EE359 Discussion Session 5Performance of Linear Modulation in Fading, Diversity
November 1, 2017
EE359 Discussion 5 November 1, 2017 1 / 32
Outline
1 Recap
2 Performance analysis of linear modulation
3 Moment Generating Functions
4 Linear modulation in fading
5 Diversity
EE359 Discussion 5 November 1, 2017 2 / 32
Announcement
Midterm review on Monday, November 6, 4-6 pm, in Packard 364
Midterm on Thursday, November 9, 6-8 pm, in Thornton 102
OH hour changed for next week — see calendar!
EE359 Discussion 5 November 1, 2017 3 / 32
Last discussion session
Capacity formulas with CSIT and CSIR
Optimal power and rate adaptation policies
Suboptimal power adaptation policies
This session
Performance analysis of linear modulation in fading
Diversity
EE359 Discussion 5 November 1, 2017 4 / 32
Outline
1 Recap
2 Performance analysis of linear modulation
3 Moment Generating Functions
4 Linear modulation in fading
5 Diversity
EE359 Discussion 5 November 1, 2017 5 / 32
What is linear modulation
Definition
Any modulation where the data is encoded in real or complex symbols (i.e.in amplitude or phase) is linear modulation
Observations
Performance dependent on constellation (not on baseband waveform)
Examples are BPSK, QPSK, MPAM, MQAM, etc.
FSK (frequency shift keying) is not linear, why ?
EE359 Discussion 5 November 1, 2017 6 / 32
Linear modulation in AWGN
Idea
Noise is additive, hence
y[i] = γ ∗ x[i] + n[i]
Pb = Q(√
2γb) if n[i] ∼ N (0, 1), and x[i] ∈BPSK
In general, well approximated by Ps ≈ αMQ(√βMγ), where αM , βM
are constellation dependent
For differential PSK (DPSK) systems, Pb = 12e−γb
EE359 Discussion 5 November 1, 2017 7 / 32
Outline
1 Recap
2 Performance analysis of linear modulation
3 Moment Generating Functions
4 Linear modulation in fading
5 Diversity
EE359 Discussion 5 November 1, 2017 8 / 32
Moments of a Random Variable
Given a random variable X, and its probability density function f(x), itsnth moment is given by:
µn = E[Xn] =
∫ ∞−∞
xnf(x)dx
Why do we care?
n = 1 is the mean
n = 2 is the variance
Higher order moments often interesting
EE359 Discussion 5 November 1, 2017 9 / 32
Moments of a Random Variable
Given a random variable X, and its probability density function f(x), itsnth moment is given by:
µn = E[Xn] =
∫ ∞−∞
xnf(x)dx
Why do we care?
n = 1 is the mean
n = 2 is the variance
Higher order moments often interesting
EE359 Discussion 5 November 1, 2017 9 / 32
Moment Generating Function
Many ways to compute the moment, one useful way is throught theMoment Generating Function (MGF):
MX(t) = E[etX]
Intuition:
MX(t) = E[etX]
= 1 + tE [X] +t2E
[X2]
2!+ · · ·
Differentiating and setting t = 0 gives moments!
EE359 Discussion 5 November 1, 2017 10 / 32
Moment Generating Function
Many ways to compute the moment, one useful way is throught theMoment Generating Function (MGF):
MX(t) = E[etX]
Intuition:
MX(t) = E[etX]
= 1 + tE [X] +t2E
[X2]
2!+ · · ·
Differentiating and setting t = 0 gives moments!
EE359 Discussion 5 November 1, 2017 10 / 32
Moment Generating Function
More concretely:
µn = E[Xn] =dnMX
dtn
∣∣∣∣t=0
Another useful observation:
L{fX} (s) =
∫ ∞−∞
e−sxfX(x)dx
MX(t) = L{fX} (−t) =
∫ ∞−∞
e−sxfX(x)dx
MGF can be found from two-sided Laplace tranform of PDF
EE359 Discussion 5 November 1, 2017 11 / 32
Moment Generating Function
More concretely:
µn = E[Xn] =dnMX
dtn
∣∣∣∣t=0
Another useful observation:
L{fX} (s) =
∫ ∞−∞
e−sxfX(x)dx
MX(t) = L{fX} (−t) =
∫ ∞−∞
e−sxfX(x)dx
MGF can be found from two-sided Laplace tranform of PDF
EE359 Discussion 5 November 1, 2017 11 / 32
Sums of Random Variables
MGFs are also useful for dealing with linear combinations of randomvariables
Sn =
n∑i=1
aiXi,
ai arbitrary constants
Xi independant (not necessarily identical) random variables.
Moments of Sn are given by:
MSn(t) =
n∏i=1
MXi(ait)
EE359 Discussion 5 November 1, 2017 12 / 32
Outline
1 Recap
2 Performance analysis of linear modulation
3 Moment Generating Functions
4 Linear modulation in fading
5 Diversity
EE359 Discussion 5 November 1, 2017 13 / 32
Performance in fading
System model
y[i] =√γ[i]x[i] + n[i]
Different metrics
Average probability of error: Relevant when channel is fast fading
Outage probability: Relevant when channel is slow fading
Combined Outage + Avg. probability of error: shadowing (slow) andfading (fast)
EE359 Discussion 5 November 1, 2017 14 / 32
Homework 5
Problem 1
Outage probability relevant when symbol time is much less than coherencetime of channel
EE359 Discussion 5 November 1, 2017 15 / 32
Computing the average probability of error
Idea
Integrate the Q function over fading distributions
Use change of integration order to try to get closed form expressions
Some useful relations
Pb for BPSK in Rayleigh ≈ 14γ̄ (Closed form also possible)
Pb for DPSK in Rayleigh ≈ 12γ̄ (Closed form possible)
Q(x) =∫∞x
1√2πe−z
2/2dz =∫ π/2
0 e−x2/(2 sin2 φ)dφ
Average BER computation is often MGF computation asQ(√γ) ≤ 1
2e−γ/2
EE359 Discussion 5 November 1, 2017 16 / 32
P̄s using MGF (Mγ(s) =∫∞
0 esγp(γ)dγ)
Idea
Use fact that
Q(x) =
∫ ∞x
1√2πe−
y2
2 dy =1
π
∫ π/2
0e−x
2/2 sin2 φdφ
P̄s ≈∫ ∞
0
αMπQ(√βMγ)p(γ)dγ
=αMπ
∫ γ=∞
γ=0
∫ φ=π/2
φ=0e−βMγ/2 sin2 φp(γ)dφdγ
=αMπ
∫ φ=π/2
φ=0
∫ γ=∞
γ=0e−βMγ/2 sin2 φp(γ)dγdφ
=αMπ
∫ φ=π/2
φ=0Mγ(−βM/2 sin2 φ)dφ
EE359 Discussion 5 November 1, 2017 17 / 32
P̄s using MGF (Mγ(s) =∫∞
0 esγp(γ)dγ)
Idea
Use fact that
Q(x) =
∫ ∞x
1√2πe−
y2
2 dy =1
π
∫ π/2
0e−x
2/2 sin2 φdφ
P̄s ≈∫ ∞
0
αMπQ(√βMγ)p(γ)dγ
=αMπ
∫ γ=∞
γ=0
∫ φ=π/2
φ=0e−βMγ/2 sin2 φp(γ)dφdγ
=αMπ
∫ φ=π/2
φ=0
∫ γ=∞
γ=0e−βMγ/2 sin2 φp(γ)dγdφ
=αMπ
∫ φ=π/2
φ=0Mγ(−βM/2 sin2 φ)dφ
EE359 Discussion 5 November 1, 2017 17 / 32
Example: BPSK in Rayleigh Fading
BPSK in AWGN: Pb = Q(√
2γb)⇒ α = 1β = 2. Moment generatingfunction for Rayleigh Fading:
Mγb(s) = (1− sγ̄b)−1
Mγb
(− 1
sin2 φ
)=
(1 +
γ̄b
sin2 φ
)−1
Integral now becomes:
P̄b =1
π
∫ π/2
0
(1 +
γ̄b
sin2 φ
)−1
dφ
EE359 Discussion 5 November 1, 2017 18 / 32
Example: BPSK in Rayleigh Fading
BPSK in AWGN: Pb = Q(√
2γb)⇒ α = 1β = 2. Moment generatingfunction for Rayleigh Fading:
Mγb(s) = (1− sγ̄b)−1
Mγb
(− 1
sin2 φ
)=
(1 +
γ̄b
sin2 φ
)−1
Integral now becomes:
P̄b =1
π
∫ π/2
0
(1 +
γ̄b
sin2 φ
)−1
dφ
EE359 Discussion 5 November 1, 2017 18 / 32
Example: BPSK in Rayleigh Fading
BPSK in AWGN: Pb = Q(√
2γb)⇒ α = 1β = 2. Moment generatingfunction for Rayleigh Fading:
Mγb(s) = (1− sγ̄b)−1
Mγb
(− 1
sin2 φ
)=
(1 +
γ̄b
sin2 φ
)−1
Integral now becomes:
P̄b =1
π
∫ π/2
0
(1 +
γ̄b
sin2 φ
)−1
dφ
EE359 Discussion 5 November 1, 2017 18 / 32
On error floors
Idea
As γs →∞, Perror → 0 usually. Not true if there is an error floor!
Some reasons
Differential modulation with large symbol times and/or fast fading(due to small Tc)
Due to intersymbol interference ISI (or small Bc) Pb ≈ ( σTs )2
Some factors
Correlation function of channel (channel coherence time Tc andbandwidth Bc)
Fading statistics, symbol time Ts
Question
What happens to error floors if Ts decreases or data rate increases?
EE359 Discussion 5 November 1, 2017 19 / 32
Homework 5
Problem 2
Use average error probability requirement to get γ̄b
Use Pb requirement to set the target power in the presence ofshadowing and path loss
Use cell coverage area formula, if appropriate
Problem 3
Use formula for Rayleigh fading with DPSK, i.e. (6.90), with K = 0, withthe ρC function given by the Jakes’ formula (uniform scattering). Thisgives error floor due to doppler.
EE359 Discussion 5 November 1, 2017 20 / 32
Outline
1 Recap
2 Performance analysis of linear modulation
3 Moment Generating Functions
4 Linear modulation in fading
5 Diversity
EE359 Discussion 5 November 1, 2017 21 / 32
Diversity
Idea
Use of independent fading realizations can reduce the probability oferror/outage events
Some observations
Diversity can be in time, space, frequency, polarization, . . .
Diversity order used as a measure of diversity, defined as
M = limγ̄→∞
− logPelog γ̄
, Pe = P̄s or Pout
Can also use array gain (or SNR gain) γ̄Σ/γ̄, where γ̄Σ is the averageSNR after “diversity combining”
EE359 Discussion 5 November 1, 2017 22 / 32
Diversity order
Diversity order
Specifying diversity order M is roughly equivalent to saying that at high γ̄,
P̄e ≈ (γ̄)−M
Array gain
Array gain Ag is equivalent to ratio of average SNRs after diversitycombining
Ag =γ̄Σ
γ̄
EE359 Discussion 5 November 1, 2017 23 / 32
Diversity order
Diversity order
Specifying diversity order M is roughly equivalent to saying that at high γ̄,
P̄e ≈ (γ̄)−M
Array gain
Array gain Ag is equivalent to ratio of average SNRs after diversitycombining
Ag =γ̄Σ
γ̄
EE359 Discussion 5 November 1, 2017 23 / 32
Diversity combining techniques
In this class, we have talked about two schemes to exploit diversity, bothat the receiver
System model
r = γx+ n
Some receiver diversity combining schemes
Selection combining: Choose the largest SNR of the independentrealizations
Maximal ratio combining: Combine all the independent receivedSNRs to maximize SNR
EE359 Discussion 5 November 1, 2017 24 / 32
Selection combining (SC)
Idea
Given M i.i.d. r.v., γ1, . . . , γM ≥ 0,
P (maxi
(γi) < c) = P (γi < c)M
Some observations
Define γΣ = maxi γi
In Rayleigh fading γ̄Σ = γ̄(∑M
i=1 1/i) (γ̄: average SNR at a branch)
P̄b in general difficult, but for DPSK and Rayleigh fading,
P̄b = M/2
M−1∑m=0
(−1)m(M−1m
)1 +m+ γ̄
EE359 Discussion 5 November 1, 2017 25 / 32
Selection combining continued
Outage probability
Pout =(
1− e−γ0γ̄
)MQuestion (SC in Rayleigh fading)
What is the diversity gain?:
What is the SNR gain?:
EE359 Discussion 5 November 1, 2017 26 / 32
Selection combining continued
Outage probability
Pout =(
1− e−γ0γ̄
)MQuestion (SC in Rayleigh fading)
What is the diversity gain?: M
What is the SNR gain?:∑M
i 1/i
EE359 Discussion 5 November 1, 2017 26 / 32
Threshold Combining (SSC)
Idea
Use one RF chain with multiple antennas
PγΣ(γ) =
{Pγ1(γT )Pγ2(γ) γ < γT
p(γT ≤ γ2 ≤ γ) + Pγ1(γT )Pγ2(γ) γ ≥ γT
EE359 Discussion 5 November 1, 2017 27 / 32
Maximal ratio combining (MRC)
Idea
Instead of discarding weaker branches, combine the SNRs of all branches,i.e.
γΣ =
M∑i
γi
Nuts and bolts
Need to make the received components of the same phase (not aproblem with modern DSP)
Maximal ratio combining maximizes received SNR, i.e. solves thefollowing problem
maxa:||a||2=1
E[|aHγx|2]
E[|aHn|2]
MGF of sums decompose into product of individual MGFs so easy toanalyse P̄s
EE359 Discussion 5 November 1, 2017 28 / 32
MRC continued (Outage probability and P̄s)
Outage probability
Pout = 1− eγ0γ̄
(M−1∑i=0
(γ0
γ̄
)i/i!
)
Average probability of error P̄sThe MGF of sum decouples into product of MGFs
For DPSK and Rayleigh fading, average error probability is
1
2EγΣ [e−γΣ ] =
1
2
M∏i=1
Eγi [e−γi ] =
1
2
M∏i=1
M(−1)
For general constellations P̄s is approximately of the form
C
∫ φ=B
φ=A(M(−γ/2 sin2 φ))Mdφ
EE359 Discussion 5 November 1, 2017 29 / 32
Questions
What is the diversity order for MRC?:
What is the SNR gain for MRC?:
EE359 Discussion 5 November 1, 2017 30 / 32
Questions
What is the diversity order for MRC?: M
What is the SNR gain for MRC?: M
EE359 Discussion 5 November 1, 2017 30 / 32
Homework 5
Problem 4
Concepts in SC (Selection combining) - probability of outage
Problem 5
DPSK under Rayleigh fading using SSC. Hint: Split integral into twoparts: 0→ γT and γT →∞.
Problem 6
Concepts in MRC - Deriving optimal weights, computing P̂s. Using MGFapproach in part (b).
EE359 Discussion 5 November 1, 2017 31 / 32
Homework 5
Problem 4
Concepts in SC (Selection combining) - probability of outage
Problem 5
DPSK under Rayleigh fading using SSC. Hint: Split integral into twoparts: 0→ γT and γT →∞.
Problem 6
Concepts in MRC - Deriving optimal weights, computing P̂s. Using MGFapproach in part (b).
EE359 Discussion 5 November 1, 2017 31 / 32
Homework 5
Problem 4
Concepts in SC (Selection combining) - probability of outage
Problem 5
DPSK under Rayleigh fading using SSC. Hint: Split integral into twoparts: 0→ γT and γT →∞.
Problem 6
Concepts in MRC - Deriving optimal weights, computing P̂s. Using MGFapproach in part (b).
EE359 Discussion 5 November 1, 2017 31 / 32
Homework 5
Problem 7
Notions of combining - SC/MRC. For part (c), if A⇒ B or in other wordsA ⊆ B, then P (A) ≤ P (B).
Problem 8
Diversity gain in SC and MRC at high/low SNR.
EE359 Discussion 5 November 1, 2017 32 / 32
Homework 5
Problem 7
Notions of combining - SC/MRC. For part (c), if A⇒ B or in other wordsA ⊆ B, then P (A) ≤ P (B).
Problem 8
Diversity gain in SC and MRC at high/low SNR.
EE359 Discussion 5 November 1, 2017 32 / 32