Discrete Fourier Transform (aka DFT aka FFT)

1 Introduction

We consider vectors in CN . We will define an orthogonal basis e(0) , . . . , e(N−1) which is important in many applications.
>
A vector f = [ f0 , . . . , fN−1 ]> can be written as f = fˆ0 e(0) + · · · + fˆN−1 with a coefficient vector fˆ = fˆ0 , . . . , fˆN−1 .


Recall that for an orthogonal basis we obtain the coefficients as

f , e(k)


fˆk = (k) (k)

e ,e

1.1 1-periodic functions, trigonometric functions with frequency 0, 1, 2, . . . , M

We are interested in “periodic signals”. We say a function f is 1-periodic if

f (t + 1) = f (t) for all t ∈ R

The function f (t) may have complex values.

If a function g has period L > 0, then the function f (x) := g(Lx) has period 1. Therefore it is sufficient to consider 1-periodic
signals.
We want to approximate this signal using 1-periodic trigonometric functions of frequency 0, 1, . . . , M:

v(t) = a0 · 1 + a1 cos(2πt) + b1 sin(2πt) + · · · + aM cos(2πMt) + bM cos(2πMt)

|{z} | {z } | {z }
freq. 0 frequency 1 frequency M

with coefficients ak , bk ∈ C.
Since

eit = cost + i sint

1 1 it
cost = (eit + e−it ), e − e−it


sint =
2 2i
we can write v(t) as
M
v(t) = ∑ fˆk e2πikt
k=−M

Note that eit is on the unit circle in the complex plane. Let ω := e2πi/N . The equation zN = 1 has the N solutions

ω 0 , ω 1 , . . . , ω N−1

located on the unit circle, with angles j 2π

N , j = 0, . . . , N − 1.
1.2 Periodic functions for discrete time values, “aliasing”

We consider a step size 1/N and the discrete points t j = j/N for j ∈ Z. A signal is given by values f j at the points t j , j ∈ Z.
As the signal is periodic we have
f j+N = f j for j ∈ Z.
Hence we only need to specify the values
f = [ f0 , . . . , fN−1 ]> ∈ CN .
Let k ∈ Z be an integer. We now want to use the values of the trigonometric function e2πikt at the points t j : This function has
the values given by the vector
h i>
e(k) := e2πik·0/N , . . . , e2πik· j/N

We have so-called aliasing: We have for any k ∈ Z that e(k) = e(k+N)

e2πi(k+N)t j = e2πi(k+N) j/N = e2πi jk/N e|{z}

2πi j
= e2πikt j
1

This means that the vectors e(k) for all k ∈ Z are actually only N different vectors which keep repeating: Consider e.g. N = 5
(−2) (−1) (0) (1) (2) (3) (4)
. . . , e|{z} e(5) , |{z}
, e|{z}, e , e , e , e , e , |{z} e(6) , . . .
e(3) e(4) e(0) e(1)

In order to represent the vector f = [ f0 , . . . , f4 ]> ∈ C5 we need a basis of 5 vectors. We pick the vectors e(0) , e(1) , e(2) , e(3) , e(4) .
We could also have picked the vectors e(−2) , e(−1) , e(0) , e(1) , e(2) , but these are exactly the same vectors, just in a different
order.
We also define the corresponding vectors c(k) and s(k) for cos(k2πt) and sin(k2πt):

c(k) := [cos(k2πt0 ), . . . , cos(k2πtN−1 )]> , s(k) := [sin(k2πt0 ), . . . , sin(k2πtN−1 )]>

e(k) = c(k) + is(k) (1)
1 (k)
  1  (k) 
c(k) = e + e(−k) , s(k) = e − e(−k)
2 2i
i.e., n o n o
e(0) = c(0) , span e(k) , e(−k) = span c(k) , s(k) for k = 1, 2, 3, . . .

For N = 5 the vectors e(0) , e(1) , e(2) , e(3) , e(4) can also be written as e(0) , e(1) , e(2) , e(−2) , e(−1) because of aliasing. Hence
n o n o
C5 = span e(0) , e(1) , e(2) , e(3) , e(4) = span c(0) , c(1) , c(2) , s(1) , s(2)

This works for any odd N = 2M + 1:

n o n o
CN = span e(0) , . . . , e(N−1) = span c(0) , c(1) , . . . , c(M) , s(1) , . . . , s(M)

1.3 The discrete Fourier transform FN

We claim that the vectors e(0) , . . . , e(N−1) form an orthogonal basis of CN : We have
  N−1 N−1
e(k) , e(l) = ∑ e2πik j/N e−2πil j/N = ∑ a j a := e2πi(k−l)/N
j=0 j=0
For k = l we have a = 1 and hence  
e(k) , e(k) = N. (2)

For k, l ∈ {0, . . . , N − 1} with k 6= l we have a 6= 1 and

  N−1 aN − 1 e2πi(k−l) − 1 1 − 1
e(k) , e(l) = ∑ a j = = = = 0.
j=0 a−1 a−1 a−1

A vector f = [ f0 , . . . , fN−1 ]> ∈ CN can be written as a linear combination of e(0) , . . . , e(N−1) . Using the orthogonality we
obtain for the coefficients fˆk
f , e(k)
N−1


1 N−1
f = ∑ fˆk e (k)
fˆk = (k) (k)
= ∑ f j e−2πik j/N (3)
k=0 e ,e N j=0

Given a vector f ∈ CN in the “time domain” we obtain a vector fˆ ∈ CN in the “frequency domain”. This operation is called
the discrete Fourier transform FN , its inverse is called FN−1 :

fˆ = FN f , f = FN−1 fˆ

For odd N = 2M + 1 we have

n o n o
CN = span e(0) , . . . , e(N−1) = span c(0) , c(1) , . . . , c(M) , s(1) , . . . , s(M)

We claim that the vectors c(0) , c(1) , . . . , c(M) , s(1) , . . . , s(M) also form an orthogonal basis. Using (1) it is easy to see that
these vectors are orthogonal on each other and
      N
c(0) , c(0) = N, c(k) , c(k) = s(k) , s(k) = for k = 1, 2, 3, . . .
2

1.4 Discrete Fourier transform in Matlab

>
The discrete Fourier transform FN maps the vector f = [ f0 , . . . , fN−1 ]> in the time domain to the vector fˆ = fˆ0 , . . . , fˆN−1


in the frequency domain. The inverse discrete Fourier transform FN−1 maps the vector fˆ back to the vector ~f :

fˆ = FN f , f = FN−1 fˆ

As an example consider the vector f = [1, 1, 1, 1]> in the time domain. Since this corresponds to the function v(t) = 1 the
Fourier vector is fˆ = [1, 0, 0, 0]:

F4 [1, 1, 1, 1]> = [1, 0, 0, 0]> , F4−1 [1, 0, 0, 0]> = [1, 1, 1, 1]> .

Matlab has the built-in functions fft and ifft for the discrete Fourier transform and its inverse. But they have the factors
N, N −1 reversed, compared with our definitions. Therefore we define
four = @(f)fft(f)/length(f);
ifour = @(f) ifft(f)*length(f);

Then we can compute the discrete Fourier transform in Matlab:

>> f = [1,1,1,1]
f = 1 1 1 1
>> fh = four(f)
fh = 1 0 0 0
>> ifour(fh)
ans = 1 1 1 1
In Matlab all indices are shifted by 1: You can access components of a Matlab vector v using v(1), v(2), ... .
Note that Matlab begins counting with 1, whereas our data and Fourier vectors are indexed from 0 to N − 1. Therefore fˆj
corresponds to fh(j+1) in Matlab. In our example we have fˆ = ( fˆ0 , fˆ1 , fˆ2 , fˆ3 ) = (1, 0, 0, 0) and fˆ0 corresponds to fh(1) in
Matlab.

1.5 Interpretation of fˆ0 , . . . , fˆN−1 and the interpolating function v(t)

Case of odd N
>
Assume we have for N = 5 a vector f = [ f0 , . . . , f4 ]> and we compute the discrete Fourier transform fˆ = fˆ0 , fˆ1 , fˆ2 , fˆ3 , fˆ4 .


This means that

f = fˆ0 e(0) + fˆ1 e(1) + fˆ2 e(2) + fˆ3 e(3) + fˆ4 e(4)
z}|{ z}|{
(0) (1) (2) (−2)
ˆ ˆ ˆ
= f0 e + f1 e + f2 e + f3 e ˆ + f4 e(−1)
ˆ
= a0 c(0) + a1 c(1) + a2 c(2) + b1 s(1) + b2 s(2)

using aliasing.
For [a0 , a1 , a2 , b1 , b2 ]> ∈ R5 we obviously have that f = a0 c(0) + a1 c(1) + a2 c(2) + b1 s(1) + b2 s(2) ∈ R5 . Since
c(0) , c(1) , c(2) , s(1) , s(2) are linearly independent vectors in R5 a vector f ∈ R5 corresponds to a unique vector
[a0 , a1 , a2 , b1 , b2 ]> ∈ R5 .
For a real vector f ∈ R5 the coefficients fˆ0 , . . . , fˆ4 will be complex in general, but the coefficients a0 , a1 , a2 , b1 , b2 will be
real.
Therefore the function

v(t) = fˆ0 + fˆ1 e2πit + fˆ2 e4πit + fˆ3 e−4πit + fˆ4 e−2πit
= a0 · 1 + a1 cos(2πt) + a2 cos(4πt) + b1 sin(2πt) + b2 sin(4πt)

is a 1-periodic trigonometric function with frequency ≤ 2 which satisfies v(t j ) = f j . This is called an interpolating function.
>
For any odd N = 2M + 1 the Fourier vector fˆ0 , fˆ1 , . . . , fˆN−1 yields the interpolating function v(t) given by


v(t) = fˆ0 + fˆ1 e2πit + · · · + fˆn eM2πit + fˆN−M e(−M)2πit + · · · + fˆN−1 u(−1)2πit (4)
= a0 · 1 + a1 cos(2πt) + · · · + aM cos(M2πt) + b1 sin(2πt) + · · · + b2 sin(4πt) (5)

Therefore the entries of the Fourier vector should be interpreted as follows:

frequency h 0 1 2 ··· M M ··· 2 1 i

fˆ0 fˆ1 fˆ2 · · · fˆM fˆM+1 · · · fˆN−2 fˆN−1

Here fˆ0 corresponds to frequency 0, it gives the mean value of our data. Then fˆ1 and fˆN−1 together tell us about frequency
1. The highest frequency terms fˆM , fˆM+1 for frequency M are in the middle of the vector.

Case of even N
>
Consider for N = 6 a discrete signal a vector ~f = [ f0 , . . . , f5 ] with the discrete Fourier transform fˆ = fˆ0 , fˆ1 , fˆ2 , fˆ3 , fˆ4 , fˆ5 .


This means that

f = fˆ0 e(0) + fˆ1 e(1) + fˆ2 e(2) + fˆ3 |{z}

e(3) + fˆ4 |{z}
e(4) + fˆ5 |{z}
e(5)
e(−3) e(−2) e(−1)
We would like obtain a “balanced formula” with respect to positive and negative k. What should we do with e(3) ? The answer
is
e(3) = 21 e(3) + 12 e(−3)
yielding

f = fˆ0 e(0) + fˆ1 e(1) + fˆ2 e(2) + fˆ3 e(3) + fˆ4 e(4) + fˆ5 e(5)
 
= fˆ0 e(0) + fˆ1 e(1) + fˆ2 e(2) + fˆ3 21 e(3) + e(−3) + fˆ4 e(−2) + fˆ5 e(−1)
= a0 c(0) + a1 c(1) + a2 c(2) + a3 c(3) + b1 s(1) + b2 s(2)

This corresponds to the 1-periodic trigonometric function

v(t) = fˆ0 + fˆ1 e2πit + fˆ2 e4πit + fˆ3 12 e6πit + e−6πit + fˆ4 e−4πit + fˆ5 e−2πit

= c0 · 1 + a1 cos(2πt) + a2 cos(4πt) + a3 cos(6πt) + b1 sin(2πt) + b2 sin(4πt)

which satisfies v(t j ) = f j . Note that for the highest frequency 3 we only use cos(6πt) and NOT sin(6πt). This makes
sense since sin(6πt) is zero at the points t j :

j
sin(3 · 2πt j ) = sin(3 · 2π ) = sin( jπ) = 0
6

For general even N = 2M the Fourier vector ( fˆ0 , fˆ1 , . . . , fˆN−1 ) yields the interpolating function

v(t) = fˆ0 + fˆ1 e2πit + · · · + fˆM−1 e(M−1)2πit + fˆM 21 eM2πit + e−M2πit + fˆM+1 e−(M−1)2πit + · · · + fˆN−1 e−2πit

= c0 · 1 + a1 cos(2πt) + · · · + aM cos(M2πt) + b1 sin(2πt) + · · · + bM−1 sin((M − 1)2πt)

Therefore the entries of the Fourier vector should be interpreted as follows:

frequency h 0 1 2 ··· M−1 M M −1 ··· 2 1 i

fˆ0 fˆ1 fˆ2 · · · fˆM−1 fˆM fˆM+1 · · · fˆN−2 fˆN−1

1.6 Properties of the discrete Fourier transform

Here we summarize the key properties:

1 N−1 N−1
fˆk = ∑ f j e−2πik j/N ⇔ fj = ∑ fˆk e2πik j/N Inversion
N j=0 k=0

1 N−1  2 N−1  2
∑ fj
N j=0
= ∑  fˆk  Parseval (6)
k=0

1 N−1
qj = ∑ f` g j−` ⇔ q̂k = fˆk ĝk Convolution
N `=0

Recall that (
  N for k = l
e(k) , e(`) =
0 for k 6= l

This gave us the formula for fˆk in (3), and we obtain from f = fˆ0 e(0) + · · · + fˆN−1 e(N−1)
N−1  2  (k) (k) 
(f, f) = ∑  fˆk  e ,e
k=0 | {z }
N
yielding the Parseval formula. In the same way we obtain

1 N−1 N−1
∑ j j ∑ fˆk ĝk
N j=0
f g =
k=0

For the convolution we plug in the Fourier representation of g j−` and switch the order of the sums:

1 N−1 1 N−1 N−1 2πi( j−`)k/N

qj = =
∑ ` j−` N ∑ f` ∑ ĝk e
f g
N `=0 `=0 k=0
!
N−1 N−1
qj = ∑ ∑ f` e−2πi`k/N ĝk e2πik j/N
k=0 `=0
| {z }
fˆk

Hence we have q j = ∑N−1 2πik j/N with q̂ = fˆ ĝ . Remember that the signals are periodic (wrap-around): g = g
k=0 q̂k e k k k j j+N . So
g j−` with j − ` < 0 has to be interpreted in this way.
1.7 Efficient computation of FN and FN−1 : Fast Fourier Transform
>
Assume we are given the vector fˆ = fˆ0 , . . . , fˆN−1 ∈ CN in the frequency domain. We want to find the vector f = FN−1 f =


[ f0 , . . . , fN−1 ]> ∈ CN in the time domain.

N−1
fj = ∑ fˆk e2πik j/N
k=0
 
1 1 1 ··· 1

 1 ω1 ω2 ··· ω N−1 

f = EN fˆ, EN = 
 1 ω2 ω4 ··· ω 2(N−1) 
 with ω := e2πi/N
 .. .. .. .. 
 . . . . 
1 ω N−1 ω 2(N−1) · · · ω (N−1)(N−1)

Since ω satisfies ω N = 1 all entries of EN are one of the N values 1, ω, . . . , ω N−1 .

E.g., for N = 4 we have ω = e2πi/4 = eiπ/2 = i and
 
1 1 1 1
 1 i −1 −i 
E4 =  
 1 −1 1 −1 
1 −i −1 i

If we naively evaluate the formula f = EN fˆ we have to compute N multiplications and N − 1 additions for each component,
in addition to computing ω, ω 2 , . . . , ω N−1 . This means that the computation of f costs at least N 2 multiplications (plus
a similar number of additions etc.). For a vector of length N = 1000 this means that we have at least 106 operations. This
means that it is not practical to compute discrete Fourier transforms of long vectors.
Fortunately there is a way to reorganize the computation in such a way that only N · log2 N operations are required. E.g., for
N = 1024 we have
N 2 ≈ 106 , N · log2 N = 1024 · 10 ≈ 104

Note that e2πik j/N = wNjk where we define

wN := e2πi/N

Assume that N is even. Then we can split up the sum in even and odd terms: Using wN/2 = w2N we obtain
N−1 N/2−1 N/2−1 N/2−1 N/2−1
j(2k) j(2k+1)
fj = ∑ wNjk fˆk = ∑ wN fˆ2k + ∑ wN fˆ2k+1 = ∑ jk ˆ
wN/2 f2k + wNj ∑ jk ˆ
wN/2 f2k+1
k=0 k=0 k=0 k=0 k=0
| {z } | {z }
aj bj
N/2−1 jk ˆ −1 ˆe
>
f2k for j = 0, . . . , N2 − 1: This is FN/2 f where fˆe := fˆ0 , fˆ2 , . . . , fˆN−2 ∈ CN/2 is the vector

Consider a j := ∑k=0 wN/2
N N jk N/2
of even Fourier coefficients. For j = 2 ,...,N − 1 we can write j = 2 + J and have wN/2 = wJK
N/2 since wN/2 = 1. Hence
a N +J = aJ , i.e.,
2  
a0
ye
 
.. −1 e
with ye := FN/2
= ŷ
 
 . ye
aN−1

N/2−1 >
wNj jk ˆ −1 ˆo
f2k+1 for j = 0, . . . , N2 − 1: Consider FN/2 f where fˆo := fˆ1 , fˆ3 , . . . , fˆN−1 ∈ CN/2 is the

Consider b j := ∑ wN/2
k=0
| {z }
cj
jk N/2
vector of odd Fourier coefficients. For j = N2 , . . . , N − 1 we can write j = N
2 + J and obtain with wN/2 = wJK
N/2 and wN = −1
that b N +J = −bJ , i.e.,
2

 
  1
b0 
yo
  wN 
..  −1 o
= with yo :=   FN/2 ŷ
  
. −yo ..
.

 
bN−1 N/2−1
wN

Hence we obtain for even N

 
1

ye + yo
  wN 
FN−1 fˆ = −1 e  −1 o
with ye := FN/2 ŷ , yo :=   FN/2 ŷ

ye − yo ..
 . 
N/2−1
wN

Let cN denote the number of operations needed to compute the vector FN−1 fˆ. We need to compute 2 transforms of length
1/2 N/2−1 N
N/2, wN := e2πi/N = wN/2 , N2 − 2 multiplications to get 1, wN , . . . , wN , 2 multiplications to get yo , and N additions to
e o
compute y ± y . Let 1 operation denote [1 addition/subtraction, 1 multiplication/division/square root and memory access
for reading writing], then
cN = 2cN/2 + N

This algorithm is implemented in the Matlab functions fft and ifft and therefore used in our functions four and ifour .

2 Applications

2.1 Least squares fit with trigonometric functions of frequency 0, 1, . . . , K

Assume we are given a signal u ∈ CN , and we want to find the best least square fit using frequencies 0, 1, . . . , K
n o n o
VK = span e(−K) , e(−K+1) , . . . , e(K) = span c(0) , c(1) , s(1) , . . . , c(K) , s(K)

where dimVK = 2K + 1 ≤ N. Applying the DFT gives û := FN u, and because of the orthogonality of e(k) we get the least
squares fit w ∈ VK given by

w := û0 e(0) + û1 e(1) + · · · + ûK e(K) + ûN−K e(−K) + · · · + ûN−1 e(−1)

This corresponds to the function

v(t) = û0 + û1 e2πit + · · · + ûK eK2πit + ûN−K e−K2πit + · · · + ûN−1 e−2πit

2.2 Any linear translation invariant operator becomes diagonal in frequency space

Consider the “translation operator” T which maps [u0 , . . . , uN−1 ]> to [uN , u0 , . . . , uN−1 ]> , this corresponds to the matrix
 
1
 1 
T =
 
.. 
 . 
1

A linear mapping A ∈ CN×N is called translation invariant if it satisfies AT v = TAv for all v ∈ CN , i.e.,

AT = TA
Note that
Te(k) = e2πik/N e(k) = ω k e(k) , ω := e2πi/N
I.e., the matrix T has the eigenvalues ω 0 , ω 1 , . . . , ω N−1 (these are the N different roots of 1), and corresponding eigenvectors
e(0) , . . . , e(N−1) .
Conversely, if we have Av = ω k v we must have v = ck e(k) with some ck ∈ C.
Let v := Ae(k) . Then
T v = TAe(k) = ATe(k) = Ae(k) ω k = ω k v
hence there is ck ∈ C with v = ck e(k) . Hence we have shown

Ae(k) = ck e(k)

i.e., e(0) , . . . , e(N−1) are eigenvectors of A.

For a vector
f = fˆ0 e(0) + · · · + fˆN−1 e(N−1)
we obtain
(0) (N−1)
g = A f = fˆ0 Ae ˆ
| {z } + · · · + fN−1 Ae
| {z }
c0 e(0) cN−1 e(N−1)
i.e., the operation g = A f in the time domain corresponds in the frequency domain to

fˆ0
    
ĝ0 c0
 ..   ..   .. 
ĝk = ck fˆk , . . . , ĝN−1 = cN−1 fˆN−1 , i.e.,  . = .  . 
ĝN−1 cN−1 fˆN−1

Result: Any translation

  invariant linear mapping A in the time domain corresponds to a diagonal matrix
c0
 ..  in the frequency domain. Hence in the frequency domain

 .
cN−1
• we can apply the mapping v = Au with N operations: v̂k = ck ûk for k = 0, . . . , N − 1
• we can apply the inverse mapping u = A−1 v with N operations: ûk = c−1
k v̂k for k = 0, . . . , N − 1 (if ck 6= 0 for all k)

In general computing v = Au in the time domain takes N 2 operations, and computing u = A−1 v in the time domain using
Gaussian elimination takes 13 N 3 operations. Since the cost of the DFT is only N log N, it is much more efficient to perform
these operations in the frequency domain.

2.3 Computing discrete convolutions and deconvolutions (periodic setting)

For two signals f , g ∈ CN we define the periodic discrete convolution by

1 N−1
vj = ∑ u` g j−`
N `=0

(note that g j is considered a periodic signal with g j−N = g j here). We can interpret this as applying a “filter” to the signal u.
This filter given by g j can e.g. be a smoothing filter. By using the DFT we can compute the filtered signal v using O(N log N)
operations.
Moreover, we can get back from the filtered signal v to the original signal u using O(N log N) operations if ĝk 6= 0 for all k.
This is called “deconvolution”.
2.4 Approximating the solution of a differential equations (periodic setting)

As an example consider the ordinary differential equation

−U 00 (t) + βU(t) = f (t) (7)
with β ≥ 0. Here F(t) is a given 1-periodic function. We want to find a 1-periodic function U(t) satisfying (8).
We assume that we are given the values f j = F( j/N) for j = 0, . . . , N − 1, and we want to find values u j approximating
U( j/N) for j = 0, . . . , N − 1.
We approximate U 00 (t j ) using a difference quotient
u j−1 − 2u j + u j+1
U 00 (t j ) ≈ , with h := N −1
h2
For
N−1
uj = ∑ ûk ek2πi jk/N
k=0
we obtain
N−1 N−1 N−1
u j+1 = ∑ ûk ek2πi( j+1)/N = ∑ ûk ek2πi/N ek2πi j/N , u j−1 = ∑ ûk e−k2πi/N ek2πi j/N
k=0 k=0 k=0
N−1  
ik2π/N −ik2π/N
u j+1 + u j−1 − 2u j = ∑ e +e − 2 ûk ek2πi j/N
k=0 | {z
  }
2 cos (k2π/N) − 1
Note that cost − 1 ≥ 0 and cost − 1 > 0 if t 6= j2π, j ∈ Z .
For −U 00 (t j ) + βU(t j ) we obtain the approximation
N−1
u j−1 − 2u j + u j+1
∑ Ck ûk ek2πi( j+1)/N with Ck := N 2 2 cos(k2π/N) − 1 +β
 
− + β u j =
h2 k=0 | {z }
≥0
This means that the operator in the time domain
u j−1 − 2u j + u j+1
fj = +uj (8)
h2
becomes
fˆk = Ck ûk
in the frequency domain. We are given the vector f = [ f0 , . . . , fN−1 ]> . Solving (8) in the time domain would require solving
an N × N linear system Au = f using elimination.
Case 1: β > 0 . In this case we have Ck 6= 0 for all k. Using DFT we can obtain the unique solution in the frequency domain
as
ûk := fˆk /Ck for k = 0, . . . , N − 1
This means that the matrix A corresponding to the linear system (8) is nonsingular.
Case 2: β = 0 . In this case we have C0 = 0, but Ck > 0 for k = 1, . . . , N − 1. Hence we must have

fˆ0 = 0 · û0 = 0
This means that in order to obtain a solution the vector f = [ f0 , . . . , fN−1 ]> must have sum zero.
If this condition is satisfied, we obtain the solution û = [û0 , . . . , ûN−1 ]> given by
û0 = arbitrary, ûk = fˆk /Ck for k = 1, . . . , N − 1
This means that we can add c · [1, . . . , 1]> with arbitrary c ∈ R to the solution vector.
Therefore the matrix A has the null space span {[1, . . . , 1]}, and we have (range A)> = span {[1, . . . , 1]}.
Note that for β = 0 we have the differential equation −U 00 (t) = f (t), and it is clear that for any solution U(t) also U(t) + c
is a solution.