0, then there are not any roots, so it is not possible to find a best approximation. If b ≥ 0,
even though the roots x± exist, they happen to be minimums of (18), so the sign is always positive
and therefore, the equation (16) can never be satisfied.

3
 Now, the last case of this problem is when
b2 + 4a > 0 .
Note that in this case, if b > 0, then the 4 roots in (19) always exist in the real line, but if b ≤ 0, there
are only 2.

In the first two cases, we have not found any solutions, so all the best approximations must be in
this case. To study it, we also consider another 3 subcases.

Subcase 1: 0 roots in [−1, 1].

This occurs when either the roots (19) do not exist or they just do not belong to the interval
[−1, 1]. Hence, the sgn (f − p) is ±1 in [−1, 1] and the integral (16) is never 0, so there are not any
best approximations.

Subcase 2: 1 or 2 roots in [−1, 1].

This means that ±x1 ∈ [−1, 1]. Note that we do not state anything over the roots ±x2 . Thus, the
roots can be ordered as
−1 ≤ −x1 ≤ 0 ≤ x1 ≤ 1 .

We know that ±x1 are not extremums of f − p, so there exists a change in the sign. Therefore, we
will denote the sign in the interval [−1, −x1 ] by λ, so we can try to solve the system of equations (16)
and (17),

Z 1 Z −x1 Z x1 Z 1
sgn (f − p) dx = λ dx − λ dx + λ dx = λ (−4x1 + 2) = 0 , (21)
−1 −1 −x1 x1

Z 1 Z −x1 Z x1 Z 1
λ
sgn (f − p) · x2 dx = λ x2 dx − λ x2 dx + λ x2 dx = −4x31 + 2 = 0 .


(22)
−1 −1 −x1 x1 3

q
The solution of (21) is x1 = 21 and x1 = 3 12 for (22). Therefore, there is not a solution that satisfy
both equations at the same time, so it does not exist a best approximation either.

Subcase 3: 3 or 4 roots in [−1, 1].

In this subcase, all the roots belong to the interval [−1, 1] and can be ordered as
−1 ≤ −x2 < −x1 ≤ x1 < x2 ≤ 1 .

Since it is the last possible case, we have to find the solutions for the best approximations, so we
have to find the solutions of the system (16) and (17).

Z −x2 Z −x1 Z x1 Z x2 Z 1
dx − dx + dx − dx + dx = 4x1 − 4x2 + 2 = 0 , (23)
−1 −x2 −x1 x1 x2

Z −x2 Z −x1 Z x1 Z x2 Z 1
1
x2 dx − x2 dx + x2 dx − x2 dx + x2 dx = 4x31 − 4x32 + 2 = 0 .


(24)
−1 −x2 −x1 x1 x2 3

The non-linear system to solve is

(
2x1 − 2x2 = −1
(25)
2x31 − 2x32 = −1

4
which has 2 solutions (used the MATLAB Symbolic Toolbox). One of the solutions is negative for
both x1 and x2 , so the only valid solution is

√ √
5−1 5+1
x1 = , x2 = . (26)
4 4

Then, if rewrite (19) we can get the following expression for a and b:

2
2
x21 − x22 − x21 + x22
b = x21 + x22 , a= . (27)
4

Consequently, we have the best approximation of f (x) from M ,

3 2 1
p(x) = ·x − . (28)
4 16

We notice that even though M is not a Haar subspace, we have found that the solution is unique.
As we could have predicted, the Haar condition is sufficient but not necessary with regard to the
unicity of the solution.

Problem 3. Let M be a subspace of L1 [−1, 1] generated by {x, x3 }. Find the best L1 -approxi-
mation of f (x) = x5 from M .

In this problem we can see again that M is not a Haar subspace (we can have at most 3 roots).
Therefore, unicity is not guaranteed.

Since M = span {x, x3 }, we have the following system of equations:

Z 1
sgn(f − p) · x dx = 0 , (29)
−1
Z 1
sgn(f − p) · x3 dx = 0 , (30)
−1

with the following polynomial

f (x) − p(x) = x5 − bx3 − ax = x(x4 − bx2 − a) . (31)

Notice that f − p is an odd function, then 0 is always a root and they appear in pairs. As in
problem 2, the next roots take the form of

s √ s √
b− b2 + 4a b+ b2 + 4a
±x1 = ± , ±x2 = ± . (32)
2 2

Case 1: 1 root in [−1, 1].

5
 If√ f − p has only one root in [-1,1] it is either because it only has one real root (b2 + 4a < 0 or
b − b2 + 4a < 0) or the other roots lay outside [-1,1]. Also, we know that 0 is always a root of the
polynomial. Here we analyse the first case as the other will come in the following cases.

For the sake of notation, notice we could write 29 or 30 in a more general way as follows. Let qi
be an element of M . In this case, q1 = x and q2 = x3 . Then we can write 1 as

Z 1
λ qi dx = 0 , (33)
−1

where λ denotes the same idea as in Problem 2. Then, if we assume only one root in the interval,
we split the domain and assume changes of sign for the best approximation function

Z x1 Z 1
λ qi dx − λ qi dx = 0 . (34)
−1 x1

Solving the integrals we have

λ(2 · Ii (x1 ) − Ii (−1) − Ii (1)) = 0 , (35)

x2 x4
R
where Ii (x) = qi dx. In this exercise we have I1 (x) = 2 and I2 (x) = 4 . Replacing into 35 we
have two equations:

x21 1 1
λ(2 − x21 = 1 ,
− − )=0→ (36)
2 2 2

x41 1 1
λ(2 − x41 = 1 ,
− − )=0→ (37)
4 4 4

but we know that 0 is always a root, so it is a contradiction.

Case 2: 3 roots in [−1, 1].

We use the same idea as before. For 3 roots we have −1 ≤ x1 ≤ x2 ≤ x3 ≤ 1. In this case we have

Z x1 Z x2 Z x3 Z 1
λ qi dx − λ qi dx + λ qi dx − λ qi dx = 0 . (38)
−1 x1 x2 x3

Solving the integrals we

λ(2 · Ii (x1 ) − 2 · Ii (x2 ) + 2 · Ii (x3 ) − Ii (−1) − Ii (1)) = 0 , (39)

which leads to

x21 x2 x2 1 1
λ(2 −2 2 +2 3 − − )=0→
− x21 − x22 + x23 = 1 , (40)
2 2 2 2 2

x41 x4 x4 1 1
λ(2 −2 2 +2 3 − − )=0→
− x41 − x42 + x43 = 1 . (41)
4 4 4 4 4

6
 We know by symmetry that x1 = −x3 and x2 must be 0. Then from 40 and 41 we have that
x21 = 1/2 and x41 = 1/2 which is another contradiction.

Case 3: 5 roots in [−1, 1].

Finally, we analyse when we have 5 roots in [-1,1]. This means −1 ≤ x1 ≤ x2 ≤ x3 ≤ x4 ≤ x5 ≤ 1.

Again, we have the following integrals

Z x1 Z x2 Z x3 Z x4 Z x5 Z 1
λ qi dx − λ qi dx + λ qi dx − λ qi dx + λ qi dx − λ qi dx = 0 , (42)
−1 x1 x2 x3 x4 x5

and their solution

λ(2 · Ii (x1 ) − 2 · Ii (x2 ) + 2 · Ii (x3 ) − 2 · Ii (x4 ) + 2 · Ii (x5 ) − Ii (−1) − Ii (1)) = 0 . (43)

Then we have

x21 x2 x2 x2 x2 1 1
λ(2 −2 2 +2 3 −2 4 +2 5 − − )=0→
− x21 − x22 + x23 − x24 + x25 = 1 , (44)
2 2 2 2 2 2 2

x41 x4 x4 x4 x4 1 1
λ(2 −2 2 +2 3 −2 4 +2 5 − − )=0→
− x41 − x42 + x43 − x44 + x45 = 1 . (45)
4 4 4 4 4 4 4

Using same arguments as in last case, we know that the root in the middle must be 0 (r3 ). Also,
x1 = −x5 and x2 = −x4 , so we have a system of two unknowns and two equations. In other words

1
x21 − x22 = , (46)
2
1
x41 − x42 = . (47)
2

The solutions of the previous system are

√ √
3 1 1 3
x1 = − , x2 = − , x3 = 0, x4 = , x4 = . (48)
2 2 2 2

Finally, this yields the following best approximation

3
p(x) = x3 − x. (49)
16

7
Problem 4. Let M be a subspace of L1 [−1, 1] generated by {1, x}. Find the best L1 -
1
approximation of f (x) = x2 +10 from M .

It can be easily checked that the subspace M is a Haar subspace as the dimension is 2 but there is
only one root for the equation a − bx = 0. Thus, we know that the best approximation is unique for
this problem.

So, we can use Corollary 1, like we have done before in this assignment, to find the best approxi-
mation. This is, finding a and b so that,
Z 1
sgn(f − p) · q(x)dx = 0, ∀ q ∈ M, where p(x) = a + bx. (50)
−1

Since M = span {1, x}, we have the following system of equations:

Z 1
sgn(f − p) dx = 0 , (51)
−1
Z 1
sgn(f − p) · x dx = 0 . (52)
−1

Since the solution is unique we need to find the values of a and b that satisfy this system of
equations.

In order to find this we will study the sign of the following equation,
1
f (x) − p(x) = − a − bx = 0 . (53)
x2 + 10
Rearranging the terms we find,

g(x) bx3 + ax2 + 10bx + 10a − 1 = 0

f (x) − p(x) = = . (54)
h(x) x2 + 10

We can see that h(x) > 0 in [−1, 1], so analysing the sign of g(x) is equivalent. Thus, suppose that
g(x) has two roots in [−1, 1] denoted by x1 and x2 , and ordered as x1 < x2 . If we assume that both
roots are not extremums of f − p and that the sgn (f − p) = λ in [−1, x1 ], we can rewrite compute the
integrals (51) and (53),
Z 1 Z x1 Z x2 Z 1
sgn(f − p) dx = λ dx − λ dx + λ dx = 2λ (x1 − x2 + 1) = 0 , (55)
−1 −1 x1 x2
Z 1 Z x1 Z x2 Z 1
x dx = λ x21 − x22 = 0 .


sgn(f − p) · x dx = λ x dx − λ x dx + λ (56)
−1 −1 x1 x2

Now we have to solve the system of equations in order to find the roots x1 and x2 ,
(
x1 − x2 = −1 ,
(57)
x21 − x22 = 0 .

We have solved this system previously in Problem 1, so the solutions are x1 = − 21 and x2 = 12 . We
know the roots that satisfy the equations (51) and (53), so the next task is to find the coefficients a
and b which satisfy that g(xi ) = 0.
( (
g(− 12 ) = 0 , 41
2 b + 41a = 4 ,
1
−→ (58)
g( 2 ) = 0 , − 412 b + 41a = 4 .

8
 4
whose solution is a = 41 and b = 0.

We have find a best approximation and since the solution is unique, the best approximation of f (x)
from M in L1 [−1, 1] is

4
p(x) = . (59)
41

Problem 5. Give an example of a function f ∈ C[−1, 1] which has a nonunique best approxima-
tion in the mean by the functions of the form {cx}, in L1 [−1, 1]. Then, look at the same problem
on the intervals [0, 1] and [α, 1], for any α > 0.

Part 1: Interval [−1, 1].

Lets define f (x) = |x|. In this case, we can not use 1 since for c = 1 we have infinite numbers of
point coinciding. Still we can find a best approximation looking at

Z 1 Z 1
|f (x) − p(x)|dx = ||x| − cx|dx , (60)
−1 −1

since we know that the best should minimize this error function. Now, we integrate by dividing
the domain into two:

Z 1 Z 0 Z 1 Z 1 Z 0
r(c) = ||x| − cx|dx = | − x − cx|dx + |x − cx|dx = |1 − c| xdx − | − 1 − c| xdx, (61)
−1 −1 0 0 −1

with the last expression being equal to

| − 1 − c| + |1 − c|
r(c) = (62)
2

It’s clear to see that if −1 < c < 1 then r(c) = 1. If |c| ≥ 1, r(x) > 1, therefore, having multiple
best approximations.

Part 2: Interval [0, 1].

In the interval [0, 1], our problem reduces to analysing

Z 1 Z 1 Z 1
r(c) = |f (x) − p(x)|dx = |x − cx|dx = |1 − c| xdx . (63)
0 0 0

Notice that now the minimum (r(c) = 0) is attained when c = 1.

Part 3: Interval [α, 1], α > 0.

In this last part, M is a Haar subspace in [α, 1], so by Theorem 1, every function f (x) ∈ C[α, 1]
has a unique best approximant. Thus, the best approximation of the problem is obtained the same
way as in (63) and is achieved when c = 1.