Winter 2023 ECSE-610 1

3. Statistical Multipath
Channel Models

Prof. Benoit Champagne

Dept. of Electrical & Computer Engineering

McGill University, Montreal, Canada

January 19, 2023

Winter 2023 ECSE-610 2

Introduction
▪ In this chapter, we examine RX power variations due to
constructive and destructive addition of multipath components.
▪ These variations occur over very short distances, on the order of
the wavelength, and are referred to as small-scale fading.
▪ In practice, deterministic channel models are not available and
we must characterize multipath channels statistically, i.e., as a
random time-varying impulse response.
▪ The delay spread 𝑇𝑚 refers to the time delay between the
arrival of the first and the last multipath signal components
associated with a single transmitted pulse.
▪ We need to consider two cases separately depending on the
bandwidth 𝐵 of the information signal:
▪ If 𝑇𝑚 < 1/𝐵 ⟹ narrowband fading model.
▪ If 𝑇𝑚 > 1/𝐵 ⟹ wideband fading model
Winter 2023 ECSE-610 3

Overview
3.1 Time-varying Channel Impulse Response
3.2 Narrowband Fading Model
3.3 Wideband Fading Model
Winter 2023 ECSE-610 4

3.1 Time-varying channel

impulse response
Winter 2023 ECSE-610 5

Time-invariant (TI) multipath channel

▪ Multipath scenario:

▪ TX signal: 𝑠 𝑡 = Re 𝑢 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡
▪ RX signal: 𝑟 𝑡 = Re 𝑣 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡 + noise
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑒 𝑗𝛽𝑛 𝑒 −𝑗2𝜋𝑓𝑐 𝜏𝑛 𝑢(𝑡 − 𝜏𝑛 )

𝑛=0
▪ Main parameters:
▪ 𝑁: number of paths (𝑛 = 0 for LOS)
▪ 𝛼𝑛 : amplitude (0 < 𝛼𝑛 < 1)
▪ 𝛽𝑛 : phase shift (0 ≤ 𝛽𝑛 < 2𝜋)
▪ 𝜏𝑛 : propagation delay (𝜏𝑛 = 𝑑𝑛 /𝑐; 𝜏0 ≤ 𝜏1 ≤ ⋯ ≤ 𝜏𝑁 )
▪ In the sequel, we ignore noise term and focus on low-pass
equivalent channel model between 𝑢 𝑡 and 𝑣(𝑡).
Winter 2023 ECSE-610 6

Impulse response in TI case

▪ The relationship between 𝑣(𝑡) and 𝑢(𝑡) can be expressed as a
convolution integral:
∞
𝑣 𝑡 = 𝑐 𝑡 ∗ 𝑢 𝑡 = න 𝑐 𝜏 𝑢 𝑡 − 𝜏 𝑑𝜏
0
𝑁

𝑐 𝜏 = ෍ 𝛼𝑛 𝑒 𝑗𝛽𝑛 𝑒 −𝑗2𝜋𝑓𝑐 𝜏𝑛 𝛿(𝜏 − 𝜏𝑛 )

𝑛=0
▪ We refer to 𝑐(𝜏) as the channel impulse response:
▪ 𝑐 𝜏 gives the channel response at time 𝑡 = 𝜏 to an impulse applied
at time 𝑡 = 0: 𝑢 𝑡 = 𝛿 𝑡 ⟹ 𝑣 𝜏 = 𝑐(𝜏).

▪ Equivalently, 𝑐(𝜏) gives the response at any time 𝑡 to an impulse

applied at time 𝑡 − 𝜏.
Winter 2023 ECSE-610 7

Time-varying (TV) multipath channel

▪ In practice, radio channel is not TI:
▪ Moving RX or TX.
▪ Moving objects (cars, people, etc.) along signal path.
▪ Hence, the amplitude, phase and delay of each path are TV:
▪ 𝛼𝑛 ⇒ 𝛼𝑛 (𝑡), 𝛽𝑛 ⇒ 𝛽𝑛 𝑡 , 𝜏𝑛 ⇒ 𝜏𝑛 (𝑡).
▪ Even the number of paths could change over time: 𝑁 ⇒ 𝑁(𝑡)
▪ Low-pass equivalent multipath TV channel model:
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 (𝑡)𝑒 𝑗𝛽𝑛 𝑡
𝑒 −𝑗2𝜋𝑓𝑐 𝜏𝑛 𝑡
𝑢(𝑡 − 𝜏𝑛 (𝑡))
𝑛=0
𝑁

= ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 𝑡
𝑢 𝑡 − 𝜏𝑛 𝑡
𝑛=0
where we define the compound phase term:
𝜙𝑛 𝑡 = 2𝜋𝑓𝑐 𝜏𝑛 𝑡 − 𝛽𝑛 (𝑡)
Winter 2023 ECSE-610 8

Note on Doppler effect

▪ Doppler is due to TV nature of delay, i.e. 𝜏ሶ 𝑛 𝑡 ≠ 0.
▪ To explain this, consider simplified scenario:
▪ Single path (drop 𝑛)
▪ Constant gain 𝛼 and phase 𝛽
▪ TV delay 𝜏 𝑡 due to RX motion
▪ In this case, the I/O relation is
𝑣 𝑡 = 𝛼𝑒 𝑗𝛽 𝑒 −𝑗2𝜋𝑓𝑐𝜏 𝑡
𝑢(𝑡 − 𝜏 𝑡 )
▪ Assume RX moving at constant speed 𝑣 and angle 𝜃:
𝑣𝑡 cos 𝜃
▪𝜏 𝑡 =𝜏 0 − .
𝑐
▪ 𝜏 0 = 𝑑0 /𝑐 where 𝑑0 is the path length at 𝑡 = 0
▪ 𝑣𝑡 cos 𝜃 = variation in path length due to RX motion
Winter 2023 ECSE-610 9

Doppler effect (cont’d)

▪ It is convenient to define the Doppler phase shift:
𝜙𝐷 𝑡 = 2𝜋𝑓𝑐 𝜏 0 − 𝜏 𝑡 = 2𝜋𝑓𝐷 𝑡
𝑣 𝑣
𝑓𝐷 = 𝑓𝑐 cos 𝜃 = cos 𝜃
𝑐 𝜆
▪ Then we have for the carrier component:
𝑗 2𝜋𝑓 𝑡−𝜏 0 +𝜙 𝑡
𝑒 𝑗2𝜋𝑓𝑐(𝑡−𝜏 𝑡 ) =𝑒 𝑐 𝐷

= 𝑒 −𝑗2𝜋𝑓𝑐𝜏(0) 𝑒 −𝑗2𝜋 𝑓𝑐+𝑓𝐷 𝑡

▪ Note the inconsistency in eq. (3.2) of textbook.

▪ More generally, we define the instantaneous Doppler
frequency as:
1 𝑑𝜙𝐷 (𝑡) 𝑑𝜏(𝑡)
𝑓𝐷 𝑡 = = −𝑓𝑐
2𝜋 𝑑𝑡 𝑑𝑡
Winter 2023 ECSE-610 10

TV impulse response
▪ Consider the I/O relation for the TV channel:
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 𝑡 𝑢 𝑡 − 𝜏𝑛 𝑡
𝑛=0
𝜙𝑛 𝑡 = 2𝜋𝑓𝑐 𝜏𝑛 0 − 𝜙𝐷,𝑛 𝑡 − 𝛽𝑛 (𝑡)
fixed delay Doppler Others

▪ In analogy with the TI case, we can rewrite this as

∞
𝑣 𝑡 = න 𝑐 𝜏, 𝑡 𝑢 𝑡 − 𝜏 𝑑𝜏
0
𝑁

𝑐 𝜏, 𝑡 = ෍ 𝛼𝑛 (𝑡)𝑒 −𝑗𝜙𝑛 𝑡
𝛿(𝜏 − 𝜏𝑛 (𝑡))
𝑛=0
▪ We refer to 𝑐(𝜏, 𝑡) as the TV channel impulse response:
▪ Response at time 𝑡 to impulse applied at 𝑡 − 𝜏.
Winter 2023 ECSE-610 11

Delay spread
▪ Delay spread basically refers to the time delay between
the arrival of the first and the last signal components.
▪ There are several different measures of delay spread in
the literature:
▪ The delay spread is typically measured relative to the signal
component to which the RX is synchronized.
▪ Definitions of delay spread based on power delay profile
include the average and rms delay spread (Section 3.3.1).
▪ For the sake of discussion, simply let 𝑇𝑚 = 𝜏𝑁 − 𝜏0
▪ Some typical values (Note: 1 foot ≡ 1ns, 1000 feet ≡ 1𝜇s):
▪ Indoor: 10-100ns;
▪ Outdoor urban: 100ns-1ms;
▪ Outdoor suburban: 1-10𝜇s or more.
Winter 2023 ECSE-610 12

Channel classification
▪ Let 𝐵 denote the bandwidth of information signal 𝑢(𝑡).
▪ We say that two multipath components with delays 𝜏𝑖 and 𝜏𝑗
are non-resolvable if 𝜏𝑖 − 𝜏𝑗 < 𝐵 −1
▪ In this case, we have 𝑢(𝑡 − 𝜏𝑖 ) ≅ 𝑢(𝑡 − 𝜏𝑗 )
▪ Otherwise, the components are said to be resolvable.
▪ The impact of multipath on the RX signal depends on the size
of 𝑇𝑚 relative to inverse signal bandwidth 𝐵−1 :
▪ If 𝑇𝑚 < 𝐵 −1 , the LOS and multipath components are non-resolvable
⟹ narrowband fading model (Section 3.2).
▪ If 𝑇𝑚 > 𝐵 −1 , the multipath components are resolvable into several
discrete components ⟹ wideband fading model (Section 3.3).
▪ The channel frequency response tends to remain constant over
frequency separations of 𝛥𝑓 = 𝑇𝑚−1 , hence:
▪ For narrowband fading, we have 𝐵 < Δ𝑓 ⟹ frequency flat
▪ For wideband fading. we have 𝐵 > Δ𝑓 ⟹frequency selective
Winter 2023 ECSE-610 13

3.2 Narrowband Fading Model

Winter 2023 ECSE-610 14

TV multipath channel
▪ Low-pass equivalent I/O relation:
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 𝑡 𝑢 𝑡 − 𝜏𝑛 𝑡
𝑛=0
𝜙𝑛 𝑡 = 2𝜋𝑓𝑐 𝜏𝑛 0 − 𝜙𝐷,𝑛 𝑡 − 𝛽𝑛 (𝑡)
▪ Simplified Doppler model:
𝑣
𝜙𝐷,𝑛 𝑡 = 2𝜋𝑓𝐷,𝑛 𝑡, 𝑓𝐷,𝑛 = cos 𝜃𝑛
𝜆
▪ TV impulse response:
∞
𝑣 𝑡 = න 𝑐 𝜏, 𝑡 𝑢 𝑡 − 𝜏 𝑑𝜏
0
𝑁

𝑐 𝜏, 𝑡 = ෍ 𝛼𝑛 (𝑡)𝑒 −𝑗𝜙𝑛 𝑡 𝛿(𝜏 − 𝜏𝑛 (𝑡))

𝑛=0
Winter 2023 ECSE-610 15

Narrowband (NB) assumption

▪ This occurs when the delay spread is small compared to
inverse signal bandwidth, i.e.:

▪ In this case, the multipath components are not resolvable:

𝑢 𝑡 − 𝜏𝑖 𝑡 ≅ 𝑢 𝑡 − 𝜏0 𝑡 , ∀𝑖 ∈ {1, ⋯ , 𝑁}
▪ Convenient to redefine 𝑢(𝑡) ≡ 𝑢(𝑡 − 𝜏𝑜 𝑡 ) so that 𝜏0 𝑡 = 0 (i.e.,
we consider normalized delays relative to 𝜏0 𝑡 ).
▪ Received signal under narrowband assumption
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 𝑡 𝑢(𝑡)
𝑛=0

TV complex gain
Winter 2023 ECSE-610 16

RX signal model (cont’d)

▪ For simplicity we assume 𝑢 𝑡 = 1:
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 (𝑡)
𝑛=0
▪ Band-pass RX signal in real form:
𝑟 𝑡 = Re 𝑣 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡
= 𝑟𝐼 𝑡 cos 2𝜋𝑓𝑐 𝑡 − 𝑟𝑄 𝑡 sin(2𝜋𝑓𝑐 𝑡)
where
𝑁

𝑟𝐼 𝑡 = Re 𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 cos 𝜙𝑛 (𝑡)
𝑛=0
𝑁

𝑟𝑄 𝑡 = Im 𝑣 𝑡 = − ෍ 𝛼𝑛 𝑡 sin 𝜙𝑛 (𝑡)
𝑛=0
Winter 2023 ECSE-610 17

3.2.1 Second-order Characterization

Winter 2023 ECSE-610 18

Stochastic modeling
▪ In practice, the exact form of the path gains 𝛼𝑛 (𝑡) and
phases 𝜙𝑛 𝑡 is not accessible by measurement and
considered to be unknown.
▪ Even the number of path 𝑁 may not be known exactly.
▪ For Doppler, the path angle of arrival 𝜃𝑛 may be unknown.
▪ To overcome this difficulty in the analysis and design of
wireless systems, we model 𝛼𝑛 (𝑡) and 𝜙𝑛 (𝑡) as slowly-
varying stochastic (or random) processes.
▪ To this end, we still need to make several assumptions
whose validity has been confirmed experimentally for a
variety of (but not all) propagation environments.
▪ These assumptions are presented and discussed below.
Winter 2023 ECSE-610 19

Independence assumptions
▪ The quantities of interest are the path gains 𝛼𝑛 (𝑡) and path
phases 𝜙𝑛 (𝑡), for 𝑛 ∈ {0, ⋯ , 𝑁}
▪ Path independence: we assume that for different paths, i.e. for
any 𝑛 ≠ 𝑚, the processes {𝛼𝑛 𝑡 , 𝜙𝑛 (𝑡)} and {𝛼𝑚 𝑡 , 𝜙𝑚 (𝑡)} are
statistically independent.
▪ Gain and phase independence: we assume that for a given
path with index 𝑛, 𝛼𝑛 (𝑡) and 𝜙𝑛 (𝑡) are independent.
▪ These assumptions considerably simplify the calculation of
various moments, e.g.:
▪ 𝐸 𝛼𝑛 𝑡 𝜙𝑚 𝑡 = 𝐸 𝛼𝑛 𝑡 𝐸[𝜙𝑚 (𝑡)] for 𝑛 ≠ 𝑚
Winter 2023 ECSE-610 20

Gaussian model
▪ Recall expressions for RX signal I and Q components:
𝑁 𝑁

𝑟𝐼 𝑡 = ෍ 𝛼𝑛 𝑡 cos 𝜙𝑛 (𝑡) , 𝑟𝑄 𝑡 = − ෍ 𝛼𝑛 𝑡 sin 𝜙𝑛 (𝑡)

𝑛=0 𝑛=0
▪ We note that if 𝑁 ≫ 1, then 𝑟𝐼 (𝑡) and 𝑟𝑄 (𝑡) are obtained as the
sum of a large number of independent random variables.
▪ Hence, we can invoke the central limit theorem and claim that
the probability distribution of 𝑟𝐼 (𝑡) and 𝑟𝑄 (𝑡) should be
approximately Gaussian (normal).
▪ More generally, we shall assume that 𝑟𝐼 (𝑡) and 𝑟𝑄 (𝑡) are jointly
Gaussian random processes.
▪ Very convenient: knowledge of first and second moments (mean
and correlation functions) provides complete characterization.
▪ The Gaussian model leads to so-called Rayleigh fading model.
▪ This assumption is not valid if there is dominant LOS path: in this
case, Rician or Nakagami fading models more appropriate...
Winter 2023 ECSE-610 21

Additional assumptions
▪ Constant gain: 𝛼𝑛 𝑡 = 𝛼𝑛
▪ 𝛼𝑛 is a non-negative random variable.

▪ Linearly TV phase: 𝜙𝑛 𝑡 = 𝑎𝑛 − 𝑏𝑛 𝑡
▪ 𝑎𝑛 = (2𝜋𝑓𝑐 𝜏𝑛 0 − 𝛽𝑛 ) modulo 2𝜋 is uniform in 0,2𝜋 .
▪ 𝑏𝑛 = 2𝜋𝑓𝐷,𝑛 where 𝑓𝐷,𝑛 is a random Doppler shift.
▪ 𝑎𝑛 and 𝑏𝑛 are independent.
Winter 2023 ECSE-610 22

Calculations of 1st and 2nd order moments

▪ In class, we shall prove the following results about the
random process 𝑟𝐼 (𝑡) and 𝑟𝑄 (𝑡).
▪ Property 1: 𝑟𝐼 𝑡 and 𝑟𝑄 (𝑡) have zero-mean, i.e.,
𝐸 𝑟𝐼 𝑡 = 𝐸 𝑟𝑄 𝑡 =0
▪ Property 2a: 𝑟𝐼 𝑡 and 𝑟𝑄 (𝑡) are jointly wide sense
stationary (WSS) with auto-correlation functions:
1
𝐸 𝑟𝐼 𝑡 𝑟𝐼 𝑡 + 𝜏 = ෍ 𝐸 𝛼𝑛2 𝐸[cos(2𝜋𝑓𝐷,𝑛 𝜏)] ≜ 𝐴𝑟𝐼 (𝜏)
2
𝑛
1
𝐸 𝑟𝑄 𝑡 𝑟𝑄 𝑡 + 𝜏 = ෍ 𝐸 𝛼𝑛2 𝐸[cos(2𝜋𝑓𝐷,𝑛 𝜏)] ≜ 𝐴𝑟𝑄 (𝜏)
2
𝑛
Winter 2023 ECSE-610 23

Calculations (cont’d)
▪ Property 2b: The cross-correlation function of 𝑟𝐼 𝑡 and
𝑟𝑄 (𝑡) is given by
1
𝐸 𝑟𝐼 𝑡 𝑟𝑄 𝑡 + 𝜏 = ෍ 𝐸 𝛼𝑛2 𝐸[sin(2𝜋𝑓𝐷,𝑛 𝜏)] ≜ 𝐴𝑟𝐼 𝑟𝑄 (𝜏)
2
𝑛
from which we deduce,
𝐸 𝑟𝑄 𝑡 𝑟𝐼 𝑡 + 𝜏 ≜ 𝐴𝑟𝑄𝑟𝐼 𝜏 = 𝐴𝑟𝐼 𝑟𝑄 −𝜏 = −𝐴𝑟𝐼 𝑟𝑄 (𝜏)
▪ Property 3: The received signal
𝑟 𝑡 = 𝑟𝐼 𝑡 cos 2𝜋𝑓𝑐 𝑡 − 𝑟𝑄 𝑡 sin(2𝜋𝑓𝑐 𝑡)
is WSS with zero-mean and autocorrelation
𝐴𝑟 𝜏 = 𝐴𝑟𝐼 𝜏 cos 2𝜋𝑓𝑐 𝜏 − 𝐴𝑟𝑄 𝜏 sin(2𝜋𝑓𝑐 𝜏)
Winter 2023 ECSE-610 24

Uniform scattering model

▪ This model, also referred to as the Clarke-Jakes model,
assumes that a large number of multipaths reach the RX from
nearly every direction with nearly equal power:
▪ In our probabilistic model, the same
“effect” can be obtained by assuming
uniformly distributed arrival angles in
the Doppler model:
▪ 𝑓𝐷,𝑛 = (𝑣Τ𝜆) cos 𝜃𝑛 .
▪ 𝜃𝑛 uniform in [0,2𝜋).

▪ With this new assumption, we can evaluate

the auto- and cross-correlation functions
𝐴𝑟𝐼 (𝜏) and 𝐴𝑟𝐼𝑟𝑄 (𝜏) in closed forms:
▪ Class derivation…
Winter 2023 ECSE-610 25

Main points
▪ Under the preceding assumptions for the narrowband TV
multipath signal model with uniform scattering:
▪ 𝑟𝐼 𝑡 and 𝑟𝑄 (𝑡) are independent WSS Gaussian with zero mean.
▪ Their common autocorrelation function is
𝐴𝑟𝐼 𝜏 = 𝐴𝑟𝑄 𝜏 = 𝑃𝑟 𝐽0 (2𝜋𝑓𝐷 𝜏)
where 𝑓𝐷 = 𝑣/𝜆 is the maximum Doppler shift.
▪ Properties of the autocorrelation:
▪ 𝐴𝑟𝐼 (𝜏) ⟶ 0 as 𝜏 ⟶ ∞.
▪ The first zero occurs when 𝑓𝐷 𝜏 ≅ 0.4.
▪ Signal decorrelates over distance of first zero at
𝑓𝐷 𝜏 = 0.4
𝑣𝜏 ≅ 0.4𝜆 under the uniform assumption
for arrival angle 𝜃𝑛 .
▪ This is often used as a rule of thumb in
system design.
Winter 2023 ECSE-610 26

Power spectral density (PSD)

▪ The PSD is defined as the Fourier transform of the
autocorrelation function:
∞

𝑆𝑟𝐼 𝑓 = න 𝐴𝑟𝐼 𝜏 𝑒 −𝑗2𝜋𝑓𝜏 𝑑𝜏

−∞
▪ It represents the average distribution of power in random
process 𝑟𝐼 (𝑡) as a function of frequency 𝑓.
▪ For the narrowband TV channel with uniform scattering:
∞

𝑆𝑟𝐼 𝑓 = 𝑃𝑟 න 𝐽0 2𝜋𝑓𝐷 𝜏 𝑒 −𝑗2𝜋𝑓𝜏 𝑑𝜏

−∞
𝑃𝑟 1
, if 𝑓 < 𝑓𝐷
= ൞𝜋𝑓𝐷 1 − 𝑓Τ𝑓𝐷 2

0, else
Winter 2023 ECSE-610 27

Power spectral density (cont’d)

▪ The PSD for the narrowband TV multipath model with uniform
scattering, also known as Doppler spectrum, is plotted below:
▪ Observations:
▪ The Doppler spectrum is
bandlimited to 𝑓 < 𝑓𝐷
▪ Under the uniform assumption
for 𝜃𝑛 , this PSD is equal to the
PDF of the random Doppler
frequency 𝑓𝐷,𝑛 = (𝑣Τ𝜆) cos 𝜃𝑛 .
▪ The singularities at 𝑓 = ±𝑓𝐷 are
a consequence of the special
cos 𝜃𝑛 mapping.
▪ Measured Doppler spectra do
not exhibit such singularities.
Winter 2023 ECSE-610 28

3.2.2 Enveloppe and Power

Distributions
Winter 2023 ECSE-610 29

Narrowband TV multipath model (Recap)

▪ Received signal: 𝑟 𝑡 = Re 𝑣 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 (𝑡) = 𝑟𝐼 𝑡 + 𝑗𝑟𝑄 (𝑡)

𝑛=0
▪ Under previous assumptions of large number of independent
random paths with no dominant LOS, we found that:
▪ 𝑟𝐼 (𝑡) and 𝑟𝑄 𝑡 are WSS Gaussian processes with zero mean
2 𝐸[cos(2𝜋𝑓 𝜏) ]
▪ Autocorrelation: 𝐴𝑟𝐼 𝜏 = 𝐴𝑟𝑄 𝜏 = 12 σ𝑛 𝐸 𝛼𝑛 𝐷,𝑛
2 𝐸[sin(2𝜋𝑓 𝜏) ]
▪ Cross-correlation: 𝐴𝑟𝐼 𝑟𝑄 𝜏 = 12 σ𝑛 𝐸 𝛼𝑛 𝐷,𝑛
1 2 ,𝐴
▪ We note that 𝐴𝑟𝐼 0 = 𝐴𝑟𝑄 0 = 𝑃𝑟 ≜ 2 σ𝑛 𝐸 𝛼𝑛 𝑟𝐼 𝑟𝑄 𝜏 = 0.
▪ Hence, at any given time 𝑡, random variables 𝑟𝐼 (𝑡) and 𝑟𝑄 𝑡 are
independent Gaussian with zero mean and common variance
𝑁
1
𝜎 2 = 𝑃𝑟 = ෍ 𝐸[𝛼𝑛2 ]
2
𝑛=0
Winter 2023 ECSE-610 30

Envelope and phase distribution

▪ Let us express 𝑣 𝑡 = 𝑟𝐼 𝑡 + 𝑗𝑟𝑄 (𝑡) in polar form, i.e. 𝑣 𝑡 =
𝑧 𝑡 𝑒 𝑗𝜓(𝑡) , where we define:

▪ Envelope: 𝑧 𝑡 = 𝑟𝐼 𝑡 2 + 𝑟𝑄 𝑡 2 ≥0

▪ Phase: 𝜓 𝑡 = arctan 𝑟𝐼 𝑡 , 𝑟𝑄 𝑡 ∈ [0,2𝜋)

▪ Applying transformation theorem in 2D for pairs of random
variables, we can show the following (try it…):
▪ 𝑧(𝑡) and 𝜁(𝑡) are independent
▪ 𝜓 𝑡 is uniformly distributed in [0,2𝜋)
▪ 𝑧(𝑡) is Rayleigh distributed:
𝑥 −𝑥 2 Τ2𝜎2
𝑒 , 𝑥≥0
𝑝𝑧 𝑡 𝑥 = ൝𝜎 2
0, else
Winter 2023 ECSE-610 31

Power distribution
▪ The instantaneous power is defined as
𝜍 𝑡 = 𝑣 𝑡 2 =𝑧 𝑡 2

▪ Applying transformation theorem in 1D, we find that the

PDF of 𝜍 𝑡 is exponential with parameter 2𝜎 2 :
1 −𝑥Τ2𝜎 2
2𝑒 , 𝑥≥0
𝑝𝜍 𝑡 𝑥 = ൝2𝜎
0, else
▪ The exponential distribution will be used extensively in the
study of receiver diversity in Chapter 7.
Winter 2023 ECSE-610 32

Rician distribution
▪ In the presence of a strong LOS component, the Rayleigh
distribution for the envelope is not applicable.
▪ In this case, the RX signal is the superposition of a complex
Gaussian component and a deterministic LOS component:
𝑁

𝑣 𝑡 = 𝛼0 𝑡 𝑒 −𝑗𝜙0 (𝑡) + ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 (𝑡)

𝑛=1
▪ The components 𝑟𝐼 (𝑡) and 𝑟𝑄 (𝑡) are still independent
Gaussian but they have not zero mean.
▪ It can be shown that:
▪ 𝜓 𝑡 is not uniformly distributed
▪ 𝑧(𝑡) is Rician distributed:
𝑥 − 𝑥2 + 𝑠2 𝑥𝑠
𝑝𝑧 𝑡 𝑥 = 2 exp 𝐼0 , 𝑥≥0
𝜎 2𝜎 2 𝜎 2

▪ 𝑠 2 = 𝛼02 (𝑡) is the power in the LOS (at time 𝑡).

Winter 2023 ECSE-610 33

3.3 Wideband Fading Model

Winter 2023 ECSE-610 34

Narrowband fading model (review)

▪ Let 𝐵 denote the input signal bandwidth:
𝐵
𝑢 𝑡 ≅ න 𝑈 𝑓 𝑒 −𝑗2𝜋𝑓𝑡 𝑑𝑓
−𝐵
▪ Let 𝑇𝑚 = max 𝜏𝑁 𝑡 − 𝜏0 (𝑡) denote the delay spread.
𝑡
▪ We refer to the fading as narrowband if: 𝑇𝑚 < 𝐵 −1
▪ In this case, the multipath signal images are non-resolvable:
𝑢(𝑡 − 𝑡𝑖 𝑡 ) ≅ 𝑢(𝑡 − 𝑡0 𝑡 ) ≡ 𝑢(𝑡)
Winter 2023 ECSE-610 35

Wideband fading model

▪ We refer to the fading as wideband if: 𝑇𝑚 > 𝐵 −1 .
▪ In this case, multipath images with |𝜏𝑖 (𝑡) − 𝜏𝑗 (𝑡)| > 𝐵 −1
are said to be resolvable, i.e.
𝑢 𝑡 − 𝑡𝑖 𝑡 ≠ 𝑢(𝑡 − 𝑡𝑗 𝑡 )
▪ The RX signal is subject to intersymbol interference (ISI):
▪ Information pulses reaching the RX through multipath with large 𝜏𝑖 (𝑡)
interfere with subsequently transmitted pulses (dashed pulses in figure).
Winter 2023 ECSE-610 36

Discussion
▪ Multipath TV model (recap):
𝑠 𝑡 = Re 𝑢 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡 𝑟 𝑡 = Re 𝑣 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡 + noise
𝑁

𝑣 𝑡 = ෍ 𝛼𝑛 𝑡 𝑒 −𝑗𝜙𝑛 𝑡 𝑢 𝑡 − 𝜏𝑛 𝑡 = 𝑟𝐼 𝑡 + 𝑗𝑟𝑄 (𝑡)

𝑛=0
▪ If 𝑁 is large and the phase of each component is uniformly
distributed, then the RX signal will still be a zero-mean complex
Gaussian process with Rayleigh envelope.
▪ However, in the wideband case, the images may be resolvable
and we cannot assume that 𝑢 𝑡 − 𝜏𝑛 𝑡 ≅ 𝑢 𝑡 for all 𝑛.
▪ Hence, our previous analysis for narrowband fading channel
with uniform scattering is no longer valid.
▪ Characterization of wideband fading must take into account the
effects of TV channel parameters and delay spread.
Winter 2023 ECSE-610 37

TV impulse response
▪ The starting point for studying wideband channels is the equivalent
lowpass channel impulse response 𝑐 𝜏, 𝑡 ∈ ℂ:
∞
𝑣 𝑡 = න 𝑐 𝜏, 𝑡 𝑢 𝑡 − 𝜏 𝑑𝜏
0
▪ Two different types of model exist for multipath TV channels:
▪ Discrete (as a sum over path):
𝑁

𝑐 𝜏, 𝑡 = ෍ 𝛼𝑛 (𝑡)𝑒 −𝑗𝜙𝑛 𝑡 𝛿(𝜏 − 𝜏𝑛 (𝑡))

𝑛=0
▪ Continuous (as an integral over delay):
∞
𝑐 𝜏, 𝑡 = න 𝛼(𝑡, 𝜏) 𝑒 −𝑗𝜙 𝑡,𝜏
𝑑𝜏
0
▪ In general, the TV channel impulse response 𝑐(𝜏, 𝑡) is random
because of the random amplitudes, phases, and delays.
▪ Consequently, we adopt a statistical characterization of 𝑐(𝜏, 𝑡) in the
study of wideband wireless systems, i.e., as a random process.
Winter 2023 ECSE-610 38

Statistical characterization of 𝒄(𝝉, 𝒕)

▪ If the number of independent multipath components is large
with no dominant LOS, we can invoke the central limit theorem
to assume that 𝑐(𝜏, 𝑡) is a complex Gaussian process.
▪ Under this assumption, a complete statistical characterization of
𝑐(𝜏, 𝑡) is provided by its first and second order moments.
▪ If we further assume that the phase of each component
𝜙𝑛 (𝑡)~𝑈[0,2𝜋), then the real and imaginary parts or
𝑐 𝜏, 𝑡 = 𝑐𝐼 𝜏, 𝑡 + 𝑗𝑐𝑄 (𝜏, 𝑡)
are independent Gaussian processes with zero-mean, same
autocorrelation, and zero cross-correlation:
▪ 𝐸 𝑐𝐼 𝜏, 𝑡 = 𝐸 𝑐𝑄 𝜏, 𝑡 =0
▪ 𝐸 𝑐𝐼 𝜏, 𝑡 𝑐𝐼 𝜏 ′ , 𝑡 ′ = 𝐸 𝑐𝑄 𝜏, 𝑡 𝑐𝑄 𝜏 ′ , 𝑡 ′ ≠0
▪ 𝐸 𝑐𝐼 𝜏, 𝑡 𝑐𝑄 𝜏 ′ , 𝑡 ′ =0
Winter 2023 ECSE-610 39

Characterization of 𝒄 𝝉, 𝒕 (cont’d)
▪ A Gaussian process 𝑐(𝜏, 𝑡) whose real and imaginary parts
𝑐𝐼 (𝜏, 𝑡) and 𝑐𝑄 (𝜏, 𝑡) obey the above correlation properties is
referred to as a complex circular Gaussian process.
▪ For this this type of process, it can be shown that
𝐸 𝑐 𝜏, 𝑡 𝑐 𝜏′, 𝑡′ =0
▪ Hence, knowledge of the following autocorrelation function
𝐴𝑐 𝜏, 𝜏′, 𝑡, 𝑡′ ≜ 𝐸[𝑐 ∗ 𝜏, 𝑡 𝑐(𝜏′, 𝑡′)]
is sufficient to fully characterize 𝑐 𝜏, 𝑡 .
▪ In practice, this function of 4 parameters reduces to a simpler
function of 2 parameters, as explained below.
Winter 2023 ECSE-610 40

Wide sense stationarity

▪ To simplify analysis, most wideband models assume that the
joint statistics of a channel measured at two different times, say
𝑡 and 𝑡′, depend only on the difference Δ𝑡 = 𝑡′ − 𝑡.
▪ In the case of the autocorrelation function, this means that
𝐴𝑐 𝜏, 𝜏′, 𝑡, 𝑡′ = 𝐴𝑐 (𝜏, 𝜏′, Δ𝑡)
▪ Note: we used the same notation, 𝐴𝑐 (. ), on both sides of the
equality, although formally, these are different functions.
▪ A wideband channel whose autocorrelation satisfies this
condition (along with 𝐸 𝑐 𝜏, 𝑡 = 0 and 𝐸 𝑐 𝜏, 𝑡 𝑐 𝜏′, 𝑡′ = 0), is
referred to as wide sense stationary (WSS).
Winter 2023 ECSE-610 41

Uncorrelated scattering
▪ In practice, the complex channel gains associated to resolvable
multipath components with delays 𝜏 and 𝜏 ′ , where |𝜏 ′ − 𝜏| >
𝐵 −1 , tend to be uncorrelated.
▪ This is plausible since the two components are likely caused by
different scatterers, resulting in independent gain and phase
parameters.
▪ We say that such a channel has uncorrelated scattering (US)
and denote channels that are both WSS and US as WSSUS.
▪ In this case, the autocorrelation function further simplifies to
𝐴𝑐 𝜏, 𝜏′, 𝑡, 𝑡′ = 𝐴𝑐 𝜏, Δ𝑡 𝛿(Δ𝜏)
where Δ𝜏 = 𝜏′ − 𝜏 and 𝛿 ⋅ is the Dirac delta function.
▪ 𝐴𝑐 𝜏, Δ𝑡 provides a measure of the correlation between channel
responses with same 𝜏 but separated by Δ𝑡 in time.
▪ This model is idealized: in reality, we should interpret 𝛿(⋅) as a
pulse of short duration, i.e. < 𝐵−1 , and unit area.
Winter 2023 ECSE-610 42

Scattering Function
▪ In practice, we find that the variations of 𝐴𝑐 𝜏; Δ𝑡 with respect
to (wrt) its time arguments exhibit different behaviors:
▪ Variations wrt to 𝜏 occur over short intervals, on the order of
the delay spread (e.g. from 10ns to 100𝜇s).
▪ Variations wrt to Δ𝑡, resulting from RX or TX motion, i.e.
Doppler, occur over longer time intervals (from 100𝜇s to 1s).
▪ The scattering function for a WSSUS channel is defined as
the Fourier transform of 𝐴𝑐 𝜏; Δ𝑡 with respect to the time Δ𝑡:
∞
𝑆𝑐 𝜏, 𝜌 = න 𝐴𝑐 𝜏, Δ𝑡 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
∞
▪ We refer to 𝜌 as a Doppler frequency.
▪ 𝑆𝑐 𝜏, 𝜌 characterizes the average output power associated with the
channel as a function of the multipath delay 𝜏 and Doppler 𝜌.
Winter 2023 ECSE-610 43

Remarks
▪ The most important characteristics of the wideband
WSSUS channel are derived from its autocorrelation
function 𝐴𝑐 𝜏, Δ𝑡 and its scattering function 𝑆𝑐 𝜏, 𝜌
▪ This includes the following:
▪ Power delay profile.
▪ Delay spread.
▪ Coherence bandwidth.
▪ Doppler power spectrum.
▪ Coherence time.

▪ These characteristics are described below.

Winter 2023 ECSE-610 44

3.3.1 Power Delay Profile

Winter 2023 ECSE-610 45

Definition
▪ The power delay profile 𝐴𝑐 𝜏 , also called multipath intensity
profile, is defined as the autocorrelation 𝐴𝑐 (𝜏, Δt) with Δ𝑡 = 0:
𝐴𝑐 𝜏 = 𝐴𝑐 (𝜏, 0)
▪ In practice, 𝐴𝑐 (𝜏) is non-negative and represents the average
power associated with a given multipath delay.
▪ The power delay profile is easily measured empirically and very
useful in system design (see next slide).
Winter 2023 ECSE-610 46

Average and RMS delay spread

▪ Earlier, we defined the delay spread as the time difference
between first (LOS) and last arrivals, i.e. 𝑇𝑚 = 𝜏𝑁 − 𝜏0
▪ In practice, measures of delay spread based on the power
delay profile are more useful; this includes:
∞
‫׬‬0 𝜏𝐴𝑐 𝜏 𝑑𝜏
▪ Average delay spread: 𝜇 𝑇𝑚 = ∞
‫׬‬0 𝐴𝑐 𝜏 𝑑𝜏

∞ 2
‫׬‬0 𝜏−𝜇𝑇𝑚 𝐴𝑐 𝜏 𝑑𝜏
▪ RMS delay spread: 𝜎𝑇𝑚 = ∞
‫׬‬0 𝐴𝑐 𝜏 𝑑𝜏

▪ Can be interpreted as the mean and standard deviation of the

𝐴𝑐 (𝜏)
normalized 𝐴𝑐 (𝜏, 0), interpreted as a PDF: 𝑝 𝜏 = ∞
‫׬‬0 𝐴𝑐 𝜏 𝑑𝜏

▪ Often, the RX locks to the multipath component at 𝜇 𝑇𝑚 , in

which case 𝜎𝑇𝑚 characterizes the time spreading.
Winter 2023 ECSE-610 47

3.3.2 Coherence Bandwidth

Winter 2023 ECSE-610 48

Frequency response of TV channel

▪ In system design and analysis, we also consider the
frequency domain characterization of the channel:
▪ We define the TV channel frequency response as:
∞
𝐶 𝑓, 𝑡 = න 𝑐 𝜏, 𝑡 𝑒 −𝑗2𝜋𝑓𝜏 𝑑𝜏
−∞
▪ Under our previous assumptions (independent multipaths
with uniform phases), 𝐶(𝑓, 𝜏) is zero-mean Gaussian.
▪ Hence, it is completely characterized by its
autocorrelation function:
𝐴𝐶 𝑓1 , 𝑓2 , 𝑡1 , 𝑡2 = 𝐸[𝐶 ∗ 𝑓1 , 𝑡1 𝐶(𝑓2 , 𝑡2 )]
Winter 2023 ECSE-610 49

Properties of autocorrelation
▪ Under the WSS assumption, the autocorrelation function
of 𝐶 𝑓, 𝑡 only depends on Δ𝑡 = 𝑡2 − 𝑡1 .
▪ If we further assume uncorrelated scattering, i.e. WSSUS,
then it depends only on Δ𝑡 and Δ𝑓 = 𝑓2 − 𝑓1 :
𝐴𝐶 𝑓1 , 𝑓2 , 𝑡1 , 𝑡2 = ⋯
∞
= න 𝐴𝑐 𝜏, Δ𝑡 𝑒 −𝑗2𝜋Δ𝑓𝜏 𝑑𝜏
−∞
≡ 𝐴𝐶 (Δ𝑓, Δ𝑡)
▪ The function 𝐴𝐶 Δ𝑓, Δ𝑡 can be measured by transmitting
a pair of sinusoids separated by Δ𝑓 and calculating their
cross-correlation at the receiver for time separation Δ𝑡.
Winter 2023 ECSE-610 50

Coherence bandwidth
▪ Define 𝐴𝐶 Δ𝑓 = 𝐴𝐶 (Δ𝑓, Δ𝑡) with Δ𝑡 = 0.
▪ It follows that 𝐴𝐶 (Δ𝑓) is the Fourier transform of the power
delay profile:
∞
𝐴𝐶 (Δ𝑓) = න 𝐴𝑐 𝜏 𝑒 −𝑗2𝜋Δ𝑓𝜏 𝑑𝜏
−∞
▪ By definition, we also have that 𝐴𝐶 Δ𝑓 = 𝐸 𝐶 ∗ 𝑓1 , 𝑡 𝐶 𝑓2 , 𝑡 is
an autocorrelation function in the frequency domain.
▪ The channel frequency responses 𝐶(𝑓, 𝑡) taken at frequencies
𝑓1 and 𝑓2 are approximately independent if the frequency
separation Δ𝑓 = 𝑓2 − 𝑓1 is such that 𝐴𝐶 (Δ𝑓) ≅ 0.
▪ The frequency 𝐵𝑐 such that 𝐴𝐶 (Δ𝑓) ≅ 0 for all Δ𝑓 > 𝐵𝑐 is called
the coherence bandwidth of the channel.
Winter 2023 ECSE-610 51

Coherence bandwidth (cont’d)

▪ Due to the Fourier transform relation between 𝐴𝐶 (Δ𝑓) and
the power delay profile 𝐴𝑐 (𝜏), it follows that 𝐵𝑐 ≅ 𝜎𝑇−1
𝑚

▪ Narrowband channel assumption revisited:

▪ Recall that for narrowband assumption, we need 𝑇𝑚 = 𝜏𝑁 −
𝜏0 < 𝐵𝑢−1 where 𝐵𝑢 is the signal bandwidth.
▪ Replacing 𝑇𝑚 by 𝜎𝑇𝑚 , we obtain 𝐵𝑢 < 𝐵𝑐 .
▪ Hence, fading across signal bandwidth is highly correlated.
▪ This is referred to as flat fading.
▪ Wideband channel:
▪ For wideband channel, we have 𝑇𝑚 > 𝐵𝑢−1 , i.e. 𝐵𝑢 > 𝐵𝑐
▪ In this case, channel amplitude values at frequencies
separated by more than 𝐵𝑐 are roughly independent.
▪ This is referred to as frequency selective.
Winter 2023 ECSE-610 52

3.3.3 Coherence Time

Winter 2023 ECSE-610 53

Doppler power spectrum

▪ The time variations of the channel due to TX or RX motion
cause a Doppler shift in the RX signal.
▪ This effect can be characterized by taking the Fourier
transform of 𝐴𝐶 Δ𝑓, Δ𝑡 w.r.t. Δ𝑡:
∞
𝑆𝐶 Δ𝑓, 𝜌 = න 𝐴𝐶 (Δ𝑓, Δ𝑡) 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
−∞
▪ In particular, setting Δ𝑓 = 0, we define the so-called
Doppler power spectrum as
∞
𝑆𝐶 𝜌 = 𝑆𝐶 Δ𝑓, 0 = න 𝐴𝐶 (Δ𝑡) 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
−∞
where here, we define 𝐴𝐶 Δ𝑡 = 𝐴𝐶 Δ𝑓, Δ𝑡 |Δ𝑓=0 .
Winter 2023 ECSE-610 54

Coherence time
▪ 𝐴𝐶 Δ𝑡 = 𝐸 𝐶 ∗ 𝑓, 𝑡1 𝐶 𝑓, 𝑡2 is the autocorrelation
function of the channel response 𝐶(𝑓, 𝑡) over time.
▪ The time 𝑇𝑐 such that 𝐴𝐶 (Δ𝑡) ≅ 0 for all Δ𝑡 > 𝑇𝑐 is called
the coherence time of the channel.
▪ The channel frequency responses 𝐶(𝑓, 𝑡) taken at times
𝑡1 and 𝑡2 are approximately independent if the time
separation Δ𝑡 = 𝑡2 − 𝑡1 is such that 𝐴𝐶 (Δ𝑡) ≅ 0.
Winter 2023 ECSE-610 55

Doppler spread
▪ Recall that the Doppler power spectrum is the Fourier
transform of 𝐴𝐶 Δ𝑡 :
∞
𝑆𝐶 𝜌 = න 𝐴𝐶 (Δ𝑡) 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
−∞
▪ Let 𝐵𝐷 denote the maximum observable Doppler
frequency, i.e. such that 𝑆𝐶 (𝜌) ≅ 0 if 𝜌 > 𝐵𝐷 .
▪ 𝐵𝐷 is also called the Doppler spread.
▪ As a consequence of the above FT relationship, it follows
that 𝐵𝐷 ≅ 𝑇𝐶−1
Winter 2023 ECSE-610 56

3.3.4 Two-dimensional Transforms

Winter 2023 ECSE-610 57

Two dimensional Fourier transforms

▪ For the WSSUS channel, all the relevant physical
properties and parameters can be deduced from 𝐴𝑐 𝜏, Δ𝑡
and its 2D Fourier transforms (FT):
▪ The scattering function 𝑆𝑐 𝜏, 𝜌 is defined as the FT of
𝐴𝑐 𝜏, Δ𝑡 with respect to the time Δ𝑡:
∞
𝑆𝑐 (𝜏, 𝜌) = න 𝐴𝑐 𝜏, Δ𝑡 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
∞
▪ The FT of 𝐴𝑐 𝜏; Δ𝑡 with respect to the delay 𝜏 is
∞
𝐴𝐶 (Δ𝑓, Δ𝑡) = න 𝐴𝑐 𝜏, Δ𝑡 𝑒 −𝑗2𝜋Δ𝑓𝜏 𝑑𝜏
−∞
▪ The FT of 𝐴𝑐 Δ𝑓, Δ𝑡 with respect to Δ𝑡 is
∞
𝑆𝐶 Δ𝑓, 𝜌 = න 𝐴𝐶 (Δ𝑓, Δ𝑡) 𝑒 −𝑗2𝜋𝜌Δ𝑡 𝑑Δ𝑡
−∞
Winter 2023 ECSE-610 58

Summary of FT relationships
▪ These relations are summarized in the diagram below:
Winter 2023 ECSE-610 59

Selected sources
▪ A. F. Molish, Wireless Communications, 2nd Ed., IEEE Wiley, 2011.
▪ A. Goldsmith, Wireless Communications, Cambridge, 2005.