Syllabus

U NIVERSITÉ LIBRE DE B RUXELLES
É COLE POLYTECHNIQUE DE B RUXELLES

ELEC-H422
Wireless communication channels
Prof. Philippe D E D ONCKER
2024.2
Table of contents
1 Wireless Communication Channels 3

1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Physical channel models . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Free-space model . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Multipaths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 Time-variant channels . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4 Quasi-static approximation . . . . . . . . . . . . . . . . . . . . . 8
1.3 Tapped Delay Line models . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Wideband models . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.2 Narrowband models . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Course outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Narrowband channels 15
2.1 Deterministic model in a local area . . . . . . . . . . . . . . . . . . . . . . 15
2.1.1 Two-waves model . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.2 N -waves model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.3 1+N -waves model . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Spatial statistical characterization . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Stochastic model in a local area . . . . . . . . . . . . . . . . . . . 22
2.2.2 Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3 Mean values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Spatial correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 Stochastic model . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Time-variant channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.1 Stochastic model . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.2 Receiver moving - IOs fixed . . . . . . . . . . . . . . . . . . . . . 36
2.4.3 Receiver fixed - IOs moving . . . . . . . . . . . . . . . . . . . . . 40
3 Wideband channels 45
3.1 Deterministic wideband model . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.1 1-wave model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.2 2-waves model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Stochastic wideband model . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 Frequency correlation and delay spread . . . . . . . . . . . . . . . 48
1
2 TABLE OF CONTENTS
3.2.2 Coherence bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.3 Tapped Delay Line models . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4 Example : channel models for 60GHz WLAN systems . . . . . . . . . . . 54
3.4.1 Principal rays calculation . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.2 Intra-clusters parameters . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.3 Cluster amplitudes . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.4 Channel impulse response . . . . . . . . . . . . . . . . . . . . . . 58
4 Multi-dimensional channel analysis 59

4.1 Global duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Scattering functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 Channel functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4 Dispersion-Selectivity duality . . . . . . . . . . . . . . . . . . . . . . . . . 64
5 MIMO channels 67
5.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Channel matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Spatial multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3.1 Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3.2 Channel degrees-of-freedom . . . . . . . . . . . . . . . . . . . . . 72
5.4 Sub-channel correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.4.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.4.2 Correlation effects . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.5 Double-directional models . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.6 Wideband models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Wireless Communica-
tion Channels
1.1 Definition
1
The wireless communication channel is the environment where wireless signals propa-
gate. It also includes both the transmit and receive antennas. The wireless channel accepts
at its input the baseband signal x(t) produced by the transmitter. It delivers at its output the
baseband signal y(t) to the receiver.
1.2 Physical channel models
1.2.1 Free-space model
In the simplest case, only the Line-of-Sight (LOS) propagation path exists, and the recei-
ved baseband signal y(t) is an attenuated and delayed version of the transmitted baseband
signal x(t) :
y(t) = a ejϕ e−j2πfc τ x(t − τ ) (1.1)
where fc is the carrier frequency, τ the propagation delay, a the attenuation factor, and ϕ
models any phase shift occurring in the channel and not due to the propagation delay (e.g.
due to the antennas).
The signal at the receiver is corrupted by noise and interferences. Those are usually
modeled as a gaussian process n(t), and the free-space channel model is equivalent to the
Additive White Gaussian Noise (AWGN) model seen in Master 1 :
3
4 Wireless Communication Channels
y(t) = a ejϕ e−j2πfc τ x(t − τ ) + n(t) (1.2)
In this model, the actual propagation delay τ is of no importance, and we write
y(t) = α x(t) + n(t) (1.3)
where α = a ejϕ e−j2πfc τ is the complex factor modelling all attenuations and phase shifts
in the channel.
In wireless communications, the receiver location plays a significant role in channel

modelling. It can be introduced in (1.2) by noting that e−j2πfc τ models the phase shift
due to the propagation path length. Under plane wave propagation assumption, this latter
⃗
can also be written as e−j βc ·⃗r where β⃗c is the wave vector (at carrier frequency) and ⃗r the
receiver location (the transmitter being at the origin of the coordinate system) :
⃗
y(t) = a ejϕ e−j βc ·⃗r x(t − τ ) + n(t) (1.4)
Good To Know
Since n(t) should always be added in channel input/output relationships, for sake of conci-
seness, it will always be omitted in the sequel of this course.
1.2.2 Multipaths
The received signal is actually made of multiple copies of the transmitted signal refer-
red to as multipaths components (MPC), inducing time dispersion in the channel.
Each MPC arrives at the receiver with its own attenuation and delay, so that (1.1) be-
comes :
XN
y(t) = an ejϕn e−j2πfc τn x(t − τn ) (1.5)
n=1
where N is the number of MPC’s arriving at the receiver. The baseband channel impulse
response h(τ ) is defined as
1.2 Physical channel models 5
N
X
h(τ ) = an ejϕn e−j2πfc τn δ(τ − τn ) (1.6)
n=1
or, in more compact form

N
X
h(τ ) = αn δ(τ − τn ) (1.7)
n=1
where αn = an ejϕn e−j2πfc τn . This model is referred to as the Linear Time-Invariant (LTI)
baseband impulse response.
With (1.6) or (1.7), (1.5) can be written as

Z ∞
y(t) = h(τ ) x(t − τ ) dτ (1.8)
0
which is a convolution product, i.e. the usual input/output relationship for LTI systems. An
example of experimental impulse response is given in Figure 1.1.
Taking the Fourier transform of h(τ ), the baseband channel transfer function H(f )
is defined by
Z ∞
H(f ) = h(τ ) e−j2πf τ dτ (1.9)
−∞
where f is the baseband frequency. With (1.6) :
N
X
H(f ) = an ejϕn e−j2π(fc +f )τn (1.10)
n=1
The Fourier transform of (1.8) gives
Y (f ) = H(f ) X(f ) (1.11)
As an example, the experimental transfer function related to the impulse response of Figure
1.1 is plotted in Figure 1.2.
−60
−70
−80
−90
−100
[db]
−110
−120
−130
−140
−150
0 50 100 150 200 250
Delay [ns]
F IGURE 1.1 – Example of experimental impulse response (indoor wireless communication, the
impulse response has been measured over a 2GHz bandwidth around 60GHz).
1.2 Physical channel models 7
−50
−55
−60
−65
[db]
−70
−75
−80
−85
59 59.2 59.4 59.6 59.8 60
0
<latexit sha1_base64="NUmLeU5Z+noToX/y4u3RJY84jxw=">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</latexit>
<latexit
60.2 60.4 60.6 60.8 1
61
<latexit sha1_base64="JpkgzSaZ0dx621uvkC7w88ODhqA=">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</latexit>
<latexit
f f[GHz]
[GHz]
<latexit sha1_base64="Wcji+6e7eph2CK2p6lyFh7wSv5s=">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</latexit>
F IGURE 1.2 – Experimental transfer function related to the impulse response in Figure 1.1.
1.2.3 Time-variant channels
This is not the end of the story. Each parameter in (1.6) may vary over time, and the
time-variant impulse response is defined as
N (t)
X
h(τ, t) = an (t) ejϕn (t) e−j2πfc τn (t) δ(τ − τn (t)) (1.12)
n=1
It is usually assumed that N and τn are slowly varying as compared to an and ϕn , and both
are kept time-invariant :
N
X
h(τ, t) = an (t) ejϕn (t) e−j2πfc τn δ(τ − τn ) (1.13)
n=1
Or
N
X
h(τ, t) = αn (t) δ(τ − τn ) (1.14)
n=1
where αn (t) = an (t) ejϕn (t) e−j2πfc τn . With this time-variant model, (1.8) becomes
Z ∞
y(t) = h(τ, t) x(t − τ ) dτ (1.15)
0
which is no more a convolution product !
The time-variant transfer function is defined as

Z ∞
H(f, t) = h(τ, t) e−j2πf τ dτ (1.16)
−∞
but (1.11) cannot be used anymore.
1.2.4 Quasi-static approximation
In the quasi-static approximation, a time-variant channel is assumed to be time invariant

during the transmission of each block of symbols. As a consequence, (1.8) and (1.11) can
be used for each block of symbols separately.
1.3 Tapped Delay Line models 9
With this approximation, h(τ, t) can be intuitively understood :
— Time t measures the usual clock time that is changing slowly as compared to the
transmission duration of each block of symbols.
— Delay τ measures propagation time durations of the same order as, or faster than
the duration of the block of symbols.
Good To Know
In the quasi-static approximation, the time-variant impulse response h(τ, t) models through
τ the channel time dispersion that exists at time t.
In the sequel of this course, the quasi-static approximation is always assumed.
1.3 Tapped Delay Line models
1.3.1 Wideband models
Channel model (1.14) is correct from a physical point of view, but not from a system
point of view. Indeed, the receiver cannot resolve all copies arriving at delays τn due to the
system finite bandwidth. If the input signal x(t) is bandlimited with bandwidth B, the time
resolution is ∆τ = 1/2B.
As any bandlimited signal, x(t − τ ) can be written as

∞
X
x(t − τ ) = x(t − l∆τ ) sinc 2B(τ − l∆τ ) (1.17)
l=−∞
where sinc(x) = sin πx/πx.
By (1.15)
Z ∞ X
y(t) = h(τ, t) x(t − l∆τ ) sinc 2B(τ − l∆τ ) dτ
0 l
X Z ∞ (1.18)
= x(t − l∆τ ) h(τ, t) sinc 2B(τ − l∆τ ) dτ
l 0
We define the complex gain of the tap l as

Z ∞
hl (t) = h(τ, t) sinc 2B(τ − l∆τ ) dτ (1.19)
0
to obtain the Tapped Delay Line model of the system :
∞
X
y(t) = x(t − l∆τ ) hl (t) (1.20)
l=−∞
Actually, from (1.19), the gains hl (t) are non negligible only for a few positive or null
values of l, and the summation is limited to l = 0..L, where L depends on the channel
maximum propagation delay :
L
X
y(t) = x(t − l∆τ ) hl (t) (1.21)
l=0
which is mathematically equivalent to (1.15) if we define a Tapped Delay Line model

of the impulse response :
L
X
hTDL (τ, t) = hl (t) δ(τ − l∆τ ) (1.22)
l=0
We see that, due to the system finite bandwidth, the physical impulse response (1.14)
has been discretized into L + 1 taps of duration ∆τ . This process is illustrated in Figure
1.3.
Tap gains can be computed from (1.19) and (1.14) :

Z ∞XN
hl (t) = αn (t) δ(τ − τn ) sinc 2B(τ − l∆τ ) dτ
0 n=1
N
(1.23)
X Z ∞
= αn (t) δ(τ − τn ) sinc 2B(τ − l∆τ ) dτ
n=1 0
i.e.
N
X
hl (t) = αn (t) sinc 2B(τn − l∆τ ) (1.24)
n=1
|↵1 |
<latexit sha1_base64="x0J14bCPVOSXpqWJ7HaE5GAS3Xg=">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</latexit>
|h1 |
<latexit sha1_base64="evDof6LdujEn7W+F3d16HgG0LmM=">AAAC0HicjVHLSsNAFD3GV62vqks3wSK4KkkR1F3BjcuKxhbaWpLptA3Ni8lEKEXErV/gVn9K/AP9C++MKahFdEKSM+fec2buvV4S+Km0rNc5Y35hcWm5sFJcXVvf2CxtbV+lcSYYd1gcxKLpuSkP/Ig70pcBbyaCu6EX8IY3OlXxxg0XqR9Hl3Kc8E7oDiK/7zNXEnXdppg0h11bg26pbFUsvcxZYOegjHzV49IL2ughBkOGEBwRJOEALlJ6WrBhISGugwlxgpCv4xy3KJI2oyxOGS6xI/oOaNfK2Yj2yjPVakanBPQKUprYJ01MeYKwOs3U8Uw7K/Y374n2VHcb09/LvUJiJYbE/qWbZv5Xp2qR6ONY1+BTTYlmVHUsd8l0V9TNzS9VSXJIiFO4R3FBmGnltM+m1qS6dtVbV8ffdKZi1Z7luRne1S1pwPbPcc4Cp1o5qVjnh+VaNZ90AbvYwwGN8wg1nKEOh6wFHvGEZ+PCGBt3xv1nqjGXa3bwbRkPH2BQlJw=</latexit>
⌧
<latexit sha1_base64="DsSH6dOIgCZ6zq49cmU0tpzz1lg=">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</latexit>
<latexit
⌧
<latexit
<latexit sha1_base64="LmfIRSoPZNL+RdyhZVXOX1fS4jA=">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</latexit>
⌧
F IGURE 1.3 – Physical impulse response and TDL model.
|h1 |
<latexit sha1_base64="evDof6LdujEn7W+F3d16HgG0LmM=">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</latexit>
⌧
<latexit
⌧
<latexit
<latexit sha1_base64="LmfIRSoPZNL+RdyhZVXOX1fS4jA=">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</latexit>
⌧ <latexit sha1_base64="LmfIRSoPZNL+RdyhZVXOX1fS4jA=">AAACzHicjVHLSsNAFD2Nr1pfVZdugkVwVdIiqLuCLtxZwWixLTJJpzV08iCZCKV06xe41d8S/0D/wjvjFNQiOiHJmXPvOTP3Xi8RQSYd57Vgzc0vLC4Vl0srq2vrG+XNrasszlOfu34s4rTlsYyLIOKuDKTgrSTlLPQEv/aGJyp+fc/TLIijSzlKeDdkgyjoBz6TRN10TrmQrCNZfluuOFVHL3sW1AyowKxmXH5BBz3E8JEjBEcESViAIaOnjRocJMR1MSYuJRToOMcEJdLmlMUpgxE7pO+Adm3DRrRXnplW+3SKoDclpY090sSUlxJWp9k6nmtnxf7mPdae6m4j+nvGKyRW4o7Yv3TTzP/qVC0SfRzpGgKqKdGMqs43Lrnuirq5/aUqSQ4JcQr3KJ4S9rVy2mdbazJdu+ot0/E3nalYtfdNbo53dUsacO3nOGeBW68eV52Lg0qjbiZdxA52sU/jPEQDZ2jCJesIj3jCs3Vu5dbYmnymWgWj2ca3ZT18AGaOkw4=</latexit>
⌧
F IGURE 1.4 – Physical impulse response and TDL model under US assumption.
Due to the sinc, only propagation paths with τn ∼ l∆τ significantly contribute to the
sum, and, in a first approximation, we can say from (1.24) that the receiver sums up all
signal copies arriving during every ∆τ tap
X X
hl (t) ≃ αn (t) = an (t) ejϕn (t) e−j2πfc τn (1.25)
τn ∈ tap l τn ∈ tap l
Under this approximation, each propagation path contributes to one and only one tap.
Since the propagation paths are generally assumed uncorrelated, the taps given by (1.25)
are also uncorrelated. We speak about the Uncorrelated Scattering (US) in the delay do-
main assumption of the channel. This approximation is illustrated in Figure 1.4.
1.3.2 Narrowband models
Let us define the channel delay spread στ as 1
στ = max |τi − τj | (1.26)

i,j
and the channel coherence bandwidth ∆fc :

1
∆fc = (1.27)
στ
If B << ∆fc , i.e. ∆τ >> στ , all MPC’s fall into a single tap, and, by noting h(t) ≡ h0 (t)
in (1.22) :
hTDL (τ, t) = h(t) δ(τ ) (1.28)
where, under US assumption, from (1.25)
N
X
h(t) = an (t) ejϕn (t) e−j2πfc τn (1.29)
n=1
(1.21) becomes :
y(t) = h(t) x(t) (1.30)
which is referred to as a narrowband model. It is important to remind that the narrowband

condition depends on both the system (through ∆τ ) and the channel (through στ ) !
Good To Know
The narrowband model (1.30) looks very similar to the free-space model (1.3), but it is not
equivalent from the physical point of view : while the free-space model takes into account
a LOS component only, the narrowband model includes the interference effects among
MPC’s : h(t) defined by (1.29) will suffer from space and time fading, as we will discuss
in Chapter 2.
The narrowband channel transfer function (1.16) reads
H(f, t) = h(t) (1.31)

1. More general definitions of στ and ∆fc will be given in Chapter 3
1.4 Course outline 13
which is independent of the baseband frequency. We speak about frequency flat channels :
all spectral components of the transmitted signal are affected by the same channel gain
h(t).
This is not true in general : with the wideband model (1.22), H(f, t) depends on the
frequency. We speak about frequency selective channels : each spectral component of the
transmitted signal will be affected by a different channel gain. This is obvious, for instance,
in the example drawn in Figure 1.2.
1.4 Course outline
We can now define the goals of this course :

1. We will first study in Chapter 2 how to characterize and model channels under
narrowband condition. We already started this study in Master 1. This year, we will
assume that channel attenuation modeling is known (path loss, shadowing) and we
will focus on an in-depth study of space and time fading. This study is important
per se, but it is also a pre-requisite to be able to address wideband models : as seen
from (1.24) or (1.25), each tap of a wideband model is mathematically equivalent
to a narrowband channel (1.29), and, so, both are characterized and modeled in the
same way.
2. In Chapter 3, we will then address wideband channel characterization and model-
ling.
3. A global multi-dimensional modelling synthesis will be given in Chapter 4.
4. Finally, in Chapter 5, we will see how these models can be extended to the important
case of multi-antenna systems.
Narrowband channels
2.1 Deterministic model in a local area

2
In Telecommunication Channels, we studied how propagation models predict large
scale variations of received power. These large scale variations occur over distances equal
to several tens of wavelengths, i.e. several meters with the current mobile technologies.
We also studied that multipaths induce another kind of signal variations at wavelength
scale : multipath interferences in a given local area cause signal fading referred to as small
scale fading.
A typical experimental result of a transfer function amplitude measured in a local area

(5λ wide) is shown in Figure 2.1. The interference pattern clearly appears. This example
corresponds to a usual WiFi communication : an access point located indoor, 6m away
from the local area of the receiver, with no obstacle obstructing the Line-of-Sight (LOS).
If this WiFi communication is to be simulated, we have to be able to reproduce in Matlab
this interference pattern.
In order to characterize and model small scale fading, let us start with the time-invariant
narrowband channel model, from (1.29) :
N
X
h = an ejϕn e−j2πfc τn (2.1)
n=1
We will study the spatial dependence of this transfer function, in some local area, as we
did last year.
15
16 Narrowband channels
-44
-45
-40
-46
-45
dB
[dB]
-47
|h| in
-50
|H|
-48
-55 -49
5
4 5 -50
4
3
3 -51
y/l 2
2
1 1 x/l -52
y-axis of the grid in λ 0 0 x-axis of the grid in λ
F IGURE 2.1 – Transfer function amplitude. Experimental result.
x
~11
<latexit sha1_base64="isXYf/KQ2bifSMaqFN/kxj4q0Mw=">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</latexit>
~N
<latexit sha1_base64="raSPUpgl9yiCEtNYMQFWdRyGM9U=">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</latexit>
N
~
r
✓1 z
✓2
RX
~2 2
<latexit sha1_base64="fixKEgK9orfI8t0ICmhuzYywMVU=">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</latexit>
local a rea
F IGURE 2.2 – N (locally) plane waves incident on a receive antenna located in a local area.
2.1 Deterministic model in a local area 17
Definition
A local area is a region of space where the transfer function h can be modeled by using a
given set of N (locally) plane waves (Figure 2.2).
Typical dimensions of local areas range from a few wavelengths (indoor) to a few tens of
wavelengths (outdoor)
In a given local area, (2.1) can be written so as to explicitly introduce the phase shift
due to any small displacement ⃗r, defined in a coordinate system attached to the local area
(Figure 2.2) :
N
⃗
X
h(⃗r ) = an ejϕn e−j βcn ·⃗r e−j2πfc τn
n=1
where τn is here the propagation delay till the center of the local area, and β⃗cn the wave
vector of MPC n (at carrier frequency). Or, by defining φn = ϕn − 2πfc τn :
N
⃗
X
h(⃗r ) = an ejφn e−j βcn ·⃗r (2.2)
n=1
This transfer function can be calculated by using a ray-tracing software, for instance. It
is referred to as a deterministic channel model.
2.1.1 Two-waves model
Let us first consider a simplified model where the receiver picks up two plane waves,
with identical amplitudes and traveling in opposite directions as shown in Figure 2.3(a).
The transfer function (2.2) is, considering for sake of simplicity an = 1 and φn = 0
(n = 1, 2) :
h(z) = e−jβc z + ejβc z (2.3)
whose amplitude is
|h(z)| = 2 | cos βc z|
drawn in Figure 2.3(b). It corresponds to a typical two-waves interference pattern, with
zeros every half wavelengths. This pattern is most often drawn in dB (20 log |h|) as in
Figure 2.3(c). In this latter figure in dB, the zeros of |h| become fadings to −∞.
RX
z
(a)
h 1
0
0 1 2 3 4
z/l
(b)
-10
-20
h [dB]
-30
-40
-50
0 1 2 3 4
z/l
(c)
F IGURE 2.3 – (a) Two-waves model, (b) Interference pattern (linear scale), (c) Interference pattern
(dB scale)
The received power experiences the same fading process, at very small scale (∼ λ/2,
i.e. a few centimeters with current mobile technologies, or even a few millimeters with
mm-wave communications). Fades can be very deep, and it is usual to experience 30 dB
fades during communication. Small-scale fading is thus a major phenomenon of wireless
communication channel.
2.1.2 N -waves model
Let us now consider a more realistic case where N waves are incident onto a local area
as depicted in Figure 2.2. The transfer function h is given by (2.2).
Examples of interference patterns of this model are drawn in Figure 2.4 for N =
1, 2, 10. When N = 1 a single wave is incident and |h| is constant. When N = 3, an
interference pattern appears, but it is still regular. When N = 10, the interference pattern
has lost any regularity. The transfer function amplitude when the receiver lies in this pattern
is drawn in Figure 2.5. Fades seem impossible to predict. They look like a random process.
2.1.3 1+N -waves model
The previous model assumed that all incident waves have same amplitudes. This is not
the case when there is a LOS between the transmitter and the receiver : the direct LOS wave
is then much stronger than the multipaths since it has a shorter propagation distance, and
without any reflection or diffraction. In the local area, the transfer function is then written
as
N
⃗ ⃗
X
h(⃗r ) = a0 ejφ0 e−j βc0 ·⃗r + ai ejφn e−j βcn ·⃗r (2.4)
n=1
where β⃗c0 is the wave vector of the direct wave.
Example of interference pattern is drawn in Figure 2.6 for one direct wave and 10
MPCs. Amplitude a0 was chosen 20 times stronger than the other ai ’s. The interference
pattern looks much more regular than in the N -waves model, of course. The transfer func-
tion amplitude when the receiver lies in the 1 + N -waves interference pattern is drawn in
Figure 2.7. Fades are much shallower than in the N -waves model.
h
0
(a)
4l
h
0
(b)
4l
h
0
(c)
4l
F IGURE 2.4 – Typical interference pattern for the N -waves model, (a) N = 1, (b) N = 3, (c)
N = 10 (incidence angles were chosen randomly, the an supposed to be equal to 1/N , and the
phases φn were chosen randomly but equal to each others).
-5
-10
h [dB]
-15
-20
0 l 2l 3l 4l
Position
F IGURE 2.5 – Transfer function amplitude when the receiver moves through the N -waves model
interference pattern (N =10)
4l
F IGURE 2.6 – Typical interference pattern in the 1+N -waves model (N = 10)
-2
-4
h [dB]
-6
-8
0 l 2l 3l 4l
Position
F IGURE 2.7 – Transfer function amplitude when the receiver moves through the 1+N -waves
interference pattern
2.2 Spatial statistical characterization
2.2.1 Stochastic model in a local area
Previous deterministic method has drawbacks, however :

— Phases and amplitudes of the incident waves depend on the interactions of each
wave with the Interacting Objects (IOs). To be able to predict these values, a full
knowledge of all the IOs characteristics (position, material,..) is required, which
is not always available. A small error in these characteristics would lead to major
differences in the calculated phases φn , and so in the predicted interference patterns.
⃗
— Terms e−j βcn ·⃗r in (2.2) very quickly vary with position ⃗r. A small shift in receiver
position would lead to a received power totally different than the one predicted by
ray-tracing. So, this latter method would be impossible to validate experimentally.
As a consequence, another methodology is most often considered : a stochastic model

of the interference patterns is built, by assuming that h is a random function in a given local
area.The exact interference pattern is not looked for, but rather its statistical properties (first
and second order moments) to be able to reproduce in Matlab a similar interference pattern,
having the same statistical properties.
For h to become a random function, we must assume that it includes some random
2.2 Spatial statistical characterization 23
variable. As explained above, the MPC phases are impossible to predict in a given local
area. So we will consider that phases Φn = φn − β⃗cn · ⃗r in (2.2) are random in the local
area.
Our deterministic model (2.2) for h in a given local area is so replaced by the stochastic
model 1
X N
h = an ejΦn (2.5)
n=1
where Φn (= φn − β⃗cn · ⃗r) is a random variable with uniform distribution between 0 and
2π.
Each draw of these phases gives one possible value of h in the local area. We say that
each draw gives one realization of h in the local area. The mean value (mathematical
expectation) over these realizations will be denoted E[ ].
So, in this "engineer" approach, mean values over realizations E[ ] are to be understood
as averaging over spatial samples in the local area. For instance, E[|h|] is to be understood
as the mean value of |h| in the local area.
2.2.2 Distribution
N -waves model
By (2.5) :
N
X N
X
h = an cos(Φn ) + j an sin(Φn ) (2.6)
n=1 n=1
so that the real an imaginary parts of h are the sum of N random variables. By the central-
limit theorem, when N is high enough (typically from N = 10 onwards) the real and
imaginary parts of h have gaussian distributions, and the phase of h has a uniform distri-
bution between 0 and 2π.
In that case, the statistical distribution of |h| is referred to as a Rayleigh distribution whose
pdf is
1. For sake of notation readiness, deterministic and random functions are written in the same way through
the whole course.
PDF
PDF
0.2 0.4 0.6 0.8 1. 2. 3. 4. 5. 6.
h Arg h
F IGURE 2.8 – Histogram of |h|, Arg h and Rayleigh distribution fitting
|h| −|h|2 /2σ2

PDF (|h|) = e (2.7)
σ2
and cdf :
2 /2σ 2
Pr (|h| < γ) = 1 − e−γ (2.8)
2
For |h| :
1 −|h|2 /2σ2
PDF (|h|2 ) = e
2σ 2 (2.9)
2
Pr (|h|2 < γ) = 1 − e−γ/2σ
In these equations, σ is the Rayleigh distribution parameter.
Rayleigh distributions have the following properties :

r
π
E [ |h| ] = σ
2
E |h|2 = 2 σ 2

(2.10)
median of |h| = 1, 18 σ
Max(PDF(|h|)) = σ
The histogram of |h| obtained by the values of Figure 2.4(c) is drawn in Figure 2.8,
together with a Rayleigh distribution best fitting the histogram. The histogram of Arg h is
also drawn and is quasi uniform between 0 and 2π. In practice, there is a very good match
between experimental values and Rayleigh distributions in Non Line-of-Sight (NLOS)
cases.
PDF
0.2 0.4 0.6 0.8 1.

h
F IGURE 2.9 – Histogram of |h| and Rice distribution fitting.
Implementation
By convention, h is normalized so that
E |h|2 = 1

(2.11)
and must be multiplied by the path loss and shadowing when introduced in the Matlab
code, if necessary.
This normalized h is computed in Matlab by drawing
h = CN (0, 1)
where a random complex variable Z = X + jY is CN (0, 1) if X and Y are independent

random variables with normal distributions (zero mean and variance 1/2).
1+N -waves model
In that model, it is possible to show that |h| has a Rice distribution :
√ !
|h| 2 2 |h| 2K
PDF (|h|) = 2 e−|h| /2σ e−K I0 (2.12)
σ σ
Measured data
0.9
Ricean fit
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4
F IGURE 2.10 – CDF of |h| (normalized) and Rice distribution fitting.
where I0 is the modified Bessel function of order 0, and σ and K the two parameters of the
Rice distribution.
It is possible to show that 2σ 2 is the mean power of the mulipaths in the local area, and
K is
a2
K = PN 0 (2.13)
n=1 a2n
K is referred to as the Rice factor, usually written in dB :

a2
K[dB] = 10 log PN 0
n=1 a2n
It gives the relative

PN power of the LOS component (a20 ) compared to the mean power of the
multipaths ( n=1 a2n ).
CDF of |h| obtained from Figure 2.6 and a fitted Rice distribution are drawn in Figure
2.9. Experimental cdf deduced from Figure 2.1 and fitted Rice distribution are drawn in
K=10
2
PDF
1 K=0 K=3
0
0.5 1 1.5 2
h
F IGURE 2.11 – Rice distribution examples.
Figure 2.10. As seen, the match is perfect in this example by choosing K = 12 dB.
Figure 2.11 gives some examples of Rice distributions for a same mean value of |h|2 in
the local area. K = 0 corresponds to a Rayleigh distribution (no LOS component). With
growing K, the distribution becomes more and more tight : the LOS component becomes
more and more dominant and fading decreases.
The K factor is one of the key parameters of wireless channels.
Implementation
If one MPC dominates the other multipaths, a normalized h is obtained with

r r
K jΦ 1
h = e + CN (0, 1) (2.14)
K +1 K +1
where K is the channel Rice factor (experimentally found), and Φ is uniformly distributed.
2.2.3 Mean values
N -waves model
By (2.5), clearly
E[h] = 0
The mean power is given by :
< P > = E |h |2 = E [ h h∗ ]

" N N
#
X X
jΦn −jΦm
= E an e am e
n=1 m=1 (2.15)
N
XX N
an am E ejΦn e−jΦm

=
n=1 m=1
The MPCs are usually assumed to have uncorrelated phase terms :
E ejΦn e−jΦm = δnm

(2.16)
This assumption is also referred to as an Uncorrelated Scattering (US) assumption of the

wireless channel. In (2.5), the Φn model the random phases of the MPCs having different
wave vectors, i.e. coming from different incident angles. So, (2.16) is a US assumption in
the angular domain (while we saw a US assumption in the delay domain in Chapter 1).
The mean power so becomes
N
X
< P >= a2n (2.17)
n=1
The mean power in a given local area is independent of the MPC phases, which is a huge
advantage since the phases are impossible to predict or to measure, as we have already
seen.
2.3 Spatial correlation 29
1+N -waves model
Similarly :
E[h] = 0
N
2
X (2.18)
a20 a2n

E |h| = +
n=1
2.3 Spatial correlation
2.3.1 Stochastic model
The transfer function is random, but some spatial correlation exists in a local area. To
calculate this correlation, let us consider a displacement z ⃗1z around some point ⃗r0 in the
local area :
⃗r = ⃗r0 + z ⃗1z (2.19)
The transfer function (2.2) becomes :
N
⃗
X
h(z) = an ejφn e−j βcn ·⃗r0 ejβc cos θn z (2.20)
n=1
⃗
In the stochastic approach, ejφn e−j βcn ·⃗r0 = ejΦn is assumed random, and
N
X
h(z) = an ejΦn ej2πkn z (2.21)
n=1
1
where kn = λ
cos θn , with λ the wavelength at carrier frequency.
Let us define the MPC angular distribution by :

N
X
a(k) = an ejΦn δ(k − kn ) (2.22)
n=1
which is a random function (through the phases Φn ). It gives the MPCs incident on the
local area as a function of the projection of their wave vector on the z axis.
The transfer function (2.21) becomes :

Z ∞
h(z) = a(k) ej2πkz dk (2.23)
−∞
which is the Fourier transform of a(u). By inverting :

Z ∞
a(k) = h(z) e−j2πkz dz (2.24)
−∞
i.e. by measuring the transfer function along some axis z in the local area, it is possible to
easily compute the angular distribution of the MPC arriving in the local area ! It is worth
noting that it is impossible to discriminate all the MPCs since two MPCs arriving from the
same direction k (same incidence angle with respect to axis z, i.e. arriving from a cone with
axis z) cannot be resolved. So, we don’t really get the angular distribution, but a proxy, the
k-distribution of the MPCs.
This emphasizes a first wireless channel duality :
Channel Duality 1
Space ⇐⇒ Angle of Arrival (AoA)
Let us now calculate the correlation between two points z + ∆z and z :
R(z + ∆z, z) = E [h(z + ∆z) h∗ (z)]

Z ∞ Z ∞
′ (2.25)
= E [a(k)a∗ (k ′ )] ej2πk(z+∆z) e−j2πk z dk dk ′
−∞ −∞
By (2.22), under US assumption (2.16) :
E [a(k)a∗ (k ′ )] ≡ P (k) δ(k − k ′ ) (2.26)
where P (k) is the MPCs Power Angular Spectrum (PAS). It gives the incident power
density (per-unit k) onto the local area.
By (2.25) and (2.26) :

Z ∞
R(∆z) = P (k) ej2πk∆z dk (2.27)
−∞
which only depends on displacement ∆z (and not on the position z in the local area).
The channel is said to be Wide-Sense Stationary, WSS, w.r.t. space.
By inverting (2.27), the PAS can be written as the Fourier transform of the spatial correla-
tion : Z ∞
P (k) = R(∆z) e−j2πk∆z d∆z (2.28)
−∞
In order to model the interference pattern in a local area (like Figure 2.1), it is required
to know the statistical distribution of the transfer function, but also its spatial correlation.
By (2.27), spatial correlation is easily deduced from the PAS which has a simple physical
meaning.
The Angular spread σk characterizes the angular extension of P (k) :

s
∞
1
Z
σk = k 2 P (k) dk − km
2 (2.29)
P −∞
where P is the total power : Z ∞

P = P (k) dk
−∞
and km the mean value of k :

∞
1
Z
km = k P (k) dk
P −∞
Let us consider two points in the local area, as drawn here right. If
only one MPC is incident onto the local area, by (2.22) with a1 = 1 :
q1
a(k) = ejΦ1 δ(k − k1 )
Dz
where k1 = λ1 cos θ1 and Φ1 is a random phase, as usual. The PAS is
given by (2.26) :
P (k) = δ(k − k1 ) (2.30)
1
q1-q2=0,5
0.8 q1-q2=1
|R(Dz)|
0.6
q1-q2=1,5
0.4
0.2
0 Dz l/2
F IGURE 2.12 – Spatial correlation amplitude when two MPCs are incident onto the local area
and the correlation function is computed by (2.27) :

Z ∞
R(∆z) = δ(k − k1 ) ej2πk∆z dk
−∞ (2.31)
jβc ∆z cos θ1
= e
Its amplitude does not depend on ∆z, and is always equal to 1 : with only one incident
MPC, the interference pattern is flat (Figure 2.3(a)) and h is only affected by a phase shift
βc ∆z cos θ1 between the two points.
√Let us now consider that two MPCs are incident from angles θ1 and θ2 (with a1 = a2 =
1/ 2 for normalization) :
1 1
P (k) = δ(k − k1 ) + δ(k − k2 ) (2.32)
2 2
and
1 jβc ∆z cos θ1 1
R(∆z) =
e + ejβc ∆z cos θ2 (2.33)
2 2
Correlation now depends on ∆z, as drawn in Figure 2.12. When the angular spread θ1 − θ2
increases, correlation drops. The corresponding interference patterns are drawn in Figure
2.13. These interference patterns are still quite regular but the spatial correlation drop with
angular spread is clear.
Let us generalize to N incident MPCs :
N
1 X jβc ∆z cos θn
R(∆z) = e (2.34)
N n=1
| h|
(a)
| h|
(b)
| h|
(c)
F IGURE 2.13 – Interference pattern of h in a local area (4λ wide), when two MPCs are incident
(a) θ1 − θ2 = 0, 5, (b) θ1 − θ2 = 1, (c) θ1 − θ2 = 1, 5
JJ00(! cv t)z)
"D)
(k)
PS(
<latexit sha1_base64="ajEY8nLJz0VZMEGWu/uof1F7b5s=">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</latexit>
<latexit sha1_base64="zryV0+cNWuTxXXUuynOR6HrMvg8=">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</latexit>
0
!v
-1/
<latexit sha1_base64="UvDTG228Tb6DXeirQaA/9gdJlXM=">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</latexit>
0 !1/v
<latexit sha1_base64="zGcL8HSXT+Ye5EUfZTVRGdLSYsU=">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</latexit>
2 4 6 8
"kD ! vc t z
<latexit sha1_base64="JR3MWywhh00CI5vziX1FQr77BiU=">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</latexit>
<latexit sha1_base64="Yc714tuJRGx/OwKDWAxc0q4NR6A=">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</latexit>
F IGURE 2.14 – Clarke model : Power Angular Spectrum and spatial correlation function
When an infinite number of MPCs are uniformly (in θ) incident onto the local area, this
series converges to a Bessel function :
R(∆z) = J0 (βc ∆z) (2.35)
This model is very often used. It is referred to as the Clarke model (or sometimes Jakes
model). Correlation function and PAS of the Clarke model are drawn in Figure 2.14. We
will obtain the closed-form expression of P (k) drawn in Figure 2.14 later on in this chapter.
It is still worth noting that the PAS is not the most natural way of describing the angular
dispersion, since it depends on cos θ and not on the azimuth θ itself. In the literature, a
Power Azimuth Spectrum P (θ) is then defined as the incident power density (per-unit θ)
as a function of azimuth θ.
The Azimuth spread 2 σθ characterizes the angular extension of P (θ) :

s
∞
1
Z
σθ = θ2 P (θ) dθ − θm
2 (2.36)
P −∞
where P is the total power :

Z ∞
P = P (θ) dθ
−∞
2. Unfortunately, in the literature, both σu and σθ are often referred to as the angular spread. Definitions
must be carefully checked in every text.
2.4 Time-variant channels 35
and θm the mean value of θ :

∞
1
Z
θm = θ P (θ) dθ
P −∞
In the Clarke model, the correlation function (2.35) is zero when β∆z ≃ 2, 4. The
distance ∆zc between two uncorrelated points is then
∆zc ≃ 0, 38 λ
It is referred to as the channel coherence distance. Clarke model gives good results when
the number of IO’s around the receiver local area is very high, as is the case for mobiles
in urban areas or indoor. In other situations, correlation functions must be calculated by
(2.27), knowing the PAS. A tridimensional model equivalent to the Clarke model gives the
coherence distance
λ
∆zc = (2.37)
2
which is easier to remind.
2.4 Time-variant channels
2.4.1 Stochastic model
In a given local area, time-dependence of the transfer function h is deduced from (2.2) :
N
⃗
X
h(⃗r , t) = an ejφn (t) e−j βcn ·⃗r(t) (2.38)
n=1
It depends on time because (i) the receiver is moving in the local area (⃗r = ⃗r(t)) ; and/or
(ii) the IOs are moving (φn = φn (t)). Amplitudes an are assumed constant during the
observation time.
Once again, it is impossible to predict in a deterministic way these phase variations and
a stochastic model must be used. We will assume that the phases are random variables over
time, with uniform distribution between 0 et 2π.
Each draw of these phases gives one possible value of h over time, i.e. each draw gives
one time realization of h. Mean values over these realizations (i.e. time averages) will be
denoted E[ ].
2.4.2 Receiver moving - IOs fixed
Distribution
In this scenario, the receiver is moving through the interference pattern we modeled in
section 2.2. The statistical distributions are so identical, Rayleigh or Rice, over time now.
Mean values
All conclusions drawn for space characterization stay valid over time.
Time correlation and Doppler spread
In this scenario, (2.38) reads

N
⃗
X
h(⃗r , t) = an ejφn e−j βcn ·⃗r(t) (2.39)
n=1
We suppose that the receiver is moving at constant speed v :
⃗r (t) = ⃗r0 + v t ⃗1z (2.40)
where ⃗r0 is the position at an arbitrary initial observation time, so that

N
⃗
X
h(t) = an ejφn e−j βcn ·⃗r0 ejβc v t cos θn (2.41)
n=1
⃗
But in the stochastic approach the phase term e−j βcn ·⃗r0 is random and a stochastic is dedu-
ced :
XN
h(t) = an ejΦn ej2πkn v t (2.42)
n=1
By defining fDn = kn v :
N
X
h(t) = an ejΦn ej2πfDn t (2.43)
n=1
and each MPC is seen by the receiver with a frequency shifted by fDn : the usual Doppler
effect. If the transmitter sends a signal at frequency fc , the receiver will get N signals (the
N MPCs) with N different frequencies shifted by the Doppler effect and depending on the
incidence angle.
Similarly to the spatial case, the Doppler distribution seen by the mobile is defined as a
function of the Doppler shift fD as
N
X
a(fD ) = an ejΦn δ(fD − fDn ) (2.44)
n=1
so that Z ∞
h(t) = a(fD ) ej2πfD t dfD (2.45)
−∞
which is the inverse Fourier transform of a(fD ). By inverting :

Z ∞
a(fD ) = h(t) e−j2πfD t dt (2.46)
−∞
This emphasizes a second wireless channel duality :
Channel Duality 2
Time ⇐⇒ Doppler shift
Let us calculate the time correlation between t + ∆t and t :
R(t + ∆t, t) = E [h(t + ∆t) h∗ (t)]

Z ∞ Z ∞
′ (2.47)
= E [a(fD )a∗ (fD′ )] ej2πfD (t+∆t) e−j2πfD t dfD dfD′
−∞ −∞
The MPCs with Doppler shifts fD et fD′ are coming from two directions k and k ′ . Under
US assumption in the angular domain :
E [a(fD )a∗ (fD′ )] = P (fD ) δ(fD − fD′ ) (2.48)

In this equation, P (fD ) is referred to as the Power Doppler Spectrum (PDS). It gives the
power density (per-unit fD ) of the MPCs incident on the local area as a function of their
Doppler shift.
Under the US assumption we see that time correlation only depends on ∆t (and not on
the value of t). The channel is Wide Sense Stationary w.r.t. time :
Z ∞
R(∆t) = P (fD ) ej2πfD ∆t dfD (2.49)
−∞
and the Doppler spectrum is the Fourier transform of the time correlation function :
Z ∞
P (fD ) = R(∆t) e−j2πfD ∆t d∆t (2.50)
−∞
In the literature, both P (k) and P (fD ) are usually called Doppler spectrum due to their
equivalence with the change of variable k → fD = k v.
The Doppler spread σD characterizes the extension of P (fD ) :

s
∞
1
Z
σD = fD2 P (fD ) dfD − fDm
2
P −∞
where P is, as usual, the total power :

Z ∞
P = P (fD ) dfD
−∞
and fDm the mean Doppler shift :

∞
1
Z
fDm = fD P (fD ) dfD
P −∞
A closed form expression for the PDS can be obtained in the Clarke model, if we sup-
pose that an infinite number of MPCs are arriving uniformly (w.r.t. θ) around the receiver.
If P (θ) is the Power Azimuth Spectrum of the MPCs, this assumption is
P
P (θ) = (2.51)
2π
J0J(0(!c vv t)t)
(f"
P S( DD))
<latexit sha1_base64="cywWlL23tJylskG2mRwFfOMlX9U=">AAACyXicjVHLSsNAFD2Nr1pfVZdugkWpm5KKr2VBF4KbCvYBtZQknerYvJxMxFpc+QNu9cfEP9C/8M6YglpEJyQ5c+49Z+be60Qej6VlvWaMicmp6ZnsbG5ufmFxKb+8Uo/DRLis5oZeKJqOHTOPB6wmufRYMxLM9h2PNZz+oYo3bpiIeRicyUHE2r59EfAed21JVL1a7HWOtjr5glWy9DLHQTkFBaSrGuZfcI4uQrhI4IMhgCTswUZMTwtlWIiIa2NInCDEdZzhHjnSJpTFKMMmtk/fC9q1UjagvfKMtdqlUzx6BSlNbJAmpDxBWJ1m6niinRX7m/dQe6q7DejvpF4+sRKXxP6lG2X+V6dqkejhQNfAqaZIM6o6N3VJdFfUzc0vVUlyiIhTuEtxQdjVylGfTa2Jde2qt7aOv+lMxaq9m+YmeFe3pAGXf45zHNS3S+W90u7pTqGymY46izWso0jz3EcFx6iiRt5XeMQTno0T49q4Ne4+U41MqlnFt2U8fAD+FZDP</latexit>
<latexit sha1_base64="U2zNdK3OcwrjJRIMTvDbWKJck/g=">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</latexit>
-! v
<latexit sha1_base64="ooyeKxpwDQzaH7KFKuc4rFy2q/E=">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</latexit> <latexit sha1_base64="ncuZxHO+WdZG5YHV+9SLsdQKfp8=">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</latexit>
v/ 0 !v/v 2 4 6 8
"D
<latexit sha1_base64="lliwwbx4br79QkD8j9uvRyEg99Y=">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</latexit>
<latexit sha1_base64="QuvWw37KBB8jRP3He/XGdwfSC0E=">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</latexit>
! vv t t
fD c
F IGURE 2.15 – Doppler spectrum and normalized time correlation function in the Clarke model.
The corresponding PDS P (fD ) is deduced
dθ
P (fD ) = P (θ) (2.52)
dfD
v
With fD = λ
cos θ
P0
P (fD ) = r 2 (2.53)
fD
1 − v/λ
where P0 = P/(v/λ). By (2.35), time correlation function in the Clarke model is obtained
by the change of variable ∆z = v∆t :
R(∆t) = J0 (βc v ∆t) (2.54)
which is drawn in Figure 2.15.
Coherence time
By (2.37), typical channel coherence time ∆tc is obtained by the change of variable
∆zc = v ∆tc :
1λ
∆tc = (2.55)
2v
If ∆tc is larger than the symbol (or frame) duration TS , channel can be assumed static
during each symbol (frame). We speak about slow fading. Otherwise, the fading is referred
to as fast fading.
It is important to note that the same channel could be modeled either in slow fading or
in fast fading depending on the symbol (frame) duration, or the mobile speed. Most often,
for sake of simplicity, channels are modeled in slow fading : it is supposed constant during
each symbol (frame), i.e. the quasi-static assumption of Chapter 1 holds.
2.4.3 Receiver fixed - IOs moving
Distribution
From (2.38), the N -waves stochastic model is given by

N
⃗
X
h = an e−j βcn ·⃗r ejΦn (2.56)
n=1
where Φn is a random phase over time, with uniform distribution between 0 et 2π.
By the central-limit theorem, when N is high, the real and imaginary parts of h have
gaussian distributions (over time). As we know, |h| has then a Rayleigh distribution (over
time).
In the 1 + N -waves model :

N
⃗ ⃗
X
h = a0 e−j βc0 ·⃗r ejΦ0 + an e−j βcn ·⃗r ejΦn (2.57)
n=1
and |h| has a Rice distribution over time, characterized by a (temporal) Rice factor K.
It is worth noting that one MPC could be constant over time :

N
⃗c0 ·⃗
X
−j β r jφ0
h = a0 e e + an ejΦn (t) (2.58)
n=1
and |h| still has a Rice distribution over time.

Good To Know
We obtain the important result here that Rayleigh and Rice distributions are the mathema-
tical tools dedicated to model both space and time fading.
Generally speaking, time characterization of h is independent of its spatial characterization.
We could for instance imagine a scenario where h has a Rayleigh distribution over space
and a Rice distribution over time.
Mean values
All conclusions drawn for space characterization stay valid over time.
If one MPC is constant over time, the mean value of h is given by the complex ampli-
tude of this MPC, of course.
Time correlation
The stochastic model (2.43) still holds, but the Doppler shifts fDn are now due to the IO
movements and no more to the receiver movements. Since the IO movements are unknown
and arbitrary, they are difficult to model. However, h(t) is still linked to a Doppler distri-
bution by (2.45) that has to be obtained experimentally. Time correlation is then deduced
from the Power Doppler Spectrum by (2.49).
Narrowband spatial characterization
Space ⇐⇒ Angle of Arrival (AoA)
In the narrowband stochastic model, transfer function h is modeled as a random function

in the local area. We characterized this random function by its distribution and its first two
moments
Distribution
N -waves model : |h| has a Rayleigh distribution
|h| −|h|2 /2σ2

PDF (|h|) = e
σ2
1 + N -waves model : |h| has a Rice distribution
√ !
|h| 2 2 |h| 2K
PDF (|h|) = 2 e−|h| /2σ e−K I0
σ σ
where K is the Rice factor.

a2
K[dB] = 10 log PN 0
n=1 a2n
The phase of h has a uniform distribution between 0 and 2π.

Mean values
E[h] = 0
N
X
2
a20 a2n

E |h| = +
n=1
where a0 = 0 in the N -waves model.

Spatial correlation
Spatial correlation depends on the incident Power Angular Spectrum P (k) :
Z ∞
R(∆z) = P (k) ej2πk∆z dk
−∞
where k = λ1 cos θ.
Channel coherence distance is of the order of
λ
∆zc =
2
for a receiver uniformly surrounded by IOs.
When the angular spread σk increases, the channel coherence distance ∆zc decreases.
Narrowband time characterization
Time ⇐⇒ Doppler shift
In the narrowband stochastic model, the transfer function is a random function over time,
having a Rayleigh or Rice distribution.
If the transmitter sends a signal with frequency fc , the receiver gets N MPCs each affected
by a Doppler shift fD . The power density of the MPCs are described by the Power Doppler
Spectrum P (fD ). If the IOs do not move, P (fD ) is identical to the angular spectrum P (k)
by replacing k → fD = k v, with v the mobile speed.
Time correlation
Time correlation is deduced from the Doppler spectrum :
Z ∞
R(∆t) = P (fD ) ej2πfD ∆t dfD
−∞
Channel coherence time is of the order of

λ
∆tc =
2v
for a mobile uniformly surrounded by fixed IOs.
When the Doppler spread σD increases, the channel coherence time decreases.
If TS is the symbol (or frame) duration :
∆tc > TS slow fading

∆tc < TS fast fading
In slow fading w.r.t symbols (frame), channel is assumed constant during each symbol
(frame)
Wideband channels
3.1 Deterministic wideband model

3
The time-invariant channel impulse response can be written as (1.6)
N
X
h(τ ) = an ejϕn e−j2πfc τn δ(τ − τn )
n=1
N
(3.1)
X
= an ejφn δ(τ − τn )
n=1
where φn = ϕn − 2π fc τn .
Or, equivalently, its Fourier transform, the time-invariant transfer function

Z ∞ N
X
−j2πf τ
H(f ) = h(τ ) e dτ = an ejφn e−j2πf τn (3.2)
−∞ n=1
3.1.1 1-wave model
If the receiver only detects one MPC in the delay domain :

H(f ) = a1 ejφ1 e−j2πf τ1 (3.3)
whose amplitude does not depend on baseband frequency f : the transmitted signal is
multiplied by the same channel gain across the whole frequency band (i.e. there is no
distortion due to the channel). It is said to experience flat fading.
45
46 Wideband channels
-10
-20
h [dB]
-30
!f
-40
F IGURE 3.1 – 2-waves model : amplitude of the transfer function (a1 = a2 = 1, and φ1 = φ2 =
0)
3.1.2 2-waves model
If two MPCs are detected in the delay domain :
H(f ) = a1 ejφ1 e−j2πf τ1 + a2 ejφ2 e−j2πf τ2 (3.4)
whose amplitude is
p
|H(f )| = (a1 ejφ1 e−j2πf τ1 + a2 ejφ2 e−j2πf τ2 ) (a1 e−jφ1 ej2πf τ1 + a2 e−jφ2 ej2πf τ2 )
q
= a21 + a22 + 2 a1 a2 cos(φ1 − φ2 ) cos (2πf ∆τ )
(3.5)
where ∆τ = τ2 − τ1 .
Now, the amplitude of H depends on the frequency, and on the difference of propaga-
tion delays between the two MPCs. One example of (3.5) is drawn in dB scale in Figure
3.1. It can be seen that certain frequencies will be heavily damped by the channel so that
the transmitted signal is distorted when detected by the receiver. The channel experiences
frequency selective fading.
Frequency selectivity comes from the difference in propagation delays as shown by

(3.5), i.e. from signal time dispersion during propagation. Eq (3.5) shows that the fre-
3.2 Stochastic wideband model 47
quency shift ∆f between two fadings of |H| is

1
∆f = (3.6)
∆τ
i.e. the larger the time dispersion ∆τ , the faster the frequency fading.
The channel impulse response h(τ ) (complex !) lists the detected

MPCs in the delay domain. For instance in the 2-waves model : ~
h (! )
jφ1 jφ2
h(τ ) = a1 e δ(τ − τ1 ) + a2 e δ(τ − τ2 ) (3.7)
a1
as drawn here right (in amplitude). a2
!
The effect of time dispersion can thus be studied, equivalently, in !1 !2
two domains : either in the delay domain τ by modeling the channel
by its impulse response h(τ ), or in the (baseband) frequency domain,
by modeling the channel by H(f ).
This emphasizes the third wireless channel duality :
Channel Duality 3
Baseband frequency ⇐⇒ Delay
3.2 Stochastic wideband model
The time-variant transfer function is, from (3.2),

N
X
H(f, t) = an ejφn (t) e−j2πf τn (3.8)
n=1
where, as usual, the amplitude an are assumed constant during the observation time.
If the receiver and/or the IO’s are moving, or if the channel is measured at different
places of a local area, we know from Chapter 2 that the phases φn (t) in (3.8) are random.
A stochastic wideband model can thus be deduced :
N
X
H(f ) = an ejΦn e−j2πf τn (3.9)
n=1
where Φn is a random phase with uniform distribution between 0 and 2π. Each draw of
these phases gives one possible value of H(f ), i.e. each draw gives one temporal or spatial
realization of H(f ). Mean values over these realizations (i.e. time or space averages) will
be denoted E[ ].
3.2.1 Frequency correlation and delay spread
Let us derive the frequency correlation between f et f + ∆f :
R(f + ∆f, f ) = E [H(f + ∆f ) H ∗ (f )]

Z ∞ Z ∞
′ (3.10)
= E [h(τ )h∗ (τ ′ )] e−j2π(f +∆f ) τ ej2π f τ dτ dτ ′
−∞ −∞
MPCs arriving at different delays τ and τ ′ are usually assumed to be uncorrelated :
E [h(τ )h∗ (τ ′ )] ≡ P (τ ) δ(τ − τ ′ ) (3.11)
where P (τ ) is referred to as the Power Delay Profile (PDP). It gives the power density
(per-unit τ ) incident onto a local area as a function of propagation delay τ . Assumption
(3.11) is the Uncorrelated Scattering assumption in the delay domain. By (3.10) and (3.11) :
Z ∞
R(∆f ) = P (τ ) e−j2π∆f τ dτ (3.12)
−∞
Correlation only depends on ∆f (not on f ). Channel is Wide Sense Stationary w.r.t.

frequency.
The PDP is the inverse Fourier transform of the frequency correlation function :
Z ∞
P (τ ) = R(∆f ) ej2π∆f τ d∆f (3.13)
−∞
The PDP is of outmost importance for high data rate communications. It indicates that
the signal arrives with some delay spread at the receiver. If different symbols are sent
very close to each others, they can overlap at the receiver. We speak about Inter-Symbol
Interference, ISI.
Cell στ [µs] ∆fc
Indoor 0,03 5 MHz

Urban 3 50 kHz
Countryside 5 30 kHz
TABLE 3.1 – Typical delay spreads and coherence bandwidths for different cell sizes.
The absolute value of propagation delay is of no importance. By convention, the pro-

pagation delay of the first MPC is chosen as the reference τ = 0. The PDP is so defined
∀ τ ≥ 0.
The Delay spread στ characterizes the PDP duration :

s
∞
1
Z
στ = τ 2 P (τ ) dτ − τm
2 (3.14)
P 0
where P is the total power :

Z ∞
P = P (τ ) dτ (3.15)
0
and τm the mean delay :

∞
1
Z
τm = τ P (τ ) dτ (3.16)
P 0
Table 3.1 gives typical values for στ . The larger the communication range, the larger the
delay spread since the MPCs follow propagation paths with important length differences.
Experiments show that the PDP can be very often modeled by an exponential :
P (τ ) = P0 e−τ /στ τ ≥0 (3.17)
Frequency correlation function can be deduced from (3.12) :

Z ∞ Z ∞
−j2π∆f τ P0 στ
R(∆f ) = P (τ ) e dτ = P0 e−τ /στ e−j2π∆f τ dτ =
−∞ 0 1 + jστ 2π ∆f
(3.18)
Those PDP and frequency correlation function are drawn in Figure 3.2.
1 1
0,8 0,8
|R( Df )|
0,6 0,6
PDP
0,4 0,4
0,2 0,2
0 st 2st 3st 4st 0 1/2pst 2/2pst 3/2pst 4/2pst

t Df
F IGURE 3.2 – Power Delay Profile et frequency correlation function amplitude in the exponential
model (normalized).
3.2.2 Coherence bandwidth
The channel coherence bandwidth ∆fc is defined as the bandwidth over which corre-
lation is above a given threshold (typically 0.7). As shown in the exponential model, ∆fc
is inversely proportional to the delay spread. The precise relation between both parameters
depends on the PDP shape, but typically :
1
∆fc ≃ (3.19)
2π στ
If the signal bandwidth B is smaller than ∆fc , all signal frequencies will experience about
the same channel gain H(f ). We are in flat fading. Otherwise, the channel must be modeled
taking into account frequency selective fading.
This property can be equivalently seen in the delay domain : if the symbol duration TS
is larger than the channel delay spread στ , there is no inter-symbol interferences. We are
in flat fading. Otherwise, symbols overlap at the receiver : we are in frequency selective
fading.
Frequency characterization
Baseband frequency ⇐⇒ Delay
Channel frequency dependence is due to the differences in propagation delays of the MPCs,
i.e. to the channel time dispersion. It can be studied either in the frequency domain via
H(f ) or in the delay domain via h(τ ), which, in the time-invariant or quasi-static approxi-
mations, is Z ∞
h(τ ) = H(f ) ej 2πf τ df
−∞
Frequency correlation
Channel frequency correlation is deduced from the Power Delay Profile P (τ ) :
Z ∞
R(∆f ) = P (τ ) e−j2π∆f τ dτ
−∞
When the Delay spread στ increases, the channel Coherence bandwidth ∆fc decreases.
Typically
1
∆fc ≃
2π στ
where στ is the delay spread of the PDP.
If B is the signal bandwidth :
∆fc > B flat fading

∆fc < B frequency selective fading
Equivalently, if TS is the symbol duration :
στ < T S flat fading

στ > T S frequency selective fading
In case of flat fading, there is no inter-symbol interference (ISI)

3.3 Tapped Delay Line models
We saw in Chapter 1 that the receiver cannot discriminate all MPCs : if B is the signal
bandwidth, all waves arriving during a delay tap of duration ∆τ = 1/2B sum up at the
receiver. In slow fading, the channel impulse response can be written as a Tapped Delay
Line model (1.22) :
L
X
hTDL (τ, t) = hl (t) δ(τ − l∆τ ) (3.20)
l=0
where
N
X
hl (t) = αn (t) sinc 2B(τn − l∆τ ) (3.21)
n=1
or under the uncorrelated scattering assumption :

X X
hl (t) ≃ αn (t) = an (t) ejϕn (t) e−j2πfc τn (3.22)
τn ∈ tap l τn ∈ tap l
The channel impulse response is so made of L taps, where each tap includes several MPCs
that the receiver cannot discriminate. The propagation delay of the first tap τ1 is conventio-
nally chosen as τ1 = 0.
Since each tap is made of several waves that sum up at the receiver, it experiences
fading. Each tap must so be modeled by a narrowband model : Rayleigh or Rice fading,
random phase with uniform distribution, and specific Doppler spectrum. A typical example
of tap spatial fading is drawn in Figure 3.3. The Power Delay Profile, PDP, is defined as
the average power of each tap :
L
X
E |hl |2 δ(τ − l∆τ )

P (τ ) = (3.23)
l=0
For standardization purposes, wideband models are given in terms of an impulse res-
ponse model which is typical to the communication scenario : PDP, tap fading (Rayleigh
or Rice), Doppler spectrum of each tap.
~ ~
h (! ) h (! )
! !
~ ~ zone locale
h (! ) h (! )
! !
F IGURE 3.3 – Typical spatial fading of channel impulse response
Example
Channel models for 3G communication have been standardized by ITU. For communica-
tion to a pedestrian in a urban environment, the channel impulse response includes 6 taps :
Tap τn [µs] P (τn ) [dB] Fading Doppler spectrum
1 0 0 Rayleigh Clarke
2 0,2 -0,9 Rayleigh Clarke
Experiments show that the MPCs propagate as bundles in the delay and angular do-
mains : the receiver sees different packets of MPCs arriving around the same delay and
coming from about the same direction. Those packets are referred to as clusters. Clus-
ters are intuitive on physical grounds : each IO has a complex shape that reflects waves
with a slight delay and angular dispersion, so producing packets of MPCs. Models that
incorporate clusters are referred to as Cluster Delay Line models : each incident cluster is
characterized by its mean delay, delay spread, mean angle of incidence and angular spread.
In each cluster, the MPCs are then modeled by a TDL model.
F IGURE 3.4 – 3D model of a conference room used for ray-tracing. Both TX and RX are located
on a table. (A. Maltsev. et al., doc. : IEEE 8021.11-09/0334r8, 2010)
3.4 Example : channel models for 60GHz WLAN systems
Specific channel models are required for 60GHz systems due to the small wavelength
compared to classical indoor communication systems. One standardized channel model has
been developed in the framework of the IEEE 802.11ad group. We will give here the key
ingredients of this model, in the SISO case, and neglecting the polarization aspects.
3.4.1 Principal rays calculation
As we know, 60GHz MPCs suffer from very high losses due to reflection and diffrac-
tion. By using ray-tracing, single and double reflected rays are enough in order to provide
accurate results w.r.t propagation delays and incidence angle of the MPCs onto the receiver.
So, the first step in generating a 60GHz channel impulse response is to place a trans-
mitter and a receiver in a given indoor environment, and to calculate the LOS, single and
double reflected rays by classical ray-tracing approach. A set of realizations can be ob-
tained by moving the transmitter and/or the receiver through the environment. A typical
example is shown in Figure 3.4 for a conference room.
Due to people and furniture, some of these rays will, however, be blocked. Channel mo-
3.4 Example : channel models for 60GHz WLAN systems 55
dels include the probability for each ray to be blocked. This probability was derived from
ray-tracing simulations and random-walk models for people moving through the environ-
ment. This probability is highly dependent on the communication scenario. For instance, a
ray reflected off the ceiling has a quasi null probability of being blocked. On the contrary,
double reflected rays off the walls are more subject to blockage due to longer propagation
paths.
Example
Let us consider the conference room scenario in Figure 3.4. The probability for the rays to
be blocked depends on the chosen communication scenario, LOS or NLOS, that must be
fixed a priori. It has been found that :
— Each single reflected ray on the ceiling has a null probability of being blocked
— Single reflected rays off the walls have equal probability to be blocked (0.24) but
simultaneous blockage of several rays has a null probability.
— The same holds for double reflected rays off wall and ceiling with blockage proba-
bility equal to 0.037.
— Several second order reflected rays off the walls can be blocked simultaneously.
The number k of simultaneously blocked rays is given by a binomial distribution,
with the total number of rays n = 8, and the individual probability of being blocked
p = 0.175 : !
n k
f (k; n; p) = p (1 − p)n−k (3.24)
k
3.4.2 Intra-clusters parameters
If the IO’s, the walls and ceiling were perfect reflectors, with no roughness, no hetero-
geneity, and if no other small reflectors (small objects) were present in the environment,
the rays given by ray-tracing should be the only MPCs propagating from TX to RX. In
real-world environments, it is rather found that rays arrive at the receiver as bundles refer-
red to as clusters. Each cluster is made of several rays grouped in the delay and the angular
domain. The central characteristics of these clusters (arrival time τi and incidence angle of
the central ray) are well predicted by ray-tracing as described here above, but the spread of
the rays inside each cluster can only be derived by experiments.
A cluster structure in the delay domain is shown in Figure 3.5. The cluster i consists of
(i)
a central ray with fixed amplitude α0 (given by ray-tracing) and a set of pre-cursors with
F IGURE 3.5 – Delay domain model of a cluster (A. Maltsev. et al., doc. : IEEE 8021.11-
09/0334r8, 2010)
(i) (i) (i) (i)

amplitudes α−1 ..α−Nf and post-cursors α1 ..αNb . The number of pre-cursors and post-
cursors depends on the scenario.
The arrival time of the central ray is chosen as the reference τ = 0. The mean amplitude
(i) (i)
of the pre-cursors Af (τ ) and post-cursors Ab (τ ) decays exponentially :
(i) (i)
Af (τ ) = Af (0) e|τ |/γf
(i) (i)
(3.25)
Ab (τ ) = Ab (0) e−τ /γb
(i)
where the decay rates γf and γb are experimentally found. The peak amplitudes Af (0)
(i) (i)
and Ab (0) are deduced from the amplitude of the central ray α0 by two experimental
parameters :
(i)
α0
Kf = 20 log | (i)
|
Af (0)
(i)
(3.26)
α0
Kb = 20 log | (i)
|
Ab (0)
(i)
The actual values of the pre- and post-cursors amplitudes αk obey Rayleigh distributions.
Finally, for normalization, the total power of the cluster (sum of the ray powers) is set to
one.
3.4 Example : channel models for 60GHz WLAN systems 57
Parameter Value
Kf 10 dB
γf 3.7 ns
λf 0.37 ns−1
Nf 6
Kb 14.2 dB
γb 4.5 ns
λb 0.31 ns−1
Nb 8
TABLE 3.2 – Clusters parameters for the conference room channel model (A. Maltsev. et al.,
doc. : IEEE 8021.11-09/0334r8, 2010)
(i)
The propagation delays τk around the central ray are deduced from the fact that the
pre-cursors and post-cursors are modeled as two Poisson processes with arrival rates λf
and λb (also depending on the scenario).
We finally get the mathematical expression of the normalized impulse response for the
cluster i :
X (i) (i)
c̃(i) (τ ) = αk δ(τ − τk ) (3.27)
k
(i)
where the propagation delay of the central ray τ0 is zero.
In the angular domain, the pre- and post-cursors are spread around the central ray whose
incidence angles are deduced from ray-tracing. It is experimentally found that the azimuth
and elevation angles of each ray can be drawn from a normal distribution with zero mean
and 5 degrees std around the central ray angles.
Table 3.2 summarizes the experimental parameters for the scenario in Figure 3.5.
3.4.3 Cluster amplitudes
With the procedure described above, we have now a set of clusters whose total powers
are normalized, and whose central propagation delay is zero. The last step in the modeling
process is to affect an amplitude to each cluster, and the correct propagation delay .
LOS ray
The first "cluster" is the LOS path, which is modeled by a single ray with amplitude
given by the Friis formula with the antenna gains excluded :
λ
A0 = (3.28)
4π d
NLOS clusters
The amplitude of each cluster i is also deduced from ray-tracing :

λ
Ai = g (i) (3.29)
4π d
where d is total propagation distance of cluster i, and g (i) a reflection loss given by ray-
tracing (typical value around -10dB for single reflected rays and -16dB for double reflected
rays).
3.4.4 Channel impulse response
We finally get the global (unnormalized) impulse response of the 60GHz channel, by
summing the cluster impulse responses shifted in the delay domain :
X
h(τ ) = Ai c̃(i) (τ − τi ) (3.30)
i
where τi is the propagation delay of the central ray of cluster i (computed by ray-tracing).
Multi-dimensional chan-
nel analysis
4.1 Global duality

4
In the two previous chapters, we saw that channel models depend on minimum six
parameters : t, τ , f , z, fD , k, and are described in terms of many different functions. Let
us review the notations we used so far :
Symbol Description
f Baseband frequency.
z Position in a local area.
t Usual clock time.
τ Propagation delay.
k Wave vector projection on the z axis. Proxy of the angle of arrival.
fD Doppler shift.
h(.) Used when the baseband frequency is not an argument of the channel
function, either in narrowband models (h(t), h(z)..) or in wideband
models in the delay domain (h(τ, t),..).
H(.) Used when the baseband frequency is an argument of the channel
function (H(f, t),..).
a(.) Amplitude distribution of the MPCs.
P (.) Power density spectrum of the MPCs.
R(.) Channel auto-correlation.
59
60 Multi-dimensional channel analysis
The parameters can actually be split into two parts :
— (f, z, t) span the observation space : the channel is always measured/modeled at

some baseband frequency f , point z, and time t. Any measurement/model of the
channel can so be located at some coordinates (f, z, t) in a space referred to as the
observation space.
— (τ, k, fD ) span the MPC spectrum space : every MPC arrives at the receiver with
some delay τ , direction of arrival k, and frequency shift fD . Every MPC can thus be
located at coordinates (τ, k, fD ) in a space referred to as the incident MPC "spec-
trum" space.
We learned from the two previous chapters that these two spaces are dual : we can go
back and forth between the two spaces by applying Fourier transforms.
Duality of the two parameter spaces
Observation space MPC spectrum space
f τ
z ⇐⇒ k
t fD
The objective of a global multi-dimensional approach is to formalize this duality.
4.2 Scattering functions
The distribution of the MPC amplitudes in the spectrum space is given by the glo-
bal scattering function a(τ, k, fD ). By using a discrete model for this distribution, as in
Chapters 2 and 3 :
X
a(τ, k, fD ) = ατn ,kn ,fDn δ(τ − τn ) δ(k − kn ) δ(fD − fDn ) (4.1)
n
where ατn ,kn ,fDn is the complex amplitude of MPC n located at (τn , kn , fDn ) in the spec-
trum space. Two MPCs located at two different coordinates (τn , kn , fDn ) have experienced
4.2 Scattering functions 61
different propagation processes. It is thus reasonable to assume that they are not correlated,
leading to the Uncorrelated Scattering (US) assumption :
E [ a(τ, k, fD )a∗ (τ ′ , k ′ , fD′ )] = P (τ, k, fD ) δ(τ − τ ′ ) δ(k − k ′ ) δ(fD − fD′ ) (4.2)
where P (τ, k, fD ) is the global power density of the MPCs in the spectrum space.
We already saw scattering functions and power densities in the previous chapters, but in
uni-dimensional cases. To go from the global functions to uni-dimensional ones, we have
to integrate over the variables to eliminate. For instance, keeping the τ variable :
ZZ X
a(τ ) = a(τ, k, fD ) dk dfD = ατn δ(τ − τn ) (4.3)
n
where X
ατn = ατm ,km ,fDm (4.4)
τm =τn
which can be easily understood : a(τ ) is the amplitude of the MPC arriving with delay τ ,
i.e. the sum (integral) of all MPCs arriving with delay τ whatever their incidence direction
k and frequency shift fD .
It is worth noting that a(τ ) is just the channel impulse response h(τ ) defined in Chapter 1 :
X
h(τ ) = a(τ ) = ατn δ(τ − τn ) (4.5)
n
Similarly for the two other variables :

ZZ X
a(k) = a(τ, k, fD ) dτ dfD = αkn δ(k − kn )
n
ZZ X (4.6)
a(fD ) = a(τ, k, fD ) dτ dk = αfDn δ(fD − fDn )
n
where
X
αkn = ατm ,km ,fDm
km =kn
X (4.7)
αfDn = ατm ,km ,fDm
fDm =fDn
These uni-dimensional scattering functions were already defined in (2.22) and (2.44).
The uni-dimensional power densities (2.26), (2.48) and (3.11) can also be deduced by
integration :
ZZ
P (τ ) = P (τ, k, fD ) dk dfD
ZZ
P (k) = P (τ, k, fD ) dτ dfD (4.8)
ZZ
P (fD ) = P (τ, k, fD ) dk dτ
i.e. the Power Delay Profile, Power Angular Spectrum, and Power Doppler Spectrum, res-
pectively.
And, finally, the total power of the MPCs :

ZZZ
P = P (τ, k, fD ) dτ dk dfD (4.9)
4.3 Channel functions
The scattering functions describe how the MPCs are distributed in the spectrum space.
Knowing this distribution, it is then possible to deduce channel models in the observation
space by applying Fourier. Summarizing what we obtained in the uni-dimensional cases :
Observation space Spectrum space
a(τ ) e−j 2πf τ dτ

R R
H(f ) = a(τ ) = h(τ ) = H(f ) ej 2πf τ df
h(z) e−j2πkz dz
R R
h(z) = a(k) ej2πkz dk ⇐⇒ a(k) =
h(t) e−j2πfD t dt
R R
h(t) = a(fD ) ej2πfD t dfD a(fD ) =
where we notice the (unfortunate) asymmetry in the Fourier transforms for the three va-
riables.
4.3 Channel functions 63
A global channel transfer function in the observation space can be deduced, and its dual
relationship :
ZZZ
H(f, z, t) = a(τ, k, fD ) e−j 2πf τ ej2πkz ej2πfD t dτ dk dfD
~
(4.10)
ZZZ
a(τ, k, fD ) = H(f, z, t) ej 2πf τ e−j2πkz e−j2πfD t df dz dt
With (4.1) and (4.10) :

X
H(f, z, t) = ατn ,kn ,fDn e−j 2πf τn ej2πkn z ej2πfDn t (4.11)
n
But duality has not to be always applied on all variables, and channel models can be
built at will, as shown on two examples here below.
Example 4.1
Let’s build a time-variant channel impulse model h(τ, t).

Since the dual of t is fD , we start by writing the scattering function
Z X
a(τ, fD ) = a(τ, k, fD ) dk = ατn ,fDn δ(τ − τn ) δ(fD − fDn ) (4.12)
n
We then apply duality on the fD variable to get

X
h(τ, t) = ατn ,fDn ej2πfDn t δ(τ − τn ) (4.13)
n
Example 4.2
A receiver wants to measure the propagation delays and Doppler shifts of the incident
MPCs. How could it perform this task ?
We note first that the dual variables of (τ, fD ) in the observation space are (f, t). The
receiver has hence to measure the time-varying transfer function H(f, t), and then compute
ZZ
a(τ, fD ) = H(f, t) ej 2πf τ e−j2πfD t df dt (4.14)
Since the propagation velocity is c, the propagation distance is cτ , and a(τ, fD ) also gives
the propagation distance of every MPC. This scattering function is then referred to as the
Range-Doppler map of the channel, as obtained by radars, for instance.
4.4 Dispersion-Selectivity duality
The global duality between the observation space and the MPC spectrum space is very
important from the physics point of view since it states that the properties of the channel in
the observation space (fading, correlation,..) are imposed by the dispersion of the MPCs in
the spectrum space.
The dispersion of the MPCs over each dimension in the spectrum space is so of ou-
termost importance. We characterized this dispersion by defining spread parameters as the
second order moments of the power densities :
Variable Spread parameter Symbol
τ delay spread στ
k angular spread σk
fD Doppler spread σD
Due to dispersion, the observation space is selective : at some points (f, z, t), the chan-
nel fades. We characterized selectivity separately over every dimension of the observation
space by defining coherence metrics :
4.4 Dispersion-Selectivity duality 65
Variable Coherence metric Symbol
f coherence bandwidth ∆fc

z coherence distance ∆zc
t coherence time ∆tc
Coherence metrics are calculated via the channel auto-correlation functions R(.). We saw
that US assumption implies that the correlation functions are not dependent on the absolute
coordinates of the observation point (f, z, t) but only on the increments (∆f, ∆z, ∆t),
through dualities with the power densities
Observation space Spectrum space
P (τ ) e−j2π∆f τ dτ
R R
R(∆f ) = P (τ ) = R(∆f ) ej2π∆f τ d∆f
R(∆z) e−j2πk∆z d∆z

R R
R(∆z) = P (k) ej2πk∆z dk ⇐⇒ P (k) =
R(∆t) e−j2πfD ∆t d∆t

R R
R(∆t) = P (fD ) ej2πfD ∆t dfD P (fD ) =
These dualities are applications of the more general Wiener-Khinchin theorem.
The duality through Fourier also implies the general rule : the wider the spread, the
narrower the coherence, and, we can remind as practical rules :
στ · ∆fc ≃ 1
σk · ∆zc ≃ 1 (4.15)
σD · ∆tc ≃ 1
MIMO channels
5.1 Definition
5
As depicted in Figure 5.1, wireless communication systems can be sorted according to :
— Several transmit antennas, single receive antenna (Multiple-Input Single-Output,

MISO systems)
— Single transmit antenna, several receive antennas (Single-Input Multiple-Output,
SIMO systems)
— Several transmit and receive antennas (Multiple-Input Multiple-Output, MIMO sys-
tems)
Multi-antenna systems are used for many years in applications like astronomy, radars
or satellite communications. Their use in wireless communication is rather new but it is
now becoming the standard in many systems. MIMO systems allow major improvements
in terms of capacity or reliability. They also pave the way to new communication paradigms
like multi-user strategies, or massive antenna arrays technologies (massive MIMO).
To optimize communication, different signals are applied to the different antennas at

the transmitter. At the receiver, received signals are combined. Different signal processing
techniques can be applied as shown in Table 5.1. These techniques cannot be applied si-
multaneously, and they should be chosen according to the targeted improvement.
In this chapter, we will study spatial multiplexing. Beamforming was studied in the
course "Communication channels" while spatial diversity is essentially a coding problem.
67
68 MIMO channels
1 1 1 1
2 2 2 2
TX RX TX RX TX RX
...
...
...
...
NT NR NT NR
MISO NT x 1 SIMO 1 x NR MIMO NT x NR
F IGURE 5.1 – Multi-antenna systems
Technique Effect
Beamforming Increased transmitted power in some directions.

Increased received power from some directions.
Interference mitigation (null steering)
Spatial Multiplexing Increased capacity
Spatial Diversity Fading mitigation
TABLE 5.1 – Multi-antenna processing techniques
Example 5.1
Multi-antenna techniques are part of all the recent wireless network standards : Wi-Fi
(IEEE 802.11n), 5G, LTE,..
5.2 Channel matrix
We defined in Chapter 2 the SISO channel models. These models can be straightfully
generalized to the MIMO case. In the narrowband case :
y = Hx + n (5.1)
5.2 Channel matrix 69
where x is the vector of symbols sent to the NT transmit antennas :

 
x1
x 
 2 
x = 
.. 
 (5.2)
. 
xN T
y is the vector of symbols received at the NR receive antennas :

 
y1
y 
 2 
y = 
.. 
 (5.3)
. 
yNR
n is the vector of noise, and H the MIMO channel matrix :

 
h11 h12 ... h1NT
 h21 h22 ... h2NT 
H =  (5.4)
 
.. .. .. .. 
 . . . . 
hNR 1 hNR 2 . . . hNR NT
Each sub-channel hij must be modeled as explained in Chapter 2. In case of Rayleigh

fading :
hij = Aij ej Φij
where Aij is a Rayleigh faded amplitude , and Φij a uniformly distributed random phase,
i.e.
H = Hw (5.5)
where Hw is matrix whose elements are CN (0, 1).
For each sub-channel hij , correlation between realizations is fixed by an angular spec-
trum, and each sub-channel is normalized according to
E |hij |2 = 1

(5.6)
which supposes that the mean power of each sub-channel is the same.
70 MIMO channels
In case of Rice fading

 
r ejΦ011 . . . ejΦ01NT
K  .. .. ..
H=

. . .
K +1
 
jΦ0NR 1 jΦ0NR NT
e ... e
  (5.7)
r A11 ejΦ11 ... A1NT ejΦ1NT
1  .. .. ..
+

. . .
K +1
 
jΦNR 1 jΦNR NT
ANR 1 e . . . ANR NT e
where it is assumed that the Rice factor K is the same for all sub-channels. This last equa-
tion can be written as
r r
K 1
H = Hm + Hw (5.8)
K +1 K +1
where
— Hm is the coherent part : E [H] = Hm
— Hw is a Rayleigh matrix made of independent CN (0, 1) elements
5.3 Spatial multiplexing
Multi-antenna systems allow simultaneous transfer of several independent information

streams, so drastically increasing the system capacity. This technique is referred to as spa-
tial multiplexing
5.3.1 Principle
Let H be a channel matrix NR × NT . As any complex matrix, it can be written in terms

of a singular value decomposition (SVD) :
H = U Σ VH (5.9)
where .H denotes the conjugate transposed. Matrices U (NR × NR ) and V (NT × NT ) are
unitary matrices :
UH U = I
(5.10)
VH V = I
5.3 Spatial multiplexing 71
and matrix Σ (NR × NT ) has on its "diagonal" 1 the singular values of H and null elements
elsewhere. If RH is the rank of H, this latter has RH singular values σi :
p
σi = λi i = 1..RH (5.11)
where λi are the eigenvalues of HHH .
Some algebra shows that every σi is real and positive (since HHH is hermitic)
With (5.1) :
y = Hx + n (5.12)
Let us suppose that we want to send the vector of symbols x̃ and let us choose at the
transmitter
x = V x̃ (5.13)
and compute at the receiver
ỹ = UH y (5.14)
By (5.12) and (5.9) :
ỹ = UH (Hx + n)
= UH UΣVH x + n

(5.15)
= UH UΣVH Vx̃ + n

= UH UΣVH Vx̃ + UH n
Finally, with (5.10)

ỹ = Σx̃ + ñ (5.16)
where ñ = UH n.
Since Σ is diagonal, it is possible to simultaneously transmit RH independent symbol
streams by applying (5.13) and (5.14). The MIMO channel is so converted into RH inde-
pendent SISO channels with gains σi
ỹi = σi x̃i + ñi i = 1..RH (5.17)
The importance of the rank of H is clear : the number of independent symbol streams is
equal to the number of singular values, i.e. to the rank of H. As for any matrix NR × NT :
RH ≤ min (NR , NT ) (5.18)
and minimum two antennas at the transmitter and at the receiver are required to apply
spatial multiplexing.
1. Σ is not square if NR ̸= NT
72 MIMO channels
Since H is random, so are U, V and also the singular values that can be different for
each realization of the channel.
It is also worth noting that in order to apply (5.16) it is required to know the channel
matrix, which is not trivial.
Finally, since σi are the equivalent SISO channel gains, a singular value too close to
zero will give a bad symbol stream (bad SNR).
Example 5.2
Let us consider a 3×3 MIMO system, and one realization of its channel matrix :
 
0, 1 0, 3 0, 7
H =  0, 5 0, 4 0, 1 
0, 2 0, 6 0, 8
Numerically, we get
   
−0, 55 0, 37 −0, 74 −0, 28 −0, 77 −0, 57
U =  −0, 33 −0, 91 −0, 21  V =  −0, 56 −0, 34 0, 74 
−0, 76 0, 12 0, 63 −0, 77 0, 53 −0, 34
 
1, 33 0 0
Σ =  0 0, 51 0 
0 0 0, 09
The rank of H is 3 (3 non zero singular values), but since the smallest singular value is
close to zero, only two symbol streams could be sent in parallel.
5.3.2 Channel degrees-of-freedom
Let us consider a 2×2 MIMO system using omnidirectional antennas (in the plane of
Figure 5.2(a)). Antenna arrays are in the far field of each others (distance R). The distance
between antennas in each array is d, with d ≪ R.
(a)
(b)
(c)
F IGURE 5.2 – Number of degrees-of-freedom of MIMO channels (a) free space (b) multipath
environment (c) keyhole
74 MIMO channels
2
s1
1.5
s2
0.5
0
0 0.2 0.4 0.6 0.8 1
a
F IGURE 5.3 – Singular values of a free space 2×2 MIMO system
Free space channels
Let us consider first that the system is in free

√ space as in Figure 5.2(a). The channel
matrix is then purely deterministic. Let R1 = R2 + d2 . For the amplitude terms in H, it
can be assumed that R1 ∼ R, but not for the phase terms. After normalization :
−jβR
e−jβR1

e
H = (5.19)
e−jβR1 e−jβR
It can be easily computed that the eigenvalues of HHH are
λ1 = 2 (1 + cos α)
(5.20)
λ2 = 2 (1 − cos α)
where α = β(R1 − R).
Corresponding singular values are drawn in Figure 5.3. It can be seen that when α → 0
(i.e. d → 0), σ2 → 0 : the channel matrix is singular. In this case sub-channels collapse :
there is only one propagation path between the transmitter and the receiver. Spatial multi-
plexing is impossible.
As seen in Figure 5.3, σ2 is not negligible if α ∼ 1, corresponding to

r
1 R
d/λ = (5.21)
2 λ
This distance is quite high, as shown in the next example.
Example 5.3
Let a 5G base station @ 3,6 GHz communicates to a user located at distance R = 500m.
We suppose that LOS is very strong, so that free space approximation holds. By (5.21), to
use spatial multiplexing, the distance d between antennas should be
d = 3m
Rayleigh channels
Let us now consider the other extreme case where there is no LOS and a lot of IO’s
as drawn in Figure 5.2(b). MPCs are numerous, and the channel matrix should be built
according to (5.5) where fading is due to the IO’s, receiver and/or transmitter movements.
For each realization, H is rank 2 whatever the distance between the antennas. For spa-
tial multiplexing, two sub-channels exist whatever this distance ! Thanks to multipaths, an
infinite number of propagation paths exist. This will be re-examined later on by taking into
account correlation between sub-channels.
Keyhole channels
Let us now consider a third case, where an obstacle with a single aperture is placed bet-
ween the transmitter and the receiver (Figure 5.2(c)). All waves should go through this aper-
ture to reach the receiver. The matrix H can thus be split into propagation transmitter→aperture,
and propagation aperture→receiver :
h1 h′1 h1 h′2

h1 ′ ′

H = h1 h2 = (5.22)
h2 h2 h′1 h2 h′2
which is always rank 1 ! Due to the obstacle, all waves have part of their propagation
in common, and spatial multiplexing is impossible. This case is referred to as a keyhole
channel.
Our discussion illustrates what is called the number of degrees-of-freedom of a MIMO

channel. It gives the maximal spatial multiplexing level that can be reached for a given sys-
tem, in a given environment. It is equal to the rank of the H matrix. It depends on the
number of antennas, but also on the number of independent propagation paths in the envi-
ronment.
76 MIMO channels
In real-world, propagation paths are complicated to identify. But it can be claimed that
spatial multiplexing works better if the environment is rich in multipaths (indoor,..).
5.4 Sub-channel correlation
5.4.1 Model
Sub-channels of H are usually correlated, i.e.

E [hij h∗kl ] ̸= 0 (5.23)
This correlation was not taken into account in (5.8).
Let us consider Rayleigh fading (Hm = 0) and let us define the correlation matrix
R = E vec(H) vec(H)H

(5.24)
where the ’vec’ operator transforms a matrix into a vector by superposing the matrix co-
lumns. Let us note that R (NR NT × NR NT ) is hermitic and have 1’s on its diagonal.
Example 5.4
For a 2×2 MIMO system : 

h11
 h21 
vec(H) =  
 h12  (5.25)
h22
and  
1 r1 t1 s1
 r1∗ 1 s2 t2 
R = 
 t∗1 s∗2 1 r2 

s∗1 t∗2 r2∗ 1

where
— r1 = E[h11 h∗21 ] and r2 = E[h12 h∗22 ] are the receive correlations computed with
antennas 1 and 2 transmitting, respectively.
— t1 = E[h11 h∗12 ] and t2 = E[h21 h∗22 ] are the transmit correlations computed with
antennas 1 and 2 receiving, respectively.
— s1 = E[h11 h∗22 ] and s2 = E[h21 h∗12 ] are the cross-channel correlations
Six correlation coefficients are to be estimated, by computation or measurements.
5.4 Sub-channel correlation 77
A matrix H having the right correlation properties can be built starting from an uncor-
related matrix Hw by
vec(H) = R1/2 vec(Hw ) (5.26)
H
where R1/2 is given by Choleski decomposition : R1/2 R1/2 = R. Indeed :
 
H 1/2 H
= E R1/2 vec(Hw ) vec(Hw )H R

E vec(H) vec(H)  = R (5.27)
| {z }
=I
This method is, however, very heavy to implement since it requires to know the matrix
R of dimensions NR NT × NR NT . In practice, simplified models are often used. The most
common model is the Kronecker model. It is based on two assumptions :
1. Receive correlations are independent from the transmit antenna (r1 = r2 in example
5.4). The same holds for the transmit correlations (t1 = t2 in example 5.4).
2. Cross-channels correlations are equal to the product of the corresponding sub-
channels correlations (s1 = r1 t1 et s2 = r2 t2 in example 5.4).
It can be shown that a Rayleigh matrix H can then be built according to
T
1/2
H = R1/2
r H w Rt (5.28)
where the matrices Rr (Nr ×Nr ) and Rt (NT ×NT ) are the receive and transmit correlation
matrices, respectively.
Eq (5.28) should still be normalized to satisfy (5.6).
In case of Rice fading, by generalizing (5.8) :

r r
K 1
1/2
T
H = Hm + Rr1/2 Hw Rt (5.29)
K +1 K +1
where each term should be normalized.

78 MIMO channels
Example 5.5
For a 2×2 MIMO system, receive and transmit correlation matrices in the Kroncker model
are :
1 r 1 t
Rr = Rt =
r∗ 1 t∗ 1
where
— r = r1 = r2 is the correlation between receive antennas .
— t = t1 = t2 is the correlation between transmit antennas.
Only two correlation coefficients are to be estimated (as compared to 6 in the full model)
Example 5.6
Standard IEEE 802.11n (MIMO WiFi) proposes to use the Kronecker model for channel
modeling in indoor environment.
5.4.2 Correlation effects
To understand the effects of correlation, let us consider a 2 × 2 MIMO system, with

Rayleigh fading. Without correlation :

h11 h12
H = Hw = (5.30)
h21 h22
whose rank is 2. If the correlation between antennas at the transmitter and the receiver are
zero :
1/2 1/2 1 0
Rr = Rr = Rt = Rt = (5.31)
0 1
and, by (5.28) :
T
1/2
H = Rr1/2 Hw Rt = Hw (5.32)
i.e. without correlation, the rank of matrix H is maximal.
In case of full correlation :

1 1
Rr = Rt =
1 1
(5.33)
1/2 1/2 1 0
Rr = Rt =
1 0
5.4 Sub-channel correlation 79
and the channel matrix becomes
T

1/2 h11 h11
H = R1/2
r Hw Rt = (5.34)
h11 h11
whose rank is always 1 !
Correlation has thus an impact on the channel matrix rank, i.e. the number of degrees-
of-freedom of the channel. A multipath environment is not a sufficient condition to have
high rank channels, antennas should also be sufficiently far from each others so that corre-
lation is low.
How to compute matrices Rr and Rt ?
Let us consider, for instance, the case of a mobile moving into a static channel. We saw
in section 2.2 that spatial correlation can be deduced from the incident angular spectrum.
Correlations coefficients can be computed by (2.31), the angular spectrum being correctly
chosen (theoretically, numerically or experimentally). Transmit correlations are computed
by using reciprocity by inverting the role of the transmitter and the receiver and using
(2.31).
If N MPCs are incident, by (2.38) :
N
1 X jβ d cos θn
r = e (5.35)
N n=1
We see here the huge importance of the angular spectrum for MIMO systems : it defines
correlation between transmit and receive antennas, i.e. the spatial multiplexing capabilities
of the system. A broad angular spectrum is needed for a low inter-antenna correlation.
It can be shown that this conclusion also holds for a static antenna array communicating
into a time varying channel.
80 MIMO channels
propagation path n
<latexit sha1_base64="WZkJkwniH2UOrPAg+duLZogLyL4=">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</latexit>
~ TX ~ RX
cn
cn
<latexit sha1_base64="rhAmory2CQfQ391nlRDn5SyfvK8=">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</latexit>
<latexit sha1_base64="imA0GSlBp1oFByCEadKXnd2VZn4=">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</latexit>
F IGURE 5.4 – Propagation path for double-directional model.
Example 5.7
Let us consider a 5G basestation @ 3,6GHz located on a rooftop to give broadband access

to users inside a building.
Inside the building, a lot of MPCs propagate and arrive all around the user. Clarke model is
thus suitable and correlation is low for antenna separation of the order of λ/2 i.e. ∼ 4 cm
at the user terminal.
On the rooftop, all MPCs come from the same direction (the user direction) since there is
no IO surrounding the basestation. Angular spread is low and correlation is low only if the
inter-antenna distance is equal to several wavelengths, at the basestation.
5.5 Double-directional models
Instead of defining a MIMO channel model purely mathematically as we did for (5.29),
it is possible to build it on physical grounds. This kind of MIMO channel models are refer-
red to as double-directional since they include both the Angles-of-Departure (AoD) and
Angles-of-Arrival (AoA) of every propagation path. Correlation between the sub-channels
is then automatically included since all physical effects are considered, especially the de-
parture and arrival power angular spectra.
By extending (2.38), a physical double-directional narrowband channel model between

antenna i and antenna j can be derived :
N
⃗ TX·⃗
r T X (t) ⃗ RX · ⃗
r RX (t)
X
hij (t) = ej β cn j an ejφn (t) e−j β cn i (5.36)
n=1
5.5 Double-directional models 81
propagation path
<latexit sha1_base64="yQ1/psSfOiN5KkEqA721O2kPsQ8=">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</latexit>
cluster m <latexit sha1_base64="OqMZPmq9j02iqhfdPhfWAP0cGOw=">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</latexit>
Nm MPC’s
<latexit sha1_base64="wMzMjA59DTDOuok3fZ+7aj2j0hQ=">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</latexit>
~ TX ~ RX
<latexit sha1_base64="JgBkvTCH9CWc4oVl8W9D7ZYcBIQ=">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</latexit>
cn,m <latexit sha1_base64="JbZoAlWCsZwam7n83v8DNe0E/oU=">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</latexit>
cn,m
F IGURE 5.5 – Propagation paths of cluster m for double-directional model.
where N is the number of propagation paths, an ejφn (t) account for all propagation effects
(between the local area center of the transmitter to the local area center of the receiver) and
antenna effects, ⃗r RX
i (t) is the position of antenna i in the receiver local area coordinate
TX
system, ⃗r j (t) is the position of antenna j in the transmitter local area coordinate system,
β⃗ RX ⃗ TX
cn is the wave vector of arrival of path n, β cn is the wave vector of departure of path n
(see Figure 5.4).
All the propagation-related parameters in (5.36) are deduced from measurements, or

numerically, for instance by ray-tracing.
But this is not the end of the story. As we saw in the example model 3.4, MPCs are
barely found to propagate individually. They rather propagate as bundles grouped in the
AoD and AoA domains. These bundles are referred to as clusters.
When clusters are taken into account, double-directional models are built in two steps.
First, a model similar to (5.36) is written separately for each cluster m :
Nm
⃗ TX T X (t) ⃗ RX RX (t)
X
hij,m (t) = ej β cn,m · ⃗r j an,m ejφn,m (t) e−j β cn,m · ⃗r i (5.37)
n=1
where Nm is the number of MPCs belonging to the cluster m and all other propagation
parameters are related to the propagation path n of the cluster m (see Figure 5.5).
Next, all clusters are summed to obtain the total double-directional narrowband channel
model between antenna i and antenna j :
M
X
hij (t) = hij,m (t) (5.38)
m=1
82 MIMO channels
where M is the number of clusters.
5.6 Wideband models
Wideband models should take into account time dispersion of the MPCs. Channel ma-
trix then depends on propagation delay τ . The TDL model is straightfully extended to
MIMO :
XL
H(τ ) = Hl δ(τ − τl ) (5.39)
l=1
where L is the number of taps, τl propagation delay of tap l, and Hl channel matrix for tap
l. Each matrix Hl should be built with a narrowband model (like (5.29) for Kronecker, or
(5.36)/(5.38) for physical models) and normalized to satisfy a given Power Delay Profile
P (τ ) : h i
E | (Hl )ij |2 = P (τl ) (5.40)

Syllabus

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Syllabus

Uploaded by

Copyright:

Available Formats

U NIVERSITÉ LIBRE DE B RUXELLES

É COLE POLYTECHNIQUE DE B RUXELLES

Wireless communication channels

Prof. Philippe D E D ONCKER

1 Wireless Communication Channels 3

3.2.2 Coherence bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . 50

4 Multi-dimensional channel analysis 59

1.2 Physical channel models

1.2.1 Free-space model

y(t) = a ejϕ e−j2πfc τ x(t − τ ) + n(t) (1.2)

In this model, the actual propagation delay τ is of no importance, and we write

y(t) = α x(t) + n(t) (1.3)

In wireless communications, the receiver location plays a significant role in channel

or, in more compact form

With (1.6) or (1.7), (1.5) can be written as

where f is the baseband frequency. With (1.6) :

The Fourier transform of (1.8) gives

Y (f ) = H(f ) X(f ) (1.11)

1.2.3 Time-variant channels

which is no more a convolution product !

The time-variant transfer function is defined as

but (1.11) cannot be used anymore.

1.2.4 Quasi-static approximation

In the quasi-static approximation, a time-variant channel is assumed to be time invariant

With this approximation, h(τ, t) can be intuitively understood :

In the sequel of this course, the quasi-static approximation is always assumed.

1.3 Tapped Delay Line models

1.3.1 Wideband models

As any bandlimited signal, x(t − τ ) can be written as

where sinc(x) = sin πx/πx.

We define the complex gain of the tap l as

which is mathematically equivalent to (1.15) if we define a Tapped Delay Line model

Tap gains can be computed from (1.19) and (1.14) :

F IGURE 1.3 – Physical impulse response and TDL model.

1.3.2 Narrowband models

Let us define the channel delay spread στ as 1

στ = max |τi − τj | (1.26)

and the channel coherence bandwidth ∆fc :

y(t) = h(t) x(t) (1.30)

which is referred to as a narrowband model. It is important to remind that the narrowband

The narrowband channel transfer function (1.16) reads

H(f, t) = h(t) (1.31)

1.4 Course outline

We can now define the goals of this course :

2.1 Deterministic model in a local area

A typical experimental result of a transfer function amplitude measured in a local area

F IGURE 2.1 – Transfer function amplitude. Experimental result.

2.1.1 Two-waves model

2.1.2 N -waves model

2.1.3 1+N -waves model

where β⃗c0 is the wave vector of the direct wave.

2.2 Spatial statistical characterization

2.2.1 Stochastic model in a local area

Previous deterministic method has drawbacks, however :

As a consequence, another methodology is most often considered : a stochastic model

F IGURE 2.8 – Histogram of |h|, Arg h and Rayleigh distribution fitting

|h| −|h|2 /2σ2

Rayleigh distributions have the following properties :

0.2 0.4 0.6 0.8 1.

F IGURE 2.9 – Histogram of |h| and Rice distribution fitting.

By convention, h is normalized so that

where a random complex variable Z = X + jY is CN (0, 1) if X and Y are independent

1+N -waves model

In that model, it is possible to show that |h| has a Rice distribution :