Unit 2

Statistical Multipath Models and

Capacity Analysis

CO2: Utilize various signaling schemes and capacity analysis for fading channels (K3)
 Topics
• Time-Varying Channel Impulse Response
• Narrowband fading models
• Wideband Fading Models
• Capacity of Flat-Fading Channels
• Channel and System Model-Channel Distribution
Information (CDI) Known
• Channel Side Information at Receiver-Channel
Side Information at the Transmitter and Receiver
• Capacity with Receiver Diversity-Capacity
Comparisons
• Capacity of Frequency-Selective Fading Channels
• Time-Invariant Channels-Time-Varying Channels
Statistical Multipath Models
 Types
 Large scale fading
 Path loss and shadowing
 We will model the mobile
communications channel
 as a multipath channel
 with a random time-varying impulse
response.
 Different paths
 Arrive at different times (delay)
 Have different strengths (gain)
 May have different carrier frequencies
due to mobility and direction (Doppler)
v cos 
f 
v
cos 
c
c 
Signal
Model
 Transmitted bandpass signal with
equivalent
lowpass
s(t)  u(t)e jt 2f  {u(t)}cos c t  {u(t)}sin c
signal c

  t
 Communication takes place  Bs / 2, f c  Bs / 2]
 at [ fc
Processing takes place at baseband [B s / 2, Bs
/ 2]
 Bandwidth interchangeably by Bs , B, W
 Example: I/Q modulation/demodulation

with
sampling

D/ A/
A D

EE728 METU 5
AOY
 Received signal without
noise Messag
resolvable
Number e
multipathof
signal
componen
ts
(t )) 
 N (t ) 
j 2fc (t  n
Dn

r(t)      (t)u(t  (t))e
n n
(t )  

n1
Phase due
to Doppler
dela frequency
amplitu
de y n
t
 n
D (t)  2f D
( )d
0
 Resolvabilit
y?
puls
e

t f
T 1/
 Roughly, pulse duration~ 1/ T
 B linear
Interested in pulses since
modulation consists of a train of pulses
where each pulse carries information
on its amplitude/phase. |    |1/
1 2
 Two paths are resolvable if
B
 reflector
cluster

Path distances are

very close.
~ 1/ B

1 2 t
3
Individual paths are not
distinguished. => non-
resolvable

1 2 t
3
 Path distances
are much
different.

~ 1/
B ~ 1/
B

t
1
2
Individual paths are observed =>
resolvable
 N (t ) 
 j 2 f c  n
(t )  j 2fc 
 (t ) Dn u(t   n (t)) e t 
r(t)    (t)e
n
 
  n1
 (t )  2f  (t )  
(t )
n c n Dn

 
 N (t )  j n j 2fc

r(t)      n (t)e
(t ) u(t   n (t)) e t
 
  n1
  (t) a function of path loss and
n
  (t)
depends on delay and
shadowing
n

DopplerTwo random processes

 Assumption:

are independent.
  
 N (t )  j n j 2fc

r(t)      n (t)e
(t ) u(t   n (t)) e t
 
 n1



   
N

c( , t) 
(t )
n (t)e  j n
(t )  ( n (t))
 n1

Equivalent lowpass response of the
channel at time t to an impulse at
time t  
Examples
 At time t there is no physical
reflector with multipath delay n
(t )  
c( , t) 
 0
 For time-invariant
c( , t)  c( , tchannels
 T ) for
any T
 In particular
c( , t )  c( , t  t )  c( ,0)  c( )
c( )
0 
 Nonstationary channel (channel
response changes with time, i.e., it
is a function of time)

Propose an example
scenario!
 Delay
spreadT  max |   |
m 1
n
n 1st path is usually LOS
(line- of-sight) and
strong. Hence, receiver
usually synchronizes to
c( the 1st path.

)
0 1   
 If delay spread
3 2 much smaller than
pulse duration
Non-resolvable
Tm 1/ paths
Narrowband fading
B
Narrowband Fading Models
Tm 1/
B Observed
signal

t
 Delay associated with the ith multipath
has
 i  Tm  u(t   i )  u(t) for all i
  
 N (t )  j n j 2fc

r(t)      n (t)e u(t)e
(t ) t

 n1

 j n 
 N (t )
c N (t)     n (t)e
(t ) 
 n1

N
(t )
c N , I (t)   n (t) cosn (t)
n1
N
(t )

c N ,Q (t)     n (t)
sinn (t)
n1
 (t)  2f  (t)  
 Recall the central limit theorem
 If number of paths are large
 channel gain becomes a complex Gaussian
r.v.
 whose amplitude (envelope) is
Rayleigh distributed.
 In rich scattering environments, there
are many paths and thus channel gains
are taken to be complex Gaussian.
20
0

15
0
 Received power
10
variations due to
0 constructive/destr
5
0
uct ive addition
of paths
0
0 5
0
10
0
15
0
20
0
 These
distances
variations
5 occur over
 Small-

short
 scale
on the order
wavelength~
of
0  0.4
called
FADING

- EE728 METU AOY 18

0 5 10 15 20
 Large and small scale propagation
effects together
 Example
 1Ghz carrier, delay of 50ns (a typical value
for an indoor system)

2f c n  2 50 1
 A small change in delay corresponds to a
large phase change
 Even an additional delay of ns (0.3m)
corresponds to full rotation
 Hence fading may change
considerably (depending on
carrier frequency) with distance.
 N
(t )

  n (t) cosn
c N , I (t)  n
1N
(t) (t )
c N ,Q (t)     n (t) sinn (t)
n1
2f c n 
 Assumption: Phase changes
due to large carrier frequency and
1
fast
hence it is uniformly distributed
 N (t ) 
E[c N , I (t)]  E    n (t) cosn 

N
(t) (t )
n1

  E[ n (t)]E[cos n (t)]  0

n1
 N
(t )

Similarly, E[c N ,Q (t)]   E[ n (t)]E[sin n (t)]
0
 c N (t) is a zero-mean n1 Gaussian

process.
 Derivations based on key
assumptions that generally apply to
propagation scenarios without a
dominant LOS
 Assumption: Amplitude, multipath
delay, Doppler frequency change
 n (t)enough
slowly   n ton
be
(t)  f
considered
D (t)  f D
n
constant
 over time intervalsn of interest
n
 t

t

D n (t)  2f Dn ( )d  2f nD
0

 Autocorrelation for a fixed

setting
Rc (t, t   )  E[c N , I (t)c N , I (t
  )]  
N ,I
 E  n  n cos n (t)m  m cos m (t

  )
 n
 E[ n m ]E[ cos n (t) cos m (t
m

  )] E[  2
]E[cos  (t) cos  (t
   )]
n n n
m
 n (t)  2f c n  2f Dn t
 0
E[cos n (t) cos n (t   )]
 1
 )]  E[cos(4f c n  4f t  2f D  
D
2

n

E[cos(2f D
n 20 )]
n

term changes rapidly w.r.t.
4fc n
other phase terms -> uniform dist.
R
c N ,I
(t, t   ) 
2
 E[ 2 ]cos(2f
)
1 m n
Dn
cos

1  E[n2 ]cos(2v

n
)
2
m 

 By the derived mean and
autocorrelation,
 cc has
(t)the same
is WSS
NN ,I,Q(t)

 autocorrelation.
Cross-correlation
Rc ,c (t, t   )  E[cN ,I (t)cN ,Q (t
  )]
cos

N ,I N ,Q
2
 E[n ]sin(2v )
m  
 E[c N ,Q (t)c N , I (t
  )]
 Overall, c(t) is also WSS.
c(t)  c N , I (t) cos 2f c t  c N ,Q (t) sin 2f c t
Rc (t,t   )  Rc ( ) cos(2f c )  Rc ,c
( ) sin(2f c )
N ,I
N ,I
N ,Q

 We now need an assumption to

evaluate autocorrelation.
 Dense scattering environment =>
uniform scattering
 The channel consists of many
scatterers densely packed in angle
 It is also assumed that each
multipath component has the
same power
E[ 2 ]  2P /N
n r
 n  n  
, N
2
N
R
1
c N ,I
( ) 
2  cos(2v cosn )
n
1 N
Pr cos

2  cos(2v n
)
n  
 1 
 Let’s take the limit as
N  ,   0
2
Pr cos
Rc N ,I ( )
 2 cos(2

v
 
)d
0

J n (z) 
Bessel func. of the first
kind

Rc ( )  Pr J 0 (2f D )
N ,I

Total received Maximum Doppler

power freq.
 Signal
decorrelates
over a
distance of
~0.5
wavelength

Note:
Recorrelati
on
 With uniform
scattering
Pr cos
Rc N ,I ,c
 N ,Q
( )
2 
2 sin(2v
0
)d 
 0
 Real and imaginary
components are uncorrelated
and thus independent.
 Power spectral density of fading
at baseband  2P
1
| f |
r D
S (f) (f) ( )}  D 1 ( f / f 2
f
c c N ,Q c N ,I
S N ,I F{R  f 
 D) o.w
0 .
Sc
(f)
N ,I

f / fD
 This PSD can be intuitively explained as follows
 The range of angles for which their cos value

around +-1 is larger in comparison to other values.

 This PSD can actually be directly obtained from the

pdfcos
o (Woodward
f  Theorem)

 PSD is useful for generating fading in simulations.

 Generate white noise

 Pass the noise through a filter whose square is the

desired
PSD.
  Envelope and power distributions
 Real and imaginary components are
Gaussian and independent.

z(t) 2N ,I (t)  (t)

c

 N ,Qc 2
Phase uniformly distributed
 in [0,2 ] has Rayleigh
Amplitude
distribution.
0.

Z  X 2 Y 7

0.
2 6

0.
5

Ν (0, 
0.
4

z  2 z2 
exp ) 2  , z 
0.

f Z (z) 3

  2
 2 0
0.
2

 EE728 METU
0.1

Y 33
AO 0 0. 1 1. 2 2. 3 3. 4 4. 5
  Power has exponential
distribution.
Z 2  X 2 Y 1
2
0.

Ν (0, 
9
0.
8
0.

1 2
) x  7
0.

f Z ( x)
6

2
exp   , x
0.
5


0.

 0 2 2  2 4
0.
3
2
 0.
2

0.0
1 0 0. 1 1. 2 2. 3 3. 4 4. 5
5 5 5 5 5
 There is sometimes a LOS component.
 Real and imaginary parts are not zero-
mean in that case. Qk Xi  Q ~ 
2
 Chi-square distribution 2
k
Z  X Y
2 2 2 i 1

Ν ( 2 2
) Ν ( Y ,  )
X
,
 Power is not exponential now, it
has noncentral chi-square
distribution.
 Amplitude then has Rician
distribution.
 Power
distribution
0.
1

0.0
9

0.0
8

0.0
7
0.0 Rician
5
0.0
0.0 fading
6
4
0.0
s2
parameter
3 K
0.0
2
2
2
0.0
10 5 10 2 2 3
 Rician 0
15 0 5 0

distribution
z 
f Z (z) exp  z 2
 s 2

I 0 ( ), z 
 
  zs22   0 
2 2

LOS power modified Bessel
EE728 METU AOY 36
func. of zeroth order
 Some experimental data does not fit
well into any of the above
distributions.
 More general fading distributions were
developed whose parameters can be
adjusted to fit a variety of empirical
measurements.
 Nakagami-m fading distribution
 m=1 => Rayligh fading
 m=infinity => No fading

m  (K 1)2 /(2K 1) => Rician

(approximately)
 Level crossing rate and average
fade duration

 HW
 Fading in time are sometimes
probabilistically modeled for predicting
it.
 Finite-State Markov channels
(FSMC)
Time-varying SNR

0 
  zN2
 0
 Fading
E s range discretized into regions
,
Rj  { : Aj    Aj1}

FSMC assumes that  stays in the same
region for a duration of T and then
transitions to another pk,
pk,k 1 p
 
k
k p
1
pp,k

 First order models are deficient

 Fade duration statistics can be obtained
Wideband Fading Models
narrowban
d

wideban
d

 A large spreading of a pulse

 Multipaths of a pulse interfere with
subsequently transmitted pulses (Rayleigh
fading demo in matlab)
 Intersymbol interference (ISI)
 Recall the equivalent lowpass
channel impulse response
function c( , t)
 If deterministic,  deterministic
S ( ,  )   c( ,
scattering func.
t)e
c
 j 2 t
dt

 How does a multipath change in
time?
 Doppler characteristics of the
channel
 1

0.5
c I(,t

0
)

-
0.5 4
0
- 30
1 0 2
1 0
2 1
0 3 4 0
0 0 5 t
0 0
 0
 2
0

1
5
Sc(,)

1
0

4
0 0
3
0 0
1 2
0 2 0
0 3 1
0 4 0
0 5 0

0

 
In general, c( , t )is random due to
random amplitudes, phases, and
delays.
 Autocorrelation func.


Rc ( 1 , 2 ;t,t  t)  E c * ( 1 ,t)c( 2 ,t  t) 
R (We

, assume that
; t, t  t) Rthe
( channel
, model is WSS.
; t) 

c

E c * ( , t)c(
1 2 c 1
, t  t)
2
 1 2
 In real environments the channel response
uncorrelated with a multipath at a
associated with a given multipath of delay 21 is 1
different delay since two multipaths are
caused by different scatterers.  
 Uncorrelated scattering WSS model
(WSSUS)
 With
Rc (1 , 2 ; t)  Ec* (1 ,
WSSUS
2 ,t
t)c( 
t)   Rc ( 1 , 1 ; t ) ( 1   2 )  Rc ( 1 ;

t )
S ( ,Rc (),t)e

dt
 Scattering function for j 2random
 t
channels
c
 Scattering function characterizes avg
output power associated with the
channel as a func.
of delay  and Doppler 
=> PSD for 
 Power delay profile (multipath
intensity profile)
autocorrelation with
0 time
difference
Rc ( )  Rc corresponds to
average power
Rc
( ;0)
( )
0 
 Some parameters to quantify delay

spread 

 cR (   ) R c( )d

2
T m

( )d T m
0

T  0 
 Weightin
 R
m
g by
 Rc 0 c
relative
powe
Avg delay rms delay
spread

( )d  r
 Time-varying multipath channel in
freq. domain (channel freq.

response
C ( f ; at
t) t) c( ;

t)e  j 2f d

Under WSSUS,

autocorrelation of C ( f ;
t ) in freq. depends only on the freq.
difference.
 c( ; t) WSS Gaussian  C( f ; t)WSS
Gaussian

RC ( f1 , f 2 ; t)  E C * ( f1 ;t)C( f 2 ;t 

t)   
 
 E   c* (1, t)e j 2f11 d1 c( 2 , t  t)e j 2f2 2 



 
d 2  * j 2f1 1  j 2f 2 2

 E c ( 1 ,  2 , t  
e d 1d 2
t)c( 
 
 

t) e  j 2 ( f 2  f1 )

R c (  , t )e

d

RC (f ; t), f  f1  f 2
 Example: Two sinusoids transmitted
with freq. difference of f
 Calculate the cross-correlation of the
channel response with time difference
of t . That would be RC (f ; t)
. 
R (f )   R ( )e  j
 F.T. of 2power delay profile
f
d

C c
 RC (f )  EC * ( f ; t)C ( f  f ; is a

autocorrelation func. regular
t)

Correlation between frequencies at a given
time
 Channel response approximately
independent at freq. separations f
where RC (f ) 
 The freq.0 Bc where RC (f )  0 for all f
is called
Bc the coherence bandwidth.

Rc ( )  0 for   T  RC (f )  0 for f

1/ T

T typically taken to be the rms delay

spread.
 Other approximations for
coherence bandwidth exists.
 The exact form of delay spread is
not that important for
understanding the general impact
of delay spread on multipath
channels, as long as the
characterization roughly measures
multipath distribution.
 Rc F{.} | RC (f )
( ) |

f
Channel at a particular
band
 Flat fading: Fading roughly equal
entire
across band
the ifB 
 (narrowband)
Frequency selective
Bc fading: Fading
widely across the bandBif Bc
varying
 (ISI)
A signal with symbol
duration
Ts  1/ B 1/ Bc 1/ B 1/ Bc
Ts flat frequency
selective
 Time variations due to
motion
S C (f ,  ) RC (f , t)e  j 2 t
  dt
 Doppler at a certain
frequency
S (0,  )  S ()  R
(t )e  j 2 t dt
C C C

 R (t) is an autocorrelation func. defining
C

how channel response decorrelates


over time.
Doppler power spectrum S C : PSD of
channel response as a func. ofthe Doppler
()
freq.
 | RC (t) | SC (  )
F{.}

t 
 Channel coherence time Tc
 Time-varying channel
approximatelyTc seconds
decorrelates
 later.
Doppler spread Bd  1/
Bd
Tc
 Slow fading channel Ts 
 Fast fading channel Tc
Ts  Tc

 Flat fading channel Ts 1/

 Frequency selective channel Bc
Ts  1/ Bc
 We almost always employ discrete
models to study time-varying
channels.
Discrete-Time Model
 Time-varying impulse response model
is too complicated in its continuous
N
form (t )

c( , t)    n (t)e  jn (t ) (   n (t))

n1
 We know that not all paths are
resolvable.
 Combine the paths to clusters
whose responses will not be
resolvable
 Resolvabili T  1/
 ty B
Distributions ofN (a ,i ) may be
1 i
different c( , t)   k (t)e  j (t ) (  kT
 k

 Model
)
k 0
Capacity of Wireless Channels

Multiuser
Systems
 Capacity, Shannon
 THM: Communication with arbitrarily
small error rates is possible for all rates
smaller than capacity.
 Mutual information, random coding
 Capacity in AWGN channel
 Power constraint: input with P watts
 Channel of bandwidth B
N0
 Noise with 2
PSD
 Sampling at 1/
B  P/B   P  bits/
C 
y[i] B log 
2  n[i]
x[i] B 2
 1  B log 2 1 sec
1  N 0 B /  log  N0 B 
B  bps
SNR 2
Capacity of Flat Fading
Channels
 Assume a discrete-time channel
with stationary and ergodic time-
varying gain
g[i], 0  g[i]
 Channel power gain follows a
distributio
given p g ( x)
 n.
Ex: Exponential dist. with Rayleigh
fading

Gain g[i] is independent of input and
at each time i
noise.
either as an i.i.d. process
 Gain can change
or with some correlation over

time.



Gain is constant over some
 We will mainly
blocklength T work with block
fading channels.


  [i ] 
 Instantaneous SNR at time i
Pg[i]
N0 B
 Dist. of inst. SNR determined by p g ( x)
 Capacity depends on what is known
at the receiver
 Channel dist. info (CDI) at both
 Receiver channel state info (CSI), CDI

at tx Transmitter and receiver CSI at
 both
Transmitter can adapt both its power and rate in
this
case and thus this scenario has the largest capacity.
Channel Dist. Info Known
 Solving the capacity achieving dist. in
mutual information formulations are
complicated.
 Due to fading distributions
 Due to fading correlation
 Finding the capacity is still an open
problem in almost all channel
distributions.
 Even for the simplest case
(Rayleigh channel)
 Discrete input with a number of discrete
mass points
 No closed form solution
Receiver
CSI

  [i] is known to rx at
 time
We willi. only consider p () from now


pon.
g ( x) known by tx and rx
 Two channel capacity definitions

 Shannon capacity (ergodic capacity)

 Outage capacity
  Ergodic (Shannon)
capacity 
C   B log 2 1 
p( )d
 Idea: A very
0 long blocklength, all channel
states are observed.
.. ..
.  Average mutual info. (Capacity theorem .
holds for stationary and ergodic channels)
 No adaptation at tx
 A specific code is used with the channel.
 Code design should incorporate
channel correlation statistics.
 High complexity maximum-
likelihood decoder at rx
 Capacity-achieving code must be
sufficiently long so that a received
codeword experiences all possible fading
states.

E B log
Jensen’s 2 1   B log 2 1
inequality
E[ ]
 Ergodic capacity of a fading channel with rx
CSI only is less than the capacity of an
AWGN channel with the same average SNR.
 Outage capacity
 Applies to slowly varying channels (block
fading)
 There is a fixed code rate in this model.
 Outage
Rate is iflarger
channel cannot
than the accommodate
instantaneous
capacity.
the rate.


 R  for
Tx encodes B log 2 1rate
a data 
 When SNR 
min at time i is smaller than 
min  [i]   min  outag
e
RC
 Outage P( [i]   min ) 
probability Po ut
C o ut  (1 Po ut ) log 2 1 
min 
Normalized capacity usually drawn with


outage probability in plots.

C / B  log 2 1 


Ex: Rayleigh fading with avg SNR
min

20dB
C / Ba  log2 1 
data rate
min    2 C / 1
 min B
2 C / B 1

P( [i]  mi )  1 e  min /1 0 0
 1 100

 n e
CSI both at Transmitter and
Receiver
 Tx and rx observe instantaneous
channel gains.
 Originally capacity was found
[Wolfowitz] for

s[i] stationary and ergodic random
C s capacity
 process representing the channel
of a particular channel
state
 p(s) s
state (fraction
prob. of time) that channel
is in s
C C s p(s)
sAS
 P[i]

 Application

to continuous

fading
C   C p( )d   B log 2 (1  )

 p( )dpower adaptation, additional

Without
0 0
tx CSI does not increase capacity.
 No need for very long blocklength
 More practical since autocorrelation
etc. not accounted
 Alternatively, power may be adapted.
(Noise power normalized)
 [i]  P( [i])
 Average power constraint still holds. At
tx 

0
  [i]) P( ) p( )d P
P(P(
 At E 
) P  is the instantaneous  [i ]  N B
rx
0
Pg[i]
 Definition of fading channel
SNR.
capacity with average power
constraint

 2 P( 
C max B log  1
 )  p( )d
P (  ):  P (  ) p(  )d
P 0
P
0
 Capacity achieved by time diversity
 Fading values quantized to a finite set
 An AWGN encoder-decoder pair for
each fading level

Input x jfor encoder  j
and data  P( 
R  C  B log P(  ))
has avg power j j 2  j
rate  jjP 
 1 
 Optimal power adaptation
 Lagrangian and derivative set to

zero  P(   p( )d    P( )
 2 ) p( )d 
J (P( )) 0  B log 0
P
1

J (P( ))  B / ln 2 
     p( ) 

   
P( ) 1 P( ) 0/ P
P( 1  B / ln 2. / 
 P) P   1
P 
B / ln 2 1 1 1
   

P  
0 
 With the P( ) 
constraint 1 0 1
  
P( )  0
 
0
P  0 
0
 No data is transmitted
below the cutoff.
P(
)
P
 Waterfilling in time (AMC based on
this)
 When channel conditions are good,
transmit
8
 When channel conditions are bad, don’t
use the6 channel
4


0
0 10 20 30 40 50 60
0 0 0 0 0 0

2
5

2
0
1/


1
5

1 EE728 METU 20
00 AOY20
0 10 30 40 50 60
 

P(P) p( )d



0

 1
 1 1
   p( )d 
 0
0  
 1
Cutoff changes with avg power

 Usually no closed form solution for
cutoff 
   p( )d
C  B 2
 0  
 0 log

 There could be cases where suboptimal

strategies may be preferred.
EE728 METU AOY
EE 359 Lecture
Notes
Stanfor-
d
EE 359 Lecture
Notes
Stanfor-
d
  Zero-Outage Capacity/Channel
inversion
 Tx uses CSI to maintain a constant SNR at
rx
 Tx inverts the channel
P( ) 

P 
p( )d  1    1/

 Transmit
E[1/power
] constraint

 Capacity C  B log 2 (1  )  B log 2

(11/ E[1/  ])
 Quite simple, practical

 May exhibit large capacity
reductions
Outage capacity is 0 for Rayleigh
 fading.
 Truncated channel inversion
 Don’t transmit at very bad channel
cases
 Enhances capacity of inversion
significantly
Capacity of Frequency-Selective
Fading Channels
 Application of the ideas in flat
fading
 Time-invariant channels

 Block fading in freq.

 Subchannels of width B with constant
channel gain
 Parallel AWGN channels
with | Power
allocated to
|2H j Pj jth channel
SNR j 
N0 B
 Waterfilling for parallel AWGN channels (EE
533) with total power P

C  B 2 1 N B 
| H j |2 Pj 

max P1 ,P2 ,…:j Pj  j log  0 

P
 
 Why block
fading?
 Time-varying channels
 It is difficult to determine capacity due to
ISI and time variation.
 Approximation: independent bands of
width coherence BW with time-
varying gains
 Block fading
 CSI at tx and rx
 Capacity achieved by waterfilling across
time and frequency
 page 120-121
 HW questions