You are on page 1of 29

Lecture 11 Wireless Communication

Monsoon semester 2023-24,


Topics: Need for diversity with fading

Vinay Joseph

NIT Calicut

September 27, 2023

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Plan for studying fading channel
Key problem: how to send and receive bits using a wireless channel?
Step 1: Additive White Gaussian Noise channel without fading
y [m] = x[m] + w [m]
▶ x[m]: transmitted baseband symbol
▶ w [m]: additive white Gaussian noise
▶ y [m]: received signal
Step 2: Flat-fading channel (i.e., single tap)
y [m] = h[m]x[m] + w [m]
▶ h[m]: a tap (e.g., zero’th tap ho [m]) of discrete-time baseband
equivalent model.
▶ Flat-fading =⇒ Small delay spread =⇒ One-tap.
▶ What if |h[m]| is very small? This is called a deep-fade.
Step 3: More general frequency-selective channel (i.e., multi-tap)
X
y [m] = hl [m]x[m − l] + w [m]
l
▶ hl [m]: lth tap of discrete-time baseband equivalent model
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
Recap and way forward

We started by studying AWGN channel.

Given a communication channel like AWGN channel,


▶ We can obtain detection rules based on MAP rule and ML rule
uA
⋆ ML rule in terms of LLR function: Λ(y ) ≥ 0
<
uB

▶ Asses performance by analysing probability of error (e.g., in terms of


SNR)
⋆ E.g., probability of error for BPSK and QPSK decays exponentially
with SNR in an AWGN channel

Way forward: study flat-fading channel

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel: Introduction

Flat-fading (single-tap) channel

y [m] = h[m]x[m] + w [m]

▶ x[m] is transmitted in time slot m


▶ y [m] is received signal in time slot m

Statistical model
▶ w [m] ∼ CN (0, N0 ), i.e., complex Gaussian noise
▶ Rayleigh fading: h[m] ∼ CN (0, 1) with normalized variance of 1.
Detection: two types
▶ Non-coherent: no assumption on prior knowledge of h[m] at receiver
▶ Coherent: make assumptions on knowledge at receiver of h[m]

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: Introduction
Non-coherent: no assumption on prior knowledge at receiver of h[m]
Flat-fading channel: y [m] = h[m]x[m] + w [m]
How to design a digital communication system for this channel?
Involves designing transmission and reception/detection scheme.
Transmission scheme: maps digital input to baseband x[m] samples.
▶ Example 1: One input bit mapped to one x[m] sample (e.g., x0 = −a
and x1 = +a)
⋆ Digital input: b0=0, b1=1, b2=1, b3=0,...
⋆ b0=0− > −a, b1=1− > +a, b2=1− > +a, b3=0− > −a,...
▶ Example 2: One input bit mapped to two x[m] samples (e.g.,
xA = [a 0], xB = [0 a])
⋆ Digital input: b0=0, b1=1, b2=1, b3=0,...
⋆ b0=0− > [a 0], b1=1− > [0 a], b2=1− > [0 a], b3=0− > [a 0],...
Reception scheme: maps baseband y [m] samples to digital output
▶ Example 1: Each y [m] sample maps to a decoded/output bit
▶ Example 2: Two y [m] samples map to a decoded/output bit

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


End-to-end system

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: BPSK fails

Non-coherent: no assumption on prior knowledge at receiver of h[m]


Flat-fading channel: y [m] = h[m]x[m] + w [m]
How to design a digital communication system for this channel?
Can BPSK with x[m] ∈ {−a, +a} work?
▶ Fails even without noise because:
⋆ Phase of y [m] is uniformly distributed between [0, 2π]
⋆ Magnitude of y [m] is independent of transmitted symbol
▶ To understand the failure of BPSK, consider an even simpler setup:
y = hsimple x, where hsimple takes values -1 and +1 with probability 0.5
⋆ h[m] can be viewed as a more complex counterpart of hsimple
⋆ hsimple : discrete random variable
⋆ h[m]: continuous random variable

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: Information
via energy

What if we send one of xA and xB :


       
x[0] a x[0] 0
xA = = , xB = =
x[1] 0 x[1] a

Form of binary pulse position modulation over a pair of time samples


”Information” stored in which of the pair of time samples has more
”energy”

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: Intuitive take

 
y [0]
Detection is based on received signal y = , where
y [1]
y [0], y [1] ∈ C
Intuitive detection rule
▶ If xA is transmitted, y [0] = ah[0] + w [0] and y [1] = w [1]
▶ If xB is transmitted, y [0] = w [0] and y [1] = ah[1] + w [1]
▶ Thus, comparing |y [0]| and |y [1]| appears promising.
⋆ Detect xA if |y [0]| > |y [1]|
⋆ Detect xB if |y [0]| < |y [1]|
Now, let us derive a detection rule more rigorously using ML rule

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: ML rule
ML rule in terms of LLR function:

f (y |xA )

xA
Λ(y ) = log =≥ 0
f (y |xB ) <
xB

How to determine conditional distribution f of y ?


If xA is transmitted,
▶ y [0] = ah[0] + w [0]. Then, y [0] ∼ CN (0, a2 + N0 )
⋆ Follows from property of linear combination of scalar Gaussian r.v.s.
▶ y [1] = w [1]. Then, y [1] ∼ CN (0, N0 )
If xB is transmitted, similarly,
▶ y [0] = w [0] =⇒ y [0] ∼ CN (0, N0 )
▶ y [1] = ah[1] + w [1] =⇒ y [1] ∼ CN (0, a2 + N0 )
In both cases, y [0] and y [1] are independent since w [0] and w [1] are
independent

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: ML rule
If xA is transmitted, y [0] ∼ CN (0, a2 + N0 ) and y [1] ∼ CN (0, N0 )
If xB is transmitted, y [0] ∼ CN (0, N0 ) and y [1] ∼ CN (0, a2 + N0 )
In both cases,

y [0], y [1] are independent (1)


yR [0], yI [0] are independent, yR [1], yI [1] are independent (2)

Thus LLR function is:


f (y |xA ) (1) f (y [0]|xA ) f (y [1]|xA )
   
Λ(y ) = log = log
f (y |xB ) f (y [0]|xB ) f (y [1]|xB )
f (yR [0]|xA ) f (yI [0]|xA ) f (yR [1]|xA ) f (yI [1]|xA )
 
(2)
= log
f (yR [0]|xB ) f (yI [0]|xB ) f (yR [1]|xB ) f (yI [1]|xB )
 
a2 |y [0]|2 − |y [1]|2
= , subst. scalar Gaussian’s distribution
(a2 + N0 )N0

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection: ML rule

Thus LLR function is:


 
a2 |y [0]|2 − |y [1]|2
Λ(y ) =
(a2 + N0 )N0
xA
Recall: ML rule Λ(y ) ≥ 0.
<
xB
So, ML rule says that we should decide xA is transmitted if
|y [0]|2 ≥ |y [1]|2 , and decide xB is transmitted otherwise
▶ Also called an energy or a square-law detector
▶ Same as our rule from intuitive analysis
Now we have a detection rule. How well does it perform? Let us look
at its probability of error.

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel non-coherent detection using ML:
Performance analysis
Due to symmetry of setting, we can show that (similar to AWGN case)

pe = P {error|xA is transmitted}
How does an error occur when xA is transmitted? This happens when
|y [0]|2 < |y [1]|2 . Conditioned on xA being transmitted
▶ Note: y [0] ∼ CN (0, a2 + N0 ) and y [1] ∼ CN (0, N0 )
Property of w ∼ CN (0, σ 2 ): |w |2 is exponentially distributed with
mean σ 2 , i.e., u = |w |2 has PDF σ12 exp − σu2

n o
Thus, pe = P |y [0]|2 < |y [1]|2 |xA is transmitted which is equal to
Z ∞ Z ∞    
1 u0 1 u1
2
exp − 2 exp − du0 du1
u0 =0 u1 =u0 a + N0 a + N0 N0 N0
−1
a2

= 2+
N0
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
Flat-fading channel non-coherent detection using ML:
Performance
Recall
average received signal energy per (complex) symbol time
SNR =
noise energy per (complex) symbol time
1 2

2 a +0 a2
= =
N0 2N0
Probability of error for detection using ML:
1
pe =
2(1 + SNR)

Probability of error does not decay exponentially with SNR


▶ Probability of error of 10−3 requires SNR ≈ 500, i.e., 27 dB.
▶ For AWGN channel, recall BPSK probability of error ≈ exp(−SNRB ).
So, SNRB = 9dB is enough for pe = 10−3

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading non-coherent detection using ML rule:
Summary
Flat fading channel: y [m] = h[m]x[m] + w [m]
▶ w [m] ∼ CN (0, N0 ), Rayleigh fading: h[m] ∼ CN (0, 1)
Transmit one of xA and xB :
       
x[0] a x[0] 0
xA = = , xB = =
x[1] 0 x[1] a
 
y [0]
Detection based on received signal y = , y [0], y [1] ∈ C
y [1]
ML rule: decide xA is transmitted if |y [0]|2 ≥ |y [1]|2 , and decide xB
is transmitted otherwise. Its probability of error is given by:
1
pe =
2(1 + SNR)
Probability of error decays relatively slowly with SNR
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
Flat-fading channel coherent detection: Introduction
Our previous attempt to use wireless/fading channel was not
promising as pe decayed slowly with SNR. Was it because of
non-coherent detection, i.e., no knowledge of h[m]?
We next analyse coherent detection where receiver knows h[m]
▶ The knowledge is achieved by sending known signals called pilots p
(i.e., they are know to receiver also) and receiver estimating h[m].

y [m]
y [m] = h[m]p[m] + w [m] =⇒ h[m] ≈
p[m]
▶ It should perform at least as good as non-coherent detection.
We look at below one-shot version of the problem (like for AWGN):

y = hx + w

▶ w ∼ CN (0, N0 )
▶ h ∼ CN (0, 1)
▶ x, y ∈ C
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
Flat-fading channel coherent detection: Analysis using
sufficient statistic

We map the problem into a scalar detection problem using the


following mapping of y to r :
 ∗ 
h
r =R y = |h|x + z
|h|

▶ Note that R {.} is the real part operation. Hence, r ∈ R


▶ Also, z ∈ R. Further, z ∼ N (0, N0 /2). Variance is halved as we only
consider the real part.
r is a sufficient statistic of received y to detect x
▶ Applying quote from Fisher’s 1922 paper in Wikipedia page: r is
sufficient statistic of y to detect x if no other statistic calculated using
y provides any additional information about x

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel coherent detection: Probability of error
r = |h|x + z is like the one-shot analysis of y = x + w for AWGN
analysis with additional scaling term |h|
Reusing AWGN analysis, we can conclude that for a given h, if we
transmit one of x = +a or x = −a, then
! q 
a|h| 2
P {error|h} = Q p =Q 2|h| SNR
N0 /2

where SNR = a2 /N0 .


Thus, probability of error is (last step derived in next slide)
 q 
2
pe = E [P {error|h}] = E Q 2|h| SNR
r !
1 SNR
= 1−
2 1 + SNR

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel coherent detection: Probability of error

Probability of error is
 q 
R∞ √ 
2
E Q 2|h| SNR = a=0 Q 2aSNR e −a da (3)
R∞ R∞ −b 2 /2 e −a dbda
= √ √1
a=0 b= 2aSNR 2π e (4)
R ∞ R b2 /(2SNR) 1 −b2 /2 −a
= b=0 a=0 √ e

e da (5)
R ∞ 1 −b2 /2  
= √ e 1 − e −b 2 /(2SNR) db
b=0 2π
 q 
1 SNR
= 2 1− 1+SNR (6)

where (3) uses exp. distribution of |h|2 , (4) uses Q(.) definition, (5)
involves exchange of order of integration, (6) utilizes the fact that
integral over half of real axis of Gaussian PDF is equal to 0.5.

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Flat-fading channel coherent detection: Probability of error

Thus, probability of error with coherent detection is


r !
1 SNR
pe = 1−
2 1 + SNR

Using Taylor approximation for large SNR, we have


1
pe ≈
4SNR
Even with coherent detection, pe decays slowly with SNR and in
particular not decaying exponentially.

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


What explains slow decay of probability of error with fading
Coherent detection analysis indicates that poor performance with
fading is not due to lack of knowledge of channel at receiver.
Figure below is from [1]:

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


What explains slow decay of probability of error with fading
Coherent detection analysis indicates that poor performance with
fading is not due to lack of knowledge of channel at receiver.
Probability of error for coherent detection in flat fading is
 q 
pe = E Q 2|h|2 SNR
 q   n o
≥E Q 2|h| SNR | |h| SNR is small P |h|2 SNR is small
2 2

n o
≈ 0.5P |h|2 SNR is small

where first inequality above follows Law of Total Expectation [Wiki]


considering only one term, and next step follows from fact that Q(.)
is close to 0.5 if its argument is close to zero.
Now, for large SNR (using Taylor’s approximation)
n o n o Z 1/SNR 1
2 2
P |h| SNR is small ≈ P |h| SNR < 1 = e −a da ≈
0 SNR
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
Slow decay of probability of error and deep fades

Thus, for large SNR


n o 1
P |h|2 SNR < 1 ≈
SNR
Also, probability of error for coherent detection is of the same order
since
1
pe ≈
4SNR
n o
Hence, we can conclude that |h|2 < SNR 1
or ”deep fade” plays a
significant role in determining probability of error in high SNR case.

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


How to increase decay rate of probability of error: Diversity

Lessons till now


▶ Slow decay of error probability is primarily because reliable
communication depends on strength of a single signal path.
▶ The single signal path is in deep fade with non-negligible probability,
and will thus see errors with non-negligible probability.
Solution
▶ Ensure that communication uses multiple signal paths each of which
fades independently
▶ With multiple independently fading paths, reliable communication is
possible as long as one of the paths is strong
▶ Diversity: using multiple independently fading paths.
Ways to obtain diversity:
▶ Time diversity
▶ Space diversity
▶ Frequency diversity

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Takeaways
Flat-fading channel non-coherent detection: With ML based
detection, probability of error is given by:
1
pe =
2(1 + SNR)

Flat-fading channel coherent detection: Using Taylor approximation


for large SNR, we have
1
pe ≈
4SNR
n o
|h|2 < SNR
1
or ”deep fade” plays a significant role in determining
decay rate of probability of error
For better decay rate for probability of error, we need diversity which
essentially involves using multiple signal paths thus not being overly
influenced by a deep fade on a single signal path.

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Summary of results

Comparison of various settings and schemes considered


Bits/ Reception
Trans.
Channel symbol scheme (us- pe ∝
scheme
sent ing ML rule)
Closest sym-
AWGN BPSK 1 e −SNR
bol
Any non-
Flat-
BPSK 1 coherent 0.5
fading
scheme
Non-coherent
Flat- 0 : [a 0] 1
0.5 energy detec- ∝ SNR
fading 1 : [0 a]
tor
Flat- Coherent de- 1
BPSK 1 ∝ SNR
fading tection

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Some useful Math examples and exercises

Continuous random variable is completely described by its probability


density function (PDF)
▶ Example: Recall that Gaussian random variable x ∼ N (µ, σ 2 ) with
mean µ and variance σ 2 has Probability Density Function (PDF):

(x − µ)2
 
1
f (x) = √ exp − , x ∈R
2πσ 2 2σ 2
▶ Discrete random variable is completely described by its probability mass
function (PMF)
For a continuous random variable X with pdf fX (.),
Z ∞ Z 3
P(X > 3) = fX (u)du, P(X < 3) = fX (u)du
u=3 u=−∞

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication


Some useful Math examples and exercises
For continuous independent random variables X and Y with pdf
fX (.) and fY (.), and joint PDF fX ,Y (., .),
fX ,Y (u, v ) = fX (u)fY (v )
P(X > 3, Y > 2) = P(X > 3)P(Y > 2)
For continuous independent random variables X and Y with pdf
fX (.) and fY (.), Z ∞ Z ∞
P(X > 3, Y > 2) = fX (u)fY (v )dvdu,
u=3 v =2
Z ∞ Z ∞
P(X < Y ) = fX (u)fY (v )dvdu
u=−∞ v =u
For a continuous random variable X with pdf fX () and a function g (.),
Z ∞ Z ∞
E [X ] = ufX (u)du, E [g (X )] = g (u)fX (u)du
u=−∞ u=−∞

Taylor approximations [Wiki]: for small x


▶ (1 + x)a ≈ 1 + ax; e x ≈ 1 + x
Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication
References

Tse, D., Viswanath, P. (2005). Fundamentals of Wireless


Communication. Cambridge: Cambridge University Press

Vinay Joseph (NIT Calicut) Lecture 11 Wireless Communication

You might also like