Winter 2019 ECSE-610 1

4. Capacity of Wireless
Channels

Prof. Benoit Champagne

Dept. of Electrical & Computer Engineering

McGill University, Montreal, Canada

January 31, 2023

Winter 2019 ECSE-610 2

Introduction
▪ Channel capacity refers to the maximum data rates (in bps)
that can be transmitted over a wireless channel with vanishing
error probability, assuming no constraints on delay or
complexity of the encoder and decoder.
▪ The mathematical theory of communication underlying channel
capacity was pioneered by C. Shannon in his 1948s paper:
▪ In its modern form, the theory relies on the notion of mutual information
between the channel input and output, modeled as random variables.
▪ In particular, channel capacity is defined as the channel’s mutual
information maximized over all possible input distributions.
▪ Key contributions of Shannon’s work include the coding
theorem and its converse:
▪ Coding theorem: a code exists that can achieve a data rate close to
capacity with negligible probability of error.
▪ Converse: any data rate higher than capacity cannot be achieved without
an error probability bounded away from zero.
Winter 2019 ECSE-610 3

Introduction (cont’d)
▪ This chapter examines the capacity of a single-user wireless
channel where TX and/or RX have a single antenna.
▪ Multi-antenna systems ⟹ Chap. 10.
▪ Multi-user systems ⟹ Chap. 14.
▪ We first look at the well-known formula for capacity of a TI
additive white Gaussian noise (AWGN) channel.
▪ We then consider capacity of TV flat fading channels.
▪ Here, capacity formulation depends on what is known about the
channel at the TX and RX, i.e. channel state information (CSI).
▪ For different CSI assumptions, we also need to distinguish between
ergodic and outage capacity.
▪ The presentation is limited to discrete-time systems:
▪ Most of the results admit extension to continous-time (as limiting
cases of sampled system)
▪ The details are beyond the scope of this course.
Winter 2019 ECSE-610 4

Overview
4.1 Capacity in AWGN
4.2 Capacity of Flat Fading Channels
Winter 2019 ECSE-610 5

4.1 Capacity in AWGN

Winter 2019 ECSE-610 6

Discrete-memoryless channel
▪ A discrete memoryless TI channel is a simplified model of a
transmission where both the input and output are random
symbols taken from discrete alphabets.
▪ We let 𝑥 denote the random symbol transmitted at a given time.
▪ 𝑥 ∈ 𝒳 is a random variable (RV)
▪ 𝒳 denotes a discrete set possible values for 𝑥 (alphabet)
▪ RV 𝑥 is characterized by a probability mass function (PMF) 𝑝(𝑥)
▪ We let 𝑦 denote the random symbol received at the channel
output, when 𝑥 is transmitted.
▪ 𝑦 ∈ 𝒴 is a random variable (RV)
▪ 𝒴 denotes a discrete set possible values for 𝑦
▪ RV 𝑦 is characterized by a PMF 𝑝(𝑦)
▪ The channel is simply modeled in terms of the conditional PMF
of 𝑦 given 𝑥, i.e. 𝑝 𝑦 𝑥 .
▪ In the TI case, 𝑝 𝑦 𝑥 remains constant over time
Winter 2019 ECSE-610 7

Channel’s mutual information

▪ Definition of channel’s mutual information:
𝑝 𝑥, 𝑦
𝐼 𝑋, 𝑌 = ෍ 𝑝(𝑥, 𝑦) log 2
𝑝 𝑥 𝑝 𝑦
𝑥∈𝒳,𝑦∈𝒴
▪ When log 2 is used, the units are in bits per channel use.
▪ The mutual information can also be written as
𝐼 𝑋, 𝑌 = 𝐻 𝑌 − 𝐻(𝑌|𝑋)
▪ 𝐻 𝑌 = − σ𝑦∈𝒴 𝑝(𝑦) log 2 𝑝(𝑦) is the entropy of the output.
▪ 𝐻 𝑌|𝑋 = − σ𝑥∈𝒳,𝑦∈𝒴 𝑝(𝑥, 𝑦) log 2 𝑝(𝑦|𝑥) is the conditional
entropy of the output given the input.
▪ We can think of 𝐼(𝑋, 𝑌) has the amount of uncertainty
removed from output 𝑦 given the input 𝑥.
Winter 2019 ECSE-610 8

Channel capacity
▪ For the discrete-memoryless channel, the capacity is defined
as the maximum of the channel’s mutual information over all
possible input distributions:
𝐶 = max 𝐼 𝑋, 𝑌
𝑝(𝑥)
𝑝(𝑥, 𝑦)
= max ෍ 𝑝(𝑥, 𝑦) log 2 ( )
𝑝(𝑥) 𝑝 𝑥 𝑝(𝑦)
𝑥,𝑦
▪ For this channel, Shannon proved the following theorems:
▪ Coding theorem: a code exists that can achieve a data rate as
close as desired to capacity with negligible probability of error.
▪ Converse: any data rate higher than capacity cannot be achieved
without an error probability bounded away from zero.
▪ Shannon also extended these results to other types of
channels, including the continuous AWGN channel.
Winter 2019 ECSE-610 9

AWGN channel
▪ The AWGN is a simplified model of a transmission where
both the input and output are random symbols, but this
time taking continuous values.
▪ Here we consider a memoryless discrete-time formulation.
▪ We let 𝑥[𝑖] denote the channel input at discrete-time 𝑖 ∈ ℤ:
▪ 𝑥[𝑖] is a RV with values in ℝ, the set of real number.
▪ RV 𝑥[𝑖] is characterized by its PDF 𝑝(𝑥).
▪ 𝑥[𝑖] has zero-mean with known power 𝑃 = 𝐸[𝑥 𝑖 2 ]
▪ We let 𝑦[𝑖] denote the channel output at time 𝑖:
▪ 𝑦[𝑖] is a RV with values in ℝ.
▪ RV 𝑦[𝑖] is characterized by its PDF 𝑝(𝑦).
Winter 2019 ECSE-610 10

AWGN channel (cont’d)

▪ The effect of the channel on the input is to add noise:
𝑦 𝑖 =𝑥 𝑖 +𝑛 𝑖
▪ We make the following assumptions about the noise:
▪ 𝑛[𝑖]~𝒩(0, 𝜎𝑛2 ), i.e. zero mean Gaussian RV variance 𝜎𝑛2 .
▪ Noise samples 𝑛[𝑖] and 𝑛[𝑗] at times 𝑖 ≠ 𝑗 are independent
▪ We refer to such a channel as memoryless.
▪ The AWGN channel is TI if 𝜎𝑛2 is constant over time
▪ The noise variance can be calculated as 𝜎𝑛2 = 𝑁0 𝐵:
▪ 𝐵 is the channel bandwidth.
▪ 𝑁0 Τ2 is the power spectral density level of the noise
𝑃
▪ The channel SNR can be expressed as 𝛾 =
𝑁0 𝐵
Winter 2019 ECSE-610 11

AWGN channel capacity

▪ The capacity of the AWGN channel is given by Shannon’s well-
known formula:
𝐶 = 𝐵 log 𝑐 1 + 𝛾
▪ He proved that a code exists that achieves data 𝑅 rates
arbitrarily close to 𝐶, with arbitrarily small bit error probability.
▪ A proof can be given in terms of the mutual information 𝐼(𝑋, 𝑌),
where the summations are replaced by integrals.
▪ The input distribution achieving capacity (i.e. maximizing 𝐼(𝑋, 𝑌)) is
the Gaussian distribution 𝒩(0, 𝑃).
▪ Conversely, any code with rate 𝑅 > 𝑃 has a probability of error
bounded away from zero.
▪ On AWGN channels, turbo codes have come within a fraction
of a decibel of the Shannon capacity limit.
Winter 2019 ECSE-610 12

Example 4.1
▪

▪ Solution: Discussed in class…

Winter 2019 ECSE-610 13

4.2 Capacity in Flat Fading

Channels
Winter 2019 ECSE-610 14

Introduction

▪ Here, contrary to AWGN channel, capacity is not given by

single formula but depends on what is known about the
channel at TX and RX, i.e. channel state information (CSI).
▪ Furthermore, depending on the CSI assumptions, different
definitions of capacity may be employed:
▪ Ergodic capacity: which characterizes the maximum achievable
rate averaged over all fading states.
▪ Outage capacity: which characterizes the maximum constant
rate that can be maintained given some probability of outage.
Winter 2019 ECSE-610 15

4.2.1 Channel and System Model

Winter 2019 ECSE-610 16

Channel and system model

▪ We consider the following transmission model consisting of a
TX, TV fading channel and RX:

▪ The transmitter encodes the input message 𝒘 into a discrete-

time input sequence 𝑥 𝑖 .
▪ The channel applies a TV gain ℎ 𝑖 = 𝑔[𝑖] along with additive
noise 𝑛[𝑖] to generate its output:
𝑦 𝑖 = ℎ 𝑖 𝑥 𝑖 + 𝑛[𝑖]
▪ The receiver decodes the transmitted message from the
ෝ
channel output sequence 𝑦 𝑖 and reconstructs 𝒘.
Winter 2019 ECSE-610 17

Flat fading model

▪ We assume discrete-time channel with TV gain ℎ 𝑖 =
𝑔[𝑖], where 𝑔 𝑖 ≥ 0 is the power gain at time 𝑖.
▪ We model 𝑔[𝑖] as a stationary, ergodic random process
with PDF 𝑝𝑖 (𝑔) ≡ 𝑝(𝑔) (same for all 𝑖).
▪ E.g., for Rayleigh fading, 𝑝(𝑔) is exponential.
▪ We assume that 𝑔[𝑖] is independent of the input signal 𝑥[𝑖].
▪ The changes in 𝑔 𝑖 over time can be modeled differently:
▪ Memoryless (or iid): 𝑔[𝑖] and 𝑔 𝑗 are independent if 𝑖 ≠ 𝑗
▪ Channel with memory: 𝑝 𝑔 𝑖 𝑔 𝑗 for 𝑗 < 𝑖 ≠ 𝑝(𝑔 𝑖 )
▪ Block fading channel: 𝑔[𝑖] is constant over some block of
length 𝑇, after which it changes to a new value.
Winter 2019 ECSE-610 18

SNR distribution
▪ We let 𝑃 = 𝐸 𝑥 𝑖 2 denote the input TX power, which for
now is assumed to be constant.
▪ For the noise, we make the same assumption as in the
AWGN model, in particular:
▪ 𝑛[𝑖]~𝒩(0, 𝜎𝑛2 ), i.e. zero mean Gaussian RV with 𝜎𝑛2 = 𝑁0 𝐵.
▪ The instantaneous SNR at the RX is written as
𝑃𝑔 𝑖
𝛾 𝑖 ≡𝛾 𝑔 𝑖 =
𝑁0 𝐵
▪ Clearly, the distribution of 𝑔[𝑖] determines that of 𝛾 𝑖
▪ We can think of 𝛾[𝑖] as a conditional SNR, given 𝑔[𝑖]
▪ The average SNR over the ensemble of fading state is
∞ 𝑃𝜇𝑔
𝛾 = න 𝛾(𝑔)𝑝 𝑔 𝑑𝑔 =
0 𝑁0 𝐵
Winter 2019 ECSE-610 19

Different scenarios
▪ The capacity of the channel depends on what is known
about 𝑔[𝑖] at the TX and RX, i.e., the CSI.
▪ We will consider three different scenarios for the CSI:
1) Channel distribution information (CDI): Only the distribution of
𝑔[𝑖], i.e. the PDF 𝑝(𝑔), is known to the TX and RX.
2) Receiver CSI: The value of 𝑔 𝑖 is known to the receiver at time
𝑖, and both TX and RX know the PDF 𝑝(𝑔).
3) Transmitter and receiver CSI: The value of 𝑔[𝑖] is known to the
TX and RX at time 𝑖, and both TX and RX know 𝑝(𝑔).
▪ In case 3), the TX can adapt both its power and rate to the
channel gain 𝑔[𝑖] at time 𝑖, leading to higher capacity.
Winter 2019 ECSE-610 20

4.2.2 Channel Distribution

Information Known
Winter 2019 ECSE-610 21

Discussion
▪ Consider the CDI case, where only channel gain PDF 𝑝(𝑔) or,
equivalently, the SNR PDF 𝑝(𝛾), is known to TX and RX.
▪ For IID fading (memoryless channel):
▪ The capacity is given by 𝐶 = max 𝐼 𝑋, 𝑌
𝑝(𝑥)
▪ However, solving for the capacity-achieving input distribution 𝑝 𝑥
can be quite difficult depending on the fading distribution.
▪ For correlated fading (channel with memory):
▪ Above definition of capacity can be extended by considering long
blocks of input data, i.e., 𝑥 ≡ (𝑥1 , … , 𝑥𝑁 ), with 𝑁 → ∞.
▪ The capacity-achieving input distribution is found by optimizing
over input blocks, i.e.,
𝐶 = max 𝐼 𝑋𝑁 , 𝑌𝑁
𝑝(𝑥1 ,…,𝑥𝑁 )
but this is even more difficult:
Winter 2019 ECSE-610 22

Discussion (cont’d)
▪ The capacity-achieving input distribution and corresponding
fading channel capacity under CDI are known for two specific
models of interest:
▪ IID Rayleigh fading channels, but still, this requires numerical
calculations.
▪ Finite-state Markov channels where 𝑔[𝑖] takes values in a
finite set and evolves according to 𝑝(𝑔[𝑖]|𝑔 𝑖 − 1 ).
▪ In general, finding the optimal input distribution and capacity of
fading channels under CDI remains an open problem for
almost all channel distributions.
Winter 2019 ECSE-610 23

4.2.3 Channel State Information at

Receiver
Winter 2019 ECSE-610 24

Introduction
▪ We now consider the case where the CSI 𝑔 𝑖 (or equivalently
𝛾[𝑖]) is known to the RX at time 𝑖, but not to the TX:
▪ As before, both the TX and RX know the distribution of 𝑔[𝑖].
▪ The TX cannot adapt its transmission strategy relative to the CSI
and must employ a constant transmission rate.
▪ In this case there are two channel capacity definitions that are
relevant to system design:
▪ Ergodic (Shannon) capacity: “average” maximum data rate that can
be sent over the channel with vanishing error probability, where the
average is taken over the fading ensemble.
▪ Capacity with outage: maximum data rate that can be transmitted
over a channel under constraint on the SNR outage probability.
▪ These are further discussed below.
Winter 2019 ECSE-610 25

Ergodic capacity
▪ Ergodic capacity of a fading channel with receiver CSI for an
average TX power 𝑃 is given by
∞
𝐶𝑒 = න 𝐵 log 2 1 + 𝛾 𝑝 𝛾 𝑑𝛾
0
▪ This is simply the average of Shannon formula for AWGN over
the SNR distribution 𝑝 𝛾 .
▪ Recall that 𝛾 = 𝑃𝑔Τ𝑁0 𝐵, so that the TX power actually affects 𝐶
▪ Using Jensen’s inequality, we have
𝐶𝑒 = 𝐸 𝐵 log 2 1 + 𝛾 ≤ 𝐵 log 2 1 + 𝐸 𝛾
▪ Ergodic capacity of a fading channel with receiver CSI only is less
than capacity of AWGN channel with SNR = 𝐸 𝛾 ≡ 𝛾ҧ
▪ Poor channel states typically reduce ergodic capacity
Winter 2019 ECSE-610 26

Example 4.2
▪

▪ Solution: The channel has 3 possible received SNRs:

𝑃𝑡 𝑔𝑖 .01 𝑔𝑖
𝛾𝑖 = = , 𝑖 = 1,2,3
𝑁0 𝐵 (10−9 )(30 ⋅ 103 )
▪ SNR values and corresponding probabilities 𝑝𝑖 = 𝑝 𝛾𝑖 :
𝒊 𝜸𝒊 𝒑𝒊
1 0.833 (-0.8dB) 0.1
2 83.33 (19.2dB) 0.5
3 333.3 (25.2dB) 0.4
▪ The rest in class…
Winter 2019 ECSE-610 27

Capacity with outage

▪ Capacity with outage usually applies to slowly varying
channels, where the SNR can be assumed nearly constant
over long transmission bursts
▪ It accounts for the possibility that bits sent over a given burst
be decoded incorrectly, i.e. with non-zero error probability.
▪ Specifically, in this approach:
▪ The TX fixes a minimum received SNR 𝛾𝑚𝑖𝑛 and encodes for a
data rate 𝐶 = 𝐵 log 2 (1 + 𝛾𝑚𝑖𝑛 ).
▪ The RX decode the data correctly if its SNR ≥ 𝛾𝑚𝑖𝑛 ; while if SNR <
𝛾𝑚𝑖𝑛 , the data cannot be decoded correctly.
▪ We refer to the event SNR < 𝛾𝑚𝑖𝑛 as an outage and denote the
corresponding probability as:
𝑃𝑜𝑢𝑡 = Pr(𝛾 < 𝛾𝑚𝑖𝑛 )
Winter 2019 ECSE-610 28

Capacity with outage (cont’d)

▪ The outage probability can be obtained by plotting 𝐶 versus
𝑃𝑜𝑢𝑡 for different values of 𝛾𝑚𝑖𝑛 , treated as a parameter:
▪ 𝐶(𝛾𝑚𝑖𝑛 ) = 𝐵 log 2 (1 + 𝛾𝑚𝑖𝑛 )
▪ 𝑃𝑜𝑢𝑡 𝛾𝑚𝑖𝑛 = Pr(𝛾 < 𝛾𝑚𝑖𝑛 )
▪ 𝛾𝑚𝑖𝑛 ∈ [0, ∞]

▪ Fig. 4.2 shows such a plot for

Rayleigh fading (𝛾ҧ = 20dB)

▪ The average rate for “correctly received” data over

different burst is given by:
𝐶𝑜𝑢𝑡 = 1 − 𝑃𝑜𝑢𝑡 𝐵 log 2 (1 + 𝛾𝑚𝑖𝑛 )
Winter 2019 ECSE-610 29

Example 4.3
▪

▪ Solution: Discussed in class…

Winter 2019 ECSE-610 30

4.2.4 Channel Side Information at

Transmitter and Receiver
Winter 2019 ECSE-610 31

Introduction
▪ Under the transmitter and receiver CSI assumption, the
value of the instantaneous gain 𝑔[𝑖] is known to the TX
and RX at time 𝑖, as well as its PDF 𝑝(𝑔).
▪ In this case, channel capacity is achieved when the TX
adapts its power, data rate, and coding scheme to the
channel variations.
▪ The optimal power allocation is a “water-filling” in time,
where power and data rate are:
▪ Increased when channel conditions are favorable.
▪ Decreased when they are not favorable.
Winter 2019 ECSE-610 32

Ergodic capacity
▪ Under the known CSI assumption, the TX can adapt its
power data rate, and coding scheme as a function of
instantaneous SNR 𝛾[𝑖].
▪ However, this only makes sense under a constraint (i.e.
upper bound) on the average transmit power 𝑃 𝛾 :
∞
න 𝑃 𝛾 𝑝 𝛾 𝑑𝛾 ≤ 𝑃ത
0
▪ This motivates the definition of the fading channel
capacity with average power constraint as:
∞
𝑃 𝛾 𝛾
𝐶 = max න 𝐵 log 2 1 + 𝑝 𝛾 𝑑𝛾 (1)
𝑃(𝛾) 0 ത
𝑃
∞
s. t. න 𝑃 𝛾 𝑝 𝛾 𝑑𝛾 ≤ 𝑃ത
0
Winter 2019 ECSE-610 33

Solution for optimal power

▪ By applying calculus of variation with Lagrange multiplier𝜆, the
optimal 𝑃 𝛾 is obtained as:
𝑃(𝛾) 1 1 ത
𝜆𝑃𝐵
= − , 𝛾0 =
𝑃ത 𝛾0 𝛾 ln 2
▪ However, this stationary solution may lead to negative power;
hence, further enforcing the constraint 𝑃 𝛾 > 0:
1 1
𝑃(𝛾) − , 𝛾 ≥ 𝛾0
= ቐ𝛾0 𝛾 (2)
ത
𝑃
0, 𝛾 < 𝛾0
▪ The value of the Lagrange multiplier (equivalently 𝛾0 ) is
obtained by enforcing the contraint on the average power:
∞ ∞
𝑃(𝛾) 1 1
න 𝑝 𝛾 𝑑𝛾 = න − 𝑝 𝛾 𝑑𝛾 = 1 (3)
0 𝑃ത 𝛾0 𝛾0 𝛾
Winter 2019 ECSE-610 34

Solution for optimal power (cont’d)

▪ There is no simple expression for 𝛾0 in (3): it must be
determined numerically (see Example 4.4).

▪ The solution for the

optimal power in (2)
is referred to as a
water-filling solution:

▪ The final channel capacity is obtained upon substitution of

solution (2) in the capacity formula (1):
∞
𝛾
𝐶𝑤𝑓 = න 𝐵 log 2 𝑝 𝛾 𝑑𝛾
𝛾0 𝛾0
Winter 2019 ECSE-610 35

Example 4.4
▪

▪ We need to find the cutoff value 𝛾0 that satisfies the discrete version
of (4.15) given by
1 1
෍ − 𝑝 𝛾𝑖 = 1
𝛾0 𝛾𝑖
𝛾𝑖 ≥𝛾0
▪ We first assume that all channel states are used to obtain 𝛾0 (i.e., 𝛾0 ≤ 𝛾1 )
and see if the resulting 𝛾0 is below that of the weakest channel.
▪ If not, we have an inconsistency and must redo the calculation assuming at
least one of the channel states is not used (i.e. 𝛾1 < 𝛾0 ≤ 𝛾2 ), etc.
▪ Continued in class…
Winter 2019 ECSE-610 36

Channel inversion
▪ We now consider a suboptimal adaptation scheme where the
TX uses the CSI to maintain a constant received power, or
equivalently, a desired SNR 𝛾𝑑 at the RX.
▪ Recall the definition of the instantaneous SNR under constant
ത
transmit power 𝑃:
ത
𝑃𝑔
𝛾=
𝑁0 𝐵
▪ To maintain constant SNR 𝛾𝑑 at RX , the TX needs to invert
effect of channel fading by transmitting with 𝑃(𝛾), such that:
𝑃 𝛾 𝛾𝑑 𝑃 𝛾 𝑔
= (which implies = 𝛾𝑑 )
𝑃ത 𝛾 𝑁0 𝐵
▪ The channel appears to the encoder and decoder as a time-
invariant AWGN channel with capacity 𝐵 log 2 (1 + 𝑐) .
Winter 2019 ECSE-610 37

Channel inversion (cont’d)

▪ The constant SNR 𝛾𝑑 is determined from the power constraint:
∞ ∞
𝛾𝑑
න 𝑃 𝛾 𝑝 𝛾 𝑑𝛾 = 𝑃ത ⟹ න 𝑝 𝛾 𝑑𝛾 = 1
0 0 𝛾
1
⟹ 𝛾𝑑 =
𝐸 𝛾 −1
▪ Since the channel appears to the encoder and decoder as a
time-invariant AWGN channel with constant SNR 𝛾𝑑 :
𝐶𝑧𝑜 = 𝐵 log 2 (1 + 𝛾𝑑 )

▪ This is also called zero-outage capacity, since the data rate is

fixed under all channel conditions and there is no outage.
▪ In general, we find that that 𝐶𝑧𝑜 ≤ 𝐶𝑤𝑓 .
Winter 2019 ECSE-610 38

Example 4.5
▪

▪ We first need to find the value of 𝛾𝑑 :

.1 .5 .4
𝛾𝑑−1 = 𝐸 𝛾 −1 = + + = .1272
.8333 83.333 333.33
▪ The resulting capacity is:
𝐶𝑧𝑜 = 𝐵 log 2 1 + 𝛾𝑑 = 30,000 log 2 (1 + 1Τ.1272) = 94.43kbps
▪ For comparison, we previously found
▪ For water-filling approach: 𝐶𝑤𝑓 = 200.82 kbps.
▪ For AWGN with same 𝐸 𝛾 : 𝐶 = 223.8 kbps
Winter 2019 ECSE-610 39

Truncated channel inversion

▪ Zero-outage capacity may be significantly smaller than
Shannon capacity due the requirement of maintaining a
constant data rate in all fading states:
▪ Too much power can be wasted on bad channel conditions.
▪ By suspending transmission in particularly bad fading states,
we can maintain a higher constant data rate in the other states
and thereby increase capacity.
▪ The truncated channel inversion policy for power adaptation
only compensate for fading above a certain threshold 𝛾0 :
𝛾𝑑
𝑃(𝛾) , 𝛾 ≥ 𝛾0
=ቐ𝛾
𝑃ത
0, 𝛾 < 𝛾0
where 𝛾0 is based on the outage probability 𝑃out = Pr 𝛾 < 𝛾0 .
Winter 2019 ECSE-610 40

Truncated channel inversion (cont’d)

▪ Since the channel is only used when 𝛾 ≥ 𝛾0 , the power
constraint yields
∞
1
න 𝑃 𝛾 𝑝 𝛾 𝑑𝛾 = 𝑃ത ⟹ 𝛾𝑑 =
0 𝐸0 𝛾 −1
∞
−1
1
𝐸0 𝛾 ≜න 𝑝 𝛾 𝑑𝛾
𝛾0 𝛾
▪ The outage capacity associated with a given outage
probability 𝑃𝑜𝑢𝑡 or 𝛾0 is defined by
𝐶𝑜𝑢𝑡 𝛾0 = 1 − 𝑃𝑜𝑢𝑡 𝐵 log 2 (1 + 𝛾𝑑 )
▪ In turn, we can obtain the maximum outage capacity by
maximizing over all 𝛾0 :
𝐶𝑜𝑢𝑡 = argmax 𝐶𝑜𝑢𝑡 (𝛾0 )
𝛾0
Winter 2019 ECSE-610 41

4.2.6 Capacity Comparisons

Winter 2019 ECSE-610 42

Scope of comparisons
▪ Different capacity formulas for:
▪ AWGN with average SNR 𝛾ҧ = 𝐸 𝛾 as a reference.
▪ Ergodic capacity with CSI known at RX (eq. 4.4).
▪ Ergodic capacity with CSI known at RX and TX and optimal
(watter filling) power allocation (eq. 4.13).
▪ Zero-outage (channel inversion) capacity (eq. 4.18)
▪ Truncated channel inversion (eq. 4.22)
▪ Different SNR PDF:
▪ Log-normal (𝜎𝑑𝐵 = 8).
▪ Rayleigh (or exponential).
▪ Nakagami (𝑚 = 2).
▪ The different curves show 𝐶 Τ𝐵 (in bits/sec/Hz) versus
average SNR 𝛾.ҧ
Winter 2019 ECSE-610 43

Rayleigh fading
Winter 2019 ECSE-610 44

Nakagami fading
Winter 2019 ECSE-610 45

4.3 Capacity of Frequency-

Selective Fading Channels

This material is optional

Winter 2019 ECSE-610 46

Introduction
▪ In this section, we examine the Shannon capacity of
frequency-selective fading channels.
▪ We first consider the capacity of a time-invariant (TI)
frequency-selective fading channel.
▪ This capacity analysis is like that of a flat fading channel but with
the time axis replaced by the frequency axis.
▪ Then we discuss the capacity of time-varying (TV) frequency-
selective fading channels.
Winter 2019 ECSE-610 47

4.3.1 Time-invariant
Winter 2019 ECSE-610 48

System model
▪ We consider data transmission through a TI wideband
channel with frequency response 𝐻 𝑓 :

▪ We further make the following assumptions:

▪ Total transmit power constraint 𝑃.
▪ 𝐻 𝑓 known to both the transmitter and receiver.
▪ 𝑛[𝑖] is zero-mean white Gaussian noise.
Winter 2019 ECSE-610 49

System model (cont’d)

▪ To simplify the analysis, we assume that 𝐻 𝑓 is frequency
block fading:
▪ The transmission band is divided into independent narrowband
subchannels of bandwidth 𝐵, index by 𝑗 ∈ {1, … , 𝐽}.
▪ The subchannels are approximately flat fading, so that 𝐻 𝑓 ≈ 𝐻𝑗
for all 𝑓 in the 𝑗th subband.

▪ The frequency-selective channel thus consists of a set of

AWGN channels in parallel:
2
▪ The SNR of the 𝑗th channel is given by 𝛾𝑗 = 𝐻𝑗 𝑃𝑗 ൗ𝑁0 𝐵.
▪ 𝑃𝑗 denote the power allocated to the 𝑗th channel.
▪ The 𝑃𝑗 are subject to the power constraint σ𝑗 𝑃𝑗 ≤ 𝑃 .
Winter 2019 ECSE-610 50

Capacity formulation
▪ The capacity of this parallel set of channels is the sum of
the achievable rate on each channel, with power optimally
allocated over all channels:
𝐶 = max ෍ 𝐵 log 2 (1 + 𝛾𝑗 ) (1)
𝑃𝑗 𝑗

s. t. ෍ 𝑃𝑗 ≤ 𝑃ത
𝑗
▪ This is similar to the capacity and optimal power allocation
for a flat fading channel, but with 𝑃𝑗 and 𝛾𝑗 changing in a
deterministic way over the frequency index 𝑗.
▪ The optimal power allocation is found via the same
Lagrangian technique used in the flat-fading case, leading
to a water filling solution.
Winter 2019 ECSE-610 51

Water filling solution

▪ Optimal power allocation:
1 1
𝑃𝑗 − , 𝛾𝑗 ≥ 𝛾0
= ൞ 𝛾0 𝛾𝑗
𝑃
0, 𝛾𝑗 < 𝛾0
▪ The cutoff SNR value 𝛾0 is obtained by solving

1 1
σ𝛾𝑗 ≥𝛾0 − = 1.
𝛾0 𝛾𝑗

▪ The capacity then becomes

𝛾𝑗
𝐶 = ෍ 𝐵 log 2 .
𝛾0
𝛾𝑗 ≥𝛾0

▪ This capacity is achieved by transmitting at different rates and

powers over each subchannel.
Winter 2019 ECSE-610 52

4.3.2 Time-variant case

Optional reading
Winter 2019 ECSE-610 53

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