Fourier transformation — A tutorial

Ronald Starke

TU Bergakademie Freiberg

Institute for Electrical Engineering TU Freiberg

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 1 / 75
 1 Fourier transformation on the real axis
Fundamentals
General properties
Kramers-Kronig relations
Effects on Fourier transforms
Connection to Laplace transform
Uncertainty relation
Connection to Fourier sine and cosine transform
2 Fourier transformation on an interval
Definition and properties
Connection to Fourier Sine and Cosine series
Connection to continuous Fourier transform
3 Discrete Fourier transformation
Definition and properties
Connection to Fourier series
Connection to continuous Fourier transformation
Shannon-Nyquist sampling theorem
4 Commemoratio
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 2 / 75
 Fourier transformation on the real axis

Section 1

Fourier transformation on the real axis

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 3 / 75
 Fourier transformation on the real axis Fundamentals

Integral operators
An integral operator is a mapping of functions f (t) 7→ g(t) defined by its
integral kernel F (t, t 0 ) as
Z ∞
g(t) = dt 0 F (t, t 0 )f (t 0 ) . (1)
−∞

In other words, the image function g(t) is obtained by integrating out the
preimage f (t) against a function of two variables, F = F (t, t 0 ). This
operation is linear,
α1 f1 (t) + α2 f2 (t) 7→ α1 g1 (t) + α2 g2 (t) , (2)
with α1 , α2 ∈ C, as follows from the linearity of the integral.

Integral operators are essentially the continuous analogue of linear

operators in linear algebra; the integral kernel corresponds to the matrix
representing the linear operator and the integration corresponds to the
matrix multiplication.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 4 / 75
 Fourier transformation on the real axis Fundamentals

Dirac delta distribution

By definition, the identity operator maps each function onto itself,
i.e. f (t) 7→ f (t) or g(t) = f (t). For both practical and conceptual reasons,
one sometimes wants to write this as an integral operator. The
corresponding integral kernel is called Dirac delta distribution (or Dirac
delta, for short),
Z ∞ Z
f (t) = 0 0 0
dt δ(t − t )f (t ) ≡ dt 0 δ(t − t 0 )f (t 0 ) . (3)
−∞

The Dirac delta is the continuous analogue of the Kronecker delta in linear
algebra, X
fi = δij fj , (4)
j

with 
1 , if i = j ,
δij = (5)
0 , if i 6= j .
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 5 / 75
 Fourier transformation on the real axis Fundamentals

Nascent Dirac delta (Dirac sequence)

Although the Dirac delta distribution constitutes the integral kernel of an
operator, it is not given by a bona fide function; it has to be approximated
by a (possibly continuous) sequence of functions in the sense of
Z Z
dt δ(t − t )f (t ) ≡ lim dt 0 δη (t − t 0 )f (t 0 ) ,
0 0 0
(6)
η

where depending on the representation, the (continuous or discrete)

parameter η might go to infinity or zero (or to some other value). The
crucial points here are:
1 Prior to performing the limit, the functions δη are entirely regular.
2 They become singular once the limit is performed.
3 In the representation of the Dirac delta, however, the integration is
performed first. The limit is performed afterwards.
4 This interchange of limes and integral makes the difference.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 6 / 75
 Fourier transformation on the real axis Fundamentals

Realization of nascent Dirac delta I

In order to qualify as a nascent Dirac delta, the sequence δη (τ ) should
(roughly speaking) for all values of the parameter η
1 integrate to unity,
Z
dτ δη (τ ) = 1 , (7)

2 diverge at τ = 0 and converge to zero everywhere else,


∞, τ = 0,
lim δη (τ ) = δ(τ ) = (8)
η 0 τ 6= 0 ,
3 be greater than or equal to zero everywhere, δη (τ ) ≥ 0.
(The third condition can actually be relaxed, while the second one is
heuristic and would require mathematical clarification.) Intuitively,
integrating against a nascent Dirac delta means averaging around every
point with the weighting function δη . In the limit, the weighting gets ever
stronger at τ = 0 and ever weaker everywhere else.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 7 / 75
 Fourier transformation on the real axis Fundamentals

Realization of nascent Dirac delta II

Dirac sequences can easily be realized by sequences of probability
distributions ever more peaked at the origin. Examples with η → 0 are:
1 the normal distribution (also Gaussian or Bell curve),

τ2
 
1
δη (τ ) = p exp − 2 , (9)
2πη 2 2η
2 the Cauchy-Lorentz distribution (also Breit-Wigner distribution),
1 η
δη (τ ) = , (10)
π η + τ2
2

3 the uniform distribution,

δη (τ ) = (2η)−1 χ[−η,η] (τ ) . (11)
Here χ denotes the characteristic function, i.e.

1 , if τ ∈ [−η, η] ,
χ[−η,η] (τ ) = (12)
0 , if τ ∈/ [−η, η] .
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 8 / 75
 Fourier transformation on the real axis Fundamentals

Sokhotski-Plemelj identity
An extremely useful formula is given by the distributional identity
1 1
= P ∓ iπδ(τ ) , (13)
τ ± iη τ
i.e. for sufficiently well-behaved functions f (τ ) we have
Z Z Z
f (τ ) f (τ )
lim dτ = P dτ − iπ dτ f (τ )δ(τ ) . (14)
η→0 τ + iη τ
Here, P denotes the so-called principal value,
Z Z
f (τ ) def f (τ )
P dτ = lim dτ . (15)
τ η→0 R\[−η,η] τ
Essentially, this follows from the algebraic identity
1 τ η
= 2 −i 2 , (16)
τ + iη τ + η2 τ + η2
together with Eq. (10).
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 9 / 75
 Fourier transformation on the real axis Fundamentals

Fourier representation of Dirac delta I

The central theorem which encapsulates almost the entire Fourier

transformation theory is given by the Fourier representation of the Dirac
delta, Z ∞
dω
δ(t − t 0 ) = exp − iω(t − t 0 ) .


(17)
−∞ 2π
This is one of the most profound, subtle, astonishing, practical and
beautiful formulae in all of mathematics.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 10 / 75
 Fourier transformation on the real axis Fundamentals

Fourier representation of Dirac delta II

Essentially, Eq. (17) follows from the fact that
Z
def dω

δη (τ ) = exp − iωτ − η|ω| (18)
2π
constitutes a nascent Dirac delta in the limit η → 0. In fact, with

ω, ω > 0,
|ω| = (19)
−ω , ω < 0 ,
we get
Z
dω

δη (τ ) = exp − iωτ − η|ω| (20)
2π
Z 0 Z ∞
dω
dω

= exp − iωτ + ηω + exp − iωτ − ηω (21)
−∞ 2π 0 2π
 
1 1 1 1 η
= − = , (22)
2π −iτ + η −iτ − η π η + τ2
2

and this is the Cauchy-Lorentz distribution, Eq. (10).

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 11 / 75
 Fourier transformation on the real axis Fundamentals

Definition of Fourier transform

In order to motivate the definition of the Fourier transform, we consider
Z
f (t) = dt 0 δ(t − t 0 )f (t 0 ) (23)
Z Z 
dω −iω(t−t 0 )
= dt 0 e f (t 0 ) (24)
2π
dt 0 iωt 0 0
Z Z 
dω −iωt
= √ e √ e f (t ) . (25)
2π 2π
This makes it plausible to define the Fourier transform as
Z ∞
˜ dt
f (ω) = √ eiωt f (t) , (26)
−∞ 2π
such that the original function can be expressed in terms of the Fourier
transform in the form
Z ∞
dω
f (t) = √ e−iωt f˜(ω) . (27)
−∞ 2π
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 12 / 75
 Fourier transformation on the real axis Fundamentals

General logic of Fourier transformation

1 The Fourier transformation is an invertible mapping of (sufficently
regular) functions which associates with every complex valued
function on the real axis f = f (t) its Fourier transform f˜ = f˜(ω).
2 This Fourier transform f˜ = f˜(ω) is again a complex valued function
on the real axis.
3 The Fourier representation, Eq. (27), of a function in terms of its
Fourier transform, Eq. (26), amounts to a continuous superposition of
harmonic oscillations.
4 Physically, we therefore think of f (t) as a time signal, with t being
the time variable, such that f˜(ω) represents its frequency spectrum.
5 Correspondingly, ω can be considered an angular frequency. In
particular, the argument of the exponential, ±iωt, is then a
dimensionless quantity.
6 Other interpretations as e.g. in terms of positions and wave-vectors
are also possible.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 13 / 75
 Fourier transformation on the real axis General properties

Linearity
The mapping of signals f (t) to their spectra f˜(ω) is linear, i.e. for

f (t) = c1 f1 (t) + c2 f2 (t) (28)

the Fourier transform is given by

f˜(ω) = c1 f˜1 (ω) + c2 f˜2 (ω) . (29)

This simply follows from the fact the the integral constitutes a linear
functional of the integrand. Concretely, in our case we have
Z
dt
f˜(ω) = √ eiωt c1 f1 (t) + c2 f2 (t)


(30)
2π
Z Z
dt iωt dt
= c1 √ e f1 (t) + c2 √ eiωt f2 (t) (31)
2π 2π
= c1 f˜1 (ω) + c2 f˜2 (ω) . (32)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 14 / 75
 Fourier transformation on the real axis General properties

Preservation of scalar product

One of the most important properties of the Fourier transformation is the
preserveration of scalar products (theorem of Parseval),
Z Z Z
dt f ∗ (t)g(t) = dt dt 0 f ∗ (t)δ(t − t 0 )g(t 0 ) (33)
Z Z Z 
0 ∗ dω −iω(t−t 0 )
= dt dt f (t) e g(t 0 ) (34)
2π
dt 0 iωt 0 0
Z Z  Z 
dt −iωt ∗
= dω √ e f (t) √ e g(t ) (35)
2π 2π
Z
= dω f ∗ (ω)g(ω) . (36)

In particular, this implies for the norm of a function (theorem of

Plancherel)
Z 1/2
def ∗
kf k = dt f (t)f (t) = ke
fk. (37)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 15 / 75
 Fourier transformation on the real axis General properties

Real and imaginary functions

If the Fourier preimage is given by a real function,
f ∗ (t) = f (t) , (38)
we find from the defining equation (27) of the Fourier representation
Z
dω
f ∗ (t) = √ eiωt (f˜)∗ (ω) (39)
2π
Z
dω −iωt ˜ ∗
= √ e (f ) (−ω) [ω 7→ −ω] (40)
2π
Z
! dω
= √ e−iωt f (ω) . (41)
2π
Hence, the Fourier transform of a real function fulfills
(f˜)∗ (−ω) = f˜(ω) . (42)
Similarly, we get for a purely imaginary function
f ∗ (t) = −f (t) , (f˜)∗ (−ω) = −f˜(ω) . (43)
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 16 / 75
 Fourier transformation on the real axis General properties

Even and odd functions

If the Fourier preimage is given by an even function,
f (−t) = f (t) , (44)
we find
Z
dω
f (−t) = √ eiωt f˜(ω) (45)
2π
Z
dω −iωt ˜
= √ e f (−ω) (46)
2π
Z
! dω
= √ e−iωt f˜(ω) . (47)
2π
Hence, we get
f (ω) = f (−ω) , (48)
i.e. the Fourier transform is even as well. Similarly, for odd functions
f (−t) = −f (t) , f˜(−ω) = −f˜(ω) . (49)
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 17 / 75
 Fourier transformation on the real axis General properties

Pointwise products and convolutions I

If the Fourier preimage is given by a point-wise product,
f (t) = g(t)h(t) , (50)
we find for the Fourier transform
Z Z
dt
˜
f (ω) = √ dt 0 eiωt g(t)δ(t − t 0 )h(t 0 ) (51)
2π
dω 0 −iω0 (t−t 0 )
Z Z Z 
dt 0 iωt
= √ dt e g(t) e h(t 0 ) (52)
2π 2π
dω 0 dt 0
Z Z Z
dt 0 0 0
= √ √ √ ei(ω−ω )t g(t)eiω t h(t 0 ) (53)
2π 2π 2π
0 dt 0
Z Z  Z 
dω dt 0 0 0
= √ √ ei(ω−ω )t g(t) √ eiω t h(t 0 ) (54)
2π 2π 2π
dω 0
Z
= √ g̃(ω − ω 0 )h̃(ω 0 ) . (55)
2π
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 18 / 75
 Fourier transformation on the real axis General properties

Pointwise products and convolutions II

Conversely, if the Fourier preimage is given by a convolution,
Z
f (t) = dt 0 g(t − t 0 )h(t 0 ) , (56)

we find for the Fourier transform

Z Z
dt
˜
f (ω) = √ dt 0 eiωt g(t − t 0 )h(t 0 ) (57)
2π
√ Z dt Z dt 0 iω(t−t 0 ) 0
= 2π √ √ e g(t − t 0 )eiωt h(t 0 ) (58)
2π 2π
√ Z dt 0 iωt 0 0 Z dt iω(t−t 0 )
 
0
= 2π √ e h(t ) √ e g(t − t ) (59)
2π 2π
√
= 2π g̃(ω)h̃(ω) . (60)

Put differently, the Fourier transformation converts products into

convolutions and vice versa.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 19 / 75
 Fourier transformation on the real axis General properties

Important Fourier transforms I

From the Fourier representation of the Dirac delta, it follows directly that
a constant function,
f (t) = c , (61)
has the Fourier transform
Z
dt
f˜(ω) = √ eiωt c (62)
2π
√
= c 2π δ(ω) . (63)
Similarly, harmonic oscillations,
f (t) = e−iω0 t , (64)
have a peaked spectrum
Z
dt
f˜(ω) = √ ei(ω−ω0 )t (65)
2π
√
= 2π δ(ω − ω0 ) . (66)
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 20 / 75
 Fourier transformation on the real axis General properties

Important Fourier transforms II

Counterintuitively for the beginner, a sinusoidal signal,

f (t) = sin ω0 t , (67)

oscillates with two frequencies,

Z
˜ dt
f (ω) = √ eiωt sin ω0 t (68)
2π
dt iωt eiω0 t − e−iω0 t
Z  
= √ e (69)
2π 2i
1 √

= 2π δ(ω + ω0 ) − δ(ω − ω0 ) . (70)
2i
Similarly, we find for the cosine,
1√
f˜(ω) =


f (t) = cos ω0 t , 2π δ(ω + ω0 ) + δ(ω − ω0 ) . (71)
2
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 21 / 75
 Fourier transformation on the real axis General properties

Important Fourier transforms III

Another important example is the damped oscillation switched on at t = 0,

f (t) = Θ(t) e−iω0 t−ηt , (72)

with the Heaviside theta function,


1 , if t ≥ 0 ,
Θ(t) = (73)
0 , if t < 0 .
Its Fourier transform can be calculated directly,
Z ∞
dt
f˜(ω) = √ ei(ω−ω0 )t−ηt (74)
0 2π
1 −1
=√ (75)
2π i(ω − ω0 ) − η
1 i
=√ . (76)
2π ω − ω 0 + iη

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 22 / 75
 Fourier transformation on the real axis General properties

Important Fourier transforms IV

The Fourier transform of the Heaviside function itself can be obtained

from this in the limit η → 0 by setting ω0 = 0,
1 i
f˜(ω) = √ . (77)
2π ω + iη
This has to be understood in the sense of
dω e−iωt
Z
Θ(t) = lim i , (78)
η→0 2π ω + iη
i.e. once more one has to integrate before the limit is performed. A direct
verification of this formula via the explicit evaluation of the integral
requires tools from complex analysis (theory of holomorphic functions) and
is therefore omitted here.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 23 / 75
 Fourier transformation on the real axis Kramers-Kronig relations

Kramers-Kronig relations I

A particularly important class of signals is given by the “retarded”

functions, which—by definition—are identically zero on the negative axis,
i.e. f (t) = 0 for t < 0. Trivially, these fulfill

f (t) = Θ(t)f (t) , (79)

because for t > 0 the Heaviside function multiplies with unity. In the
frequency domain, this point-wise product translates into a convolution
(see Eq. (55)). Using Eq. (78), we get

dω 0 i f˜(ω 0 )
Z
f˜(ω) = √ √ (80)
2π 2π ω − ω 0 + iη
dω 0 f˜(ω 0 )
Z
=i . (81)
2π ω − ω 0 + iη

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 24 / 75
 Fourier transformation on the real axis Kramers-Kronig relations

Kramers-Kronig relations II
We apply the Sokhotski-Plemelj identity, Eq. (13), in the form
1 1
0
=P − iπδ(ω − ω 0 ) . (82)
ω − ω + iη ω − ω0
Evaluating the Dirac delta, we then find

dω 0 f˜(ω 0 )
Z
˜
f (ω) = i P . (83)
π ω − ω0
Considering real and imaginary part separately yields

dω 0 Im f˜(ω 0 )
Z
Re f˜(ω) = −P , (84)
π ω − ω0
dω 0 Re f˜(ω 0 )
Z
˜
Im f (ω) = P (85)
π ω − ω0
These are the celebrated Kramer-Kronig relations.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 25 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Fourier transform of derivative

By acting on the defining equation of the Fourier representation,
Z
d dω
f (t) = √ e−iωt (−iω)f˜(ω) , (86)
dt 2π
we find directly the effect of a derivative on the Fourier transform,

g(t) = f 0 (t) , g̃(ω) = −iω f˜(ω) . (87)

Conversely, derivatives of the Fourier transform,

Z
d ˜ dt
f (ω) = √ eiωt it f˜(t) , (88)
dω 2π
translate into multiplicative operations in the time-domain,
1 d ˜
g(t) = t f (t) , g̃(ω) = f (ω) . (89)
i dω
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 26 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Fourier transform of complex conjugate

For the Fourier transform of the complex conjugate function, we get by
evaluating the defining equation
Z
dt
∗
f (ω) =
e √ eiωt f ∗ (t) (90)
2π
Z ∗
dt
= √ e−iωt f (t) (91)
2π
Z ∗
dt i(−ω)t
= √ e f (t) (92)
2π
 ∗
= f˜(−ω) . (93)

Hence, we read off the general relation

g(t) = f ∗ (t) , g̃(ω) = f ∗ (−ω) . (94)
In case f (t) is real or imaginary valued, we recover Eqs. (42) or (43)
respectively.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 27 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Fourier transform of time shifted function

If the argument of the time signal is shifted,

g(t) = f (t − t0 ) , (95)

we find for the resulting Fourier transform

Z
dt
g̃(ω) = √ eiωt f (t − t0 ) (96)
2π
Z
dt
= eiωt0 √ eiω(t−t0 ) f (t − t0 ) (97)
2π
iωt0 ˜
= e f (ω) . (98)

In other words, a shift in the argument leads to a phase factor in the

Fourier domain.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 28 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Fourier transform of frequency modulated function

Conversely, multiplying the time signal by a harmonic function,

g(t) = e−iω0 t f (t) , (99)

shifts the Fourier transform instead,

Z
dt
g̃(ω) = √ ei(ω−ω0 )t f (t) (100)
2π
= f (ω − ω0 ) . (101)

This is the inverse effect to the shift in the time signal. Note, however, the
difference in the presigns of the respective exponential.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 29 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Consistency check
The modulation with an harmonic function as in Eq. (99) can be
interpreted as a pointwise product in the sense of Eq. (50).
Correspondingly, we now reproduce its effect on the Fourier transform,
Eq. (101), by using the result for the Fourier transform of pointwise
products, Eq. (55), and the Fourier transform of a harmonic funtion,
Eq. (66). With
√
f (t) = g(t)h(t) , g(t) = e−iω0 t , g̃(ω) = 2π δ(ω − ω0 ) , (102)
we find
dω 0
Z
f˜(ω) = √ g̃(ω − ω 0 )h̃(ω 0 ) (103)
Z 2π
= dω 0 δ(ω − ω 0 − ω0 )h̃(ω 0 ) (104)

= h̃(ω − ω0 ) . (105)
This coincides indeed exactly with our earlier result.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 30 / 75
 Fourier transformation on the real axis Effects on Fourier transforms

Fourier transform of scaled function

The effect of a rescaling in the time signal,

g(t) = f (ct) , (106)

can be calculated directly by a change in the integration variable,

Z
dt
g̃(ω) = √ eiωt f (ct) (107)
2π
Z
1 d(ct) i(ω/c)ct
= √ e f (ct) (108)
c 2π
1
= f˜(ω/c) . (109)
c
Hence, in the Fourier domain, it produces the inverse scaling up to
prefactor.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 31 / 75
 Fourier transformation on the real axis Connection to Laplace transform

Connection to Laplace transform I

We consider again the Fourier transform of a retarded function
(i.e. f (t) = 0 for t < 0),
Z ∞ Z ∞
˜ dt iωt dt
f (ω) = √ e f (t) = √ eiωt f (t) . (110)
−∞ 2π 0 2π
As the integration now extends only along the positive axis, we may
perform the transition to complex frequencies with positive imaginary part,
Z ∞ Z ∞
˜ dt i(ω+iη)t dt
f (ω + iη) = √ e f (t) = √ eiωt−ηt f (t) . (111)
0 2π 0 2π
Introducing the new variable
s = −i(ω + iη) = −iω + η , (112)
this can be written as
Z ∞
dt
f¯(s) = √ e−st f (t) . (113)
0 2π
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 32 / 75
 Fourier transformation on the real axis Connection to Laplace transform

Connection to Laplace transform II

This is none other than the Laplace transform f¯ of the time signal f , i.e.
f˜(ω + iη) = f˜(is) = f¯(s) . (114)
The inverse Laplace transform is typically written as the so-called
Bromvich–Mellin integral
1 η+i∞ ds st ¯
Z
f (t) = √ e f (s) , (115)
i η−i∞ 2π
where η has to be chosen sufficiently large in order to avoid all singularities
of the Laplace transform. Restoring the original variables, this can be
written as Z ∞
dω
f (t) = √ e−iωt+ηt f˜(ω + iη) . (116)
−∞ 2π
One shows directly that this indeed yields the inverse Fourier transform
independently of the precise value of η. In particular, for well-behaved
functions, η can be set to zero.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 33 / 75
 Fourier transformation on the real axis Connection to Laplace transform

Connection to Laplace transform III

Concretely, we find
Z ∞
dω
f (t) = √ e−iωt+ηt f˜(ω + iη) (117)
−∞ 2π
Z ∞
dω −iωt+ηt ∞ dt 0 iωt 0 −ηt 0 0
Z
= √ e √ e f (t ) (118)
−∞ 2π 0 2π
Z ∞ Z ∞ 
0 dω −iω(t−t 0 ) 0
= dt e eη(t−t ) f (t 0 ) (119)
−∞ 2π
Z0 ∞
0
= dt 0 δ(t − t 0 )eη(t−t ) f (t 0 ) (120)
0
= Θ(t)f (t) , (121)

because for t < 0, the support of the Dirac delta lies outside the
integration domain.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 34 / 75
 Fourier transformation on the real axis Connection to Laplace transform

Connection to Laplace transform IV

1 The Fourier transform of a retarded function is the limiting case of its

Laplace transform with the imaginary part of the frequency going to
zero.
2 The Laplace transform of a (not necessarily retarded) function is the
Fourier transform of its retarded part (function set to zero on the
negative axis; f (t) 7→ Θ(t)f (t)) continued to complex frequencies
whose imaginary part is positive.
3 The inverse Laplace transform yields back the retarded part of a
function, Θ(t)f (t).
4 For retarded functions, the inverse Laplace transform therefore yields
back the original function.
5 For regular, retarded functions, the imaginary part of the frequency
can be set to zero such that the inverse Laplace transform formally
coincides with the inverse Fourier transform.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 35 / 75
 Fourier transformation on the real axis Uncertainty relation

Gaussian function I
Consider again a Gaussian function,

t2
 
1
f (t) = exp − 2 . (122)
(2πσt2 )1/4 4σt

We want to show that its Fourier transform is of the same form,

ω2
 
˜ 1
f (ω) = exp − 2 , (123)
(2πσω2 )1/4 4σω

with
1
σω = . (124)
2σt
Observing, that f ∗ (t)f (t) is a normal distribution with variance σt2 and
similarly f ∗ (ω)f (ω) with σω2 , this implies that the deviations are inversely
proportional to each other, i.e. the more localized the one, the more
delocalized the other.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 36 / 75
 Fourier transformation on the real axis Uncertainty relation

Gaussian function II
To convince ourselves, consider

t2
   
d 1 t
f (t) = − 2 exp − 2 . (125)
dt (2πσt2 )1/4 2σt 4σt

Hence, f is the unique solution to the differential equation

d t f (t)
f (t) + = 0, (126)
dt 2σt2

with the initial condition

1
f (0) = . (127)
(2πσt2 )1/4

We recall that as a linear differential equation of first order with constant

coefficents, the solution space of Eq. (126) is one-dimensional, hence all
solutions are proportional to the Gaussian.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 37 / 75
 Fourier transformation on the real axis Uncertainty relation

Gaussian function III

Taking the Fourier transform of the differential equation, Eq. (126), we get
with Eqs. (87) and (89)

1 1 d ˜
−iω f˜(ω) + f (ω) = 0 . (128)
2σt2 i dω

This is equivalent to

d ˜ 4σ 2
f (ω) + t ω f˜(ω) = 0 . (129)
dω 2
This is again of the form (126), but with the deviation given by Eq. (124).
Hence, f˜ is again given by a Gaussian. The prefactor is given such that the
Fourier transform is again normalized to one in the sense of kf˜k = 1 as it
has to be on account of the Plancherel theorem, Eq. (37).

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 38 / 75
 Fourier transformation on the real axis Uncertainty relation

Gaussian function IV

The prefactor can also be verified explicitly by

Z
dω
f (0) = √ f˜(ω) (130)
2π
ω2
Z  
1 dω
= √ exp − (131)
(2πσω2 )1/4 2π 2(21/2 σω )2
1 −1/2

1/2 2
1/2
= (2π) 2π(2 σ ω ) (132)
(2πσω2 )1/4
= 2 /4 π − /4 σω/2
1 1 1
(133)
−1/4
= (2πσt2 ) , (134)

and this coincides with Eq. (127). Here, we have used the fact that the
normal distribution is normalized.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 39 / 75
 Fourier transformation on the real axis Uncertainty relation

Uncertainty relation I
We consider a normalized time signal,
Z Z
kf k2 = dt f ∗ (t)f (t) = dt |f (t)|2 = 1 . (135)

By the theorem of Plancherel, Eq. (37), the same will then hold for the
Fourier transform, Z
2
kf˜k2 = dω f˜(ω) = 1 . (136)

First, we assume that the expectation values of both the time signal and
its Fourier transformation vanish,
Z Z
2
dt t|f (t)| = 0 , dω ω|f˜(ω)|2 = 0 . (137)

The respective variances are then given by

Z Z
σt2 = dt t 2 |f (t)|2 , σω2 = dω ω 2 |f˜(ω)|2 . (138)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 40 / 75
 Fourier transformation on the real axis Uncertainty relation

Uncertainty relation II
Consider now the manifestly positive definite function
Z
d ˜ 2
I (λ) = dω ω f˜(ω) + λ f (ω) ≥ 0 , (139)
dω
with λ ∈ R. Multiplying out brings it to the form

I (λ) = aλ2 + bλ + c ≥ 0 , (140)

with
Z
d ˜ 2
a= dω
f (ω) , (141)
dω
Z  
∗ d d ∗
b = dω ω (f˜) (ω) f˜(ω) + f˜(ω) (f˜) (ω) , (142)
dω dω
Z
2
c = dω ω f˜(ω) . (143)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 41 / 75
 Fourier transformation on the real axis Uncertainty relation

Uncertainty relation III

Now, consider
Z
d ˜∗
(f ) (ω)f˜(ω)


b= dω ω (144)
dω
Z
= − dω (f˜)∗ (ω)f˜(ω) (145)

= −1 , (146)
where we have integrated by parts and used the fact that boundary terms
vanish because the spectral function has to decay sufficiently fast.
Furthermore, we obviously have
c = σω2 . (147)
Finally, by going back to the time domain and observing that Fourier
transforming translates multiplications into derivatives, we find
Z
2
a = dt t f (t) = σt2 . (148)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 42 / 75
 Fourier transformation on the real axis Uncertainty relation

Uncertainty relation IV

The condition I (λ) ≥ 0 precludes the existence of two different,

real-valued zeros, which translates into the inequality

b 2 − 4ca ≤ 0 . (149)

Using the identifications of the coefficients, Eqs. (141)–(143), this means

1
σt2 σω2 ≥ . (150)
4
This is the uncertainty relation. It is fully analogous to its counterpart in
quantum mechanics (Heisenberg uncertainty relation for position and
momentum). Essentially, it means that the more localized the time signal,
the more delocalized its spectrum and vice versa.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 43 / 75
 Fourier transformation on the real axis Uncertainty relation

Uncertainty relation V

In the general case, where the expectation value of the time signal
f = f (t) is t0 , while the spectrum is concentrated at ω0 , we simply observe
that now the function g defined by

g(t) = eiω0 t f (t + t0 ) , g̃(ω) = e−i(ω+ω0 )t0 f˜(ω + ω0 ) , (151)

has its respective expectation values at the origin (see Eqs. (98) and
(101)), such that the uncertainty (150) relation applies. On the other
hand, the respective variances of f and g and their Fourier transforms also
coincide, because shifting the variable and multiplication with phase factors
has no bearing on these. This proves the general uncertainty relation.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 44 / 75
 Fourier transformation on the real axis Connection to Fourier sine and cosine transform

Real and imaginary part of Fourier transform

For purposes which will become clear immediately, we now write the
Fourier transform as
1
f˜(ω) = a(ω) + ib(ω) ,


(152)
2
with real-valued functions a(ω) and b(ω), i.e.

a(ω) = 2 Re f˜(ω) , b(ω) = 2 Im f˜(ω) . (153)

The reality condition in the Fourier domain,

f (ω) = f ∗ (−ω) , (154)

then translates into

a(ω) = a(−ω) , (155)

b(ω) = −b(−ω) . (156)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 45 / 75
 Fourier transformation on the real axis Connection to Fourier sine and cosine transform

Representation in terms of sine and cosine oscillations

The Fourier transform of a real function can now be written as
Z ∞
dω
f (t) = √ e−iωt f˜(ω) (157)
−∞ 2π
1 ∞ dω
Z



= √ cos ωt − i sin ωt a(ω) + ib(ω) (158)
2 −∞ 2π
Z
1 dω 


= √ a(ω) cos ωt + b(ω) sin ωt + i b(ω) cos ωt − a(ω) sin ωt
2 2π
1 ∞ dω
Z


= √ a(ω) cos ωt + b(ω) sin ωt (159)
2 −∞ 2π
Z ∞
dω

= √ a(ω) cos ωt + b(ω) sin ωt . (160)
0 2π
Here, the imaginary part of the integrand has been dropped as the integral
of an even function multiplied by an odd function vanishes. (It also has to
be zero on grounds of consistency, because f (t) was assumed to be real.)
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 46 / 75
 Fourier transformation on the real axis Connection to Fourier sine and cosine transform

Fourier sine and cosine transform

Conversely, by applying the Euler theorem we find directly
Z
˜ dt
f (ω) = √ eiωt f (t) (161)
2π
Z
dt

= √ cos ωt + i sin ωt f (t) (162)
2π
! 1

= a(ω) + ib(ω) . (163)
2
Hence, the cosine and sine transform are explicitly given by
Z
dt
a(ω) = 2 √ cos ωt f (t) , (164)
2π
Z
dt
b(ω) = 2 √ sin ωt f (t) . (165)
2π
We stress that the Fourier sine and cosine transformation deals only with
positive frequencies.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 47 / 75
 Fourier transformation on the real axis Connection to Fourier sine and cosine transform

Amplitude and phase

Finally, writing the Fourier transform in terms of an amplitude and a phase
angle,

a(ω) = A(ω) cos ϕ(ω) , b(ω) = −A(ω) sin ϕ(ω) , (166)

we get
Z ∞
dω

f (t) = √ a(ω) cos ωt + b(ω) sin ωt (167)
0 2π
Z ∞
dω

= √ A(ω) cos ϕ(ω) cos ωt − sin ϕ(ω) sin ωt (168)
2π
Z0 ∞
dω

= √ A(ω) cos ωt + ϕ(ω) . (169)
0 2π
In this representation, the time signal is written as a continuous
superposition of harmonic functions whose amplitude and phase angle
depend on the respective frequency.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 48 / 75
 Fourier transformation on an interval

Section 2

Fourier transformation on an interval

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 49 / 75
 Fourier transformation on an interval Definition and properties

Periodic functions I
A somewhat different notion of Fourier transformation is used in the case
of periodic functions,
f (t + T ) = f (t) . (170)
These have a Fourier series representation,
1 X −iωn t ˜
f (t) = √ e fn , (171)
T n
where the discrete frequencies are given by
2πn
, n ∈ Z,
ωn = (172)
T
i.e. the index n traverses all integers. The qualitatively new feature here
consists in the fact that the time signal is represented by a discrete
summation of harmonic functions while in the case of the Fourier
transformation on the real axis, the time signal is represented by a
continuous integration of such contributions.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 50 / 75
 Fourier transformation on an interval Definition and properties

Periodic functions II
In Eq. (171), the contribution from n = 0 corresponds to a time
independent constant. ω1 = 2π/T is the so-called fundamental frequency
(Grundschwingung), while the ωn ; |n| > 1 are the higher harmonics
(Oberschwingungen). The frequencies are chosen in this way in order to
ensure the periodicity of the exponentials,
e−iωn (t+T ) = e−iωn t−iωn T (173)
= e−iωn t e−i2πn (174)
−iωn t
=e . (175)
Conversely, the Fourier amplitudes can be obtained from
Z T
˜ 1
fn = √ dt eiωn t f (t) , (176)
T 0
i.e. the Fourier expansion coefficents are already determined by the
behaviour of the function within a periodicity interval.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 51 / 75
 Fourier transformation on an interval Definition and properties

Resolution of identity
One verifies directly the representation of the Kronecker delta,
1 T
Z
δnm = dt ei(ωn −ωm )t , (177)
T 0
Conversely, an integral kernel for the identity operator is given by
1 X −iωn (t−t 0 )
δT (t − t 0 ) = e . (178)
T
n∈Z
Its action is restricted to sufficiently well-behaved periodic functions,
Z T
f (t) = dt 0 δT (t − t 0 )f (t 0 ) . (179)
0
Analoga of the formulae for the continuous Fourier transform, as e.g.
Z T X
dt f ∗ (t)g(t) = f˜n∗ g̃n , (180)
0 n∈Z
can easily be found.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 52 / 75
 Fourier transformation on an interval Connection to Fourier Sine and Cosine series

Real functions
In the case of a Fourier series, the reality condition reads
∗
f˜n = f˜−n . (181)

The time signal can then be reworked into

1 X −iωn t ˜
f (t) = √ e fn (182)
T n∈Z
∞
1 ˜ 1 X −iωn t ˜
f n + eiωn t f˜−n


= √ f0 + √ e (183)
T T n=1
∞
1 ˜ 1 X −iωn t ˜
f n + eiωn t f˜n∗


= √ f0 + √ e (184)
T T n=1
∞
1 2 X
= √ f˜0 + √ Re e−iωn t f˜n . (185)
T T n=1

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 53 / 75
 Fourier transformation on an interval Connection to Fourier Sine and Cosine series

Fourier sine and cosine series

Introducing again (see Eq. (152))
1
f˜n = (an + ibn ) , (186)
2
and using
  1
Re e−iωn t f˜n = Re (cos ωn t − i sin ωn t)(an + ibn )


(187)
2
1
= (an cos ωn t + bn sin ωn t) , (188)
2
we get
∞
1 ˜ 1 X

f (t) = √ f0 + √ an cos ωn t + bn sin ωn t . (189)
T T n=1

This is the Fourier sine and cosine series.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 54 / 75
 Fourier transformation on an interval Connection to Fourier Sine and Cosine series

Fourier sine and cosine coefficients I

Conversely, we calculate the Fourier coefficients by
Z T
1
f˜n = √ dt eiωn t f (t) (190)
T 0
Z T
1

=√ dt cos ωn t + i sin ωn t f (t) (191)
T 0
! 1
= (an + ibn ) . (192)
2
From this, we read off for n > 0
Z T Z T
2 2
an = √ dt f (t) cos ωn t , bn = √ dt f (t) sin ωn t , (193)
T 0 T 0

and for n = 0,

a0 = 2f˜0 , b0 = 0 . (194)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 55 / 75
 Fourier transformation on an interval Connection to Fourier Sine and Cosine series

Fourier sine and cosine coefficients II

Alternatively, consider the orthogonality relations
Z T
T
dt cos ωn t cos ωm t = δnm , (195)
0 2
Z T
T
dt sin ωn t sin ωm t = δnm , (196)
0 2
Z T
dt sin ωn t cos ωm t = 0 . (197)
0
These, in turn, follow from the identities
1

cos ωn t cos ωm t = cos(ωn − ωm )t + cos(ωn + ωm )t , (198)
2
1

sin ωn t sin ωm t = cos(ωn − ωm )t − cos(ωn + ωm )t , (199)
2
1

sin ωn t cos ωm t = sin(ωn − ωm )t + sin(ωn + ωm )t . (200)
2
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 56 / 75
 Fourier transformation on an interval Connection to Fourier Sine and Cosine series

Fourier sine and cosine coefficients III

The formulae for the Fourier sine and cosine coefficents can now also be
verified from
Z T
2
an = √ dt f (t) cos ωn t ,
T 0
∞
Z T !
2 1 1 X

=√ dt √ f0 + √ am cos ωm t + bn sin ωm t cos ωn t ,
T 0 T T m=1

by mutliplying out and using the orthogonality conditions. Equally one

proceeds in the case of bn . Hereby, we observe that the integral of a
trigonometric function over an entire period vanishes.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 57 / 75
 Fourier transformation on an interval Connection to continuous Fourier transform

Connection to Fourier transform I

A periodic function with the interval [0, T ] has the singular Fourier
transform
Z !
dω 1 X
fe(ω) = √ eiωt √ f˜n e−iωn t (201)
2π T n
 1/2 X
2π
= f˜n δ(ω − ωn ) . (202)
T n
Conversely, one obtains (cf. Eq. (171))
Z  1/2 X !
dω −iωt 2π
f (t) = √ e f˜n δ(ω − ωn ) (203)
2π T n
1 X ˜ −iωn t
=√ fn e . (204)
T n
Note, however, that the continuous Fourier transform of a periodic
function is necessarily singular.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 58 / 75
 Fourier transformation on an interval Connection to continuous Fourier transform

Connection to Fourier transform II

The Fourier transform can also be interpreted as the limit T → ∞ of the
Fourier series. For this purpose, we introduce
def 2π
∆ω = ωn+1 − ωn = (205)
T
and write in the limit T → ∞ or ∆ω → 0
1 X −iωn (t−t 0 ) 1 X 0
e = ∆ω e−iωn (t−t ) (206)
T 2π ω
n∈Z
Z n
dω −iω(t−t 0 )
→ e . (207)
2π
Similarly, we get in the limit T → ∞ (cf. Eq. (177) and Eq. (227))
Z T/2
T δnm dt i(ωn −ωm )t
δnm ≡ = e → δ(ωn − ωm ) , (208)
2π ∆ω −T/2 2π

which is the Fourier representation of the Dirac delta.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 59 / 75
 Discrete Fourier transformation

Section 3

Discrete Fourier transformation

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 60 / 75
 Discrete Fourier transformation Definition and properties

Definition of discrete Fourier transformation

Consider finite, N -dimensional vectors,

f˜0
   
f0
f =  ...  , f˜ =  ...  , (209)
   

fN −1 f˜N −1

related by the linear transformation

N −1
↔ X
f = U f˜ , fi = Uij f˜j , (210)
j =0

whose matrix is given by

 
1 2πi
Uij = √ exp ij = Uji . (211)
N N

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 61 / 75
 Discrete Fourier transformation Definition and properties

Unitarity of the transformation matrix I

The discrete Fourier transformation is the linear mapping f 7→ f˜. In order
to show its unitarity, we define
 
def 2πi
x = exp (i − j ) , (212)
N
such that x N = 1. With this, we find
N −1  −1
 NX
X 2πi
exp `(i − j ) = x` (213)
N
`=0 `=0
1 − xN
= , (214)
1−x
where we have used the geometric summation formula. Hence, the
summation yields zero except in the case x = 1 or i = j , where the
geometric summation is not applicable. In that case, it gives N . Thus, the
overall result equals N δij .
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 62 / 75
 Discrete Fourier transformation Definition and properties

Unitarity of the transformation matrix II

Unitarity means that the adjoint matrix equals the inverse. This can now
be verified directly:
N
X −1
UU † )ij = Ui` Uj∗` (215)
`=0
N −1  
1 X 2πi
= exp `(i − j ) (216)
N N
`=0
= δij . (217)
In particular, this implies that the discrete Fourier transformation is
invertible and preserves the complex scalar product,
f † · g = (f˜)† · g̃ . (218)
This is the analogue of Parseval’s theorem for continuous Fourier
transformations (see Eq. (36)).
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 63 / 75
 Discrete Fourier transformation Connection to Fourier series

Connection to Fourier series and Fourier transform

The discrete Fourier series can be interpreted as the discretization of the
Fourier series. To see this, we introduce discrete times tn = n∆t with
n = 0, . . . , N − 1 (the so-called direct lattice) and subdivide the interval
[0, T ] into subintervals [tn , tn+1 ], such that T = N ∆t. Similarly, we
introduce a discrete set of frequencies ωn = n∆ω with
∆ω = 2π/N ∆t ≡ 2π/T (the so-called dual lattice). We then get
2π T 2π
ωi tj ≡
i j = ij . (219)
T N N
In particular, there are as many discrete frequencies as there are discrete
times. A function in the time domain is now given by a finite vector,
f = ( f0 , . . . , fN −1 )T , fi = f (ti ) . (220)
The same applies to the Fourier transform,
f˜ = ( f˜0 , . . . , f˜N −1 )T , f˜i = f˜(ωi ) . (221)
The left out value fN ≡ f (tN ) does not matter in the limit N → ∞.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 64 / 75
 Discrete Fourier transformation Connection to Fourier series

The discrete Fourier representation of the Kronecker delta,

Eqs. (216)–(217), can now be written as
N −1
1 X

δij = ∆t exp it` (ωi − ωj ) . (222)
T
`=0

In the limit N → ∞, ∆t → 0 with N ∆t = T = const., we get

1 T
Z
δij = dt ei(ωi −ωj )t . (223)
T 0
This coincides with Eq. (177). Conversely, in the time domain we can write
N −1
δij 1 X

= ∆ω exp iω` (ti − tj ) , (224)
∆t 2π
`=0

and this converges to the Dirac delta for ti , tj ∈ [0, T ]. Hence, we recover
the Fourier transform on a finite interval from the discrete Fourier
transformation in an appropriate limit.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 65 / 75
 Discrete Fourier transformation Connection to continuous Fourier transformation

Connection to continuous Fourier transform I

The transition to the continuous Fourier representation of the Dirac delta
distribution can also be performed directly by rewriting the discrete Fourier
representation (216)–(217) in the form
1 X N ∆t δij
∆t eit` (ωi −ωj ) = δij ≡ → δ(ωi − ωj ) , (225)
2π 2π ∆ω
1 X δij
∆ω eit` (ωj −ωi ) = → δ(ti − tj ) . (226)
2π ∆t
This shows that the Fourier transformation and its inverse can be obtained
from their discrete counterparts in appropriate limiting cases. Conversely,
the discretization of the Fourier transform leads back to the notion of the
discrete Fourier transform. We note, however, that in all these cases one
has to pay attention to the respective conventions in the defining
equations, which make themselves felt in the presence of irksome
prefactors (as in the relation between the Fourier transform on an interval
and the Fourier transform on the real axis, Eq. (202)).
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 66 / 75
 Discrete Fourier transformation Connection to continuous Fourier transformation

Connection to continuous Fourier transform II

The continuous Fourier transform corresponding to the Fourier

representation of the Dirac delta distribution, i.e. Eq. (17), can also be
obtained by rewriting Eq. (223) in the form (see also Eq. (208))
Z T/2
1 T δij
dt ei(ωi −ωj )t = δij ≡ (227)
2π −T/2 2π ∆ω

and performing the limit T → ∞. Here, by dint of the periodicity we have

transferred the integration interval [0, T ] to the interval [−T /2, T /2],
which is symmetric around the origin. As the resulting formula is actually
central to the Fourier analysis on an interval but has been obtained in the
form of Eq. (223) as an appropriate limit of the discrete Fourier transform,
it also highlights again the interconnectedness between the Fourier analysis
of periodic functions and the continuous Fourier transformation.

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 67 / 75
 Discrete Fourier transformation Connection to continuous Fourier transformation

Further pedagogical works related to Fourier transforms

For a rather exhaustive survey of the relations between different notions of

the Fourier transform, the interested reader is referred to Table 1 on p. 92
of our pedagogical article on electron-phonon coupling:

https://www.researchgate.net/publication/303747580

Sensible conventions for the generalization of the Fourier transformation to

spatial variables and to integral kernels are given in Sct. 1.1 of our
paradigmatic article on electrodynamics in media:

https://www.researchgate.net/publication/275208133

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 68 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem I
Assume we want to calculate numerically the Fourier transform of a
periodic function by sampling the time signal at discrete points
tj = j ∆t , j = 0, . . . , N − 1; T = N ∆t, i.e. the interval [0, T ] is sampled
with the frequency fsample = 1/∆t. We then get
Z T
1
f˜i ≡ √ dt eiωi t f (t) (228)
T 0
N −1
1 X
≈√ ∆t eiωi tj f (tj ) (229)
T j =0
 1/2 N −1
T 1 X iωi tj
= √ e f (tj ) . (230)
N N j =0

Up to the prefactor of (T /N )1/2 , the Fourier expansion coefficient f˜i is

therefore approximated by the discrete Fourier transform of the vector
constructed from the discretized time signal.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 69 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem II
However, the exact expression for the periodic time signal is given by (see
Eq. (171))
1 X ˜ −iωn tj
f (tj ) = √ fn e . (231)
T n
For the discrete Fourier transform of this time signal, we then get
N −1 N −1
!
1 X iωi tj 1 X iωi tj 1 X ˜ −iωn tj
√ e f (tj ) = √ e √ fn e (232)
N j =0 N j =0 T n
 
N −1
1 X X
=√ f˜n  ei(ωi −ωn )tj  . (233)
NT n j =0

On the other hand, if n = k mod N , i.e. n = k + `N for some integer

` ∈ Z and k = 0, . . . , N − 1, then
ei(ωi −ωn )tj ≡ ei(ωi −ωk +`N )tj = ei(ωi −ωk )tj . (234)
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 70 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem III

The reason for this is (cf. Eq. (219))
2π T 2π
ωk +`N tj = (k + `N )j = kj + 2πkj . (235)
T N N
and thus
2π
eiωk +`N tj = ei(2πkj /N +2πkj ) = ei N kj = eiωk tj . (236)
In other words, if the time signal is only evaluated at the discrete times ti ,
the frequencies
2π 2π
ωk =
k , and ωk +`N = (k + `N ) (237)
T T
will be indistinguishable in the corresponding Fourier analysis. This is the
so-called “aliasing”. It can also be put as follows: integer multiples of
2π T =N ∆t 2π
ωN ≡ N = = 2πfsample (238)
T ∆t
cannot be detected.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 71 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem IV
The reason for this is that a signal with frequency corresponding to an
integer mutliple of the sampling frequency,
2π
ω`N = `ωN = `N , (239)
T
will appear constant if only probed at the sampling points. In fact, due to
fsample = 1/∆t we get
eiω`N tj = e2πi`fsampl j ∆t = e2πi`j = 1 , ∀j ∈ Z . (240)
For the discrete approximation of the Fourier transform, Eq. (228) or
(233), this implies for our discrete approximation
X
fi ≈ fk +`N δik . (241)
`

This will only be true, if fk +`N ≈ 0 for ` 6= 0, i.e. the contributions from
the higher frequencies have to vanish.
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 72 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem V

Consequently, we can only sample the periodic function reliably if the exact
time signal roughly involves only frequencies corresponding to one dual
lattice. E.g., for ` = 0 we get ωn = nω1 ; n = 0, . . . , N − 1; ω1 = 2π/T ,
which for large N samples the frequency interval [0, 2π/∆t] = [0, 2πN /T ].
We now ask the central question: which is the maximum frequency
detectable provided we know that the exact signal contains both positive
and negative frequencies (which it does for real valued functions)? To
answer this, we choose an alternative representation of the dual lattice
which is symmetric around the origin and hence corresponds to the
frequency interval [−π/∆t, π/∆t]. Strictly speaking, this is only possible
if N is odd, i.e. ωn = nω1 ; n = −(N − 1)/2, . . . , (N − 1)/2. The switch
in the frequency interval is possible, because the interval [−π/∆t, 0] is
equivalent to the interval [π/∆t, 2π/∆t] for discrete Fourier transforms:
the involved frequencies differ by 2π/∆t = 2πfsample .

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 73 / 75
 Discrete Fourier transformation Shannon-Nyquist sampling theorem

Sampling theorem VI
The symmetric dual lattice samples the frequency interval
[−π/∆t, π/∆t] = [−πfsample , πfsample ] and hence stretches out between

ωmin ≡ 2πfmin = −πfsample , ωmax ≡ 2πfmax = πfsample . (242)

We now read off that the modulus of the maximum frequency is just

|ωmin | = ωmax ≡ 2πfmax = π/∆t (243)

or
1
|fmin | = fmax =
fsample , (244)
2
with the sampling frequency fsample = 1/∆t. This is the celebrated
Shannon-Nyquist theorem. We stress again that the validity of this
equation is based on the assumption that the time signal may involve
positive and negative frequencies. Otherwise one could detect with
fmax = fsample .
Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 74 / 75
 Commemoratio

Commemoramus illos qui haec omnia invenierunt

Pierre-Simon de Laplace, French mathematician (1749–1827)

Marc-Antoine Parseval, French mathematician (1755–1836)
Johann Carl Friedrich Gauß, German mathematician (1777–1855)
Jean-Baptiste Joseph Fourier, French physicist (1768–1831)
Augustin Cauchy, French mathematician (1789–1857)
Julian Karol Sokhotski, Russo-Polish mathematician (1842–1927)
Josip Plemelj, Slovene mathematician (1873–1967)
Michel Plancherel, Swiss mathematician (1885–1967)
Harry Nyquist, Swedish electrical engineer (1889–1976)
Hendry Anthony Kramers, Dutch physicist (1894–1952)
Gregory Breit, Judeo-Russian physicist (1899–1981)
Paul Adrien Maurice Dirac, British theoretician (1902–1984)
Eugene Paul Wigner, Hungarian mathematician (1902–1995)
Ralph de Laer Kronig, German physicist (1904–1995)
Claude Elwood Shannon, American electrical engineer (1916–2001)

Ronald Starke (Bergakademie Freiberg) Fourier transformation — A tutorial April 23, 2021 75 / 75