You are on page 1of 7

4404

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008

Generalized BICM for Block Fading Channels


Santosh Nagaraj, Member, IEEE

AbstractIn this paper, we propose a novel form of bit


interleaved coded modulation (BICM) that we call generalized
BICM (GBICM) for block fading channels. The achievable
diversity order on block fading channels is limited by the channel
itself and not by the code used. Conventional BICM attempts to
maximize the Hamming free distance of the underlying convolutional code. Although this maximizes diversity on independently
fading channels with ideal interleaving, its effect on block fading
channels is not clear. In this paper, we analyze code performance
on block fading channels and seek to optimize it by suitably
modifying the BICM paradigm. The resulting GBICM system is
just as simple and as flexible as BICM is when coding for different
modulations and different kinds of channels. In fact, the same
decoder can be used to decode both BICM and GBICM. We show
code design principles and analysis techniques with examples.
Simulation results supporting the arguments are shown.
Index TermsBlock fading channels, bit interleaved coded
modulation, multivariate transfer function.

I. I NTRODUCTION

ANDWIDTH efficient error control coding techniques


for communication systems have been investigated by
several authors (see [1]). Trellis coded modulation (TCM) was
first proposed by Ungerboeck [2], [3] for additive white Gaussian noise (AWGN) channels. TCM involves joint codingmodulation wherein modulation is an inherent part of the code
itself. Bit interleaved coded modulation (BICM) [4][10] is
a well known technique for bandwidth efficient coding on
fading channels. BICM is today used in several systems (IEEE
802.11g, for example) to provide coding and diversity gains at
high spectral efficiencies. BICM is also extremely flexible in
the sense that encoding and decoding operations are independent of the modulation used. BICM is ideal for independently
fading Rayleigh channels with perfect interleaving so that each
codebit in an error event experiences an independently fading
channel.
BICM separates the aspects of coding and modulation,
thereby allowing concatenation of any convolutional code to
any modulation. BICM is preferred on fading channels for the
reason that it achieves a higher order of diversity than other
techniques of equal decoding complexity [5]. For techniques
with joint coding-modulation, diversity order is equal to the
minimum number of differing symbols between two distinct
code sequences and is always smaller than that achieved by
BICM. With BICM, diversity order is equal to the minimum
number of differing bits (Hamming free distance) between two
distinct code sequences.

Manuscript received August 20, 2007; revised October 24, 2007 and December 29, 2007; accepted January 6, 2008. The associate editor coordinating
the review of this paper and approving it for publication was H. Nguyen.
Dr. Nagaraj is with the Department of Electrical and Computer Engineering,
San Diego State University (e-mail: snagaraj@mail.sdsu.edu).
Digital Object Identifier 10.1109/T-WC.2008.070927

There are several applications today where the channel


falls under neither the independently fading channel model
or the additive white Gaussian noise (AWGN) channel model.
Such channels exhibit fading and provide diversity, but the
number of independent fading realizations within a code block
is small. Several codebits of an error event experience the
same fading gain. For example, the IEEE 802.11g standard
involves encoding of data blocks that are confined to one
OFDM symbol. Due to correlated fading across the OFDM
subcarriers, the maximum diversity order achievable is limited
by the coherence bandwidth of the channel [11]. Performance
of codes over block fading channels have been considered in
[12], [13] among others. Similar channels are also encountered
in systems with slow time-frequency hopping (like GSM,
EDGE, etc. [12]) and are called quasi-static or block fading
channels.
Block fading channels have received considerable attention
in the recent past. Information theoretic aspects and coding
theorems have been provided in [11], [12], [14], [15]. Performance of convolutional codes over block fading channels
have been analyzed in papers such as [16], [17]. Block fading
channels have also been considered in a different scenario
(not applicable to this paper) by other authors recently. They
focus on the multiantenna or the multiuser scenarios [18], [19].
Over block fading channels, the diversity order achieved by
the BICM system is limited by the channel itself and not by
the convolutional code used. Increasing the code complexity
(thereby, the free distance) has no direct effect on performance.
This is because, the code optimization criterion for block
fading channels is not equivalent to increasing the free distance
[20]. The question arises if it would be possible to modify the
BICM model to optimize it for the block fading channel while
still maintaining the flexibility that BICM offers.
In this paper, we propose a novel form of BICM that we call
generalized BICM (GBICM) for block fading channels. The
decoding complexity of GBICM is identical to that of BICM.
In fact, the same decoder can be used to decode both BICM
and GBICM. GBICM is just as flexible as BICM is when
encoding for different modulations. We show code design
principles with examples that achieve performance gains of
12 dB over BICM. We also show performance analysis
techniques for GBICM. Finally, we introduce a novel measure
of code performance on block fading channels to evaluate
BICM and GBICM.
In Section II, we present the block fading channel model
and some preliminaries. In Section III, we present the GBICM
system model and code design principles with examples and
comparisons with BICM. In Section IV, we present performance bounding techniques for GBICM.

c 2008 IEEE
1536-1276/08$25.00 

NAGARAJ: GENERALIZED BICM FOR BLOCK FADING CHANNELS







Fig. 1.

4405



 

xj,k

System model of a BICM encoder with a binary convolutional code C, an interleaver , and, M -ary modulator X

II. P RELIMINARIES

Let L be the maximum diversity order afforded by the


channel. The maximum diversity order achievable by any
rate rc code is
= L(1 rc ) + 1.
(1)

symbol, the correlation introduced by the modulator is only


partially broken. However, the b-interleaver ensures that there
is no modulation-introduced correlation between successive
b-bit groups. This fact differentiates GBICM from TCM,
where the entire correlation introduced by the modulation is
preserved. GBICM is, therefore, a hybrid of BICM and TCM;
(b) Alternatively, one can simply consider a GBICM system as
yet another BICM system, but, with a different bit interleaver
than the conventional one. Note that the b-interleaver is also
a specific case of a bit interleaver.
From a practical perspective, the 2nd picture is attractive.
It shows us that GBICM can be implemented in a system that
currently runs BICM with only one small change the interleaver. Everything else remains the same. From an analysis
perspective, the 1st approach is more relevant. Incorporating
the partial correlation that exists between successive codebits
allows us to obtain tighter performance bounds than that
possible by completely ignoring the correlation.
The receiver we use is identical to the BICM receiver in [5],
but with the modified deinterleaver. We want to mention here
that it is possible to design a better receiver for GBICM by
considering the partial correlation and obtaining metrics for bbit groups instead of bit metrics. However, we choose to work
with the conventional BICM receiver itself in order to maintain
an identical complexity with BICM. The performance loss due
to this is small. We will also assume that the receiver has
perfect channel state information.

This result follows from the Singleton Bound and is valid for
any code of any complexity.

A. Why GBICM Works Product Euclidean Distance (PED)

A. BICM System Model


We will first describe the model of a BICM system (Figure
1) before proceeding to GBICM. Consider a block fading
channel with L independent Rayleigh fading realizations
hj , 1 j L over each code block. A rate rc = kc /nc
convolutional encoder C encodes a block of information bits
and the resulting codebits of the codeword c are interleaved
by . Groups of m interleaved codebits are then modulated
onto M -ary signal constellations x X (either M -PSK or
M -QAM), where M = 2m . The average transmitted energy
per symbol is S, so that E[|x|2 ] = S. We will use a double
index notation (j, k) to emphasize the block fading nature of
the channel. The k-th M -ary symbol transmitted on the j-th
channel is represented by xj,k and is received at the receiver
as a faded and AWGN corrupted symbol yj,k = hj xj,k + j,k ,
where j,k are Gaussian noise samples with variance N0 /2.
The transmitted and received signals can be considered to be
matrices represented by x = [xj,k ] and y = [yj,k ], respectively
for simplicity of notation.
B. Diversity Limit (from Knopp and Humblet [14])

III. G ENERALIZED BICM


With the description of a BICM system above, it is now
very easy to describe a GBICM system. GBICM differs from
conventional BICM in that the interleaver is no longer a bit
interleaver, but is a b-interleaver (b-) for some integer
1 < b < m. The interleaver first groups b successive bits of
the codeword c and then interleaves the b-bit groups. That is,
the b-bit groups are not broken up by the interleaver. Groups of
m/b interleaved bit groups are modulated onto M -ary signals
X . Equivalently, GBICM is simply BICM with a modified
interleaver structure. The b bits of any one bit group are
transmitted on the same M -ary symbol.
NOTE: The difference between GBICM and either BICM
or TCM is in the fact that 1 < b < m. With b = 1, GBICM
is equivalent to BICM and with b = m, GBICM is equivalent
to TCM.
GBICM can be pictured in either of the following two
ways: (a) In a conventional BICM system, the bit interleaver ensures that the correlation introduced by the M -ary
modulator is completely broken. With GBICM, since b-bit
groups of successive codebits are transmitted on the same

c be any two distinct codewords that differ in


Let c and
Nj symbol positions along the j-th channel and the squared
Euclidean distance at the k-th position be d2j,k for 1 k Nj .
The squared Euclidean
distance
c and
c as seen by

Nj between
d2j,k . We have assumed here
the receiver is S j=1 h2j k=1
that dj,k correspond to distances on a unit symbol energy
Nj 2
signal set, i.e., for S = 1. Let
k=1 dj,k = fj . Applying
the Chernoff bound [21] to the conditional pairwise error
probability P2 (c
c|h), we get
P2 (c
c|h)
P2 (c
c)

exp(

S  2
h fj ), and, after averaging,
N0 j=1 j

= Eh [P2 (c
c|h)]

 

1
N0


S
j=1 fj

(2)

for Rayleigh gains hj . Among codes that achieve the same


diversity order ,the one
the higher product Euclidean
Nwith

j
2
distance (PED)
(
d
i=1
k=1 j,k ) for all pairs of distinct
codewords achieves the lower average bit error rate Pb on
the block fading Rayleigh channel [20], [22].

4406

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008


2
5

nearest error in
underlined position

 

 0



 





 







Fig. 2. Distance for single bit error c c 1 with 16-QAM signal sets
depends on the other three bits in the label; If 0010 is transmitted, an error in
the underlined position is at least 4 away, whereas, if 1010 is transmitted,
5
an error in the underlined position is only 2 away.
5

Minimum PED: The Euclidean distance d associated with


the nearest single bit error (c c 1) is fixed with BPSK
or QPSK modulations. However, with 16-QAM modulation,
it is not. The distance depends on the position i of the bit c
on the 16-QAM label and on the values of the other 3 bits
in the symbol. By enumerating the m2m1 possible distances
associated with a single bit error, we can obtain the probability
mass function (PMF) of the distance d. For the unit energy
Gray-mapped 16-QAM constellation used in IEEE 802.11g
(see Figure 2), it is
2
6
3
1
(d ) + (d ),
4
4
5
5

(3)

where, () is the Kronecker delta function.


For GBICM, single bit group errors c c q are more
relevant than single bit errors. There are 2b 1 kinds of single
bit group errors. For example, with b = 2, the single bit group
errors can be (c c 01), or, (c c 11), or, (c c 10)
for any bit group c. Euclidean distance for any bit group error
q depends on the position of the bit group as well as the
values of the m/b 1 other bit groups in the symbol. By an
enumeration of the bit group errors, it is possible to obtain the
PMFs of the distances.The distance PMFs for Gray-mapped
16-QAM are:
pd (c c 01) =
pd (c c 11) =
pd (c c 10) =

2
3
(d ) +
4
5
4
(d )
5
2
3
(d ) +
4
5

6
1
(d )
4
5
6
1
(d ). (4)
4
5

We define the minimum PED (MPED) of an error event as


MPED(c
c) =

min
dj,k

Nj


(
d2j,k ).
j=1 k=1

Fig. 3. Shortest error event (in bold) with the 2-state convolutional code
(generator polynomials (1, 3) in octal) spans two trellis stages.

nearest error in
underlined position

4
5

pd (c c 1) =

(5)

Let w be the Hamming weight of an error event. Suppose that wj of these bitsoccur over the j-th channel for

1 j so
that w =
1 wj . The MPED for BICM is
2 
simply dmin 1 wj , where, dmin is the minimum distance
of the signal constellation X . With GBICM, let wj,q be the
number of bit-groups of the q-th
 kind on the j-th channel for
1 q 2b 1 so that wj = q wj,q f (q) where f (q) is the
Hamming weight of the b-bit error label q. Now, if dmin,q is
the minimum Euclidean distance between any two constellation points differing
by a Hamming weight of f

(q), MPED


for GBICM is j=1 q wj,q d2min,q . If dmin,q > f (q)dmin
for the considered signal constellation and labeling map, the
MPED of GBICM exceeds that of BICM. Consequently,
GBICM has a lower bit error probability than BICM.
Intuitively, the constellation labeling must be such that
neighboring points differ in only one bit-group for good
GBICM
performance. Further, the requirement that dmin,q

f (q)dmin necessitates that f (q) = 1 for all immediate


neighbors of any constellation point. Only Gray mapping
satisfies these requirements. It appears that Gray-maps are
ideal for use in GBICM due to their ability to provide
increased dmin,q with increasing f (q) for commonly used
signal constellations like QPSK, 8-QAM, 16-QAM, and 64QAM.
Example 1: Consider a channel with L = 2 independent
fading realizations h1 and h2 inside every code block. Also
consider BICM and GBICM generated by the 2-state convolutional code with generator polynomials (1, 3) in octal.
Although the free distance of this code is 3, the diversity
order BICM achieves is only = 2 since L itself is 2. The
modulation used is 16-QAM with b = 2 for the GBICM
system. Assume that the all-zero codeword is transmitted. The
shortest error event that can occur with the 2-state code is
shown in Figure 3. This error event corresponds to the decoded
codeword 1110. We will next evaluate the PED for this error
event. As will be shown later, the corresponding GBICM code
also achieves = 2.
BICM: The Hamming weight of the error event is 3 and
the two codewords will differ in three symbols due to bit
interleaving. Since the code achieves diversity = 2, at least
one symbol must be transmitted on each channel hj for j =
1, 2. So, N1 = 2 and N2 = 1, or, N1 = 1 and N2 = 2. In

NAGARAJ: GENERALIZED BICM FOR BLOCK FADING CHANNELS

4407

10

10

64state BICM, L=4


64state GBICM, L=4
64state BICM, L=5
64state GBICM, L=5

2state BICM, L=2


2state GBICM, L=2

10
2

Pb

Pb

10

10

10

10

10

Fig. 4.

11

12

13

14

15
16
Eb/N0 (dB)

17

18

19

20

Performance of 2-state BICM and GBICM on L = 2 channels.

either case, the MPED between c and


c would be
2
MPEDB (000 111) = min (d (c c 1)+


d2 (c c 1)) d2 (c c 1)
32
, for 16-QAM.
(6)
=
25
GBICM: The b-interleaver (with b = 2) ensures that the two
2-bit groups of the error event are transmitted on only two
different symbols. Again, since the code achieves diversity
= 2, at least one of the symbols must be transmitted on
each channel hj for j = 1, 2. So, N1 = 1 and N2 = 1. The
MPED between c and
c in that case would be
MPEDG (00 00 11 10) =

min [d2 (c c 11)

d2 (c c 10)]
64
,
(7)
=
25
for the Gray-mapped 16-QAM signal set. Gray mapping also
maximizes the MPED for the GBICM system. The above
1
analysis shows a 10 log10 (MPEDG /MPEDB ) = 1.5 dB
SNR gain over BICM with 2-state codes and 16-QAM sets
on L = 2 channels. Although we considered only the shortest
error event in our analysis above, we expect similar gains for
other error events as well. We have simulated the performances
of BICM and GBICM on L = 2 channels and the results are
shown in Figure 4. GBICM provides about 1 dB gain over
BICM. This has been supported by our simulations.
Example 2: We have also simulated BICM and GBICM on
L = 4 and L = 5 channels with the industry standard, rc =
0.5, convolutional code of complexity 64 states. The results
are shown in Figure 5 for 16-QAM signal sets and b = 2.
GBICM provides about 1.52.0 dB improvement over BICM.
Note that the gains are higher at higher code complexities.
IV. P ERFORMANCE B OUNDS FOR GBICM
GBICM is inherently a non-linear code since there is partial
modulation-induced correlation between successive codebits.
For any two distinct codewords c and
c, the pairwise error
probability P2 (c
c) depends on (c,
c) and not on c
c.

10

10

12

14

16

18

20

Eb/N0 (dB)

Fig. 5. Performance of 64-state BICM and GBICM on L = 4 and L = 5


channels.
TABLE I
D IVERSITY ORDER () FOR TCM, GBICM, AND BICM WITH 16-QAM
SETS .
Trellis Complexity
2 states
4 states
8 states
16 states

TCM

Diversity Order
GBICM (b = 2)

BICM

1
2
2
3

2
3
4
5

3
5
6
7

The analysis can be simplified by linearizing this code with the


introduction of an additional parameter u such that u = 0 or
u = 1 for each symbol randomly. If u = 1, the M -ary labeling
for the symbol is complemented. The receiver is expected to
know the sequence u so that decoding is not affected.
A. Transfer Function
The generalized transfer function (TF) [16], [21] for
GBICM is different from that for BICM in that, the distance
properties are now required separately for each of the 2b 1
kinds of bit group errors (see Section III-A).
Consider the rate nkcc code as a rate nkcc/b
/b code in terms
of the number of bit groups. To simplify the analysis, we
will assume that nc /b = L. That is, each bit group generated
by the encoder during one trellis transition is transmitted on
a different channel. Note that in many cases, this can be
achieved by considering an equivalent code instead of the
original code. Each of the nc /b bit groups corresponding to a
trellis transition is represented by the term Dj,q , where index j
represents the channel over which the bit group is transmitted
and index q can take values from 1, . . . , 2b 1 to denote the
particular kind of non-zero bit group error. For example, a
rate 1/2 code on an L = 2 channel can be expressed as an
equivalent rate 2/4 code.
Example 3: Consider the GBICM code generated in Example 1 for 16-QAM with L = 2, b = 2 and rc = 1/2. Since
we need nc /b = L, we get nc = 4, i.e., the code should
be considered as an equivalent rc = 2/4 code with generator

4408

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008

00/00 00 + 10/11 01

can be upper bounded as




00/01 00 + 10/10 01

Pb

min (0.5, Pb (h)) p(h)dh,

(10)

01/00 11 + 11/11 10
01/01 11 + 11/10 10

N D1,3D2,1

N D2,3 + N 2 D1,3D2,2

D1,1 + N D1,2D2,1

N 2D1,2D2,2 + N D1,1D2,3

Fig. 6. Trellis and state transition diagram for GBICM with b = 2 and rate
2/4, 2-state convolutional code with generator polynomials (1, 3, 4, 14) in
octal; parameter N denotes the presence of an information bit 1 (see [21]).

polynomials (1, 3, 4, 14) in octal. This means that the basic


GBICM trellis stage corresponds four output bits or two 2-bit
groups (see Figure 6). The first bit group is transmitted over
the j = 1 channel and the second over the j = 2 channel, as
is the case with interleaving over block fading channels [15].
The GBICM transfer function can be obtained by solving the
state diagram in Figure 6.
Diversity Order: The maximum diversity order that can be
achieved by GBICM of a given b is the minimum number of
bit groups an error event spans and is always less than that of
BICM. So, GBICM is not suitable for use on independently
fading channels with ideal interleaving. However, GBICM
may be an excellent choice for indoor wireless LANs where
the achievable diversity orders are limited by the channel
coherence bandwidth. Table I lists the maximum diversity
orders of BICM, GBICM and TCM for rc = 1/2 and 16QAM.
Let w = (w1,1 , . . . , wL,2b 1 ) be the weight profile of
an error event with wj,q bit groups of the q-th kind on
the j-th fading channel. An expansion of the generalized
transfer function gives the weight enumeration factors a(w)
that correspond to the total number of information bit errors
associated with error events of profile w as
b

L 2
1


T (D1,1 , . . . , DL,2b 1 , N )
w
=
a(w)
Dj,qj,q

N
N =1
w
j=1 q=1
(8)
The union bound on the bit error probability of the convolutional code conditioned on the channel realization h is given
by
b 
a(w)P2 (w|h),
(9)
Pb (h)
kc w

where, P2 (w|h) is the conditional pairwise error probability


of two code sequences c and
c that are separated by a weight
profile of w. Consequently, the average bit error probability

where, p(h) is the probability density function of the channel


gains. The min (0.5, Pb (h)) term is necessary to make the
bound tight as the union bound may not converge if h values
are small [16]. Owing to the presence of the min operator
inside the integral, Pb can only be obtained numerically. A
closed form solution does not exist. We will next solve for
P2 (w|h).

B. Expurgated Bound
Let c and
c be any two distinct codewords that differ in
wj,q bit groups of the q-th kind along the j-th channel so
that cj,k = cj,k q for 1 k wj,q . Since GBICM with a
Gray mapped signal set can be considered to be BICM with a
specific kind of interleaver, it satisfies all the conditions listed
in [5] for expurgation of the union bound
Pb,EX (w|h)

   b
2 1
L
= Eu,i 2 1 q=1 wj,q (bf (q))


P (x
z|h, i, u) ,
(11)
x

where, the summation is over all signal sequences x with bit


group cj,k in the i-th bit group position of the (j, k)th symbol
and
z is the sequence nearest to x with cj,k in the i-th bit
group position of the (j, k)th symbol. Here, f (q) denotes the
number of error bits (ones) in the binary label of q. Equation
(11) can be greatly simplified for GBICM on block fading
channels.
Consider the term (12)
P (x
z|h, i, u)
 
2b 1 wj,q 2 2
L
S j=1 q=1
k=1 hj dj,q,k
= Q
(12)
2N0
under ML decoding with perfect channel state information at
the receiver. Here dj,q,k is the Euclidean distance between x
and
z at the k-th symbol of the q-th kind of bit group error on
the j-th channel. Note that dj,q,k > 0 for all j, q, k considered
in the sum and can take only a few different values depending
on i and u. With some algebra, (11) can be simplified as [7]

Pb,EX (w|h)

 

S j,q,k h2j d2j,q,k


.
Edj,q,k Q
2N0
(13)

Now, dj,q,k is not fixed for a given q (see Section III-A).

NAGARAJ: GENERALIZED BICM FOR BLOCK FADING CHANNELS

4409

Using the fact that dj,q,k dmin,q , (12) can be written as

Pb,BB upper bound


Pb,TF upper bound
Pb,EX upper bound
Simulation
2

10

Pb

P (x
z|h)

 

2b 1
L
S j=1 h2j q=1 wj,q d2min,q

Q
2N0


2b 1
L
S j=1 h2j q=1 wj,q d2min,q

exp
4N0

L 2b 1 wj,q 2 2
S j=1 q=1
k=1 hj dj,q,k
exp
,
4N0

10

10

(14)

by using the relation Q( x + y) Q( x)ey/2 . By using


(14) in (13), we get

 
2b 1
L
S j=1 h2j q=1 wj,q d2min,q

Pb,EX (w|h) Q
2N0


2b 1
L
S j=1 h2j q=1 wj,q d2min,q

exp
4N0




S j,q,k h2j d2j,q,k
Edj,q,k exp
4N0
 

2b 1
2
2
S L
h
w
d
j,q
min,q
j=1 j
q=1
Q

2N0


2b 1
L
S j=1 h2j q=1 wj,q d2min,q

exp
4N0
L 2
1

b

j=1 q=1

j,q
w
(
q

Sh2j
),
4N0

(15)

10

11

12

13

14

15
16
E /N (dB)
b

17

18

19

20

Fig. 7. Simulation and analysis results for 2-state GBICM and L = 2


channels with 16-QAM signal sets and b = 2.

which is independent of i, u, etc. Equation (16) can be used


to bound the conditional error probability

b T (D1,1 . . . , N )
Pb,T F (h)
Sh2 d2

j min,q

2kc
N
4N0
N =1,Dj,q =e
(17)
and Pb can be obtained by limiting and averaging numerically
as shown in (10).
Figure 7 shows the performance of 2-state GBICM on
L = 2 channels of Example 1, along with the numerically averaged (using (10)) transfer function bound Pb,T F and expurgated bound Pb,EX . Also graphed is the Bhattacharya bound
Pb,BB from [12] which is obtained by ignoring the partial
modulation-induced correlation between successive codebits.
V. C ONCLUSIONS

where, q () is the characteristic function of pd (c c q)


from equation (4).
Note: The above analysis has to be modified if the nonnearest neighbors cannot be expurgated from the union bound
for the considered signal constellation. In that case, the BICM
Union Bound (BICM-UB) from [5] should be the starting
point for the analysis. We, however, ignore such a scenario in
this paper as it was argued in Section III that GBICM works
best with Gray mapping which lends itself to expurgation of
non-nearest neighbors.

C. Transfer Function Bound


Unlike ideal interleaving channels considered in [5], block
fading channels allow us to obtain transfer function based
bounds and approximations to the conditional error probabil
ity. Using the upper bound Q( y) 12 ey/2 [21] in (12),

L 2b 1
S j=1 q=1 wj,q h2j d2min,q
1
,
exp
P (x
z|h)
2
4N0
(16)

We have shown in this paper that when coding for block


fading channels, BICM in its conventional form may not
provide the best performance. We proposed a novel form of
coding called GBICM, which provided 12 dB improvement
in performance over BICM. A key aspect here is that implementing GBICM in a system that is already implementing
BICM is very simple. Only the interleaver has to be modified.
The same decoder can be used. We feel that GBICM is feasible
to be implemented in a practical system if the transmitter were
to know the kind of channel the receiver faces. Reasonable
performance improvements can be expected with GBICM
then. We have shown that the right measure of performance
on block fading channels is not the Hamming free distance,
but, the product Euclidean distance.
R EFERENCES
[1] E. Biglieri, Introduction to Trellis-Coded Modulation, with Applications.
New York: Maxwell Macmillan International, 1991.
[2] G. Ungerboeck, Channel coding with multilevel/phase signals, IEEE
Trans. Inform. Theory, pp. 5567, Jan. 1982.
[3] , Trellis-coded modulation with redundant signal setspart 1:
introduction, IEEE Commun. Mag., pp. 511, Feb. 1987.
[4] E. Zehavi, 8-psk trellis codes for a rayleigh channel, IEEE Trans.
Commun., vol. 40, pp. 873884, May 1992.

4410

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008

[5] G. Caire, G. Taricco, and E. Biglieri, Bit-interleaved coded modulation, IEEE Trans. Inform. Theory, vol. 44, pp. 927946, May 1998.
[6] V. Sethuraman and B. Hajek, Comments on Bit interleaved coded
modulation, IEEE Trans. Inform. Theory, vol. 52, pp. 17951797, Apr.
2006.
[7] P.-C. Yeh, S. A. Zummo, and W. E. Stark, Error probability of bit
interleaved coded modulation (BICM) in wireless environments, IEEE
Trans. Veh. Technol., vol. 55, pp. 722728, Mar. 2006.
[8] A. Martinez, A. G. i Fabregas, and G. Caire, Error probability analysis
of bit interleaved coded modulation, IEEE Trans. Inform. Theory,
vol. 52, pp. 262271, Jan. 2006.
[9] A. Chindapol and J. Ritcey, Design, analysis, and performance evaluation for BICM-ID with square QAM constellations in rayleigh fading
channels, IEEE J. Select. Areas Commun., vol. 19, pp. 944957, May
2001.
[10] X. Li, A. Chindapol, and J. Ritcey, Bit-interleaved coded modulation
with iterative decoding and 8PSK signaling, IEEE Trans. Commun.,
vol. 50, pp. 12501257, Aug. 2002.
[11] A. Ekbal, K.-B. Song, and J. M. Cioffi, Outage capacity and cutoff rate
of bit-interleaved coded OFDM under quasi-static frequency selective
fading, in Proc. IEEE GLOBECOM, 2003, pp. 10541058.
[12] A. G. i Fabregas and G. Caire, Coded modulation in the block fading
channel: coding theorems and code construction, IEEE Trans. Inform.
Theory, vol. 56, pp. 91114, Jan. 2006.
[13] P. C. Yeh, S. Zummo, and W. Stark, A union bound on the error
probability of binary codes over block fading channels, IEEE Trans.
Veh. Technol., vol. 52, pp. 20852093, Nov. 2005.
[14] R. Knopp and P. A. Humblet, On coding for block fading channels,
IEEE Trans. Inform. Theory, vol. 46, pp. 189205, Jan. 2000.
[15] E. Malkamaki and H. Leib, Coded diversity on block fading channels,
IEEE Trans. Inform. Theory, vol. 45, pp. 771781, Mar. 1999.
[16] , Evaluating the performance of convolutional codes over block
fading channels, IEEE Trans. Inform. Theory, vol. 45, pp. 16431646,
July 1999.
[17] M. Chiani, A. Conti, and V. Tralli, Further results on convolutional
code search for block fading channels, IEEE Trans. Inform. Theory,
pp. 13121318, June 2004.
[18] W. Song, B. Li, and Y. Li, High performance space time trellis coded
modulation, in Proc. IEEE Intl. Conf. on Micr., Ant., Prop. and EMC
Tech., 2005, pp. 11311134.

[19] W. Zhang, C. DAmours, and A. Yongacoglu, Trellis coded modulation


design for multiuser systems on AWGN channels, in Proc. IEEE Veh.
Tech. Conf., 2004, pp. 17221726.
[20] S. Nagaraj and M. Bell, Concatenated trellis coded modulation for
quasi-static fading channels, IEEE Trans. Wireless Commun., vol. 6,
pp. 32203228, Sept. 2007.
[21] S. B. Wicker, Error Control Systems for Digital Communication and
Storage, 1st ed. Prentice Hall, 1995.
[22] Y. S. Leung, S. G. Wilson, and J. W. Ketchum, Multifrequency trellis
coding with low delay for fading channels, IEEE Trans. Commun.,
vol. 41, pp. 14501459, Oct. 1993.
Santosh Nagaraj received his B. Tech. degree in
Electrical Engineering from the Indian Institute of
Technology, Madras, India, in 2000, and his PhD
degree from Purdue University, West Lafayette, IN
in 2005.
Since 2005, he has been on the faculty of San
Diego State University, San Diego, CA, where is
currently an Assistant Professor of Electrical and
Computer Engineering. His research interests are
primarily in the areas of communication system
design, broadband modulation and demodulation
techniques, and signal processing. He is currently working on adaptive
modulation and coding (AMC) for orthogonal frequency division multiplexing
(OFDM) systems with feedback and on design of efficient coding techniques
for block fading channels. His other research interests include multiple input
multiple output (MIMO) systems and power control.
Dr. Nagaraj has served as a reviewer for several prestigious journals such as
IEEE T RANSACTIONS ON C OMMUNICATIONS , IEEE T RANSACTIONS ON
W IRELESS C OMMUNICATIONS, etc. He has served as a reviewer for several
professional conferences as well. He has also served on the technical program
committee for IEEE W IRELESS T ELECOMMUNICATIONS S YMPOSIUM.
In addition to his active research and professional activities, Dr. Nagaraj
has been a dedicated teacher committed to excellence in undergraduate and
graduate teaching. In 2007, he received the Outstanding Faculty Award from
San Diego State Universitys College of Engineering.

You might also like