Professional Documents
Culture Documents
i
={1,.,i} to be a subset of of size j, B
i,j
= i \ B
i,j,
and
the product over an empty set to be one ie.,
( ) . ., 1
k k k
k
i e p p
|
|
e
e
= =
[ [
1 The transition matrix of
the Markov chain is given by
10 11
20 21 22
,0 ,1 ,2 ,
1 0 0 0
0 0
0
M M M M M
p p
p p p P
p P p p
(
(
(
( =
(
(
(
(2)
Where
, ,
.
c
i i j i j
ij
Bi j k B k B
P pk qk
cO e e
| |
| =
|
\ .
[ [
(3)
for i=1,...,M and j=0,....,i. The transition matrix of the
Markov chain is given by
1 1
2 21 22
1 ,2 ,
1 0 0 0
0 0
1 0
0 (4)
M M M M
q p
q q q Q
q T
qm Q Q Q
(
(
(
(
( = =
(
(
(
(
Where
ij i
Q p =
1, 1 1, 1
c
i i i
k B k B j
pk qk
e e
| |
|
|
\ .
[ [
For
2,..., 1,.... . i M and i i = =
we note that the upper - left sub
matrix of P , which is lower triangular, is the transition
matrix of the Markov chain This implies that for
in the state space of the Markov chain . This implies that for
j, l in the state space of the Markov chain
( ) { } : 0
i
X t t >
( ) ( )
i i
X t n X t j i P + = = (
( )
( )
( )
n
M i jl
X t n X t j p i ( =P + = = =
Where
( ) n
jl
P
is in nth step transition probability from I to l of
the Markov chain
( ) { } : 0
i
X t t >
This is alos true for the Markov chain
{ } ( ): 0
M
Y t t >
As illustrated, the constructed markov chains are used to
an alyze the resequencing buffer occupancy and the
resequencing delay in steady state.
VI. RESEQUENCING BUFFER OCCUPANCY
We lett B
r
represent the resequencing buffer occupancy at
the beginning of a slot, referred to as the observation instant.
At the end of the slot, a block of M PACKETS Are received.
The probability generating function G
B
, (Z) B
r
of is given
in the following theorem.
( )
1
1 2
1
1 1 0 0
...
r O
x M i
B M
i n x x
G z P
= = = =
= +
( ) ( )
2
1
1
0 1
, , ,...., 1 (7)
n
n
x M
j
j j n
x j
p q z C n i x x
= =
| |
+ P (
|
\ .
[
Where
( )
1
, , ,....., 1
n
C nj i x x P (
Is given in (9) and we have (8) show
Proof: The packet received over channels I at the end of the
slot, at which the resequencing buffer is observed, is
denoted by Co. Due to the dynamic assignment rule, Co has
the smallest sequence number among all packets that have
not been correctly received so far. Therefore, only those
correctly received with Co will contribute to B
r
. We defined
A
n
to be the event that Co has been ( incorrectly) received for
times until the observation instant. if Ao is true, the oldest
packet Co has never been received before, and therefore, the
resequencing buffer is empty at the observation instant. When
A
n
is true for n 1, we let A
n,j
be the subvent that Co is
received for the first time over channel
, 1,.... , i for i M =
and
has been ( incorrectly) received for n times. We denote by tm
the slot at whose end is received for (t + 1)
st
time. When the
event A
n,j
is true, we have the process { X
M
(t) : t = 0,.n} by
considering the packet received over channel in slot 0 as the
packet of our interest ( see Fig.4.). Moreover, and By
conditioning on a sample path of the Markov chain , we have
Where
( )
1 1
, , ,....,
n
C n i x x
(
P
is given in [20] as
1 2 1,
1, 1, 0
.....
n
i x i x x
XP P P
1
1
1
..... (9)
k
n
i x
k
p p
+
=
| |
|
\ .
[
when the events and
( ) { }
,
1 1
M k
X k x k n = s s
fo
( )
1,.... 1
, ,
n
C N i x x
(i.e, the joint event are true, the packets that
are received correctly between slot 0 and slot and have a
larger sequence number than that associated with contribute
to the resequencing buffer occupancy For each channel
,....,M , the number of slots in which packets with a larger
sequence number are received ( either erroneously or
( ) { } : 0 .
i
X t t >
correctly) between lot 0 and slot - max and otherwise, . An
example of is shown in Fig. 4.
If we let B
j
r
represent the number of packets that contribute
to over channel j, it is a Binomial random variable with
parameters
j
and q
j
. According to the assumption of mutual
in dependence between different channels
1
,.....,
M
r r
B B
, are
independent. therefore, when is true, the conditional pgf of
is
( ) / .
,
z n i
r n i
B A
( )
2 1
1 2 1 0
1
, , ,
1
0 0
.... 1 1
n
n
x x i
M
j
r n i m k
j
z
x x x
B A X k x k n
=
= =
(
E = s s
(
, ,
( ) 1 1
M k n i
X X k x k n A ( P = s s
| |
( )
2 1
1 2 1
1
0 0 0 1 .
1
......
n
n
x x M i
j
j j
x x x j n i
p q z
A
= = = =
| |
= +
|
P
\ .
[
( )
1 1
, , , ....,
n
C n i x x
X P (
10
where the last equality is due to te fact that a Binomial
random variable
,1 N p
B
( with parameter N and 1-p) has the
pgf.
.
,1
( ) ( (1 ))
N
N p
B z p z p
= +
11
Since is true if and only if all M packets received in slot -
m (m slots before slot 0) are received successfully.
| |
0
(0) 0 ( 1)
M M Mo
A X X M P P = P = = ( =
12
By conditioning on the events and , the pgf z B
r
of B
r
is
obtained and given in 98). theorem 3 is proved.
B. resequencing Delay
with the construction of Markov chain { X
M
(t) : t 0} the
probability that C is transmitted for the first time over
channel i , for i =1,.....,M, is given in the following lemma.
Lemma 2: For each i=1,2,.M we denote by, the event
that C is transmitted for the first time over channel . Then,
| |
( 1)
1 ,
1
M
i
M i MM
PM M
P
o
o
o
= +
P A =
13
Proof: if occurs, in the block of M packets recieved in .slot
- ( slots before also 0), there is at least one packet that is
correctly received. We let be the number of packets received
correctly in slot - m. In other words,is the number of packets
that are received for the first time in slot 0. By the definition
of
| |
( )
( ) ( )
1
(0) 1
0 1
M M
M
M M
k
X M X M
X M k X M
o
o
=
( P = =
P A= =
( P = =
14
,
, 1,..... .
1
M M
MM
P
M
P
o
o
= =
if = the packets received over channels (M +
1),..M are new packets ( received for the first time) in slot
0. Since C is an arbitrarily chosen packet that must be
received for the time in slot 0
( )
1
1 , ....,
0,
i
if i M M
otherwise
o
o o
= +
PA A= ( =
`
)
15
By conditioning on the event { = } we have the
probability of ^
i
as given in (15). Lemma 2 is proved.
Clearly, the markov chain{Y
M
(t) : t 0} is transient with
the absorbing state 0. if S denotes the fundamental matrix
{26] of the Markov chain {Y
M
(t) : t 0} , then we have.
( )
1
0
n
n
T
=
=
S I T
(16)
we denote by the resequencing delay of the packet C then,
the distribution function D
r
of is given in the following
theorem.
Theorem 3: In steady state, if
| | ( ) , 0,1,.....
r
d x D xm x then P = =
( ) | |
1,0
1 1
0
M i
i qk ik
i k
d Pk S
= =
= P A
and
( ) | |
( )
1
1,
2 2 1
P
M i k
x
i qk ik
k j jo
i k j
d x S P
= = =
| |
= P A
|
\ .
For x 1 where
| |
i
P A
VII IID CHANNEL
Independent identically distributed complex Gaussian
inputs are highly relevant in channels impaired by Gaussian
noise. Some of the arguments for this relevance are that, with
side information in the form of CSI Channel state
information at the receiver.
These are the unique capacity-achieving inputs.
Their mutual information represents very well the mutual
information of proper complex discrete constellations
commonly used in wireless systems. Expression for the
perfect CSI capacity achieved by IID Complex Gaussian
inputs are available. No expressions are available for the
mutual information achieved by IID complex Gaussian inputs
save for the very special case of memory less channels and
straight MoneteCarlo computation is not feasible because it
would entail large-dimensional histograms. Although no
longer capacity Gaussian inputs remain highly relevant. The
mutual information they achieve represents the highest
spectral efficiency that can be attained by Gaussian code
books.
The capacity of achieving in the absence of perfect CSI
become largely unfeasible in certain cases. A simple outer
Monte-Carlo in conjunction with these distributions yields a
semi analytical method that allows quickly and efficiently, the
mutual information. The method accommodates not only
scalar channels, but also MIMO settings.
MODEL OF CHANNEL
Consider n
r
transmit and n
R
receive antennas and let n
R
x n
r
Matrix H represent the discrete time fading channel. Under
block Ray leigh-fading, channel entries are drawn from a zero
mean unit variance complex Gaussian distribution at the
beginning of each fading block and they remain constant for
the n
b
is the coherence time in symbols bandwidth if it is the
frequency domain modeled. This process is repeated for
every block in an IID fashion. There is no antenna
correlation and thus the entries of H are also independent.
Assembling into matrices the input, the output, and the
noise for the symbols within each block, their relationship
becomes
(1)
r
SNR
y HX N
n
= +
Where the input X is an n
r
x n
b
matrix while the output y
and the noise N are n
R
x n
r
matrices. Both X and N have IID
zero - mean- unit -variance complex Gaussian entries. with
that, SNR indicate the average signal-to- noise ratio receive
antenna. Each codeword spans a large number of fading
blocks, in the time and / or frequency domains, which
endows ergodic quantities with operational meaning.
Although a block - fading structure admittedly represents
a drastic simplification of reality, it does capture the essential
nature of fading and generally yields results that are
remarkably similar to those obtained with continuous -fading
models. in fact, for a rectangular Doppler spectrum, an exact
correspondence in terms of the estimation of H can be
established between block and continuous - fading models
whereby
where and are the maximum Doppler frequency and the
symbol period, respectively. Typically, with the velocity and
the carrier frequency. The mapping in (2) is in terms of the
minimum mean square error in the estimation of H, and thus
it is exact for pilo-based schemes that rely on such explicit
estimation, but more broadly we take it as indicative of the
fading rate represented by a given value of n
b
.
A. COMPUTATTION OF THE MUTUAL INFORMATION
The mutual information under investigation can be
expressed as
1
( ) ( )
b
I h Y h Y X
n
= (
where (.) denotes the differential entropy operator.
Our first results leverages the derivations in (13) to obtain
a closed -form expression for
.
(.) exp int , . .,
q
LetE denotethe onential egral i e
1
( ) . .
q
q
E f t e dt Then
,
,
=
Proposition 1
( )
2 1 2 2
/
2
0 0 0
( )
j nT i i j
nT SNR
R
i j
i i l
h Y X n Log e e
n n n
= = =
( | |
=
( |
\ .
E E E
2
2
( 1) (2 )!( )!
2 ! !( )!
2 2 2
l
b T
i
b T j l
j n n l
b r
j n n j
j n n
| |
+
|
|
+
\ .
+
Having expressed ( ) h Y X in (3), we now turn our attention
to (Y) The output Y is affected by a combination of
multiplicative and additive noise, and thus Y is not Gaussian
distributed. The unconditional output distribution
P
(Y),
which had not been reported in the literature to the best of
our knowledge, constitutes the central result in this paper.
while the formula we obtain appears too involved to allow for
a closed - form expression of
h
(Y), the formula is easy to
evaluate and it thereby enables computing
( ) ( )
2
log ( ) (5) h Y p Y p Y dY =
}
| |
2
( ) . (6) E Log p Y =
through straight Monte - carlo averaging. Our formula for
P
(Y), is given in the next proposition, where we use the
standard notation
| | { } max , 0 . z z
+
=
Proposition 2 for 1 k min(n
R
x n
r
) define the functions.
| |
( )
1
( ) exp .
/ 1 1
T R
k n n
k
T T b R
x SNR z z
f x z dz
z SNR n z SNR n n n
+ +
=
`
+ + + =
)
}
(8)
| |
1
.....,
nR R
Let d d d betheeignevaluesof YY and definethen XnRmatr ix Z withentries =
Let d=[d
1
,d
nr
] be the eigenvalues of YY and define the
n
R
x n
r
define the functions
1
,
,
1,
1 1
1 1
i
R
R
j nT
i
i n j s
ij
i n s j s
d
T
Z
n
SNR
s s s s
s s + s s
`
)
| |
|
\ .
Where s = min{ n
r,
n
R
} then
2
!
1
1 ,
( )
( ) det . (9)
( )
T
k
R R
Y
b R
n
i j nR
k n n
n n
p Y Z
dj di
t
s s
= +
c
=
H H
Proof : See Appendix .
We note that, due to the rotational invariance of H, only
the eigenvalues of YY are relevant to the distribution in (9).
Using (4) and ( 9), an algorithm to compute can be put
forth as follows.
Algorithm 1: Evaluation of
1. Pre - compute on a discrete set wirh a suitable step size .
2. Generate a sufficiently large number of input and output
vectors according to (1)
3. For each input and output pair, apply (9) to obtain .
4. Compute the sample average of -
2
log ( ) , ( ). p Y via Montecarlo thereby Obtaining h Y
5. Compute from (4) and apply (3).
In the special case of memoryless channels, i.e., for the
solution in proposition 2 reduces to the one in .
The accuracy can be made as high as desired by average
over more input / output sample pairs and by increasing the
precision in step 1. For the results presents in Section V, the
number of samples and the value of were chosen such that
two decimal digits are correct with 90% probability. With a
standard workstation, the entire computation process is a
matter of seconds.
as a final remark, we mention that proposition 2 is easily
extendable to include all input distribution X that are
rotationally invariant and where the eigenvalue distribution of
X X (or of X X ) is of the form.
2
1
( ) det ( ) 9 ( ),
s
k
k
p V k
=
=
[
where V (.) denotes a Vandermonde matrix and the
functions 9k (.) are arbitrary. Further details are given in the
Appendix.
A1. Baselines
Before exemplifying the method described in Section III,
we introduce the perfect - CSI capacity, a lower bound to ,
and the spectral efficiency achievable with pilot based
communication, all of which serve as baselines.
2
( ) det
T
SNR
c SNR E Log I HH
n
( | |
= +
( |
\ .
If thclosed forms for which can be round in (13) , (14)
B. MUTUAL INFORMATION LOWER BOUND
A simple application of Jensens inequality to the bound in
( 15, Theorem 2) yields the following.
I (SNR) ( )
lower
I SNR with >
2
( ) ( ) (1 ). (11)
b T R
LOWER
b T
n n n
I SNR C SNR Log SNR
n n
= +
C. PILOT - BASED COMMUNICATION.
In pilot - based communication, pilot symbols are inserted
within each fading block, leaving symbols available for data.
The channel is estimated on the basis of the pilot observations
at the receiver, and this estimate is subsequently utilized to
detect the data. We analyze here the spectral efficiency
achievable with separate processing of the pilots and the
data symbols, which refers to estimating the channel on the
basis of only the received pilots and then decoding the data (
through nearest neighbor decoding) as if that estimate was
perfect (16). During the transmission of pilot symbols,
(12)
p p
T
SNR
Y HP N
n
= +
where the output, , and the noise, are matrices. The entries
of are zero - mean unit variance complex Gaussian while P is
deterministic and satisfies .
During the transmission of data symbols, in turn, (1)
applies with X and N of dimension n
R
x (n
b
n
p
). The value
of , which can be optimized by solving a convex problem,
depends on SNR, n
b
and n
r
. This optimization, and the
ensuing spectral efficiency, has been studied extensively, e.g.,
in bits /s/Hz, such spectral efficiency equals.
( )
2
:1
/
1 (13)
1 1 /
max
p p b
p p T
n n n
b p T
n SNR n n
C
n SNR n n s s
| |
| |
|
|
|
+ +
\ .
\ .
where is the perfect - CSI capacity in (10). if the pilot and
data symbols are not required to have the same power, i.e., if
piolt power - boosting is allowed, then it is optimal to set and
to optimize only over the relative powers of pilots and data.
This results in a different convex optimization, which in this
case can be solved explicitly (17) leading to
1
( )
2
1 1 (14)
2
nT nbSNR
C
nb nb nT
| |
| |
| |
\ . \ .
In Bits / s/ Hz and with
2
b T
b T
b
b T
n SNR n
n n
n SNR
n n
+
=
The spectral efficiency in (14) is superior to that in ( 13).
However, Pioly power boosting increase the peakiness of the
overall signal distribution, rendering it less amenable to
efficient amplification.
D. SOME EXAMPLES.
Recalling (2), and in order to calibrate the relevant values
of , the following observation can be made in the context of
emerging systems such as 3GPP LTE or IEEE 802.16
WiMAX:
The carrier frequency Typically lies between 1 and 5 GHz.
The symbol period is T
s
100S.
However, it could be
shortened to T
s
|
\ .
in ,bits /s/Hz/ (3 dB). By activating min { n
r
, n
R
} transmit
antennas and min { n
r
, n
R
} receive antennas, a straight
forward computation shows that the lower bound in
proposition 3 achieves the same high - SNR slope. Since
(SNR) increase with the number of receive antennas ( by the
chain rule, additional outputs are never harmful), implies that
the optimal high - SNR slope is achieved if IID Gaussian
inputs are sent from min { n
r
, n
R
} transmit antennas and all
available receive antennas are used. Based Upon (13), it is
also straightforward to confirm high SNR slope if min
transmit antennas are used with one pilot per antenna.
Exampel 8 Let n
r
, = n
R
, = 2 and Let n
b
, = 10 is the high -
SNR mutual information.
The mutual Information achieved by IID Gaussian inputs
satisfy the perfect CSI capacity and the high SNR expansion
of the tru capacity, for which for n
r
= n
R
is given.
( ) ( ) ( )
2 2
1
1 log , 1
b T
T b T
b b
n n
C SNR n Log G n n o
n e n t
| || |
+ + +
| |
\ . \ .
( )
( )
( )
1
1
2
1 !
, .
2
1 !
i
t
i t n
i
t
i
i
G t n
i
t
t
= +
=
H
+ =
H
Although for n
b
n
R
+ min{ n
r
, n
R
} the input X that achieves
the true capacity for is an isotropically random unitary
matrix IID Gaussian inputs seem to perform very well at
high SNR, even in relatively fast fading. Non - asymptotically
in the SNR, the optimum X is no longer just a unitary matrix
but rather the product of a unitary matrix but rather the
product of a unitary matrix and a nonnegative real diagonal
matrix. No expressions are then available for the true
capacity.
{ } . min ,
b R T R
B Case n n n n Z +
In this case, the optimum X is agian the product of a
unitary matrix and a nonnegative real diagonal matrix, even
for . The high - SNR of the true capacity equals.
{ }
{ }
, , / 2
min , , / 2 1
T R b
T R b
b
n n n
n n n
n
| |
(
| (
|
\ .
but no further expressions are available for the true capacity.
Our method to compute the mutual information of IID
Gaussian inputs continues to apply.
We have presented a method ( part analytical, part Monte
Carlo) to compute the mutual information achieved by IID
complex Gaussian inputs on block Rayleigh-faded channels,
both scalar and MIMo. This mutual information is highly
relevant with Gaussian codebooks.
The method presented may be of further interest to other
multivariate problems involving combinations of
multiplicative and additive Gaussian noise, either with
respect to the mutual information or to the constituting
differential entropies.
{ }
2
2
exp exp
4
B
x A xB dx
A A
t
+ =
`
)
}
The second one is an inteegral due to Itzykson and Zuber.
Given an diagonal matrix B with diagonal entries b, n
arbitary matrix matrix D with eigenvalues d, and an
isotropically distributed unitary reandom matrix U,
{ }
( )
( ) ( )
( ) ( )
1
1 !det ,
det det
m
Tr UDU Z
m
m E d b
e p U dU
V d V b
=
= H
}
where the (i, j) the entry of the M X M matrix E(d,b) equals
{ }
exp
ij i j
E d b =
while V(.) denotes a vandermonde matrix, i.e., such that
3
1
det ( ) ( )
i
i j m
V d d d
s s s
=
The third identify treat when B is of reduced rank,
0 ,
n
if b for n N = Z
1,...., 0
det ( , )
lim
det ( ) N M
b b
E d b
V b
+
| |
( ) ( 1) / 2
1...
1
det ( , ) ( 1)
det ( )
( 1)!
M N M N
N
M N
N
k
k
E d b
V b
b k
[
where the (i,j) th entry of
{ }
1
( , ) exp , 1 .
j N
i j i
E d Z equals d z for j N and d for N j M
s + s s
The final identity was proved in by Cgiani, Wn and
Zanella. Given two arebitrary matrices respectively, and an
arbitrary function
( ) ( ) ( , ) ( ) ( ),
i j i j
x and x with i j thentries x and x + u + u
1
... det ( ) det ( ) ( )
M
m
Dord
m
x x x dx
=
+ u
[
} }
{ }
, 1...
det ( ) ( ) ( )
b
i j
a
i j M
x x x dx
=
| |
= + u
|
\ .
}
Proof of Proposition 1
Conditioned on X, the output Y is complex gaussian
furthermore, the rows of Y are IID conditioned on X. Hence,
to obtain it suffices to evaluate its value for an arbitrary row
of Y and then scale it by the number of rows, i.e., by n
R..
Let y be an arbitrary row of Y. The conditional co-
variance of the n
B
-dimensional coloumn vector equals
T
SNR
y y X I X X
n
E ( = +
and thus
( ) ( )
h y X h y X =
( )
2
log det 1
b
n
T
SNR
E e X X
n
t
( | |
= +
| (
\ .
with expectation over the distribution of X. factoring out the
term , what remains coincides with the perfect - CSI capacity
of a MIMO channel, only with the role of the channel played
by X. Since the entries of X are IID complex gaussian with
zero mean and unit variance, we may directly apply the
closed form in with appropriate dimensioning. Scaling the
end result by
B. Proof of Proposition 2
{ } / , min ,
T T R
define SNR n s n n and r = = =
Max {n
r
, n
R
}. Then denoting by x
t
the column of X,
( ) ( ) ( ) , H p y E p Y H X p X dX
(
=
}
( )
{ }
2
1
1 1
exp
b
n
t t
t T b R b
H E y Hx
n n n n t t
=
(
=
(
[
}
{ }
2
/ ]
t t
X exp x dx
Using the singular - value decomposition
, H U V = E
absorbing V into X through the variable substitution
t t
x V x =
,
and assembling the diagonal entries of
EE into | |
1
,...... , s =
( )
{ }
( )
{ }
2
1
exp
exp
b
n n
T b
n
t t
n n R b
t
Y
p Y EH x x
t
t =
(
= A
(
[
}
{ } { } |
2 exp
t t t
x
X EXP Re y U x dx
`
)
{ }
( )
2
,
1 1
exp b
n s
n n
t k T b
H
Y
E
n n R b t t
= =
(
=
(
[ [
}
{ }
{ }
2
, ,
1
exp 2Re 2Re
t k k t k t k k
X y u x x
| |
E +
` |
\ . )
2
,
,
1
exp .
T
R
n
t k
t k
k n
x
X dx
= +
| |
|
`
|
) \ .
[
Applying to each variable
, t k
x
{ } { } ( )
2
2
1 1
exp Re
( ) exp
nb s
k t k
n n
t k
H
Y y u
p Y E
R b
t
= =
(
(
=
`
(
(
)
[ [
{ } ( )
2
Im
1
exp
1 1
k t k
k
k
y u
X
| |
` |
+ | |
\ .
+
|
\ . )
{ }
( )
2
8
1
exp
exp
1
b R
k k k
n n
k
k
H
Y
U YY u
E
t
=
(
( =
`
+ (
)
[
1
1
b
n
X
k
| |
|
+
\ .
Let ( ) A be a n
R
X n
R
diagonal matrix with kth diagonal
entry
( ) / 1 ,1 0,
k k R
a k k k s and a s k n = + s s = Z s
. It is known
that U, and V in the singular - value decomposition of H
are independent reandom matrices and that both U and V are
isotropically distributed. Thus, we can express
{ }
( )
2
1
exp
1
( )
1
b
R
R b
n
n
n n
k k
Y
p Y p
t
=
| |
=
|
+
\ .
[ }
( ) ( ) { } { }
exp X P U Tr A U U dU d
(
}
where the density distribution of the (ordered) eigenvalues in
equals
( ) ( )
( ) ( )
8 8
2
1
det
! !
r
k k
k R T
e
p V
n k n k
=
=
[
and scaling by 1/s ! yields the corresponding density of the
unordered eigenvalues.
Henceforth, we separately consider the cases
. T R T R
n n and n n s Z
Let ( ) | |
1
,....,
T R nR
n n A a a > =
contain the diagonal elements
of ( ) A and let d contain the eigenvalues of YY
+
. Then,
( )
( )
( )
( )
( ) ( )
( ) ( )
2
1
1
1 ! det ,
1
det 1 det
R
b
R
R b
n
n Y
n
k
n n
k
k e E a d
p Y p d
V d k V a
=
=
| |
=
|
+
\ .
[
[ }
With E (.,.) V(.) respectively More over
( ) ( )
1
1
1 1 det
k
k k
V a
| |
=
|
+ +
\ .
[
( ) ( )
2
1 1
1
k
k k e
+ +
( )
( ) ( )
( )
2
/ 2
1 1
1
det R R
k k
n n
V
+ +
=
[
( )
( )
2
1
1
/2
1
1
det ( )
R R
R
n n
k
n nR
k
V
+
=
[
Plugging into yields
{ }
( )
2
2
2
1
exp
( ) det ( )
det ( ) !
R R
R
n n
n
T
k
Y
p y V
V d n k
}
[
( ) ( )
( )
1
1
det ,
1
T R R
b R
n n k n
k
n n
k
k
e
X E a d d
+
= +
[
{ }
( )
( )
2
1
exp
det /
det ( ) !
R
R b
n
n n
T
k
Y
V
V d n k
t
=
}
[
( ) ( )
( )
1
1
, .
1
n n
T R
R k
b R
k n
n n
k
k
e
X Det E a d d
+
= +
[
The Multiple integral is an instance simply identify
( ) ( ) det / V u =
( ) ( ) ( )
det , E a d =
( )
( ) 1 1
T R
b
n n x
n
R
e x
x
x n
=
+ +
To obtain
( )
{ }
( )
2
1
exp
det
det ( ) !
R R b
n n n
T
k
Y
Z
p Y
V d n k
t
=
[
Where
( )
( )
1
1
/ exp .
1
1
T R
b R
n n
i
ij j n n
o
x x
Z x d x dx
x
x
+
=
`
+
+ )
}
Remark : The proposition can be generalized to also
encompass non-Gaussian inputs. If X has rotational
invariance and the corresponding P ( ) of the form
( ) ( )
( )
( ) ( )
2
1
9
det
! !
R
n
k
k R T
k
p V
n k n k
=
=
[
( )
{ }
( )
( )
2
1
exp
det /
det ( ) !
R b
nR
n n
T
k
Y
p Y V
V d n k
t
=
}
[
( ) ( )
( )
( )
1
1
9
det , .
1
R
b R
n
k
n n
k
k
k
X E a d d
+
= +
[
Hence, proposition 2 applies , but with a different matrix Z. If
the functions g
k
(x) are distinct, needs to be replaced with,
Theorem 2]. This gives a similar, but not identical, result to
Proposition 2 ( the term det Z is replaced by a general tensor -
operator). These observations for non - Gaussian inputs hold
for the case nr Z nR
as well.
2) case nr Z nR :
( )
1,.....
T
n
Let a a a ( =
contain the diagonal
elements of and let d contain the eigenvalues of . The r.h.s of
is precisly the setup in, but in this case the limit in must be
taken. taking that limit and plugging in the density yields,
after a few manipulations.
( )
{ }
( )
( )
2
1
exp
det /
det !
R T
T
R b
n n
n
n n
k
Y
p Y V
V d k
=
}
[
( ) ( )
( )
1
1
,
1
T
b R
n
k
n n
k k
e
X Det E a d d
+
=
+
[
where the density of unordered eigenvalues has been used
and thus the integration limits are 0 to for all variables.
due to the structure of the matrix E ( ( ), ) a d However, the
same result as in still applies,
( ) ( ) ( )
( )
1
1
det / det ,
1
T
b R
n n
n n
n n
e
V E a d d
+
=
+
[
}
( )
1
'
'
' 1
1 exp
1
n T
T nR
n
n n
pn p n n
d
o
o o
o
o o
+
e e =
=
`
+
)
[
}
( )
1
' 1
1 1
R
T
n b R
T
n
n n n
n n n
n n
n
e
X d d
o
+
= +
+
[
( ) ( )
'
1
1 1
1 ' '
T R
T
n
n n T
T R
n n
n n
n n
p p n n n
f d d
o o
o
o o
o o
+
e e = = +
=
[ [
( ) ( )
'
1
' '
1 1
! 1
T R
T
n n
n T R
n n
n n
T n
p n n n
n f d d
o
o o
o
e = = +
=
[ [
!det ,
T
n Z =
Where
( ) ( )
( )
1
1
0
1
/ exp ,
1
1
b R
n
n n n
zy
f x z x z dz
z
z
+
=
`
+
) +
}
denotes the sign of the permutation denotes the set of all
permutations of the integers 1.N, and (i, j) the th entry Z of
equals
( )
ij j i T
Z f d if j n = s
and
1
T
j n
ij i T
Z d if j n
=
.
By combing both cases, the complete statement in
the proposition is obtained.
VIII. CONCLUSION
We conducted performance analysis of the resequencing
buffer for SW-ARQ-inS over a generic number of parallel
channels with both time-varying and time-invariant packet
error rates. With the dynamic assignment rule applied in the
protocol, exact statistical results of the resequencing buffer
occupancy with both channel models were derived in steady
state. The distribution function of the resequencing delay for
the model with time-invariant error rates and the mean
resequencing delay for the model with time-varying error
rates were also obtained. For the model with time-invariant
error rates, we numerically computed the pmf of the
resequencing buffer occupancy using its probability
generating function and the pmf of the resequencing delay.
Through numerical and simulation results, we discussed the
impact of the packet-to-channel assignment rules, the
variance in the error states, the average error rate, and the
number of parallel channels on the mean resequencing buffer
occupancy and the mean resequencing delay. The dynamic
assignment rule always outperforms the static assignment rule
for both channel models. For MSW-ARQ-inS over parallel
channels with both possibly different time-invariant error
rates and the GilbertElliott model, the mean resequencing
buffer occupancy and the mean resequencing delay increase
with the average error rate and the number of parallel
channels even though the mean resequencing buffer
occupancy decreases with the variance in the error states. In
future work, we can apply the modeling and analytical
approach presented in this paper to conducting performance
studies on the selective-repeat ARQ protocol over parallel
ACKNOWLEDGMENT
I m very much thankful to all My Faculty Members who
helped me, regarding my M.Tech Project, and A special
Thanks to Faculties of KITS College, Allur Institute Of
Management Sciences. ( i.e.,) Sony Madhu Latha Madam,
Ayesha Bhanu Madam and Rajendera Prasad Sir , for helping
me in this Project and Faculties of SR Engineering College.
REFERENCES
[1] Z. Ding and M. Rice, Throughput analysis of ARQ protocols for
parallel multichannel communications, in Proc. IEEE GLOBECOM,
2005, pp. 12791283..
[2] M. E. Anagnostou and E. N. Protonotarios, Performance analysis
of the selective repeat ARQ p.
[3] X. Deng and A. M. Haimovich, Achievable rates over timevarying
Rayleigh fading channels, IEEE Trans. Communications,
R. Fantacci, Queueing analysis of the selective repeat automatic repeat
request protocol wireless packet networks, IEEE Trans. Veh. Technol.,
[4] E. H. Miller, A note on reflector arrays (Periodical styleAccepted for
publication), IEEE Trans. Antennas Propagat., to be published.
[5] W. Stallings, Data and Computer Communications, 6th ed. Upper
Saddle River, NJ: Prentice-Hall, 2000, ch. 2.
[6] S. B. Wicker, Error Control Systems for Digital Communication and
Storage. Upper Saddle River, NJ: Prentice Hall, 1995, ch. 5. of feasibility
(Periodical style), IEEE Trans. Electron Devices, vol. ED-11, pp. 3439,
Jan. 1959.
[7] J.-F. Chang and T.-H. Yang, Multichannel ARQ protocols, IEEE
[8] www.classle.net/projects/project_ideas/resequencing-analysis-st...
[9] http://en.wikipedia.org/wiki/Stop-and-wait_ARQ
[10] http:// en.wikipedia.org/wiki/Automatic_repeat_request
[11] Jin Fu Chang, Senior Member, IEEE and Tzu Hui Yang Multichannel
ARQ Protocols.
[12] M. Moeneclaey Throughput optimization for a generalized Stop-and-
wait ARQ scheme, IEEE Trans. Commun, Vol COM-34,pp 205-207, Feb
1986.