Full course overview

1. Fourier analysis and analog filtering

1.1 Fourier Transform
1.2 Convolution and filtering
MAP 555 : Signal Processing 1.3 Applications of analog signal processing
2. Digital signal processing
Part 1 : Fourier analysis and analog filtering
2.1 Sampling and properties of discrete signals
2.2 z Transform and transfer function
2.3 Fast Fourier Transform
R. Flamary
3. Random signals
3.1 Random signals, stochastic processes
3.2 Correlation and spectral representation
December 1, 2023 3.3 Filtering and linear prediction of stationary random signals
4. Signal representation and dictionary learning
4.1 Non stationary signals and short time FT
4.2 Common signal representations (Fourier, wavelets)
4.3 Source separation and dictionary learning
4.4 Signal processing with machine learning

1/108 2/108

Course overview Signal and function in Lp (R) space

Fourier Analysis and analog filtering 4
Signals and definitions 4
Properties of signals
Lp space
Dirac distribution and convolution Lp (S) is the set of functions whose absolute value to the power of p has a finite
Linear Time-Invariant systems integral or equivalently that
Fourier transform 15 Z
Fourier series ∥x∥p = |x(t)|p dt < ∞ (1)
Fourier Transform and properties S
Properties of the Fourier Transform
Frequency response and filtering 29
Convolution and Fourier Transform ▶ L1 (R) is the set of absolute integrable functions
Filtering and frequency representation ▶ L2 (R) is the set of quadratically integrable functions (finite energy)
Representation of the FT and frequency response
First and second order systems ▶ L∞ (R) is the set of bounded functions
Applications of analog signal processing 66
Analog filtering
Modulation Signal and images
Fourier optics In this course we will mostly study
Digital signal processing 105 ▶ 1D temporal signal with x(t) ∈ R, ∀t ∈ R (or complex valued function).

Random signals 105 ▶ 2D images with x(v) ∈ R, ∀v ∈ R2

Signal representation and dictionary learning 105

3/108 4/108
Some properties of signals Classical signals (1)

Signal x(t)
Causality Heaviside function Signal Γ(t)
A signal x(t) is causal if 0.5
 1.0
0.4 
0 if t < 0
0.8
x(t) = 0, ∀t < 0 0.3
Γ(t) = 1/2 if t = 0 (2)
0.2


0.6

Example: 1 if t > 0 0.4

( 0.1
0  2 for t < 0 0.0
Also known as the step function. 0.2

x(t) = 0.0
sin(t) exp − t2 for t ≥ 0 0.1 4 2 0 2 4
t 0.2 4 2 0 2 4
t

1.1 Signal x(t)

Periodicity Rectangular function Signal PiT (t)
1.0
A signal x(t) is periodic of period T0 is 0.9
 0.25

1/T if |t| < T /2 0.20
0.8
x(t − kT0 ) = x(t), ∀t ∈ R, ∀k ∈ N ΠT (t) = 1/2T if |t| = T /2 (3)
0.7 0.15


Example:  0.6 0 else
0.10
2
 0.5 ▶ Π(t) = T1 Γ(t + T2 ) − Γ(t − T2 ) .
x(t) = exp − (t−kT20 −1) for kT0 < t < 0.4
0.05

▶ Finite energy signal (finite support). 0.00

(k + 1)T0 , ∀k ∈ N 0.3 4 2 0
t
2 4
0.05 4 2 0 2 4
t

5/108 6/108

Classical signals (2) Classical signals (3)

Complex exponantial with z = τ + wi

Complex exponential τ \w −2 0 2
Signal exp((−0.5 + −2j) ∗t) Signal exp((−0.5 +0j) ∗t) Signal exp((−0.5 +2j) ∗t)
let ez (t) be the following function R → C 5
12
5
10

ez (t) = exp(zt) (4) 0

6
0

−0.5 5 4 5

where z is a complex number. When z = τ + wi the, 10

2
10
0
4 2 0 2 4 4 2 0 2 4 4 2 0 2 4
t t t

ez (t) = (cos(w ∗ t) + i ∗ sin(w ∗ t)) exp(τ t) Signal exp((0.0 + −2j) ∗t) Signal exp((0.0 +0j) ∗t) Signal exp((0.0 +2j) ∗t)
1.0 1.0
1.0

Special cases: 0.5 0.8 0.5

0.6

▶ z = τ real, then we recover the classical exponential.

0.0 0.0
0 0.4

0.5 0.2 0.5

0.0
ez (t) = exp(τ t) 1.0
4 2 0
t
2 4 4 2 0
t
2 4
1.0
4 2 0
t
2 4

Signal exp((0.5 + −2j) ∗t) Signal exp((0.5 +0j) ∗t) Signal exp((0.5 +2j) ∗t)
▶ z = wi imaginary then 5
12
5
10

0 8
0
ez (t) = cos(w ∗ t) + i ∗ sin(w ∗ t) 6
.5 5 4 5

2
10 10
0
4 2 0 2 4 4 2 0 2 4 4 2 0 2 4
t t t

7/108 8/108
Dirac delta Dirac delta (2)

Main properties of Dirac delta Signal δ(t)

Dirac delta definition
▶ Model point mass at 0. R∞
1.0
▶ Let ϕ a function supported in [−1, 1] of unit mass: −∞
ϕ(u)du = 1
▶ Value outside 0 : δ(t) = 0, ∀t ̸= 0 0.8

▶ ϕT (t) = 1
0.6 T
ϕ( Tt ) has support on [−T, T ] and unit mass.
▶ δ is a tempered distribution. 0.4 ▶ We can define the dirac delta δ as
▶ Very useful tool in signal processing 0.2

δ(t) = lim ϕT (t)

▶ Can be seen as the derivative of the 0.0
T →0
Heavyside function 1t≥0 (t) 0.2
4 2 0
t
2 4

▶ Integral Dirac delta in practice

Z +∞ Z +∞
δ(t)dt = 1, x(t)δ(t)dt = x(0) (5) ▶ Theoretical object in signal processing (impulse).
−∞ −∞
▶ Used to model signal sampling for digital signal processing.
▶ Dirac and function evaluation for signal x(t) and t0 ∈ R :
▶ Used to model point source in Astronomy/image processing, point charge in
δ(t − t0 )x(t) = δ(t − t0 )x(t0 ) Physics.
Z +∞ ▶ Has a bounded discrete variant.
⟨x(t), δ(t − t0 )⟩ = x(t)δ(t − t0 )dt = x(t0 ) (6)
−∞

9/108 10/108

Convolution operator Linear Time-Invariant (LTI) systems

Definition x(t) y(t)
Let two signals x(t) and h(t). The convolution between the two signals is defined as System
Z +∞
x(t) ⋆ h(t) = x(τ )h(t − τ )dτ (7)
−∞ Definition
▶ Convolution is a bilinear mapping between x and h. ▶ A system describes a relation between an input x(t) and an output y(t).
▶ It models the relation between the input and the output of a Linear Time ▶ Properties of LTI systems:
Invariant system. ▶ Linearity x1 (t) + ax2 (t) → y1 (t) + ay2 (t)
▶ If f ∈ L1 (R) and h ∈ Lp (R), p ≥ 1 then ▶ Time invariance x(t − τ ) → y(t − τ )
∥f ⋆ h∥p ≤ ∥f ∥1 ∥h∥p ▶ A LTI system can most of the time be expressed as a convolution of the form:

▶ The dirac delta δ is the neutral element for the convolution operator: y(t) = x(t) ⋆ h(t)
Z +∞ where h(t) is called the impulse response (the response of the system to an input
x(t) ⋆ δ(t) = x(τ )δ(t − τ )dτ = x(t) (8) x(t) = δ(t))
−∞

▶ It can also be used to model a temporal delay: Examples

▶ Passive electronic systems (resistor/capacitor/inductor) .
x(t) ⋆ δ(t − t0 ) = x(t − t0 ) (9) ▶ Newtonian mechanics, Fluid mechanics, Fourier Optics.
11/108 12/108
LTI systems and Ordinary Differential Equation Signal and frequencies

Ordinary Differential Equation (ODE)

The system is defined by a linear equation of the form:

dy(t) dn y(t) dx(t) dm x(t)

a0 y(t) + a1 + · · · + an n
= b0 x(t) + b1 + · · · + bm (10)
dt dt dt dtm

▶ ODE based system with linear relations are an important class of LTI systems. ▶ A signal is x(t) a function of time, an image x(v) a function of space.

▶ Also called homogeneous linear differential equation. ▶ Those functions are what we measure/observe but can be hard to
interpret/process automatically.
▶ n is the number of derivatives for y(t) and m for x(t).
▶ Another representation for a signal is in the frequency domain (1/t).
▶ max(m, n) is the order of the system.
▶ Better representation for numerous applications.
▶ The output of the system can be computed from the input by solving Eq. (10).
▶ Linearity and time invariance are obvious from the equation. Applications
▶ Signal processing (biomedical, electrical).
▶ Image processing (2D signals), filtering, reconstruction.
▶ Colors are combination of waves of different frequencies.
13/108 14/108

Fourier Series (1) Fourier series (2)

Decomposition as trigonometric series

2π
One can express periodic x(t) of period T0 = w0
integrable on the period as
∞
a0 X
x(t) = + [ak cos(kw0 t) + bk sin(kw0 t)]
2
k=1

where ak and bk are the Fourier coefficients that can be computed as

Z Z
2 2
ak = x(t) cos(kw0 t)dt bk = x(t) sin(kw0 t)dt
T0 T0 T0 T0

▶ Representation of a periodic signal as an infinite number of coefficients

corresponding to harmonic frequencies.
▶ Can be interpreted as a change of basis from temporal to frequencies.
History ▶ Functions can be approximated with a finite number N of terms.
▶ Trigonometric series used by Euler, d’Alembert, Bernoulli and Gauss. ▶ Gibbs phenomenon appears for discontinuous functions
▶ Introduced by Joseph Fourier in [Fourier, 1807]. [Hewitt and Hewitt, 1979].
▶ Fourier claimed that these series could approximate any function.
15/108 16/108
Example of Fourier series Complex Fourier series

Complex harmonic decomposition

2π
One can express periodic x(t) of period T0 = w0
integrable on the period as
∞
X 2π
x(t) = ck ejkw0 t avec w0 =
T0
k=−∞

where the coefficients ck are called the complex Fourier coefficients and can be
computed with
Z Z T0 Z T0 /2
1 1 1
1.0 ck = x(t)e−ikw0 t dt = x(t)e−ikw0 t dt = x(t)e−ikw0 t dt
0.8 T0 T0 T0 0 T0 −T0 /2
0.6
0.4
0.2 Relations between decompositions
0.0
0.2 Using the Euler formula we can show that ak and bk and the ck coefficients are related
0.4
0.6 by
5 a0
4 = c0 ak = ck + c−k bk = i(ck − c−k )
0 3 2
1 2 2 Time t
Freque3ncy 4 5
1 Note that if x(t) is an even function then the bk = 0 , and if x(t) is odd then ak = 0.
6 0
Example : Square wave
▶ Square wave with T0 = 2 17/108 18/108
P∞
1
▶ x(t) = i=−∞ 1[iT0 ,iT0 +T0 /2] (t)
0.8
▶ a0 = 1, ak = 0 ∀k > 0
Fourier transform
0.6
0.4
▶ bk = 2
Interpretation of the Fourier transform
0.2
πk
for k odd else bk = 0
0
0 1 2 3 4 5
Z ∞

Definition (Fourier Transform) x(t) = ei2πf t X(f )df

−∞
The Fourier Transform (FT) of a signal x(t) can be expressed as
Z ∞ Harmonic representation
F[x(t)] = X(f ) = e−i2πf t x(t)dt (11) ▶ The FT represents the signal in the frequency domain.
−∞

When it exists the inverse Fourier transform is defined as ▶ |X(f )| is the magnitude of a sinusoidal signal for frequency f .
Z ∞ ▶ Arg(X(f ) is the phase of the sinusoidal signal.
F −1 [X(f )] = x(t) = ei2πf t X(f )df (12)
−∞
▶ For a real signal x(t), X(f ) = X(−f )∗ and an informal interpretation would be
Z +∞ Z +∞
▶ Note that the ˆ operator is also often used for the Fourier transform x̂ of x. x(t) = X(f )ei2πf t df = |X(f )|ei2π(f t+Arg(X(f ))) df (13)
▶ In signal processing and electrical engineering the references often use j instead −∞
Z
−∞
+∞
of i for the imaginary number (i is a measure of current). ≈ X(0) + 2 |X(f )| cos(2π(f t + Arg(X(f )))) (14)
▶ The FT is a change of representation for the function x from the temporal 0+

representation to the harmonic (frequency) representation. ▶ The modulus and argument of the FT allow identification of the frequency
content of the signal and its phase.

19/108 20/108
Examples of Fourier Transform (1) Examples of Fourier Transform (2)
Rectangular function
Decreasing exponential
Signal x(t)
 Signal ΠT (t)  1.0

1/T if |t| < T /2 
1 for t > 0
0.5 0.8
ΠT (t) = 1/2T if |t| = T /2 (15) x(t) = e−at Γ(t), Γ(t) = 1/2 for t = 0

 0.4 
 0.6
0 else 0.3
0 else
0.4
0.2
The Fourier transform is with a > 0 0.2
0.1
Z The Fourier transform is
1 T /2 −i2πf t 0.0
Z ∞
0.0
F[ΠT (t)] = e dt 6 4 2 0
t
2 4 6
T −T /2 0.12.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 F[e−at Γ(t)] = e−at e−i2πf t dt
t
 −i2πf t T /2 0 0.6 X(f)
−e F[ΠT (t)]
Z ∞ Partie reelle
= 3.5 Partie imaginaire
i2πf T −T /2 3.0
Partie reelle
Partie imaginaire = e−(a+i2πf )t dt 0.4
2.5
0
eiπf T − e−iπf T  ∞
= 2.0 e−(a+i2πf )t 0.2

i2πf T 1.5 =
−(a + i2πf ) 0
1.0 0.0
sin(πf T ) 1
= = sinc(πf T ) 0.5
=
πf T 0.0 a + i2πf 0.2
0.5
4 2 0 2 4
sin(t) f
with sinc(t) = t
and sinc(0) = 1 1.0 4 2 0 2 4
f

21/108 22/108

Properties of the Fourier Transform Properties of the Fourier Transform (2)

Linearity Frequency shift

Let x1 (t) and x2 (t) two signals of TF X1 (f ) and X2 (f ) respectively. Let x(t) be a signal of FT X(f ) then we have
For a ∈ R and b ∈ R, we have : h i
F ei2πf0 t x(t) = X(f − f0 )
F[ax1 (t) + bx2 (t)] = aX1 (f ) + bX2 (f )
Multiplication by a complex exponential of frequency f0 , translates the TF by f0 .
Proof. Comes from the linearity of the integration. Proof. Regroup exponentials in the integral.
Time shift Time scaling
Let x(t) be a signal of FT X(f ).
Let x(t) be a signal of FT X(f ) and a a scaling a ̸= 0 then we have
For t0 ∈ R, let x(t − t0 ) a time shift of x(t) then we have:
 
1 f
F[x(t − t0 )] = e−i2πt0 f X(f ) F[x(at)] = X
|a| a
Proof. Change of variable in the integral.
Proof. Change of variable for separate cases a > 0 and a < 0.

23/108 24/108
Properties of the Fourier Transform (3) Properties of the Fourier Transform (4)
Derivation
Let x(t) be a signal of FT X(f ) then we have
  Even and odd signals
dx(t)
F = i2πf X(f )
dt x(t) X(f )
Even real Even real
Odd real Odd imaginary
Integration Even imaginary Even imaginary
R∞
Let x(t) be a signal of FT X(f ) such that −∞ x(t)dt = 0 then we have Odd imaginary Odd real
Z 
t
1 For a real signal x(t) : X(f ) = X(−f )∗
F x(u)du = X(f )
−∞ i2πf
Conjugate signal
R∞
If −∞
(x(t) − c)dt = 0 where c is often called the constant term, we have Let x(t) be a signal of FT X(f ) and x∗ (t) its complex conjugate, then we have
Z 
t
1 F[x∗ (t)] = X ∗ (−f )
F x(u)du = X(f ) + cδ(f )
−∞ i2πf

where δ(f ) is the Dirac delta.

Those two properties can be used to solve Ordinary Differential Equations (ODE).
25/108 26/108

Duality of the Fourier Transform Fourier Transform in Lp (R)

▶ For 1 ≤ p ≤ 2 the FT maps from Lp (R) to Lq (R) with 1
p
+ 1
q
= 1.
Let x(t) be a signal of FT X(f ). When the inverse Fourier transform exists we can
write ▶ Consequence of the Riesz–Thorin theorem.
Z +∞ Z +∞ ▶ The TF of an absolute integrable function is bounded (Example : rectangle).
x(−t) = X(f )ej2πf (−t) df = X(f )e−j2πf t df = F[X(f )]
−∞ −∞
Parseval-Plancherel identity in L2
▶ The last term is the TF of function X(f ). The TF of an L2 function is L2 . Note that L2 is a Hilbert space of inner product:
Z ∞
▶ This means that if F[x(t)] = X(f ) then
< x, y >= x(t)y ∗ (t)dt
−∞
F[X(t)] = x(−f )
For two functions x, y ∈ L2 (R)2 of respective TF X, Y ∈ L2 (R)2 the
▶ Applying twice the TF operator to x(t) returns x(−t): F[F[x(t)]] = x(−t) Parseval-Plancherel identity states that
Z ∞ Z ∞
Example < x, y >= x(t)y ∗ (t)dt = X(f )Y ∗ (f )df (16)
For the rectangular function ΠT (t) : −∞ −∞
Z ∞ Z ∞
ΠT (t) → sinc(πf T ) < x, x >= |x(t)|2 dt = |X(f )|2 df (17)
sinc(πT t) → ΠT (−f ) = ΠT (f ) −∞ −∞

which means that the energy of a signal is preserved by FT.

More details in [Hunter, 2019, Chap. 5.A] and [Mallat et al., 2015, Chap. 1]
27/108 28/108
Convolution and Fourier Transform Dirac delta and Fourier Transform
Convolution and Fourier Transform
Let two signals x(t) and h(t) of respective Fourier transform X(f ) and H(f ) then Fourier transform and Dirac delta
F[x(t) ⋆ h(t)] = X(f )H(f ) (18) ▶ Fourier Transform of δ(t) and δ(t − t0 ):
Z +∞
▶ The TF of a convolution is a pointwise multiplication in frequency.
F[δ(t)] = δ(t)e−i2πf t dt = e0 = 1
▶ The complex exponential function is the eigenvector for the convolution operator. −∞

▶ Easy interpretation of the effect of a linear filtering. F[δ(t − t0 )] = e−i2πf t0

▶ By duality of FT we have:
Proof F[1] = δ(f )
Z ∞ Z ∞ i2πf0 t
F[e ] = δ(f − f0 )
F[x(t) ⋆ h(t)] = e−2iπf t x(u)h(t − u)dudt
Z
−∞
Z
−∞ ▶ Convolution
∞ ∞
= e−2iπf (u+v)
x(u)h(v)dudv F[x(t) ⋆ δ(t)] = 1X(f ) = X(f )
−∞ −∞
Z ∞
Z ∞  Z ∞  F[x(t)δ(t)] = X(f ) ∗ 1 = X(f )df = x(0)
−2iπf u
= e x(u)du e−2iπf v h(v)dv = X(f )H(f ) −∞
−∞ −∞

with the change of variable v = t − u.

29/108 30/108

The dirac comb Fourier transform of periodic signals

Cosine
▶ The dirac comb is expressed as
∞ x(t) = cos(2πf0 t) with f0 > 0
X
XT (t) = δ(t − kT ) (19)
k=−∞
▶ Bounded signal with unbounded energy.
where X is the Cyrilic Sha symbol. ▶ Intuitively this signal contains only one frequency (f0 )
▶ The Fourier Transform of the dirac comb is ▶ Its TF can be computed using to the dirac distribution.
∞
X ∞
X  
1 k 1
F[XT (t)] = e2iπkT f = δ f− = X 1 (f ) (20) FT of trigonometric functions
T T T T
k=−∞ k=−∞
 
where the second equality comes from the Poisson summation formula. ej2πf0 t + e−j2πf0 t 1 1
F cos(2πf0 t) = = δ(f − f0 ) + δ(f + f0 )
▶ The dirac comb is used to perform a regular temporal sampling. 2 2 2
 j2πf0 t −j2πf0 t

▶ Multiplying a signal by the dirac comb corresponds to a convolution by a dirac e −e 1 1
F sin(2πf0 t) = = δ(f − f0 ) − δ(f + f0 )
comb in the Frequency domain (and vice versa). 2i 2i 2i
The FT of sine and cosine is equal to 0 everywhere except on the frequency f0 of the
functions.

31/108 32/108
Fourier transform of periodic signals (2) Fourier Transform in Rd

The Fourier Transform can be naturally extended to functions in Rd .

Fourier transform of periodic signal Fourier Tansform in Rd
Let x(t) be a periodic signal of period T0 , it can be expressed as the following complex Let x(v) : Rd → C, the Fourier Transform of x can be expressed as
Fourier series: X Z
i2π k t
x(t) = ck e T0 F[x(v)] = X(u) = x(v)e−2iπ<v,u> dv (21)
k Rd

Its Fourier transform can be expressed as When it exists the Inverse FT is defined as
  Z
X k
X(f ) = F[x(t)] = ck δ f − F −1 [X(u)] = x(v) = X(u)e2iπ<v,u> du (22)
T0 Rd
k

▶ u ∈ Rd is a directional frequency.
▶ The FT of a periodic signal of period is null except on frequencies k
, k ∈ N.
T0 ▶ All the properties of the 1D FT are preserved (duality, convolution, ...)
▶ 1
is the fundamental frequency, k
with |k| ≥ 2 are called the harmonics.
T0 T0 ▶ With d = 2, frequency representation of black and white images.
▶ The TF of a periodic function is a weighted sum of diracs.
▶ With large d, approximation for efficient kernel approximation in machine learning
[Rahimi and Recht, 2008].

33/108 34/108

Examples of Fourier Tranform in 2D Examples of Fourier Tranform in 2D (Modulus and

Phase)

35/108 36/108
Fourier transform and angular frequency How to compute a Fourier Transform ?

▶ The FT in this course is a function of frequency f (in Hz).

▶ Another common way to represent frequency is the angular frequency w (in
rad/s) such that Usual steps
w
w = 2πf, f=
2π 1. Use known FT pairs if possible.
▶ When using angular frequency the FT is non-unitary meaning that : 2. Express the function as a composition of operations with known properties:
Z ∞ ▶ Linearity, time shift
F̃[x(t)] = X̃(w) = e−iwt x(t)dt ▶ Convolution
−∞
Z ▶ Duality
∞
1
F̃ −1 [X̃(f )] = x(t) = eiwt X̃(w)dw 3. Use the properties of FT on the composition.
2π −∞
4. Check properties (FT of even/odd function) to detect easy mistakes.
▶ There exists a unitary angular frequency FT that scales both FT and IFT by √1 .
2π As a rule : try to avoid computing the integral but sometime you have to do it.
▶ In the following we will sometime use the FT as a function of the angular
frequency: w
X̃(w) = X
2π

37/108 38/108

Frequency response of LTI systems Frequency response and static gain

x(t) y(t) Response to a mono-frequency signal
h(t)
▶ For a system of impulse response h(t) with an input x(t) = e2jπf0 t
Z +∞
Impulse response and frequency response y(t) = h(τ )e2jπf0 h(t−τ ) dτ
−∞
Z
▶ Most LTI systems can be expressed as a convolution of the form: +∞
= e2jπf0 t h(τ )e−2jπf0 hτ dτ
−∞
y(t) = x(t) ⋆ h(t)
= e2jπf0 t H(f0 ) = x(t)H(f0 )
where h(t) is called the impulse response (the response of the system to an input
x(t) = δ(t)) ▶ An input signal with unique frequency f0 is multiplied by H(f0 ).
▶ The Fourier transform of the LTI system relation between x and y is ▶ Its amplitude is multiplied by |H(f0 )| and a phase Arg(H(f0 )) is added.
▶ The complex exponential is an eigenvector of the convolution operator.
Y (f ) = H(f )X(f ) (23)

▶ The frequency response H(f ) (also called transfer function) of the LTI system is Static gain
the Fourier transform of h(t): The complex static gain is the constant K such that
Z +∞
Y (f )
H(f ) = (24) K = H(0) = h(t)dt
X(f ) −∞

39/108 40/108
LTI systems and Ordinary Differential Equation Representation of the frequency response

Frequency interpretation of the frequency response

Ordinary Differential Equation (ODE)
The system is defined by an equation of the form: ▶ The frequency response of a system gives information on the transformations due
to the system.
dy(t) dn y(t) dx(t) dm x(t)
a0 y(t) + a1 + · · · + an n
= b0 x(t) + b1 + · · · + bm (25) ▶ Quantities that can be plotted :
dt dt dt dtm
H̃(w) = Re(H̃(w)) + jIm(H̃(w))
Frequency response of an ODE = |H̃(w)|ejArg(H̃(w))
▶ We recall the properties of the FT for th n-th derivative of a function:
▶ |H̃(w)| modulus of the frequency response.
 (n)   
d x(t) ▶ Arg(H̃(w)) = ∠H̃(w) = tan−1 Im(H̃(w) phase in radian.
F = (2iπf )n X(f ) = (iw)n X(w) Re(H̃(w)
dtn

▶ The Frequency response of the ODE can be expressed as Graphical representation of systems
▶ Bode plot (Modulus+Argument).
Y (w) b0 + b1 jw + · · · + bm (jw)m
H(w) = = (26) ▶ Nichols/Black plot (Modulus VS Argument).
X(w) a0 + a1 jw + · · · + an (jw)n
▶ Nyquist plot (Real VS Imaginary)

41/108 42/108

Bode plot Properties of the Bode plot

The logarithm and the argument allows for simple diagrams for combination of systems
Multiplication
If two LTIs H̃1 (w) and H̃2 (w) are in series the the equivalent system is
H̃(w) = H̃1 (w)H̃2 (w)
▶ G̃(w) = G̃1 (w) + G̃2 (w)
▶ Φ̃(w) = Φ̃1 (w) + Φ̃2 (w)
Definition
The Bode plot of a system is composed of two plots that are function of w: Division
▶ Magnitude (or gain) in decibels (dB) H̃1 (w)
If and LTI can be expressed as H̃(w) = then
H̃2 (w)
G̃(w) = 20 log10 (|H̃(w)|) ▶ G̃(w) = G̃1 (w) − G̃2 (w)
▶ Phase in degrees or radians ▶ Φ̃(w) = Φ̃1 (w) − Φ̃2 (w)
This is particularly useful for rational frequency responses such as ODE.
Φ̃(w) = Arg(H̃(w)|) = ∠|H̃(w)|

The scale of the radial frequency w is logarithmic, which means that for a rational
frequency response H one will be mostly piecewise linear.
43/108 44/108
Diagramme de Black Nyquist plot

Définition Definition
The Nichols plot (Diagramme de Black in France) is a parametric plot of H̃(w) with The Nyquist plot is a parametric plot of H̃(w) with Real(H̃(w)) on x-axis and
20 log10 |H̃(w)| on y-axis and phase Φ̃(w) on x-axis. Imag(Φ̃(w)) on y-axis.
▶ Show the Modulus/Phase trajectory as a function of w. ▶ Show the trajectory of H̃ in the complex plane.
▶ Can be plotted following the Bode plot w. ▶ Used in system control to study the stability of systems.

45/108 46/108

Frequency response of electronic systems First order system (1)

Principle ▶ System
Ohm’s law can be extended to capacitors and inductors using what is called complex x(t) = Ri(t) + y(t)
electrical impedance. Z
1 t
The linear system i(t) → u(t) is expressed as y(t) = i(v)dv
C −∞
˜
Ũ (w) = H̃(w)I(w) ˜
= Z̃(w)I(w) x(t) = RCy ′ (t) + y(t)

For electronic systems j is used instead if i as the imaginary number. ▶ Frequency response
Y (f ) 1
Resistor Capacitor Inductor H̃(f ) = =
Rt X(f ) 1 + RC2jπf
▶ u(t) = Ri(t) ▶ u(t) = 1
i(u)du ▶ u(t) = L di(t)
C −∞ dt ▶ Using complex impedance
▶ Ũ (w) = RI(w)
˜
▶ Ũ (w) = 1 ˜ ▶ Ũ (w) = jLwI(w)
˜
jCw
I(w) ˜ ˜
▶ ZR = R ▶ ZL = jLw Ỹ (w) = Zc I(w) et X(w) = (ZR + ZC )I(w)
▶ ZC = 1
jCw Ỹ (w) Zc 1 1
The frequency response of passive electronic systems can be computed with simple H̃(w) = = = ZR
=
X̃(w) ZC + ZR 1+ Z 1 + RCjw
C
computation of complex numbers.

47/108 48/108
First order system (2) First order system (3)
Normalized system
We reformulate the frequency response as as:
1 Bode plot
H(w) = (27)
1 + j ww0
Argument
1 1
where w0 = τ
= RC
. 1. H̃(w) = 1
w
1+j w
0
Diagramme de Bode
2. Φ̃(w) = arg(H(w)) = −arg(1 + jw) = −tan−1 (w)
Modulus
3. limw→0 Φ̃(w) = 0
1
1. H̃(w) = w
1+j w
0 4. limw→∞ Φ̃(w) = −π/2
2. |H̃(w)| = r 1 5. When w = w0 , Φ̃(w) = −tan−1 (1) = −π/4 (−45◦ )
2
1+ w2
w0
when w = 10w0 , Φ̃(w) = −84◦
w2
3. G̃(w) = 20 log10 (|H(w)|) = −10 log10 (1 + 2)
w0
when w = .1w0 Φ̃(w) = −6◦

4. limw→0 G̃(w) = 0
2
5. limw→∞ G̃(w) = −10 log10 ( w
w2
) = −20 log10 (w) + 20 log10 (w0 )
0

6. When w = w0 , G̃(w) = −10 log10 (2) = −3dB

49/108 50/108

First order system (4) First order system (5)

Bode plot Nichols plot

Bode Diagram Nichols Chart

0

−10 10 3 dB
Magnitude (dB)

6 dB −3 dB

−20 0 −6 dB

Open−Loop Gain (dB)

−30 −10 −12 dB

−40 −20 dB
0 −20

−30
Phase (deg)

−45
−40 −40 dB

−50
−90
−2 −1 0 1 2 −360 −315 −270 −225 −180 −135 −90 −45 0
10 10 10 10 10
Frequency (rad/sec)
Open−Loop Phase (deg)

51/108 52/108
First order system (6) Second order system (1)
▶ Complex Impedance

˜ L
Ỹ (w) = Zc I(w)
Nyquist plot ˜
X̃(w) = (ZL + ZR + ZC )I(w)
Nyquist Diagram

0.5
6 dB4 dB2 dB0 dB
10 dB −2 dB
−4 dB −10 dB −6 dB ▶ Frequency response
Imaginary Axis

1
20 dB −20 dB Ỹ (w) ZC jCw
H̃(w) = = = 1
0 X̃(w) ZL + ZR + ZC jCw
+ R + jLw

▶ Normalized frequency response

−0.5 1 k
−1 −0.5 0 0.5 1 H̃(w) = = jw jw 2
Real Axis 1 + RCjw + LC(jw)2 1 + 2z w + (w )
n n

▶ k Static gain : k = 1
q
▶ z damping ratio of the system : z = R C
2 L
▶ wn natural frequency of the system : wn = √1
LC
53/108 54/108

Second order system (2) Second order system (3)

Linear differential equation Response of the system for z > 1

The second order differential equation corresponding to the system is
▶ c1 and c2 are real coefficients.
d2 y(t) dy(t) ▶ The FT can be expressed as
+ 2zwn + wn2 y(t) = kwn2 x(t) (28)
dt2 dt
M M
H̃(w) = − (32)
jw − c1 jw − c2
Factorization
The second order system can be factorized as with M = √wn2 ,
2 z −1

kwn2 ▶ The impulse response of the system is

H̃(w) = (29)
(jw − c1 )(jw − c2 )
h(t) = M (ec1 t − ec2 t )Γ(t)
with
▶ The step response of the system is
c1 = (30)   c1 t 
p e ec2 t
c2 = −zwn − wn z2 − 1 (31) e(t) = 1 + M − Γ(t)
c1 c2
c1 and c2 are called the poles of the transfer function.

55/108 56/108
Second order system (4) Second order system (5)

Response of the system for z < 1

Response of the system for z = 1
The FT becomes: ▶ In this case the damping is weak and oscillations appear.
kwn2
H̃(w) = (33) ▶ This comes from the fact that when z < 1 coefficients c1 and c2 are complex.
(jw + wn )2
The impulse response is
that is the square of one first order system.
The impulse response fo the system can be expressed as h(t) = M (ec1 t − ec2 t )Γ(t)

h(t) = wn2 te−wn t Γ(t) ▶ The step response is

The step response can be expressed as wn e−zwn t  p 

h(t) = √ sin wn t 1 − z 2 Γ(t)
1−z 2
e(t) = (1 − e−wn t − wn te−wn t )Γ(t)
that is a sine with an exponentially decreasing magnitude.

57/108 58/108

Second order system (6) Second order system (7)

Impulse and step responses Bode plot
We can plot the Bode plot using the normalized frequency response:
Réponse impulsionnelle d’un systeme du second ordre
0.6
z=0.5 k
H(w) =  2   (34)
0.4 z=1 jw jw
z=1.5 wn
+ 2z wn
+1
0.2
Modulus
k
0 1. H̃(w) = 
jw
2 
jw
 .
wn
+2z w +1
n
−0.2 k
−5 0 5 10 15 20 2. |H̃(w)| = s
 2 2  2
.
1− ww +4z 2 ww
n n
Réponse indiciele d’un systeme du second ordre
1.5
z=0.5 3. G̃(w) = 20 log10 (|H̃(w)|) = !
z=1   2 2  2
1 z=1.5 w 2 w
−10 log10 1 − wn + 4z wn + 20log(k)

0.5 4. limw→0 G̃(w) = 20log(k)

4
w
5. limw→∞ G̃(w) = −10 log10 ( w 4 ) = −40 log10 (w) + 40 log10 (wn )
0 n
−5 0 5 10 15 20
6. En w = w0 , G̃(w) = −20 log10 (2z) + 20 log(k).
59/108 60/108
Second order system (8) Second order system (9)

Properties of the modulus Bode plot

p
▶ The modulus of the frequency response for z < (2)/2 has a maximum at the
Argument
following frequency p k
wmax = wn 1 − 2z 2 1. H̃(w) = 
jw
2 
jw
 .
wn
+2z w +1
n
▶ The value of the modulus at this frequency is !
 2   2z ww
jw jw −1
2. Φ̃(w) = arg(H(w)) = −arg( wn
+ 2z wn
+ 1) = −tan n
2 .
k 1− w2
|H̃(wmax )| = √ wn
2z 1 − z 2
3. limw→0 Φ̃(w) = 0
▶ The cutoff frequency at -3dB is equal to 4. limw→∞ Φ̃(w) = −π(−180◦ )
q p 5. En w = w0 , Φ̃(w) = −tan−1 (1) = −90◦ ,
w−3 = wn 1 + 2z 2 + 2 − 4z 2 + 4z 4

61/108 62/108

Second order system (10) Second order system (11)

Bode plot Nichols plot

Bode Diagram Nichols Chart

50 40
0 dB z=0.1
0.25 dB
0.5 dB z=1
0 20 1 dB −1 z=10
dB
Magnitude (dB)

3 dB −3 dB
6 dB
−50 0 −6 dB

Open−Loop Gain (dB)

−12 dB
−100 −20 −20 dB

−150 −40 −40 dB

0
z=0.1
−60 −60 dB
−45 z=1
Phase (deg)

z=10
−80 −80 dB
−90

−135 −100 −100 dB

−180 −120 dB
−120
−3 −2 −1 0 1 2 3 −180 −135 −90 −45 0
10 10 10 10 10 10 10
Frequency (rad/sec) Open−Loop Phase (deg)

63/108 64/108
Second order system (12) Applications of analog signal processing
Nyquist plot R2

Nyquist Diagram -

R +
6
0 dB z=0.1
z=1 C R1
z=10
4

2 dB −2 dB
2
Applications of analog signal processing
Imaginary Axis

4 dB −4 dB
6 dB −6 dB
10 dB−10 dB ▶ Analog signal filtering.
0
▶ Electronic passive and active filters.
▶ Modeling and filtering with physical systems.
−2
▶ Telecommunications.
▶ Amplitude modulation.
−4 ▶ Multiplexing.
▶ Fourier optics
−6
−6 −4 −2 0 2 4 6 ▶ Light propagation in perfect lens/mirror systems.
Real Axis ▶ Point spread functions of telescope and cameras.

65/108 66/108

Analog filtering Applications of analog filtering

x(t) y(t)
Filter

Definition
Signal processing system that aim at selecting part of the signal and attenuating
another part (noise).
Analog filtering as opposed to digital filtering (next course) ▶ High end audio, amplifiers, (equalizer, echo).
▶ Car suspension.
Objectives
▶ Seismic protection.
▶ Find a system that transform a signal x(t) to extract pertinent information.
▶ Band-pass before Analog-to-Discrete conversion.
▶ Attenuate noise in a signal.
▶ Fourier optics, telescope modeling.
▶ Separate several components of a signal (when different frequency bands).

67/108 68/108
Signal to Noise Ratio (SNR) Filtering and bandwidth
Gain and Attenuation
▶ In order to characterize a filter one uses its Gain/Phase (Bode plot).
Additive noise
The recording of a signal often contains additive noise: G̃DB (w) = 20 log10 (|H̃(w)|) et Φ̃(w) = Arg(H̃(w))

y(t) = x(t) + n(t) ▶ Attenuation is also often used Ã(w) = −G̃DB (w)

y(t) is the recorded signal, x(t) is the signal of interest and n(t) is the noise.
Bandwith and passband
Signal to Noise Ratio The band with of a filter is the set of frequency for which the Gain is over a reference
(usually -3dB). Bandwith at −3dB:
Px    
SN R = ou SN R(dB) = 10 log10 (RS/B ) (35) |H̃(w)|
Pn BW = w|20 log ≥ −3
▶ Px is the power of the signal and Pn is the power of the noise. max(|H̃(w)|)
A2
▶ When signals are cosine the SNR is SN R = x
where Ax and An are the
A2
n Types of filters
amplitudes. ▶ Low-pass, BW = [O, fc ] with fc cutoff frequency
▶ The objective of filtering is often to maximize the SNR. ▶ High-pass, BW = [fc , ∞]
▶ Band-pass, BW = [fc1 , fc2 ]
▶ Band-stop, BW = [0, fc1 ] ∪ [fc2 , ∞]
69/108 70/108

Filter distortion Filter distortion (2)

Undistorted transmission
A signal is considered undistorted when the output of the system is
With
y(t) = Cx(t − t0 ) ▶ C a constant gain.
▶ t0 > 0 is a delay.

A system with no distortion has the following FT and impulse response Phase distortion
Let a system of frequency response
X̃(w)
H̃(w) = = Ce−jwt0 et h(t) = Cδ(t − t0 ) H̃(w) = |H(w)|ejϕ(w)
Ỹ (w)
We can deduce that for
With
x(t) = cos(ωt)
▶ |H̃(w)| = C or else amplitude distortion.
y(t) = |H̃(ω)| cos(ωt + ϕ(ω)) = |H̃(ω)| cos(ω(t + ϕ(ω)/ω))
▶ Arg(H̃(w)) = −wt0 or else phase distortion.
Note that the argument of the frequency response varies linearly with the frequency. The delay ϕ(ω)/ω is also called propagation time of frequency delay. For it to be
independent from frequency it is necessary that

ϕ(ω)
= cte = τ → ϕ(ω) = ωτ
71/108 ω 72/108
Ideal low pass filter Filter design
Definition
▶ The ideal low-pass filter is often a theoretical object in signal processing. Real filter
▶ Perfect to use when the noise and signal have non-overlapping spectra. ▶ Ideal filters are non causal and cannot be implemented in practice .

▶ The frequency response of the ideal filter is ▶ We search for an approximation of the ideal filter.
 ▶ the approximation has to respect constraints (Gabarit in french).
1 if |f | < fc
H(f ) =
0 else
Constraints of a filter
where fc is the cutoff frequency.
Parameters:
▶ The impulse response of the filter is
▶ Bandwidth BP and rejected band
sin(2πfc t) ▶ Oscillations :
h(t) = 2fc = 2fc sinc(2πfc t)
2πfc t
▶ ε in passing bandwidth
Realizable filter ▶ δ in attenuated bandwidth
▶ A realizable temporal filter is causal and stable (absolute integrable).
▶ Ideal filter is neither of those and cannot be used for 1D (time) filtering. The constraints define the area that are acceptable for a given application.
▶ For images (2D) causality is not necessary.

73/108 74/108

Simple example of filter design Simple example of filter design

▶ Application to Brain computer interface.

▶ Interesting signal for event related potentials
below ≈ 12Hz (ws = 2π ∗ 12).
▶ SNR= G̃(ws ) − G̃(wedf )
▶ Electrical noise (EDF) at 50Hz
▶ SNR is a decreasing function of w0 .
(wedf = 2π ∗ 50).
▶ What is the best value for w0 ?
▶ Two low power signals As ≈ An .
▶ Maximum attenuation of signal at -3dB.
▶ Filtering with first order filter. Choix que w0

▶ Frequency response ▶ Before filtering:  ▶ With the maximum attenuation of 3dB constraint. → ws ≤ w0 ≤ ∞.
SNR= 20 log10 A As
=0 ▶ For w0 = wedf → RS/B = 2.76dB
1 n
H̃(w) = ▶ For w0 = (wedf + ws )/2 = 37 ∗ 2 ∗ π → RS/B = 4.07dB
1 + j ww0 ▶ After filtering :
SNR= G(ws ) − G(wedf ) ▶ Pour w0 = ws → RS/B = 9.63dB
▶ Gain in Db
  ▶ Choice of w0 ? → w0 = ws respects the constraint and maximizes the SNR.
w2
G̃(w) = −10 log10 1+ 2
w0

75/108 76/108
Approximating a low pass filter (1) Approximating a low-pass filter(2)
▶ Need for an approximation function that respects the constraints constrained
optimization.
▶ Criterion is optimized (for instance maximization of SNR).
▶ Two approaches are usually used:

Maximally flat frequency response

▶ Minimal distortion is achieved when the passband is flat.
▶ Let |H̃(w)| be the modulus of the frequency response of an order k filter.
Constraints for a low-pass filter
▶ |H̃(w)| is maximally flat in w = 0 if all the K th derivatives are null
▶ Passband: 1 − ε ≤ |H̃(w)| ≤ 1 pour w < wp
▶ wp : passing frequency. dK |H̃(w)|
=0
▶ ε: passband margin parameter (ε = 1/2 → −3dB). dwK

▶ Stopband : |H̃(w)| ≤ δ pour w > wa Equiripple filter

▶ wa : attenuation frequency. ▶ A better rollof (sharper decrease) can be achieved at teh cost of oscillations.
▶ δ: stopband margin parameter. ▶ Oscilations can occur in the passband (leading to distortion) of cutband (limited
▶ wa − wc is the transition band. attenuation).
▶ An equiripple filter has constant magnitude for its oscillations in the bandpass.

77/108 78/108

Butterworth filter (1) Butterworth filter (2)

▶ Butterworth filters are maximally flat [Butterworth et al., 1930]. ▶ The Butterworth filter is monotonically decreasing with the frequency.
▶ The amplitude of the frequency response can be expressed as ▶ The amplitude of the frequency response can be expressed as
Butterworth
 2n  4n  6n
1 1 w 3 w 5 w
|H̃(w)| = r (36) 1 |H̃(w)| = 1 − + − + ...
 2n 2 wc 8 wc 16 wc
w 0.8
1+ wc
▶ The derivative in w = 0 is then null up to order k = 2n − 1.
0.6
with ▶ The frequency response of a (normalized) Butterworth filter can be expressed as
0.4
▶ n: order of the filter. H̃(w) = Bn1(w) where Bn (w) is a Butterworth polynomial :
0.2
▶ wc : cutoff frequency. Q n h   i
 2 2 2k+n−1
k=1 (jw)) − 2jw cos π +1 if n = even
0
2n
0 0.2 0.4 0.6 0.8 1 , Bn (w) = Q n−1 h   i
▶ The passing wp and attenuation wa frequencies are: (jw + 1) 2 2
(jw)) − 2jw cos 2k+n−1
π +1 if n = odd
k=1 2n

For |H̃(w)| = 1 − ε For |H̃(w)| = δ Order Polynomial

 1/2n  1/2n 1 1 +√jw
ε 1−δ 2 (jw)2 + 2jw + 1
wp = wc wa = wc
1−ε δ 3 (jw + 1)((jw)2 + jw + 1)

79/108 80/108
Chebyshev filter Filter implementation
Chebyshev type 1 Chebyshev type 2
Implementation of the filter consist in finding the physical components that recovers
the selected frequency response H̃(w).
1 1

0.8 0.8 Passive filter

0.6 0.6
▶ Only passive components (R, C, L).

0.4 0.4
▶ No energy source, no amplification (conservation of energy).

0.2 0.2
▶ The input and output impedance has an effect on the frequency response
(impedance matching).
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Active filter
▶ Better rolloff than Butterworth of same order but leads to oscillations in the ▶ Use an energy source and Operational Amplifiers (OA).
bandpass (type 1) or in the stopband (type 2).
▶ OA has near infinite impedance but limited bandwidth (typically 100KHz).
▶ Equiripple filter.
▶ Saturation can occur (non-linearity).
▶ Amplitude of the frequency response:
1 ▶ Stability can be a problem (due to feedback)
|H̃(w)| = r   ▶ Tn (·): Chebyshev polynomial of
1 + ε2 Tn2 wwc Rarely use inductors in practice (price, resistance, space, mutual inductance) !
order n.

81/108 82/108

Passive filters (1) Passive filters (2)

Example filter
Butterworth Filter
▶ Brain-Computer Interface application. ▶ Corresponding frequency response with the Cauer topology.
▶ w0 = ws = 2π ∗ 12 ▶ For an order n filter with cutoff frequency wc = 1 the following structure:
▶ w0 = 1
→ RC = 1
≈ 0.01326
RC 2∗π∗12 L2 L4 Ln-1
▶ What to choose for R and C ?
▶ Price and space constraints.
C1 C3 Cn
Filter transformation
▶ low-pass → high-pass
With the values :
1/jCw → jLw et jLw → 1/jCw
▶ Ck = 2 sin( 2k−1 π) for k odd.
2n
▶ low pass → band-pass ▶ Lk = 2 sin( 2k−1 π) for k even.
2n
▶ Assuming the input and output have a 1 Ohm resistance.
1/jCw → B/C(jw + 1/jw) et jLw → L/B/(jw + 1/jw)

83/108 84/108
Active filters (1) Active filters (2)
First order active filter (with amplification)

R2 Second order active filter (Structure from [Sallen and Key, 1955])
r2
C2

-
R R
-
R + K
+

r1
C R1 C1

▶ Frequency response
▶ Frequency response K
A H̃(w) = 2
H̃(w) = jw 1 + 2zjw + (jw)
2
1+ w 0
wn wn

where r
where 1 C1 3 − K r1 + r2
R1 + R2 1 wn = √ et z= et K=
A= et w0 = R C1 C2 C2 2 r1
R1 RC
▶ Parameters: R, C1 , C2 , r1 , r2 .
▶ Parameters: R, C, R1 , R2
▶ Permute R and C for a high-pass filter.
85/108 86/108

Analog filtering : mechanical filter Modulation

Modulation is an encoding method that allows to transport a band-limited signal.
Virgo Gravitational waves detector [Acernese et al., 2014] Demodulation is the reverse operation.
▶ Interferometer to detect gravitational waves. Motivations
▶ Attenuate vibrations from the earth [Braccini et al., 1996] ▶ Raw signal transmission often not efficient (electromagnetic waves).
▶ Objective : attenuations of 10−9 for high frequencies. ▶ The change in frequencies allow transmitting several band-limited signals in
▶ Use a mirror in a chamber with mechanical filters. parallel.
▶ Use a series of mechanical filters for the attenuation. ▶ Use only of an authorized bandwidth.
▶ Active correction for remaining low frequencies.
Definitions
▶ Modulating signal x(t) is a band limited signal we want to transmit
(X(f ) = 0 pour |f | > fx ).
▶ Carrier is the periodic base signal p(t) used for transportation often :

p(t) = cos(2πfp t)

▶ Modulated signal y(t) is a band-limited signal that can be transported in the

physical medium (cable, air, optical fiber)

87/108 88/108
Amplitude Modulation (1) Amplitude Modulation (2)
h <1
y(t)
Modulation index
a(t)
▶ Envelope of the modulated signal.

x(t) y(t) a(t) = Ac |1 + ks x(t)|

h =1
▶ Maximum amplitude of modulating
signal:
Definition: Amplitude Modulation (AM)
Mx = max |x(t)|
The carrier is multiplied by the modulating signal x(t) t

▶ The index of modulation is defined as

y(t) = Ac (1 + ks x(t)) cos(2πfp t + ϕm ) h >1
h = ks Mx
▶ ks : modulation factor
▶ fp : carrier frequency ▶ h < 1: under-modulation.
▶ h > 1: over-modulation.
▶ ϕm : phase (usually added during transmission).

89/108 90/108

Amplitude Modulation (3) Amplitude Modulation (4)

Synchronous demodulation
Interpretation in the Fourier domain Done with multiplying the signal with the carrier:
▶ Multiplication → Convolution. X(f) w(t) = y(t) cos(2πfp t + ϕd )

Y (f ) = X(f ) ⋆ P (f ) = As (1 + ks x(t)) cos(2πfp t + ϕm ) cos(2πfp t + ϕd )

-fx fx As As
= (1 + ks x(t)) cos(ϕm − ϕd ) + (1 + ks x(t)) cos(4πfp t + ϕm + ϕd )
▶ The spectrum of the modulating P(f) 2 2
signal is moved around the
After low pass filtering (and removing of the constant) one can recover
frequency fp .
-fp fp
▶ Simple way to transmit a band Y(f) x̂(t) =
As
ks x(t) cos(ϕm − ϕd )
limited signal in a given bandwidth. 2
▶ Modulated signal spectrum is
-fp fp
contained in fp ± fx . ▶ cos(ϕm − ϕd ) = 1 if ϕm = ϕd .
▶ Very important to have a good synchronization (requires active components).

91/108 92/108
Amplitude Modulation (5) Applications of amplitude modulation (1)

y(t)
a(t)
x̂(t)
C R

Low frequency radio broadcasting

Asynchrone demodulation ▶ First AM transmission R. Fessenden on 23 December 1900 at Cobb Island,
Maryland (1.6Km).
▶ Synchronous demodulation can require complex active components.
▶ 1907 Lee de Forest invents the triod vacuum tube allowing for a better
▶ A coarse approximation of the envelope of the signal can be done with a simple amplification [De Forest, 1908].
diode/RC system.
▶ Weather bulletin emitted from the Eiffel Tower in february 1922.
▶ Requires under-modulation because if h < 1 then
▶ France Inter grandes ondes
a(t) = Ac |1 + ks x(t)| = Ac + Ac ks x(t) ▶ Emitted between 1 January 1947 and 31 December 2016.
▶ Allouis longwave transmitter (2000KW), now used for TDF time signal.
▶ Can require a lot of power for transmission.

93/108 94/108

Applications of amplitude modulation (2) Frequency modulation (1)

X1 (f) X2 (f) X3 (f)

-fx fx -fx fx -fx fx

Definition
Frequency modulation (FM) consists in modifying the frequency of the carrier using
x(t). The modulated signal has the following form:
-2fp -3fp -fp 0 fp 2fp 3fp  Z t 
y(t) = cos 2π f (τ )dτ
Frequency-division multiplexing 0

▶ Multiplexing: transmission of several signals in parallel. ▶ f (t) = fp + f∆ x(t) is the instantaneous frequency of the signal.
▶ If x(t) = 0 we recover the carrier.
▶ Use of a different fp for each signal.
▶ When x(t) ̸= 0 the instantaneous frequency is modified by x(t)
▶ Every signal is band limited : if ∆fp > 2fx then no loss of information.
▶ f∆ is the frequency deviation (equivalent to ks in AM).
▶ Frequency Hoping: Experimented by G. Marconi, Patent by N. Tesla
[Tesla, 1903], proposed for secret communication by [Kiesler and George, 1942].
95/108 96/108
Frequency modulation (2) Fourier Optics

Properties of Frequency Modulation

Principle
▶ More robust than AM (noise, atténuation) but propagation distance limited.
▶ More complex to implement (requires a Voltage Controled Oscillator VCO). ▶ Introduction to Fourier Optics: [Perrin and Montgomery, 2018, Goodman, 2005]

▶ Intuitively the spectrum of the modulated signal should be ̸= 0 only in the band ▶ Huygens–Fresnel principle for wave propagation.
fp ± f∆ Mx , BUT ▶ When the source is at infinity, one can use the Fraunhofer diffraction (far field).
▶ Continuous variation of the frequencies imply a spectrum on all frequencies. ▶ Several optical elements corresponds to linear operations and can be defined as
▶ The Carson bandwidth rule states that most of the signal power (98%) is in the LTI and modeled/interpreted through Fourier Transform.
band ▶ Difference between coherent VS incoherent sources.
b = 2(f∆ + fx )
Applications of Fourier transform in optics
Application of Frequency Modulation ▶ Analog image processing techniques.
▶ FM radio broadcasting. ▶ MRI : sampling of an image in the Fourier domain.
▶ Frequency modulation synthesis (chiptunes). ▶ Astronomy : modeling of telescopes, source detection, coronarography.
▶ Magnetic tape storage.

97/108 98/108

2D Fourier Transform using a lens The 4F correlator

▶ Place a transmissible object at

the focal length of a lens.
▶ The FT of the object is formed
on the focal plane behind the
lens.
▶ FT computed at the speed of
light, depends on the precision
of the optics.
▶ Let i(v) be the 2D image in the focal plane before the lens and I(u) its FT.
▶ Here f is the focal length of the lens (in optics ν or υ are often used to denote
frequency).
▶ λ is the wavelength of the light source. ▶ One can use the FT provided by Optical lenses to perform analog image filtering.
▶ The Filtering is done with a mask in the Fourier plane of the image.
▶ The the image formed in the right focal plane will be I( λf
p
) where p is the
position in the focal plane. ▶ Equivalent to a convolution (correlation).
▶ The output image is mirrored due to the two FT instead of an inverse FT.
▶ Active research domain in Optical Neural Networks [Zuo et al., 2019]
99/108 100/108
Telecope and Point Spread Function (PSF) Airy disk and angular resolution

Circular Aperture
▶ The PSF for a circular aperture often called the Airy disk
comes from the FT of the circle
F2D [circ(r)] = J1 (2p ir′ )/r′ .
▶ The diameter of the circle defines the maximum resolution.

▶ The source point in the images are considered incoherent to the observed image Angular resolution
(intensity) is the sum of the responses of each source. ▶ Minimal angle that allows discriminating two point sources.
▶ A telescope can be considered as a LTI system (at least close to the axis). ▶ Given by the Rayleigh criterion
▶ The relation between the true image and the image observed in the focal plane is
always a convolution by what is called the Point Spread Function: λ
θ = 1.22
D
y(v) = x(v) ⋆ h(v)
λ is the wavelength and D is the diameter of the telescope.
▶ The PSF h can be obtained as ▶ It corresponds to the first zero of the Bessel J1 function.
2
h(v) = F −1 [A(u)]

where A(u) is the aperture shape of the telescope.

101/108 102/108

Fourier Optics in astronomy Bibliography

▶ Signals and Systems [Haykin and Van Veen, 2007].

▶ Signals and Systems [Oppenheim et al., 1997].
▶ Signal Analysis [Papoulis, 1977].
▶ Fourier Analysis and its applications [Vretblad, 2003].

Real life telescopes ▶ Polycopiés from Stéphane Mallat and Éric Moulines [Mallat et al., 2015].
▶ New telescopes have several small mirrors : more complex PSF. ▶ Théorie du signal [Jutten, 2018].
▶ Fourier Optics model only for perfect optics. ▶ Distributions et Transformation de Fourier [Roddier, 1985]
▶ Lenses/mirrors have optical aberrations and a surface roughness introducing
scattering.
▶ Ground telescope have to compensate for atmospheric turbulence (deformable
mirrors with adaptive optics).

103/108 104/108
References I References II
[Goodman, 2005] Goodman, J. W. (2005).
[Acernese et al., 2014] Acernese, F., Agathos, M., Agatsuma, K., Aisa, D., Allemandou, N., Introduction to Fourier optics.
Allocca, A., Amarni, J., Astone, P., Balestri, G., Ballardin, G., et al. (2014). Roberts and Company Publishers.
Advanced virgo: a second-generation interferometric gravitational wave detector.
[Haykin and Van Veen, 2007] Haykin, S. and Van Veen, B. (2007).
Classical and Quantum Gravity, 32(2):024001.
Signals and systems.
[Braccini et al., 1996] Braccini, S., Bradaschia, C., Del Fabbro, R., Di Virgilio, A., Ferrante, John Wiley & Sons.
I., Fidecaro, F., Flaminio, R., Gennai, A., Giassi, A., Giazotto, A., et al. (1996).
[Hewitt and Hewitt, 1979] Hewitt, E. and Hewitt, R. E. (1979).
Seismic vibrations mechanical filters for the gravitational waves detector virgo.
The gibbs-wilbraham phenomenon: an episode in fourier analysis.
Review of scientific instruments, 67(8):2899–2902.
Archive for history of Exact Sciences, 21(2):129–160.
[Butterworth et al., 1930] Butterworth, S. et al. (1930).
[Hunter, 2019] Hunter, J. (2019).
On the theory of filter amplifiers.
Notes on partial differential equations. 2014.
Wireless Engineer, 7(6):536–541.
URL https://www. math. ucdavis. edu/˜ hunter/m218a 09/pde notes. pdf.
[De Forest, 1908] De Forest, L. (1908).
[Jutten, 2018] Jutten, C. (2018).
Space telegraphy.
Théorie di signal.
US Patent 879,532.
Univ. Grenoble Alpes - Polytech’ Grenoble.
[Fourier, 1807] Fourier, J. B. J. (1807).
[Kiesler and George, 1942] Kiesler, M. H. and George, A. (1942).
Mémoire sur la propagation de la chaleur dans les corps solides.
Secret communication system.
US Patent 2,292,387.
105/108 106/108

References III References IV

[Mallat et al., 2015] Mallat, S., Moulines, E., and Roueff, F. (2015).
Traitement du signal.
Polycopié MAP 555, École Polytechnique. [Sallen and Key, 1955] Sallen, R. P. and Key, E. L. (1955).
A practical method of designing rc active filters.
[Oppenheim et al., 1997] Oppenheim, A. V., Willsky, A. S., and Nawab, S. H. (1997).
IRE Transactions on Circuit Theory, 2(1):74–85.
Signals and systems prentice hall.
Inc., Upper Saddle River, New Jersey, 7458. [Tesla, 1903] Tesla, N. (1903).
System of signaling.
[Papoulis, 1977] Papoulis, A. (1977).
US Patent 725,605.
Signal analysis, volume 191.
McGraw-Hill New York. [Vretblad, 2003] Vretblad, A. (2003).
Fourier analysis and its applications, volume 223.
[Perrin and Montgomery, 2018] Perrin, S. and Montgomery, P. (2018).
Springer Science & Business Media.
Fourier optics: basic concepts.
arXiv preprint arXiv:1802.07161. [Zuo et al., 2019] Zuo, Y., Li, B., Zhao, Y., Jiang, Y., Chen, Y.-C., Chen, P., Jo, G.-B., Liu,
J., and Du, S. (2019).
[Rahimi and Recht, 2008] Rahimi, A. and Recht, B. (2008).
All-optical neural network with nonlinear activation functions.
Random features for large-scale kernel machines.
Optica, 6(9):1132–1137.
In Advances in neural information processing systems, pages 1177–1184.

[Roddier, 1985] Roddier, F. (1985).

Distributions et transformée de fourier.
107/108 108/108