— Lecture Notes —

Wireless Communications I
CO25-300311

- Module 1 -
The Physical Wireless Channel

A Course of the Undergraduate Program

in Electrical Engineering at Jacobs University Bremen
School of Science and Engineering

Prof. Dr. Giuseppe Abreu

February 8, 2022
Lecture 1
Taking to the Air and Back to the Wire
Emission of Linear Antenna in Free Space
• A short linear antenna of length L excited with an alternate current I =
I0 ej2πf t , produces at a point at a distance r making an angle θ with the
antenna, the following electro-magnetic waves:
 
I0 L cos θ 1 c
E† = + ej2πf (t−r/c) (0.1)
2πε0 c r2 j2πf r3
c2
 
I0 L sin θ j2πf c
E= + + e−j2πf (t−r/c) (0.2)
4πε0 c2 r r2 j2πf r3
 
I0 L sin θ j2πf c
H= + ej2πf (t−r/c) (0.3)
4πε0 c2 r r2

where ε0 is the permittivity of the medium (vacuum/air).

In the near field things are “messy”!

• Far away (far-field), only E and H are significant, and simply to

jI0 Lf sin θ −j2πf (t−r/c)
E −−−−−→ e (0.4)
r1 2rε0 c2

jI0 Lf sin θ j2πf (t−r/c)

H −−−−−→ e (0.5)
r1 2rε0 c2

In the far field E and H are equivalent!

Response of Linear Antenna in Free Space

• Although it is clear from equations (0.4) and (0.5) that E and H are
mathematically equivalent (in the far field), it is actually the magnetic
field H that induces a current at the receiver’s antenna, specifically

dH πI0 Lf 2 sin θ

Ir = − = exp j2πf (t − r/c)
dt rε0 c2
α

= exp j2πf (t − r/c) (0.6)
r
where α = πI0 Lf 2 sin θ/ε0 c2 .
 Note: Many (or most!) text books are, strictly speaking, incorrect in
talking about propagation over the electric field E. While not an issue
on far field, in near field (e.g. body area networks) more care must be
taken.

Large Scale Effects

Power Loss: Free space, stationary
Let us study some of the effects the physical environment impose over the ef-
fectively received signal under stationary, free space conditions i.e.,
α
Er , < {Ir } = cos (2πf (t − r/c)) (0.7)
r
• The amplitude of the signal reduces with 1/r:
• The power of the signal reduces with 1/r2 :

α2
E[Er2 ] = (0.8)
2r2

Free space is very similar to the wire channel!

E[V 2 ] P
E[I 2 ] = 2
= 2
Ω Ω

Doppler Distortion: Free space, mobile

Consider a mobile receiver with velocity ~v . Let v denote the scalar speed on
direction of the line between transmitter and receiver. Then

r0 = r + vt (0.9)
α
Er = cos (2πf (1 − v/c)t − 2πf r/c) (0.10)
r + vt
• The power of the electric field stays the same:

α2
E[Er2 ] = E[cos2 (2πf (1 − v/c)t − 2πf r/c)]
E[(r + vt)2 ]
α2
−− −→ (0.11)
r1
T ∝1/f
2r2

2
 • The frequency of the electric field varies with v:

f 0 = f · (1 − v/c) (Doppler shift) (0.12)

Doppler shift causes distortion!

Shadowing: Free space, two paths, stationary

If two paths of lengths r1 and r2 exist between transmitter and receiver, then:
α α
Er = cos (2πf (t − r1 /c)) − cos (2πf (t − r2 /c)) (0.13)
r1 r2
• A deterministic phase difference is observed:

φ1 = −2πf r1 /c; φ2 = −2πf r2 /c + π (0.14)

∆r
∆φ = 2πf +π (0.15)
c

• The power of the electric field varies with the location:

α2 α2
E[Er2 ] −−−→ − E[cos (2πf t) · cos (2πf t + ∆φ)] (0.16)
r1
T ∝1/f
r2 r2

Shadowing is perceived as “blind spots.”

The phenomenon is technically known as spatial selectivity!

Note: Simplistic definitions of delay spread and coherence band-

width can be made at this point:
∆r 1
(delay spread) ∆τ , ⇐⇒ Bc , (coherence bandwidth)
c ∆τ

3
 Figure 1: Interference of two sinusoidal waves of same frequency.

The reciprocal relationship between delay spread and coherence

bandwidth is general. However, different proportionality constants
may be adopted (see e.g. [1, pp. 15]).

Path Loss Exponent: Ground plane reflection

Consider the special case where the second path is due to a ground plane
reflection, as depicted in figure . Then

r1
hs − hr

hs

r2
hs + hr
hr
θ θ
r

Figure 2: Geometry of the ground plane two-path channel.

• The Delay spread varies with the inverse of r.

4
 Homework 1: Show that
2hs hr
∆τ ,−−−−−−−→ (0.17)
r(hs ,hr ) rc
√
Hint: Use the asymptotic result 1 + x −−−→ 1 + x/2.
x1

• The amplitude of the received signal decreases with 1/r 2

Homework 2: Show that

α α
Er = cos (2πf (t − r1 /c)) − cos (2πf (t − r2 /c))
r1 r2
4πf αhs hr
−−−−−−−→ − sin[2πf (t − r1 /c)] (0.18)
rf /c cr2

Hint: Assume that r1 ≈ r2 .

Homework 3: Show the same without assuming that r1 ≈ r2 .

Hint: Use the asymptotic result (1 + x)−1 −−−→ 1 − x.

x1

• The power of the received signal decreases with 1/r 4

β2
E[Er2 ] −−−−−→ (0.19)
r1
T ∝1/f
2r4

4πf αhs hr
where β = .
c

Fresnel Diffraction: Knife-edge Obstacle

Consider the special case where the second path is due to diffraction at a
knife-edge obstacle, as depicted in figure . Then

5
 ψ
ψ1
h h0
T

d1
ψ2
d2
R

hs
hr

Figure 3: Geometry of knife-edge two-path channel.

• A deterministic phase difference is observed:

π 2
∆φ = ν (0.20)
2
with
d1 + d2
ψ=h (0.21)
d d
r 1 2
2d1 d2
ν=ψ (0.22)
λ(d1 + d2 )
where ν, ψ are the Fresnel coefficient and Fresnel diffraction angle,
respectively.
• The amplitude of the received signal varies with ν:

Er −−−−−−−→ F (ν) cos(2πf t − φ) (0.23)

h(d1 ,d2 )

where F (ν) is the Fresnel diffraction loss

1+j ∞ −jπx2
Z  
F (ν) = exp dx (0.24)
2 ν 2

6
 Homework 4: Show that for ν > 1, the Fresnel loss |F (ν)| is tightly
approximated by
1
|F (ν)| ≈ √ (0.25)
2πν
Hint: Start by showing that
Z ∞
−jπx2
     
1 1
exp dx = − C(ν) − j − S(ν) (0.26)
ν 2 2 2

where Z ν Z ν
C(ν) = cos x2 dx; S(ν) = sin x2 dx (0.27)
0 0

Fresnel Diffraction Loss

1.2

1.1

0.9

0.8

0.7
Loss: |F (ν)|

0.6

0.5

0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Coefficient: ν

Figure 4: Behavior of Fresnel Diffraction Loss/Gain.

7
Mandatory Reading
- Goldsmith [2]: Chapter 2

Advised Reading
- Tse&Viswanath [1]: Chapter 2 (Section 2.1)

Recommended
- MIT Course: Principles of Digital Communications [3],
- Lecture 21 by Prof. Gallager (MIT)

8
Lecture 2
Small Scale Effects
Two-path Fading: Free space, two paths, mobile
Let v1 and v2 denote scalar velocities projected on the lines between transmitter
and receiver, and between the specular image of the transmitter and the receiver,
respectively. Then

r10 = r1 + v1 t; r20 = r2 + v2 t (0.28)

α
Er = cos (2πf (1 − v1 /c)t − 2πf r1 /c)
r1 + v1 t
α
− cos (2πf (1 − v2 /c)t − 2πf r2 /c) (0.29)
r2 + v2 t
• The power of the signal varies rapidly with time:

α2 α2
E[Er2 ] −r−−→ − E[cos (2πf t(1 − v1 /c))·cos (2πf t(1 − v2 /c) + ∆φ)] (0.30)
i 1
T ∝1/f
r2 r2
Note: Simplistic definitions of Doppler spread and coherence time can be
made at this point:
1
(Doppler spread) ∆f , f · (v1 − v2 )/c ⇐⇒ Tc , (coherence time)
∆f

Time selectivity (fading)!

Model 1: Ray tracing

Consider the folowing generalization of the two-path model
X

Er = βi (t, f ) cos 2πf (t − τi (t))
i
( )
X

= < βi (t, f ) exp j2πf (t − τi (t)) (0.31)
i

For example, for the two-path model:

α α
β1 = ; β2 =
r1 + v1 t r2 + v2 t
v1 t + r1 φ1 v2 t + r2 φ2
τ1 = − ; τ2 = −
c 2πf c 2πf
φ1 = 0; φ2 = π

9
Model 2: Ray tracing in narrowband complex domain
Consider further the following simplification and generalization to the complex
domain X

Er = βi (t) exp j2πf (t − τi (t)) (0.32)
i

whereby
• We neglect the antenna’s frequency selectivity

βi (t, f ) −→ βi (t) (0.33)

• We neglect the bandwidth

F[cos 2πf t] = 21 [δ(2πf ) + δ(−2πf )]

jF[sin 2πf t] = 21 [δ(2πf ) − δ(−2πf )]

F[exp(j2πf t)] = δ(2πf )

Attention to the short-hand notation: δ(2πf ) is a spike at 2πf

Model 3: Channel impulse response

Define the time-variant channel response function in frequency, which in
terms of Fourier transform is
Z∞
h(f, t) = h(τ, t) exp(−j2πf τ )dτ (0.34)
−∞
F −1
h(f, t) ←−−−→ h(τ, t) (0.35)
F

Attention to the short-hand notation: F{h(t)} = h(f )

• The response of the channel to an arbitrary input:

F −1
x(t) ←−−−→ x(f ) (0.36)
F

10
 is given by
Z∞
Er (t) = x(f )h(f, t) exp(j2πf t)df (0.37)
−∞
Z∞ Z∞
 

Er (t) = x(f )  h(τ, t) exp(−j2πf τ )dτ  exp(j2πf t)df

−∞ −∞
Z∞ Z∞
 

Er (t) = h(τ, t)  x(f ) exp(j2πf (t − τ ))df  dτ

−∞ −∞
Z∞
Er (t) = h(τ, t)x(t − τ )dτ (0.38)
−∞

A signal in the wireless channel undergoes a time-varying

convolution with the channel’s response function!

• For the ray tracing case:

X

h(f, t) , βi (t) exp − j2πf τi (t) (0.39)
i
X

h(τ, t) −−−−−→ βi (t)δ τ − τi (t) (0.40)
F
i
such that the response to an input x(t) is
X

Er (t) = βi (t)x τ − τi (t) (0.41)
i

Where did the frequency dependence go?!

Z

The impulse response model arises as a consequence of ignoring

frequency dependence in each path response!

Homework 5: Explain why the absence of frequency dependence on

the ray tracing response to a narrowband signal as shown in (0.41)
is intuitively correct.

Hint: Study equation (0.37) qualitatively.

Note: Compare equation (0.41) with equation (0.32), and see that the
former is indeed the response to the input x(t) = exp(j2πf t)!

11
Model 4: Baseband Impulse Response
Define the baseband time-variant response function in frequency
 
h(f + f c, t) if f + fc > 0
hb (f, t) = (0.42)
0 if f + fc 6 0
F −1
hb (f, t) ←−−−→ hb (τ, t) (0.43)
F

• The response of the channel to an arbitrary input is:

Z∞
Er,b (t) = x(f )hb (f, t) exp(j2πf t)df (0.44)
−∞
Z∞ Z∞
 

Er,b (t) = x(f )  hb (τ, t) exp(−j2π(f + fc )τ )dτ  exp(j2πf t)df

−∞ −∞
Z∞ Z∞
 

Er,b (t) = h(τ, t)  x(f ) exp(j2πf (t − τ ))df  exp(−j2πfc τ )dτ

−∞ −∞

Z∞ Z∞
Er,b (t) = h(τ, t)x(t − τ ) exp(−j2πfc τ )dτ = hb (τ, t)x(t − τ )dτ (0.45)
−∞ −∞

Note: Recall the Fourier transform frequency shift property

x(t) exp(−j2πfc t) ←−−−→ x(f + fc )

F

• For the ray tracing case:

X

hb (f, t) , βi (t) exp − j2π(f + fc )τi (t) (0.46)
i
X

hb (τ, t) −−−−−→ βi (t)δ τ − τi (t) exp(−j2πfc τi (t)) (0.47)
F
i
X

Er,b (t) = βi (t)x τ − τi (t) exp(−j2πfc τi (t)) (0.48)
i

• Recall that the propagation delay is

 
vi t ri φi fD
τi (t) = + − = τi0 + i t (0.49)
c c 2πf fc

12
 where fDi (fc , vi ) , fc vi /c is the Doppler shift.
• For the ray tracing case:
X
βi (t) x τ − τi (t) exp(−j2π(fc τi0 + fDi t))


Er,b (t) = (0.50a)
i
X
βi (t) exp(−j2πfc τi0 ) x τ −τi (t) exp(−j2πfDi t) (0.50b)


=
i
X
βi0 (t, τi0 ) x τ − τi (t) exp(−j2πfDi t)


= (0.50c)
i
α
where βi0 (t, τi0 ) , βi (t) exp(−j2πfc τi0 ) with βi (t) , such that
ri + vi t
|βi0 (t)| = |βi (t)|

Frequency carrier affects the Doppler shift, but

does not affect the path coefficient!

Note: Of course different frequencies have different absorption

factors, but this is not a small scale effect!

• Impulse response, i.e., response to input x(t) = exp(j2πf t)

X
βi0 (t, τi0 ) δ τ − τi (t) exp(−j2πfDi (fc , vi ) t)


Er,b (t) = (0.51)
i

Doppler shift causes violent and random phase rotations!

1
Specifically, exp(−j2πfDi (fc , vi ) t) goes from 1 to 0 every t =
4|fDi |

Doppler shift is random, since it is affected by the carrier

frequency and the relative velocity between the
transmitter/receiver pair!

Multiple paths with mismatched Doppler shifts cause fading!

Statistics of Small Scale Effects

Let us now pursue a small scale statistical perspective of the wireless channel
(valid for short periods of time). Inspired in the models discussed in the previous
lecture, consider the more abstract, if simpler, baseband multi path channel
model below X
hb (t) = βi exp(−jφi (t)) (0.52)
i

13
where the dependence of βi with time ignored, while all the other physical pa-
rameters of interest such as Doppler shifts, propagation delay etc. are embedded
in the time-varying phase φi (t)

φi (t) , 2πfc τi0 + 2πfDi t (0.53)

Autocorrelation Function under Multipath Narrowband Model

Let us furthermore split hb (t) into I and Q components
P
hI (t) = βi cos(φi (t)) (0.54)
i
P
hQ (t) = − βi sin(φi (t)) (0.55)
i

• The autocorrelation function of the in-phase component is given by

X  
AI (τ ) , E [hI (t)hI (t + τ )] = E βi2 E [cos(φi (t)) cos(φi (t + τ ))]
i
(0.56)
which can be simplified to
1 X  2
AI (τ ) = E βi E [cos(2πfDi τ )] (0.57)
2 i

Homework 6: Prove equation (0.57).

Hint: Substitute equation (0.53) into (0.56) and simplify.

Autocorrelation Function under Clarke’s Multipath Model

• The average powers of all paths are the same:

E βi2 = 2P/N
 
(0.58)

• The Doppler shift of each path is proportional to its angle-of-arrival:

fc vi fc v
fDi = ≈ cos θi (0.59)
c c
• The angles-of-arrival are uniformly distributed:
2πi
θi ≈ = i∆θ (0.60)
N

where ∆θ , 2π/N .

14
 Bessel Function of the First Kind
1

0.8

0.6

0.4
J0 (2πfD τ )

0.2

−0.2

−0.4
0 0.5 1 1.5 2 2.5 3
fD τ

Figure 5: Normalized autocorrelation function of uniformly dense wireless chan-

nel (Clarke’s) model.

• Substituting the approximations above we obtain

2πfc vτi0
  
P X
AI (τ ) = E cos cos i∆θ
N i c
2πfc vτi0
 
P X
= cos cos i∆θ ∆θ (0.61)
2π i c
• Taking the number of terms to infinity (N → ∞)
Z2π  
P 2πfc vτ
AI (τ ) = cos cos θ dθ = P J0 (2πfD τ ) (0.62)
2π c
0

where we finally dropped the “ i ” and “ ’ ” from the notation for simplicity,
and since those no longer have meaning in the continuous model above.

Note: It is trivial to show that AQ (τ ) = AI (τ ).

Coherence Time and Doppler Spread

Let us start by emphasizing that the expectation taken in equation (0.61), and
the integration performed in equation (0.62), fD represents the aggregate con-

15
tribution of all phase mismatches caused by the Doppler shifts of all paths.
In other words, fD is now to be interpreted (and referred to) as a Doppler
Spread.

• The Coherence Time of a (dense) multi path channel is defined as the

time it takes for the channel to become uncorrelated, i.e.

Tc = min{τ }|A(τ ) = 0 (0.63)

3
Homework 7: Prove that Tc = .
8fD
r
2
cos x − π2 η − π4 .


Hint: Use the approximation Jη (x) ≈
πx

• The Fourier Transform of the Autocorrelation Function of a chan-

nel is referred to as the Power Spectral Density. For the narrowband
dense multi path model above (Clarke’s):

 P p 1

for |f | 6 fD
SI (f ) , F{AI (τ )} = 2πfD 1 − (f /fD )2 (0.64)
0 otherwise


Notice how as fD increases, the Power Spectral Density SI (f ) tends

to a pair of spikes at −fD and fD , implying that the
Autocorrelation Function tends to a sinusoidal waveform.
What does that mean?

Conversely, as fD → 0, the Power Spectral Density SI (f ) tends to a

rectangular function at the origin. What does that mean?
sin(πτ )
sinc(x) = πτ −−−−−→ rect(f ) = u (f + fD ) − u (f − fD )
r1

Question: What is the integral of SI (f ) over its entire domain,

and what does it represent?

16
 Power Spectral Density
3.5

3 fD = 2

2.5

2 fD = 4
S(f )

1.5
fD = 8

0.5

0
−9 −7 −5 −3 −1 1 3 5 7 9
f

Figure 6: Power Spectral Density of Narrowband Dense Multipath Channel.

Mandatory Reading
- Goldsmith [2]: Chapter 3

Additional Reading
- Rappaport [4]: Chapter 5
- Tse&Viswanath [1]: Chapter 2

Recommendend
- Principles of Digital Communications [3],
- Lecture 21 by Prof. Gallager (MIT)
- At: http://www.youtube.com/watch?v=2DbwtCePzWg

17
Lecture 3
Statistical Characteristics of Fading Channels
Let us return to the baseband channel’s time-domain response function
X

hb (τ, t) = βi (t)δ τ − τi (t) exp(−j2πfc τi (t)) (0.65)
i

Figure 7: Illustration of a baseband impulse response.

• The function P(τ ) below is known as the Power Delay Profile of the
channel, at the location z.

P(τ ) , γEt|z [|hb (τ, t)|2 ] (0.66)

where γ is normalizing constant.

18
 P

β1

βN

τ1 τN Delay

Figure 8: Illustration of the Power Delay Profile function.

Note: Recognize that the Power Delay Profile is actually a large scale
characteristic of the channel!

• Given a certain PDP, its corresponding average and rms delays can also
be defined as
Z ∞
τ P(τ )dτ
µτ = Z0 ∞ (0.67)
P(τ )dτ
0
Z ∞  21
2
 (τ − µτ ) P(τ )dτ 
0
στ = 
 Z ∞ 
 (0.68)
P(τ )dτ
0

• A widely used PDP model (supported by empirical and analytical evi-

dence) is the exponential profile
1
P(τ ) = exp(−τ /γ). (0.69)
γ

Homework 8: Prove that the mean and rms delays of an exponential

PDP shown in equation (0.69) as are both equal, i.e., µτ = στ = γ.

19
 Homework 9: Prove that under an exponential PDP as shown in
equation (0.69), the number of paths N that arrive in a time T is a
Poisson random variable with intensity 1/γ. Or, mathematically
 N
e−T /γ T
P {no. paths in T = N } = . (0.70)
N! γ

What is the average number of paths?

nP o nP o
N +1 N
Hint 1: Study P i τi > T and P i τi < T .

Hint 2: Recall that the sum of exponential variates follow the

Erlang distribution which yield a Poisson.

τ1 τ2 τn τn+1

Pm
Tm = i=1 τi

Figure 9: Illustration of exponential arrivals.

Envelope Spectral Crosscorr. and Power Delay Profile

Let us now consider the following simplification of the channel response function
N
X

hb (τ ) = hi δ τ − τi (0.71)
i=1

where we have neglected the temporal dependence of τi with time, and included
the time-varying phase rotations caused by Doppler shifts into the complex
gains hi .

Recall that N is typically also a random variable!

• Let the Power Delay Profile of the channel be given by

P , [|h1 |2 , · · · , |hN |2 ] = [P1 , · · · , PN ] (0.72)

• If the channel gains hi are independent, the associated covariance matrix

is
Rh , E[hh† ] = diag[P1 , · · · , PN ] (0.73)

20
 †
where denotes transpose conjugate, and h , [h1 , · · · , hN ].
• Due to the fast rotation of the phases of hi , these random variates can
also be assumed to be zero mean.
Next, consider the frequency domain response function of the channel, given
by the Fourier transform of hb (τ ), i.e.
Z ∞ N
X
hb (f ) , hb (τ ) exp(−j2πf τ )dτ = hi exp(−j2πf τi ) (0.74)
−∞ i=1

• Consider a pair of sinusoidal signals with frequencies f1 and f2 , and define

the associated vectors
wk , [exp(j2πfk τ1 ), · · · , exp(j2πfk τN )] (0.75)

Notice the positive arguments!

• The responses to each of these sinusoidal signals in the frequency do-

main can be written in matrix form as
yk = wk† · h (0.76)

• Let us define also the quantities

y = [y1 , y2 ] (0.77)
rk = |yk | (Envelope Response) (0.78)
∆f = f1 − f2 (0.79)

• The cross-correlation coefficient between the envelope responses in

the frequency domain is given by.
E[r1 r2 ] − E[r1 ] E[r2 ]
C(r1 , r2 ) = (0.80)
E[r12 ] − E[r1 ]2

• It can be shown [5] that

√
πP
E[r1 ] = (0.81)
2
E[r12 ] = P (0.82)
2 2

πP (1 − λ )
E[r1 r2 ] = 2 F1 − 12 , − 12 ; 1; λ2 (0.83)
4
P
where P , Pi , 2 F1 (a, b; c; z) is the Gaussian hypergeometric func-
tion and λ2 is the Fourier transform of the sample correlation function
of the Power Delay Profile, specifically
PN PN
Pn Pm exp(j2π∆f (τn − τm ))
λ2 = n m (0.84)
P2

21
 • From equation (0.83) one obtains [5]
− 12 , − 12 ; 1; λ2 − 1


2 F1
C(r1 r2 ) = (0.85)
4/π − 1

Note: The result holds for arbitrary Power Delay Profiles!

• Finally, let us show that the Envelope Spectral Cross-correlation

Coefficient is the same as the Fourier transform of the Power Delay
Profile.

Homework 10: Show from equation (0.85) that C(r1 r2 ) ≈ λ2 .

Hint 1: What is the maximum value of λ2 ?.

Hint 2: Use the approximation 2 F1 − 21 , − 12 ; 1; λ2 ≈ (4/π − 1)λ2 + 1,

for 0 6 λ2 6 1.

The Envelope Spectral Cross-correlation Coefficient is the

Fourier Transform of the Power Delay Profile!

Gaussian Hypergeometric Function

1.3

1.25

1.2
)
2
2 F1 (−1/2, −1/2; 1; λ

1.15

1.1

1.05

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
λ2

Figure 10: Gaussian Hypergeometric Function 2 F1 − 12 , − 12 ; 1; λ2 and a linear

approximation (4/π − 1)λ2 + 1.

22
Coherence Bandwidth and RMS Delay
The coherence bandwidth of the channel is defined as the bandwidth over
which the envelope spectral cross-correlation coefficient is greater than a given
threshold η.

 
f∗ = f C(r1 , r2 ) = η (0.86)

• From the above, we may replace the envelope spectral cross-correlation co-
efficient by the Fourier Transform of the PDP. Consider the Exponential
PDP of equation (0.69)

|F {exp(−τ /στ )} | 1
P(f ) = =p . (0.87)
στ 1 + (2πf στ )2

• Assuming that 2πf στ  1, we have

1
f∗ = . (0.88)
2πηστ

The coherence bandwidth is proportional to the

inverse of the rms delay!

Statistical Fading Channel Models

Let us now pursue a large scale statistical perspective of the small scale
variations of amplitude (envelope) and phase observed in the wireless channel.
Again, start form the baseband multipath channel model below
X
hb (t) = βi exp(−jφi (t)) (0.89)
i

Then, consider the extension of the model above to the following

P
z(t) = αi cos(φi (t)) + jβi sin(φi (t)) (0.90)
i
z(t) = x(t) + jy(t) (0.91)

where in (0.90) we accommodate for the possibility (so far ignored) that the
path gains may be distinct on the different complex dimensions (polarization),
while in (0.91) we assume that the contributions of the sum of a large number
of sinusoidal functions with random phases and amplitudes amount to
the independent random processes x(t) and y(t).

23
 The problem of determining the probability density functions
(PDF’s) of the amplitude r , |z| and phase θ , ∠z of random
processes with the structure shown above is often
referred to as the Random Phase Problem.

The Central Limit Theorem

Let {x1 , · · · , xN } be independent, random variates with arbitrary distribu-
2
tions. Define µX , EN [xi ] and σX , EN [x2i ] − E2N [xi ], where EN denotes
expectation over the set of N samples. Then, the quantity
N
N X N − N µX X Yi
ZN = √ = √ (0.92)
σX N i=1
N
converges to the standard Normal distribution.

Proof: Notice that Yi implicitly defined above has a zero-mean and unit vari-
ance. By force of Taylor’s theorem,
t2
ϕY (t) = 1 − + o(t2 ), t → 0 (0.93)
2
where ϕY (t) denotes the characteristic function of the distribution of Y and
o(t2 ) denotes a term that decreases faster then t2 as t → 0.
Since the characteristic function of a sum is the product of charac-
teristic functions, and due to a simple change of variables, it follows that
n   2 n
t2
 
t t 2
ϕY √ = 1− +o → e−t /2 , n → ∞. (0.94)
n 2n n
2
It is then identified that the e−t /2 is the characteristic function of the stan-
dard Normal distribution, which concludes the proof.


If paths are independent, in large quantities, and statistically

equivalent (that is, no dominant), then by force of the
Central Limit Theorem, equation (0.91) describes
a complex Gaussian process.

Side Note: Maximum Entropy Distributions

Let us momentarily depart from the discussion about the Random Phase Prob-
lem and highlight the important property of entropy maximality that a few
important distributions (two of which already appeared so far) exhibit.
In what follows we shall make use of the following optimization technique known
as Lagrange Multiplier method.

24
Lagrange Multiplier Method
Consider the following constrained optimization problem

minimize/maximize f (x) (0.95)

subject to g(x) = c (0.96)

Define the function  

Λ(x, λ) , f (x) + λ · g(x) − c (0.97)
Then, the solution of the problem is amongst the solutions of

∇Λ(x, λ) = 0 (0.98)

where ∇ denotes gradient.

The solutions of equation (0.98) are in general inflection points.

One must therefore verify/choose an admissible solution.

Example 1: Entropy Maximality of Uniform Distribution

Let p(x) be a non-negative function in [a, b] such that
Z b
p(x) dx = 1. (0.99)
a

Due to the non-negativity of p(x) and the normalization constraint given by

equation (0.99), p(x) is a probability distribution1 .
Next, consider the entropy of p(x)
Z b Z b
H(p) , − p(x) ln p(x) dx = − p ln p dx. (0.100)
a a

So, we are interested in the problem:

Z b
maximize − p ln p dx (0.101)
a
Z b
subject to p dx = 1 (0.102)
a

Applying the Lagrange Multiplier method, the Lagrangean (on p) is

!
Z b Z b
Λ(p) = − p ln p dx + λ p dx − 1 (0.103)
a a

1 Rigorously speaking these are only two of the three Kolmogorov axioms required of proba-

bility distributions. The third (countable additivity) is implied in the (assumed) integrability
of p(x).

25
In this case the gradient reduces to a simple derivative on p, namely
∂Λ(p)
= − ln p(x) − 1 + λ (0.104)
∂p(x) dx
from which it follows (solving the root) that

p(x) = eλ−1 . (0.105)

Notice from equation (0.105) that p(x) does not depend on x!

Finally, substituting equation (0.105) into (0.99) so as to find the value of λ

that satisfied the normalization constraint yields

λ = 1 − ln(b − a) (0.106)

Thus, finally we obtain

1
p(x) = , a 6 x 6 b (Uniform Distribution) (0.107)
b−a

The maximum entropy distribution in [a, b] without

prior information is the Uniform Distribution!

Example 2: Entropy Maximality in R with Known µ and σ

Now, what is the maximally entropic distribution in R with known mean and
variance?
In other words, what is the solution of the problem:
Z∞
maximize − p(x) ln p(x) dx (0.108)
−∞
Z∞
subject to p(x) dx = 1 (0.109)
−∞
Z∞
and x · p(x) dx = µ (0.110)
−∞
Z∞
and (x − µ)2 p(x) dx = σ 2 (0.111)
−∞

Before we continue, notice that

Z Z
p(x) ln p(x) dx = p(x − µ) ln p(x − µ) dx (0.112)
R R

26
which can be proved by simple substitution of variables.

The mean constraint is irrelevant!

In other words, we may consider, w.l.g., the spacial case where µ = 0, such that
the maximization problem reduces to In other words, what is the solution of the
problem:
Z∞
maximize − p(x) ln p(x) dx (0.113)
−∞
Z∞
subject to p(x) dx = 1 (0.114)
−∞
Z∞
and x2 p(x) dx = σ 2 (0.115)
−∞

The Lagrange function, taking into account this constraint is then

Z∞
 ∞   ∞ 
Z Z
Λ(p) , − p(x) ln p(x) dx+λ1 p(x) dx − 1 −λ2 x2 p(x) dx − σ 2
−∞ −∞ −∞
(0.116)
And the derivative on p is
∂Λp
= − ln p(x) − 1 + λ1 − λ2 x2 = 0, (0.117)
∂p(x) dx
such that 2 2
p(x) = e(λ1 −1)−λ2 x = e(λ1 −1) · e−λ2 x . (0.118)
Substituting equation (0.118) into the normalization and variance constraints,
respectively, yields
Z∞ r 
(λ1 −1) −λ2 x2 π (λ1 −1)

e e dx = ·e =1 1

 
λ2


 
 λ2 =
2σ 2
−∞
 
∞
=⇒
Z
2
r
π

  e(λ1 −1) = √ 1

e(λ1 −1) x2 e−λ2 x dx = · (λ1 −1) 2  
e = σ  2πσ
4λ32



−∞
(0.119)
Substituting equation (0.119) into (0.118) we have
1 x2
p(x) = √ · e− 2σ2 . (0.120)
2πσ

The maximum entropy distribution in R with prior knowledge of

variance is the Gaussian Distribution!

27
Example 3: Entropy Maximality of Exp. Distribution

Homework 11: Prove that the maximum entropy distribution in

R+ (non-negative Real support) with a known mean is the Exponen-
tial Distribution.
Hint: Use the Lagrance Multiplier method.

Recall from Lecture 4 that the Exponential model is commonly used

to describe the Power Delay Profile of wireless channels.

Back to the Random Phase Problem...

Let us return to the Random Phase Problem
X
z(t) = αi cos(φi (t)) + jβi sin(φi (t)) = x(t) + jy(t).
i

With the conviction that x(t) and y(t) can be assumed to be Gaussian ran-
dom functions, let us attempt to derive corresponding Fading distributions.
In so-doing, pay close attention to the impact of different choices that µ
and σ exercise on the resulting fading models!

Rayleigh Solution of the Random Phase Problem

• Assume that x and y are independent and identically distributed
(i.i.d) Gaussian variates with zero mean and variance σ 2
2
1 −x2 1 −y
x ∼ p(x) = √ e 2σ2 y ∼ p(y) = √ e 2σ2 (0.121)
σ 2π σ 2π

• Due to independence, the joint distribution of x and y is

2 2
1 −(x +y )
p(x, y) = e 2σ 2 (0.122)
σ 2 2π

• Going from density to probability

P {x, y} = p(x, y)dxdy (0.123)

28
• Going from cartesian to polar coordinates

θ = atan xy
p
r = x2 + y 2 (0.124)

dxdy ←→ rdr dθ (0.125)

r −r 2
P {r, θ} = e 2σ 2 dr dθ (0.126)
2πσ 2

r −r 2
p(r, θ) = 2
e 2σ2 (0.127)
2πσ

• Marginalizing

Z2π
r −r 2 r −r22
p(r) = e 2σ 2 dθ = e 2σ , r > 0 (Rayleigh PDF) (0.128)
2πσ 2 σ2
0

Z∞
r −r 2 1
p(θ) = 2
e 2σ 2 dr = , 0 6 θ 6 2π (0.129)
2πσ 2π
0

Notice that the envelope and phase distributions turn up to be

independent since p(r, θ) = p(r) · p(θ).

Rayleigh channels arise when no line-of-sight between transmitter

and receiver exists, and waves are scattered equally in I and Q.

29
 Rayleigh Fading Distribution
0.7

0.6

σ=1

0.5

0.4
p(r; σ)

0.3 σ=2

0.2 σ=3

0.1

0
0 1 2 3 4 5 6 7 8 9 10
r

Figure 11: Examples of Rayleigh PDFs.

• Discrete-time Rayleigh variates can be generated via the Box-Müller

Transform
√
x = − ln u1 sin(2πu2 ) (0.130a)
√
y = − ln u2 cos(2πu1 ) (0.130b)

where ui are i.i.d. standard uniform random variates.

• Continuous-time Rayleigh variates can be generated via the so-called Jakes
Method
"N #
√ X jβn 1
R(t, k) = 2 2 e cos (2πfn t + θn,k ) + √ cos 2πfD t (0.131)
n=1
2

where fD is the Doppler spread and the remaining parameters are

πn
βn = N +1 (0.132)
θn,k = βn + 2π(k−1)
N +1 (0.133)

There are many improved variations of the Jakes Model!

30
 Figure 12: Illustration of probability regions of Rayleigh fading.

Rice Solution of the Random Phase Problem

• Assume that x and y are independent Gaussian variates with the
same variance σ 2 , but unequal means
2 2
1 −(x−µ) 1 −y
x ∼ p(x) = √ e 2σ2 y ∼ p(y) = √ e 2σ2 (0.134)
σ 2π σ 2π

• Due to independence, the joint distribution and joint probability of

x and y are
2 2 2
1 −(x +y ) −(µ −2xµ)
p(x, y) = 2
e 2σ2 e 2σ 2 P {x, y} = p(x, y)dxdy (0.135)
σ 2π
• Going from cartesian to polar coordinates
2 2
r −(r +µ ) 1 rµ cos θ
P {r, θ} = 2
e 2σ2 e σ2 dr dθ (0.136)
σ 2π
2 2
r −(r +µ ) 1 rµ cos θ
p(r, θ) = 2
e 2σ2 e σ2 (0.137)
σ 2π

31
• Marginalizing

Z2π 2 2 Z2π
r −(r +µ ) 1 rµ cos θ
p(r) = p(r, θ)dθ = 2 e 2σ2 e σ2 , r>0
σ 2π
0 0

2 2
r −(r +µ )  rµ 
= e 2σ 2 I 0 (Rice PDF) (0.138)
σ2 σ2

Z∞
p(θ) = p(r, θ)dr, 0 6 θ 6 2π
0
√ 2
1+2 πK·cos(θ−φ)·e4K cos (θ−φ) √
= 2πe4K · (1+erf(2 K cos(θ−φ))), (0.139)

where K , µ2 /2σ 2 (i.e., ratio between the power of direct compo-

nent and the power of indirect component) is referred to as the Rice
Factor, and the the phase shift φ = π4 , 3π 5π 7π

4 , 4 , 4 .

Notice that if we define Ω = µ2 + 2σ 2 (total channel power), the

following variation of the Rice PDF can be written
2
r !
2(1 + K)r −K− (1+K)r K(1 + K)
p(r) = e Ω I0 2r (0.140)
Ω Ω

• Rice variates can be generated similarly to Rayleigh simply by adding

the constant µ.

Rice fading occurs when a dominant LOS path

exists amidst the scattered paths.

The Rice PDF reduces to the Rayleigh PDF is µ = 0.

Notice that equation (0.139) implies that Rice fading

is quadrant-stable for large µ.

32
 Rice Fading Distribution
0.7

0.6

µ = 0; σ = 1

0.5

µ = 1; σ = 1
0.4
p(r; µ, σ)

0.3
µ = 1; σ = 2

0.2 µ = 2; σ = 2

0.1

0
0 1 2 3 4 5 6 7 8 9 10
r

Figure 13: Examples of Rice PDFs.

Hoyt Solution of the Random Phase Problem

• Assume that x and y are independent zero-mean Gaussian variates
with the distinct variances

1 x2
− (1+b)σ 2
x ∼ p(x; b, σ 2 ) = p e (0.141a)
π(1 + b)σ 2
y2
2 1 − (1−b)σ 2
y ∼ p(y; b, σ ) = p e (0.141b)
π(1 − b)σ 2

where 0 6 b 6 1.
• The variances of the I and Q components are

σ2
σx2 = (1 + b) (0.142)
2
σ2
σy2 = (1 − b) (0.143)
2
such that σ 2 = σx2 + σy2 .

33
• Proceeding as before, the PDFs of the envelope and phase are
r2 br2
 
2r − σ 2 (1−b2 )
p(r; b, σ) = √ e I0 (Hoyt PDF)
σ 2 1 − b2 σ 2 (1 − b2 )
√ (0.144)
1 − b2
p(θ; b) = (0.145)
2π(1 − b cos(2θ))

Homework 12: Prove equations (0.144) and (0.145).

Hint: Start by showing that the joint PDF is

1−b cos(2θ) 2
r − r
p(r, θ) = √ e σ2 (1−b2 )
σ 2 π 1 − b2

• Alternatively, Hoyt PDFs can be expressed as

2 2 2
(1 + q 2 )r − (1+q ) r
(1 − q 4 )r2
 
p(r; q, σ) = e 4q 2 σ 2 I (0.146)
0
qσ 2 4q 2 σ 2
q2 σ2
p(θ; q) = (0.147)
2π(1 + q 2 )(sin2 θ + q 2 cos2 θ)

where −1 6 q 6 1.

Homework 13: Prove equation (0.146).

Hint: Compare equations (0.144) and (0.146). What is the relationship

between b and q?.

34
 Hoyt Fading Distribution
0.7

b = 0.8; σ = 0.5
0.6

0.5
b = 0.6; σ = 1

0.4
p(r; b, σ)

b = 0.2; σ = 2
0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
r

Figure 14: Examples of Hoyt PDFs.

Figure 15: Examples of the phase distributions of Rician and Hoyt fading.

35
 (a) Hoyt (b = 0.5) (b) Hoyt (b = −0.75)

Figure 16: Probability regions of signals under Hoyt fading.

Hoyt channels are found in satellite communications.

Notice the pola rization effect of Hoyt channels.

The Hoyt PDF also reduces to the

Half-normal PDF when b = 0.

Nakagami Solution of the Random Phase Problem

• No assumption over the statistics of x and y!
• Derivation is very hard but yields
2 m m 2m−1
exp(− σm2 r2 )


p(r; m, σ 2 ) = Γ(m) σ2 r (Nakagami PDF)

Γ(m) m−1
p(θ; m) = 2m Γ(m/2)2 | sin(2θ)| (Nakagami Phase PDF)

• Discrete-time Nakagami-m variates can be generated with

r
Ω √ √
z= · (ux γx + juy γy ), (0.148)
2m
where ux and uy are independent and uniformly distributed random num-
bers belonging to the set {−1, 1}, while γx and γy are i.i.d. Gamma
variates with shape parameter m/2 and scale parameter 2, i.e.,
γi ∼ Γ(m/2, 2) (0.149)

36
 The parameter m is known as the fading figure

• Piecewise continuous Nakagami-m processes with half-integer fading

figure m can be constructed via
 v v 
s udme ubmc
Ω uX uX
z(t|T ) = ux (t|T ) x2i (t) + juy (t|T )t yi2 (t) ,
 t 
dme + bmc i=1 i=1

(0.150)
where d·e and b·c denote the ceil and floor operators, and the processes
x2i (t) and yi2 (t) are as prescribed in the Jakes method.
• For arbitrary real m, the best approximate method so far is [6]
(
zL with probability ρ
z= (0.151)
zU with probability 1 − ρ

where mL , b2mc/2, mU , mL + 0.5 and

zL ∼ pz (z; mL , σ) (0.152)

zU ∼ pz (z; mU , σ) (0.153)
mL
ρ=2 (mU − m) (0.154)
m

37
 Nakagami-m Fading Distribution
1.5

m = 0.2; σ = 0.5
1.25

m = 1.5; σ = 0.5

1
p(r; m, σ)

m = 0.5; σ = 1
0.75

m = 1; σ = 2

0.5

0.25

0
0 0.5 1 1.5 2 2.5 3 3.5 4
r

Figure 17: Examples of the Nakagami-m PDFs.

(a) Nakagami (m = 1.25) (b) Nakagami (m = 2.5)

Figure 18: Probability regions of signals under Nakagami-m fading.

38
 Trajectory Plots of Piecewise Continuous Complex−Valued
Nakagami− m Channels (4 Snapshots; m = 2)
2

1.5

0.5
Imag(z)

−0.5

−1

−1.5

−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Real(z)

Figure 19: Snapshots of the piecewise continuous Nakagami-m processes.

Nakagami-m channels reduce to Rayleigh for m = 1,

approximate Hoyt for 0.5 6 m < 1, and Rice for m > 1.

Due to this flexibility and also to the simple expression,

the Nakagami-m model is the most popular after Rayleigh.

Circular Central-Limit Theorem: Tikhonov PDF

Circular variates are variates with 2π-periodic density

r ∼ p(mod(r, 2π)) (0.155)

• The Tikhonov or von Mises distribution is the equivalent of the

Gaussian distribution for circular variates

- Closed-form

39
 1
pT (x; α, ξ) , · eα cos(x−ξ) , x ∈ S , [−π, π]
2πI0 (α)

- Jacobi-Anger infinite series representation

∞
1 X Ik (α)
pT (x; α, ξ) = +2 · cos(k(x − ξ))
2π I (α)
k=1 0

• Tikhonov variates can be generated efficiently as described in [7].

The Tikhonov distribution is also known as

von Mises distribution.

It is associated with the phase error at the output of

a Phase Locked Loop (PLL) with input power
proportional to the parameter alpha.

Tikhonov Phase Distribution

1

0.9 α = 5; ξ = −π/4

0.8

0.7 α = 3; ξ = 0

0.6
p(θ; α, ξ)

0.5

0.4
α = 1; ξ = π/2

0.3

0.2

0.1

0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
θ

Figure 20: Examples of Tikhonov PDFs.

40
Mandatory Reading
- Goldsmith [2]: Chapter 3 (Sections 3.2 and 3.3)

Additional Reading
- Rappaport [4]: Chapter 5 (Sections 5.4 to end)
- Tse&Viswanath [1]: Chapter 2 (Section 2.3 to end)
- Proakis [8]: Chapter 2 (Section 2.1.4)

Recommendend
- Website: Lagrange Multiplier Method
. http://en.wikipedia.org/wiki/Lagrange_multiplier
- Article: Q. T. Zhang and S. H. Song, “Exact expression for the
coherence bandwidth of Rayleigh fading channels,” IEEE Trans.
Commun., vol. 55, no. 7, pp. 1296 – 1299, 2007.
- Article: G. T. F. de Abreu, “On the moment-determinance and
random mixture of Nakagami-m variates,” IEEE Transactions on
Communications, vol. 58, no. 9, pp. 2561 - 2575, Sep. 2010.
- Article: G. T. F. de Abreu, “On the generation of Tikhonov
variates,” IEEE Transactions on Communications, vol. 56, no. 7, pp.
1157 - 1168, Jul. 2008.

41
Lecture 4
Looking Deeper into Classic Fading Models
We discussed in class that when working with random variates we are often re-
quired to know their moments. In principle, we can always refer to the definition
in order to obtain such quantities, i.e.,
Z ∞
µn , rn p(r) dr. (0.156)
−∞

It can be inferred from the above that the Moments of the classic
envelope fading distributions can be complicated functions.

Examples: Moments of Classic Envelope Fading PDFs

Consider the moments of the 4 main fading channel models discussed earlier.

• Moments of Nakagami-m envelope fading PDF

Γ(m+1/2)
E[r] = σ (Mean) (0.157)
Γ(m)
E[r2 ] = σ 2 (Power) (0.158)
 n
n Γ(m+n/2) σ 2 2
E[r ] = (0.159)
Γ(m) m

where it is common to denote the power by Ω , E[r2 ] = σ 2 .

• Moments of Rayleigh envelope fading PDFs

E[r] = σ π2
p
(Mean) (0.160)
E[r2 ] = Ω = 2σ 2 (Power) (0.161)
√
E[rn ] = 2n σ n Γ(1 + k/2) (0.162)

• Moments of Ricean envelope fading PDF

E[r] = µ (Mean) (0.163)

E[r2 ] = Ω = µ2 + 2σ 2 (Power) (0.164)
p
E[rn ] = (2σ)n Γ(1+n/2) Ln/2 (−µ2 /2σ 2 ) (0.165)

where L0ν (z) = 1 F1 (−ν; 1; z) is a Laguerre polynomial.

42
 • Moments of Hoyt envelope fading PDF
s
2σ 2
E[r] = ε(1 − q 2 ) (Mean) (0.166)
π(1 + q 2 )

E[r2 ] = σ 2 (Power) (0.167)

E[rn ] = too complicated ! (0.168)
Z π
2 p
where ε(z) , 1 − z sin2 t dt is the complete Elliptic Integral of
0
type E.

Is there a systematic and easier way to obtain specific moments of a

distribution, without having to derive the moment function?

The Moment Generating Function

The moment generating function of a continuous distribution is defined with
basis on the two-sided Laplace Transform as
Z ∞
MR (t) , ER ert = p(r)etr dr
 
(0.169a)
−∞
∞
X tn t2 µ2
= µn = 1 + tµ1 + + ··· (0.169b)
n=1
n! 2!

Notice that, while the definition is two-sided, all envelope fading

distributions are single-sided.

The MGF owes its name to the following property

dn MR (t)
µn = (0.170)
dtn t=0

where µn denotes the n-th moment of p(r).

Notice that equation (0.170) is an imediate consequence

(or Corollary) of equation (0.169b) !

43
 Homework 14: Prove equation (0.169b).

Hint 1: Recall that the Taylor series representation of a function g(x)

in the vicinity of a point x = a is given by

dg(x) d2 g(x)
g(x) = g(a) + (x − a) · + (x − a)2 · + ···
dx x=a dx2 x=a

Hint 2: Look at equations (0.169a) and (0.169b)! Where should you

apply the Taylor series?

The Cumulant Generating Function

There exists also a cumulant generating function, which is defined as
∞
   X 1
n
KR (t) , log ER ert = − 1 − E[ert ] (0.171a)
n=1
n
∞ ∞
! n
X (−1)n X ti
=− µi (0.171b)
n=1
n i=1
i!

t2
t3
= µ1 t + µ2 −µ1 2 + µ3 −3µ2 µ1 +2µ1 3 + ···
| {z } 2! | {z } 3!
κ2 κ3

Again, the CGF has the following property

dn KR (t)
κn = (0.172)
dtn t=0

where κn denotes the n-th cumulant of p(r).

Notice that the second cumulant is the variance !

A classic result by Petrov which relates moments and cumulants directly has
been recently generalized by Blinnikov and Moessner [9, Eq. 32], yielding
n
X
sn −1
Y 1  µi αi
κn = n! (−1) (sn − 1)! , (0.173)
α ! i!
i=1 i
{αn }

where sn = α1 + α2 + · · · + αn , and {αn } is the set of all non-negative integer

solutions of the Diophantine equation of order n,

α1 + 2 α2 + · · · + n αn = n. (0.174)

It was mentioned earlier that it is often (but not necessarily always) possible to
reconstruct a distribution from a number of known moments. A good method

44
to do so is the so-called Edgeworth expansion [9]. Specifically
 !αj 
n i
X i/2
X Y 1 κj+2
p(r; {µn }) , pN (x) 1 + κ2 Hi+2si (x) (n+1)

i=1 j=1
αj ! (j + 2)!κ
{αi } 2
(0.175)
where pN (x) here denotes the Standard Normal distribution and Hi (x) is the
Hermite polynomial of i-th order

Nakagami-m Fading Distribution and Edgeworth Approximations

0.7

0.6 m = 1, Ω = 2, n = 6

0.5 m = 2, Ω = 5, n = 4

0.4
p(r; m, Ω)

m = 3, Ω = 15, n = 3

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8
r

Figure 21: Illustration of the accuracy of the Edgeworth moment-based PDF

reconstruction.

45
Kurtosis and Skewness
In addition to the mean (generalized by moments, and the variance (general-
ized nu cumulants), two important parameters of distributions are the kurtosis
and skewness.
Skewness measures the symmetry of a distribution about its own mean and
is defined by
E[(r − µ)3 ] µ3 − 3µ1 κ2 − µ31
S, 3/2
= 3/2
(0.176)
(E[(r − µ)2 ]) κ2

Generally,

S = 0 =⇒ Symmetric distribution
S > 0 =⇒ Distribution with most “mass” to the left of median
S < 0 =⇒ Distribution with most “mass” to the right of median

By and large fading distributions have positive skewness !

Figure 22: Illustration of distributions with negative and positive S.

46
The kurtosis measures the “peakyness” of a distribution and is defined by

E[(r − µ)4 ] (µ4 + µ41 ) + µ1 (κ3 + 2µ3 )

K, 2 = (0.177)
(E[(r − µ)2 ]) κ22

Since the kurtosis of the Gaussian distribution is zero, the measure of

“peakness” associated with the kurtosis is also often associated with Gaussianity.
In particular,

K = 0 =⇒ Gaussian-like distribution
K > 0 =⇒ Distribution with heavy tails (super-Gaussian)
K < 0 =⇒ Distribution with thin tails (sub-Gaussian)

Standard Wireless System Model

Let si represent the i-th signal of an alphabet S, and consider the follow-
ing model of the response to its input into a wireless channel, aka, “received
signal”.
yi = hi · si + wi (0.178)
where hi is a complex random variate carrying the amplitude and phase varia-
tions studied earlier (see Lesson 5), while wi model noise

The notation implies that the signal alphabet is finite!

• Notice that the effect of noise is additive

• It is often assumed that noise outcomes are fully uncorrelated, e.g,

E[wi · wj∗ ] = 0, ∀ i, j (0.179)

1
- Recall that Tc ∝ W , such that equation (0.179) implies that the
bandwidth of noise is infinite. In other words, noise is white!
• Since noise originates from a large number of effects

w ∼ CN (0, σw ) (0.180)

where CN (µ, σ) denotes the Complex Circular Normal Distribution

with mean µ and deviation σ.

Due to the above, one typically refers to noise as additive white

Gaussian noise or AWGN, for short!

47
Received Signal and Signal-to-noise-ratio (SNR)
Notice that an immediate consequence of the Standard System Model of equa-
tion (0.178) is that signals in the noisy wireless channel look like noise,
i.e
y −−−−−→ CN (µ, σy ) (0.181)
CLT
• Mean
E[y] = E[h · s] + E[w] = 0 (0.182)
- This implies the assumption that h and s are independent!
• Variance

E[y · y ∗ ] = E[(h · s + w)(h · s + w)∗ ] = E[|h|2 ]E[|s|2 ] + E[|w|2 ]

= ΩEs + N0 (0.183)

- Notice the importance of the second order moment Ω of the chan-

nel envelope r , |h|!

Following previous notation, in equation (0.183) we have

implicitly defined Ω , E[|h|2 ] (Channel Power).

• The average (large-scale) signal-to-noise-ratio (SNR) associated with the

model of equation (0.178) is defined as

Ω · Es
γ̄ , (average SNR) (0.184)
N0
N0 Ω
from which we take the relationship = .
Es γ̄

It is very common to use N0 to denote average noise power.

• Likewise, an instantaneous (small-scale) SNR can also be defined as

r2 · Es
γ, (instantaneous SNR) (0.185)
N0
r
γΩ
from which we take the relationship r = .
γ̄

48
Adjusting Fading Models and to the System Model
Theorem: Consider a pair of random variables (r, γ) such that

x ∼ pX (x) (the PDF of x is known),

y = f (x) (a mapping between x and y is known).

Then
pY (y) = g 0 (y) · pX (g(y)) (0.186)
where g(y) = f −1 (x) is the functional inverse of f (x).

Example: We know that if x ∼ N (0, 1), then the variable y = σx, is

such that y ∼ N (0, σ). Why? Apply Theorem 1:
1 −x2 1 −g(y)2 1 −y 2
x ∼ √ e 2 −−−−−−1−→ y ∼ g 0 (y) √ e 2 = √ e 2σ2
2π g(y)= σ y 2π σ 2π

Homework 15: Referring to equation (0.185), prove that

p
pr ( Ωγ/γ̄)
pγ (γ) = p (Envelope to SNR PDF mapping)
2 γγ̄/Ω
(0.187)

Classic Fading PDFs in SNR

Let us now revise the most important fading channel models studied before
in light of Theorem 1 and equation (0.187).

• Rayleigh SNR channel model:

r −r22 1 −γ
p(r; σ) = 2
e 2σ −−−−−→ pγ (γ; γ̄) = e γ̄ (0.188)
σ γ̄

• Rice SNR channel model:

(1+K)r 2  q 
2(1+K)r −K−
p(r;K,Ω)= Ω e Ω I0 2r
K(1+K)
Ω −−−→
γ −K−(1+K) γγ̄
 r 
γ
pγ (γ; K, γ̄) = 2(1 + K) e I0 2 K(1+K) (0.189)
γ̄ γ̄

49
 • Hoyt SNR channel model:
(1+q 2 )2 r 2  
(1+q 2 )r − 4q 2 σ 2 I0 (1−q4 )r2 −−−→
p(r;q,σ)= qσ 2
e 4q 2 σ 2

2 2
(1 + q 2 ) − (1+q ) γ
(1 − q 4 ) γ
 
pγ (γ; q, γ̄) = e 4q 2 γ̄ I0 (0.190)
2qγ̄ 4q 2 γ̄

• Nakagami-m SNR channel model:

m 2  m γ
−m γ̄
2 m m −
p(r;m,Ω)= Γ(m) ( Ω ) r 2m−1 e Ω −−−→ pγ (γ; m, γ̄)
r
= 1
γΓ(m) m γγ̄ e
(0.191)

Moment Generating Functions of SNR Fading Models

In comparison, the MGFs associated with the SNR variations of the fading
models of interest are substantially simpler.

• MGF of Rayleigh SNR fading PDF

1
M (s) = = (1 − sγ̄)−1 (0.192)
1 − sγ̄

• MGF of Rice SNR fading PDF

 
1+K Ksγ̄
M (s) = exp (0.193)
1 + K − sγ̄ 1 + K − sγ̄

Notice that the Rice SNR MGF reduces to Rayleigh’s for K = 0.

• MGF of Hoyt SNR fading PDF

 2 !− 21
2sγ̄q
M (s) = 1 − 2sγ̄ + (0.194)
1 + q2

Notice that the Hoyt SNR MGF reduces to Rayleigh’s for q = 1.

• MGF of Nakagami SNR fading PDF

 sγ̄ −m
M (s) = 1 − (0.195)
m

Notice that the Nakagami SNR MGF reduces to Rayleigh’s for m = 1.

50
 Homework 16: What are the variance and the skewness of the
Nakagami SNR distribution?

Hint: Use the Nakagami MGF and equation (0.176)!

Moments of Classic SNR Fading PDFs

Now let us consider the moments of other SNR variations of the 4 main fading
channel models discussed above.

• Moments of Rayleigh SNR fading PDFs

E[γ n ] = Γ(1 + n)γ̄ n (0.196)

• Moments of Ricean SNR fading PDF

n
Γ(1 + n)
E[γ n ] = 1 F1 (−n; 1; −K)γ̄
n
(0.197)
1+K
where 1 F1 (a; b; z) is the Kummer confluent hypergeometric func-
tion.
• Moments of Hoyt SNR fading PDF
  2 
n 1−n −n 1−q 2
E[γ ] = Γ(1 + n) 2 F1 2 , 2 ; 1; 1+q2 γ̄ n (0.198)

where 2 F1 (a, b; c; z) is the Gauss hypergeometric function.

• Moments of Nakagami-m SNR fading PDF

Γ(m + n) n
E[γ n ] = γ̄ (0.199)
Γ(m)mn

51
Mandatory Reading
- Alouini [10]: Chapter 2 (Section 2 until 2.2.1.4)

Recommendend
- http://en.wikipedia.org/wiki/Moment-generating_function
- http://en.wikipedia.org/wiki/Cumulant

52
References
[1] D. Tse and P. Viswanath, Fundamentals of Wireless Communications.
Cambridge University Press, 2008.
[2] A. Goldsmith, Wireless Communications. Cambridge University Press,
2005.
[3] Gallager, http://www.youtube.com/watch?v=2DbwtCePzWg.

[4] T. S. Rappaport, Wireless Communications: Principles and Practice.

Prentice-Hall, 2002.
[5] Q. T. Zhang and S. H. Song, “Exact expression for the coherence bandwidth
of rayleigh fading channels,” IEEE Trans. Commun., vol. 55, no. 7, pp. 1296
– 1299, 2007.
[6] G. T. F. de Abreu, “On the moment-determinance and random mixture of
Nakagami-m variates,” IEEE Trans. Commun., vol. 58, no. 9, pp. 2561 –
2575, Sep. 2010.
[7] ——, “On the generation of Tikhonov variates,” IEEE Trans. Communi-
cations, vol. 56, no. 7, pp. 1157–1168, Jul. 2008.
[8] J. G. Proakis, Digital Communications, Fourth Edition. New York, NY:
Mc-Graw-Hill, 2000.
[9] S. Blinnikov and R. Moessner, “Expansions for nearly gaussian distribu-
tions,” Astron. Astrophys. Suppl. Ser., no. 130, pp. 193–205, MAY 1998.
[10] M. K. Simon and M.-S. Alouini, Digital Communication over Fading Chan-
nels: A Unified Approach to Performance Analysis. New York, NY: Wiley,
2000.

53