=
+ =
K
i
i
m Z m X m Y
1
) ( ) ( ) (
) 1 ( log
2
1
1
2
2
=
+ =
K
i
i
P
C
o
R
1
R
2
C
1
C
2
Fading channels (single timescale)
Single user channel
Multiple access channel
K number of users, H
i
fading of user i, random
ergodic process, independent between users, X
i
transmitted information of user i, Z noise (with variance
1)
Assume receiver and transmitters can track all of the
fading components perfectly, ie. full channel state
information (CSI)
) ( ) ( ) ( ) ( m Z m X m H m Y + =
=
+ =
K
i
i i
m Z m X m H m Y
1
) ( ) ( ) ( ) (
Ergodic capacities without power control
Transmit with constant power (of 1)
Single user capacity
Multiuser sum capacity
where
(

.

\

+
=
K
i
i
H
1
2
1 log
2
1
E
H
( )
(
+ H 1 log
2
1
E
2 H
) , , , (
2 1 K
H H H = H
Ergodic capacities with power control
Suppose we allow dynamic power allocation, subject to
an average power constraint (of 1) for each user
Can get improvements in capacity
Single user [Goldsmith, Varaiya, 93]
Multiuser [Knopp, Humblet, 95]
 power allocated to user i, function of
They solve the following optimization problem
) (H
i
  K i
H C
i i
K
i
i i
, , 1 , 0 ) ( , 1 ) ( E to subject
) ( 1 log
2
1
E max
1
2
= > s
(

.

\

+ =
=
H H
H
H
H
H
) (
H
Outline
Introduction and description of work done
Adaptive coding with power control for optimizing average spectral
efficiency
Power control in multiple access channels
with two timescale fading
Future work Estimation problems in
wireless sensor networks
Adaptive coded modulation over fading
channels
Single user capacities
How to achieve capacity? 2 ways
1. Continually adapt the rate of transmission to match the
fading conditions. Possibly need an infinite number of
different codes.
2. Use one rate but code over many different fading
realisations [Caire, Shamai, 99]. Can result in long delays.
Motivating question: How close to capacity can we get
by adapting the rate but using a finite number of codes?
( ) ( )
(
+ =
(
+ = H H C H C
H
H
H
) ( 1 log
2
1
E max , 1 log
2
1
E
2
) (
pc 2 npc
=
+
< s +
N
n
n n n
s H s s
1
1 2
) Pr( ) 1 ( log
2
1
N
s s , ,
1
Adaptive coding with power control
Power control can depend on h and
ASE is now
with constraint
Maximizing this over and is hard
Impose additional structure on
=
+
< s +
N
n
n n n
s H H s s
1
1 2
) ) , ( Pr( ) 1 ( log
2
1
s
}
=
0
1 ) ( ) , ( dh h p h
H
s
) , , (
1 N
s s = s
) , ( s h
s ) , ( s h
) , ( s h
Adaptive coding with power control
Choose the following power allocation policy
Channel inversion over the partitioned intervals, eg. N = 3
< s
< s
<
=
h t h s
t h t h s
t h
h
N N
, /
, /
, 0
) , (
2 1 1
1
s
t
1
t
2
t
3
s
1
s
2
s
3
0
0
h
) , ( s h h
Adaptive coding with power control
Problem becomes
Now a maximization over and
Can derive necessary conditions using Lagrangian
methods
}
}
=
=
+
+
=
+
N
n
t
t
H
n
t
t
H n
N
n
n
n
n
n
dh h p
h
s
dh h p s
1
2
1
1
1
1 ) ( to subject
) ( ) 1 ( log
2
1
max
t s ,
s t
Example
Rayleigh fading, noise variance 1, power constraint 1
Adaptive coding with power control
Other work
Asymptotic analysis of the ratio ASE/Capacity for Rayleigh fading
Simple multiuser extension, can get even closer to capacity as
number of users increases
A. Leong, J. Evans, S. Dey, Adaptive coded modulation with power control
for optimizing average spectral efficiency, submitted to SPAWC 2005
Outline
Introduction and description of work done
Adaptive coding with power control for optimizing average spectral
efficiency
Power control in multiple access channels
with two timescale fading
Future work Estimation problems in
wireless sensor networks
Two timescale fading channels
Fading is often due to multiple components, which can
change on very different timescales
G
i
slow fading of user i, F
i
fast fading of user i,
assume independent of each other
Might not be feasible for transmitter to track quickly
varying components, can/will only track slower parts
Assume receiver knows both the fast and slow fading,
but transmitters only the slow G
i
components
Motivating question: How much do we lose in capacity
when compared to the full CSI case?
=
+ =
K
i
i i i
m Z m X m F m G m Y
1
) ( ) ( ) ( ) ( ) (
Power control with two timescale fading
Single user [Dey, Evans, 04]
Our work  Multiuser case
 power allocated to user i, can only depend on the slow
fading.
Transmitters only know the statistics of the fast fading, ie. mean
Average power constraint of 1
Make an additional assumption that the fast fading distributions
are identical between users, with
) (G
i
i F
i
= , 1 ] [ E
Power control with two timescale fading
We have the following problem
Objective can be shown to be concave
Introduce Lagrange multipliers
With assumption that the fast fading distributions are
identical, can solve this using the (generalised) Kuhn
Tucker conditions
  K i
F G
i i
K
i
i i i
, , 1 , 0 ) ( , 1 ) ( E to subject
) ( 1 log
2
1
E E max
1
2
= > s
(
(

.

\

+
=
G G
G
G
G
G
G  F
) (
K i
i
, , 1 , =
Solution
Power allocation is found by inverting , does not have
a closed form in general (example later)
At most one user transmits, user with best slow fade
and above the threshold
i
F g
F g
e
i j g
g
g
e
i i
i i i
i i
F i
i i
j
j
i
i
i
i
i
=
(
+
=
= > >
=
, 1 )] ( [ E satisfy s ' and
) ( 1
E ) ( where
otherwise , 0
, , ), (
) (

1
G
g
g
G
g g
g
i i
g /
g
e
i
= > >
=
, 1 )] ( [ E satisfy s ' where
otherwise , 0
, , ,
1 1
) (
H
h
H
GF H =
Multiuser Diversity
Full CSI case a form of diversity which takes
advantage of random fluctuations in fading, user only
transmits when it is the best. For a large number of users
with independent fading, the best user will usually have
good channel conditions.
Also have multiuser diversity effects in two timescale
case, but only on a slower timescale
Example
Noise variance 1, average power constraint 1
Slow fading lognormal shadowing
Fast fading Rayleigh with
Determining requires solving
for given
where
) (g
i


.

\



.

\

= =
i i i i i i i i
i i i
g
E
g g g
g e
) (
1
) (
1
exp
) ) ( (
1
) (
1
) (
1
2
g g g g
g
g ,
i
dt
t
e
x E
x
t
}
= ) (
1
i F
i
= , 1 ] [ E
Example
Symmetric case (slow fading identically distributed)
16 users
4 users
1 user
Example
Asymmetric case  2 users, user 1 fixed at average SNR
of 0dB
Suboptimal power allocation schemes
Power allocation function in two timescale fading does
not take on a closed form in general, inverting
transcendental equations numerically not that nice.
One simple suboptimal scheme (symmetric case, K is
the number of users)
Only the best user transmits, with transmission at the constant
power K.
This policy satisfies the average power constraint of 1 since each
user is best around 1/Kth of the time
Some Numerical Results
users C
tts
R
subopt
2 1.29957 1.25836
4 2.10578 2.09873
8 3.02345 3.02273
16 3.94552 3.94545
32 4.83875 4.83874
64 5.69960 5.69960
C
tts
R
subopt
10 0.50693 0.41762
0 1.29957 1.25836
10 2.60070 2.59484
20 4.17340 4.17307
30 5.81884 5.81883
40 7.47751 7.47751
(dB) SNR
We have proved convergence of the two quantities
under certain conditions, when either the number of
users or average SNR goes to infinity
Average SNR = 0 dB 2 users
Idea of proof  large users case
Let
Show
Then show
Other work Derived an upper bound on the growth rate
of the sum capacity, cant be any faster than log K
A. Leong, J. Evans, S. Dey, Power Control and Multiuser Diversity in
Multiple Access Channels with Two TimeScale Fading, Proc. WiOpt 05, to
appear
) , , max(
1 K m
G G G =
0 some for as 1 ) ) ( Pr( > + < < c c c K K G K
m
K R C
subopt tts
as
Outline
Introduction and description of work done
Adaptive coding with power control for optimizing average spectral
efficiency
Power control in multiple access channels
with two timescale fading
Future work Estimation problems in
wireless sensor networks
Sensor networks
Key technology for the 21
st
century
Sensors are/will be cheap, small, have sensing,
computation and communication ability
No central coordination, adhoc
Communication between sensors usually over wireless
links
Challenges in wireless sensor networks
Many sensors sharing same communication medium
Communication constraints limited bandwidth, energy,
computational power
Noise and fading
Node mobility
Limited node lifetime  battery power, hostile forces
Algorithms probably need to be decentralized
Resource allocation issues also, eg. power/energy, rate
allocation/quantization
Distributed signal processing
Much work on distributed detection/hypothesis testing,
less so on distributed estimation, especially for
dynamical systems.
Some problems in the design and analysis of
decentralized estimation algorithms for dynamical
systems via networked sensors
Decentralized estimation with random packet loss
Asymptotic performance analysis of decentralized estimation
Decentralized estimation with random
packet loss
Packets may not arrive due to collisions or fading
Can do Kalman filtering with random packet loss
[Sinopoli et al., 2004],
Discrete time linear state space model
Packet loss Bernoulli process
Probability of reception must be above a critical value for
meaningful state estimation
Treat analogue for hidden Markov models (HMMs)
Allow correlation in packet loss process, eg. modelled by
finite state Markov chain
Asymptotic performance analysis of
decentralized estimation
For distributed binary hypothesis testing, having identical
sensors (ie. same thresholds) is optimal as number of
sensors goes to infinity [Tsitsiklis, 88]
Also holds with capacity and power constraints [Chamberland,
Veeravalli, 04]
No such results currently exist for distributed estimation
of ergodic dynamical systems
Are there performance bounds (eg. CramrRao bounds)
when the number of sensors and time goes to infinity?
Identical sensors results?
Summary
Work done
Adaptive coding with power control for optimizing average
spectral efficiency
Power control in multiple access channels with two timescale
fading
Future work
Decentralized estimation with random packet loss
Asymptotic performance analysis of decentralized estimation